Incorporating Attention Mechanism And Graph Regularization Into Cnns For Hyperspectral Image Classification

Yang Chu, Minchao Ye, Yuntao Qian

Abstract: The rich spatial and spectral features contained in a hyper-spectral image (HSI) play an important role in the classification of land-cover classes. In this paper, we incorporate attention mechanism and graph regularization into CNN-based HSI classification model to improve the classification performance. The proposed spatial-spectral CNN with attention mechanism and graph regularization (ATGR-SS-CNN) has three advantages: 1) 3D-2D cascade CNN. The 3D-CNN facilitates the joint spatial-spectral features representation from 3D data cube and the 2D-CNN further learns more abstract spatial representation; 2) Convolutional Block Attention Module (CBAM). CBAM gives greater attentions to discriminative features; 3) graph regularization. Graph regularization aims to maintain the similarity between samples belonging to the same land-cover class. Experiments on three datasets show that the proposed method can effectively improve the classification accuracy of HSI, which is better than the compared models.

External IDs:dblp:conf/whispers/ChuYQ22