3D convolution attention-based multi-scale fusion network for hyperspectral image classification

Deep learning (DL) has significantly advanced pattern recognition and hyperspectral image (HSI) classification owing to its strong capability for hierarchical feature representation. However, existing DL-based HSI classification methods are often limited by scarce labeled samples, high parameter complexity, and the difficulty of learning discriminative features from high-dimensional spectral-spatial data. To address these challenges, this paper proposes a 3D convolution attention-based multi-scale fusion network (3D-CAMFN) for HSI classification. Specifically, a 3D convolutional attention block is first designed to jointly capture spatial-spectral dependencies, enabling the network to adaptively emphasize informative spectral channels while suppressing redundant or noisy responses. Then, shallow, middle, and deep features extracted from different network depths are integrated through a multi-scale fusion mechanism, which effectively balances fine spatial details and high-level spectral abstractions. Finally, a broad learning system (BLS) is introduced to replace the conventional deep classifier head. By using random feature mapping and pseudoinverse-based analytical solutions instead of time-consuming back-propagation, BLS reduces trainable parameters and improves training efficiency. Experimental results on three widely used HSI datasets, namely Salinas (SA), Indian Pines (IP), and WHU-Hi-HanChuan (WHU-HC), show that the proposed 3D-CAMFN achieves competitive or superior classification performance compared with several recent state-of-the-art methods. These results demonstrate that the proposed framework can effectively enhance spatial-spectral feature representation while maintaining computational efficiency, making it suitable for accurate and efficient HSI classification.