Science Review Panel Meeting Biosphere 2 Tucson AZ
Science Review Panel Meeting Biosphere 2, Tucson, AZ - January 4 -5, 2011 Vegetation Phenology and Vegetation Index Products from Multiple Long Term Satellite Data Records Cross-Sensor Continuity Science Algorithm Tomoaki Miura and Javzan Tsend-Ayush Alex Dale (M. S. Graduate Student) Joshua Turner (Federal Work Study Student) University of Hawaii at Manoa NASA MEASURES #NNX 08 AT 05 A
Factors Affecting Cross-Sensor Continuity & Considered in Science Algorithm 1. 2. 3. 4. Spectral bandpass Spatial resolution (point spread function) Algorithm Sun-target-view geometry -2 - MEASURES VIP ESDRs Science Review Panel
Science Algorithm – Two Approaches Top-down, direct image comparison using overlapping period of observation (Javzan Tsend-Ayush & Alex Dale) I. � � Derivation of multi-sensor translation equations Development of an evaluation methodology Bottom-up, simulation analysis using hyperspectral imagery (Tomoaki Miura & Joshua Turner) II. � � Derivation of multi-sensor translation equations Characterization of factorial effects, error budgets -3 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison Methods for Deriving Multi-sensor Translation Equations 1. 2. 3. 4. Spatial aggregation to the CMG (. 05 o) grid Data screening on a per-pixel basis Extraction of a random sample Regression analysis to derive equations -4 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison 2. Data Screening AVHRR LTDR. ver 2 -1998, 1999 SPOT-4 VEGETATION Terra MODIS - 2001, 2002 - 1998, 1999, 2001, 2002 I. Screening - QA -Cloudy -Partly Cloudy -Shadow I. Screening - SM - Cloud - Shadow -5 - I. Screening - QA -Cloudy -Fire -Dust -Cloud shadow , … MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison 2. Data Screening � Water mask from 2001 expanded by one pixel (Chen, 1999) -6 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison 2. Data Screening % total observations 0 Percent difference less than -45 165 Number of clear observations Percent difference between -5 and 5 0 1 11 21 Percent difference larger than 45 31 41 51 61 -7 - 71 81 91 100 MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison 3. Sample Preparation VGT 4 -NDVI vs. AVH 14–NDVI -8 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison MODIS-NDVI vs VGT-NDVI ( 40 percent of all datasets ) -9 - MEASURES VIP ESDRs Science Review Panel
Footprint Simulation Satellite � Changes in footprint size and PSF modeled using a satellite orbital model (Tan et al. , 2006) � Daily basis over a 16 -day period in June 2002 Terra MODIS � 250 m/500 m NOAA-16 AVHRR � 1. 1 at nadir km at nadir SPOT-4 VEGETATION � 1. 1 km at nadir Center of the Earth -10 - MEASURES VIP ESDRs Science Review Panel
1: View Zenith Angle Effects AVHRR Off Nadir vs Nadir -11 - MEASURES VIP ESDRs Science Review Panel
2: Geolocation Error Effects AVHRR Off Nadir vs Nadir Geolocation Error (m) X=-480, Y=30 (Angle=55) X=540, Y=-900 (Angle=5) -12 - MEASURES VIP ESDRs Science Review Panel
2: View Zenith Angle & Geolocation Error Effects Sensor Comparison ~ NDVI DOY=175 MOD 500 = 26º(X=-2, Y=0), AVH 16 = 43º (X=0, Y=-8), VGT = 50º (X=1, Y=-6) DOY=181 MOD 500 = 37º(X=1, Y=-1), AVH 16 = 45º (X=-27, Y=-18), VGT = 36º (X=68, Y=-15) (Miura et al. , 2010, in prep. ) -13 - MEASURES VIP ESDRs Science Review Panel
3: Pixel Averaging NDVI EVI MOD 500 vs VGT vs AVH 16 MOD 500 vs AVH 16 -14 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison 4. Regression Analysis � Ordinary Least Square � Geometric Mean Functional Relationship � Least Median of Square �Agreement Analysis � Mean product-difference � RMSD of systematic and unsystematic � Agreement coefficient �Error Analysis � Classified percentage error and visual interpretation -15 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison 4. Regression Analysis -16 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison 4. Regression Analysis � Deriving translation equations - Least square regression technique (95% confidence). 