Spectral-Spatial Classification of Hyperspectral Image Based on Self-Adaptive Deep Residual 3D Convolutional Neural Network

In this paper, a deep residual 3 D convolutional neural network(3 D-CNN) framework is proposed for hyperspectral images classification in order to realize fast training, classification and parameter optimization. Rich spectral and spatial features can be rapidly extracted from samples of complete hyperspectral images using the Network, because the three-dimensional feature extraction algorithm is more suitable for three-dimensional data structure. In addition, the classification accuracy can be effectively improved by data augmentation which rotates and flips hyperspectral image samples from the spatial direction, and TPE algorithm which adjusts hyper-parameters in an artificial search space. Experimental results on the standard hyperspectral data sets, such as Pavia University, Indian Pines and KSC, show that the proposed framework not only deepens the network but also improves the classification accuracy compared with SSRN and other methods. The network optimized by TPE adaptive hyper-parameter optimization algorithm reduces the number of parameters by half and the training time by about 10%, compared with manual settings.