Enhancing Leaf Disease Classification Using GAT-GCN Hybrid Model
To address low classification accuracy and poor generalization under small-sample conditions in crop leaf disease identification, this paper proposes a collaborative graph learning framework integrating Graph Attention Networks (GAT) and Graph Convolutional Networks (GCN). Methodologically: (i) superpixel segmentation (SLIC) is employed to construct robust local graph structures; (ii) an edge-enhancement strategy is designed to improve discriminability for sparse disease classes; and (iii) Xavier initialization is adopted to accelerate model convergence. Key contributions include: the first GAT-GCN collaborative architecture for fine-grained lesion modeling; deep integration of superpixel segmentation with graph learning to enhance spatial localization; and significant improvement in small-sample robustness via edge enhancement. Experimental results on apple, potato, and sugarcane datasets achieve F1-scores of 0.9818, 0.9743, and 0.8799, respectively—consistently outperforming individual baseline models—and demonstrate strong cross-crop generalization capability.