关键词: 高光谱图像, 监督分类, 深度学习, 稠密网络, 空-谱卷积

Abstract: Due to the massive parameters and complex structure, deep learning networks are usually trained in a long time with large-scale training samples. In this paper, we propose a spatial-spectral convolutional dense network (SSCDenseNet) which mainly targets limited samples for hyperspectral image classification. Three novel strategies are proposed to construct the proposed network. First, a spatial-spectral separable convolution method is adopted to make up a hidden layer unit with a spectral one-dimensional convolutional layer and a spatial two-dimensional convolutional layer; then the deep network is constructed by stacking multiple units. Second, we use batch normalization before each hidden layer unit to reduce covariance drift of data and accelerate the network training procedure. Finally, a direct connection between every two units is adopted to reuse hierarchical features, and solve the problem of gradient vanishing. The comprehensive evaluation of experiments on different datasets such as Indian Pines, Pavia University and Salinas are conducted to show the performance of the SSCDenseNet, and the results show that the proposed method outperforms several state-of-the-art deep learning based methods in terms of classification performance, especially under small-scale samples.

Key words: hyperspectral image, supervised classification, deep learning, dense network, spectral-spatial convolution