A Depthwise Separable Fully Convolutional ResNet With ConvCRF for Semisupervised Hyperspectral Image Classification

Abstract

Hyperspectral images classification relies on the accurate and efficient extraction of discriminative features, detail preservation, and efficient learning with limited training samples. This article, therefore, presents an advanced neural network architecture combined with convolutional conditional random fields (ConvCRF) and region growing (RGW) approaches to address these key issues. First, a depthwise separable fully convolutional residual network (DFRes) is proposed for efficient feature learning, where a fully convolutional operation ensures a larger field of view, and residual learning and depthwise separable convolution can mitigate the problem of vanishing gradient and overfitting. Second, because the collection of ground-truth labels is usually difficult, the proposed architecture integrates the RGW method to effectively overcome the problem of limited training samples. Third, ConvCRF is used to preserve the image details for fine-grained predictions. Finally, the abovementioned key components are coherently integrated into the new semisupervised framework, i.e., DFRes with conditional random fields and RGW. Experimental results on three hyperspectral datasets demonstrate that the proposed approach outperforms other state-of-the-art methods.