Visual Interpretation and Classification of Apple Leaf Diseases via Grad-CAM and Convolutional Neural Networks

Abstract

Apple cultivation is a crucial agricultural activity in various mountainous regions, playing a vital role in supporting the local economy and sustaining the livelihoods of farmers. Several prominent mountain districts are known for leading apple production. However, apple orchards in these areas are often threatened by numerous diseases that reduce fruit yield and quality. In this research, we suggest a machine learning-based technique to automate the detection and classification of common apple diseases based on images of apple leaves collected from various regions. Through the use of Convolutional Neural Networks (CNN), the system can classify diseases with 97.36% precision. For post hoc explainability, Grad-CAM is used, which highlights the important regions that influenced CNN’s decision. The automated disease detection tool provides farmers in Nepal’s rural mountain areas with an affordable real time solution to monitor orchard health, minimize crop loss, and improve apple production. The dataset used in this study is originally derived from the United States based PlantVillage dataset, which is widely used for apple leaf disease classification research. Although the dataset is not collected from Nepal, the visual characteristics of apple leaf diseases remain largely consistent across regions due to similar biological infection patterns. Therefore, the model trained on this dataset is applicable to Nepali apple cultivation environments as well. At present, a publicly available or annotated Nepali specific apple leaf disease dataset is not available, which limits region-specific training and evaluation.

Plant phenotyping relevance

リンゴ葉画像から病害状態を分類するCNNベースの手法とGrad-CAMによる解釈を中心に扱うため、植物病害フェノタイピング手法として該当する。

Code and data availability

The paper's apple leaf disease image dataset (9,696 images, four classes) is publicly available on Kaggle and explicitly cited by the authors as the dataset used for training and evaluation. No author code, trained model, or other paper-specific assets are reported.