80 % of the paired observations were used. � Single Equation: All samples with all land cover types Land stratification methods were examined for minimizing each VI differences across multi-sensors n Stratification Method-1 : based on the Landcover map( 15 of IGBP 17 classes) n Stratification Method-2: based on the Landcover map & latitudinal zoning n Stratification Method-3: based on the phenological region map (2956 classes ) -17 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison IGBP Land cover map (MODIS 2001) Water Evergreen Needleleaf forest Evergreen Broadleaf forest Deciduous Needleleaf forest Deciduous Broadleaf forest Mixed forest Closed shrublands Open shrublands Woody savannas Savannas Grasslands Permanent wetlands -18 - Croplands Urban and built-up Cropland/Natural vegetation mosaic Snow and ice Barren/sparsely vegetated MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison Result of Stratification method-2 30 o Water Evergreen Needleleaf forest Evergreen Broadleaf forest Deciduous Needleleaf forest Deciduous Broadleaf forest Mixed forest Closed shrublands Open shrublands Woody savannas Savannas Grasslands Permanent wetlands -19 - Croplands Urban and built-up Cropland/Natural vegetation mosaic Snow and ice Barren/sparsely vegetated MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison Result of Stratification method-3 �A phenological region map derived from MODIS data (2956 classes) -20 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison of Stratifications -21 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison MODIS-NDVI vs VGT-NDVI ( 40 percent of all datasets ) -22 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison MOD-NDVI vs VGT-NDVI ( the same direction of viewing and difference VZA is less than 10 degree) -23 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison MOD-EVI 2 vs VGT-EVI 2 LC-1 LC-2 LC-3 LC- 6 LC-7 LC-8 LC-11 LC-12 LC-4 LC-9 LC-13 LC-14 -24 - LC-5 LC-10 LC-15 MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison Translation Equations Derived from the Top-down Analyses -25 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison Translation Equations Derived from the Top-down Analyses (cont. ) -26 - MEASURES VIP ESDRs Science Review Panel
I. Top-down, Direct Image Comparison Translation Equations Derived from the Top-down Analyses (cont) -27 - MEASURES VIP ESDRs Science Review Panel
II. Bottom-up, Hyperspectral Simulation Analysis Locations of Study Sites -28 - MEASURES VIP ESDRs Science Review Panel
II. Bottom-up, Hyperspectral Simulation Analysis 1. Spectral Compatibility – NDVI -29 - MEASURES VIP ESDRs Science Review Panel
II. Bottom-up, Hyperspectral Simulation Analysis 1. Spectral Compatibility – EVI 2 -30 - MEASURES VIP ESDRs Science Review Panel
II. Bottom-up, Hyperspectral Simulation Analysis 2. Scaling Uncertainties between CMG and GAC Resolutions -31 - MEASURES VIP ESDRs Science Review Panel
II. Bottom-up, Hyperspectral Simulation Analysis 3. Atmospheric Correction - NDVI -32 - MEASURES VIP ESDRs Science Review Panel
Evaluation Methods � Derived VIs translation equations are applied to translate � � � AVHRR VIs to “SPOT 4 Vegetation-like” VIs SPOT 4 Vegetation VIs to “MODIS-like” VIs Validation: 20 % of the paired observations were used where: is an estimator of parameter -33 - MEASURES VIP ESDRs Science Review Panel
Agreement Analysis & Coefficient (Ji & Gallo, 2006) � The agreement coefficient (AC) considers that both x- and yvariables are subject to random errors. � The AC measures the RMPDS RMPDU systematic (RMPDS) and unsystematic (RMPDU) components of the root mean square difference (RMSD): -34 - MEASURES VIP ESDRs Science Review Panel
Global Coarse Resolution (0. 05 o) Daily Products: Terra MODIS vs. SPOT-4 VEGETATION NDVI Sensor (Y vs. X) RMPDS (RMPDU) GMFR R 2 . 025 (±. 045) Y =. 016 +1. 034 X . 95 <. 001 (±. 043) Y = -. 001 +1. 004 X . 95 RMPDS (RMPDU) GMFR R 2 . 027 (±. 032) Y =. 010 +1. 102 X . 91 . 001 (±. 032) Y = -. 002 +1. 010 X . 91 Original (5%) MOD vs. VGT 4 Translated (5%) MOD vs. ML-VGT 4 EVI 2 Sensor (Y vs. X) Original (5%) MOD vs. VGT 4 Translated (5%) MOD vs. ML-VGT 4 (Tsend-Ayush et al. , 2010, in prep. ) -35 - MEASURES VIP ESDRs Science Review Panel
Error Analysis Min 1 max 241 0 Cluster -45 100*(vgt-avh)/vgt Cluster 45 240 Cluster 0 0 -36 - 1 11 21 31 41 51 61 71 81 91 100 MEASURES VIP ESDRs Science Review Panel
-0. 05≤ Mod_NDVI – Mod. Like. VGT_NDVI < 0. 05 -37 - MEASURES VIP ESDRs Science Review Panel
-0. 05≤ Mod_EVI 2 – Mod. Like. VGT_EVI 2 < 0. 05 -38 - MEASURES VIP ESDRs Science Review Panel
Root Mean Square Difference of VGT-NDVI and MODIS-NDVI -39 - MEASURES VIP ESDRs Science Review Panel
Root Mean of Square Difference of VGT-EVI 2 and MODIS-EVI 2 -40 - MEASURES VIP ESDRs Science Review Panel
Current Issues & Future Work � LTDR AVHRR algorithm improvements � BRDF-adjustment � Aerosol correction � Terra vs. Aqua comparisons [to establish error bounds] � Expansion of geographic & temporal coverage -41 - MEASURES VIP ESDRs Science Review Panel
- Slides: 41