Skip to main content

Full text of "DTIC ADA322037: Support of Environmental Requirements for Cloud Analysis and Archive (SERCAA): Final Report."

See other formats

19970307 060 



Gary B. Gustafson 
Robert P. d'Entremont 
Ronald G. Isaacs 

Atmospheric and Environmental Research, Inc. 
840 Memorial Drive 
Cambridge, MA 02139-3794 

14 August 1996 
Final Report 

14 August 1992 - 13 August 1996 




Directorate of Geophysics 



"This technical report has been reviewed and is approved for publication." 

Contract Manager 

Acting Chief, Satellite Analysis and Weather 
Prediction Branch 
Atmospheric Sciences Division 

DONALD A. CHISHOLM. Acting Director 

Atmospheric Sciences Division 

This report has been reviewed by the ESC Public Affairs Office (PA) and is releasable 
to the National Technical Information Service (NTIS). 

Qualified requestors may obtain additional copies from the Defense Technical 
Information Center (DTIC). All others should apply to the National Technical 
Information Service (NTIS). 

If your address has changed, or if you wish to be removed from the mailing list, or if the 
addressee is no longer employed by your organization, please notify PL/IM, 29 
Randolph Road, Hanscom APB, MA 01731-3010. This will assist us in maintaining a 
current mailing list. 

Do not return copies of this report unless contractual obligations or notices on a specific 
document requires that it be returned. 

Di ciAnai Nom 



Form Approved 

0MB No. 0704-0188 

Public reDomna burden for this collection of information ts estimated to a»/erage t hour per response, including the lime for reviewing instructions, searching existing data sources. 
oath#rinaand maintaininq the data needed, and completing and reviewing the colleaton of information. Send comments regarding this burden estimate or any other aspect of this 
collection of information including suggestions for reducing this burden, to Washington Headquarters Services. Oireaorate for information Operations and Reports. 12IS Jefferson 
Oal^H?ghwa?. Su?Jl l204!^^^^ and to the Office of Managen?ent and Budget. Paperwork Reduction Projea (0704^ 188). Washington. DC 20503. 


14 August 1996 | Final 14 August 1992-13 Aug.1996 


Support of Environmental Requirements for Cloud 
Analysis and Archives (SERCAA): Final Report 


PE 35160F 





G.B. Gustafson, R.P. d'Entremont and 

R.G. Isaacs 


Atmospheric and Environmental Research, Inc. 

840 Memorial Drive 

Cambridge, MA 02139-3794 




Phillips Laboratory 

29 Randolph Road 

Hanscom AFB, MA 01731-3010 

Contract Manager: Allan Bussey/GPAB 





Sponsored by joint Department of Defense, Department of Energy, and 
Environmental Protection Agency Strategic Environmental Research and 
Development Program (SERDP) 


Approved for public release; distribution 


13. ABSTRACT (Maximum 200 words) 

Support of Environmental Requirements for Cloud Analysis and Archives (SERCAA) was a 
four-year, two-phase research and development program. Program goals were to produce a 
global multispectral, multisensor cloud analysis capability for retrieval of spatial, radiative, 
microphysical, and environmental cloud properties through analysis of sensor data from multiple 
operational environmental satellites. Program accomplishments during Phase 1 included 
development of three customized cloud-detection algorithms for analysis of NOAA/AVHRR, 
DMSP/OLS, and geostationary sensor data; a new algorithm for discrimination of vertical cloud 
layers within a 25 km grid cell; and an innovative analysis-integration algorithm for optimdly 
combining the results of the source-specific cloud analysis modules. During Phase 2, retrieval 
capabilities were expanded to include cloud particle size, emissivity/transmissivity, optical depth, 
and altitude. Additional work toward characterization of cloud-free bidirectional reflectance 
provides a basis for enhanced analysis of visible-channel data in the cloud detection process. 

SERCAA algorithms have been selected for operational implementation at the Air Force Global 
Weather Central. 


cloud, nepnanalysis, satellite meteorology, cirrus 






Unclassified Unclassified Unclassified 



NSN 7540-01*280-5500 Standard Form 298 (Rev 2-89) 

Pfe«f«be<J by AN$i Sid Z39-18 




Executive Summary.viii 

1.0 Introduction. 1 

2.0 Program Requirements. 2 

3.0 SERCAA Phase I Overview. 3 

3.1 DMSP/OLS. 4 

3.2 NOAA/AVHRR. 6 

3.3 Geostationary. 8 

3.4 Cloud Layering and Typing. 9 

3.5 Cloud Analysis Integration. 10 

4. SERCAA Phase n. 12 

4.1 Cloud Detection Using Visible and Near-Infirared Bidirectional 

Reflectance Distribution Models. 12 

4.1.1 Spectral Tests for Cloud-Clearing. 14 Bright Clouds Test. 14 Cold Cloud Tests. 14 NIR-Visible Ratios. 15 Low Clouds and Fog. 15 Thin Cirrus. 15 Cloud Shadows. 16 Adjacency Effects. 16 

4.1.2 BRDF Models. 17 

4.1.3 BRDF Retrieval Estimates..'.. 17 

4.1.4 Cloud Detection Results. 18 

4.2 Cirrus Radiative and Spatial Properties. 21 

4.2.1 Observation of Cirrus from Satellite. 21 

4.2.2 Passive Irrfrared Physics of Cirrus Cloud Signatures. 22 

4.2.3 Cirrus Retrievals Using 6.7-p,m Water Vapor Imager Data. 24 

4.2.4 Cirrus Retrieval Validation Results. 24 

4.2.5 Summary. 26 

4.3 Optical and Microphysical Parameters for Water Droplet Clouds. 29 

4.4 Cloud Environment. 34 

4.5 Implementation Changes for Cloud Typing and Layering. 40 

4.6 Analysis Consistency. 41 

4.7 Daytime Cirrus/Low-Cloud Discrimination Using MWIR Data. 49 




5. Future Research Topics. 50 

5.1 New Sensors. 50 

5.2 Advanced Retrieval Techniques. 52 

5.2.1 Interpretation of Visible to Midwave-Infrared Data. 52 

5.2.2 Transmissive Cirrus over Warm and Reflective 

Backgrounds. 54 

5.2.3 Cloud Typing. 55 

5.2.4 Cloud Altitude. 55 

5.2.5 Temporal Consistency. 57 

5.3 Validation. 59 

6. Summary. 61 

7. References. 61 

Appendix A AER SERCAA Related Papers and Technical Presentations for 

1993-1996 . 69 



Figure Page 

1 The New England Cartesian grid used for the BRDF study. 14 

2 AVHRR sample image valid 1221 UTC 12 Sept 1994 (left), and its 

associated manually digitized cloud mask generated using Eqs. (l)-(6). 17 

3 17-Day composite cloud mask for AVHRR morning overpasses 

between 3-20 September 1994. 19 

4. Map of Ambrals kernel combinations obtained for the 17-day AVHRR 

Sept 94 dataset. 20 

5 TPQ-11 35-GHz radar returns and coincident satellite-based SERCAA 
MWIR/LWIR retrievals for cirrus clouds over Hanscom AFB, 2315 

UTC 15 Sept 95. 25 

6 TPQ-11 35-GHz radar returns and coincident satellite-based SERCAA 
Water-Vapor/LWIR retrievals for cirrus clouds over Hanscom AFB, 

2315 UTC 15 Sept 95. 27 

7 Plots of SERCAA and RTNEPH-derived cirrus altitudes for the cirrus 

samples shown in Figures 5 and 6. 28 

8 Modified gamma droplet size distributions for stratocumulus (dot,dot) 
and altostratus (solid). Distributions are fi:om Stephens (1979) and 

Diem (1948), respectively. 30 

9 Plots of nighttime MWIR/LWIR brighmess temperature differences as 
a function of cloud top altitude for the mid-latitude summer 

atmosphere. 31 

10 Plots of nighttime MWIR/LWIR brighmess temperature differences as 
a ftmction of satellite view angle for altostratus in a tropical 

atmosphere. 32 

11 Schematic of a prototype infi’ared cloud detection scheme that 

integrates microwave derived information. 37 

12 Typical Tb signatures for clear over ocean (solid), partly cloudy over 
ocean (dash-dot), cloudy over ocean (dash), clear over land (solid), 
partly cloudy over land (dash-dot), and cloudy over land (dash) for 
data collected during May and July, 1992 for the east and west coast 

of the United States. 39 



Figure Page 

13 Typical Tb signatures for clear (dash), light precipitation (solid) and 
heavy precipitation (dot) observed in Typhoon Oliver, February 4, 

1993 in the western Pacific Ocean. 39 

14 NOAA-14 10.7-pm brightness temperature image over New England 

for November 1994. High tones denote low brightness temperatures. 42 

15 SERCAA Phase-1 cloud layer analysis for the image in Figure 14. This 
analysis was generated using LWIR brightness temperatures as input 
to the clustering and unsupervised classification routines. Each square 
represents a 16th-mesh box. White boxes contain two cloud layers, 

and gray boxes contain only one. 43 

16 SERCAA Phase-2 cloud layer analysis for the image in Figure 14. This 
analysis was generated using SERCAA Phase-2 pixel-resolution 
corrected altitudes as input to the clustering and unsupervised 
classification routines. Each square represents a 16th-mesh box. 

White boxes contain two cloud layers, and gray boxes contain only 

one. 44 

17 SERCAA evaluation regions of interest. 45 

18 SERCAA analyses of total cloud firaction over the western Pacific 

Ocean and eastern Asia on 26 March 1993. Top row contains hourly 
analyses starting at 1700 UTC and ending at 2000 UTC. Bottom row 
shows a simple persistence forecast based on the analysis at 1600 
UTC. Greyshade is proportional to analyzed cloud amount. 48 



Table Page 

1 BRDF cloud detection accuracy (%), measured against the manually 

digitized 1-km cloud mask. 21 

2 Regression coefficients for Eq. (14) that yield droplet mode radius as a 
function of the MWIR-LWIR brightness temperature pair and the 

satellite look angle, for each model atmosphere. 33 

3 Sensor data suite. 35 

4 Summary of operational environmental satellite imaging sensor 

characteristics. 46 

5 Planned environmental satellite systems with primary sensor 

characteristics. 51 


Executive Summary 

This final report describes a two-part research and development program to produce a 
global multispectral, multisensor cloud analysis capability for retrieval of spatial, radiative, 
microphysicd and environmental cloud properties through analysis of sensor data from 
multiple environmental satelhtes. Jointly sponsored by the Department of Defense, 
Department of Energy, and Environmental Protection Agency through the Strategic 
Environmental Research and Development Program (SERDP) and the Air Force P hilli ps 
Laboratory (PL), the Support of Environmental Requirements for Cloud Analysis and 
Archive (SERCAA) program achieved the project goals through development and testing of 
a series of cloud analysis algorithms. A key attribute of the SERCAA algorithms is then- 
ability to analyze and integrate data from multiple environmental satellite systems consisting 
of the U.S. military (DMSP) and civilian (NOAA/POES) polar orbiting sensor platforms 
and international geostationary systems from the U.S. (GOES), Japan (GMS), and Europe 
(METEOS AT), "nie program was broken into two development phases. Phase I focused 
on analysis of imaging-sensor data for cloud detection and retrieval of cloud spatial 
properties including location, extent, fractional amount, cloud-top altitude, and cloud type. 
Important elements are multiple cloud detection algorithms designed to best exploit the 
information content from each class of satellite, a new layering and typing algorithm based 
on generalized image processing classification techniques, and a unique analysis integration 
algorithm that optimally combines analyzed cloud products from each data source into a 
single consistent analysis. Phase n used the cloud spatial information as the first step in 
retrieval of cloud emissivity, phase, and particle size. Emissivity algorithms use 
multisj^tral infrared data to simultaneously solve multiple radiative transfer equations for 
the emission term. From emission, a first-order estimate of optical depth inferred. For 
accurate estimation of cloud emission, it is first necessary to know cloud phase. A 
deterministic phase algorithm was developed that exploits unique cloud signatures at mid¬ 
wave and long-wave infrared wavelengths from liquid water and ice cloud. While imaging 
sensors provide insufficient information to retrieve a cloud drop distribution, a useful bulk- 
property parameterization (e.g., mode radius) can be retrieved. A mode-radius retrieval 
algorithm was develop^ for water droplets based on theoretically-derived look-up tables of 
expected monochromatic infrared radiance values as a function of atmospheric profile, view 
angle, and cloud altitude. Monochromatic radiances were convolved with sensor response 
functions for IR channel confi^rations on available environmental satellites. Where 
available, collocated sounding instruments were used in addition to the imager data to 
retrieve temperature and moisture profile information in the vicinity of cloud. A unified 
retrieval algorithm was used wherein aU available sensor information is used to produce an 
optimal retrieval. 

1.0 Introduction 

A global cloud detection and analysis technology, operating continuously and in real¬ 
time, which more completely exploits the information in present satellite sensors, has been 
a recognized and stated requirement to support the operations of the United States 
Department of Defense (DoD) for a number of years. Also, the environmental monitoring 
and global change research communities have identified clouds as having a preponderant 
impact on climate and climate change. They have given top priority to improved global 
cloud specification although real-time cloud product generation is not a requirement for 
these groups. Progress firom the global change community has occurred, specifically the 
International Satellite Cloud Climatology Project (ISCCP). However, fiscd and 
technological constraints have prevented the development and implementation of a multi¬ 
satellite multispectral cloud an^ysis technology capable of utilizing in an operational mode 
the abundant sensor information presently on orbit. The need has been identified, the 
measurement capabilities have been in place, the missing component has been an advanced 
satellite-based operational cloud retrieval technology. 

This report describes the Support of Environmental Requirements for Cloud Analysis 
and Archive (SERCAA) research and development program which has produced such a 
technology and provides a technical description of the component algorithms. The program 
was accepted by the DoD portion of the Strategic Environmental Research and 
Development Program (SERDP) and supported under the Remote Sensing Thrust as a 
project of the “Definition and Demonstration of Remote Sensing Capability to Contribute to 
Environmental Understanding and Support for Environmental Issues”, administered by the 
Defense Support Projects Office (DSPO) through a contract with the Air Force Phillips 

The overall SERCAA program objective was to provide a satellite-based global cloud 
analysis capability for determining the spatial, radiative and microphysical properties of 
clouds. The two-phases of the program have resulted in development and demonstration of 
cloud algorithms for retrieval of: (1) continuously updated cloud spatial properties, and (2) 
augmented cloud optical, microphysical, and environmental properties. The combined 
algorithms form a cloud analysis model that is intended to serve a large number of 
environmental users, both within and outside of DoD. The SERCAA model can be used to 
help in determining and validating the distribution of clouds and their associated 
hy^ological and radiative effects on climate and on global climate change. It will also 
constitute a substantial upgrade to the presently operating Real-Time Nephanalysis 
(RTNEPH) cloud analysis model at the Air Force Globd Weather Center (AFGWC). In 
this role SERCAA is a fundamental component in the Cloud Depiction and Forecast System 
(CDFS) being carried out by the Air Weather Service of the U. S. Air Force. 

Under Phase I, the objective of improved satellite cloud detection capabilities for 
global applications was addressed through development of an integrated ensemble of global 
cloud andysis algorithms applicable to sensor data from both civilian and military satellites. 
The SERCAA multiplatform, multisensor, multispectral cloud product analysis algorithms 
(Isaacs et al., 1993; Gustafson et al., 1994) incorporate high-resolution sensor data into a 
real-time cloud analysis. Multispectral cloud analysis techniques are used for the detection 
and specification of clouds, especially cirrus and low clouds. Geostationary sensor data 
augment polar-orbiting data to improve the timeliness of analyses, particularly in tropical 
regions. The steps required to process the raw sensor data, collected from each of the 
satellite platforms, into each of the individual cloud analysis products include total cloud 
algorithms for the Defense Meteorological Satellite Program (DMSP) Operational Linescan 
System (OLS), the National Oceanic and Atmospheric Administration (NOAA) Advanced 
Very High Resolution Radiometer (AVHRR), and geostationary platforms; cloud layer and 


type algorithms; and a new analysis integration algorithm. SERCAA data products 
include: total cloud cover fraction, number of cloud layers (up to four floating layers), 
cloud layer coverage fraction, cloud type, cloud height, and analysis confidence level. 
These products will be provided on a 16* mesh (24 km grid tme at 60 degrees latitude) 
polar stereographic grid. 

In Phase n, the goal was to extend the multiple-source optimal satellite retrieval 
strategy that was successfully apphed for cloud spatial characteristics (cloud amount, 
height, layers, and type), to Ae augmented cloud properties (optical, microphysical, 
environmental) specification. This was accomplished by emphasizing infrared and 
microwave sounder data in conjunction with cloud detection from visible and infrared 
satellite imagers. Optical properties are emissivity, reflectivity, and optical depth; 
microphysical attributes are particle size distribution and phase; environmental parameters 
are water vapor profiles, temperature profiles and cloud liquid water. 

2.0 Program Requirements 

In the U.S. Global Change Research Program, the role of clouds has been identified 
as the number one science priority and also the most important environmental factor 
(Correll, 1990) affecting global climate change processes. The reliable determination of 
cloud parameters on a global scale is fundamental to understanding that role (Schiffer and 
Rossow, 1983). The U.S. Interagency Committee on Earth and Environmental Sciences 
(CEES) has also specified clouds and the hydrologic cycle as elements sharing the highest 
scientific priority for global change research. The two-decade-long Global Energy and 
Water Cycle Experiment (GEWEX), part of the World Climate Research Programme 
(WCRP), has singled out the cloud-radiation feedback process as primary among its most 
critical problems (WMO, 1989). GEWEX emphasizes that satellite detection of cloud 
physical and radiative properties on a global basis will require significant enhancement if 
the goal is to achieved of specifying the complex role clouds have in climate and rlimatp! 
change (Chahine, 1992). Also, the American Meteorological Society Council (AMS, 

1992) has stated, “Clouds are the most critical factor causing uncertainties in model 
estimates of global warming.” 

The requirements for detailed representative global cloud analyses have increased 
over the last decade. Advancement of cloud forecasting models and the accurate simulation 
of cloud effects on the radiative feedback processes in global/medium-range numerical 
weather prediction (NWP) and climate models is impeded by the lack of accurate cloud 
analyses (Isaacs et al., 1983). Applied meteorology’s goal of improved short and medium 
range cloud forecasting is limited in significant part to Ae accuracy of cloud parameteriza- 
tions employed by forecast modelers. In the summer of 1983, the ISCCP started as a part 
of the WCRP. One of the prim^ objectives of the ISCCP is to derive and validate a 
global cloud climatology that will allow for improved parameterizations of clouds and 
therefore their associated radiation feedback processes in climate models (Schiffer and 
Rossow et al., 1985). 

The impact of clouds on Department of Defense (DoD) activities is also of first order. 
Improved cloud forecasting and improved cloud data handling and analysis are among the 
highest prioritized Air Force requirements. The DoD policy of quantifying the 
environment’s impact on new systems before they enter the acquisition process has placed 
increased emphasis on having accurate and multi-parameter cloud archives against which 
system performance can be evaluated. The need for superior global cloud analysis and 
archive products is critical. Presently the best, and for certain cloud characteristics, the 
only operational real-time global cloud analysis is the RTNEPH (Hamill et al., 1992), 


residing at the Air Force Global Weather Central (AFGWC). It became operational in 
1983, replacing the earlier 3-Dimensional Nephanalysis, or 3DNEPH. Unfortunately, the 
cloud analysis techniques and database management contained in the RTNEPH make 
minimal use of existing advanced technologies. Currently, only Operational Linescan 
System (OLS) visible and infrared channel radiances, or single-channel Advanced Very 
High Resolution Radiometer (AVHRR) infrared data, are accepted for processing. Even 
this limited data set is then degraded in both spatial and spectral resolution and 
geometrically transformed to fit a standardized fixed grid structure. Restrictions on the 
satellite data sources made available for processing and the amount of ad hoc tuning and 
manual interaction required by the current RTNEPH introduce serious limitations for 
climate research purposes. Additionally, RTNEPH products are immediately available only 
to government, and in particular DoD weather forecast facilities. RTNEPH information is 
archived at the National Climatic Data Center (NCDC), Asheville, NC by the United States 
Air Force Combat Climatology Center (USAFCCC). From a climate standpoint, the 
quality and therefore the usefolness of these archive products is limited by the lack of data 
source information and the lack of continuity in analysis algorithms in the RTNEPH over 
the past two decades. 

Recent advances in satelhte cloud detection using multispectral analysis techniques 
have demonstrated a superior quantification of cloud characteristics utilizing near-in^ed 
(NIR) and medium-wavelength infrared (MWIR) data from the AVHRR sensor. For 
example, the location, height, amount, and optic^ properties of cirms clouds can be better 
quantified; low cloud and fog can also be more confidently detected at night and during the 
da 5 ^me over melting snow (d’Entremont, 1986; d’Entremont and Thomason, 1987; 
d’Entremont et al., 1990; Stowe et al., 1991; d’Entremont et al., 1992). Application of 
these retrieval techniques was successfully achieved within DoD under the TACNEPH 
development project conducted by researchers at the Phillips Laboratory, Geophysics 
Directorate (PL/GPA) and AER (Gustafson and d’Entremont, 1992). The objective of 
TACNEPH was to produce a stand-alone, deployable satellite cloud analysis capability on a 
regional scale to support tactical operations (Gustafson et al., 1993b). The TACNEPH 
heritage lies in the global Real Time Nephanalysis or RTNEPH model (Hamill et al., 1992) 
run operationally at the Air Force Global Weather Central (AFGWC). 

Incorporation of improved satellite cloud detection capabilities for global applications 
was a primary goal of SERCAA. The three principal accomplishments of SERCAA were: 

1) to incorporate high-resolution sensor data from multiple military and civilian satellites, 
polar and geostationary, into a real-time cloud analysis model, 2) to demonstrate 
multispectral cloud andysis techniques that improve the detection and specification of 
clouds, especially cirrus and low clouds, and 3) to provide augmented parameter, 
algoritlun, and data base specifications for an improved cloud retrieval model including 
cloud environmental, radiative, and microphysical properties. 

3.0 SERCAA Phase I Overview 

The focus of the Phase I effort (Gustafson et al., 1994) was development of an 
integrated nephanalysis model to retrieve cloud spatial properties using imaging sensor 
measurements from three categories of environmental satelUtes as the primary data source. 
NOAA Polar Orbiting Environmental Satellites (POES) carry the five-chaimel AVHRR 
imaging sensor. DMSP satellites use the two-channel OLS. The constellation of 
international geostationary satellites include two United States Geostationary Operational 
Environment^ Satellites (GOES) which carry five-channel imagers, the Japanese 
Geostationary Meteorological Satellite (GMS) which uses a four-channel instrument, and 
the European Space Agency’s Meteorological Satellite (METEOSAT) with a three-charmel 


sensor. Data from each category of satellite are analyzed using separate cloud detection 
algorithms specifically designed to exploit the specific information content of each system. 
Satellite sensor data are processed as diey are received from the Satellite Data Handling 
System by the CDFS n. For polar-orbiting satellites, the basic processing unit is one orbit 
of data, for geostationary satellites the amount of data received from any one satellite varies 
according to a data acquisition schedule defined by the operating nation(s). 

The SERCAA nephanalysis model consists of four processing levels. Level 1 is 
ingest processing that includes unpacking of the telemetry stream, sensor data calibration, 
and Earth location. Level 2 is cloud detection performed on a pixel-by-pixel basis using 
sensor-specific algorithms to analyzed data transmitted from each satellite. Level 3 is cloud 
layering and typing to provide a vertical stratification of the cloudy pixels detected in Level 
2. Level 3 output is remapped to the standard AFGWC polar-stereographic grid projection 
at a resolution of 16^ mesh. Level 1,2, and 3 are event driven processes triggered by the 
receipt of new data from any of the satellite sources. Level 4 is merge processing wherein 
the most recent analyzed products from each satellite are combined to produce a single 
optimal integrated analysis. When implemented in CDFS H, a fifth level of processing will 
be added that uses hourly integrated cloud analyses to initialize a short-term and a long-term 
cloud forecast model. Levels 4 and 5 will be clock driven processes run hourly under a 
user-defined schedule. 

