Dual-Branch Attention-Assisted CNN for Hyperspectral Image Classification

Abstract

Convolutional neural network (CNN)-based hyperspectral image (HSI) classification models have developed rapidly in recent years due to their superiority. However, recent deep learning methods based on CNN tend to be deep networks with multiple parameters, which inevitably resulted in information redundancy and increased computational cost. We propose a dual-branch attention-assisted CNN (DBAA-CNN) for HSI classification to address these problems. The network consists of spatial-spectral and spectral attention branches. The spatial-spectral branch integrates multi-scale spatial information with cross-channel attention by extracting spatial–spectral information jointly utilizing a 3-D CNN and a pyramid squeeze-and-excitation attention (PSA) module. The spectral branch maps the original features to the spectral interaction space for feature representation and learning by adding an attention module. Finally, the spectral and spatial features are combined and input into the linear layer to generate the sample label. We conducted tests with three common hyperspectral datasets to test the efficacy of the framework. Our method outperformed state-of-the-art HSI classification algorithms based on classification accuracy and processing time.