Comparison of a novel machine learning approach with dynamical downscaling for Australian precipitation

Abstract

Abstract Dynamical downscaling (DD), and machine learning (ML) based techniques have been widely applied to downscale global climate models and reanalyses to a finer spatiotemporal scale, but the relative performance of these two methods remains unclear. We implement an ML regression approach using a multi-layer perceptron (MLP) with a novel loss function to downscale coarse-resolution precipitation from the Bureau of Meteorology Atmospheric high-resolution Regional Reanalysis for Australia from grids of 12–48 km to 5 km, using the Australia Gridded Climate Data observations as the target. A separate MLP is developed for each coarse grid to predict the fine grid values within it, by combining coarse-scale time-varying meteorological variables with fine-scale static surface properties as predictors. The resulting predictions (on out-of-sample test periods) are more accurate than DD in capturing the rainfall climatology, as well as the frequency distribution and spatiotemporal variability of daily precipitation, reducing biases in daily extremes by 15%–85% with 12 km prediction fields. When prediction fields are coarsened, the skill of the MLP decreases—at 24 km relative bias increases by ∼10%, and at 48 km it increases by another ∼4%—but skill remains comparable to or, for some metrics, much better than DD. These results show that ML-based downscaling benefits from higher-resolution driving data but can still improve on DD (and at far less computational cost) when downscaling from a global climate model grid of ∼50 km.