Applying linear spectral unmixing to airborne hyperspectral imagery for mapping yield variability in grain sorghum and cotton fields

Abstract

This study examined linear spectral unmixing techniques for mapping the variation in crop yield for precision agriculture. Both unconstrained and constrained linear spectral unmixing models were applied to airborne hyperspectral imagery collected from a grain sorghum field and a cotton field. A pair of crop plant and soil spectra derived from each image was used as endmember spectra to generate unconstrained and constrained plant and soil cover abundance fractions. For comparison, the simulated broad-band normalized difference vegetation index (NDVI) and narrow-band NDVI-type indices involving all possible two-band combinations of the 102 bands in the hyperspectral imagery were calculated and related to yield. Statistical results showed that plant abundance fractions provided better correlations with yield than the broad-band NDVI and the majority of the narrow-band NDVIs, indicating that plant abundance maps derived from hyperspectral imagery can be used as relative yield maps to characterize yield variability in grain sorghum field and cotton fields without the need to choose the best NDVI. Moreover, the unconstrained plant abundance provided essentially the same results for yield estimation as the constrained plant abundance either with the abundance sum-to-one constraint only or with both the sum-to-one and non-negativity constraints, indicating that the more computationally complex constrained linear unmixing does not offer any advantage over the simple unconstrained linear unmixing for this application.