Inferring multiple coffee flowerings in Central America using farmer data in a probabilistic model

Abstract

Coffee (Coffea arabica L.) is a climate-sensitive crop; rainfalls may trigger flowering event occurrences, and extreme rainfall during a flowering day can cause considerable yield reductions. Multiple flowering events can occur in the span of 12 months; the number varies from year to year. This paper introduces a Bayesian network model capable of inferring coffee flowering events in coffee areas in the Pacific Region of Central America based on observed data for coffee flowering and precipitation. The model structure was determined based on expert knowledge, and the model parametrization was learned from 53 years of data registered in the region. Data from four farms in the region were used for model validation. The model's performance in the inference of flowering intensity was good (spherical payoff of 0.78 out of maximal 1.00), and the model was able to depict expected behaviors for single and multiple flowerings. Further, comprehensive new details on the dynamics of multiple flowerings within a crop season were obtained, e.g., that a large flowering event tends to occur more quickly (8 to 10 days) after rain than a small flowering (10 to 13 days). We believe that this Bayesian network model has the potential to evolve and support the development of agricultural index-based insurance to deal with yield losses due to extreme rainfall during flowering. The use of longer farm records for model building may also serve to increase farmers' trust in the reliability of the tool.