Machine learning assessment of hydrological model performance under localized water storage changes through downscaling

The coarse spatial resolution of the Gravity Recovery and Climate Experiment (GRACE) data has limited its application in the management of local-scale water resources. To address this limitation, we developed a new downscaling approach using predictors from regional and global hydrological models for a 15-year period (2002-2017) and tested it in the northern Great Artesian Basin, Australia. We used four different machine learning algorithms (support vector machine, partial least squares, gaussian process and random forest) to downscale the original GRACE estimate of 0.5 degrees to a spatial grain size of 0.1 degrees (global) and 0.05 degrees (regional). This was based on precipitation, evapotranspiration and runoff estimates from the Famine Early Warning Systems Network Land Data Assimilation System (FLDAS) and Australian Water Outlook (AWO) hydrological models, respectively. The downscaled products were validated using 42 in-situ precipitation observations spread across the test region. We further evaluated which of the downscaled products best mimicked local-scale hydrology using a range of statistical metrics. Our results showed that regional hydrological models best characterized the dynamics of local scale hydrology (rainfall v. downscaled product), and the gaussian process regression algorithm made the best predictions for both models. The correlation coefficients for the raw values varied from 0.45 to 0.49 while that of the standardized values varied from 0.46 to 0.52 with the random forest model providing the best fitting for the regional-based products. The regional downscaling approach employed in this study may be readily integrated into local water resources planning programs.