Prediction of coffee rust disease using Bayesian networks

In this paper we present an agricultural case study for learning Bayesian networks (BNs), namely prediction of coee rust. Wide-spread in all major production areas, coee rust causes premature defoliation, weakening the plant and reducing subsequent yield. It is typically controlled by use of chemical fungicides which must be applied before symptoms of infection are observed. Improved prediction would reduce the use of fungicides, producing healthier quality product and decreasing both economic costs and environmental impact. We use a dataset obtained from an experimental farm in Brazil over 8 years. Our preliminary data analysis informed our pre-processing of the original dataset. We also identified a number of structural priors, which our BN learner, CaMML (Causal Minimum Message Length), incorporates into its scoring metric and hence into its structure learning. Previous research has applied other classification methods to this coee rust dataset. We compare the results from a range of BNs learnt by CaMML with these previous approaches. The incorporation of structural priors in the BN learning yielded better models in terms of accuracy and interpretability. Although the BN’s predictions were comparable with some of the other techniques, they were clearly worse than decision trees, which seem to be taking advantage of context sensitive cases; this suggests avenues for improving the quality of the BN models.