Hyperspectral Image Classification Based on Dual-Scale Dense Network with Efficient Channel Attentional Feature Fusion

Abstract

Hyperspectral images (HSIs) have abundant spectral and spatial information, which shows bright prospects in the application industry of urban–rural. Thus, HSI classification has drawn much attention from researchers. However, the spectral and spatial information-extracting method is one of the research difficulties in HSI classification tasks. To meet this tough challenge, we propose an efficient channel attentional feature fusion dense network (CA-FFDN). Our network has two structures. In the feature extraction structure, we utilized a novel bottleneck based on separable convolution (SC-bottleneck) and efficient channel attention (ECA) to simultaneously fuse spatial–spectral features from different depths, which can make full use of the dual-scale shallow and deep spatial–spectral features of the HSI and also significantly reduce the parameters. In the feature enhancement structure, we used 3D convolution and average pooling to further integrate spatial–spectral features. Many experiments on Indian Pines (IP), University of Pavia (UP), and Kennedy Space Center (KSC) datasets demonstrated that our CA-FFDN outperformed the other five state-of-the-art networks, even with small training samples. Meanwhile, our CA-FFDN achieved classification accuracies of 99.51%, 99.91%, and 99.89%, respectively, in the case where the ratio of the IP, UP, and KSC datasets was 2:1:7, 1:1:8, and 2:1:7. It provided the best classification performance with the highest accuracy, fastest convergence, and slightest training and validation loss fluctuations.