Explainable Deep Learning for Plant Leaf Diseases: A Comparative Study of Grad CAM

Abstract

Plant disease detection is critical for sustainable agriculture and food security. While deep learning models achieve high accuracy in leaf disease classification, their black box nature poses limitations for trust and adoption among agricultural practitioners. This study presents a comparative evaluation of three convolutional neural network architectures (ConvNeXt-Tiny, MobileNetV2, and VGG16) for classifying potato, maize, and pepper leaf diseases, with emphasis on explainability through Gradient-weighted Class Activation Mapping (Grad-CAM). The experimental results demonstrate that ConvNeXt-Tiny achieves 99-100% accuracy across all plant species, MobileNetV2 attains 97-100% accuracy with lower computational requirements, and VGG16 yields 97-99.5% accuracy. Grad-CAM visualizations reveal that modern architectures precisely focus on lesion regions, whereas older models occasionally attend to irrelevant features such as leaf veins and edges. Misclassification analysis identifies shadows and natural leaf patterns as primary error sources. This research demonstrates that explainable artificial intelligence is not merely complementary but essential for developing trustworthy agricultural decision support systems.

Plant phenotyping relevance

植物葉の病変領域を画像から分類・可視化する手法を比較評価しており、病害状態の表現型抽出が研究の中心です。

Code and data availability

The paper uses the public PlantVillage segmented dataset (a cited third-party resource, not a paper-specific deposit) and trains ConvNeXt-Tiny, MobileNetV2, and VGG16 models with Grad-CAM analysis. No authors' public code, trained checkpoints, or supplementary data repository is provided. The only availability language

No evidence-backed public reproduction asset is currently recorded.