ResNet-18 Convolutional Neural Network Framework for Tomato Disease Detection with Agrochemical Dosage Recommendation

Abstract

This study presents a deep learning-based framework for automated tomato leaf disease detection using a transfer learning approach built on ResNet18 architecture. The system is designed to classify 10 disease categories using image inputs resized to 224×224 pixels, leveraging ImageNet pre-trained weights to enhance feature extraction. The dataset consists of approximately 18,345 training samples with a batch size of 16 and trained over 10 epochs using the Adam optimizer and CrossEntropy loss function. Experimental results demonstrate strong classification performance, achieving a peak validation accuracy of 98.2% at epoch 9, with validation loss reduced to 0.64 from an initial value of approximately 1.8. The model shows high confidence predictions, with over 90% of test samples falling within the 90–100% confidence range. Per-class F1-scores range between 0.82 and 0.95, indicating consistent performance across multiple disease categories. The confusion matrix reveals strong diagonal dominance, confirming correct classification for most classes; however, specific misclassification patterns were observed. For instance, Bacterial_spot and Early_blight exhibit mutual confusion, while Septoria_leaf_spot shows complete misclassification into Tomato_mold, indicating limitations in distinguishing visually similar disease patterns. A balanced test dataset with class distribution ranging between 9% and 11% ensures unbiased evaluation. Additionally, Grad-CAM visualization confirms that the model focuses on biologically relevant regions such as lesion areas and leaf textures, improving interpretability. Despite achieving high accuracy under controlled conditions, the model’s generalization to real-world environments remains a challenge. The study highlights the need for improved robustness against variations in lighting, background, and disease severity. Overall, the proposed system demonstrates strong potential for precision agriculture applications, particularly in mobile-based disease diagnosis systems for farmers.

Plant phenotyping relevance

トマト葉の病徴を画像から分類する深層学習手法を開発・評価しており、植物の病害状態を直接推定する方法が研究の中心です。

Code and data availability

植物フェノタイピング解析を再現する公開資産であることを、入力本文と直接リンクから確認できなかったため保留しました。

No evidence-backed public reproduction asset is currently recorded.