A Hybrid Convolutional Neural Network with Anisotropic Diffusion for Hyperspectral Image Classification

Recent research has shown that methods based on deep convolutional neural networks (DCNN) can achieve high accuracy in the classification of hyperspectral image (HSI). However, convolution operations with different dimensions in deep neural networks usually perform in an isotropic structure, resulting in the loss of extracting deep feature of anisotropic neighborhood. Thus how to improve the ability of discriminative features learning is the key issue in DCNN. In this paper, we propose an anisotropic diffusion partial differential equation (PDE) driven hybrid CNN framework, named PM-HCNN. The proposed framework uses 2D convolution and 3D convolution layers to extract of spectral and spatial contexts in HSI. And a PDE based diffusion layer is cascaded as feature propagation layers after hybrid convolution layers to propagate the intrinsical discriminative features of various classes. Due to the anisotropic diffusion on the feature space, the classification mistakes of traditional CNNs with a small number of training data can be further eliminated while object boundaries can be preserved. Experimental results on several popular datasets show that the proposed PM-HCNN achieved state-of-the-art performance compared with the existing deep learning-based methods.