1. BOSTON UNIVERSITY GRADUATE SCHOOL OF ART AND SCIENCES LAI AND FPAR ESTIMATION AND LAND COVER IDENTIFICATION WITH MULTIANGLE MULTISPECTRAL SATELLITE DATA by YU ZHANG Submitted in partial fulfillment of the Requirements for the degree of Doctor of Philosophy (Total of 31 visuals) DISSERTATION

2. KEYWORDS Multiangle Remote Sensing Land Cover LAI and FPAR 2/31

3. Multiangle Remote Sensing Multiangle remote sensing is simultaneous measurement along different look angles of reflected radiation from a target. Examples: ATSR-2 (2 observation angles, 1km resolution) POLDER (up to 14 observation angles, 6km resolution) MISR (9 observation angles, 1.1km resolution) 3/31

4. Land Cover (1) What is land cover? Land cover is simply a description of the kind of vegetation at a location at a given time. Shrubs Grasses Broad Leaf Crops Forests

5. Land Cover (2) Why is land cover important? Land cover and land use changes inferred from vegetation maps are a direct evidence of the human and climate impact on the land. Most climate and biogeochemical models, as well as algorithms that estimate surface biophysical variables from remote sensing data, utilize vegetation maps to assign certain key parameters to reduce the number of unknowns. 5/31

6. LAI and FPAR (1) What? LAI – Green Leaf Area Index = one-sided green leaf area per unit ground surface area FPAR – Fraction of incident Photosynthetically Active Radiation Absorbed by the vegetation canopy = APAR / IPAR 6/31

7. LAI and FPAR (2) Why? LAI is a key state variable in all land parameterization of climate, ecology, and hydrology models. FPAR is a key variable in terrestrial carbon models. 7/31

8. Objectives The objective of my research is to demonstrate the utility of multiangle multispectral remote sensing for estimation of LAI, FPAR and land cover. Specifically, Prototype the MISR LAI/FPAR algorithm (Part I) Empirical and theoretical analysis f multiangle, multispectral data (Part II) Land cover classification with multiangle multispectral data (Part III) 8/31

9. PART I:Prototyping MSIR LAI/FPAR Algorithm POLDER Data: –~6km resolution –Africa –Nov. 1996 –Up to14 multiangle data per pixel Biome Classification Map

10. BCM-Africa Biome Classification Map derived from AVHRR data (8km)

11. The Algorithm Metrics of multiangle observations and uncertainties Algorithm LAI & FPAR Solution Distribution Functions: mean variance LUT based inverse solution of the 3D transport equation 11/31

12. Saturation Frequency Saturation Frequency decreases using multiangle data 12/31

13. Dispersion of LAI Dispersion of LAI for single-angle retrievals and multiangle retrievals for broadleaf crops. 13/31

14. LAI 14/31

15. Part I: Conclusions The MISR LAI/FPAR algorithm performs satisfactorily Retrieval accuracy increases in the case of multiangle inputs Note: This work is published in Zhang et al, Prototyping of MISR LAI and FPAR algorithm with POLDER data over Africa. IEEE Trans. Geosci. Remote Sens. 38:2402- 2418, 2000. 15/31

16. Part II:Investigations of Multiangle Data Empirical Analysis BRDF? Angular signatures in spectral space? Theoretical Analysis (will not be presented here)

17. BRDF backscattering forward scattering B.S. F.S. B.S. F.S. 17/31

18. Angular Signature in Spectral Space MultiangleSingle-angle Location LengthNo OrientationNo InterceptNo

19. Interpretation of the Angular Signature Indices 1)Location — Biome type 2)Intercept Indices — Vegetation ground cover 3)Length Indices — Canopy structure 4)Slope Indices — LAI 19/31

20. IGBP-AS Angular signatures in the red-NIR (near-infrared) spectral space of the ten land covers from Hansen et al. (2000) 1 km land cover map of North America. 20/31

21. Part II: Conclusions We developed metrics that characterize the BRDF for use in land cover classification These metrics have a basis in transport theory Note: These works is described in a two-part series: Zhang et al., Required consistency between definitions and signatures with the physics of remote sensing I: empirical arguments. And II: theoretical arguments. Remote Sens. Environ. (Submitted in January 2001). 21/31

22. Part III: Land Cover Classification with Angular Signature Indices Data North America land cover training sites POLDER Data (June 1997, North America) Methods MANOVA, PCA, Correlation Matrix Classification Techniques Decision tree classification Maximum likelihood classification

23. Classification Variables Spectral Location (2) Red, NIR Angular Length, Slope, Intercept (3) 3 measurement patterns (3  3=9) Total 9+2=11 variables 23/31

24. Statistic1 24/31

25. Variance of PCA Data information content is larger than spectral variables only 25/31

26. 2-Classify The maximum classification accuracies as functions of the number of variables used in the decision tree and maximum likelihood classification methods.

27. Part III: Conclusions The statistical analyses confirm the idea that incorporating angular signature variables will improve biome classification. The maximum likelihood classification result indicates a improvement of classification accuracy using directional variables. Note: These works is prepared for publication: Zhang and Woodcock, Improve the land cover classification accuracy with multiangle remote sensing data. (In preparation, 2001). 27/31

28. CONCLUDING REMARKS My research demonstrates: Satisfactory performance of the MISR LAI/FPAR algorithm Multiangle data improve accuracy of LAI/FPAR retrievals It is possible to define simple metrics that characterize the BRDF – a complicated 4D function Multiangle data contain information useful for land cover classification

29. FUTURE DIRECTIONS Comprehensive analysis of MISR data to further develop these ideas ( It is not my job! :) Introducing temporal domain in land cover classification activity. 29/31

30. ACKNOWLEDGEMENTS Committee Fellow Graduate Student Data provider: Leroy, Diner, McIver 30/31

31. Thank you all! Questions Please… 31/31