Leveraging in-scene spectra for vegetation species discrimination with MESMA-MDA

We describe an approach to improve Multiple Endmember Spectral Mixture Analysis (MESMA) results for applications involving discrimination among spectrally-similar species, and commonly occur in multispectral and hyperspectral vegetation remote sensing studies. Such applications are inherently difficult, due to the high degree of similarity between distinct species, coupled with potentially high intra-species variability caused by factors such as growing conditions, canopy structure, ambient illumination, or substrate characteristics. We describe a method to map spectra to a feature space where distinctions between plant species are emphasized using a transformation based on Multiclass Discriminant Analysis. We compute this transformation using groups of pixels that represent individual plant canopies similar to the endmembers in MESMA's spectral library, and describe a technique to automatically select such spectra from a given image. Compared to conventional MESMA, and also to several alternative MESMA formulations, we observe up to twofold increases in accuracy, along with a factor of ten reduction in computation time using our MESMA approach in several species discrimination applications. We demonstrate the effectiveness of our approach for agricultural species discrimination applications using spectra captured by two different imaging spectrometers.