HMS-Mamba: A hybrid multi-scale Mamba network for hyperspectral tobacco mildew detection

Hyperspectral imaging (HSI) provides rich spatial–spectral information for non-destructive tobacco leaf mildew detection. However, early mildew symptoms are weak, spatially localized, and irregularly distributed, while inter-cultivar spectral variability reduces the separability between healthy and mildewed leaves. Existing deep learning methods struggle to jointly enhance local mildew features, suppress redundant spectral responses, and integrate long-range context, limiting their robustness under multi-cultivar conditions. To address these challenges, this paper proposes HMS-Mamba, a hybrid architecture integrating multi-scale convolutions with the Mamba framework. Specifically, a progressive 3D–2D convolutional module extracts coupled spatial–spectral features and captures multi-scale spatial textures through residual 3D convolutions and dilated 2D convolutions. A Self-Attention Channel Gating (SACG) mechanism adaptively recalibrates channel-wise responses to enhance mildew-discriminative features while suppressing redundant spectral information. Finally, a Mamba module equipped with Convolutional Positional Encoding (CPE) preserves spatial position information in serialized features and efficiently models global spatial–spectral dependencies with linear computational complexity. Experimental results on a tobacco mildew hyperspectral dataset demonstrate that HMS-Mamba outperforms current methods in classification, confirming the framework's effectiveness.