Mineral Detection from Hyperspectral Images Using a Spatial-Spectral Residual Convolution Neural Network

Mineral detection from hyperspectral images is an interesting yet challenging research topic in the remote sensing community. Although many efforts have been put into this field, mineral detection is still far from solved, because compositions of minerals are very complex. The features of an interested mineral are very hard to be distinguished from others. In this paper, we attack this problem by introducing recently developed deep learning techniques. Beyond the spectral feature, considering that minerals coexist with each other, which could be regarded as a piece of special context information and thus can be captured by a spatial convolutional neural network (CNN), we design a CNN architecture that is able to exploit both spatial and spectral features of minerals. We test the proposed network on Indiana Pines dataset and a large-scale hyper-spectral image, the results demonstrate that the proposed method is good at representing both spatial and spectral information of hyper-spectral images and it can successfully detect minerals from hyperspectral images.