A Bayesian Causal Model to Support Decisions on Treating of a Vineyard

Abstract

Plasmopara viticola is one of the main challenges of working in a vineyard as it can seriously damage plants, reducing the quality and quantity of grapes. Statistical predictions on future incidence may be used to evaluate when and which treatments are required in order to define an efficient and environmentally friendly management. Approaches in the literature describe mechanistic models requiring challenging calibration in order to account for local features of the vineyard. A causal Directed Acyclic Graph is here proposed to relate key determinants of the spread of infection within rows of the vineyard characterized by their own microclimate. The identifiability of causal effects about new chemical treatments in a non-randomized regime is discussed, together with the context in which the proposed model is expected to support optimal decision-making. A Bayesian Network based on discretized random variables was coded after quantifying the expert degree of belief about features of the considered vineyard. The predictive distribution of incidence, given alternative treatment decisions, was defined and calculated using the elicited network to support decision-making on a weekly basis. The final discussion considers current limitations of the approach and some directions for future work, such as the introduction of variables to describe the state of soil and plants after treatment.