Skip to Main Content
This paper demonstrates a new method called progressive discrimination (PD) for mapping an individual spectral class within an image. Given training data for a target, PD iteratively samples nontarget image pixels using a collapsing distance threshold within the space of an evolving discriminant function. This has the effect of progressively isolating the target class from similar spectra in the image. PD was compared to Bayesian maximum likelihood classification, mixture-tuned matched filtering, spectral angle mapping, and support vector machine methods for mapping three different invasive species in two types of high-spatial-resolution airborne hyperspectral imagery, AVIRIS and AISA. When tested with 20 different randomly selected groups of training fields, PD classification accuracies for the two spectrally distinct plant species in these images had an average of 98% and a standard deviation of 1%. These randomized trials were capable of providing higher classification accuracies than the best results obtained by two expert analysts using existing methods. For the third species that was less distinct, PD results were comparable to the results obtained by experienced analysts with existing methods. Despite requiring less input from the user than many techniques, PD provided more consistent high mapping accuracy, making it an ideal tool for scientists and land use managers who are not trained in image processing.