Traditional linear models can be too simplistic for capturing the myriad of interactions that occur among the triad of plant, pathogen and environment that results in plant disease epidemics. Tree-based machine learning (ML) algorithms are an attractive modeling solution because they automatically capture interactions, but work best when trained on a large input predictor matrix. In this study, multiple environmental and soil factors were collated from freely available gridded datasets and downscaled to better match the locations of snap bean fields evaluated for white mold across the central and western New York (NY) landscape. Functional data analysis of downscaled weather time series relative to planting was used to extract succinct summaries associated with disease occurrence. Summaries were input into a data matrix for training an ensemble tree model (XGBoost) fitted to prevalence of white mold. Environmental variables were ranked based on the contribution of their SHapley Additive exPlanations values to the fitted model. Nonlinear effects of weather and soil variables within the flowering period were detected. Most unexpectedly, air and soil temperatures at planting and in the weeks thereafter were associated with disease, which is not manifest until at least 40 days later. This study used a workflow of downscaled, gridded weather data, predictor selection, and ML model interpretation to enhance the understanding of environmental effects at a regional scale on one of the most prevalent and economically important diseases of snap beans. The proposed conceptual framework is broadly applicable to the regional prediction of other plant diseases.