Perithecia detection from images of stubble using deep learning models

Authors: AZIMI, HILDA - National Research Council - Canada; XI, PENGCHENG - National Research Council - Canada; CUPERLOVIC-CULF, MIROSALVA - National Research Council - Canada; Vaughan, Martha

Submitted to: Meeting Proceedings
Publication Type: Proceedings
Publication Acceptance Date: 10/1/2021
Publication Date: N/A
Citation: N/A

Interpretive Summary:

Technical Abstract: Fusarium Head Blight (FHB) is a serious fungal disease of cereal crops that can not only reduce grain yield and quality but can also contaminate grain with hazardous mycotoxins. In North America, FHB is predominantly caused by Fusarium graminearum (Fg). The primary form of Fg inoculum is ascospores, which are produced within small (<0.5mm), darkly pigmented fruiting bodies known as perithecia. The density of the Fg inoculum (i.e., the total number of perithecia) is associated with the potential for FHB severity. In order to provide growers with a timely tool to assess local FHB risk, we have developed machine learning models capable of detecting perithecia and estimating inoculum density from stubble images. We have implemented two deep learning-based object detection approaches, namely faster Region-based Convolutional Neural Networks (R-CNN) and You Only Look Once (YOLO). We have trained and tested a series of deep learning models through a lab-collected data set, with the best model achieving an average precision of 73% in perithecia detection. This work demonstrates the feasibility of applying machine learning to precision agriculture in the context of estimating pathogen inoculum density for disease forecasting.