Tomato Leaf Disease Identification Using EfficientNetB3 Transfer Learning and Grad-CAM Explainable Analysis

Abstract

Abstract Tomato leaf diseases (TLDs) such as Early Blight (EB), Late Blight (LB), and Leaf Mold (LM) have a negative impact on crop yield and quality, result-ing in economic losses in agriculture. Traditional diagnosis methods depend on the visual recognition of a disease, which are time-consuming, subjective and may be difficult to reach in rural settings. Keeping these drawbacks in mind, the present work introduces a Transfer Learning model for tomato leaf disease classification based on EfficientNetB3 network. A pretrained Ef-ficientNetB3 network, trained on ImageNet, is used to obtain discriminative features like lesion boundaries, discoloration, fungal textures, and infection spots. Images are cropped to 224 × 224 pixels and split into training, validation and test set. The proposed architecture employs Global Average Pooling, a dense layer with ReLU activation, drop out regularization and Softmax classifier for classification of tomato leaf images into four classes: Early Blight, Late Blight, Leaf Mold and Healthy. Experimental results show the high classification accuracy and low training, validation and test-ing losses. Inter-class misclassifications of the confusion matrix show very few, which supports good generalization. Moreover, Grad-CAM visualiza-tions can generate interpretable heat maps that point to the regions of the image affected by the disease, and multi-class ROC analysis gives high AUC values, which means that it has excellent class separability. The proposed framework provides a precise, reliable, and interpretable approach for auto-mated tomato leaf disease diagnosis, contributing to precision agriculture by assisting in early detection and prompt management of tomato diseases.

Plant phenotyping relevance

トマト葉画像から病害状態を推定する深層学習手法が研究の中心であり、Grad-CAMによる病変領域の解釈と性能評価も行っているため、植物表現型計測手法として採択。

Code and data availability

The paper's tomato leaf disease classification uses a publicly available Kaggle dataset (PlantVillage-derived tomato leaf images, 4000 images across 4 classes), explicitly declared in the Data Availability statement with a public URL. No author code, trained models, or other paper-specific assets are disclosed.