Estimating Fusarium head blight severity in winter wheat using deep learning and a spectral index

Abstract

Fusarium head blight (FHB) of wheat (Triticum aestivum L.), caused by the fungal pathogen Fusarium graminearum (Fg), reduces grain yield and quality due to the production of the mycotoxin deoxynivalenol. Manual rating for incidence (percent of infected wheat heads/spikes) and severity (percent of spikelets infected) to estimate FHB resistance is time‐consuming and subject to human error. This study uses a deep learning model, combined with a spectral index, to provide rapid phenotyping of FHB severity. An object detection model was used to localize wheat heads within boundary boxes. Corresponding boxes were used to prompt Meta's Segment Anything Model to segment wheat heads. Using 2576 images of wheat heads point inoculated with Fg in a controlled environment, a spectral index was developed using the red and green bands to differentiate healthy from infected tissue and estimate disease severity. Stratified random sampling was applied to pixels within the segmentation mask, and the model classified pixels as healthy or infected with an accuracy of 87.8%. Linear regression determined the relationship between the index and visual severity scores. The severity estimated by the index was able to predict visual scores (R2 = 0.83, p = < 2e‐16). This workflow was also applied to plot size images of infected wheat heads from an outside dataset with varying cultivars and lighting to assess model transferability. It correctly classified pixels as healthy or infected with a prediction accuracy of 85.8%. These methods may provide rapid estimation of FHB severity to improve selection efficiency for resistance or estimate disease pressure for effective management.