Enhancing Generalization of Tomato Leaf Disease Classification via TDR Model and Field-Conditioned Data Augmentation

Abstract

Tomato leaf diseases significantly affect agricultural productivity, particularly when detection systems are deployed under real-field conditions characterized by illumination variation, background clutter, and image noise. Although deep learning-based models have achieved high accuracy on laboratory datasets such as PlantVillage, their generalization performance often degrades when applied to real-world environments. This study proposes a lightweight CNN-based tomato leaf disease recognition model, referred to as the TDR-Model, combined with field-conditioned data augmentation strategies. The proposed model integrates MobileNetV3 with Convolutional Block Attention Module (CBAM) and Omni-Dimensional Dynamic Convolution (ODC) to enhance feature representation while maintaining computational efficiency. Field-conditioned augmentation using the Albumentations library to simulate real-world visual variations during training. The model is evaluated on the real-world tomato set consisting of 10 classes and 885 leaf images. Experimental results show that the proposed model achieves an overall test accuracy of 82.94%, with precision, recall, and F1-score of 85.06%, 83.04%, and 83.03%, respectively. Furthermore, the model requires only 3.47 million parameters, 0.23 GFLOPs, and an average inference time of 5.15 ms, making it suitable for real-time and resource-constrained agricultural applications. These results indicate that the proposed approach effectively balances accuracy and efficiency for practical tomato leaf disease detection.