Deep Learning-Based Tomato Leaf Disease Classification Using CNN, EfficientNetB0, and InceptionResNetV2

Abstract

Tomato leaf diseases threaten agricultural productivity because symptoms such as early blight, late blight, leaf mold, septoria leaf spot, and yellow curl virus often produce visually similar color changes, necrotic lesions, and leaf deformation. Manual visual diagnosis is subjective and depends heavily on expert experience; therefore, image-based deep learning is a relevant approach for supporting preliminary disease identification. This study compares five deep learning architectures, namely a custom convolutional neural network, EfficientNetB0, MobileNetV2, DenseNet121, and InceptionResNetV2, for classifying six tomato leaf categories using 7,192 images from a Kaggle dataset. The research workflow includes dataset preparation, image resizing and normalization, model training using the Adam optimizer, and evaluation through accuracy, loss, precision, recall, F1-score, and confusion matrix analysis. Based on the notebook results, EfficientNetB0 achieved the best validation accuracy of 89.44% after 20 epochs, followed by MobileNetV2 at 85.12%, DenseNet121 at 82.77%, the custom CNN at 70.69% test accuracy, and InceptionResNetV2 at 45.76% test accuracy. The results indicate that lightweight transfer learning models are more suitable for medium-sized agricultural image datasets than large architectures trained for only a few epochs. Future work should validate the model using real field images, harmonize all models on the same test set, and report class-wise metrics to ensure reliability before deployment as a farmer-oriented diagnostic support system.