1) global coefficients: derived from all available cloud-free aster scenes for a given network...

1
All Scenes 2.75 2.78 3.66 3.41 2.74 34683 Daytim e Only 2.87 2.63 3.21 3.17 2.63 15858 Nighttim e Only 2.48 2.69 2.34 2.97 2.63 18825 Spring 2.81 2.86 3.60 3.40 2.82 8619 Sum m er 2.71 2.98 3.77 3.90 3.06 8719 Fall 2.74 2.43 3.73 3.11 2.41 11400 Winter 2.73 2.53 3.37 2.57 2.47 5945 Group to W hich Regression w as Applied Unscaled CoefficientType Num berof in situ and M ODIS LST Com parisons Global Day/Night- Specific Season- Specific Site- Specific All Scenes 2.75 2.66 2.68 2.67 2.69 34855 Daytim e Only 2.87 2.97 2.96 3.00 3.05 15958 Nighttim e Only 2.48 2.34 2.41 2.37 2.29 18897 Spring 2.81 2.74 2.79 2.84 2.74 8696 Sum m er 2.71 2.67 2.79 2.73 2.76 8723 Fall 2.74 2.57 2.57 2.57 2.52 11418 Winter 2.73 2.48 2.50 2.50 2.72 6018 Num berof in situ and M ODIS LST Com parisons Unscaled CoefficientType Group to W hich Regression w as Applied Global Day/Night- Specific Season- Specific Site- Specific 1) Global Coefficients: Derived from all available cloud-free ASTER scenes for a given network SURFRAD: 246 scenes USCRN: 371 scenes 2) Day/Night-Specific Coefficients: ASTER scenes grouped by network and daytime or nighttime overpass Upscaling of in situ Land Surface Temperature for Satellite Validation Robert Hale (CIRA/Colorado St. Univ.), Yunyue Yu (NOAA/NESDIS STAR), and Dan Tarpley (Short & Assoc.) Conclusions and Future Activities At USCRN sites, regression-based upscaling of in situ LSTs can reduce scale-induced errors, thereby rendering in situ LSTs more appropriate for use in validating coarse-resolution satellite LSTs While statistically significant reduction of error is observed in many USCRN cases, the absolute reduction is typically fairly small (~0.2 K), and SURFRAD sites generally realize little benefit from upscaling Scale-induced error reduction is highly variable between models and coefficient groups, as well as between individual sites (not shown) Better performance from more generalized coefficients versus site-specific coefficients 2 km x 2 km average LST avg = 318.59 K Central pixel LST pixel = 318.80 K LST in situ = 312.76 K Validation of satellite-derived Land Surface Temperature (LST) poses challenges due both to the paucity of in situ measures against which the satellite-derived LSTs may be compared and because of the mismatch in spatial scale between the two. In an effort to address these issues, multiple linear regression models were derived using high-resolution LSTs from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) to characterize the relationship between “point” measurements at ground validation sites and the average LST over a larger area representing a coarse-resolution satellite pixel. The derived models were then used to upscale in situ LSTs from Surface Radiation (SURFRAD) and U.S. Climate Reference Network (USCRN) sites. Unscaled and scaled LSTs were subsequently compared with LSTs from the Moderate Resolution Imaging Spectroradiometer (MODIS). + USCRN (Newton 11SW, GA) Field of View of IRT: 1.3 m http://www.esrl.noaa.gov/gmd/grad/surfrad/bonpics/tower01c.jpg http://www1.ncdc.noaa.gov/pub/data/uscrn/documentation/site/photos/stationsbystate_hires.pdf SURFRAD (Bondville, IL) FOV of Radiometer: ~30 m 1 km 1 km ASTER LST (Stillwater 5WNW, OK; 7/23/2005) Resolution: 90 m MODIS LST (Stillwater 5WNW, OK; 7/23/2005) Resolution: 1000 m Upscaling Model Coefficients Group SURFRAD USCRN Daytime Scenes 186 193 Nighttime Scenes 60 178 3) Season-Specific Coefficients: ASTER scenes grouped by network and climatological season 4) Site-Specific Coefficients: ASTER scenes grouped by individual site SURFRAD: 13-70 scenes, depending on site USCRN: 4-68 scenes, depending on site Group SURFRAD USCRN Spring (MAM) 53 96 Summer (JJA) 72 96 Fall (SON) 80 115 Winter (DJF) 41 64 One-Predictor Regression Model Single ASTER pixel encompassing the ground station used as predictor of large-area average LST Once the and coefficients are determined using ASTER scenes, the formula is applied to in situ LSTs to determine scaled values Scaled values then are compared with MODIS coarse- resolution LSTs to determine efficacy of model – reduced standard deviation of differences for scaled vs. unscaled LSTs used as indicator of model performance Standard deviation of differences of MODIS – unscaled or scaled in situ LSTs (K) USCRN Sites (22 sites) All Scenes 2.37 2.36 2.37 2.35 2.48 12714 Daytim e Only 2.59 2.59 2.59 2.59 2.53 5596 Nighttim e Only 2.18 2.16 2.18 2.15 2.44 7118 Spring 2.53 2.53 2.55 2.53 2.65 3041 Sum m er 2.39 2.39 2.40 2.39 2.46 3931 Fall 2.36 2.36 2.33 2.32 2.47 4164 Winter 1.94 1.92 1.92 1.87 2.18 1578 Group to W hich Regression w as Applied CoefficientType Unscaled Num berof in situ and M ODIS LST Com parisons Global Day/Night- Specific Season- Specific Site- Specific Significantly different from unscaled at: = 0.05 = 0.01 No significant difference for any coefficient type SURFRAD Sites (7 sites) Two-Predictor Regression Model 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 NDVI D ay ofYear Asheville 13S (NC) 2003-2010 Mean NDVI Multi-year average of MODIS NDVI used as additional predictor to better capture seasonal changes in vegetation amount that strongly influence LST Two-Predictor Regression Model (Continued) All Scenes 2.75 2.67 2.52 2.73 2.91 34683 Daytim e Only 2.87 2.93 2.78 2.86 2.98 15858 Nighttim e Only 2.48 2.40 2.29 2.62 2.84 18825 Spring 2.81 2.76 2.62 2.94 2.90 8619 Sum m er 2.71 2.66 2.59 2.80 3.07 8719 Fall 2.74 2.58 2.38 2.62 2.83 11400 Winter 2.73 2.50 2.51 2.46 2.79 5945 Group to W hich Regression w as Applied Unscaled CoefficientType Num berof in situ and M ODIS LST Com parisons Global Day/Night- Specific Season- Specific Site- Specific Significantly different from unscaled at: = 0.05 = 0.01 USCRN Sites (22 sites) Results at SURFRAD sites similar to those for 1- predictor model – No significant reduction in scale- induced error with any of the coefficient types Air Temperature-Based Regression Model Newton 8W, GA http://www1.ncdc.noaa.gov/pub/data/uscrn/documentation/site/photos/stationsbystate_hires.pdf Air temperature used as a proxy for canopy temperature ASTER single-pixel or in situ LST used as indicator of soil surface temperature NDVI used to weight the two temperatures USCRN Sites (22 sites) Significantly different from unscaled at: = 0.05 = 0.01

