Hyperspectral image segmentation using FSMLR with Jeffreys prior

The segmentation of satellite images is a valuable tool to obtain useful information about the spatial distribution of different land cover types. The use of segmentation algorithms instead of the traditional pixel-by-pixel classifiers used to produce land cover maps results on images that exhibit a more homogeneous distribution of classes, showing the piecewise spatial continuity of the real world. Several segmentation and classification methods are being developed to properly handle the high dimensionality of hyperspectral images. An example is a Bayesian segmentation procedure based on discriminative classifiers with a Multi-Level Logistic Markov-Gibbs prior. This method adopts the Fast Sparse Multinomial Logistic Regression as discriminative classifier, a method that promotes sparsity by including a Laplacian prior. However, the use of this type of prior requires an extensive search to for the best parameter of sparsity. In this work, a modification to this method is introduced. Instead of using the Laplacian Prior to enforce the sparsity of FSMLR classifier, the Jeffreys prior is used. This prior avoids the need to proceed to an extensive search for the best parameter, and also keeps the sparsity of the densities estimators, resulting on a faster and competitive segmentation procedure. The results of the application of this new approach to the benchmarked dataset Indian Pines show the effectiveness of the proposed method when compared with that using the Laplacian prior.