Enhancing detection of common bean diseases using Fast Gradient Sign Method-trained Vision Transformers.

Abstract

Common bean production in Tanzania is threatened by diseases such as bean rust and bean anthracnose, with early detection critical for effective management. This study presents a Vision Transformer (ViT)-based deep learning model enhanced with adversarial training to improve disease detection robustness under real-world farm conditions. A dataset of 100,000 annotated images augmented with geometric, color, and FGSM-based perturbations, simulating field variability. FGSM was selected for its computational efficiency in low-resource settings. The model, fine-tuned using transfer learning and validated through cross-validation, achieved an accuracy of 99.4%. Results highlight the effectiveness of integrating adversarial robustness to enhance model reliability for mobile-based plant disease detection in resource-constrained environments.

Plant phenotyping relevance

植物の病徴を画像から検出するVision Transformer手法の開発・頑健性検証が中心であり、植物病害状態の表現型推定に該当する。

Code and data availability

The paper's field-collected common bean disease image dataset (59,072 images, annotated, four classes) was published on Zenodo, with the exact URL given in the data availability statement and footnotes. This is a paper-specific, public, directly actionable asset. No code or trained model deposit is explicitly stated.