UNMIXING THE DIRECTIONAL REFLECTANCES OF AVHRR SUB-PIXEL LANDCOVERS

Recent progress in canopy bidirectional reflectance distribution funct ion (BRDF) model inversions has allowed accurate estimates of vegetati on biophysical characteristics from remotely sensed multi-angle optica l data, Since most current BRDF inversion methods utilize one-dimensio nal (1-D) models, surface homogeneity within an image pixel is implied , The Advanced Very High Resolution Radiometer (AVHRR) is one of the f ew spaceborne sensors capable of acquiring radiometric data over the r ange of view angles required for BRDF inversions. However, its relativ ely coarse spatial resolution often results in measurements of mixed l andcovers, and thus the data may not be ideal for BRDF inversions, We present a three-step spectral unmixing method for retrieving AVHRR sub -pixel directional reflectances in regions of high spatial heterogenei ty. The reflectances of individual vegetation types are deconvolved us ing co-located Landsat TM and AVHRR data. The three major steps in the model include: 1) unmixing of vegetation endmember concentrations in TM imagery; 2) correction of dissimilar shadow fractions between TM an d AVHRR data; and 3) unmixing of AVHRR sub-pixel reflectances of veget ation types for any sun-sensor geometry, We tested the method using si mulated TM and AVHRR data. A savanna landscape simulation, comprised o f a canopy radiative transfer model and a crown geometric-optical mode l, was used to create images containing mixed pixels of tree, grass, a nd shade endmembers. TM and AVHRR spectral response functions, viewing geometries, and off-nadir pixel shape calculations were incorporated into the simulations, Following the successful testing of the unmixing method on error-free simulations, random noise representing atmospher ic perturbations and co-registration inaccuracies was added to the dat a. The method Is stable when errors resulting from either the first un mixing step or image co-registration inaccuracies are introduced, Pote ntial errors in the AVHRR data may result in inaccurately retrieved re flectances if the image scene contains a spatially homogeneous mix of landcovers. A method for detecting and mitigating this problem is pres ented.