A post-processing technique for Lagrangian artificial neural network approach to hyperspectral image classification

Lagrangian Artificial Neural Network (LANN) has been proposed recently for hyperspectral image classification. It is an unsupervised technique that can simultaneously estimate the endmembers and their abundance fractions without any prior information. Since the implementation of the LANN is completely unsupervised, the number of estimated abundance fraction images (AFI) is equal to the number of bands, which display the distribution of the corresponding endmember materials in the image scene. We find out that many AFIs are highly correlated and visually similar. In order to facilitate the following data assessment, a two-stage post-processing approach will be proposed. First, the number of endmembers n_s resident in the image scene is estimated using a Neyman-Pearson hypothesis testing-based eigen-thresholding method. Next, an automatic searching algorithm will be applied to find the most distinct AFIs using the divergence as criterion, where the threshold is adjusted until the number of selected AFIs equals the n_s estimated in the first stage. The experimental results using AVIRIS data shows the efficiency of the proposed post-processing technique in distinct AFI selection.