Upload: bendek

Post on 25-Feb-2016

26 views

Category:

Documents


0 download

DESCRIPTION

Significantly different from unscaled at:. Significantly different from unscaled at:.  = 0.05 .  = 0.05 .  = 0.01 .  = 0.01 . http://www.esrl.noaa.gov/gmd/grad/surfrad/bonpics/tower01c.jpg. Significantly different from unscaled at:.  = 0.05 .  = 0.01 . 1 km. 1 km. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: 1) Global Coefficients: Derived from all available cloud-free ASTER scenes for a given network SURFRAD: 246 scenesUSCRN: 371 scenes

All Scenes 2.75 2.78 3.66 3.41 2.74 34683Daytime Only 2.87 2.63 3.21 3.17 2.63 15858Nighttime Only 2.48 2.69 2.34 2.97 2.63 18825Spring 2.81 2.86 3.60 3.40 2.82 8619Summer 2.71 2.98 3.77 3.90 3.06 8719Fall 2.74 2.43 3.73 3.11 2.41 11400Winter 2.73 2.53 3.37 2.57 2.47 5945

Group to Which Regression was Applied

UnscaledCoefficient Type Number of in

situ and MODIS LST Comparisons

GlobalDay/Night-

SpecificSeason-Specific

Site-Specific

All Scenes 2.75 2.66 2.68 2.67 2.69 34855Daytime Only 2.87 2.97 2.96 3.00 3.05 15958Nighttime Only 2.48 2.34 2.41 2.37 2.29 18897Spring 2.81 2.74 2.79 2.84 2.74 8696Summer 2.71 2.67 2.79 2.73 2.76 8723Fall 2.74 2.57 2.57 2.57 2.52 11418Winter 2.73 2.48 2.50 2.50 2.72 6018

Number of in situ and MODIS LST Comparisons

UnscaledCoefficient TypeGroup to Which

Regression was Applied

GlobalDay/Night-

SpecificSeason-Specific

Site-Specific

1) Global Coefficients: Derived from all available cloud-free ASTER scenes for a given networkSURFRAD: 246 scenes USCRN: 371 scenes

2) Day/Night-Specific Coefficients: ASTER scenes grouped by network and daytime or nighttime overpass

Upscaling of in situ Land Surface Temperature for Satellite ValidationRobert Hale (CIRA/Colorado St. Univ.), Yunyue Yu (NOAA/NESDIS STAR), and Dan Tarpley (Short & Assoc.)