Three cloud detection algorithms are used to analyze data from the different imaging 
sensor systems utilized in CDFS n. Each cloud algorithm is designed to exploit the unique 
strengths of the respective sensor system. DMSP/OLS data have the highest spatial 
resolution and a constant instantaneous field of view (IFOV) across a scan line but are 
limited to two broad band channels, one visible and one infrared. NOAA/AVHRR data 
have a relatively coarse spatial resolution that degrades toward the edge of scan balanced by 
five iiarrow band channels with high spectral resolution. Geostationary sensors vary by 
satellite but have at least one visible and one infrared channel in common. GOES carries a 
five-channel imager, GMS a four channel, METEOSAT three, and INSAT two. 


The OLS two-channel radiometer has broad-band channels in the visible to near 
infrared (0.40-1.10 |xm) and the long wave infrared (10.5-12.6 ^im). Analysis of OLS 
data is accomplished through a dual threshold technique that operates on two dimensional 
visible-infrared data during daytime and on single-channel IR data at night. Separate cutoff 
thresholds are established to discriminate completely cloud free and cloud filled pixels. 
Threshold determination, and hence algorithm accuracy, is critically dependent on an ability 
to accurately predict the background visible and infrared radiance that the sensor would 
measure in the absence of cloud. An independent analysis of surface skin temperature 
produced by an analysis model such as SFCTMP (Kopp et al., 1994a) is used as a first 
guess to this dear-column infrared brightness temperature prediction. The first guess is 
adjusted by using a nearby cotemporal satellite measurement to establish a reference 
temperature correction to account for differences between the modeled and satellite 
observed teinperatures. Once established, the reference temperature correction is used to 
adjust the skin temperature field to predict the dear-column brighmess temperature for all 
pixels within a local analysis region. 

The reference temperature correction factor calculation requires collocated IR sensor 
data and modeled skin temperature fields. Satellite data are segmented into a series of 
small, local analysis regions. The size of each region is determined by the relative spatial 
scales of the AFGWC skin temperature database and the satellite sensor data. For each 


cloud analysis grid box a separate surface skin temperature correction is computed by 
obtaining Ae difference between the observed satellite brightness temperature for a 
reference pixel and the corresponding collocated and time interpolated skin temperature. If 
the reference pixel can be established to be cloud free then the reference temperature 
difference is used to predict the clear scene brightness temperature for all remaining pixels 
in the analysis region. To maximize the likelihood that the reference pixel is cloud free it is 
selected as the pixel with the largest IR brightness temperature within the analysis region. 

The critical step in estimation of the clear scene brightness temperature is insuring that 
the reference pixel is free from cloud contamination. This is done by comparing the 
magnitude of reference temperature difference against the range of expected temperature 
differences for a given time, location and satellite. Statistics on clear scene temperature 
differences are produced as byproducts of the nephanalysis algorithm and are maintained in 
a database stratified by location, surface type (land, water, desert), satellite, and time of 
day. Cloudy pixels are eliminated by running the nephanalysis algorithm using 
conservative cloud detection thresholds. Assumed natural variability of this quantity for a 
given satellite pass is calculated from a time weighted average of difference statistics 
accumulated over the previous ten days. If the reference temperature difference lies within 
the averaged hmits, it is assumed to fall within the naturally occurring variability for cloud 
free conditions and the reference value is assumed to be cloud free. 

If the reference pixel can be established to be cloud free, then the predicted clear scene 
brightness temperature for any pixel within the analysis box is computed by adding the 
reference temperature difference to the time interpolated AFGWC surface skin temperature 
at the desired pixel location. If the reference pixel is determined to be cloud contaminated 
then a default correction based on the time weighted mean of the historical temperature 
differences is used as the reference temperature. 

Visible channel dear-scene statistics are maintained globally on an 8* mesh grid 
using a technique developed for the RTNEPH (Kiess and Cox, 1988). Similar to the IR 
procedure described above, visible-channel clear scene statistics are computed as a 
byproduct of the nephanalysis. For each 8* mesh grid cell, all clear scene visible-channel 
pixel values are accumulated and the mean calculated. The visible background database is 
updated continuously as new data are received. To avoid anomalous vdues from 
corrupting the database, information is added only if the new value does not vary from the 
magnitude of the stored value by more than a preset threshold. If the newly computed 
mean reflectance value passes this criteria then the database is updated using a weighted 
average of the previous and new values. Thus the clear column visible background 
database contains a continuously updated record of the satellite observed clear column 
visible count at each grid cell location. 

The OLS cloud algorithm is based on a statistical threshold approach designed to 
operate using either a single infrared thermal window channel alone or in combination with 
a visible channel. Each requires the predicted clear scene bandpass weighted brightness 
temperature and, during daytime, the visible background values described above. An 
empirically derived dynamic correction factor is used to collectively account for all sources 
of error in the predicted dear-scene values without the need to understand and quantify the 
individual contributions. Error sources include sensor calibration errors, incomplete 
modeling of atmospheric transmission, and poor or unrepresentative estimates of the dear- 
scene skin temperature. 

Separate cloud detection criteria are used for processing combined visible-infrared 
data and infrared data alone. Use of visible data is restricted to conditions where the solar 
zenith angle is less than a preset threshold that is defined to discontinue visible-data proces- 


sing well before the daylight terminator. Objective processing of daytime OLS visible data 
beyond this cutoff is problematic because of rapid changes made to visible sensor gain 
control performed on-board the satellite. Visible sensor gain is controlled by the value of 
the scene solar elevation (i.e., the solar elevation at the viewed point on the Earth). Gain 
changes are relatively smooth at high solar elevations, however, as the scan approaches the 
terminator the granularity of incremental changes in the solar elevation increases resulting in 
rapid changes in sensor gain adjustment (Bunting and d'Entremont, 1982). Under daytime 
conditions, when usable visible and infrared data are available, a bispectrd algorithm is 
executed. During nighttime conditions a single channel IR test is used. 

The OLS single channel IR algorithm is a three step procedure: 1) predict the clear 
scene bandpass weighted brightness temperature; 2) compare satellite observed IR 
brightness temperatures to the predicted temperature; and 3) based on the magnitude of the 
difference between the two temperatures, classify each pixel as either cloud-filled or cloud- 
free. The dynamic threshold approach described above is used to predict the brightness 
temj^rature that the satellite would measure for a clear scene. Two separate thresholds are 
applied to the corrected temperature to account for any remaining uncertainties in the clear 
scene estimate. One threshold is used to establish a cutoff value for completely cloud filled 
pixels and the second for completely cloud free. Data points that lie between the two cutoff 
values are treated as cloud filled but the analysis is assumed to be less reliable. 

The two channel OLS algorithm is an extension of the single channel approach into 
two spectral dimensions. When usable visible data are available, data from both OLS 
sensor channels are analyzed simultaneously using two pairs of cutoff values, one pair for 
each channel. Pixels witih lower brightness temperatures than the IR cutoff and/or higher 
brightness values than the visible cutoff are classified as cloud. Warm bright regions 
require an additional a priori background type classification to remove ambiguity caused by 
similarities in the radiative signatures of desert, snow or ice backgrounds with low cloud. 
Data points that fall between all four cutoff values are classified as cloud filled but are 
considered less reliable. 


The AVHRR sensor is a five-channel radiometer with two channels sensitive to 
reflected solar, chaimels 1 centered at 0.6 |im and 2 at 0.9 pm; two channels that measure 
only emitted IR energy, channels 4 at 10.5 |xm and 5 at 12.0 |im; and one channel that is 
pnsitive to both reflected solar and emitted IR, channel 3 at 3.7 pm. The multispectral 
information content from these channels was used to develop a cloud algorithm that is less 
dependent on accurate characterization of the cloud-free background than the OLS 
algorithm. The AVHRR multispectral algorithms are based on the approach of Saunders 
and Knebel (1988) wherein a hierarchy of cloud detection and background discrimination 
tests, each exploiting a different cloud radiative signature, are run in series on each pixel 
within the scene. The algorithms are capable of operating on all five AVHRR channels 
simultaneously or on any combination of channels. A cloud analysis is produced through 
interpretation of results from all available cloud and background tests. 

Separate snow and sun glint tests are used to identify problematic background 
conditions. Generally, tests that rely on reflected solar radiation (i.e., AVHRR channels 1, 
2, or 3) can misclassify sun glint, snow, or ice as cloud. The main exception are tests that 
use channel 3 data over frozen water backgrounds since ice and snow are relatively non- 
reflective at these wavelengths, however, sun glint affects all three channels. Some 
information on snow cover and sea ice is available from the SNODEP model and from the 
Navy, however, experience has demonstrated that these databases have insufficient 


accuracy to support the nephanalysis. To supplement the available databases snow, ice and 
sun glint algorithms were developed using the multispectral sensor data alone. Snow and 
ice are detected by comparing the solar component of daytime channel 3 data to channel 1 
visible data. Cloud surface tend to be reflective at both wavelengths, however snow and 
ice surfaces have a significantly lower reflectivity at the channel-3 mid wave IR (MWIR) 
wavelengths than in die visible. Thus for cloud, both channels will exhibit a strong 
reflected solar signature while for snow or ice surfaces the signature is much greater in the 
visible channel. Similar to snow and ice, sun glint from water surfaces can produce a 
cloud-like signature in the visible to MWIR channels. Specular reflection from water 
surfaces occurs when the sensor is scanning toward the sun and the satellite zenith angle 
matches the solar zenith angle. Although these angles can be calculated with sufficient 
precision to identify the specular point, accurate identification of sun glint from water 
surfaces remains a problem. This is due to the well-known occurrence of sun glint well 
away from the specular point under conditions when the water surface is agitated. Since 
information on sea state and wind direction is not direcfly available, a multispectral sun 
glint algorithm was developed to test for conditions that resemble cloud in the reflected 
solar tests but do not in tests that rely on emitted IR radiation only. However, this 
condition alone is not sufficient to detect glint since some liquid water clouds can exhibit 
the same characteristics. An additional criteria requires that the magnitude of the channel 3 
data approach sensor saturation, a condition that normally occurs only in ghnt conditions 
since the derived channel 3 brighmess temperature is very sensitive to even small amounts 
of reflected solar. 

AVHRR cloud tests generally rely on channel differences or ratios to discriminate 
cloud signatures from those of terrestrial back^ounds. Due to the unique radiative charac¬ 
teristics of low clouds and fog at 3.7 |im relative to long wave IR channels, comparison of 
channel 3 and 4 brightness temperatures is a powerful low cloud discrimination technique 
during both day and nighttime. At 3.7 |im liquid water clouds radiate as gray bodies (i.e., 
both emit and reflect energy) whereas at 10.5 ^m they are nearly black (d'Entremont et al., 
1987). As a result, liquid water cloud emissivity can be significantly less than 1 resulting 
in a lower nighttime equivalent blackbody brightness temperature computed from channel 3 
relative to channel 4. However, during the daytime the combined emitted and reflected 
solar components cause the computed brighmess temperature to be relatively large 
compared to channel 4 where there is only emitted radiation. Additionally, for a partially 
filled IFOV (e.g., cold cloud and warm background in the same IFOV) the relative 
contribution to the derived brighmess temperature from warm and cold surfaces differs 
between me channels 3 and 4 due to me highly non-linear shape of me Planck function 
between me two channel wavelengms. At 3.7 |jm me emitted energy from me warm 
background is me dominant contribution to me computed brighmess temperamre whereas at 
11 pm me relative contributions from me background and cloud are roughly equal. This 
signature is useful for detecting broken and transmissive high (cold) cloud at night, 
particularly optically thin cirrus. 

Omer cloud tests use relative differences between me split visible (1 and 2) and split 
long wave IR channels (4 and 5). Relative visible and near IR clear scene albedo measure¬ 
ments will differ depending on background. Over water bom channel reflectances tend to 
be low, but enhanced atmospheric aerosol scattering at me shorter channel 1 wavelengths 
generally results in a slightly higher scene reflectance. Over land me signature reverses 
except in cases of extreme aerosol loading since vegetated surfaces reflect preferentially in 
me longer channel 2 wavelengms. Clouds reflect approximately equally in bom channels 
and as such tend to obliterate the background signatures. Thus a split visible cloud test 
checks for roughly equal channel 1 and 2 reflectance values. However, me absolute 
magnirnde of the measured channel 1 and 2 radiances can vary significantly over a scene 
depending on me relative reflectivity of the observed surface, solar geometry, and 


anisotropic effects making selection of a channel-difference threshold problematic. To 
cancel these effects out of the cloud detection algorithm the ratio of the two channels is used 
as opposed to a channel difference to discriminate the background signatures from the 
cloud. A cloud signature is assumed to be a channel ratio of approximately 1.0. Split IR 
data (channels 4 and 5) are used to detect ice cloud and small particles along cloud edges, 
independent of time of day. Even in cloud-free conditions, inter-channel brightness 
temperature differences are expected to exist due primarily to preferential water vapor 
absoiption at channel 5 wavelengths. However, in the presence of cloud the differences 
increase beyond the level predicted theoretically. Inoue (1987) recognized that this 
departure was caused by unequal extinction from thin ice clouds at 11 and 12 jim, with the 
^eater extinction occurring at 12 )nm. Prabhakara et al. (1988) extended this signature to 
include both liquid water and ice clouds when the droplet or particle size was smaller than 
the channel wavelength. Saunders and Kriebel (1988) developed a test to exploit these 
signatures through a theoretically derived look-up table of expected clear scene channel 
differences. When the channel differences exceed the theoretically predicted amount cloud 
is assumed to be present in the volume. 

3.3 Geostationary 

The cloud detection algorithm for geostationaty satellite platforms employs a hybrid 
approach to identify cloudy pixels within an analysis scene. The approach consists of three 
algorithms that ^e applied in series: temporal differencing, dynamic thresholding, and 
spectral discriminant tests. The procedure is applicable to any geostationary satellite with at 
least one visible and one long wave IR channel. 

The first level of processing utilizes a temporal differencing technique to identify new 
cloud development or cloud motion over either previously clear background or lower level 
cloud. The algorithm exploits the high temporal resolution of geostationary data by testing 
for rapid changes in infrared brightness temperature and/or visible reflectance in collocated 
pixels taken from sequential satellite images. Cloud detection is accomplished by identi¬ 
fying pixels that change by an amount greater than the dear-scene conditions are expected 
to change. Expected changes in dear-scene background characteristics are computed from 
time series of surface skin temperature and visible background brightness databases. 

During daytime conditions a bispectral technique is employed that simultaneously evaluates 
the change in infrared brightness temprature and visible reflectance. At night only the 
change m IR brightness temperature is evaluated. A pixel is flagged as cloud filled if the 
change in collocated pixel values over a one hour time period exceeds the expected change 
by greater than a predetermined threshold amount. Otherwise, the pixel is flagged as cloud 

The temporal difference test is only expected to identify some fraction of the cloud in 
a scene, however the confidence level for these clouds is very hi gh The dynamic 
thresholding algorithm uses information on the radiative characteristics of cloud identified 
through temporal differencing to detect nearby cloud with similar characteristics. The 
image is segmented into a regular grid and the minimum and maximum brightness tempera¬ 
tures (visible reflectance) of the pixels classified as cloudy by the temporal differencing test 
within each grid cell are identified. The maximum temperature is used to define an infrared 
brightness temperature cloud threshold for the remaining pixels within a cell. The 
threshold is set at the maximum temperature minus an offset to eliminate anomalous values. 
Separate cloud thresholds are calculated for each grid cell and all pixels within the cell with 
a lower brightness temperature or higher visible count are classified as cloudy. 


Following the temporal differencing and dynamic threshold processing, it is still 
possible that clouds with spatial and spectral attributes that have not changed over the 
analysis time interval remain undetected. The final step in the geostationary processing 
applies a series of spectral discriminant tests to the remaining unclassified pixels. Spectral 
discriminant tests use sensor channel information in the same manner as the OLS and 
AVHRR algorithms to detect spectral signatures of cloud. Since geostationaiy sensors have 
different configurations depending on platform, the suite of tests employed depends on the 
set of data channels available. Since aU geostationary imaging sensors have at least one 
visible and one long wave IR chaimel, single-channel visible and IR threshold tests can be 
applied universally. 

The recently launched GOES-8 and 9 satellites, along with GMS-5 are the first in a 
series of next-generation geostationary systems that have introduced multispectral imaging 
sensors that have potentid for significantly improving geostationary cloud detection. New 
channels include two split-window LWIR channels at 10.2 -11.2pm and 11.5 - 12.5pm, 
and on GOES, a mid-wavelength infrared (MWIR) channel at 3.8 - 4.0pm. The new satel¬ 
lites provide all channels simultaneously on a full-time basis. The geostationary algorithm 
exploits these new channels, when available, through the addition of multispectral cloud 
tests similar to the AVHRR tests described above. The algorithm automatically determines 
which tests are applicable based on the individual sensor configurations. 

3.4 Cloud Layering and Typing 

Vertical stratification of the cloud data into layers is accomplished using a single cloud 
layering algorithm for all geostationary and polar orbiter satellite data. The layering 
algorithm only operates on LWIR channel data from pixels previously classified as cloudy 
by the appropriate Level 2 processing algorithm. A three-step procedure is used: first a 
maximum likelihood classification algorithm is applied to separate the LWIR brightness 
temperature data into thermally homogeneous layers. Next each pixel is assigned a cloud 
type using local spatial information. The final step consists of accumulating die pixel data 
into 16* mesh grid cells, clustering of the pixels into local cloud layers, and assigning a 
predominant cloud type to each layer. 

Cloud layering begins by stratifying the cloudy LWIR brightness temperature data 
over a large region (on the order of a few thousand scan lines) using a generic 
unsupervised clustering algorithm. The clustering process defines homogeneous 
groupings within the spectral domain of the input data. User input controls the minimum 
spectral distance between pixel temperatures and the cluster centroid but the maximum 
number of layers is left a free parameter. Once clusters are defined a Bayesian maximum 
likelihood classifier is used to assign each pixel to a cluster. Clusters are now treated as 
vertical slices through the atmosphere based on an inverse relationship between cluster- 
mean brightness temperature and height. 

Cloud typing occurs by analyzing data one layer at a time starting from the top-most 
(coldest) layer. Each pixel and its contiguous neighbors are grouped using a region 
growing routine. The number of contiguous pixels in each group is used to assign a cloud 
type classification of either stratiform or cumuliform. A number threshold is used to make 
Ae assignments such that groups with pixel counts below the threshold are classified 
cumuliform and those above as stratiform. The region growing and type assignment 
process continues for each group in the layer until all pixels assigned to that layer by the 
maximum likelihood classifier have been processed. Subsequent processing moving from 
high to lower (cold to warmer) layers is accomplished by combining all the pixels from 
higher layers with the pixels of the current layer before the region growing process occurs. 


Cloud type assignment occurs as before but only for pixels in the current layer, pixels in 
higher layers that were previously assigned a cloud type retain that classification even if the 
new grouping places them in a different category. The inherent assumption is that 
cumuhform clouds have a greater vertical extent than stratiform while stratiform have the 
larger horizontal range. By processing one layer at a time from the top down, features such 
as cumulus towers embedded in a lower stratiform deck will be correctly classified. 

The final processing step occurs after all cloudy pixels have been assigned a cloud 
type. In this step cloudy pixel LWIR data are first accumulated into 16* mesh grid cells by 
simply identifying, based on latitude and longitude, which grid cell each pixel belongs to. 
Umupervised clustering of the LWIR data is then repeated but this time on a floating 
window of 3X3 grid cell arrays and with the maximum number of clusters set to four. 
Individual pixels are then assigned to each cluster using a maximum likelihood classifier, 
but only for the pixels in the center grid cell of the 3X3 array. The moving 3X3 window 
of 16* mesh grid cells is used to establish local cloud cluster attributes for the center grid 
cell that mitiimize artificial discontinuities across grid cell boundaries. The populated 
clusters are interpreted as cloud layers and cluster statistics are used to compute fractional 
cloud amount, mean cloud top temperature, and predominant cloud type for each layer. 

Note that in this process each grid cell will have a maximum of four cloud layers. In this 
way each 16* mesh grid cell is populated, cloud layers defined, and layer statistics 
computed. Level 3 output parameters for each grid cell include the valid time of the satellite 
data, total cloud amount and, for each layer, fractional cloud amount, mean cloud top 
temperature, and a stratiform/cumuliform designation. 

3.5 Cloud Analysis Integration 

As discussed above, separate cloud analysis algorithms are used to analyze cloud 
spatial properties from each sensor system. At the end of Level 3 processing, three 
independent gridded cloud analyses have been produced: one derived from DMSP/OLS 
data, one from NOAA/AVHRR data, and the third from the constellation of geostationary 
sensors. The analyses at each cell in the three 16* mesh hemispheric grids may have a 
different valid time based on the time of the input satellite measurements. To create the 
integrated analysis, analyzed cloud data from each of the three analysis grids are input into 
M analysis integration algorithm to produce a single, optimal analysis based on the relative 
timeliness and accuracy of the input data. 

The general conceptual approach to the inte^ation problem is one that utilizes both 
rule-based concepts as well as principles from statistical objective analysis (Lorenc, 1981; 
H^ll and Hoffman, 1993). The unique nature of the satellite-derived cloud parameters 
drive the way in which the cloud analysis data are combined. For example, some cloud 
parameters such as cloud type and number of layers are discrete quantities and cannot be 
"averaged" in any physically meaningful way. Computational concerns also argue for 
applying rule-based ideas which allow the preferential selection of one satellite analysis 
over dl others thereby avoiding weighted averaging of the data as much as possible. 
Additionally, the tradeoff between the relative timeliness and accuracy of the input analyses 
is a major consideration in the development of the integration algorithms. For instance, 
analyses derived from DMSP or AVHRR visible and infrared imagery are potentially more 
accurate than those obtained from lower spatial resolution geostationary data. However, 
there may be large time gaps between consecutive analyses derived from polar orbiting da t a 
while new geostationary analyses are generally available at least once per hour. 

Integration of total cloud amount precedes integration of layer quantities since the 
estimates of total cloud fraction are believed to be more reliable than any individual layer 


fraction (due to small sample sizes and the potential for height assignment errors). 
Processing occurs independently for each grid cell. First a series of rules are applied to 
determine if any one of the input analyses is superior to the other two. If none of the input 
grids have been updated since the time of the previous Worldwide Merged Analysis then 
die previous analysis is persisted. If new analyses are available, a check is made to 
determine if more than one are timely. If only one timely analysis is available, the merged 
total cloud fraction is set to the value of this analysis. If more than one analysis satisfies 
timeliness requirements, these analyses are examined to determine if they are all either 
completely cloudy or completely cloud free. If so, total cloud fraction is set to either 100 or 
0 percent, respectively. If multiple timely analyses exist that are neither all completely clear 
nor completely cloudy, then an estimated error of each sensor analysis is used to determine 
if the most recent analysis also has the lowest estimated error. Only when all these 
conditions fail is an 01 algorithm used to obtain a blended estimate of total cloud fraction 
from multiple input analyses. Averaging weights for the OI are based on estimated analysis 
errors computed for each available sensor an^ysis. 

