Attention-enhanced GNN model for fungal disease classification in spinach leaves using monospectral imaging.

Abstract

Plant leaf diseases must be detected and treated early to improve crop yield and reduce agricultural losses. However, pixel-level representations and the inability to be read limit the applicability of existing deep learning approaches to the agricultural sector. A graph neural network termed the Attention-Enhanced Graph Neural Network (AE-GNN) may explain and diagnose multi-plant leaf disease. The proposed framework models leaf pictures as a graph with nodes representing discriminative leaf areas and edges representing their spatial connection. Before creating the global context vector and classification, graph features are aggregated, and an attention weighting method is applied to refocus on disease-relevant nodes obscured by less informative background characteristics. Final disease prediction uses a multilayer perceptron classifier. A curated dataset of half-spinach and curry leaf pictures is used to assess the proposed method for fifteen illnesses and their healthy classifications. Grad-CAM-based explainable AI methods make the model predictions' most important areas clearer. The dataset and source code from this work are available on GitHub for reproducibility and openness. Experimental results reveal that the proposed AE-GNN outperforms convolutional neural networks and graph-based models in classification. Graph-structured learning, attention enhancement, and explainability create a robust and interpretable framework for multi-plant leaf disease diagnosis.

Plant phenotyping relevance

葉画像から植物病害状態を分類・診断する画像解析手法を提案し、既存モデルとの比較評価と説明可能性解析を行っているため、植物フェノタイピング手法が中心である。

Code and data availability

The paper's Data availability section explicitly links a public GitHub repository containing the paper's spinach/curry leaf fungal disease image dataset used for the AE-GNN phenotyping/classification analysis.