Next generation crop protection: A systematic review of trends in modelling approaches for disease prediction

Digital agriculture tools and advances in modelling approaches have the potential to deliver precise decision support systems for more effective, efficient and sustainable crop disease management. Historically, disease prediction in agriculture has relied on knowledge of the relationships between a few key environmental parameters and crop disease development. The emergence of new sensor technologies is now expanding the range of input data readily accessible for use in modelling. In addition, Artificial Intelligence (such as machine learning and deep learning algorithms) offers the capacity to process the large datasets available from a wider range of input variables relating to the three components of disease development: host, pathogen and environment. This review examined the rate and extent to which machine learning has replaced traditional modelling approaches for disease predictive model development. A systematic protocol was developed to investigate trends in modelling approaches for disease prediction in four major crop types: cereals, grape, potato and citrus. A total of 104 publications, reporting on the development of 146 disease predictive models were evaluated for modelling approach, data inputs and model performance. The results from this review indicate that the application of machine learning for predictive model development has greatly increased over the past two decades. Increased application of machine learning models (including Support Vector Machine and Random Forest) was associated with the development of more high-performance models and incorporation of higher numbers of predictor variables. The potential of deep learning models to deliver more precise and adaptable models for next generation disease management will be determined by applying these methods to large datasets. Further research is needed to investigate multi-model, machine learning approaches for disease prediction and to ensure model design captures important input variables relating to environment, host and pathogen.