Conclusions and Future Activities At USCRN sites, regression-based upscaling of in situ LSTs can

reduce scale-induced errors, thereby rendering in situ LSTs more appropriate for use in validating coarse-resolution satellite LSTs

While statistically significant reduction of error is observed in many USCRN cases, the absolute reduction is typically fairly small (~0.2 K), and SURFRAD sites generally realize little benefit from upscaling

Scale-induced error reduction is highly variable between models and coefficient groups, as well as between individual sites (not shown)

Better performance from more generalized coefficients versus site-specific coefficients suggests lack of ASTER scenes for coefficient determination may be limiting model performance

To address the above, Landsat data are being acquired and utilized for improved model development

2 km x 2 km average LSTavg = 318.59 K

Central pixel LSTpixel = 318.80 K

LSTin situ = 312.76 K

Validation of satellite-derived Land Surface Temperature (LST) poses challenges due both to the paucity of in situ measures against which the satellite-derived LSTs may be compared and because of the mismatch in spatial scale between the two. In an effort to address these issues, multiple linear regression models were derived using high-resolution LSTs from the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) to characterize the relationship between “point” measurements at ground validation sites and the average LST over a larger area representing a coarse-resolution satellite pixel. The derived models were then used to upscale in situ LSTs from Surface Radiation (SURFRAD) and U.S. Climate Reference Network (USCRN) sites. Unscaled and scaled LSTs were subsequently compared with LSTs from the Moderate Resolution Imaging Spectroradiometer (MODIS).

+

USCRN (Newton 11SW, GA)Field of View of IRT: 1.3 m

http://www.esrl.noaa.gov/gmd/grad/surfrad/bonpics/tower01c.jpghttp://www1.ncdc.noaa.gov/pub/data/uscrn/documentation/site/photos/stationsbystate_hires.pdf

SURFRAD (Bondville, IL)FOV of Radiometer: ~30 m

1 km 1 km

ASTER LST(Stillwater 5WNW, OK; 7/23/2005)Resolution: 90 m

MODIS LST(Stillwater 5WNW, OK; 7/23/2005)Resolution: 1000 m

Upscaling Model Coefficients

Group SURFRAD USCRNDaytime Scenes 186 193Nighttime Scenes 60 178

3) Season-Specific Coefficients: ASTER scenes grouped by network and climatological season

4) Site-Specific Coefficients: ASTER scenes grouped by individual siteSURFRAD: 13-70 scenes, depending on siteUSCRN: 4-68 scenes, depending on site

Group SURFRAD USCRNSpring (MAM) 53 96Summer (JJA) 72 96Fall (SON) 80 115Winter (DJF) 41 64

One-Predictor Regression ModelSingle ASTER pixel encompassing the ground station used as predictor of large-area average LST

 Once the and coefficients are determined using ASTER scenes, the formula is applied to in situ LSTs to determine scaled values

 Scaled values then are compared with MODIS coarse-resolution LSTs to determine efficacy of model – reduced standard deviation of differences for scaled vs. unscaled LSTs used as indicator of model performance

Standard deviation of differences of MODIS – unscaled or scaled in situ LSTs (K)

USCRN Sites (22 sites)

All Scenes 2.37 2.36 2.37 2.35 2.48 12714Daytime Only 2.59 2.59 2.59 2.59 2.53 5596Nighttime Only 2.18 2.16 2.18 2.15 2.44 7118Spring 2.53 2.53 2.55 2.53 2.65 3041Summer 2.39 2.39 2.40 2.39 2.46 3931Fall 2.36 2.36 2.33 2.32 2.47 4164Winter 1.94 1.92 1.92 1.87 2.18 1578

Group to Which Regression was Applied

Coefficient TypeUnscaled

Number of in situ and MODIS LST Comparisons

GlobalDay/Night-

SpecificSeason-Specific

Site-Specific

Significantly different from unscaled at: = 0.05 = 0.01

No significant difference for any coefficient type

SURFRAD Sites (7 sites)

Two-Predictor Regression Model

0.00.10.20.30.40.50.60.70.8

ND

VI

Day of Year

Asheville 13S (NC) 2003-2010 Mean NDVI

Multi-year average of MODIS NDVI used as additional predictor to better capture seasonal changes in vegetation amount that strongly influence LST

 

Two-Predictor Regression Model (Continued)

All Scenes 2.75 2.67 2.52 2.73 2.91 34683Daytime Only 2.87 2.93 2.78 2.86 2.98 15858Nighttime Only 2.48 2.40 2.29 2.62 2.84 18825Spring 2.81 2.76 2.62 2.94 2.90 8619Summer 2.71 2.66 2.59 2.80 3.07 8719Fall 2.74 2.58 2.38 2.62 2.83 11400Winter 2.73 2.50 2.51 2.46 2.79 5945

Group to Which Regression was Applied

UnscaledCoefficient Type Number of in

situ and MODIS LST Comparisons

GlobalDay/Night-

SpecificSeason-Specific

Site-Specific

Significantly different from unscaled at: = 0.05 = 0.01

USCRN Sites (22 sites)

Results at SURFRAD sites similar to those for 1-predictor model – No significant reduction in scale-induced error with any of the coefficient types

Air Temperature-Based Regression ModelNewton 8W, GA

http://www1.ncdc.noaa.gov/pub/data/uscrn/documentation/site/photos/stationsbystate_hires.pdf

Air temperature used as a proxy for canopy temperature

ASTER single-pixel or in situ LST used as indicator of soil surface temperature

NDVI used to weight the two temperatures

  USCRN Sites (22 sites)

Significantly different from unscaled at: = 0.05 = 0.01