Adaptive Dual-Stage Feature Segmentation and Hierarchical Residual Attention Network for Accurate Corn and Soyabean Leaf Disease Prediction

Accurate disease prediction in corn and soybean crops is required to enhance agricultural productivity and ensure food security. The changing texture of the leaf and illumination create errors in the segmentation process of the conventional disease identification methods. Even though deep learning methods suggest promising outcomes, they tend to get stuck in local solutions and overfit, thus they generalize poorly across different illnesses. Lack of data and unbalanced classes are problems for deep learning in plant disease diagnosis, which leads to biased models and inaccurate generalization. This paper introduces an original approach of Adaptive Feature Segmentation and Dual-Stage Filtering (AFS-DST) and a Hierarchical Residual Attention Network (HRAN) to address the above-mentioned problems. The AFS-DST technique is better because it guarantees the accuracy of the segmentation in the identification of the disease, and it dynamically optimizes features based on color and texture. The HRAN model has a residual attention block and hierarchical feature fusion to improve the generalization and robustness of the model by focusing on disease-specific patterns. Different with the previous methods, the Adaptive Feature Segmentation (AFS) with Dual-Stage Filtering and HRAN is a superior method because it is dynamically updated to adapt the segmentation as per the lighting, the disease severity, and the textures of the leaves. HRAN enhances the predictive accuracy and resistance of the models to environmental variation by attending to multi-level model-specific patterns and therefore enhancing the generalization as well as the robustness of the models. As results of the experiment on the corn and soybean datasets indicate, the proposed model is far more superior to other classic models, with a precision of 97.6 and a little loss incurred in corn, and 88.1 and better results in soybean. These results indicate that the proposed model reduces overfitting and local minima but also sets a new benchmark for disease detection on these crops while providing a strong basis for future research in agricultural disease prediction.