AMnet: An Explainable Graph Based Inception Framework for Aegle marmelos Leaf Disease Classification

Abstract

Abstract Aegle marmelos (bael) is a medicinally important tropical crop that remains severely underrepresented in computational plant pathology research. This study proposes AMnet, a deep learning framework integrating an InceptionV3 backbone with a fixed graph convolutional network to capture both within-region disease texture and between-region spatial propagation patterns for automated four-class classification of Aegle marmelos leaf diseases: Cercospora leaf, healthy leaf, leaf curl, and leaf spot. The graph module constructs a 25-node spatial graph directly from CNN feature maps, enabling end-to-end training without external preprocessing. Compared against MobileNetV2, InceptionV3, VGG19, and DenseNet201, AMnet achieved the best overall performance with 98.83% accuracy, 98.84% F1-score, 99.95% PR-AUC, and 98.45% MCC, alongside the fastest inference time of 1.83 ms. Robustness was confirmed through bootstrap confidence interval estimation and four-fold cross-validation. PCA-based clustering analysis with silhouette scoring and Davies–Bouldin indexing demonstrated clear class separability in the learned embeddings. Grad-CAM and LIME visualizations confirmed that predictions were grounded in biologically meaningful leaf regions rather than background artifacts. A Gradio-based prototype further demonstrated practical deployment potential. Although broader field validation remains necessary, AMnet provides an accurate, interpretable, and reproducible framework for diagnosing Aegle marmelos leaf disease.

Plant phenotyping relevance

葉画像から植物の病害状態を分類する説明可能な深層学習手法を開発・比較・検証しており、植物表現型（病徴・病害状態）の取得・推定が中心的です。

Code and data availability

The paper analyses a publicly available Aegle marmelos leaf disease image dataset deposited on Mendeley Data, explicitly linked in the Data availability statement and reference [36]. No author analysis code or trained model checkpoint is reported as publicly available.