Analysis errors for each sensor analysis are computed assuming an empirically 
defined initial analysis error plus an additional error growth function that is linear with 
time. Values for initial analysis error and error growth rate are sensor-dependent and are 
main ta ined for both polar orbiting and geostationary platforms. These are tunable 
parameters, and as such can be used to adjust the merge processing to correct for 

Once integration of total cloud fraction is complete, merging of all other layer 
parameters is performed. A rules-based procedure similar to diat used for total cloud is 
used to determine if the layer parameters from one analysis is superior to the others. In 
cases of only one timely analysis, all analyses indicating clear conditions, or when the most 
timely analysis is also the most accurate, Aen layer cloud parameters are set in the same 
way as total cloud amount described above. However, when all timely analyses indicate 
100 percent cloud cover or in cases where the OI technique is used to determine total cloud 
amount the integration of layer parameters requires a more complex algorithm. This is 
because the vertical distribution of cloudiness and type are determined independently for 
each analysis and, given the different sensor characteristics, are likely to be different. In 
these cases one sensor analysis is selected as most reliable and is designated as a master or 
template profile into which all other timely analyses are blended. For discretely varying 
quantities such as number of layers and cloud type the integrated analysis profile will 
simply assume the values of the master analysis. For the continuously varying parameters 
of layer cloud fraction and cloud top temperature, an OI blending is performed by 
combining layers that most closely match in cloud top temperature. 

The integration algorithm is designed to use additional information provided by 
multispectral observations from the NOAA/AVHRR and multispectral geostationary 
sensors. In particular, cloud algorithms developed for these sensors are assumed to better 
detect the presence of low cloud and transmissive cirrus. Once the integration of all cloud 
parameters has been performed, the analysis grid box is then checked for the possible 
addition of these cloud types. This is achieved by imposing separate, less strict timeliness 
constraints for these special-case clouds. This effectively extends the length of time that 
these observations will affect subsequent merged analyses. Similarly, conventional, 
surface-based observations of low cloud are also used to extend the length of time satellite- 
based low-cloud analyses are used in the merged analysis. 

Final output parameters for each 16* mesh grid cell are now computed. Total cloud 
amount is the total number of cloudy pixels in the grid cell divided by the total number of 
pixels. Layer fractional cloud amount is computed in the same way as total cloud for the 


coldest (top) layer. Cloud top altitude is computed by interpolating temperature/height 
profile information to the mean cloud top temperature. Cloud type is assigned the most 
prominent cloud type in the master analysis. A representative time is assigned as the time 
of the most recent sensor-specific analysis used to produce the merged analysis. 

4. SERCAA Phase II 

Existing cloud analysis technology does not combine the cloud information measured 
by diverse sensors. For example, the input satellite data which support the SERCAA 
Phase I algorithm suite (Gustafson et al., 1994) utilizes only cloud imager data. 
Furthermore, outputs of the SERCAA Phase I algorithm treat only cloud spatial properties 
including cloud coverage, cloud layering, and cloud type. The success of overall cloud 
depiction and forecast improvements for short and extended ranges is dependent on 
characterization of atmospheric state. The SERCAA Phase I algorithms provide the basis 
to address requirements for improved specification of cloud initial state for short range 
forecasts by improving initialization of cloud macrophysics: cloud cover, layers, and 
heights. For these reasons one objective of the Phase n effort has been to identify 
approaches to synergistically combine information from both cloud imagers and other 
sensors such as sounders aboard current platforms to enhance overall cloud analyses. 

These algorithms complete the cloud depiction with enhanced parameters such as 
cloud environment, microphysics, and radiative parameters which are required for cloud 
forecasting over longer time scales. 

4.1 Cloud Detection Using Visible and Near-Infrared Bidirectional 
Reflectance Distribution Models 

The Earth's surface varies widely on a global scale, from deserts to tropical forests to 
permanent snow and ice cover. Vegetated surfaces and snow cover also vary for a given 
Earth location as a function of time, both diumally and seasonally. Surface-based 
observations of these characteristics are extremely limited both spatially and temporally, 
arid are of limited use in monitoring changes in the Earth's surfaces over longer term, 
climatic time scales. Satellite sensor observations provide the only truly global 
measurements of the Earth's surface. Thus observations and monitoring of the Earth's 
surface are best provided by satellite radiance measurements. 

The manner in which a particular surface type reflects incident solar energy (as a 
function of solar position and the direction from which that surface is viewed) is a 
fundamental property of the surface. This prope^ is depicted by means of a bidirectional 
reflectance distribution function (BRDF), which is a mathematical representation of a 
surface's dependence of reflected energy on the illumination and viewing geometries. 
BRDFs describe the manner in which surfaces reflect incident solar radiation as a function 
of the satellite view and solar illumination geometries. Knowledge of BRDFs allows for 
directional correction (normalization) of satellite radiance data via anisotropic correction 
factors to a standard view and illumination geometry. Such a capability helps to reduce 
errors in computed derivatives from visible and near-infrared radiance data such as 
vegetation indices. BRDFs are of importance in remote sensing for computing location- 
dependent albedos of the Earth's surface on a global scale. 

Modeled BRDFs allow for prediction of visible and near-infrared satellite radiances 
for a particular location on the Earth's surface prior to the actual measurements of the 
satellite radiances themselves. Thus, the BRDF-predicted dear-scene anisotropic 


correction factors are of particular use in the SERCAA cloud detection model for analyzing 
visible data from multiple satellite observations of the Earth's surface over varying times 
and viewing geometries. 

In order to estimate an Earth-surface BRDF, a series of reflected radi^ce 
measurements from the surface over a range of viewing and solar illumination geometries 
are required. It is important that these observations not be influenced (contaminated) by 
clouds or their shadows. Thus corrections are performed to ensure that the cloud-free 
AVHRR radiance observations used to compute surface BRDFs contain only surface 
effects. In this study, BRDF models are retrieved to estimate the dear-scene radiance for 
each 1-km pixel in a series of AVHRR satellite image scenes of New England during a 17- 
day period in September 1994. Each scene has differing solar illumination and satellite 
viewing geometries. The observed radiance in each AVHRR pixel is compared to its 
corresponding BRDF-predicted dear-scene radiance for the appropriate view and 
illumination geometries and any value that exceeds the predicted clear-scene value is 
flagged as cloud. Cloud detection results are then compared digitally to manually cloud- 
cleared AVHRR data taken over New England. 

T his study has five main steps: (1) collocation of all AVHRR data to a common 1-km 
Cartesian grid projection using orbital prediction models and visual landmarks (e.g., Cape 
Cod) observable in the images to map the data; (2) cloud clearing using spectral signature 
cloud detection tests similar to those of Saunders and Kriebel (1988) and Gustafson et al. 
(1994); (3) correction of the clear-scene 0.63-|j,m visible and 0.86-p.m near-infrared 
AVH]^ radiances for atmospheric scattering effects using a coupled surface-atmospheric 
radiative transfer model; (4) fitting six semi-empirical BRDF models to the cloud-cleared 
and atmospherically corrected clear-scene radiances; and (5) application of the BRDF 
models to cloud detection. 

The NOAA AVHRR dataset used in this study consists of 17 morning-descender 
orbits of NOAA-10 and NOAA-12 between 3 and 20 September 1996. These data are first 
mapped to a 400 x 402 Cartesian grid centered over New England, shown in Figure 1. 

Grid resolution is 1.1 and 0.75 km in the column (north-south) and row (east-west) 
directions, respectively, corresponding to the nominal along- and cross-track AVHRR 
HRPT scanner FOV resolutions. Choosing these grid attributes minimizes pixel over- and 
undersampling during the remapping process. Using orbital prediction models and manual 
rubber-sheeting software, the two-dimensional remapping residuals for over 500,000 
pixels were on average only 0.3 km, or approximately 30 percent of pixel size. This high 
remapping accuracy minimizes any BRDF retrieval inaccuracy that may be caused by pixel 
collocation errors. 

The visible and NER data are next atmospherically corrected using surface visibility 
observations to define the aerosol profile in the MODTRAN atmospheric radiative transfer 
code. In general, atmospheric scattering effects cause the upwelling satellite-measured 
reflectance to be less than that of the surface itself. Next are described the cloud-clearing 
tests used on the New England data. 

4.1.1 Spectral Tests for Cloud-Clearing 

Seven tests that exploit the spectral attributes of the AVHRR data were used in this 
study to detect cloud and cloud-contaminated pixels. Each test was run individually on a 
pixel-by-pixel basis. If any of the tests detected cloud, then that pixel was considered 
cloudy and not used in subsequent BRDF clear-scene characterizations. In the following 


Figure 1 . The New England regional area analyzed in the BRDF study. 

paragraphs, Ai denotes the AVHRR Chi 0.63-iim visible albedo; A 2 denotes the Ch2 
0.86-pm near-infrared (NIR) albedo; T 3 denotes the 3.7-|im Ch3 middle-wavelenglh 
infrared (MWIR) brightness temperature; and T 4 and T 5 denote the 10.7- and 1 l.S-pm 
Ch4/Ch5 long-wavelength infrared (LWIR) brightness temperatures, respectively. The 
nadir resolution of all five AVHRR channels is 1.1 km. 

4.1.1 Spectral Tests for Cloud-Clearing 

Seven tests that exploit the spectral attributes of the AVHRR data were used in this 
smdy to detect cloud and cloud-contaminated pixels. Each test was run individually on a 
pixel-by-pixel basis. If any of the tests detected cloud, then that pixel was considered 
cloudy and not used in subsequent BRDF dear-scene characterizations. In the following 
paragraphs, Ai denotes the AVHRR Chi 0.63-p.m visible albedo; A 2 denotes the Ch2 
0.86-jxm near-infrared (NIR) albedo; T 3 denotes the 3.7-pm Ch3 middle-wavelength 
infrared (MWIR) brightness temperature; and T 4 and T 5 denote the 10.7- and 11.8-|im 
Ch4/Ch5 long-wavelength infrared (LWIR) brightness temperatures, respectively. The 
nadir resolution of all five AVHRR channels is 1.1 km. Bright Clouds Test 

The visible brightness, or “obvious bright” cloud detection test is a single-channel 
threshold test that is used to discriminate relatively high cloud albedos from a uniformly 
low-albedo New England background. If 

Ai > Ai,Land, (1) 


then the pixel is cloudy. Ai,Land was chosen to be 0.05, which is a liberally high estimate 
for vegetated surfaces. This’ test is designed to ensure that clouds with obviously high 
reflectivities do not go undetected by subsequent spectral tests. Cold Cloud Tests 

The “obvious cold” cloud detection test is a single-channel LWIR threshold test that is 
used to discriminate very cold clouds from a warm background. If 

T4 < TLand, ( 2 ) 

then the pixel is cloudy. Ttand was chosen to be 273 K, which for September in New 
England is liberally low. Like the bright test (Eq. 1), this test is designed to ensure that 
clouds with obviously low temperatures do not go undetected by subsequent spectral tests. NIR-Visible Ratios 

This ratio test compares the relative magnitudes of visible and NIR albedos using an 
interchannel ratio. For clouds, the AVHRR Chi and Ch2 albedos are very close to each 
other, while for vegetated land surfaces the Ch2 NIR albedos are much higher then the 
corresponding Chi visible albedos. Thus if 

Rlo < A 2 / Ai < Rni, (3) 

then the pixel is cloudy. Clear-land ratios generally have values greater than one, while 
water ratios have values less than one. In this study Rlo = 0-80 and Rhi =1.25 generated 
very good cloud detection results. The ratio test defined by Eq. (3) also detects coastlines 
and sub-pixel-sized ponds and lakes quite well, since they are mixed fields of view (i.e., 
water and land). Low Clouds and Fog 

The low cloud and fog test exploits the different radiative characteristics of water 
droplet clouds in the MWIR and LWIR spectral regions. Low clouds effectively reflect 
M\\TR solar energy as well as emit MWIR thermal energy, while at LWIR wavelengths the 
clouds essentially only emit thermal energy. Since the upwelling MWIR radiance contains 
both a solar and terrestrial emission component, the result is that the Ch3 observed 
brightness temperature for a daytime low cloud is higher than the corresponding Ch4 
brightness temperature. Thus if 

T 3 - T 4 > 5 K, (4) 

then the pixel is classified as cloud-filled. All liquid water droplet clouds will reflect 
enough solar MWIR energy to make T 3 higher than T 4 in the daytime. Thin Cirrus 

Split-window LWIR T 4 - T 5 brightness temperature differences exhibit small but 
persistent thin cirrus cloud signatures. There are three radiative effects that combine to 


provide this result: (1) ice particle emissivities are slightly lower at 11.Slim than at 10.7|xm; 
(2) atmospheric water vapor attenuation is stronger at 11.8|xm; and (3) there is a slightly 
stronger dependence of Planck radiances on temperature at 10.7)i.m, so that for thin cirrus 
pixel radiances that have contributions from warm backgrounds and cold clouds, the 
10.7p.m T4 is higher for any given combination of cloud and underlying surface 
temperatures, all other effects neglected. Thus if 

T4 - T5 > Tci, (5) 

then the pixel is classified as cloud-filled. For this study the thin-cirrus LWIR brightness 
temperature difference Tq was chosen to be between 2 and 2.5 K, with higher differences 
for longer atmospheric path lengths and higher water vapor concentrations. Cloud Shadows 

Cloud shadows as well as clouds contaminate upwelling surface-reflected energy, so 
their effects must be removed from the AVHRR data before using them to retrieve surface 
BRDF models. Cloud shadows have distinctively low NIR albedos in comparison to 
vegetated land albedos at 0.86p,m. Thus if 

A 2 < A2,Land, (6) 

the pixel is classified as containing cloud shadow, or a pond or lake. For this study, 
A2,Land was chosen to be 0.042. Adjacency Effects 

To ensure that the scattering effects of cloud edges into adjacent cloud-free pixels do 
not enter into the Earth-surface BRDF model calculations, any pixel that is adjacent to a 
cloud edge is flagged as cloud-contaminated and is not used in subsequent BRDF 
retrievals. Figure 2 illustrates the cloud mask obtained for the NOAA-12 AVHRR image 
collected at 1221 UTC on 12 September 1994. The mask represents the combined results 
of each of the cloud tests, and in this case masks out 48 percent of the land pixels in the 

4.1.2 BRDF Models 

Each of the BRDF models tested under this study is semi-empiiical in nature; such a 
model has the general form 

I — fiso + fgeo Kgeo + fvol Kvol> (7) 

where I is the directionally dependent radiance; the scalars fx are multivariate regression co¬ 
efficients whose values are influenced by the sub-pixel morphologic and spectral reflec¬ 
tance attributes of the vegetated surface; and the kernels Kx are deterministic predictors that 
carry information pertaining to the basic shape of the BRDF, and are analytic functions of 


Figure 2. AVHRR sample image valid 1221 UTC 12 September 1994 (left), and its 
associated manually digitized cloud mask generated using Eqs. 1-6. 

the solar illumination and satellite viewing geometries. Kernels are independent of wave¬ 
length and the parameters fiso, fgeo. and fyoi- If the surface being viewed reflects isotropi¬ 
cally, I is constant for all observations at a ^ven solar zenith angle and Eq. (7) would 
reduce to I = fjso (the iso subscript denotes isotropic), independent of view and illumination 
geometries. The component of diffuse reflectance by the vegetated surface is represented 
fgeo> and includes such effects as shadowing by vegetation canopy elements arranged in 
particular geometric fashions. Scattering of radiation (e.g., by leaves or pine needles) 
within the canopy volume is represented by fyoi- In general, a BRDF that is estimated 
using satellite measurements will be a nontrivial linear combination of these isotropic, geo¬ 
metric, and volume-scattering components. Alternatively stated, the fx parameters in Eq. 

(7) will be nonzero. In theory, the relative magnitudes of these parameters designate the 
comparative importance of the three reflective effects that define the overall characteristics 
of the BRDF of a surface (after Roujean et al., 1992). The kernels Kx are approximations 
to deterministic, physical models of radiative transfer in vegetated canopies. Accordingly, 
BRDF models of the form given by Eq. (7) are semi-empirical in nature. For a given scene 
they describe the association between observed cloud-free radiances and the view and 
illumination geometries under which those radiances were measured. Although the 
association is (strictly speaking) statistical in nature, the hope is that the deterministic origin 
of the kernels argues for the cause of association in a more physically rigorous maimer. 

Using the cloud-cleared and atmospherically corrected New England AVHRR LAC 
data taken during September 1994, six semi-empirical BRDF models were tested on a 
pixel-by-pixel basis for their ability to accurately characterize the directional dependences 
observed in the satellite radiance measurements. The models are (1) the Ross Radiative 
Transfer (RT) model (Ross and Marshak, 1985) for thick canopies; (2) the Ross RT model 
for thin canopies; (3) die Roujean geometric optical (GO) model with no mutual shadowing 
(Roujean et 1992); (4) the Li GO model for dense canopies, which is based on the Li- 
Strahler GO model (Li and Strahler, 1992) with mutual shadowing, driven by illuminated 
crowns; (5) the Li GO model for sparse canopies, which is based on the Li-Strahler GO 


model (Li and Strahler, 1992), driven by shadows on the ground; and (6) the modified 
Walthall model (based on W^thall et al., 1985). Each of Siese BRDF models has the 
general form specified by Eq. (7). 

4.1.3 BRDF Retrieval Estimates 

Cloud-free AVHRR data were then fit to the BRDF models to determine their respec¬ 
tive abilities to describe the observed directional dependence of the radiances. Recall that 
the kernels Kx in Eq. (7) are analytic functions of scene-solar and satellite zenith and azi¬ 
muth angles. Each cloud-free AVHRR radiance I for a given pixel is first associated with 
its set of kernels given that pixel's view and illumination geometries. Next the 17-day full 
set of these radiance-kernel triplets is assembled for each pixel in the New England 402 x 
400 Cartesian grid. Finally, the fx parameters in Eq. (7) are computed in a least-squares 
sense to provide the best available overall fit to the semi-empirical BRDF models. Compar¬ 
ative goodness-of-fit measures include statistical R2 hypothesis testing that demonstrates 
both the strength (via test statistic) and significance of association (via the magnitude of R) 
between the observed directional AVHRR radiance measurements lobs the overall 
BRDF models that predict lobs as a function of the satellite view and solar illumination 
geometries (Eq. 7). 

Figure 3 shows the cloud-clearing results for the 17-day New England dataset. In 
order to determine the semi-empirical BRDF model that best describes surface anisotropic 
reflection attributes, a minimum of eight clear- pixel observations must be available for a 
given pixel. Only 35 percent of the pixels in the New England grid have this minimum 
number of cloud-free observations. The majority of pixels with enough cloud-free radiance 
observations occurred in southern and eastern New England. Clouds inhibited BRDF 
retrievals primarily in the Green and White Mountains. 

Figure 4 shows Ae results of the best-fit BRDF models in the New England grid 
projection. At this point the best-fit multivariate regression coefficients fx, the kernels Kx, 
and the model choices “x” have been determined for each pixel. Thus a mathematical 
representation of the anisotropic reflectance attributes of each surface in the grid has been 
obtained. Tfre representation is written in the form given by Eq. (7) as an analytic function 
of the satellite view and solar illumination geometries, and dierefore can be used to predict 
the dear-scene reflectance of a pixel’s surface prior to observation of that surface by 
satellite. Next, the satellite-observed radiance will be compared to the predicted radiance 
and any values that exceed the predicted dear-scene value will be flagged as cloud. 

4.1.4 Cloud Detection Results 

The retrieved BRDFs were next used to cloud-clear an AVHRR test image taken 
during the same 17-day time period. First, Eq. (7) was used to predict the visible (Chi) 
and NIR (Ch2) satelUte-observed reflectances for a number of pixels in the test image. 
Next, the BRDF prediction was directly compared to the radiance observation of each test 
pixel. Any observed value that exceeds the predicted dear-scene value is flagged as cloud. 
Cloud detection results are then compared digitally to the manuaUy digitized AVHRR cloud 
mask obtained during the cloud-clearing process. 

Table 1 lists the results of the BRDF cloud clearing. Let “r” denote the BRDF dear- 
scene reflectance (which is linearly proportional to lobs Eq. (7)). When r is compared 
directly to the visible satellite observation, a correct determination of cloud presence is 
achieved 66 percent of the time. If the threshold r is allowed to vary to within 5 percent of 


Figure 3. 17-Day composite cloud mask for AVHRR morning overpasses between 3-20 
September 1994. 


AMBRALS KERNELS 0.63-0.86 um 

AVHRR 3-20 SEPT 1994 

|| R 0 S S “ T H I C K L I - S P A R S E 


Figure 4. Map of Ambrals kernel combinations obtained for the 17-day AVHRR Sept 94 

its actual value, cloud detection accuracy increases dramatically to 89 percent. Table 1 lists 
a nearly perfect cloud detection accuracy when r is allowed to vary to up to 20 percent of its 
BRDF-predicted value. Even higher accuracies are listed for AVHRR NIR Ch2, which 
usually has brighter backgrounds that are more easily and more often confused with cloud 
than those of AVHRR Chi. 

Table 1. BRDF cloud detection accuracy (%), measured against the manually digitized 

1 -km cloud mask. 











1 .20r 



These test results offer promise for single-channel, threshold-type cloud detection 
using reflected solar radiance data. BRDF techniques present alternatives to current state- 
of-the-art single-channel visible cloud detection algorithms that have trouble performing in 
scenes with Wgh solar zenith angles and surfaces &at exhibit strong anisotropic reflectance 
attributes, since the BRDF cloud^no-cloud thresholds explicitly depend on the solar illumi¬ 
nation and satellite viewing geometries. Current operational cloud detection models such as 
the RTNEPH (Hamill et al., 1992) attempt to define dear-scene background reflectances 
for polar-orbiting satellite data using a nearly constant value that does not account for 
variations in solar zenith angle and view geometries over the course of several days. 

The test results also hold promise for significantly increasing the amount of early- 
morning and late-aftemoon visible and NIR satellite data that are used operationally by 
cloud models such as the RTNEPH. This is because dear-scene reflectances at these times 
of day vary significantly for the smallest of changes in sun-satellite positions, so much so 
that constant-threshold background reflectivities simply cannot provide accurate cloud 
detection results. 

4.2 Cirrus Radiative and Spatial Properties 

The transmissive nature of cirrus clouds turns out to be its most important (in a 
climate sense) and elusive (in a retrieval sense) attribute to specify. If the semi-transparent 
nature of cirrus clouds is not accounted for, its altitude is consistently underestimated when 
using passive infrared brightness temperature data. Although it is generally agreed that 
cirrus has a net warming effect on climate, determination of the magnitude of this effect 
depends critically on the accurate specification of cirrus radiative and spatial attributes. For 
example, in the case of very thin (sub-visual) cirrus, ice particles have a more significant 
interaction effect with incident solar and upwelling thermal radiation than does upper 
tropospheric water vapor (Smith et al., 1990). 

4.2.1 Observation of Cirrus from Satellite 

There are many sources of passive satellite data that can be used to detect and analyze 
cirrus attributes. Among the earliest are visible and infrared data of the 1960s from the 
TIROS series of polar orbiting satellites, augmented in the early 1970s by geostationary 
GOES data. Current GOES-Next Imager and Sounder channels useful for detection of 


cirrus include 3.9, 6.7, 11.2, 12.7, and CO 2 13.3 - 14.5|ini spectral bands. The 3.9, 6.7, 
and 11.2 and 12.7|j,m channels will be discussed in detail shortly; the water vapor 
band has proven useful for detection of very thin cirrus over warm backgrounds such as 
deserts and tropical oceans, and has use in discriminating water-droplet clouds from ice 
clouds in daytime 3.7/6.7-|Xm imagery. 

More recent TIROS sensors include the Advanced Very High Resolution Radiometer 
(AVHRR), a five-channel passive radiometer with detectors that measure upwelling visible 
(0.63p,m), near-infrared (NIR, 0.86|xm), middle wavelength IR (MWIR, 3.7p,m), and split 
longwave IR (LWIR, 10.7 and 11.8|xm) energy both day and night. The sounder instru¬ 
ments collectively known as TOYS (JIROS Operational Vertical Sounder) also collect data in 
the wings of the ISjim CO 2 absorption band that are useful for detection of thin cirrus and 
specification of their height. There are also very high spatial resolution (500 m) Defense 
Meteorological Satellite Program data available in visible/NIR (0.4 - l.l^m) and LWIR (10 - 
12 |xm) bands that are helpful in ascertaining the small-scale spatial attributes of cirrus. 

