MSCM-net: A hyperspectral image classification method based on multi-scale convolution and Mamba

Hyperspectral image (HSI) classification is a critical task in remote sensing, requiring the effective fusion of spatial and spectral information. While Convolutional Neural Networks (CNNs) excel at local feature extraction and Transformers capture long-range dependencies, they often suffer from limited receptive fields or high computational costs, respectively. This paper introduces MSCM-net to address these limitations by combining the strengths of multi-scale CNNs and the Mamba state space model.

The MSCM-net architecture consists of three primary components:

1. Multi-Scale Convolutional Feature Extraction (MCSE): This module uses three parallel branches (1x1, 3x3, and 5x5 convolutions) to capture spatial information at different scales. It incorporates an SENet (Squeeze-and-Excitation) mechanism to adaptively weight these features, ensuring the model focuses on the most discriminative information.
2. Mamba-based Sequence Modeling: The spatial features are flattened into sequences and processed by stacked Mamba blocks. This allows the model to capture long-range dependencies across the image patch more efficiently than traditional self-attention mechanisms.
3. Dual-Branch Feature Aggregation: The model uses two parallel paths to extract features: one focusing on the central pixel (which typically contains the most relevant category information) and another performing global average pooling to capture overall image statistics. These are fused to produce a robust final representation.

The authors evaluated MSCM-net on three standard HSI datasets: Indian Pines, WHU-Hi-HongHu, and Salinas. The results demonstrate that the model achieves state-of-the-art classification accuracy while maintaining lower computational complexity than Transformer-based alternatives. The ablation studies confirm that the combination of multi-scale convolution, SENet, and the dual-branch aggregation strategy significantly contributes to the model's performance.