Bayesian network-informed conditional random forests for probabilistic multisite downscaling of precipitation occurrence

This work introduces Bayesian network-informed conditional random forests (BNICRF): a novel multiresponse classification method for downscaling the joint probability distribution of precipitation occurrence at multiple geographical locations from large-scale reanalysis predictors. BNICRFs combine a Bayesian network to model spatial dependencies and a set of random forests to predict local precipitation from large-scale inputs. Extending prior studies on Bayesian networks and random forests in climate downscaling, this method is validated under the experimental framework of the COST action VALUE (the largest, most exhaustive intercomparison study of statistical downscaling methods to date). Results demonstrate that BNICRFs effectively capture spatial relationships while maintaining single-site predictive performance comparable to single-site random forests and a well-performing generalized linear model in VALUE. Additionally, BNICRFs outperform a robust multisite approach proposed in Chandler (2020) in predictive capability while matching spatial performance. Incorporating temporal structures further enables BNICRFs to generate temporarily and spatially realistic precipitation occurrence fields.