4.2.2 Passive Infrared Physics of Cirrus Cloud Signatures 

The upwelling spectral thermal radiance lobs measured by a downward poin tin g 
radiometer for a field of view completely filled by a non-reflective, thin cirrus cloud is 

lobs ~ (1 ■ £) Isfc + £ Icld > (8) 

where e is the bulk cirrus emissivity, Igfc is the upwelling radiance emitted by the under¬ 
lying surface and clear atmosphere, and Lid includes the cirrus blackbody radiance plus the 
radiance emitted by the atmosphere above the cloud. In practice, reflection from cirms 
clouds at thermal infrared wavelengths is ignored. This is reasonable not only because the 
cinus bulk reflectivity is low, but also because there is only minor downwelling thermal 
emission incident on the top of cirrus clouds to be reflected back to space. In theory the 
specification of Lfc in equation ( 8 ) requires information on many of the physical properties 
of the atmosphere and surface that underlies the cirrus cloud: the temperature Tsfc and 
atinospheric traiismittance t (for water vapor attenuation) are two of the more important 
attributes. As discussed later, Lfc is specified using nearby measurements of cirrus-free 

The two unknowns of interest in Eq. ( 8 ) are the cirrus bulk emissivity £ and the cirrus 
Planck blackbody emission Lid. which is a known function of the cirrus effective 
temperature Tdd- In contrast there is only one known in Eq. ( 8 ), namely the radiance 
measurement Lbs- In order to specify these two unknowns, additional measurement 
information is needed. This is achieved first by considering Eq. ( 8 ) for simultaneous 
radiance measurements at two different infrared wavelengths. 

For purposes of discussion, assume that the radiance data are being measured 
AVHRR MWIR Channel 3 and LWIR Channel 4 sensors. The two cirrus radiance 
equations are then 

, Iobs,3 = (1 - £ 3 ) Isfc,3 + £3 Icld,3 


lobs,4 = (1 ■ £ 4 ) Isfc,4 + £4 Icld,45 

where the 3 and 4 subscripts denote the 3.7 and 10.7|j.m AVHRR Channels 3 and 4 
radiances, respectively. Eqs. (9a) and (9b) are two equations, but with the second equation 
a tlmd unknown £4 has been introduced. A third equation is needed that contains no new 
variables and that relates at least two of the three unknowns already established. This is 

by the 




done by assuming a relationship between the cirrus bulk optical depths 63 and 64 as 
follows. First, radiative transfer calculations are available that compute bulk cirrus optical 
depth as a function of wavelength and hexagonal ice particle size for varying cirrus cloud 
thicknesses (Takano and Liou, 1989; Hunt, 1973). Once computed, a simple linear 
regression between corresponding pairs of the two optical depths is performed to obtain a 
relationship of the form 

63 = m 84 +b , ( 10 ) 

where m and b are the regression slope and intercept, respectively. The slope m is 
nonzero; the intercept b, however, turns out to be very close to zero since the two optical 
depths are close to each other for optically very thin cirrus. Thus Eq. (10) is generally of 
the form 

53 = m54. (11) 

The value of the regression slope (m) is dependent on the effective ice crystal size which in 
turn is dependent on cirrus temperature (after Ou et al, 1993). Values range from m=2.603 
for Tcid= 210 K to m=1.088 for Tcid=253 K. 

Considering the radiative properties of citrus within a satellite field of view in a bulk 
sense, the cirrus cloud optical depth 8 is related to the citrus transmissivity t by the relation 

8 = -Int, ( 12 a) 

so that Eq. (11) can be rewritten 

lnt 3 = mlnU. ( 12 b) 

Since it is assumed that the citrus cloud is non-reflective, e +1 = 1 so that 

In (1 - £ 3 ) = m In (1 - £ 4 ) . ( 12 c) 

Finally, solving for £3 in terms of £4 yields 

£3 = l-(l-e4)“. (13) 

This is the third of the three-set equation, so that there are now three equations (9a), 
(9b), and (13) in three unknowns £ 3 , £ 4 , and Tdd- The three measurements consist of the 
linear regression slope m, and the satellite-measured radiances Iobs,3 and Iobs,4- This three- 
equation system forms the basis for the SERCAA cirrus retrieval techniques that analyze 
multispectral infrared satellite radiances. 

In practice, nearby cirras-free pixels are used to obtain accurate estimates of Isfc ,3 and 
Isfc,4- Atmospheric emission above the cirrus is neglected. Cirrus reflectivity is negl^ted 
as well. As previously mentioned, this does not introduce into the retrieval process a major 
source of error at night, but during the daytime incident solar radiances at the shorter 3.7- 
pm wavelengths noticeably affect cirrus radiance measurements. The daytime problem is a 
challenging one. Although cirrus reflectivities are relatively small, incoming 3.7-pm solar 
radiation is strong enough so that measured radiances are solar-contaminated. Subsequent¬ 
ly, Eq. (9a) is no longer accurate and the retrieval process becomes considerably more 
complex because of it. Thus,the use of Eqs. (9a), (9b), and (13) are presently restricted to 
nighttime scenes when there is no incident solar energy being reflected back to space by 
either the ciirus cloud itself or the underlying background. Research is ongoing to separate 
out the solar component in the 3.7pm daytime data (Ou et al., 1993). Another major 
constraint is that in assigning a single cloud temperature Tdd to the cirrus, it is assumed that 


the cloud is a thin sheet that lies precisely at one atmospheric level. Clearly this is not the 
case. Lidar backscatter returns from cirrus clouds consistently show their complex 
structure on both horizontal and vertical scales. In midlatitudes their altitudes range from 6 
to 13 km, and their thicknesses anywhere from 1 to 5 km and, on occasion, even higher. 
Thus the assignment of a single cirrus temperature is a gross one which in the case of 
multispectral infrared radiance retrievals results in the assigmnent of one effective cirrus 
cloud dtitude. However, the severity of this constraint affects only the cirrus altitude 
determination. Its effects on the bulk optical properties of the cloud are far less detrimental. 
Nonetheless, it is important to remember that current satellite-retrieved cirrus altitudes are 
not an accurate assessment of the true levels at which the cirrus lie, but rather are only 
correct in a radiatively bulk, energy-average sense. For this reason the cirrus temperature 
Tcid Md corresponding altitude ZcW are labeled as effective properties, since they afford 
little inference on the detailed vertical structure of the cloud. 

In practice Eqs. (9) and (13) can also be used with AVHRR channels 3 and 5. Thus 
for every triplet of AVIBIR infrared satellite radiance measurements Iobs,3» Iobs,4» and 
Iobs,5 it is possible to retrieve at sensor resolution the 3.7,10.7, and 11 . 8 ^m cIots bulk 
emissivities £3, £ 4 , £5 and optical depths 83, 84 , and 85 along with cirrus effective 
temperature (altitude) Tdd (Zdd). 

4.2.3 Cirrus Retrievals Using 6.7-p,m Water Vapor Imager Data 

In hght of the 3.7-)Lim daytime solar contamination issue, ongoing SERCAA studies 
are revealing that 6.7-|xm water-vapor channel data can be used in place of those taken at 
3.7pm. This finding is significant in that (a) the complexity of retrieving cirrus properties 
during daytime (due to mixed solar/thermal energy at 3.7-pm) is eliminated, and (b) there 
are 6.7-pm water vapor chaimels on all currently operational meteorological satellites, both 
geostationary and polar, save DMSP. Thus wiA the development of a 6.7-pm cirrus 
optical depth and effective altitude algorithm comes a capability for cirrus spatial and optical 
property analyses that are truly global in nature and that span the full diurnal cycle. The 
water-vapor cirms retrieval physics is identical to the MWIR retrieval physics in that Eq. 
(9a) is written for 6.7-pm (vice 3.7-pm) satellite radiance observations. Initial retrieval 
results over a 10 -case data set compare quite well with ground- and aircraft-based 
observations of cirrus clouds. 

4.2.4 Cirrus Retrieval Validation Results 

The figures that follow compare directly the SERCAA sateUite-based cirms retrievals 
with coincident ground-based 35-GHz active radar observations of cirrus base, top, and 
optical thickness. The ground-based data were collected by the PL Atmospheric Sciences 
Division TPQ -11 radar as a part of the Space-Based Infrared Systems (SBIRS) field 
experiment conducted over Hanscom AFB, MA and Madison, WI during mid-September 
1995. Three examples are shown of cirms altitude retrievals plotted against radar altitude 
observations for 2330 UTC at Hanscom AFB on 16 September 1995. Figure 5 has a 
predominantly blue background, is labeled G08_ORH_259_2315 across the top, and has 
the label 3.7-iim - 10.7-|im Emissivity in the vertical on the right-hand side. The colors on 
the shde are representative of the reflected power received as a function of altitude by the 
active 35-GHz radar, with warm colors (reds, oranges, yellows) corresponding to strong 
returns (optically thick clouds) and cool colors (green, cyan) corresponding to weaker 
returns (optically thin clouds). Dark blue denotes cloud-free radar observations. Yellow 
diamonds denote the SERCAA 3.7-10.7-|im satellite-retrieved cirms altitudes that 
correspond with their respective coincident radar observations. As can be seen, the 


Altitude (km) 

OGS ORH_259_2315 


Time (HHMM) 

^ ^altitude 


Figure 5 . SERCAA cirrus altitude retrievals using MWIR/LWIR-channel data compared 
to coincidence TPQ-11 35 GHz radar returns over Hanscom AFB, 2315 UTC 
15 Sept 95. 


3.7/xm — Emlsslvii 

retrievals match well with observation. Corresponding cirrus effective emissivities are 
plotted in red squares against the right-hand vertical (emissivity) axis. It is seen that the 
retrieved emissivities are higher where the radar returns are stronger and vice versa. 

Figure 6 is similar to the first, except it has the label 6.7-|xm - 10.7-|im Emissivity 
next to the vertical axis on the right-hand side. Yellow diamonds denote the SERCAA 6.7- 
10 .7-|im satellite-retrieved cirrus altitudes that correspond with their respective coincident 
radar observations. (The 3.7-|im data and the 6.7-|i,m data from these two examples are 
both from the GOES-8 Imager.) The water-vapor retrievals also match well with 
observation. However, there is a bias of sorts in the water-vapor altitude retrievals that 
weights them much more closely to the true tops of the transmissive cirrus, in comparison 
to the respective 3.7-|jm MWIR retrievals. This bias is due primarily to the fact that in the 
cirrus environment there is more atmospheric water vapor than in the dear-column (i.e., 
cirrus-free) regions. The net result is that the 6.7-|im cirrus brightness temperature 
observations are more strongly influenced by this excess water vapor than are the 
corresponding 3.7-|J,m MWIR measurements (since 3.7-)i.m is an atmospheric window 
region wherein water-vapor absorption effects are minimal). However, there are two 
useful attributes to this bias. The first is that for the SBIRS program, cirrus cloud top 
altitudes are of more value than the radiative center-of-mass altitudes that the 3.7-|im data 
provide. The second is that in differencing the MWIR and water-vapor altitude retrievals, 
information on cirrus physical thickness can be inferred that has to now been unobtainable 
using passive infrared retrieval techniques. 

Figure 7 is a simple plot of AVHRR-retrieved 3.7-|im cirrus effective altitudes with 
the corresponding 35-GHz radar observations. The radar-observed cirrus tops and bases 
are plotted as brown and cyan lines, respectively. The blue diamonds are the AVHRR 
altitude retrievals; they compare well with observation, as did their GOES Imager counter¬ 
parts. The green squares correspond to blackbody altitude retrievals obtained using 10.7- 
)j,m brightness temperatures that have been uncorrected for cirrus transmissive effects (i.e., 
assuming the cirrus is a blackbody cloud). This is the technique typically used by most 
cloud analysis models to compute cloud altitude for both opaque and transmissive clouds. 
Altitudes retrieved in this manner agree reasonably well with tme cirrus altitudes wherever 
cirrus transmissivities are low (optically thick cirrus), as is seen on the left side of the plot. 
However, where cirrus is optically thin, the comparison differences are dramatic: on the 
right half of the plot are citrus blackbody retrievals that differ by as much as 6 km from the 
true, SERCAA-retrieved cirrus altitudes. It is important to note that the green squares 
denote state of the art in the current Air Force operational global cloud analysis model, the 
RTNEPH (Hatnill et al., 1992), as well as in the Phase-I SERCAA algorithms being 
implemented as a part of CDFS-II. 

4.2.5 Summary 

As a part of SERCAA cloud analysis development efforts, new and irmovative multi- 
spectral infrared cirrus analysis techniques are being developed and successfully applied 
and validated to real-time cirrus detection and analysis scenarios. Integration of these and 
other retrieval techniques into the overall SERCAA cloud analysis allows the strongest and 
most reliable attributes of each technique to be combined into one comprehensive cloud 
analysis product that is more realistically representative of cirrus spatial and optical proper¬ 
ties than the current RTNEPH. With increasing amounts of multispectral infrared satellite 
data becoming available, the goal is to continue to retrieve from these data high quality 
augmented radiative, spatial, and microphysical properties of cirrus clouds in red-time and 
for climatological purposes at finer spatial and temporal resolution. 


Altitude (km) 



Time (HHMM) 

A A 



Figure 6. SERCAA cirrus altitude retrievals using Water-Vapor/LWIR-channel data 

compared to coincidence TPQ-11 35 GHz radar returns over Hanscom AFB, 
2315 UTC 15 Sept 95. Contrast to Figure 5 whose MWIR-channel data were 
used in place of Water-Vapor-channel data. 


6.7^m — ^0.7fLm Emissivity 

22.40 22.60 22.80 23.00 23.20 23.40 23.60 23.80 

Time (z) 

« Ch3,Ch4 effective altitude ■ blackbody altitude-cloud top-cloud bottom 

Figure 7. Plots of SERCAA and RTNEPH-derived cirrus altitudes for the cirrus samples 
shown in Figures 5 and 6. 



4.3 Optical and Microphysical Parameters for Water Droplet Clouds 

A comprehensive radiative transfer scattering theory was applied to model the 
response of NOAA AVHRR and GOES-Next satellite imager data to realistic variations in 
cloud microphysical, optical, and radiative processes. In this maimer the sensitivity of 
multispectral MWIR-LWIR measurements to variations in the microphysical properties of 
liquid water-droplet clouds was assessed. Results of these theoretical computations can be 
applied in the development of retrieval concepts to infer cloud drop size distributions, 
MWIR emissivity, and normalized cloud liquid water content (LWC). 

Calculations were made for the nighttime MWIR and LWIR chaimels of NOAA 
AVHRR. Atmospheric properties including relevant gas absorption profiles and aerosol 
scattering extinction profiles (parameterized by surface visibility) are evaluated using the 
multiple scattering version (Isaacs et al., 1987) of the Phillips Laboratory LOWTRAN-7 
radiative transfer model (Kneizys et al., 1983). LOWTRAN standard atmospheres (tropi¬ 
cal, mid-latitude summer/winter, subarctic, and U.S. Standard) drive the calculation of gas 
absorption as a function of altitude within the atmosphere. Aerosol properties (optical 
thickness, single-scattering albedo, phase functions) are also specified within LOWTRAN- 
7 by selecting an aerosol model (e.g., rural, maritime). Analogous cloud properties are 
based on Mie scattering calculations (Lon^n and Shettle, 1988) using specified cloud- 
type-dependent modified gamma droplet size distributions for the cloud particles. Complex 
index of refraction data of Hale and Querry (1973) are used. These input data are 
processed by an interface routine that generates an input file for a discrete ordinate method 
(DOM) multiple scattering radiative transfer model that computes scattering, absorption, 
and thermal emission in a vertically inhomogeneous, non-isothermal atmosphere (Stanmes 
et al., 1988). This interface merges the gaseous, aerosol, and cloud (if present) optical 
properties profiles and produces the required input profiles for use by the radiative transfer 
model. The DOM model also accounts for bidirectional reflection and thermal emission 
from the Earth's surface. Reflectance functions for ocean, soil, ice, and vegetation are 
used to provide the required surface properties. AVHRR upwelling radiances have been 
calculated as a function of viewing angle, solar zenith angle, and atmospheric state (e.g., 
clouds, water vapor). These results were then mapped to the input cloud microphysicS 
properties in order to determine whether a measurable multispectral MWIR/LWER signature 
could be expected to uniquely specify cloud water droplet size distributions. 

The DOM adding-doubling radiative transfer model was used to simulate the 
dependence of AVHRR radiances on cloud droplet size and cloud LWC. The effects of 
upwelling bandpass-weighted MWIR-LWIR radiances were determined using the modified 
gamma distribution of Deirmendjian (1969). This distribution characterizes cloud 
microphysical attributes in terms of two parameters: cloud LWC and droplet mode radius 
rm- This droplet distribution is then used by the Mie scattering code to generate the angular 
distribution of monochromatic optical depth for the desired cloud layer attributes. Mie 
calculations are subsequently used by the adding-doubling model to generate theoretically- 
expected AVHRR upwelling radiances. 

For low- and mid-level clouds, the particle density distributions of Stephens (1979) 
for stratocumulus (Sc) and Diem (1948) for altostratus (As) were used to specify the cloud 
microphysical parameters rm and LWC. Each distribution is plotted in Figure 8. 

Radiative transfer model results indicated little if any LWIR sensitivity to cloud liquid 
water content (i.e., cloud thickness) for a given cloud type and particle size distribution. In 
other words, water-droplet-cloud thermal emission at MWIR and LWIR wavelengths is 
independent of LWC starting at values of total liquid water as low as 17 g m-2. Thus for 
clouds with integrated LWC of 17 g m-2 or higher, the modeled MWIR-LWIR brightness 



Figure 8 . Modified gamma droplet size distributions for stratocumulus (dot,dot) and altostratus 
(solid). Distributions are from Stephens (1979) and Diem (1948), respectively. 

temperatme differences are not strongly dependent on water content. The dependence of 
particle size distribution was investigated by parametrically varying the mode radius rm 
from 2 to 10 |xm. For a given view angle, there is considerable sensitivity of brightness 
temperature to droplet mode radius. Figure 9 shows this sensitivity for a mid-latitude 
summer atmosphere and a cloud of 1 -km thickness whose top is at 1 , 2 , 3 , 4 , 5 , and 6 km. 

Note that in general, larger drop sizes result in lower absolute-magnitude brighmess 
temperature differences. 

Figure 10 shows the angular dependence of brightness temperature to cloud droplet 
mode radius in a tropical atmosphere, for As clouds. Note in general that the absolute 
magnitude of the MWIR-LWIR brightness temperature difference is larger for higher 
satellite viewing angles, i.e., for longer atmospheric paths. This is primarily due to the 
increased and propo^onally higher atmospheric water vapor attenuation effects at the 
LWIR wavelengths in comparison to the MWIR wavelengths. 

Results of the radiative transfer computations were next summarh^d in multivariate 
linear regression form to obtain a closed-form relation between droplet mode radius rm and 
the MWIR-LWIR brighmess temperature difference T 3 -T 4 and satellite view angle 0 as a 
function of atmospheric type: tropical, mid-latitude summer, mid-latitude winter, subarctic 
summer, and subarctic winter. The best associations were obtained using a regression 
equation of the form 

rm = ao + ai (T 3 - T 4 )2 cosB + a 2 cos 0 + a 3 (T 3 - T 4 ) + a 4 (T 3 - T 4 ) 2 . (14) 

The coefficients aj are listed in Table 2 as a function of atmosphere (e.g., tropical, 
mid-latitude winter) and cloud top altitude Zcid> which is obtainable from the water-vapor- 
attenuation corrected LWIR brighmess temperature. The regression errors are smaller than 
10 percent. It is then possible to use Eq. (14) for a given satellite scene to determine the 
drop size distribution and cloud-thickness-normalized LWC given the measured MWIR- 
LWIR brightness temperature difference from AVHRR. 


T3 - T4 : Mid-latitude Summer Atmosphere 

2 4 6 

Mode Radius ((xm) 

Figure 9. Plots of nighttime MWIR/LWIR brightness temperature differences as a 
function of cloud top altitude for the mid-latitude summer atmosphere. 

T3 - T4 : Tropical Atmosphere - As Clouds 

Table 2. Regression coefficients for Eq. (14) that yield droplet mode radius as a 

function of the MWIR-LWIR brightness temperature pair and the satellite look 
angle, for each model atmosphere. 

Mid-Latitude Summer 












































Subarctic Summer Atmosphere 

Cld Hgt 






0 -lkm 




































Tropical Atmosphere 

Cld Hgt 






0 -lkm 





































4.4 Cloud Environment 

Current operational, automated cloud analysis approaches underutilize available 
meteorological satellite remote sensing data resources. For example, onboard the two U.S. 
polar orbiting satellite platforms, DMSP and NOAA, there are two different multispectral 
imagem, four different microwave sounders, and a multi-channel infrared sounder. Table 
3 provides a compilation of the multispectral, multisensor data sources currently available 
from the DMSP, NOAA, and geostation^ platforms. Current approaches also lack the 
potential to fully exploit the next generation of satellite sensors. The plaimed National 
Polar-orbiting Oj^rational Environmental Satellite System (NPOESS) and Earth Orbiting 
System (EOS) will add an even greater number of sensor suites. The amount and accuracy 
of cloud information potentially obtainable by passive remote sensing is dependent upon 
the number of quasi-independent radiance measurements used in the analysis. The spectral 
diversity of satellite sensors is great, providing the potential to infer significant high quality 
cloud information. 

This section provides a discussion of algorithmic approaches to utilize combined data 
sets of DMSP cloud imager and microwave imager (SSNhl) and microwave sounder 
(SSM/T-1 and SSM/T-2) data to support the inference of cloud environment parameters 
including vertical temperature and moisture profiles and cloud liquid water. We also 
address approaches to deal with some current deficiencies in nephanalysis which can be 
treated by combining cloud imager and other sensor data sets. These include the "black" 
stratus problem over land, inference of land surface temperatures in the presence of cloud, 
and the identification of precipitation (including thunderstorms). All of die above 
approaches might be considered for eventual operational implementation as extensions to 
the existing SERCAA Phase 1 algorithm suite. Here we focus on DMSP based 
approaches. Much of this work is based on applications of the unified retrieval approach 
for DMSP developed under separate funding. 

The cloud spatial properties (coverage, layering, and type) resulting from the objec¬ 
tive analysis of cloud imager data by the SERCAA Phase I algorithm suite result from 
threshold, contrast, multispectral, and temporal tests. For example, the determination of 
cloud presence in an individual pixel might be determined by whether satellite-measured 
radiance at visible wavelengths is greater than a predetermined background value or 
whether the radiance in neighboring channels is similar indicating a reflective (white) cloud. 
The analysis does not consider meteorological factors in determining the presence of cloud 
although it is understood that cloud environment, i.e. temperature, moisture, and attendant 
circulations will determine the thermodynamic characteristics for cloud presence. This 
problem, i.e., that of diagnosing cloud presence and properties from the properties of the 
generating cloud environment, is a vety active area of research for general circulation 
modelers including those specifically interested in cloud forecasting (see Zivkovic and 
Louis, 1992; Nehrkom and Zivkovic, 1996). The availability of simultaneous measures of 
cloud spatial properties afforded by the SERCAA Phase I algorithms and cloud environ¬ 
ment data such as the vertical profiles of temperature and moisture alone would be 
extremely useful to this community. 

There are other reasons for supplementing the information available from cloud 
imaging sensors with that from collocated sounding sensors. The specification of 
simultaneous vertical temperature and moisture profiles provides important information in 
the determination of other desired cloud parameters such as cloud top height and more 
complete information of cloud layering and type. Additionally, cloud liquid water content 
available from microwave sensors can also be used to help in cloud typing. 


