Hyperspectral Mixture Modeling for Quantifying Sparse Vegetation Cover in Arid Environments

A linear mixture model based on calibrated, atmospherically corrected Probe-1 hyperspectral imagery was compared with three vegetation indices to test its relative ability to measure small differences in percent green vegetative cover for areas of sparse vegetation in arid environments. The goal of this research was to compare multispectral and hyperspectral remote sensing approaches for detecting human disturbance of arid environments. The normalized difference vegetation index (NDVI) was tested using both narrow and broad bandwidths. Broadband NDVI provided results r2=0.63 similar to NDVI derived from individual hyperspectral channels r2=0.60. While the soil-adjusted vegetation index (SAVI) was designed as an improvement to NDVI for sparse vegetation, in this study SAVI performed significantly worse than NDVI r2=0.51. The modified soil-adjusted vegetation index (MSAVI) provided an insignificant improvement over NDVI r2=0.64. Linear mixture modeling provided significantly better results, r2 of 0.74. Cross-validation was used to test the significance of differences between the various methods and to determine the standard error associated with each method. Results suggest that any improvements provided by adjusted vegetation indices over NDVI may be strongly dependent on those adjustments being derived from local conditions. The use of a linear mixture model with multiple soil endmembers appears to provide the best method for quantifying sparse vegetative cover. Though present in small amounts, a single plant species, Krameria erecta, was strongly correlated with residuals of the mixture model. Inclusion of a spectral endmember for this species increased the r2 of the fit with percent green cover to 0.86. However, it is not clear if the explained variation was actually due to K. erecta or a correlated phenomena. Problems were also identified with the use of multiple vegetation endmembers.