Linear spectral unmixing assisted by probability guided and minimum residual exhaustive search for subpixel classification

摘要

For subpixel analysis of remotely sensed imagery, the requirement for a sufficient number of image bands is often a constraint to a linear mixing model, because the number of end members for land cover classification is usually larger than the number of bands when Thematic Mapper (TM) or SPOT imagery is utilized. This paper proposes two algorithmic approaches to resolve the constraint that only consider a subset of end members in each pixel's unmixing process: a minimum residual exhaustive search (MRES) algorithm and a probability-guided (PG) process. The MRES method tests each possible subset and measures its residual. Once all possible scenarios have been tested, the subset which yields the minimum residual is found and its unmixing result is accepted as this pixel's constituents. The PG method calculates a pixel's posterior probability to each end member first and takes the end members which have the highest probabilities to form a subset. Then, a linear spectral unmixing procedure is applied to unmix the pixel into the subset. Case studies have shown that the PG method outperforms the MRES method.