Table 3. Sensor data suite 



Channel Wavelength/Frequency 



0.4-1.1 and 10.5-12.6 |i,m 


19.35 V&H, 22.235 V, 37.00 V&H, 85.5 V&H GHz 


50.5, 53.2, 54.35, 57.9, 58.4, 58.825, 59.4 GHz 
(O 2 absorption band) 


91.5, 150, 183*1, 183*3, and 183*7 
(H 2 O absorption band) 



0.58-0.68, 0.725-1.10, 1.58-1.641, 3 . 55 . 3 . 93 , 

10.3-11.3, 11.5-12.5 pm 


0.69, 3.76, 4.00, 4.13, 4.40, 4.46, 4.52, 4.57, 6.72, 7.33, 9.71, 

11.11, 12.55, 13.35, 13.64, 13.97, 14.22, 14.49, 14.71, 14.95 pm 
(CO 2 , N 2 O, O 3 , and H 2 O absorption bands) 


50.30, 53.74, 54.96, and 57.95 GHz 
(O 2 absoiption band) 



0.55-0.75, 3.8-4.0, 6.5-7.0, 10.2-11.2, 11.5-12.5 pm 



0.5-0.75, 10.5-12.5 pm 



0.5-0.9, 5.7-7. 1 , 10.5-12.6 pm 

1 1.6 |xm channel will first fly on NOAA K -1997 

Temperature profile data can be used to determine cloud top heights (or equivalently, 
pressures). When cloud is determined to be present in a pixel, the cloud imager equivalent 
black body brightness temperatures (EBBT) for that cloudy pixel is available. In order to 
transform this EBBT into a cloud top height, two procedures are required. First, 
estimation of water vapor continuum effects on the infrared window measurements can be 
made based on the relevant column integrated water vapor amount obtained from the 
moisture profile (see below). Second, Ae temperature profile can be used to assign a cloud 
top height (or pressure) to the corrected EBBT. 

The SERCAA Phase I cloud layering approach clusters cloud imager equivalent black 
body brightness temperatures (EBBT) into defined layers. Layers are characterized by then- 
presence and the resultant cloud top EBBT. In analog to the pixel-by-pixel discussion of 
cloud top height in the preceding paragraph, temperature profile information can be used to 
make layer height (pressure) assignments as required. 

Vertical moisture profiles provide both the vertical distribution of water vapor and the 
total integrated water vapor. In conjunction with the retrieved temperature profiles, the 
vertical moisture profiles provide simultaneous relative humidity information. Integrated 
water vapor from the cloud top to space is necessary to estimate water-vapor-continuum 


effects on infrared window measurements. The impact of water vapor absorption is to 
decrease the EBBT seen by the sensor. This effectively raises the cloud top height deter¬ 
mined if water vapor absorption were not considered. In tropical atmospheres where the 
burden of water vapor is large, this effect may be important in the determination of cloud 
top heights, especially for low clouds (since most of the water vapor is near the surface). 

Microwave imager (SSM/I) and millimeter wave moisture sounder (SSM/T-2) data 
can be used to infer integrated cloud liquid water content. Integrated cloud liquid water 
content available from microwave sensors can also be used to help in cloud typing. In 
general integrated cloud liquid water content is largest for bright, low clouds and smallest 
for darker, higher clouds. 

In our previous work, we have described a unified retrieval (UR) algorithm 
applicable to the Defense Meteorological Satellite Program SSM/T-1, SSMyT-2, SSM/I 
microwave sensors which employs physical retrieval concepts and utilizes the frill 
multispectral information content of the available data (Isaacs, 1987; Moncet and Isaacs, 
1992,1994; Moncet et al., 1996). The UR algorithm is a general non-linear physical 
retrieval algoritlm for the simultaneous retrieval of temperature profiles, water vapor 
profiles, cloud liquid water content, surface temperature and surface emissivity from the 
DMSP multi-sensor platforms. The UR concept offers the distinct advantage of 
simultaneously characterizing the atmospheric and surface state vector from the available 
multifrequency sensor observations by combining the traditionally separate functions of 
atmospheric sounding and surface property imaging. Through the physical retrieval 
process, it intrinsically recognizes the dependence of atmospheric properties on the 
observed background and surface properties on atmospheric transmission effects. 

The potentid usefulness of cloud environment information available from DMSP 
microwave sounding sensors to enhance IR based automated cloud detection is illustrated 
by examining the problem of detecting so-called "black stratus" clouds (Moncet et al., 
1996). Black stratus, which is problematic to most IR cloud detection schemes including 
the SERCAA Phase 1 algorithms, refers to low level clouds that form within an inversion 
layer at night as the surface cools. Common at high latitudes, particularly during winter, 
these clouds are difficult to detect using satelhte data because their radiative signature is 
greater than of the underlying surface. IR threshold satellite cloud detection schemes 
generally require surface-to-cloud contrast in either an infrared-channel brightness 
temperature or visible-channel brightness in order to detect a cloud. However, since the 
black stratus are warmer than the surface any infrared signature that can be detected by the 
satellite will be in the wrong direction (i.e., clouds are assumed to be colder than the 
surface). Similarly, since black stratus tend to occur under limited or no sunlight 
con(htions (e.g., high latitude winter) or over reflective snow or ice backgrounds, reliance 
on visible cloud signatures is severely restricted. 

Working collaboratively with the SERCAA Phase n team, Moncet et al. (1994) 
formulated an algorithm to treat “black stratus” detection. The algorithm illustrated in 
Figure 11 uses both IR imager brightness temperatures applied to a simple threshold 
algorithm similar to the one currently used in SERCAA Phase I for the processing of the 



IR Brightness 

Microwave Derived 
Temperature Profile 

Surface Skin Temperature (Tskin) 



Tskln - A T^ 

IR<Tskin’ ^^0 

iR >Tskin+AT q 

lR<Tskin- AT-j 

<IR < 

■ikin+ AT2 

IR >Tskin+A T 2 



Cold Cloud 


Warm Cloud 

(Type 4) 



(Type 2) 


Figure 11. Schematic of a prototype infrared cloud detection scheme that integrates 
microwave derived information. 

OLS infrared data. Additionally, however, it uses microwave sensor retrieved temperature 
profiles and surface skin temperatures to identify and classify inversion conditions. Like 
the SERCAA Phase 1 OLS algorithm, a cloud is detected if the IR measured brightness 
temperature is colder than the estimated surface skin temperature by some pre-set margin 
under so-called “normal” conditions. However, when an inversion is present, the 
algorithm also detects a low cloud if the IR brightness temperature is warmer than the 
surface temperature (i.e. the black stratus condition). 

Moncet et al. (1994) illustrate an example applying this algorithm over snow/ice 
fields based on data from the DMSP SSM/T-2 Calibration/validation effort (Falcone et al., 
1992) using DMSP 5D-2 data. While additional validation of the algorithm is required, the 
approach integrates well into the overall SERCAA algorithm suite framework. 

The existing imager based cloud analysis approaches for the infi*ared LWIR channels 
are based on discriminating cloud (usually cold; except in the case of black stratus 
described above) from the warmer, cloud-free background. The determination of cloud 
free background temperatures is supported by the cloud imager LWIR data itself (using 
adjacent cloud free pixels). This becomes problematic in overcast situations. One 
approach is to rely on model predicted surface temperatures (as in the use of the SCFTMP 
model by AFGWC). Practice indicates that these may be inaccurate resulting in over or 
underanalysis of cloud. Kopp et al. (1994b) discussed the use of SSM/I derived surface 
temperatures in improving results of the RTT^PH model at AFGWC. In their application, 
SSM/T surface temperatures are used along with those of SFCTMP based on a weighting 
determined within SCFTMP. 

As a result of the UR approach described above, surface temperatures may be 
retrieved over ocean and land in fully overcast situations using the combined DMSP micro- 
wave sensor suite. Over the ocean standard deviations of between 2 and 3 K are obtained 
between the retrieved sea surface temperatures (SSTs) and Navy SST fields. Over land, 
there is little sensitivity of retrieved surface skin temperatures to cloud. This provides the 
basis for retrieval of surface skin temperature in completely overcast conditions. Results 
indicate that it would be worth investigating whether coupling OLS information with the 
microwave skin temperature retrievals would be helpful in enhancing IR based low cloud 


detection capabilities over land (Moncet et al., 1996). An advantage of the UR approach 
over that of using SSM/I derived temperatures alone is the retrieval of surface emissivity 
which decouples the surface temperature results from those of surface type. 

Convective storms represent a significant weather aviation hazard and are given 
special treatment in operational cloud typing. At AFGWC, the approach adopted to identify 
thunderstorms is the convective-stratiform technique (CST) of Adler and Negri (1988). 

The CST technique is based on the use of satellite imageiy alone. In particular it uses the 
spatial properties of LWIR EBBT's to differentiate between convective and stratiform 
situations. A description of the approach applied can be found in Kopp et al. (1996). One 
immediate suggestion based on our experience with the DMSP microwave sounder and 
imager data is the use of these data to help characterize the presence and vertical extent and 
hence intensity of thunderstorms and convective precipitation. 

A number of studies have indicated that multifrequency microwave observations of 
convective systems can be used to identify both liquid and glaciated hydrometeors (see 
Wilheit et al., 1982; and Hakkarinen and Adler, 1988; Adler et al., 1990; Kummerow et 
al., 1991, Petty, 1994; Petty, 1995; Petty and Miller, 1995). These studies have generally 
tended to focus on window channel observations (i.e. 91 GHz and channels away from the 
center of the 183.31 GHz water vapor line). Microwave brighmess temperatures (Tbs) 
tend to be dommated by emission from rain below the freezing level, while higher 
frequency millimeter brightness temperatures are affected by scattering from frozen 
hydrometeors. These signatures thus provide gross information on the vertical structure of 
storms (Spencer et al., 1989). 

The key to using window channel data is associating these signatures with consistent 
vertical temperature stracture information and exploiting sounding channel sensitivity to the 
s^e phenomenon. Notably, precipitation estimates from SSM/I are obtained using 
microwave window channels and thus assume a model of the vertical distribution and 
phase of precipitation. Recent aircraft observations of precipitating convective systems at 
microwave and millimeter wave frequencies indicate the potential for inferring the vertical 
distribution of precipitation proj^rties (Wilheit et al., 1982; Hakkarinen and Adler, 1988). 
The channel set of the DMSP microwave sensor suite (Falcone and Isaacs, 1987) provides 
information content from 19.35 to 183 GHz which can be exploited to provide such 
inforaiation. Since the multifrequency response of liquid precipitation differs from that of 
glaciated precipitation, an indication of the freezing level is also possible. Examining the 
effect of clouds and precipitation on SSM/T-2 channels, non-precipitating clouds over 
water generally displayed warmer 91 GHz Tbs compared to clear FOVs (Figure 12). There 
was no difference between Tb signatures in the SSM/T-2 channels for cloud-filled FOVs 
over land. Griffin et al. (1994) showed a distinct SSM/T-2 Tb signature for precipitation 
(data set used was that from a developing typhoon in the western Pacific Ocean) (Figure 
13). The presence of light rain over water caused the warmest Tb to shift to 150 GHz. As 
the rain rate and scattering in the FOV increased, the 183il GHz Tb became the warmest 
of the thr^ atmospheric channels. This work and that of Pickle et al. (1996) suggest that 
DMSP microwave imager and sounder data would be useful in delineating precipitating 
cloud types, particularly thunderstorms. 


Tb [K] § Tb [K] 

SSM/T.2 Channels [GHz] 

12. Typical Tb signatures for clear over ocean (solid), partly cloudy over ocean 
(dash-dot), cloudy over ocean (dash), clear over land (solid), partly cloudy 
over land (dash-dot), and cloudy over land (dash) for data collected during 
May and July, 1992 for the east and west coast of the United States. 

Figure 13. Typical Tb signatures for clear (dash), light precipitation (solid) and heavy 
precipitation (dot) observed in Typhoon Oliver, February 4,1993 in the 
western Pacific Ocean. 



4.5 Implementation Changes for Cloud Typing and Layering 

The SERCAAI cloud typing algorithm provided a broad classification of pixels into 
one of ^o types of clouds, cumuliform or stratiform, based on their spectral and spatial 
properties (Gustafson et al., 1994). Cloud-filled longwave IR pixels were segregated into 
cloud classes using an unsupervised classification approach. The classes were then 
processed in a top-down fashion by applying a multi-pass, region growing segmentation 
routine in concert with variance-based texture information. The segmentation routine 
resulted in the formation of regions, or groups of connected pixels, whose size was the 
basis of the cloud type determination. The underlying assumption gove rnin g the process is 
that cumuliform clouds are represented by small, connected and compact regions of pixels 
while stratiform clouds extend over larger regions of connected pixels wi thin a given 
temperature class. Therefore, a size threshold was applied to each new region and regions 
were then sorted into cumuliform or stratiform types. 

While this implementation provided an accurate reflection as to the nature of cloud 
formations in a scene, the calculation of the texture derivative and the region-growing 
feature of the segmentation routine proved to be computationally intensive. An improved 
technique has been developed and implemented in which composite temperature classes are 
segmented using a fast, single-pass algorithm based solely on pixel contiguity. Pixels in 
each composite mask are analyzed in turn with respect to their imm ediafe. neighbors and are 
assigned either a new region identification (ED) number or that belonging to their top or left 
neighbor. Ultimately, aU pixels that ^e connected in one of the four adjoining spatial 
^ections (left, right, top, bottom) will belong to the same region. As in the ori gin al 
implementation, the number of pixels in a region determines Ae cloud type label it will 
receive. Regions less than a given threshold size are assigned a cumuliform label; aft others 
receive a stratiform label. By omitting both the multi-pass segmentation technique and the 
derivation of a texture layer, the new routine successfully operates several orders of 
magnitude faster than its predecessor despite the continued iterative, top-down processing. 

In addition to the improved speed and performance of the SERCAA Phase I cloud 
typing algorithm, a new techmque for discriminating among cloud layers was also 
developed and tested. The SERCAA Phase-1 cloud layering algorithm was achieved by 
performing a second, local-scale unsuj^rvised classification (clustering) layering process 
that operates on the cloudy pixels within a 3x3 floating window of l/16th-mesh grid cells 
centered on the grid cell currently being analyzed (Gustafson et al., 1994). This process is 
conceptually simil^ to the large-scale layering procedure just described; unsupervised 
clustering and maximum likelihood classification are performed on the cloudy pixels that 
lie within the floating window. Under Phase 1, these pixel values were linearly 
proportional to the cloudy pkel's LWIR brightness temperature. This technique worked 
well for opaque clouds whose brightness temperatures are very close to the true physical 
teinperature of the cloud. However, for thin cirrus the LWIR brightness temperature, 
being a mix of cold cloud and warm background, is not representative of the physical 
temperature of the cloud. Thus cirrus that has varying optical thicknesses in the horizontal 
but whose tops are essentially at the same vertical level was being separated into too many 
layers because of the large variation in cirrus brightness temperatures. In this case the 
relatively large spread of LWIR brightness temperatures misleads the clustering algorithm 
into separating the cirrus into too many layers when the spread in cloud top altitudes is not 
nearly as large. 

To alleviate this problem, tests were conducted using the SERCAA Phase 2 retrieved 
cloud top altitudes instead of the raw brightness temperatures as input to the layer 
unsupervised clustering and maximum likelihood classification (MLC) routines. Recall that 
SERCAA Phase 2 cloud top altitudes include water-vapor-attenuation corrected altitudes for 


opaque clouds, and transmissive cirrus corrections for thin ice-particle clouds. These 
altitudes were passed to the clustering and MLC routines in tandem with the original LWIR 
brightness temperatures, and the results were subjectively compared. 

Figure 14 contains an LWIR NOAA-14 image over New England in November 
1994. Note that the image contains a large cirrus shield of varying brightness 
temperatures. The corresponding LWIR brightness-temperature layer analysis is shown in 
Figure 15. In Figure 15, gray 16th-mesh boxes have one cloud layer in them while white 
16th-mesh boxes have two. Note that in the left side of the image the SERCAA Phase 1 
algorithm has found two layers where there is a large variation in cirrus LWIR brightness 
temperatures. However, the Phase 2 height analysis indicates that the cirrus tops are all 
within 1-2 km of each other over this region. When the Phase-2 altitudes were fed into the 
clustering procedure, far fewer 16th-mesh boxes were analyzed as containing two layers, 
as is shown in Figure 16. Note that in the heart of the thin cirrus shield, the Phase 2 
altitude-generated layer algorithm assigns only one layer to the cirrus. In Figure 15, the 
corresponding brightness-temperature layer analysis seems overdone, again being confused 
by the large variations in brightness temperature that are caused by variation in cirrus 
optical properties across the scene. Overall the results that were obtained for two dozen 
cases using the altitude-generated layers were noticeably improved in areas of thin cirrus. 
However, it is recommended that the technique be tested over many more samples and that 
some sort of objective measure of goodness be established to judge whether layers should 
be generated from altitudes or brightness temperatures. 

4.6 Analysis Consistency 

The existing RTNEPH cloud analysis process (Kiess and Cox, 1988; Hamill et al, 
1992) was designed to produce quarter orbit based cloud analysis products based on 
individual polar orbiting passes from two-charmel (VIS and IR) data, primarily supplied 
from the DMSP/OLS. To address known deficiencies in the RTNEPH's polar platform a 
primary objective of the SERCAA effort was to incorporate both polar multispectral data 
from the NOAA TIROS Advanced Very High Resolution Radiometer (AVHRR) and 
hourly geostationary data from GOES, Meteosat, and GMS (Neu et al., 1994). The 
availability of the multispectral data significantly improves the ability to detect low stratus 
and thin cirrus, while geostationaty data complements that of the polar platforms by 
providing higher temporal sampling of diurnal cloud development and dissipation, 
particularly in the tropics. A key element of SERCAA is the integration algorithm to 
optimally combine the most recently available cloud product analyses from all platforms 
into a single optimal analysis. 

As discussed in Section 3 above, the design philosophy for the SERCAA analysis 
inte^ation algorithm was to provide an optimal 16* mesh merged cloud analysis valid at a 
specific instant in time based on the analyzed cloud products derived from the most recent 
observations available from the three independent classes of environmental satellites. The 
key attribute of the integration scheme is that the RMS error at each grid cell will be 
minimized independent of the cells around it. Thus, the algorithm design does not include 
logic to enforce meteorological consistency between temporally or spatially adjacent grid 
cells. This was neither the requirement nor the design philosophy during SERCAA. 
However, evaluation of SERCAA-derived integrated cloud products using real satellite data 
has revealed that meteorological consistency between hourly-generated products can be 
degra,ded by numerous factors including: changes in the mix and relative age of input 
satellite data, variations in solar illumination and satellite viewing geometry, and the t 5 q)e of 
cloud present. 


Figure 14. NOAA-14 10.7-ixm brightness temperature image over New England for 
November 1994. High tones denote low brightness temperatures. 


Figure 15. SERCAA Phase-1 cloud layer analysis for the image in Figure 14. This 

analysis was generated using LWIR brightness temperatures as input to the 
clustering and unsupervised classification routines. Each square represents a 
16th-mesh box. White boxes contain two cloud layers, and gray boxes contain 
only one. 


Figure 16. SERCAA Phase-2 cloud layer analysis for the image in Figure 14. This 
andysis was generated using SERCAA Phase-2 pixel-resolution corrected 
altitudes as input to the clustering and unsupervised classification routines. 
Each square represents a 16th-mesh box. White boxes contain two cloud 
layers, and gray boxes contain only one. 


The SERCAA analysis and integration algorithms were evaluated using a series of 
ten-day satellite data sets obtained from NOAA 11 and 12, DMSP FIO and FI 1, GOES-7, 
GMS-4, and METEOSAT 3,4, and 5 for the months of March and July from 1993 and 
1994. The algorithms were applied over three regions of interest including eastern Asia 
and tropical Indonesia, central and northern South America, and north-central Africa and 
the Mediterranean that were selected for geographic and climatic diversity (Figure 17). 
Results from this real-data study were analyzed to establish baseline characteristics of both 
the analysis and integration algorithms. 

Generally, results of the source-specific analysis algorithms showed good agreement 
with manual interpretation of satellite imagery generated from the input sensor data 
(Heideman et al., 1994; 1995). Examination of the final integrated results showed the 
complementary nature and relative strengths of the different data sources. However, under 
some conditions, inconsistencies were found to exist between contemporaneous polar- and 
geostationary-based cloud analyses. These are manifested as either spatially and/or 
temporally inconsistent cloud fields in the integrated products. Gustedson et al. (1995) 
investigated the source of these inconsistencies and concluded that they were due, at least in 
part, to differences in the inherent information content provided by each of the satellite 
systems used as data sources. Table 4 summarizes some of the relevant satellite sensor 
characteristics. The unique information content of each source with respect to cloud 
detection capability is a well understood (see Isaacs and Barnes, 1988) function of number 
of channels, channel band passes, sensor resolution, and solar illumination geometry. 

Figure 17. SERCAA evaluation regions of interest. 


Table 4. Summary of operational enviromnental satellite imaging sensor characteristics 













Cycle (h)^ 

Cycle (h)^ 

































































































Ipield of View on the ground varies from satellite subpoint (highest resolution) out to edge of scan 
(lowest resolution). 

^Assumes a two satellite constellation for each polar system; DMSP in early and mid morning orbits and 
NOAA in early morning and afternoon orbits. Geostationary minimum repeat cycles are for selected 
subregions, maximum cycles are for full disk. 

Also given the current constellation of gaps of polar satellites, up to four hour gaps can 
occur between the crossing times of consecutive orbits. While it is these differences that 
the SERCAA analysis integration algorithm exploits in producing a combined analysis, the 
different sensor characteristics can lead to different analyses of the same scene from 
multiple platforms. 

Figure 18 illustrates the essential nature of the temporal consistency issue using 
SERCAA analyses produced under contract with the DNA. The analyses shown cover a 
large region over the western Pacific Ocean and eastern Asia and are valid for the period 
1600 to 2200 UTC on 26 March 1993. The top row of panels show a 4-h sequence of 
integrated SERCAA analyses (gray shade is proportional to total cloud fraction with white 
representing completely overcast) beginning 1700 and ending at 2200 UTC. The bottom 
row contains a simple persistence forecast based on the SERCAA analysis valid at 1600 
UTC. The 1 and 2-h persistence forecasts correspond fairly well with the verifying 
SERCAA analyses because the available input satellite analyses during these time, period 
were limited to GMS only. However at 1900 and 2000 UTC the SERCAA analysis 


abruptly changes in appearance, and this is directly attributable to the addition of low cloud 
detected by the available AVHRR analyses valid at 1820 and 1958 UTC, respectively. 

While it is true that the later analyses in the above example provide a more complete 
picture of the actual clouds, the sudden appearance of analyzed low cloud fields over areas 
covered by the AVHRR sensor are not consistent with the analyses from other sensors. 
Note that this is strictly a consequence of the increase sensitivity to those types of clouds in 
the multispectral AVHRR data relative to the other data sources. However, since the 
overall time series is not meteorologically consistent, even though the analysis is 
compatible with the SERCAA goal of providing the best analysis based on the most 
recently available data, it may negatively impact user assessment of the accuracy and 
usefulness of the cloud analyses. This problem, however, is not unique to SERCAA: 
even in the current RTNEPH with its much more restricted input data, cloud fields are 
updated abruptly over areas covered by more recent satellite passes, and the cloud analyses 
also exhibit spatial artifacts of the data coverage. A potentially more serious concern is the 
effect of temporal inconsistencies on the initialization and verification of cloud forecast 

In adopting the SERCAA algorithm suite as the basis of the CDFSII-NEPH, an often 
stated assumption was that the most accurate cloud depiction would support the best cloud 
forecast. In the current suite of forecast models in use at GWC, which are all advective in 
nature, there are no technical difficulties in using cloud analyses with temporal or spatial 
discontinuities in the cloud fields for initialization. The difficulty arises when one tries to 
verify the resulting forecasts against subsequent cloud analyses. As an illustration of this 
problem, a simple persistence forecast from the hour preceding the cloud analyses is shown 
in Figure 18 alongside the verifying analyses. It is clear that for hours 1 and 2 of the 
forecast, neither Ae analysis nor the forecast show the low cloud fields because of the lack 
of multispectral AVHRR data. As a result, the verification provides an unrealistically 
optimistic assessment of the forecast accuracy. Only at hours 3 and 4 would the detection 
of the low cloud in the verifying analyses result in an apparent sudden reduction in forecast 
skill to its more realistic value. In this case the apparent reduction in forecast skill is an 
artifact of the verification procedure, caused by an unrealistically high apparent forecast 
skill at hours 1 and 2. Other scenarios are possible in which the apparent forecast skill is 
reduced because of temporal inconsistency in the analysis. For example, in a situation 
where an analysis at the initial time contains multispectrally-detected low cloud, but later 
analyses do not as the multispectral data age out of the an^ysis. In this scenario, forecasts 
generated from the initial andysis will be penalized for (correctly) predicting low clouds not 
present in the (later) verification analyses. Here the apparent forecast skill reduction is 
caused by an unrealistically pessimistic assessment of the forecast skill by the verification 
procedure. Similar problems would occur with advective forecasts since analysis 
inconsistencies at the start and end point of the cloud trajectories would result in artificially 
low verification score. 

