Hyperspectral Image Classification With Spatial Consistence Using Fully Convolutional Spatial Propagation Network

In recent years, deep convolutional neural networks (CNNs) have demonstrated impressive ability to represent hyperspectral images (HSIs) and achieved encouraging results in HSI classification. However, the existing CNN-based models operate at patch level, which a pixel is separately classified into classes using of around it. This patch-level classification will lead large number repeated calculations, it hard identify appropriate size that beneficial accuracy. addition, conventional CNN convolutions with local receptive fields, cause failure contextual spatial information modeling. To overcome these aforementioned limitations, we propose novel end-to-end, pixel-to-pixel, fully propagation network (FCSPN) for Our FCSPN consists 3-D convolution (3D-FCN) (CSPN). Specifically, 3D-FCN first introduced reliable preliminary classification, dual separable residual (DSR) unit proposed effectively capture spectral simultaneously fewer parameters. Moreover, channel-wise attention mechanism adapted grasp most informative channels from redundant channel information. Finally, CSPN correlations HSIs via learning linear propagation, allows maintaining consistency further refining results. Experimental on three benchmark data sets demonstrate achieves state-of-the-art performance