CNN MODEL OPTIMIZATION USING MULTI-STAGE DATA AUGMENTATION FOR LOCAL PLANT LEAF DISEASE CLASSIFICATION

Abstract

Plant leaf diseases are a major factor in reducing agricultural productivity, particularly for local commodities that often lack adequate artificial intelligence-based disease detection systems. This study aims to optimize the performance of a Convolutional Neural Network (CNN) model using the Inception V3 architecture through the application of multi-stage data augmentation to improve the classification accuracy of local plant leaf diseases. The dataset used is PlantifyDR from Kaggle, which has limited data volume and visual variation, requiring an effective augmentation strategy to improve the model's generalization ability. The proposed multi-stage augmentation approach consists of three stages—geometric, photometric, and texture-noise augmentation—that systematically enrich the diversity of training images. Evaluation results show that the proposed model provides significant performance improvements compared to the baseline model. The Inception V3 model with multi-stage augmentation achieved an accuracy of 0.762, an F1-score of 0.727, and a perfect AUC (1.00) across all classes, while the baseline model only achieved an accuracy of 0.595 and an average AUC of 0.877. Accuracy, loss, ROC curve, and confusion matrix analyses confirmed that multi-stage augmentation reduced overfitting and enhanced the model's ability to differentiate disease symptoms across leaf types. Therefore, this study concludes that multi-stage data augmentation is an effective approach for optimizing deep learning models on small and complex datasets, while also providing a significant contribution to the development of more accurate and reliable AI-based plant disease detection systems.

Plant phenotyping relevance

葉画像から植物病害症状を分類するCNNと、性能改善のための多段階データ拡張を開発・評価しており、植物の病害状態を推定する画像ベースのフェノタイピング手法が中心です。

Code and data availability

The paper's plant leaf disease classification experiments use the publicly available PlantifyDR Kaggle dataset (Apple, Berry, Guava leaf images), which is the image input for the study's phenotyping measurements. No author analysis code or trained model checkpoints are reported as publicly deposited.