As noted, the potential for a high level of temporal inconsistency in CDFS n has 
potentially significant implications for the apparent accuracy of cloud forecasts initialized 
from these fields. It is important to point out that, given the above example, the issue of 
cloud forecast accuracy is as much an issue of validation data set characteristics as it is one 
of forecast quality itself. A possible mitigation strategy would be development of a new 
integration dgorithm with modified design criteria that emphasized consistency over fidelity 
to the input data. Validation methodology should also be reviewed to determine if 
requirements are better served by validation criteria that consider temporal consistency of 
both the input and verification data. This is addressed in Section 6.2.5 below. 


Figure 18. SERCAA analyses of total cloud fraction over the western Pacific Ocean and eastern Asia on 26 March 1993. Top 
row contains hourly analyses starting at 1700 UTC and ending at 2000 UTC. Bottom row shows a simple 
persistence forecast based on the analysis at 1600 UTC. Grey shade is proportional to analyzed cloud amount. 

4.7 Daytime Cirrus/Low-Cloud Discrimination Using MWIR Data 

Nighttime cirrus cloud detection tests compare the 3.7-|im and 10.7-|xm brightness 
temperatures; generally, MWIR temperatures are higher then corresponding LWIR 
temperatures, with the magnitude of the difference dependent on cirrus cloud altitude and 
optical depth. At night, thin cirrus MWIR-LWIR spectral signatures are unambiguous. 

There are three main reasons why citrus detection using multispectral infrared 
measurements is successful at night. The most dominant effect has to do with the nature of 
the dependence of the Planck function on temperature at the MWIR and LWIR 
wavelengths. Proportionally less energy comes from the warmer part of the scene as 
wavelength increases; this effect is due solely to the exponential dependence of the Planck 
function on temperature at 3.7 and 10.7|jm. The second effect is that of the varying 
emissivities of ice particle clouds themselves among the three wavelengths. In general, 
cirrus emissivities increase with increasing wavelength from 3.7 to 12|j,m. This means 
that, on the basis of emissivity alone, increasingly more of the upwelling radiant energy in 
a cirras-filled pixel comes from the colder cloud. This is an analogous but weaker effect to 
that of the Planck function in that it amplifies the brightness temperature differences that 
comprise the thin cirrus signature. Finally, a third effect that causes brightness 
temperatures for cirrus pixels to decrease with increasing wavelength is that of varying 
atmospheric water vapor attenuation. In general, atmospheric water vapor attenuation is 
stronger at longer wavelengths. This operates in the same sense as do the previous two, 
increasing the difference between 3.7 and 10.7-|im-channel brighmess temperatures, 
although typically it is the weakest effect of the three. 

In the daytime, solar energy adds complexity to the cirrus MWIR-LWIR signature. 

At 3.7|im there is a noticeable solar component to incident MWIR energy; this component 
can be either reflected or transmitted by thin cirrus. Any transmitted energy continues 
downward to the underlying surface where it is reflected upward and re-transmitted by the 
cirrus. These effects cause the 3.7|im brighmess temperature to be even higher for thin 
cirrus during the day, so that the MWIR-LWIR difference remains positive. However, 
during daytime the magnitude of this positive 3.7-10.7fim difference is ambiguous. Low 
water droplet clouds such as cumulus, stratus, and fog also have very high brightness 
temperatures at 3.7p,m because they are relatively efficient reflectors of solar energy at 
MV^ wavelengths. Thus low clouds have a combination of thermal emission and solar 
reflection of MWIR radiation that causes their brighmess temperamres to be high as well. 

In turn, me MWIR-LWIR brighmess temperature difference is often as high for low, 
opaque clouds as it is for high, transparent citrus. 

For meteorological satellites wim 6.7-|xm water vapor sensors, this problem is 
surmountable. Such channels sense predominantly upwelling thermal energy mat has been 
absorbed and re-emitted by atmospheric water vapor, wim little or no contribution to me 
upwelling radiance coming from low clouds or me Earth's surface. The only type of 
clouds mat influence strongly me 6.7-|im merrnal energy are cirrus clouds. The weighting 
function of a 6.7-nm sensor peaks relatively high in me atmosphere, not often any lower 
man cirrus altitudes. Thus water-vapor brighmess temperatures for thin cirrus are similar 
to me true physical temperamre of the cirrus itself, while for low clouds they are very 
different. It is very conceivable merefore mat a 6.7-10.7-|im brighmess temperamre 
difference will be routinely small for cirrus, but large for low clouds and fog. 

It is be possible to quantify this relationship through manual cloud-clearing of 
daytime multispectral GOES-Next imagery and comparing me observed cirrus water- 


vapor/LWIR differences with those that correspond to water-droplet clouds. Thus a 
potential daj^time ice/water-cloud discrimination test is to first check for large MWIR-LWIR 
differences; this detects thin cirrus and stratus and fog. The second check is then to 
determine how large the water-vapor/LWIR difference is: if small, then cirrus; if large, 
then low cloud. Such a suite of daytime tests will allow for confident daytime cloud phase 

5. Future Research Topics 

The scope of the SERCAA program was large and multifaceted. However, as with 
any large research program, issues came up during the period of the project that are of 
interest to Air Force mission requirements but that were outside the scope of the current 
program. Many of these issues are not being adequately addressed by existing or currently 
planned capabilities. There were also lessons learned that could potentially improve future 
operational capabilities. This section provides an overview of these topics and a discussion 
of possible approaches and how they might be used to address Air Force requirements. 

5.1 New Sensors 

Over the coming decade a number of new environmental satellite and sensor systems 
are scheduled for launch. Table 5 provides a list of proposed satellites and respective 
characteristics of the primary sensors that are potentially useful for cloud applications. 
Extant satellites, depending on the suite of onboard sensors, provide varying levels of 
information that can be exploited for cloud property retrieval. The SERCAA cloud analysis 
m(^el currently being implemented for operational Air Force use as part of CDFS n is 
unique in that it analyzes data from multiple satellites and accommodates differences in 
information content through the use of multiple cloud analysis algorithms, each tailored to a 
specific satellite sensor system (Gustafson et al., 1994). Thus as new satellites are 
launched, the SERCAA model is intended to be easily expandable to include the new data 
sources simply through the addition of another source-specific analysis algori thm 
However, for this or any other plan to improve cloud analysis accuracy using new sensor 
data to succeed it is necessary that the appropriate algorithms be developed prior to 
introduction of the data into the analysis system. 

High spectral resolution data from new satellites offer the potential for significantly 
advancing the state of the art for analysis of cloud from space. Numerous studies have 
addressed the advantages of adding specific channels or higher resolution sensors to the 
existing constellation of environmental satellites (e.g., Di Girolamo and Davies, 1995; 

King et al., 1992; Eyre and Menzel, 1989; Arking and Childs, 1985). These new data 
sources are applicable for addressing specific weaknesses in current cloud analysis 
algorithms. For example, discrimination and accurate altitude assignments for optically 
thin cirrus is extremely difficult during daytime conditions. However, 8.55 fxm channel 
data, has been demonstrated to provide sufficient information, when combined with data 
from split long wave IR channels, to develop a robust, daytime thin cirrus detection 
algorithm (Strabala et al., 1994; Ackerman et al., 1990). An 8.55 pm channel is planned 
for instraments on the ADEOSI and EOS/AM platforms, and data in case-study quantities 
are available now from the MODIS Airborne Simulator (MAS). Similarly, improved cloud 
altitude assignments of high cloud can be made from temperature sounding instruments 
(Wylie et al., 1994; Menzel et al., 1992) and multispectral IR instruments (dEntremont et 
al., 1995; dEntremont et al., 1992) such as HIRS-S, MODIS, OCTS, AVHRR-3, and 


Table 5. Planned environmental satellite systems with primary sensor characteristics 




Launch Date 






moderate-resolution 5 channel radiometer in visible to thermal IR 


12-channel infrared sounding radiometer 




moderate-resolution 4 channel radiometer in visible to thermal IR 




high-resolution 6-channel radiometer in visible-thermal IR 


15-channel microwave temperature sounding radiometer with 12 channels 
in the 50-60 GHz oxygen absorption band 


20-channel infrared temperature and moisture sounder with channels along 
CO 2 and water vapor absorption lines 


5-channel microwave moisture sounder with 3 channels along the 183 

GHz water vapor absorption line 




Ku band active scatterometer for measurement of ocean-surface wind 
velocity (speed and direction) 


high-resolution 12-channel radiometer in visible-thermal infrared 


very-high-resolution 5-channel radiometer in visible to near infrared with 
daytime cross-track pointing, stereo capability 




moderate resolution radiometer with 3 channels covering broad band 
visible to long wave IR, visible to short wave IR and long wave IR 


staring telescope/filter imager at 0.777 fim for lightning detection 




moderate resolution dual radiometer, each with 3 channels, covering broad 
band visible to long wave IR, visible to short wave IR and long wave IR 


moderate resolution, multiple-angle imaging radiometer with 4 channels 
in visible to near-IR 


moderate resolution scanning radiometer with 36 channels covering 
visible to long wave IR 


very-high-resolution, 3-instrument radiometer with 3 visible to near IR 
channels, 6 short wave IR channels, and 5 thermal IR channels 


low resolution scanning spectrometer with 4 channels in gaseous 
absorption bands 




limb scanning and solar and lunar occultation spectroradiometer with 9 
channels covering visible to long wave IR 




Ku band active scatterometer for measurement of ocean-surface wind 
velocity (speed and direction) 


scanning microwave radiometer with 6 frequencies in 6-90 GHz for 
humidity profiling and precipitation/surface temperature retrieval 


scanning radiometer with 35 channels covering ultraviolet to thermal IR 




high spectral resolution infrared sounder using a 2300 channels 
spectrometer between 0.4 and 15.4 |j,m for retrieval of temperature and 
humidity profiles and surface temperature 


see NOAA K-N above 


5-channel microwave humidity sounding radiometer with 3 channels 
along the 183 GHz water vapor absorption line 


see EOS/AM above 


see EOS/AM above 




see NOAA K-N above 


see NOAA K-N above 


see NOAA K-N above 


see NOAA K-N above 








Changes and advances in satellite sensor design, and how they affect the cloud 
algorithm performance, is an ongoing issue. The recent launches of GOES 8 and 9, the 
first two satellites in the GOES I-M series of next generation of the U.S. operational geo¬ 
stationary program, provide the opportunity to evaluate the impact of new sensor channels 
using data available now. These satellites have a sensor design that is a significant depar¬ 
ture from the previous series (Menzel and Purdom, 1994). The principal instrument 
package consists of completely separate and autonomous imaging and sounding sensors 
with additional channels and enhanced resolution. Existing geostationary cloud analysis 
algorithms will require modifications to exploit the additional channels. The potential 
benefits from making these modifications are both large and relatively easily identified since 
the new sensor configuration is similar to the current AVHRR sensor which has been 
studied extensively. Advantages of a multispectral sensor over the previously-available 
two-channel system include improved low cloud detection, cloud phase determination, 
transmissive cirrus identification and characterization of cloud radiative and microphysical 

The Japanese Meteorological Agency has also recently launched a new geostationary 
satellite, GMS-5. The imaging sensor is modified over the previous version with the 
addition of a pair of split long wave infrared window channels in place of a single broad 
band channel. This offers the potential for improved thin cirrus detection, one of the 
recognized weaknesses of current geostationary-data analysis algorithms. 

5.2 Advanced Retrieval Techniques 

Existing and planned objective cloud analysis capabilities at the Air Force Global 
Weather Central (AFGWC) make use of multispectral data from multiple satellite systems. 
However in a number of import areas, gaps or weakness exist in the analysis 
methodology that affect the ability to adequately address Air Force requirements; 

1) quantitative interpretation of da54ime sensor channel data in the visible to midwave- 
infrared (MWIR) spectrum; 2) discrimination of transmissive cirrus over hot and bright 
backgrounds; 3) accmate cloud typing as a first-order determinant of cloud microphysical 
and radiative properties; 4) use of remotely sensed cloud radiative and environment 
characteristics to improve retrieval accuracy for cloud altitude; and 5) use of inferred cloud 
physical characteristics to provide temporal consistency within time series of analyzed 
cloud parameters derived from multiple sateUites. 

5.2.1 Interpretation of Visible to Midwave-Infrared Data 

All satellite-based cloud analysis techniques require some knowledge of the terrestrial 
back^ound radiative characteristics at the various sensor wavelengths in order to accurately 
discriminate cloud. Reliance on this knowledge vanes by technique and information 
content of the satellite data. However, quantitative background information is difficult to 
obtain globally since Earth surface radiative characteristics vary widely with geography and 
surface cover (e.g., deserts, tropical forests, permanent snow and ice cover). Vegetated 
surfaces and snow cover also vary for a given Earth location as a function of time, both 
diumally and seasonally. 

The problem of background surface characterization is particularly acute for data 
channels sensitive to reflected solar radiation. Changes in solar illumination with time of 
day, anisotropic reflection, and similarities between reflective terrestrial backgrounds and 
cloud all contribute to make objective cloud detection using these channels complex and 


often problematic. However, visible to midwave infrared sensor data are extremely useful 
for objective cloud detection since they generally exhibit a strong liquid-water cloud 
signature (i.e., the clouds are good reflectors). For two-channel sensors like the OLS and 
NffiTEOS AT satellites, the visible channel adds significantly to low cloud detection 
accuracy during daytime. Also for the new AVHRR-3 sensor scheduled for launch in 1996 
(Table 5), loss of the 3.7 |im channel during sunlight hours will place increased reliance on 
Ae visible and near-IR (MR) channels for daytime low cloud detection. 

Proposed research topics related to improved use of visible to MWIR data fall 
generally in the area of better cloud-free background characterization. The ratio of two- 
channel visible/near IR data has been used successfully to detect reflective clouds 
(Gustafson et al., 1994; Stowe et al., 1991; Saunders and Kriebel, 1988). The principle 
advantage of using a ratio, instead of the absolute magnitude of a single visible data channel 
compared to some threshold, is that as long as the channels are spectrally close, some 
uncertainties in the measurements (e.g., solar illumination, geometry) tend to cancel out. 
However, problems persist with reflective backgrounds erfiibiting ratio signatures similar 
to cloud (e.g., deserts, snow and ice). Development of databases of dear-scene ratio 
values to characterize the problematic backgrounds offer the potential to improve cloud 
discrimination. This approach retains the advantages of using channel ratios while relying 
on locally adaptive thresholds to account for changes in background characteristics. 
Single-charmel algorithms have traditionally used this type of approach to characterize the 
background reflectance (e.g., Hamill et al., 1992; Rossow and Schiffer, 1991). 

The AVHRR is the only extant sensor system that contains narrow-band chaimels in 
both the visible and near IR. All other environmental satellite systems are limited to a 
single, broad band "visible" chaimel covering that spectral range. Reflectance of most 
terrestrial surfaces at these wavelengths is anisotropic resulting in potentially large changes 
in measured cloud-free radiance as a result of relatively small changes in the solar and/or 
viewing geometry. This uncertainly in the cloud-free radiance has constrained the accuracy 
of single-chaimel techniques. Bi-directional Reflectance Distribution Functions (BRDF) 
are potentially useful for describing changes in surface reflectivity as a function of 
geometric considerations. Earlier work investigating the application of deterministic BRDF 
models to real-world conditions (d'Entremont et al., 1995; d'Entremont et al., 1996) has 
demonstrated their feasibility of using BRDF models for predicting the bandpass weighted 
dear-scene visible and near IR reflectance of known surfaces. BJ^F-derived dear-scene 
reflectance information can be applied to visible and NIR cloud detection algorithms to 
improve cloud-retrieval accuracy. Improved characterization of dear-scene reflectance 
conditions is also expected to increase the amount of visible to MWIR data that can be used 
for quantitative evaluation by objective cloud analysis algorithms. 

Knowledge of MWIR surface reflectance characteristics provides the potential for 
significantly improving the accuracy of operational cloud analysis models that use daytime 
3.7-3.9 |Xm data. During daylight hours surfaces both reflect solar energy and emit thermal 
energy at these wavelengths. If surface-reflectance effects can be successfully separated 
from thermal effects, improved detection and characterization of cloud (particularly thin 
transmissive cirrus) and more accurate skin temperature retrievals can be expected. MWIR 
reflectivity (r) can be expressed as a function of emissivity (e) only (i.e., r = 1-e), since the 
transmissivity of terrestrial surfaces at these wavelengths can be assumed to be zero. Thus, 
if the directional dependence of the surface emissivity can be calculated, then the directional 
dependence of the surface reflectivity is also known. LWIR emissivity is assumed to be 
1.0 for most terrestrial surfaces, therefore the physical surface skin temperature can be 
retrieved by applying an atmospheric correction to the observed LWIR brightness tempera¬ 
ture. This can then be used to compute the coincident MWIR emissivity from the corres¬ 
ponding atmospherically corrected MWIR brightness temperature. Directional dependence 


could be determined from a series of observations of the same geographic location over 
multiple orbits (i.e., varying view angles). GOES-8 and GOES-9 data along with NOAA 
AVHQRR data are all useful sources of coincident LWIR and MWTR observations. 

5.2.2 Transmissive Cirrus over Warm and Reflective Backgrounds 

When cirrus optical depth is sufficiently small, transmission of background radiation 
through the cloud can reduce the cloud-back^ound contrast to a level that cannot be 
detected by automated cloud algorithms. This can occur in visible, MWIR, and LWIR 
image bands, and is particularly severe over reflective and warm backgrounds such as 
desert. Under such conditions it may be possible to determine the presence of these 
transmissive cirrus using wavelengths wWe the atmosphere is opaque such as the 6.7-p,m 
water-vapor band. 

Water vapor imaging chaimels centered near 6.7 pm are available on most 
geostationary satellites and on IR sounders on the NOAA polar platforms. At these 
wavelengths the sensors measure upwelling thermal radiation emitted primarily by water 
vapor in the atmosphere, with little or no contribution from the underlying terrestrial 
surface. Under cloud-free conditions, the absolute magnitude of a 6.7-pm brightness 
temperature is dependent on the amount of water vapor that is present along the 
atmospheric path. If an estimate of the water vapor profile is available (e.g., from IR and 
microwave satellite sounders), a predicted dear-scene 6.7 pm brightness temperature can 
be calculated using radiative transfer code. Any cloud in a satellite-observed 6.7 pm field 
of view would produce a negative departure from the corresponding predicted brightness 
temperature. \\^e this approach is identical in form to standard single-channel threshold 
cloud tests, it effectively eliminates any problematic surface contributions that can act to 
mask weak cloud signatures. Sensitivity of this test can be evaluated using nighttime cases 
where conventional multispectral tests perform well for detecting thin cirms. This will 
provide an experience base that can be used to determine the usefulness of the water vapor 
channels for upper-level cloud detection. It is possible to employ this technique not only 
with 6.7-pm data but with any water-vapor imagery in the 6.5-7.5-pm spectral regions 
such as are typically available on infrared moisture sounders. Using adthtional water vapor 
sounding channels also offers the potential for discriminating multiple cloud layers at 
different levels in the atmosphere. 

Water vapor imagery is also potentially useful for improving cloud forecasts. Cross¬ 
correlation short-term cloud forecasting techniques have been demonstrated to produce 
useful 0 to 6-hour forecast products (Hamill and Nehrkom, 1993). These techniques 
generate short-term trajectory forecasts for a particular region using cloud-track wind 
vectors that have been computed by correlating cloudy areas in time series of geostationary 
data. Regions in later images that correlate most strongly with a cloudy region in an earlier 
image are used to define a displacement vector that is used to generate a trajectory cloud 
forecast. One of the problems in relying on cloud-track winds from satellite imagery is that 
no displacement vector data can be retrieved in cloud-free areas. No vectors can be 
obtained since there are no wind tracers in the clear regions from which to infer horizontal 
atmospheric motion. This in turn poses a problem when trying to advect clouds into 
previously cloud-free regions. One possibility for inferring motions in cloud-free regions 
is to study the correlation of water-vapor imagery patterns from one time to the next in 
cloud-free regions. This would provide a truly satelUte-only short-term cloud forecast 
capabiUty and could potentially be extended to derive wind profiles using water vapor 
channels that peak at differing levels in the atmosphere. With wind profile information it 
could then be feasible to advect low, middle, and high clouds separately with multiple wind 
fields that vary with altitude. 


5.2.3 Cloud Typing 

Cloud type information is often used as a surrogate to infer cloud physical and optical 
properties (e.g., physical and optical depth, drop or ice Ctystal size distribution, transmis¬ 
sion) due to the difficulty of retrieving these parameters directly from environmental 
satellite data. However, the cloud type retrievals themselves may be suspect. For 
example, the RTNEPH model currently identifies eight cloud types from only one visible 
and one IR channel (Hamill et al., 1992). Numerous approaches to cloud typing have been 
tried with varying degrees of success. A common technique discriminates type through a 
measure of the smface spatial texture generally evaluated through the local variance of 
infrared brightness temperature or visible reflectance (e.g., Lamaei, 1994; Hawkins and 
d'Entremont, 1990; Ebert 1989). Recently, neural networks have been used to assimilate 
multispectral satellite sensor data to retrieve a number of parameters including cloud type 
(e.g., Bankert, 1994, Lee et al., 1990). Garand (1988) stratified multispectrd satellite data 
into a taxonomy of classes and used a knowledge-base system to identify different cloud 
types. In SERCAA a hybrid technique is used 5iat first vertically stratified LWIR data and 
ttien applies standard image processing segmentation and classification techniques to dis¬ 
criminate cumuliform from stratiform cloud. Current Air Force requirements call for nine 
cloud types; cirrus, ckrostratus, altocumulus, altostratus, stratocumulus, stratus, cumulus, 
cumulonimbus, and nimbostratus. Robustness is a major issue in that any typing algorithm 
must be applicable to data from all available sensor systems over globally varying condi¬ 
tions. While most cloud typing techniques use some form of nephanalysis algorithm to 
provide cloud spatial information (e.g., location, extent, altitude), typically no use is made 
of cloud physical or optical properties available from advanced andysis algorithms. How¬ 
ever, numerous approaches exist to retrieve this type of information from existing sensor 
systems (e.g., Ou et al., 1993; d'Entremont et al., 1993; Nakajima and King, 1989). 

Processing time and efficiency are primary factors in operational implementations of 
cloud typing algorithms. If neural nets were to be shown feasible for cloud type retrieval 
this could significantly reduce the system resources required to perform that function. 

