From pixels to plant health: accurate detection of banana Xanthomonas wilt in complex African landscapes using high-resolution UAV images and deep learning

Abstract

Abstract Bananas and plantains are vital for food security and smallholder livelihoods in Africa, but diseases pose a significant threat. Traditional disease surveillance methods, like field visits, lack accuracy, especially for specific diseases like Xanthomonas wilt of banana (BXW). To address this, the present study develops a Deep-Learning system to detect BXW-affected stems in mixed-complex landscapes within the Eastern Democratic Republic of Congo. RGB (Red, Green, Blue) and multispectral (MS) images from unmanned aerial vehicles UAVs were utilized using pansharpening algorithms for improved data fusion. Using transfer learning, two deep-learning model architectures were used and compared in our study to determine which offers better detection capabilities. A single-stage model, Yolo-V8, and the second, a two-stage model, Faster R-CNN, were both employed. The developed system achieves remarkable precision, recall, and F1 scores ranging between 75 and 99% for detecting healthy and BXW-infected stems. Notably, the RGB and PAN UAV images perform exceptionally well, while MS images suffer due to the lower spatial resolution. Nevertheless, specific vegetation indexes showed promising performance detecting healthy banana stems across larger areas. This research underscores the potential of UAV images and Deep Learning models for crop health assessment, specifically for BXW in complex African systems. This cutting-edge deep-learning approach can revolutionize agricultural practices, bolster African food security, and help farmers with early disease management. The study’s novelty lies in its Deep-Learning algorithm development, approach with recent architectures (Yolo-V8, 2023), and assessment using real-world data, further advancing crop-health assessment through UAV imagery and deep-learning techniques.

Plant phenotyping relevance

UAV画像と深層学習により、バナナの健全・感染茎および植物病害状態を直接推定する手法を開発・比較・評価しており、フェノタイピング手法が中心である。

Code and data availability

The paper's UAV imagery and field-sampling data (the core phenotyping/disease-detection inputs) are only available upon reasonable request from the corresponding author, so no public paper-specific dataset exists. The authors do publicly release their FALCONCV/CvStudio deep-learning platform used for model development,

No evidence-backed public reproduction asset is currently recorded.