Hyperspectral Image Classification Based On A Convolutional Neural Network And Discontinuity Preserving Relaxation

In this paper, we present a novel method for hyperspectral image classification to take advantage of the merits of a convolutional neural network (CNN) and the spatial contextual information of hyperspectral imagery (HSI). We built a novel network consisting of several convolutional, pooling and activation layers to extract the effective features and predict the class membership probability distribution vectors for HSI pixels. Furthermore, in order to fully exploit the spatial contextual information and improve the classification accuracy under the condition of limited training samples, a promising discontinuity preserving relaxation (DPR) algorithm is applied to process the probabilistic results obtained by the CNN work. The proposed method was tested on two widely-used hyperspectral data sets: the Indian Pines and University of Pavia data sets. Experiments revealed that the proposed method can provide competitive results compared to some state-of-the-art methods.