Applications of Machine Learning to Understand and Predict Potato Blackleg

Abstract

Blackleg is a bacterial disease responsible for significant economic losses in many seed potato producing countries. Blackleg is mainly seedborne, meaning that latently infected tubers are key for pathogen transmission, while conducive weather and environmental conditions determine disease expression in field. Various control strategies have been tested with limited success, and the more effective are based on avoidance of contamination and production of healthy certified seeds.

This study sought to advance our knowledge of the influential drivers of blackleg and their relative importance in disease development to identify factors that limit control and provide solutions for better management. Longitudinal data obtained from statutory Scottish crop inspections were combined with host, weather, soil, and geographic data to construct a comprehensive epidemiological database for interrogation. Analysis of historical datasets using cutting-edge modelling techniques (machine learning) identified and ranked important predictor variables of blackleg on a landscape scale.

Given that there is a lack of models to predict blackleg in the field, a comparative analysis of several machine algorithms was performed to obtain spatiotemporal predictions of blackleg. Results highlighted that interannual prediction of the disease is challenging, although an algorithm that maximised prediction accuracy (AUROC = 0.84) was identified. The model can aid decision-making at the end of the growing-season (i.e., at harvest and post-harvest stages, which are critical for disease prevention). The algorithm was further tested for its ability to predict blackleg ahead of the upcoming growing-season (forecasting). A key result of this work is that satisfactory forecasting can be obtained by combining readily available data with bioclimatic variables representative of years with low and high disease pressure.

The model obtained was then developed into a user-friendly risk assessment tool that can be used by farmers, industry, and research institutions as an educational aid or a tool to support decisions about agronomic choices that could mitigate risk.

Date of Award	2023
Original language	English
Awarding Institution	University of Dundee
Supervisor	Peter Skelsey (Supervisor), Sonia N. Humphris (Supervisor) & Ingo Hein (Supervisor)