A Graph-Attentive and Texture-Refined Lightweight Network for Robust Cassava Leaf Disease Recognition

Timely identification of cassava leaf infections is important to protect yield quality and support sustainable crop management in low-resource agricultural regions. This research work introduces an advanced vision framework that utilizes a depth-efficient dual-path design with structural reasoning for high-accuracy disease categorization. The proposed model includes a connectivity-aware graph encoder with a directional texture reasoning block which provides comprehensive learning of spatial patterns and lesion propagation across the leaf surface. A dynamic token selection mechanism in the proposed work reduces redundant processing by prioritizing visually critical regions. Additionally, a cascaded classifier refines predictions across multiple representation scales. Experimental analysis of proposed model utilized cassava leaf dataset and detects diseases with superior accuracy of 94.92%, precision, recall, and F1-score values of 94.07, 94.66, and 94.36 respectively. The proposed model achieves a low inference latency of $\mathbf{1 2. 4}$ milliseconds which provides rapid support and reliable deployment in cassava lead disease diagnosis.