Hyperspectral Remote Sensing Image Classification Based on Convolutional Neural Network

Abstract

Remote sensing hyperspectral imaging can obtain rich spectral information of terrestrial objects, which allows the indistinguishable matter in the traditional wideband remote sensing to be distinguished in hyperspectral remote sensing. Hyperspectral image has the characteristics of “combining image with spectrum”. Making full use of spectral information and spatial information in hyperspectral image is the premise of obtaining accurate classification results. At present, most of hyperspectral data feature extraction algorithms mainly utilize local spatial information in the same channel and spectral information in the same spatial location of different channels. However, these methods require a large amount of prior knowledge, it is difficult to fully grasp the hyperspectral data of all spatial and spectral information, and the model generalization ability is poor. With the development of deep learning, convolutional neural network shows superior performance in all kinds of visual tasks, especially in the two-dimensional image classification, and could get a high classification accuracy. In this paper, an image classification method based on three-dimensional convolution neural network is proposed based on the structural properties of hyperspectral data. In the proposed method, first the stereo image blocks of hyperspectral data are intercepted, then multi-layer convolution and pooling operation of extracted blocks by convolutional neural network are implemented to obtain the essential information of hyperspectral data, finally the classification of hyperspectral data is completed. The experimental results show the proposed method could provide better feature expression and classification accuracy for hyperspectral image.