Lightweight Graph Attention Network-Based Framework for Intelligent Frangipani Leaf Disease Classification

Frangipani (Plumeria) is a valuable ornamental and medicinal herb, whereas the foliar diseases may seriously affect the quality of leaves and plant health, whereas the diagnosis is slow and subjective by hand. The paper focuses on an effective image-to-graph model to develop Frangipani leaf disease classification based on a frozen ResNet-18 feature extractor and a tailored lightweight Graph Attention Network (GAT). The embeddings of every leaf image are faced in a k-nearest-neighbor graph (k=5, cosine similarity), and a small two-layer single-head GAT (1,563 trainable parameters) is used to classify each node in the k-nearest-neighbor graph. The lightweight GAT achieved a validation accuracy of 99.62% and a test accuracy of 99.37% on the proposed Frangipani leaf dataset, while maintaining high precision, recall, and F1-score across all classes, it can be deployed on the edges. Grad-CAM visualization is interpretable, where the disease-relevant areas are emphasized, and the trained model is presented in an Android-based application to make real-time on-device diagnoses.