1. Validating Satellite Land Surface Temperature Products for GOES-R and JPSS Missions
Yunyue Yu, Mitchell Goldberg, Ivan Csiszar
NOAA/NESDIS
Center for Satellite Applications and Research
Presentation for IGARSS 2011

2. Motivation -- LST validation needs
LST Products Derived from Different Sensors for Decades
POES
NOAA AVHRR (TIROS-N to NOAA-19) since 1978
ESA ATSR/AATSR since 1991
EOS MODIS since 1999
MetOp AVHRR since 2006
GOES
US GOES/Imagers since 1975
Meteosat (MVIRI since 1995 and then) SEVIRI since 2004
Future LST products
JPSS/NPP
GOES-R
LST Climate Data Record
Just listed major LST sources, not include microwave sensors, LEO sensors.
GOES generation:
SMS – GOES 1 (1975) to 3
1st - GOES 4 to 7
2nd – GOES 8 to 12
3rd – GOES 13 to 15
4th – GOES R, S, T, U

3. Motivation -- LST validation issues
LST Validation Difficulties
In situ data limitation
Measurement difficulty
Cloud contamination effect
particularly the partial or thin cloudy pixels
Spatial and temporal variations
Spot vs pixel difference
Sub-pixel heterogeneity
Accurate match-up process (different sampling rates and sampling timing)
Others
i.e. angle effect
In situ data limitation
Different spatial scales
Different sampling rates and sample timing
Sub-pixel cloud contamination
Sub-pixel heterogeneity
Limited samples of clear case
Surface heterogeneity is shown in a 4km x 4km Google map (1km x 1km, in the center box) around the Bondville station area

4. Approach -- Strategy Using Existing Ground Observation Data
Cost consideration
Site representativeness and selection: characterization analysis
Match-up Dataset Generation
Stringent cloud filtering: additional measures
Data pair quality control
Site-to-pixel Model Development
Synthetic pixel analysis using high resolution sensor data
Proxy data testing
Real satellite data evaluation
Validation Methodology
Direct comparisons
Indirection comparisons

5. Approach
Synthetic pixel analysis using ASTER data— an integrated approach for
site representativeness analysis and site-to-pixel model development
Quantitatively characterize the sub-pixel heterogeneity and decide whether a ground site is adequately representative for the satellite pixel. The sub-pixels may be generated from pixels of a higher-resolution satellite.
For pixel that is relatively homogeneous, analyze statistical relationship of the ground-site sub-pixel with the surrounding sub-pixels: {T(x,y) } ~ T(x0,y0)
Establish relationship between the objective pixel and its sub-pixels (i.e., up-scaling model), e.g., Tpixel = T(x,y) + DT (time dependent?)
T(x,y,t)
T(x0,y0,t0)
ASTER pixel
The site pixel
MODIS pixel
Issues:
1) different t and t0 resulting different Ts.
2) the ground homogeneity maybe location/time dependent
3) the up-scaling model
4) the uncertainty of up-scaling estimation
Solutions:
1) develop or use proper random statistic model for heterogeneity and interpolation analysis
2) develop or use proper up-scaling model to perform point-to-area estimation
The Synthetic pixel/sub-pixel model

6. Approach
A Site Characterization Simulation Model – synthesizing VIIRS pixel using higher-resolution ASTER TIR pixels.
ASTER scene (90m) pixel
Each synthetic pixel has the target ground site enclosed, but the distance between the ground site and the center of synthetic pixel varies, which mimics the possible over-passing VIIRS swaths.
Distance of every synthetic pixel center from the ground site is within the pixel size (~1Km).
Different colors are used for the 9 synthetic pixels, and the center of each pixel is marked with a small numbered square of the same corresponding color.
The numbers on the squares are the pixel IDs used in the relevant analysis.
Colored squares:
Ground site synthetic VIIRS pixels

7. Approach
Quantification of the difference between the Synthetic Pixel and Ground Measurement, that is,
Evaluation: and with
the model:
Note that:
Why Tc – Tg is systematic error?
Sub-pixel heterogeneity
Systematic bias between ASTER and ground measurements
Tsat – satellite pixel LST
Tg -- ground site LST
Tc central sub-pixel LST

8. Data Sets Data Acquisition
Inventory preparation of clear-sky ASTER scenes passing over 23 ground sites (SURFRAD and CRN) during
Collection of the clear-sky ASTER data sets associated with six SURFRAD sites and two CRN sites.
AST_04, AST_05, AST_08 and L1B (in total, about 2000 ASTER swath scenes).
Collection of the six SURFRAD and the two CRN ground data.
Collection of the two CRN ground data.
Collection of MODIS LST product data (MOD11_L2).
All the swaths passing over the SURFRAD sites in 2001
All the swaths corresponding to the ASTER scenes during
Collection of the narrow-band emissivity data sets
UW-Madison Baseline Fit Emissivity Database
North American ASTER Land Surface Emissivity Database (NAALSED)
Collection of NCEP reanalysis TPW datasets

