A posteriori random forests for stochastic downscaling of precipitation by predicting probability distributions

Abstract

This work presents a comprehensive assessment of the suitability of random forests, a well-known machine learning technique, for the statistical downscaling of precipitation. Building on the experimental and validation framework proposed in the Experiment 1 of the COST action VALUE—the largest, most exhaustive intercomparison study of statistical downscaling methods to date—we introduce and thoroughly analyze a posteriori random forests (AP-RFs), which use all the information contained in the leaves to reliably predict the shape and scale parameters of the gamma probability distribution of precipitation on wet days. Therefore, as opposed to traditional random forests, which typically provide deterministic predictions, our AP-RFs allow realistic stochastic precipitation samples to be generated for wet days. Indeed, as compared to one particular implementation of a generalized linear model that exhibited an overall good performance in VALUE, our AP-RFs yield better distributional similarity with observations without loss of predictive power. Noteworthy, the new methodology proposed in this paper has substantial potential for hydrologists and other impact communities which are in need of local-scale, reliable stochastic climate information.

Statistical downscaling methods aim to improve the limited spatial resolution of current climate models by linking a set of key large-scale predictor variables (e.g., geopotential, winds, etc.) to the predictand of interest (e.g., precipitation). Recently, the Experiment 1 of the COST action VALUE carried out the most comprehensive intercomparison of statistical downscaling methods to date. However, it lacked the inclusion of machine learning techniques, whose popularity has rapidly grown during the last years. Therefore, building on the same data and experimental framework used in VALUE, this work aims to partially fill this knowledge gap by introducing a modification of random forests—a well-known machine learning technique—for stochastic downscaling of precipitation at 86 European locations. As opposed to traditional random forests, which typically provide deterministic predictions, our proposed model predicts a probability distribution of precipitation for each predictors’ state. This is key to appropriately characterize the uncertainty of the downscaled predictions, allowing us to produce realistic samples of precipitation for wet days and to answer questions such as “What is the probability of getting more than 40 mm of precipitation today?,” relevant for many impact activities.