Also, depending on the quality of the training set and the information content in the satellite 
data, neural nets may be able to discriminate more cloud t 5 q)es than traditional retrieval 
techniques. It is important to note that the accuracy and robustness of neural networks are 
strongly tied to the amount and accuracy of the available training data; they are positively 

5.2.4 Cloud Altitude 

Cloud altitude is typically derived from LWIR brightness temperature data that are 
assumed to be representative of the cloud-top temperature (usually either an average or 
median brightness temperature computed from all cloudy pixels). Some temperature/height 
profile is then used to establish the cloud top altitude or pressure. Inherent assumptions are 
that derived LWIR brightness temperatures are representative of the cloud top temperature, 
that atmospheric transmission between the cloud and satellite can be ignored, and Aat the 
available temperature/height profile information is sufficiently accurate for the desired 
retrieval accuracy. 

More accurate and timely temperature and moisture profiles would improve not only 
fixing the height of the retrieved cloud-top temperature but also the accuracy of that 
temperature through modeling of the LWIR attenuation caused by the atmosphere above the 
cloud. Use of satellite-derived soundings provide much improved coverage in comparison 
to conventional radiosonde measurements, both in terms of geographic density and 
temporal frequency. Also, since they are coincident with the satellite imaging sensor 


measurements used for the cloud analyses, satellite-derived soundings are the optimal 
information source for quantifying the environment in which clouds are embedded. The 
sounder and imager data are best analyzed together since 1) the accuracy of the retrieved 
soundings is dependent on the presence of cloud (microwave-based soundings are less 
affected than IR based, particularly for non precipitating clouds) and 2) both cloud 
detection and altitude calculations are potentially improved with better understanding of the 
vertical structure of the atmosphere. 

Cloud top altitude is typically determined by comparing a measured cloud-top LWIR 
brightness temperature with its coincident atmospheric temperature profile and choosing the 
profile altitude that corresponds with the brightness temperature observation. Not ac¬ 
counted for are the possible transmissive nature of the clouds and the effects of atmospheric 
water vapor attenuation on upwelling LWIR brightness temperatures. In-situ or NWP- 
derived profiles of atmospheric temperature and water vapor are not always available in 
real-time. However, retrievals of atmospheric temperature and water vapor from collocated 
satellite LWIR and microwave sounder data ehminate the need for external data sources. 

The first step in correcting LWIR brightness temperature measurements for water 
vapor attenuation is to compute the atmospheric transmittance profile. Such computations 
can be quickly made using band-model transmittance codes such as those of Weinreb and 
Hill (1980) and Weinreb and Neuendorffer (1973). These models are well suited for this 
purpose since they can be customized for a satellite band-pass and response function, and 
they are fast, computationally efficient, and accurate. Input for these models includes 
atmospheric temperature and water vapor profiles that can be obtained from coincident 
satellite sounder data and/or NWP upper-air analysis fields. 

Once a transmittance profile is obtained, predicted infrared brightness temperature can 
be computed as a function of altitude for a cloud at that altitude. If a blackbody cloud is 
placed at some level in the atmosphere, the measured brightness temperature for that cloud 
will in general be lower than the actual cloud top temperature due to attenuation by the 
atmosphere above the cloud. This attenuation is computable by evaluating the radiative 
transfer equation in the forward direction using the band-model transmittance profile 
con-esponding to cloud-satellite atmospheric path. Thus for black clouds, true cloud top 
altitude as a function of satellite-measured brightness temperature can be computed a priori 
and used in lookup-table form to adjust satellite-measured cloud top temperature, and 
subsequently the cloud top height, for water vapor attenuation. 

For transmissive cirrus clouds, the non-blackbody effects on upwelling LWIR 
radiances must also be accounted for. This has been accomplished u si ng physical 
techniques such as those employed by d'Entremont et al. (1993), Ou et al. (1993) 
dEntremont et al. (1990) and Rao et al. (1995). These models all use a deterministic 
approach to account for the low bias in retrieved cirrus altitudes caused by an incorrect 
blackbody assumption. However, the topic of cirrus radiative and spatial property retrieval 
using daytime MWIR and LWIR data remains problematic. During nighttime, the physical 
retrieval of cimis emissivity and effective altitude (two unknowns) is accomplished nstn g a 
coincident pair of MWIR and LWIR thermal-only radiance measurements (two knowns). 
The retrieval process capitalizes on the thermal properties of cirrus clouds at the 3.7-|im 
and 10.7-p.m wavelengths. During daytime this retrieval is complicated by the addition of a 
reflected solar component to the measured MWIR radiances, essentially adding a thir d 
unknown to the system without the addition of a third measurement. Thus for daylight 
observations there are three unknowns (emissivity, altitude, and 3.7-|j,m solar reflection) 
and only two measurements (a combined solar-plus-thermal MWIR radiance and a thermal- 
only LWIR radiance). Before the cirrus optical and radiative properties can be retrieved, 
the solar component of the MWIR radiance observation must be removed. Rao et al. 


(1995) have presented and successfully implemented a deterministic technique for removal 
of the MWIR solar component. Their technique requires a priori computations of very 
complex processes that model the MWIR bidirectional scattering properties of ice crystals. 

An alternative solution to the daytime MWIR-LWIR cirrus retrieval problem requires 
separation of the solar and thermal MWIR components from the overall MWIR 
measurement. Recall that the fundamental cirrus retrieval model requires as input the 
upwelling MWIR and LWDR thermal-only radiances. Cirrus effective altitude can be 
independently retrieved both day and night using data along the wing of an LWIR 
absorption band using a technique such as CO 2 Slicing (Wylie and Menzel, 1989; Menzel 
et al., 1992,1993). CO 2 slicing generates cirrus effective altitude analyses using 15-|xm 
sounder data that are available with GOES and NOAA sounder data. Using this approach, 
the effective altimde can be eliminated from the list of unknowns and the imager M^TR- 
LWIR retrievals of the cloud effective emissivity can proceed. 

A cirrus BRDF for 3.7-|im reflected solar energy using coincident measures of CO 2 
effective altitude and MWIR-LWIR radiances could be used to specify the reflected-solar 
MWIR radiance component in situations where CO 2 Slicing analyses are not available. As 
discussed above, daytime retrieval of cirrus attributes is initially posed as a problem with 
three unknowns and two measurements, which in general has no mathematically unique 
solution. Using CO 2 Slicing to eliminate effective citrus altitude from the unknowns, the 
reflected-solar portion of the MWIR cirrus radiance can then be retrieved whenever 
coincident CO 2 Slicing analyses are available. Tabulated retrievals over a range of view 
and illumination angles will be used to constmct a cirrus MWIR BRDF that will be used 
where CO 2 Slicing analyses are not available. The resultant two-channel cirrus retrievals 
can be compared to results using the technique of Rao et al. (1995). Also, to check for 
consistency, the cirrus BRDFs can be directly and quantitatively compared with the full¬ 
blown radiative transfer (scattering by ice crystals) calculations of Rao et al. (1995). Other 
issues that will require investigation are the potential dependencies of measured cirrus 
BRDFs on effective ice crystal size (effective temperature) and optical depth (effective 

5.2.5 Temporal Consistency 

As discussed in Section 4.6 above, integrated analyses produced by combining data 
from multiple satellite platforms to produce a single unified or integrated cloud analysis can 
exhibit inconsistencies between collocated results obtained from different satellite platforms 
at approximately the same time. The problem consists of combining data from high spatial 
and spectral resolution polar-orbiting satellites with high temporal resolution geostationary 
data. Under some conditions geostationary-based analyses tend to identify less cloud than 
corresponding polar-based andyses. Investigation of a limited number of cases has 
revealed no apparent problems with the accuracy of any of the individual, source-specific 
analyses. However, when combined the temporal consistency of the resultant integrated 
analysis can be degraded. This in turn may negatively impact subsequent applications that 
use tire cloud analysis data as input, such as cloud forecast models. Depending on the 
source, age and mix of satellite data available over a given time period, certain clouds can 
appear in an integrated analysis only to later disappear and then reappear as different data 
sources are used and subsequently aged out of the analysis. 

Inconsistencies between polar- and geostationary-based analyses are most likely due 
to inherent differences in their respective scan methodologies and major sensor 
characteristics (i.e., the available information content varies with platform). Table 4 
summarizes the imaging sensor characteristics of the currently operational environmental 


satellites. From the table it can be seen that each data source exhibits characteristic 
strengths that contribute to the relative accuracy of derived cloud products. OLS data have 
high spatial resolution but are relatively low in spectral information content and temporal 
resolution. AVHRR has the greatest spectral information content and its repeat cycle is 
consistent with synoptic scale events, but the spatial resolution is relatively coarse. 
Geostationary data exhibit the greatest tempor^ resolution, but the spatial resolution is 
significantly de^aded for a large fraction of the coverage area, and spectral information 
content varies widely with satelhte. 

An investigation of the magnitude and frequency of inconsistencies between cloud 
analyses derived from different satellite systems and of their root causes is needed to 
characterize the problem more completely. To date only a few cases have been documented 
where differences between the polar- and geostationary-based analyses exist. It is not 
known how often this situation occurs or whether it is limited only to specific types of 
cloud (e.g., low cloud, transmissive cirrus). The cases that have been identified are 
consistent with conditions described above where the accuracy of the individual source- 
sp^ific analyses have been confirmed through visual inspection yet discrepancies between 
coincident geo- and polar-based analyses exist. One possible explanation is the relatively 
coarse resolution, both spatially and spectrally, of geostationary compared to the polar- 
orbiter data. Geostationary sensors used in the earlier studies have a spatial resolution of 5 
km at the subpoint that degrades to approximately 17 km at the edge of scan. Also they 
were only two-channel instruments with a radiometric resolution of no better than 8 bits. 
This is in contrast to the five-channel AVHRR data with a 10 bit radiance quantization 
range and 4 km footprint degrading only near the edges of scan, and the OLS with an 8 bit 
quantization and a roughly fixed 2.7 km footprint throughout the scan. The data quality 
hypothesis can be tested using the new GOES-8 data now available in real-time at PL. The 
imaging sensor on GOES-8 has additional charmels plus improved spatial and spectral 
resolution that should increase cloud detection sensitivity and possibly reduce differences in 
the derived cloud analyses as compared to polar-based analyses. However, even if the 
improved data quality from the new GOES satellites does resolve the geostationary vs. 
polar consistency issue for those platforms, the problem will remain for the other 
international geostationary satellites (i.e., Meteosat and GMS). 

To the extent inconsistencies exist between cloud data analyzed from different 
satellites, and the inconsistencies are detrimental to subsequent applications, mitigation is 
probably best achieved through modifications to the integration algorithm used to combine 
the separate analyses. This could be made more effective if inconsistency proble ms are 
linked to a particular range of conditions or cloud types that can be identified algorithmi¬ 
cally. An approach capable of quantifying system related impacts on analysis charac¬ 
teristics derived from geostationary satelhtes is to selectively degrade GOES 8 capabilities 
and observe the impact on cloud analysis accuracy and consistency. GOES 8 has improved 
radiometric and spatial resolution plus additional spectral capabilities (i.e., additional 
channels) relative to other geostationary systems. Sensitivity to system capabilities can be 
derived by parametrically varying GOES 8 system parameters (e.g., sensor or spectral 
resolution, channel configuration) and tabulating the change in retrieved cloud properties. 

The principal objective of such sensitivity studies would be to identify cloud types or 
characteristics for which retrieval accuracy are most heavily impacted by the degraded 
system capabilities of geostation^ satellites. Using this information, a mitigation strategy 
could be developed to best exploit the relative strengths of the individual satellites. For 
example, if it is demonstrated that geostationary analyses poorly resolve, say, low stratus 
cloud, then the inte^ation algorithm could be modified to selectively persist low cloud 
identified from earlier polar passes, even in cases where more recent geostationary analyses 
indicate clear. Another strategy is to use a short-term cloud forecast algorithm such as 


cross correlation (Hamill and Nehrkom, 1993) to extend the influence of the polar-based 
analyses for longer time periods. A similar approach can be applied to analyzed data from 
nearly-coincident NOAA and DMSP scans (e.g., at high latitudes) to identify any cloud 
types that are sensitive to sensor-specific system characteristics. 

5.3 Validation 

A recurring criticism of objective satellite-based cloud analysis algorithms is the 
general lack of formal validation. There are several reasons for the paucity of quantitative 
validation work, however, most derive from a single cause: the absence of independently 
derived or measured cloud parameter databases with an established accuracy sufficient to be 
classified as ground truth. 

A need exists for an established procedure and accepted objective measures of 
algorithm accuracy over a range of conditions. The initial step in development of an 
algorithm validation methodology is to first identify certain fundamental study attributes: 

• goal: what information is required from the validation study, what is the end 
user's application and what information is needed to determine if the subject 
algorithm can provide that information with sufficient accuracy, does the 
accuracy metric need to be absolute or relative to some known standard (e.g., an 
existing legacy technique); 

• scope: does the subject cloud algorithm need to be applied globally or 
regionally, only over certain background conditions (e.g., ocean, land, desert, 
snow), day and night, to specific satellite or sensor types (e.g., polar, 
geostationary, infrared, microwave); 

• supporting data: what supporting data are available for use as part of the 
validation study, is there a truth data set that is accepted by the end user, what is 
the scale, resolution, projection of the supporting data relative to the satellite 
data, are the supporting or truth data parameters directly relatable to the algorithm 
output parameters; and 

• results: is there specific quantitative information required by the end user, what 
statistical measures or form of results are required, are these results sufficient to 
meet the goals. 

Comparisons with trath sets, if they are available, provide an absolute metric that is 
potentially useful for Air Force programs such as SERCAA and CDFS n that need to 
gauge algorithm performance in terms of absolute accuracy. Given the requirements placed 
on cloud algorithms for near real time global analyses, any ground truth data set would 
need to be representative of globally varying surface and atmospheric conditions for a full 
range of seasonal and diurnal conditions. The only practical data source capable of 
providing the required level of coverage is environmental satellites. Thus to date, most 
validation studies that require an absolute accuracy metric have involved some form of 
subjective comparison between satellite-derived algorithmic results and the corresponding 
satellite imagery; and even those studies tend to be limited to validation of cloud spatial 
properties (e.g., coverage, amount, height). 

Other Air Force programs have requirements for cloud climatologies, rather than real¬ 
time analyses. This drives a need to establish a measure of relative accuracy between 
available climatological data and to quantify any differences. Intercompaiisons over 


extended time periods between results from real-ti me cloud analysis algorithms and 
coincident data being ingested into candidate climatological data sets are potentially useful 
for establishing any characteristic biases or artifacts in the climatologies, particularly if the 
absolute accuracy of the cloud algorithms is previously established. Radiative and 
microphysical retrieval algorithms can realistically only be validated against in-situ and/or 
active remotely sensed data. Well calibrated lidars and radars can provide necessary 
information on cloud altitude and optical properties but are typically only available from 
intensive field study programs for a limited range of meteorological conditions and a single 
geographic region for perhaps one season. These data need to be collected from as many 
locations and time periods as possible to provide a more complete validation of the retrieval 

An acceptable vahdation methodolo^ must identify factors or techniques that can 
minimize problems associated with the limited availability of truth data and that can provide 
the required accuracy information. There is evidence that using a man-computer technique 
to analyze satellite imagery is an acceptable approach for developing ground truth 
information for some applications. Gustafson et al. (1993) produced a quantitative 
me^ure of the absolute accuracy of total cloud amount retrievals produced by a set of 
tactic^ cloud analysis algorithms. Their approach used direct comparison with ground 
tmth images produced through manual analysis of satellite imagery following a formalized 
technique developed by Gustafson and Felde (1988). This technique has also been 
successfully applied to previous studies of the RTNEPH (d'Entremont et al., 1989) and an 
SSM/I cloud algorithm (Gustafson and Felde, 1989). 

Comparison of results from multiple cloud models to determine the optimal approach 
for a paiticidar application is another common validation requirement (e.g., Rossow et al., 
1985). During the recent cloud algorithm development program conducted at the Phillips 
Laboratory, a validation study was designed and executed to establish a relative accuracy 
metric for comparison of cloud algoritlun results against a known standard (P hilli ps 
Laboratory, 1994). The focus of diis study was to identify the impact on analysis accuracy 
of the differing analysis approaches used by the respective algorithms. Two validation 
criteria were established: 1) did differences between the two techniques exceed a specified 
level of statistical significance and 2) where they were different, which was more accurate. 
To minimize the impact on comparison statistics of minor algorithmic differences such as 
output resolution and precise positioning of cloud edges, the output data were first gridded 
and stratified into three categories: clear, partly cloudy, and completely cloudy. Direct 
comparison statistics were then computed using collocated output pairs taken from the two 
algorithms at matching analysis times. A ^2 test was used to demonstrate that the 
computed differences were not attributable to random chance. A dichotomous test using a 
single-blind manual evaluation was then applied to establish if one analysis was superior 
for those cases where the results differed. A similar approach was used by Welch et al. 
(1992) to intercompare results from three independent cloud classification schemes and 
establish a relative accuracy metric for each algorithm. 

The above-cited studies provide useful insight into possible validation approaches to 
detemune the accuracy or applicability of candidate cloud algorithms to a given application. 
Note that in each case, available supporting data were used in the way best suited to meet 
the goals of the individual studies. This insured that the information required to establish 
the usefulness of the analysis approach to the desired application would be provided. In 
some cases the amount of available data was too small to allow for comprehensive 
conclusions, however in each case the study met the stated goals. 


6. Summary 

Development of improved automated nephanalysis capabilities from multiplatform 
sensor data is an important goal of the Strategic Environmental Research and Development 
Program (SERDP). One such SERDP initiative is the Support of Environmental Require¬ 
ments for Cloud Analysis and Archive (SERCAA) project. SERCAA as a two-phase 
research and development program has provided both the next generation nephanalysis 
model for CDFSII and a new global cloud algorithm for use in determining the radiative 
and hydrological effects of clouds on climate and global change. SERCAA cloud analysis 
products will be available to a wide community of users both within and outside of the 
Department of Defense. The principal accomplishments of SERCAA are: 1) incorporation 
of high-resolution sensor data from multiple military and civilian satellites, polar and geo¬ 
stationary, into an integrated real-time cloud analysis model, 2) demonstration of multi- 
spectral cloud analysis techniques that improve the detection and specification of clouds, 
especially cirrus and low clouds, and 3) development of cloud microphysical and radiative 
property retrieval algorithms and support database specifications for an improved cloud 

7. References 

Adler, R.F. and A. J. Negri, 1988: A satellite Infrared technique to estimate tropical 
convective and stratiform rainfall. J. Appl. Meteor, 27, 30-51. 

Adler, R.F., R.A. Mack, N. Prasad, H.-Y.M. Yeh and I.M. Hakkarinen, 1990: Aircraft 
microwave observations and simulations of deep convection from 18 to 183 GHz. 
Part I: Observations, J. Atmos. Ocean. Technol., 7, 377-391. 

Ackerman, S.A., W.L. Smith and H.E. Revercomb, 1990: The 27-28 October 1986 FIRE 
DFO cirms case study: spectral properties of cirrus clouds in the 8-12 micron 
window. Mon. Wea. Rev., 118, 2377-2388. 

AMS Council, 1992. Needs and Opportunities in Mesoscale Research - A Call for Action. 
Bull. Amer. Meteor. Soc., 73, 63-65. 

Arking, A. and J. Childs, 1985: Retrieval of cloud cover parameters from multispectral 
satellite images. J. Clim. and Appl. Met, 24:322-333. 

Bankert, R.L., 1994: Cloud classification of AVHRR imagery in maritime regions using a 
probabilistic neural network. J. Appl. Meteor., 33, 909-918. 

Bunting, J.T. and R.P. d’Entremont, 1982: Improved cloud detection utilizing Defense 

Meteorological Satellite Program near Infrared measurements. AFGL-TR-82-0027, 
Air force Geophysics Laboratory, ADAl 18751. 

Chahine, M.T., 1992: GEWEX: The Global Energy and Water Cycle Experiment. EOS, 
Trans, of the Amer. Geophys. Union, 73, 9-14, 1992. 

Corell, R.W., 1990: The United States Global Change Research Program (US/GCRP): 

An Overview and Perspectives on the FY 1991 Program. Bull. Amer. Meteor. 
Soc., 71, 507-511. 


d’Entremont, R.P., 1986: Low and mid-level cloud analysis using nighttime multispectral 
imagery. J. Clim. andAppl. Met., 25:1853-1869. 

d’Entremont, R. P., and L. W. Thomason, 1987: Interpreting Meteorological Satellite 

Images Using a Color-Composite Technique. Bull. Amer. Meteor. Soc., 68, 762- 

d’Entremont, R. P., L. W. Thomason, and J. T. Bunting, 1987: Color-composite Image 
Processing for Multispectral Meteorological Satellite Data. Proceedings of Digital 
Image Processing and Visual Communications Technologies in Meteorology, 
Cambridge, MA. SPIE, the International Society for Optical Engineering, 846, 

d’Entremont, R. P., G.B. Gustafson, J.T. Bunting, M.K. Griffin, C. Barker Schaff, P.L. 
Nowak and J.M. Ward, 1989: Comparisons between the RTNEPH and AFGL 
cloud layer analysis algorithms. AFGL-TR-89-0175, ADA216637. 

d'Entremont, R. P., M. K. Griffin, and J. T. Bunting, 1990: Retrieval of Cirrus Radiative 
Properties and Altitudes Using Multichannel Infrared Data. Proceedings, Fifth 
Conference on Satellite Meteorology and Oceanography, Amer. Meteor. Soc., 3-7 
September 1990, London, England, 4-9. 

d'Entremont, R. P., D. P. Wylie, J. W. Snow, M. K. Griffin, and J. T. Bunting, 1992: 
Retrieval of Cirrus Radiative and Spatial Properties Using Independent Satellite 
Data Analysis Techniques. In Preprints, Sixth Conference on Satellite Meteorology 
and Oceanography, Jan. 5-10, Atlanta, GA, AMS, Boston, pp. 17-20. 

d'Entremont, R.P., D. P. Wylie, D. C. Peduzzi, and J. Doherty, 1993: Retrieval of Cirrus 
Radiative and Spatial Properties Using Coincident AVHRR and HIRS Satellite 
Data. Proceedings, Passive Infrared Remote Sensing of Clouds and the 
Atmosphere. SPIE 1934, 13-15 April 1993, Orlando, 180-196. 

d'Entremont, R. P., D. P. Wylie, S-C. Ou and K-N. Liou, 1995: Retrieval of cirrus 

radiative and spatial properties using coincident multispectral imager and sounder 
satellite data. Preprints, Cloud Impacts on DoD Operations and Systems, Oct. 24- 
26, Hanscom AFB, MA, pp. 155-156. 

d'Entremont, R. P., D. P. Wylie, S-C. Ou and K-N. Liou, 1996: Retrieval of cirrus 

radiative and spatial properties using coincident multispectral imager and sounder 
satellite data. Preprints, Eighth Converence on Satellite Meteorology and 
Oceanography, Jan 28 - Feb 2,1996, AMS, Boston, MA. 

Deirmendjian, D., 1969: Electromagnetic scattering on spherical polydispersions. 

Elseveier, New York. 

Diem, M., 1948: Messungem der Frosse Von Walkendementen H. Meteor. Rundschau, 9, 

Di Girolamo, L. and R. Davies, 1995: The image navigation cloud mask for the Multi¬ 
angle Imaging SpectroRadiometer (MISR). J. Atmos, and Oceanic Tech., 12. 

Ebert, E.E., 1989: Analysis of polar clouds from satellite imagery using pattern 

recognition and a statistical cloud analysis scheme. J. Appl. Meteor., 28, 282-299. 


Eyre, J.R. and W.P. Menzel, 1989: Retrieval of cloud parameters from satellite sounder 
data: A simulation study. J. Appl. Meteor., 28, 267-275. 

Falcone, V.J. and R.G. Isaacs, 1987: The DMSP microwave suite. In Proceedings, 
NOAA Conference on Passive Microwave Observing from Environmental 
Satellites, Tech. Rpt. NESDIS 35, pg. 174-185. U.S. Dept, of Commerce, 
Washington, DC. 