9. Clear Cases (additional cloud filter)
Data Sets
Candidate Ground Sites and Database
Dataset Used
SURFRAD data
ASTER Data
Data period:
ASTER data is courtesy by Shunlin Liang
Satellite LST: MODIS LST, ASTER, and GOES LST
Ground LST: Derived from SURFRAD site measurements
Table: Matched ASTER Data
Stations
Clear Cases
(ASTER cloud mask)
Clear Cases (additional cloud filter)
Desert Rock 63 46
Bondville, IL
115 51
PennState, PA 61 20
Boulder, CO 35 13
Fort Peck, MT 12 8
Should use my one (CRN sites are not used in the table, therefore not need to be shown)

10. Data Sets Ground Site Broadband Emissivity
Regression based on the UW-Madison Baseline Fit Emissivity Database
( Seemann et al., 2008).
Ground Site Broadband Emissivity
Regression of Broadband emissivity from well-developed narrowband emissivity database:
UW-Madison baseline Fit Emissivity Database
a=0.2122, b=0.3859,c= (Wang, 2004)
Jul. 2011

11. Processing General Components of Validation Processing Ground Data
Satellite Data
Match-up Datasets
Time Match-up
Geolocation Match-up
Ground Data Mask
Satellite Cloud Mask
Manual Cloud Control
Ground LST Estimation/Extraction
Satellite LST Calculation/Extraction
Synthetic Analysis and Correction
Ground Data
Ground Data Reader
Satellite Data Reader
Indirect Comparison
Direct Comparison
Statistical Analysis
Outputs
(Plots, Tables, etc.)

12. Upwelling, Downwelling
Processing
Sample Match-up Flow Chart
Geolocation
Match-up
SURFRAD
Data
Satellite
Time
(< 5 mins)
Time Series
Smoothness Check
(if available):
Upwelling, Downwelling
Irradiances
Spatial
Difference Test:
BT -- 3X3 pix STDs,
Visual deg
Channel BT
(Ts, T10mm), (T10mm, T3.9mm)
(T10mm, T12mm)
Matched
Dataset
Manual
Tuning
Cloud Mask
Additional cloud filter
Note: this flow chart is specifically for GOES Imager
Similar procedure is/will be applied for the ASTER and MODIS/VIIRS data

13. Processing Data Processing/Analysis Clear-sky cases analysis
Cloud and clear-sky climatology analysis (for site selection)
ASTER Clear-sky swath selection from the ASTER inventory from the Warehouse Inventory Search Tool
Ground broadband emissivity regression analysis
SURFRAD LST estimation from PIR measurements
Spatial and temporal match-up among ground sites, ASTER scenes and MODIS scenes
Geolocation mapping of ASTER pixels as the sub-pixels of a MODIS pixel
Quality-control and enhanced cloud filtering
Processing of ASTER LST QC information
Processing of ASTER emissivity QC information
Processing MODIS cloud masks
Processing of MODIS LST QC information
Surface observations
Statistical testing

14. Processing Data Processing/Analysis (con’t)
Site representativeness analysis and site-to-pixel difference characterization
Semi-variance analysis
Synthetic analysis
Site-to-pixel model testing
Testing with all the MODIS Terra LST swaths passing over SUFRAD sites in 2001
Testing with the MODIS LST swaths corresponding with ASTER scenes during
VIIRS LST case studies on NPP land LPEATE platform
Development of VIIRS LST algorithm modules for flexible offline testing and algorithm improvement
Visualization tools of the analysis and results disolay

15. Results
Additional cloud filtering is need for obtaining high quality satellite-ground match-up dataset
Left: ATSER cloud free dataset. Right: possible cloud contamination.
Cloud

16. Results
Comparison of the temperatures calculated from synthetic pixel average (top-right), center-pixel (bottom-left), and nearest pixel (bottom-right) with the ground site temperature. Note the different colors represent for the 9 different synthetic pixels shown previously.
For this particular site the ground site location within the satellite pixel does not have significance impact to the validation process, simply because the land surface thermal emission at Desert Rock is fairly homogeneous.
SURFRAD Station: Desert Rock

17. Results Sample statistical analysis result on the Desert Rock site.
0.61
0.08
2.20
-1.74 2
-1.81
Mean
Ts - Tc
0.76
-0.07
2.21
-1.88
Average
0.60
-0.16
2.37
-1.97 8
0.65
-0.26
2.40
-2.07 7
0.80
-0.34
2.30
-2.15 6
0.98
-0.24
2.18
-2.05 5
0.96
0.06
2.03
-1.75 4
0.92
0.20
1.99
-1.61 3
-0.01
2.26
-1.82 1
2.46
0.69
0.04
2.13
-1,78
STD
STD`
Tc-Ta
Ts – Ta
Case
Site=Desert Rock, NV
Impact of pixel location bias to the ground site
The Ta and Ts difference is tested by comparing its spatial structure to the site geographic structure.
It shows that such Ta and Ts difference matches the site geographic feature well, which implies that the synthetic pixel temperature calculation is reasonable.
Ts: LST of SURFRAD site
Ta: average LST over 13x13 ASTER pixels
Tc: LST of ASTER pixel nearest to the site

