Accuracy assessments of hyperspectral waveband performance for vegetation analysis applications

Abstract

The main objectives of this research were to: (a) determine the best hyperspectral wavebands in the study of vegetation and agricultural crops over the spectral range of 400–2500 nm; and (b) assess the vegetation and agricultural crop classification accuracies achievable using the various combinations of the best hyperspectral narrow wavebands. The hyperspectral data were gathered for shrubs, grasses, weeds, and agricultural crop species from the four ecoregions of African savannas using a 1-nm-wide hand-held spectroradiometer but was aggregated to 10-nm-wide bandwidths to match the first spaceborne hyperspectral sensor, Hyperion. After accounting for atmospheric widows and/or areas of significant noise, a total of 168 narrowbands in 400–2500 nm was used in the analysis. Rigorous data mining techniques consisting of principal component analysis (PCA), lambda–lambda R 2 models (LL R 2M), stepwise discriminant analysis (SDA), and derivative greenness vegetation indices (DGVI) established 22 optimal bands (in 400–2500 nm spectral range) that best characterize and classify vegetation and agricultural crops. Overall accuracies of over 90% were attained when the 13–22 best narrowbands were used in classifying vegetation and agricultural crop species. Beyond 22 bands, accuracies only increase marginally up to 30 bands. Accuracies become asymptotic or near zero beyond 30 bands, rendering 138 of the 168 narrowbands redundant in extracting vegetation and agricultural crop information. Relative to Landsat Enhanced Thematic Mapper plus (ETM +) broadbands, the best hyperspectral narrowbands provided an increased accuracy of 9–43% when classifying shrubs, weeds, grasses, and agricultural crop species.