Wavelet-prior-guided mamba network for accurate and efficient rice disease recognition

Rice disease recognition is of great importance for sustainable agricultural production. Ho wever, most existing methods mainly rely on spatial-domain feature modeling and insufficiently exploit frequency-domain information, making it difficult to fully characterize the texture and structural patterns of diseased images across different frequency components. In addition, conventional convolutional neural networks have limitations in modeling long-range dependencies, which may reduce their ability to recognize lesion regions with uneven distributions or highly variable appearances. To address these challenges, this study proposes a Wavelet-Prior-Guided Mamba network, termed WPMamba, for rice disease recognition. Specifically, discrete wavelet transform is introduced to decompose image features into different frequency components, thereby explicitly incorporating frequency-domain prior information. The decomposed features are further integrated with Mamba modules to enhance global feature modeling and representation capability. Moreover, a prior-guided selective fusion (PGSF) module is designed to adaptively fuse spatial-domain and frequency-domain features, while a lesion-aware spatial attention (LASA) module is introduced to guide the network toward disease-relevant regions. Experimental results on a rice disease dataset demonstrate that the proposed WPMamba achieves an accuracy of 95.72%, outperforming several mainstream models. Meanwhile, WPMamba contains only 1.21 M parameters and requires 0.49 GFLOPs, indicating its favorable balance between recognition performance and computational efficiency. These results suggest that WPMamba provides an effective and efficient solution for intelligent rice disease recognition.