18. Results Sample statistical analysis result on the Bondville site.
1.07
-0.05
2.05
-0.64 2
-0.59
Mean
Ts - Tc
1.05
-0.09
2.08
-0.80
Average
0.97
-0.18
2.02
-0.77 8
0.95
-0.03 7
-0.001
2.12
-0.62 6
1.10
2.14
-0.60 5
1.15
2.10
-0.68 4
1.27
2.17 3
1.04
-0.14
2.01
-0.73 1
0.92
-0.07
2.04
-0.66
STD
STD`
Tc-Ta
Ts – Ta
Case
Site=Bondville, IL
Impact of pixel location bias to the ground site
Jul. 2011

19. Results Sample statistical analysis result on the Boulder site.
Site=Boulder, CO
0.91
-0.27
2.30
-1.03 2
-0.77
Mean
Ts - Tc
0.72
-0.13
2.58
-0.90
Average
0.70
-0.25
2.70
-1.02 8
-0.10
2.80
-0.87 7
0.69
-0.00
2.75 6
0.61
2.64
-0.67 5
-0.03
2.54
-0.80 4
0.84
-0.14
2.27
-0.91 3
0.85
-0.38
2.61
-1.15 1
2.60
0.58
-0.07
2.62
-0.84
STD
STD`
Tc-Ta
Ts – Ta
Case
Impact of pixel location bias to the ground site
Jul. 2011

20. Results
Sample scatter plots show the linear relationship between satellite LST and ground LST.
Tsat = A Tg + B + e

21. Results 11/11/2018 Site-to-Pixel Statistical Relationship Site C σ
Desert Rock, NV
1.88
2.21
Boulder, CO
0.90
2.58
FortPeck, MT
-0.15
2.62
Bondville,IL
0.68
2.08
PennState, PA
-0.17
1.99
WolfPoint 34, MT
1.47
1.98
WolfPoint 29, MT
0.64
3.86
Oakley, KS
0.24
1.03
Mercury, NV
3.99
1.25
Site A B σ
Desert Rock, NV
1.115
-33.66
1.73
Boulder, CO
1.138
-38.76
1.32
FortPeck, MT
1.089
-25.51
2.02
Bondville, IL
1.014
-3.29
2.01
PennState, PA
1.002
-0.66
1.97
WolfPoint 34, MT
0.768
65.99
WolfPoint 29, MT
0.593
114.48
3.23
Oakley, KS
1.374
1.33
Mercury, NV
1.044
-9.77
1.14
11/11/2018

22. Results
Site-to-Pixel Statistical Relationship – MODIS Proxy
Site
MODIS Pixel- SURFRAD
Synthetic Pixel- SURFRAD
Nearest Aster Pixel – Synthetic pixel
Match-Up Pairs (SURFRAD, ASTER, MODIS)
Mean
StdDev
 Desert Rock, NV
-0.44
1.84
2.09
1.69
0.04
0.44 34
 Boulder,CO
-0.49
2.08
1.49
2.90
-0.09
0.67 7
 Fort Peck, MT
0.35
2.52
0.78
1.95
0.38
1.15 4
 Bondville, IL
-0.17
1.51
1.04
1.48
0.03
0.90 26
 Penn State, PA
-1.53
1.91
0.61
1.96
1.07
The standard deviation of the difference between real MODIS pixel LST and SURFRAD LST are consistent with the standard deviation of the difference between synthetic pixel and SURFRAD.
The synthetic pixel LST is generally much warmer (about 1.5K) than SURFRAD LST, while MODIS LST is slightly cooler than SURFRAD LST.
Implication: High resolution ASTER data be used for site representativeness analysis, but a further verification of ASTER LST quality seems essential for the parameter estimation of the site-to-pixel model (e.g., A,B,C) and the application of the model in real cases.

23. Results –Summary
Synthetic pixel analysis model is created for analyzing ground site temperature heterogeneous feature. ASTER data are used to generate synthetic VIIRS pixel data (LST) and compared to the SURFRAD site data
LST of SURFRAD measurements may be used as good references for VIIRS/ABI LST cal/val if the measurements are of high-quality and the in-situ estimation of LST is accurate enough.
Directional variations are small, so small geo-referencing (within 1Km) bias may not be an issue affecting VIIRS/ABI LST cal/val at the above SURFRAD sites.
Application of the site-to-pixel model depends on the ASTER LST data quality.
Directional variation of the potential sub-pixel heterogeneity is found to be consistent with the physical topographic features, even if it is small.
The limited datasets doesn’t allow us to characterize the seasonal variation of heterogeneities, which is more desirable than a simple mean difference. More datasets are expected. And about 1K difference seems unavoidable in practice.

24. Future Plan
Analysis over time scales of interest, e.g., seasonal variation.
Analysis over sites of different surface types
Up/down-scaling models
Emissivity uncertainties
Test of the scaling model using real satellite data