CSA-Kansformer : Cross-scale aggregation and Kansformer network for hyperspectral image classification.

Abstract

Hyperspectral image (HSI) classification is essential in remote sensing, leveraging rich spectral and spatial information. Convolutional neural networks (CNNs) excel at extracting local features, while transformers capture global semantic information. However, CNNs struggle with global dependencies, and transformers face high computational costs for hyperspectral data. To combine the strengths of both, this paper introduces the CSA-Kansformer model, which incorporates three key modules to enhance performance and efficiency. First, the spatial and channel reconstruction convolution (SCConv) block is introduced to reduce redundant features by employing spatial redundancy units (SRUs) and channel redundancy units (CRUs), thereby reducing computational cost and extracting abstract spatial-spectral features. Second, we build an innovative cross-scale aggregation module (CSAM) combining the fusion convolution module (FCM), channel attention module (CAM), and spatial attention module (SAM). This novel design enables efficient multi-scale feature aggregation, selectively emphasizing key spatial and spectral information, thereby enhancing feature representation and improving accuracy in HSI classification tasks. Finally, we introduce the Kansformer block, which replaces layer normalization with batch normalization to enhance training stability, accelerate convergence, and improve generalization in HSI classification. Furthermore, it optimizes feature learning by replacing the multi-layer perceptron (MLP) with the Kolmogorov-arnold network (KAN), boosting both training efficiency and model performance. We conducted extensive experiments on four widely recognized hyperspectral datasets (Botswana, Houston2013, WHU-Hi-HanChuan, and WHU-Hi-HongHu). The results show that the proposed CSA-Kansformer model significantly outperforms nine state-of-the-art methods in both classification accuracy and computational efficiency. The code of this work is available at https://github.com/wanxiaoqing/CSA-Kansformer.