Hyperspectral Image Classification Via Sample Expansion for Convolutional Neural Network

摘要

Convolutional neural network (CNN) has already attracted great attention in hyperspectral classification. However, for hyperspectral imagery, it is often with few training samples in small-size classes. In this paper, a novel hyperspectral classification framework combining CNN and the orthogonal complement subspace projection (OCSP) balance solution is proposed for addressing this problem. Moreover, the existing synthetic minority over-sampling technique (SMOTE) is also applied onto our framework for further improving the classification accuracy. Experimental results demonstrate that the proposed framework can improve hyperspectral classification performance of CNN.