Detection of fusarium head blight in wheat under field conditions using a hyperspectral camera and machine learning

Fusarium head blight (FHB) is among the most devastating fungal diseases in cereal crops, reducing yield, and affecting human and livestock health through the production of mycotoxin. Despite application of fungicides, complete eradication of disease is virtually impossible in the field. There is a need for a disease detection technology during late growing stage for estimation of yield affected with FHB and for potential selective harvesting. Most published studies have focused on FHB detection during the milk growth stage using hyperspectral cameras. This preliminary study attempted to fill the knowledge gap by detecting FHB at the ripening stage. A spectral library of healthy and infected ears was collected with a hyperspectral camera in the visible and near-infrared region, over the canopy of eight different wheat varieties. The ears were segmented from the background using a simple linear iterative clustering (SLIC) superpixel algorithm on the normalized difference vegetation index (NDVI) images. Three different machine learning methods, namely, support vector machine (SVM), artificial neural network (ANN), and logistic regression (LR), were utilized for classification. To visualize the FHB distribution in the hypercube, the best performing model was applied for predicting the infected ears in the canopy images. The percentage area coverage of FHB for each hypercube was estimated. Results showed that the SVM algorithm produced the best classification accuracy (CA) of 95.6 % in the test set, followed successively by ANN and LR with CA values of 82.9 and 82.5 %, respectively. Interestingly, the preliminary study shows significant differences in spectral reflectance according to the variety of different resistance levels. The study also proves the feasibility of FHB detection using the developed prediction model during late growth stage with the potential of yield loss estimation before harvest.