Detection of Coffee Leaf Diseases Using Lightweight Deep Learning: A Comparative Study of EfficientNet-B0 and Vision Transformer

Abstract

This study compares lightweight deep learning models for coffee leaf disease detection using field-acquired images. Two architectures were evaluated: EfficientNet-B0, trained from scratch, and Vision Transformer (ViT), fine-tuned from pretrained weights. The dataset consisted of 843 balanced RGB images representing three major coffee leaf diseases: Leaf spot, Rust, and Sooty mold. Images were captured under natural field conditions without synthetic augmentation to ensure realistic evaluation. On the held-out test set, EfficientNet-B0 achieved an accuracy of 88.37%, precision of 87.9%, recall of 88.02%, and F1-score of 87.96%. ViT achieved an accuracy of 85.12%, precision of 84.76%, recall of 84.93%, and F1-score of 84.85%. Error analysis indicated that both models struggled to differentiate rust and sooty mold due to overlapping textural patterns. EfficientNet-B0 showed faster convergence and higher robustness, making it more suitable for mobile and edge deployment. ViT, while slightly less accurate, demonstrated stable learning behavior and potential benefits from larger or more diverse datasets. The results demonstrate feasibility for mobile deployment in real-time field diagnosis, providing a practical benchmark for lightweight AI in precision agriculture.