Hyperspectral Image Classification Based on Center-Guided Weighted Sparse Attention Mechanism Network

Yuanyuan Zhao¹Lan Li²Long Wang^2*Xiaobin Zhao^3*

• ¹Henan University of Science and Technology, Luoyang, China
• ²Zhengzhou University of Aeronautics, Zhengzhou, China
• ³University of Science and Technology Beijing, Beijing, China

With the widespread application of hyperspectral images in remote sensing, agriculture, and environmental monitoring, efficient detection and classification have become important research directions. In this paper, we propose a hyperspectral image classification method based on the Center-Guided Weighted Sparse Attention Mechanism Network (CGWSAM-Net). Firstly, the center-guided weighted sparse attention mechanism is introduced, which can effectively enhance the attention to key regions and suppress the interference of irrelevant regions by dynamically assigning weights to each pixel and randomly removing some feature points. Secondly, in the feature extraction stage, a combination of multi-scale convolutional kernels (1x1, 3x3 and 5x5) is used to realize the fusion of different scales of information and effectively capture multi-level features from details to macroscopic. Finally, fusing the convolution results using residual connections alleviates the gradient vanishing problem and accelerates network convergence. Experiments show that CGWSAM-Net achieves classification accuracies of 98.17% and 94.77% on Gaofeng State-Owned Forest Farm and Yellow River Estuary datasets, respectively, demonstrating the effectiveness and robustness of the method.