Combining ultralow-altitude drone phenotyping with deep learning analytics to assess resistance and disease dynamics of Fusarium head blight in wheat

Abstract

Fusarium head blight (FHB) is a serious fungal disease that affect small grain cereals, causing significant wheat ( Triticum aestivum L.) yield and quality losses globally. Breeding disease-resistant wheat varieties is key to address FHB-related challenges, but its progress is delayed by traditional methods due to the small-scale, laborious and relatively subjective nature of manual assessment. This study presents a new approach that combines ultralow-altitude drone phenotyping with an optimized You Only Look Once (YOLO) model to examine FHB in wheat, enabling us to perform large-scale and automated symptomatic analysis of this disease. We first established an Open FHB (OFHB) training dataset, consisting of 4867 diseased and 106,801 healthy spikes collected from 132 commercial breeding lines during FHB progression. Then, a deep learning model called YOLOv8-WFD was trained for detecting healthy and diseased spikes, followed by an adaptive Excess Green method to identify symptomatic regions and thus FHB-related traits on spikes. To study resistance levels, we employed an unsupervised SHapley Additive exPlanations (SHAP) method to pinpoint key traits between 10 and 20 d after inoculation (DAIs), resulting in the classification of 423 varieties trialed during the 2023–2024 growing seasons into four resistance levels (i.e., highly and moderately susceptible, and moderately and highly resistant), which were highly correlated with field specialists’ evaluations. Finally, we derived disease developmental curves based on measures of key traits during 10–20 DAI, quantifying varietal disease progression patterns over time. To our knowledge, this work represents a significant advancement in large-scale disease phenotyping and automated analysis of FHB in wheat, providing a valuable toolkit for breeders and plant researchers to assess resistance levels, select disease-resistant varieties, and understand dynamics of the fungal disease.

Plant phenotyping relevance

ドローン画像、深層学習、症状領域抽出を組み合わせ、コムギ穂のFHB症状・関連形質と病害進展を自動定量する手法が研究の中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.