Advanced Plant Disease Recognition using Graph Neural Networks and Semantic Segmentation

Plant diseases threaten global crop yields, necessitating efficient detection methods. We propose a hybrid framework combining semantic segmentation, automated feature extraction, and graph-based classification for multi-class disease identification. Using the Kaggle Plant Diseases Dataset (87,900 images, 38 classes), our method integrates U-Net (intersection over union: 0.85) for lesion localization, a neural architecture search-optimized convolutional neural network (1.5M parameters) for 128D features, and a graph attention network (GAT) to model spatial relationships. It achieves 99% average precision on the validation set, runs efficiently on a standard principal component (PC), and is suitable for edge deployment—offering a lightweight, robust solution for early disease detection in precision agriculture.