The Detection of Fusarium Head Blight in Multiple Species of Wheat Using a Multispectral UAV in Southern Manitoba

Abstract

Fusarium Head Blight (FHB) is a fungal disease that affects cereals such as wheat, severely damaging the plant, reducing its yield and value, and potentially rendering it unsafe for human or animal consumption. Detection of FHB in wheat fields is essential due to the threat it presents to Canada’s agricultural production. UAV’s and remote sensing techniques have been adopted within precision agricultural approaches for canopy scale detection of the disease. These approaches have focused on the detection of the disease in a singular species of wheat in an experimental field. The goal set out by this thesis is the detection of FHB across a diversity of wheat species using a multispectral UAV in an experimental field, and the transfer of the detection model to a monoculture commercial wheat field. We collected multispectral UAV imagery, and ground-based measures of the state of FHB in randomly sampled plots in both the experimental and commercial fields. A selection of 15 Vegetation Indices (VIs) were then extracted from the multispectral imagery, chosen for their past FHB detection performance. Three supervised machine learning classification models were selected, support vector machines, random forest, and extreme gradient boosting based on their prior applications in disease detection in wheat. We determined a set of Key VIs sensitive to FHB across wheat species using Spearman’s Ranked correlation, feature importance in RF and XGB models, and replacement sampling. These key VIs along side the ground-based measures were used for the training and testing of the FHB detection models, and applied to the experimental and commercial field. In our results Green Leaf Index (GLI), Anthocyanin Reflectance Index (ARI), Plant Senescence Reflectance Index – red-edge (PSRI-RE), and Enhanced vegetation index (EVI) emerged as the Key VIs for their effectiveness across wheat species. Amongst the three models, XGB offered the greatest overall accuracy at 87% for disease detection, but all models were successful in distinguishing healthy and FHB diseased wheat in the experimental field. When applied to the commercial field, the models successfully distinguished healthy from non-healthy stressed wheat, but had difficulty with the interrow spacing from the experimental site. In this thesis we successfully detected FHB across a diverse collection of wheat species, and identify challenges when transferring from an experimental to commercial field.