A precision grading method for walnut leaf brown spot disease integrating hierarchical feature selection and dynamic multi-scale convolution

Walnut leaf brown spot disease, caused by <i>Ophiognomonia leptostyla</i>, is among the most destructive fungal diseases in walnut cultivation. In the development of smart agriculture, precision grading of plant diseases remains a core technical challenge; specifically, this disease is plagued by blurred lesion edges and inefficient extraction of complex features, which directly limits the accurate grading of the disease. To address these issues, this study proposes a disease grading method integrating hierarchical feature selection and adaptive multi-scale dilated convolution, and develops the CogFuse-MobileViT model. This model overcomes the limitations of the standard MobileViTv3 model in capturing blurred edges of tiny lesions via three innovative modules: specifically, the Hierarchical Feature Screening Module (HFSM) enables hierarchical screening of disease-related features; the Edge Feature Focus Module (ECFM) works in synergy with the HFSM to enhance the focus on lesion edge features; and the Adaptive Multi-Scale Dilated Convolution Fusion Module (AMSDIDCM) achieves dynamic multi-scale fusion of lesion textures and global structures. Experimental results demonstrate that the proposed model achieves an accuracy of 86.61% on the test set, representing an improvement of 7.8 percentage points compared with the original MobileViTv3 model and significantly outperforming other mainstream disease grading models. This study confirms that the CogFuse-MobileViT model can effectively resolve the issues of blurred edges and inefficient feature extraction in this disease, provides a reliable technical solution for its precision grading, and holds practical application value for the intelligent diagnosis of plant diseases in smart agriculture.