Leaf-by-Leaf Diagnosis: A Custom CNN with Pyramidal Feature Extraction for Plant Disease Classification

Abstract

Plant diseases threaten global agriculture, causing 20–40% yield losses and food insecurity. Current diagnostic methods are costly and lack scalability. While deep learning advances plant disease detection, there remains a need for CNNs with simpler architectures, better generalizability, and lower computational cost. This study presents a novel CNN for multi-class classification of 38 diseases. Trained on a public dataset of over 87,000 RGB images, the architecture comprises five convolutional blocks (filters 32–512) with max pooling and dropout (0.25, 0.4), followed by a 1,500-unit dense layer and SoftMax output. Optimized with Adam (lr=0.0001) and categorical cross-entropy, the model achieved 98% training and 96% validation accuracy with approximately 28.7 million parameters significantly fewer than transfer learning architectures. These results demonstrate an effective balance between predictive performance and computational efficiency, positioning the model as a promising tool for real-world agricultural deployment.

Plant phenotyping relevance

葉画像から植物病害を分類するCNNの開発と性能評価が研究の中心であり、植物の病害状態を直接推定するため、植物フェノタイピング手法に該当します。

Code and data availability

The paper's sole qualifying asset is the plant disease image dataset used for all its CNN training/validation measurements: the publicly available New Plant Diseases Dataset (Augmented) on Kaggle, explicitly declared in the Data availability statement. No author code, trained model checkpoints, or other paper-specific