Comparative Evaluation of Deep Learning Models for Multi-Class Tomato Leaf Disease Classification

Abstract

Tomato leaf diseases severely reduce yield and quality, particularly in smallholder farming systems where access to expert diagnosis is limited. Automated image-based detection using deep learning offers a scalable solution, but appropriate model selection requires balancing predictive performance against computational complexity. Although many deep learning studies report high accuracy on curated datasets, fewer works discuss the practical accuracy–efficiency trade-off that determines whether a model can run on mobile or low-power devices. This study presents a comparative evaluation of three convolutional architectures: a custom CNN, MobileNet, and InceptionV3 for multi-class tomato leaf disease classification. A ten-class dataset comprising healthy leaves and nine common tomato leaf diseases was used. Images were resized to 224 × 224, normalized, segmented using Otsu thresholding to isolate the leaf region, and augmented during training to reduce overfitting. Under a common classification head which includes global average pooling, dropout, dense layer with 256 units, and a 10-class softmax output, the models were compared in terms of overall accuracy, per-class precision, recall, F1-score, depth, parameter count, and model size. MobileNet achieved the best overall accuracy of 88%, outperforming the custom CNN and InceptionV3, while also requiring the fewest parameters of 3.49M and the smallest memory footprint of 13.33MB. InceptionV3, despite its higher capacity of 22.33M parameters, showed inferior generalization under the given training method. The results indicate that MobileNet provides the most favourable balance between accuracy and efficiency and is particularly suitable for deployment in resource-constrained or on-device tomato disease diagnostic systems.