A hybrid CNN–Mamba architecture with multiscale and dynamic window aggregation for hyperspectral image classification

Hyperspectral image (HSI) classification presents significant challenges due to the high dimensionality and complex spectral–spatial dependencies inherent in the data. To tackle these challenges, we propose a hybrid multiscale CNN–Mamba network with dynamic window aggregation for efficient and accurate HSI classification. First, a multiscale CNN module is employed to extract shallow spectral–spatial features, capturing both local textures and structural details. Subsequently, a dynamic window aggregation module, consisting of fully connected layers and a global pooling layer, assigns windows of varying sizes to different attention heads in a window-based multihead self-attention mechanism. This allows the model to effectively capture multiscale contextual information. The outputs from multiple scales are dynamically fused by adaptive weighting of the window branches, and the concatenated features are forwarded into a single-scan Mamba module. Leveraging a three-directional scanning strategy, the Mamba module enhances long-range dependencies and global feature learning. Further improve the extracted feature we proposed spatial channel attention module. To further improve representation, we introduce a feed-forward module augmented with depthwise convolution, which enriches spatial correlations while maintaining computational efficiency. Extensive evaluations conducted on five publicly available HSI datasets—Salinas, Pavia University, Holden Carter, WHU-Hi-HongHC, and the Chinese satellite GF-5 Yancheng present the validity of the proposed approach. The results indicate that it consistently delivers excellent performance, substantially surpassing existing approaches in terms of overall accuracy (OA), average accuracy (AA), and the Kappa coefficient.