Hyperspectral image classification with dual attention dense residual network

Abstract

ABSTRACT Recent years, convolutional neural networks (CNNs) have attracted broad attention in hyperspectral image (HSI) classification. Most CNN-based HSI classification methods use image patches as the inputs of network. As the size increases, input patches may contain inhomogeneous pixels, which will cause damage to the classification results. In general, it needs to develop hierarchical architectures for CNNs to learn high-level feature representation. However, excessively deep architectures will cause overfitting and gradient vanishing, thus degrades the generalization performance of CNN. To address the above issues, a dual attention dense residual network (DADRN) is proposed for HSI classification. First, a dual attention module (DAM) is designed to reduce the damage of inhomogeneous pixels. It consists of a spectral attention unit and a spatial attention unit, and can achieve adaptive feature refinement. Second, a dense residual subnetwork (DRN) composed of dense convolutional blocks (DCBs) is proposed to extract more discriminative features from the output feature maps of the DAM. Specifically, the DCB achieves feature reuse mechanism through dense connection and channel concatenation operation, and the DRN adopts dense residual connections to alleviate overfitting and gradient vanishing. Experimental results on three benchmark HSI datasets demonstrate that the competitive advantage of proposed method over several state-of-the-art classification methods.