Falcone, V.J., M.K. Griffin, R.G. Isaacs, J.D. Pickle, J.F. Morrissey, A.J. Jackson, A. 
Bussey, R. Kakar, J. Wang, P. Racette, D.J. Boucher, B.H. Thomas, A.M. 
Kishi, 1992: SSM/T-2 Calibration and Validation Data Analysis, PL-TR-92-2293, 
Phillips Laboratory, Hanscom AFB, MA 01731-5000, ADA265817. 

Garand, L., 1988: Automated recognition of oceanic cloud patterns. Part I: Methodology 
and application to cloud climatology. J. Climate, 1, 20-39. 

Griffin, M.K., V.J. Falcone, J.D. Pickle, and R.G. Isaacs, 1994: SSM/T-2 Brightness 
Temperature Signatures, AMS 7th Conference on Satellite Meteorology and 
Oceanography Proceedings, Monterey, CA, 110-113. 

Gustafson, G.B. and G.W. Felde, 1988: Interactive satellite image processing applied to 
cloud detection. Preprints: 4th Conference on Interactive Information 
Processing Systems, American Meteorological Society, Boston, MA. 

Gustafson, G.B. and G.W. Felde, 1989: Validation of automated cloud detection from 
microwave imagery. Preprints: 5th Conference on Interactive Information 
Processing Systems, American Meteorological Society, Boston, MA. 

Gustafson, G.B. and R.P. d’Entremont, 1992: Single channel and multispectral cloud 
algorithm development for TACNEPH. In: Preprints, Sixth Conference on 
Satellite Meteorology and Oceanography, American Meteorological Society, pp. 

Gustafson, G. B., R. G. Isaacs, J. M. Sparrow, J. T. Bunting, and R. P. d'Entremont, 
1993: Validation of infrared cloud detection algorithms developed for TACNEPH. 
Proceedings Passive Infrared Remote Sensing of Cloud and the Atmosphere. 
Orlando, 13-15 April 1993, SPIE, the International Society for Optical 
Engineering, 1934, 62 - 71. 

Gustafson, G.B., R.G. Isaacs, R.P. d’Entremont, J.M. Sparrow, T.M. Hamill, C. 
Grassotti, D.W. Johnson, C.P. Sarkisian, D.C. Peduzzi, B.T. Pearson, V.D. 
Jakabhazy, J.S. Belfiore, and A.S. Lisa, 1994: Support of Environmental 
Requirements for Cloud Analysis and Archive (SERCAA): Algorithm 
Descriptions. PL-TR-94-2114, Phillips Laboratory, Hanscom AFB, MA, 

ADA 283240,100 pp. 

Gustafson, G.B., R.P. d'Entremont and D.C. Peduzzi, 1995: Cloud data sets derived 
from combined geostationary and polar-orbiting environmental satellite sensors 
using the SERCAA cloud model. Preprints, Cloud Impacts on DoD Operations and 
Systems, Oct. 24-26, Hanscom AFB, MA, pp. 79-90. 

Hakkarinen, I.M. and R. F. Adler, 1988: Observations of precipitating convective systems 
at 92 and 183 GHz: aircraft results. Meteorol. Atmos. Phys., 38, 164-182. 


Hale, G. M., and M. R. Querry, 1973: Optical constants of water in the 200 nm to 200|im 
wavelength region, Apl. Optics, 12, 555-563. 

Haimll. T. M., and R. Hoffman, 1993: SERCAA Cloud Analysis Integration: Design 
Concepts and Interaction with Cloud Forecast Models. Phillips Laboratory Tech. 
Report PL-TR-93-2100, Phillips Laboratory Geophysics Directorate, Hanscom 
AFB, MA, 54 pp., ADA 269104. 

Hamill, T. M., and T. Nehrkom, 1993: A Short-Term Cloud Forecast Scheme Using 
Cross Correlations. Wea. and Forecasting, 8, 401-411. 

Hamill, T.M., R.P. d'Entremont, and J.T. Bunting, 1992: A Description of the Air Force 
Real-Time Nephanalysis Model, Wea. and Forecasting, 7,288-306. 

Hawkins, R. S., and R. P. d’Entremont, 1990: Automated Cloud Typing Using Satellite 
Imagery. GL-TR-90-0326, Hanscom AFB, MA, 6 pp, ADA230492. 

Heideman, K. F., J. M. Sparrow, and T. S. Lisa, 1994: Validation of the SERCAA Cloud 
Analysis Algorithm. Proceedings, Seventh Conference on Satellite Meteorology 
and Oceanography, Amer. Meteor. Soc., 6-10 June 1994, Monterey CA, 326 - 

Heideman, K. F., R. P. d'Entremont, J. M. Sparrow, T. S. Lisa, and G. B. Gustafson, 
1995: SERCAA Integrated Cloud Analysis Validation. Preprint of the Cloud 
Impacts on DoD Operations and Systems 1995 Conference (CIDOS-95), D. 
Grantham, ed., Phillips Laboratory Tech. Paper PL-TR-95-2129, Hanscom AFB 
MA, 1 Oct 95, ADA300914. 

Hunt, G. E., 1973: Radiative Properties of Terrestrial Clouds at Visible and Infrared 

Thermal Window Wavelengths. Quart. Joum. Royal Meteor. Soc., 99,346-369. 

Inoue, T., 1987: A Cloud Type Classification with NOAA-7 Split-Window Measurements. 
Joum. Geophys. Res., 92, No. D4, 3991-4000. 

Isaacs, R. G., W, C. Wang, R. D. Worsham, and S. D. Goldenberg, 1987: Multiple 
scattering treatment for use in the LOWTRAN and FASCODE models. Appl 
Optics, 26, 1272-1281. ' 

Isaacs, R.G., E. Kalnay, G. Ohring, R. McClatchey, 1993: Summary of the 

NMC/NESDIS/DoD Conference on DMSP Retrieval Products. Bull. AMS. 74-1 

onr rvi —^—y • j 

Kiess, R.B. and W.M. Cox, 1988: The AFGWC automated real-time cloud analysis 
model. Tech. Note 88/001, AFGWC. 

King, M.D., Y.J. Kaufman, W.P. Menzel, and D. Tanre, 1992: Remote sensing of cloud, 
aerosol, and water vapor properties from the Moderate Resolution Imaging 
Spectrometer (MODIS). IEEE Trans. Geosci. Rem. Sens., 30, 2-27. 

Kneizys, F. X., E. P. Shettle, W. O. Gallery, J. H. Chetwynd Jr., L. W. Abreu, E. A. 
Selby, S. A. Clough, and R. W. Perm, 1983: Atmospheric transmittance/radiance’ 
Computer Code LOWTRAN6. AFGL-TR-83-0187, Air Force Geophysics 
Laboratory. ADA137786. 


Kopp, TJ., T. J. Neu, and J.M. Lanicci, 1994a: A Description of the Air Force Global 
Weather Central's Surface Temperature Model. Preprints, 10th Conf. Num. Wea. 
Pred., Portland, Amer. Meteor. Soc., 435-437. 

Kopp, T. J., T. J. Neu, and P. J. Broil, 1994b: Application of SSM/I to a real time 

automated satellite nephanalysis model. 7th Conference of Satellite Meteorology 
and Oceanography. American Meteorological Society, Boston MA, pp 247-250. 

Kopp, T. J., S. G. Zahn, and R. B. Kiess, 1996: Application of a convective stratiform 
technique in an operational global nephanalysis. 8th Conference of Satelhte 
Meteorology and Oceanography. American Meteorological Society, Boston MA, pp 

Kummerow, C., I.M. Hakkarinen, H.F. Pierce, and J.A. Weinman, 1991: Determination 
of precipitation profiles from airborne passive microwave radiometric 
measurements. J. Atmos. Oceanic Tech., 8, 148-158. 

Lamaei, N., 1994: Cloud-type discrimination via multispectral textural analysis. Opt. 
Engineer., 34, 1303-1313. 

Lee, J., R. Weger, S.K. Sengupta, and R.M. Welch, 1990: A neural network approach to 
cloud classification. IEEE Trans. Geosci. Rem. Sens., 28, 846-855. 

Li, X., and A. H. Strahler, 1992: Geometric-Optical Bidirectional Reflectance Modeling of 
the Discrete Crown Vegetation Canopy: Effect of Crown Shape and Mutual 
Shadowing. IEEE Trans. Geosci. and Remote Sens., 30, 276-292. 

Lorenc, A., 1981: A global three-dimensional multivariate statistical interpolation scheme. 
Mon. Wea. Rev., 109: 701-721. 

Menzel, W.P. and J.F.W. Purdom, 1994: Introducing GOES-I: The first of a new 
generation of geostationary operational Environmental satellites. Bull. Amer. 
Meteor. Soc., 75, 757-781. 

Menzel, W.P., D.P. Wylie, and K.I Strabala, 1992: Seasonal and diurnal changes in 
cirrus clouds as seen in four years of observations with the VAS. J. Appl. 

Meteor., 31, 370-385. 

Menzel, W.P., D.P. Wylie, and K.I. Strabala, 1993: Trends in global cirrus inferred from 
three years of HIRS data. Proceedings, Seventh Int'l TOYS Conference, Igls, 
Austria, 10-16 February 1993. 

Menzel, W. P., and J.F. W. Purdom, 1994: Introducing GOES-I: The First of a New 
Generation of Geostationary Operational Environmental SateUites. Bull. Amer. 
Meteor. Soc., 75, 757-781. 

Moncet, J.L. and R.G. Isaacs, 1994: Unified retrieval of atmospheric temperature, water 
substance, and surface properties from the combined DMSP sensor suite. 

Preprints, 7th Conference on Satellite Meteorology and Oceanography, American 
Meteorological Society, June 6-10, Monterey, CA, 489-492. 


Nakajima, T. and M.D. King, 1989: Cloud optical parameters as derived from the 

multispectral cloud radiometer. In Lenoble and Gelyn, eds, IRS ‘88, pgs. 18-21. 
A. Deepak Pub. 

Nehrkom, T. and M. Zivkovic, 1996: A comparison of cloud cover schemes. Mon. 
Weather Rev. (in press). 

Neu, T.J., R.G. Isaacs, G.B. Gustafson and J.W. Snow, 1994: Improved cloud analysis 
for CDFS n through the SERCAA research and development program. Preprints, 
7th Conference on Satellite Meteorology and Oceanography, June 6-10, 
Monterey, CA, 239-242. 

Ou, S. C., K. N. Liou, W. M. Gooch, and Y. Takano, 1993: Remote Sensing of Cirrus 
Cloud Parameters Using Advanced Very High Resolution Radiometer 3.7- and 
10.9-mm Channels. Appl. Opt., 32, 2171-2180. 

Petty, G. W., 1994: Physical retrievals over-ocean rain rate from multichannel microwave 
imagery. Part II: Algorithm Implementation. Meteorol. Atmos. Phys., 54, p 
101 - 121 . ^ 

Petty, G. W., 1995: The status of satellite-based rainfall estimation over land, Remote 
Sensing Environment, 51, p 125-137. 

Petty, G. W., and D. K. Miller, 1995: Satellite microwave observations of precipitation 
correlated with intensification rate in extratropical oceanic cyclones, Mon. Weather 
Rev., 123, No. 6, p. 1904-1911. 

Pickle, J. R. G. Isaacs, V. Jakabhazy, M. Griffin, and V. J. Falcone, 1996: Detection of 
precipitation using SSM/T-2 Measurements. 8th Conference of Satellite 
Meteorology and Oceanography. American Meteorological Society, Boston MA, pp 

Phillips Laboratory, 1994: SERCAA integrated cloud analysis validation using satellite 
data from May 1993. Phillips Laboratory Technical Memorandum, 3 June 1994. 

Prabhakara, C. J.-M. Yoo, G. Dalu, and R. S. Fraser, 1990: Deep Optically Thin Cirrus 
Clouds in the Polar Regions. I - Infrared Extinction Characteristics. Joum. Appl 
Meteor., 29, 1313 - 1329. 

Rao, N. X., S. C. Ou, and K. N. Liou, 1995: Removal of the Solar Component in 

AVHRR 3.7-nim Radiances for the Retrieval of Cirrus Cloud Parameters. Joum. 
Appl. Meteor., 34, 482-499. 

Ross, J.K., and A. Marshak, 1985: A Monte Carlo Procedure for Calculating the 

Scattering of Solar Radiation by Plant Canopies. Sov. Joum. Remote Sens 4 
783-801. ■’ ’ 

Rossow, W.B. and R.A. Schiffer, 1991: ISCCP Cloud Data Products. Bull. Amer. 
Meteor. Soc., 72, 1, 2-20. 


Rossow, W., F. Mosher, E. Kinsella, A. Arking, M. Debois, E. Harrison, P. Minnis, E. 
Ruprecht, G. Seze, C. Simmer and E. Smith, 1985: ISCCP Cloud Algorithm 
Intercomparison. J. Clim. Appl. Met, 24, 877-903. 

Roujean, J-L., M. Leroy, and P-Y. Deschamps, 1992: A Bidirectional Reflectance Model 
of the Earth's Surface for the Correction of Remote Sensing Data. Joum. Geophys. 
Res., 97 D18, 20455-20468. 

Saunders, R.W. and K.T. Kriebel, 1988: An improved method for detecting clear sky and 
cloudy radiances from AVHRR data. Int. J. Remote Sensing, 9:1, 123-150. 

Schiffer, R.A. and W.B. Rossow, 1983: The International Satellite Cloud Climatology 
Project (ISCCP): The First Project of the World Climate Research Program. Bull. 
Amer. Meteor. Soc., 64, 779-784. 

Schiffer, R.A. and W.B. Rossow, 1985: ISCCP Global Radiance Data: A New Resource 
for Climate Research. Bull. Amer. Meteor. Soc., 66, 1498-1505. 

Smith, W.L., H. E. Revercomb, H. B. Howell, and M.-X. Lin, 1990: Multi-spectral 

Window Radiance Observations of Cirrus From Satellite and Aircraft, November 2, 
1986 Project FIRE. FIRE Science Results 1988, NASA Langley Res. Ctr., 89 - 

Spencer, R.W., H.M. Goodman, R.E. Hood, 1989: Precipitation retrieval over land and 
ocean with the SSM/I: identification and characterization of the scattering signal. J. 
Atmos, and Ocean. Tech., 6, 254-274. 

Stamnes, K., S. C. Tsay, W. Wiscombe, and K. Jayaweera, 1988: Numerically stable 
algorithm for discrete ordinate method radiative transfer in multiple scattering and 
emitting layered media. Appl. Optics, 18,994-1008. 

Stephens, G. L., 1979: Optical properties of eight water cloud types. Commonwealth 

Scientific and Industrial Research Organization, Division of Atmospheric Physics, 
Tech. Paper #36, Australia. 

Stowe, L.L., E.P. McClain, R. Carey, P. Pellegrino, G.G. Gutman, P. Davis, C. Long 
and S. Hart, 1991: Global distribution of cloud cover derived from 
NOAA/AVHRR operational satellite data. Adv. Space Res., 3,51-54. 

Strabala, K.I., S.A. Ackerman and W.P. Menzel, 1994: Cloud Properties inferred from 8- 
12 micron data. J. Appl. Meteor., 33, 212-229. 

Takano, Y., and K. N. Liou, 1989: Solar Radiative Transfer in Cirrus Clouds. Part I: 

Single Scattering and Optical Properties of Hexagonal Ice Crystals. Joum. Atmos. 
Sci., 46, 3 - 19. 

Walthall, C. L., J. M. Norman, J. M. Welles, G. Campbell, and B. L. Blad, 1985: Simple 
Equation to Approximate the Bidirectional Reflectance for Vegetative Canopies and 
Bare Soil Surfaces. Appl. Optics, 24, 383-387. 

Weinreb, M. P., and A.C. Neuendorffer, 1973: Method to Apply Homogeneous-Path 

Transmittance Models to Inhomogeneous Atmospheres. Joum. Atmos. Sci., Amer. 
Meteor. Soc., 30, 662-666. 


Weinreb, M.P., and M. L. Hill, 1980: Calculation of Atmospheric Radiances and 

Brightness Temperatures in Infrared Window Channels of Satellite Radiometers. 
NOAA Tech. Rep. NESS 80, U.S. Dept, of Commerce, Washington, D.C., 40 


Welch, R.M., S.K. Sengupta, A.K. Goroch, P. Rabindra, N. Rangaraj, and M.S. Navar, 
1992: Polar cloud and surface classification using AVHRR imagery: an 
intercomparison of methods. J. Appl. Meteor., 31,405-420. 

Wilheit, T.T., A.T.C. Chang, J.L. King, E.B. Rodgers, R. A. Nieman, B.M. Kjupp, A. 
S. Milm^, J. S. Stratigos, H. Siddalingaiah, 1982: Microwave radiometric 
observations near 19.35,92 and 183 GHz of precipitation in tropical storm Cora. 
J.Appl. Meteor., 21, 1137-1145. 

World Meteorological Organization (WMO), 1989: Report of the Tenth Session of the 
Joint Scientific Committee. IRD-No. 314, World Meteorological Organization, 57 
pp., Geneva Willefranche-sur-mer, France. 

Wylie, Donald P., and W. P. Menzel, 1989: Two Years of Cloud Cover Statistics Using 
VAS. J. Clim., 2, 380 - 392. 

Wylie, D.P., W.P. Menzel, H.M. Woolf, and K.I. Strabala, 1994: Four years of global 
cirrus cloud statistics using HIRS. J. Clim., 1, 1927-1986. 

Zivkovic M. and J.-F. Louis, 1992: A new method for developing cloud specification 

schemes in general circulation models. Mon. Weather Rev., 120, pp 2928-2941. 


Appendix A 

AER SERCAA-Related Papers and Technical Presentations for 1993 - 1996: 

AMS 8*** Conference on Satellite Meteorology and Oceanography - 28 

January - 2 February 1996, Atlanta, GA 

Evaluation of SERCAA integration algorithm for analysis of multiplatform/multisensor 
satellite-derived cloud parameters. Gary Gustafson, Christopher Grassotti, Robert 
dEntremont (AER). 

Cloud detection and land-surface albedos using visible and near-infrared bi-directional 
reflectance distribution models. Robert d'Entremont (AER), Crystal Schaaf (PL), 
and Alan Strahler (BU). 

Retrieval of cirrus radiative and spatial properties using coincident multispectral imager 
and sounder satellite data. Robert d'Entremont (AER), Donald Wylie (UW), Szu- 
Cheng Ou, Kuo-Nan Liou (UU). 

Retrieval of cloud spatial, microphysical, radiative and environment parameters from 
multisource using SERCAA. Ronald Isaacs, Gary Gustafson, Robert 
d'Entremont (AER). 

Cloud Impacts on DoD Operations and Systems (CIDOS 95) - 24-26 

October 1995, Hanscom AFB, MA 

Cloud data sets derived from combined geostationary and polar-orbiting environmental 
satellite sensors using the SERCAA cloud model. Gary Gustafson, Robert 
d'Entremont, Daniel Peduzzi (AER) 

Enhanced satellite cloud analysis by the development of a higher resolution (6-km) 
global geography data set. Lt. Rad Robb (PL), Daniel Peduzzi (AER), Joan Ward 

Cloud detection using visible and near-infrared bidirectional reflectance distribution 
models. Robert d'Entremont (AER), Crystal Schaaf (PL), Alan Strahler (BU) 

SERCAA Phase 11: cloud radiative, microphysical, and environmental properties. 
Ronald Isaacs, Gary Gustafson, Robert d'Entremont, David Hogan (AER), Maj. 
Michael Remeika, James Bunting (PL) 

Retrieval of cirrus radiative and spatial properties using coincident multispectral imager 
and sounder satellite data. Robert d'Entremont (AER), Donald Wylie (UW), Szu- 
Cheng Ou, Kuo-Nan Liou (UU). 

Support of enviromnental requirements for cloud analysis eind archive (SERCAA) 
integrated cloud analysis validation. Kermeth Heideman (PL), Robert dEntremont, 
Jeanne Sparrow, Anthony Lisa, Gary Gustafson (AER). 


European Symposium on Satellite Remote Sensing II, Conference on Passive 
Infrared Remote Sensing of Clouds and the Atmosphere III, 25-29 September 
1995, Paris, France. 

Multi-spectral multi-platform satellite cloud and cloud environment retrievals for 
SERCAA. Ronald G. Isaacs, Gary B. Gustafson, and Robert P. d'Entremont 

Strategic Environmental Research and Development Program Symposium - 
12-14 April 1994, Washington, D.C. 

Multi-spectral multi-platform satellite cloud and cloud environment retrievals from 
SERCAA. Ronald G. Isaacs, Gary B. Gustafson, and Robert P. d'Entremont 
(AER), James T. Bunting and Maj. Michael F. Remeika (PL) 

European Symposium on Satellite Remote Sensing - 26-30 September 
1994, Rome, Italy 

Integration of multiplatform/multisensor satellite global cloud analyses within 
SERCAA. C. Grassotti, R. Isaacs, G. Gustafson (AER). 

Improved cloud analysis for CDFS n through the SERCAA research and development 
program. R.G. Isaacs, G.B. Gustafson, R.P. d'Entremont (AER), T.J. Neu 
(AFGWC), J.W. Snow (PL). 

AMS Tth Conference on Satellite Meteorology and Oceanography, 6-10 
June 1994, Monterey, CA 

Cloud Cover Determination Using the DMSP OLS. Gary B. Gustafson, Daniel C. 
Peduzzi, and Jean-Luc Moncet (AER). 

Validation of the SERCAA Cloud Analysis Algorithm. Kenneth F. Heideman (PL) and 
Jeanne M. Sparrow (AER). 

Analysis of Geostationary Satellite Imagery Using a Temporal Differencing Approach. 
Robert P. d'Entremont, Gary B. Gustafson, and Brian T. Pearson (AER). 

Analysis Integration Within SERCAA: Optimizing the Analysis of 

Multiplatform/Multisensor Satellite-Derived Cloud Parameters. C. Grassotti, T. 
Hamrll, R. Isaacs, G. Gustafson, D. Johnson (AER). 

Improved Cloud Analysis for CDFS E Through the SERCAA Research and 

Development Program. Thomas J. Neu (AFGWC), Ronald G. Isaacs and Gary B. 
Gustafson (AER), and J. William Snow (PL). 


Cloud Impacts on DoD Operations and Systems (CIDOS 93) - 16-19 
November 1993, Ft. Belvoir, VA 

Satellite Cloud Analysis Programs at the Air Force Phillips Laboratory: An Overview: 
Part 1 Tactical Nephanalysis (TACNEPH). Gary B. Gustafson, Ronald G. Isaacs 
(AER), Robert P. d'Entremont and James T. Bunting (PL). 

Satellite Cloud Analysis Programs at the Air Force Phillips Laboratory: An Overview: 
Part 2 Support of Environmental Requirements for Cloud Analysis and Archives 
(SERCAA). R. G. Isaacs and G. B. Gustafson (AER) and J. W. Snow and R. 
P. d'Entremont (PL). 

Validation of TACNEPH Cloud Detection Algorithms. Jeanne M. Sparrow, Gary B. 
Gustafson, Anthony S. Lisa and Robert P. d'Entremont (AER). 

Remote Sensing of Cloud Thickness and Base from Multispectral Cloud Imager Data. 
Ronald G. Isaacs, Alberto Bianco, Gary Gustafson, and Charles Sarkisian (AER). 

A Short-Term Cloud Forecast Scheme Using Cross Correlations. Thomas M. Hamill 
and Thomas Nehrkom (AER). 

Robust Database Management for Virtual-Application Environments. James S. 
Belfiore (AER). 

SPIE Conference on Passive Infrared Remote Sensing of Clouds and the 
Atmosphere - 13-15 April 1993, Orlando, FL 

Remote sensing of cloud for defense and climate studies: an overview (invited paper). 
R.G. Isaacs (AER). 

Validation of infrared cloud detection algorithms developed for TACNEPH. G.B. 
Gustafson, R.G. Isaacs, J.M Sparrow (AER), J.T. Bunting, and R.P. 
d'Entremont (PL).