Classification of hyperspectral images with convolutional neural networks and probabilistic relaxation.

In this paper, an integrated framework for the classification of hyperspectral images is presented. Firstly, two convolutional neural networks (CNNs) were developed for the extraction of representative features. In particular, a pixel-wise CNN and a patch-based CNN were designed to extract spectral features and spectral–spatial features, respectively. The two neural networks consist of several convolutional, pooling and activation layers, and are able to predict the class membership probabilities of test pixels. Secondly, two probabilistic relaxation methods, namely Markov random fields and discontinuity preserving relaxation were integrated into the framework in order to refine the probabilistic results from a Bayesian perspective. This framework enhances the classification performance by exploiting the contextual information available from neighboring pixels. This is particularly advantageous when only limited training samples are available. The proposed framework was tested on both simulated and real-world data sets. The experimental results suggest that the proposed methods outperform several state-of-the-art methods.