Evaluation of compositing algorithms for AVHRR data over land

The objective of this study was to evaluate alternative methods for producing both Normalized Difference Vegetation Index (NDVI) and single-channel composite images of land surfaces from the Advanced Very High Resolution Radiometer (AVHRR). As a goal, it was specified that the composite image should approximate as much as possible a single-date image with a constant, near-nadir geometry. The comparative performance of three single-step [maximum NDVI (MaN), maximum apparent temperature (MaT), maximum difference of channels 2 minus 1 (MaD)] and two two-step [maximum NDVI followed by maximum temperature (MaNMaT) or by minimum scan angle (MaNMiSc)] criteria in creating composite images was evaluated for three land cover types, namely, cropland, coniferous forest, and deciduous forest within a boreal ecosystem. The assessment was carried out using 18 images obtained with various acquisition geometries in mid-summer over a 1000/spl times/1000 km area in Manitoba, Canada. In addition, MaT performance was compared with that of MaN for all Canada during two compositing periods. It was found that among the five criteria tested, MaT and MaNMiSc were the most effective one- and two-step criteria, respectively. MaN preferentially selected off-nadir pixels from the forescatter region, the degree varying with land cover type; the overall NDVI values were higher than for MaT. MaD showed a very strong preference for high backscatter region pixels, regardless of land cover type. Depending on cover type, the NDVI values resulting from Mao were higher or lower than those from a nadir image. Based on ranking of the five techniques using statistics of the differences between composite and reference images, it was found that MaT and MaNMiSc performed similarly, and better than the remaining criteria. Results of the tests show that although statistically reasonable approximations of the reference image could be produced by one or more methods, none of the criteria could consistently yield composites closely resembling the nadir image on a pixel basis, even for a reasonably long compositing period. Therefore, pixel-specific applications of the composites relying on individual channels will likely have to be based on data corrected for bidirectional effects. The results suggest that such corrections are also required for NDVI and, by analogy, for other AVHRR channel combinations.< >