Deep stacking network with coarse features for hyperspectral image classification

Hyperspectral image (HSI) classification attracts increasing attentions in remote sensing community for its academic significance and potential wide applications. Although most of used existing “intelligent” methods extracting the features of original hyperspectral data are based on shallow-layer networks such as neural network (NN) and support vector machine (SVM) for its simplicity in realization, two deep neural networks (DNN), i.e. the deep convolutional network (DCN) and the deep belief network (DBN), have been used for HSI classifications with improved performance, as deep learning recently achieves great success for its ability in deep feature extraction and representations. In this paper, another DNN, a deep stacking network (DSN) based approach with coarse features for hyperspectral image classification is proposed, as its advantages over shadow networks and other deep models in its simplicity and batch-mode learning — not requiring stochastic gradient descent that other DNNs require. In this approach, the inputs to the DSN are coarse features, i.e. coarse spectral features by band reduction, coarse spatial features extracted by PCA from original HSI, or combination of both. The fine deeper features in the hyperspectral data are gradually obtained by employing the nonlinear activation function on the hidden layer nodes of each module, which is different from the current DSN that usually uses linear weights between hidden layer to output layer. A closed limit on its node's input — output is also exerted. The theoretical analysis and experimental results with AVIRIS hyperspectral image have shown that, (1) the proposed DSN approach achieves improved classification performance compared with shadow networks, (2) the new DSN approach with both the coarse spectral feature and coarse spatial feature has higher classification accuracy compared to that with either only, (3) this study has insight of necessity to develop new deep neural network for HSI classification from direct HSI data.