1. -IKONOS, ETM, MODIS NDVI: comparison -Jeff Morisette, MODLAND, SSAI -Positive Systems for Appalachian Transect -Rob Sohlberb, MODLAND, UMd -Report from Stennis Space Center on SDP validation activities -Mary Pagnutti, SSC -One Ikonos DEM/Stereo Pair for Barton Bendish site -J. Peter Muller, MODLAND, ULC NASA’s: Science Data Purchase

2. One approach to scaling Comparing ETM+, IKONOS, and MODIS NDVI products Framed in the context of statistical hypothesis testing J. Morisette

3. General validation procedure: correlative analysis (slide from 1999 validation mtg.) Field data “Tasked” acquisitions: Airborne and high res. Satellite Automatic acquisitions: reference data and products to be validated Compare: Need to consider all three elements as samples from unknown distributions, use each component to estimate the respective distribution, and compare distributions points to pixels parameters and distributions relationships surfaces PointFine resolution

4. Spectral Bands, Red and NIR Difference may be important: Gitelson and Kaufman, 1998; Bo-Cai Gao, 2000; which compared MODIS to AVHRR and found large differences in NDVI *ASD spectrum from grass area near GSFC AVHRR IKONOS ETM+ MODIS grass reflectance*

5. Study Area: Konza Prairie Data: MODIS daily products: Sept. 11 500m surface reflectance 500m pointer file 1km viewing geometry LDOPE tools to combine (available through EDC EDG) ETM+, Sept. 11 IKONOS, Sept. 15 Aeronet (Meyer) (available through Konza Prairie Core Site web page) Vermote et al.’s Six S code (for ETM+ and IKONOS)

6. IKONOS area on MODIS 500m and ETM+

7. Subset area in Sinusoidal Projection

8. Sampling with MODIS “Tile Mapper” Done for both IKONOS and ETM+

9. Sampled imagery IKONOS (30m)ETM+

10. Comparison at Multiple-scales MODIS Pixel 1, 1 ETM+ 14, 16 n=224 IKONOS 116, 120 n=13,920 Considering all three as variable and subject to errors, consider MODIS pixel relative to the distribution from the higher resolution data

11. IKONOS vs ETM+ Correlation =.5639 Reject hypothesis of zero correlation Using standard Pearson method (p value ~0)

12. IKONOS vs MODIS Correlation =.3114 Reject hypothesis of zero correlation Using standard Pearson method (p value ~0)

13. ETM+ vs MODIS Correlation =.3401 Reject hypothesis of zero correlation Using standard Pearson method (p value ~0)

14. Do the data follow a normal distribution? Null Hypothesis: Normally distributed Test: Kolmogorov-Smirnov Goodness-of-Fit Test: MODIS data: Reject (p =.0079) ETM+ at 500m: Reject (p =.0004) IKONOS at 500m: Reject (p ~ 0) ETM+: Reject (p ~ 0) IKONOS at 30m: Reject (p ~ 0) So, should consider testing correlation with non-parametric methods.

15. Non-parametric correlation Null Hypothesis: Zero Correlation Test: Spearman's rank correlation IKONOS vs ETM+:Reject (rho =.5791, p ~ 0) ( corr =.5639) IKONOS vs MODIS: Reject (rho =.3099, p ~ 0) ( corr =.3114) ETM+ vs MODIS: Reject (rho =.3362, p ~ 0) ( corr =.3401) But we still might want to question the hypothesis being tested.

16. Test for Paired Differences Null Hypothesis: average paired difference is zero Test: T test (assume normality and homogeneity of variance) Test: Wilcoxon Rank Sum Tests IKONOS vs ETM+ IKONOS vs MODIS Reject all three pair-wise combination ETM+ vs MODIS based on either test. So, for these data we are somewhere in the middle: There is positive correlation, but the average difference is not zero

17. Normalized differences to include variability in validation data MODIS – IKONOS(average) Std. Dev (IKONOS ave.) = “z score”

18. Z score analysis IKONOS vs ETM

19. Z score analysis IKONOS vs MODIS

20. Z score analysis ETM vs MODIS

21. Do the z-scores follow a normal distribution? Null Hypothesis: Normally distributed Test: Kolmogorov-Smirnov Goodness-of-Fit Test: Z from IKONOS vs ETM+:Reject (ks = 0.1955, p ~ 0) Z from IKONOS vs MODIS: Reject (ks = 0.0537, p = 0.0142) Z from ETM+ vs MODIS: Reject (ks = 0.0538, p = 0.013) So, should consider testing z-scoresw with at least both parameteric and non-parametric methods

22. Test of z-score “centered” on zero Non-Parametric Null Hypothesis: Median value is zero Test: Wilcoxon Signed Rank Sum Tests Z from IKONOS vs ETM+:Reject (Z =-46.493, p ~ 0) Z from IKONOS vs MODIS: Reject (Z = -9.9305, p ~ 0) Z from ETM+ vs MODIS: Reject (Z = -6.2677, p ~ 0) Parametric Null Hypothesis: Mean value is zero Test: T test Z from IKONOS vs MODIS: Reject (t = -10.143, p ~ 0) Z from ETM+ vs MODIS: Reject (t = -5.4727, p ~ 0)

23. Conclusions Assumption of normality is not always met Non-parametric methods are available Z-score method shows one possible way to scale up; which incorporates variability and considers the validation data with respect to its distribution There is a fundamental difference between the null hypothesis of the correlation being zero and the difference being zero There is closer statistical agreement between MODIS and either IKONOS and ETM+ than between IKONOS and ETM+ There is a difference between statistical and practical difference

24. Comments ETM+, IKONOS, MODIS and Sun photometer data were easily available Major difficulty was ISIN projection and georeferencing – coordination of Jacqueline Le Moigne, GSFC might prove helpful. Results are planned to be communicated in the validation article in the Special Issue of RSE.