Al Bahri, Safrizal Razali, Aulia Rahman
Tomato leaf disease detection using deep learning faces significant challenges due to limited and imbalanced datasets, leading to overfitting and poor generalization. This study presents a comprehensive evaluation of two advanced data augmentation techniques—CutMix and MixUp—on tomato leaf disease classification using the EfficientNetV2-M architecture. The research compares three experimental configurations: baseline (no augmentation), CutMix augmentation, and MixUp augmentation, using a balanced dataset of 4,000 images across eight disease classes. Results demonstrate a clear trade-off between accuracy and regularization. The baseline model achieved the highest test accuracy of 93%, establishing a strong performance benchmark but showing signs of overfitting. CutMix augmentation provided a balanced approach with 86% accuracy while offering improved regularization and stability, particularly benefiting distinct disease classes like Healthy (F1=0.94) and Tomato Yellow Leaf Curl Virus (F1=0.90). MixUp exhibited the strongest regularization effect but at the cost of significantly reduced performance (70% accuracy), with extreme class-dependent effects that severely impacted challenging categories like Early Blight (F1=0.45) and Late Blight (F1=0.51). The findings suggest that CutMix offers the most favorable balance for practical applications, providing reasonable accuracy while enhancing model robustness. This study provides valuable insights for selecting appropriate augmentation strategies in agricultural computer vision applications, particularly for tomato disease detection systems in real-world conditions. ©2025 IEEE.
Department of Electrical and Computer Engineering, Universitas Syiah Kuala, Banda Aceh, Indonesia
Research at a Glance
Register to unlock
Topics & SDG Alignment
Register to unlock
Collaboration
Register to unlock
Author Profile (Selected)
Register to unlock
References Overview
Register to unlock
Journal & Source
Register to unlock
Metadata & Integrity
Register to unlock