Towards improving the state of the art in statistical downscaling of regional precipitation and their extremes with emerging developments in machine learning and data mining

Regional projections of precipitation extremes are important for water resources and flood hazards management. Global climate models often cannot predict regional precipitation or their extremes with the precision and accuracy required by stakeholders. A lack of understanding of precipitation physics, intrinsic space-time variability, and the fact that precipitation is not a state variable of the physical models contribute to the challenge. Dynamical downscaling, which may improve processes like orographic precipitation, may cause a cascade of uncertainties as results from one complex model are fed into another. While statistical downscaling appears attractive, the criticisms include lack of interpretability and an inability to generalize under non-stationarity conditions expected under climate change. This study explores whether recently developed methods in data mining and machine learning can help improve the state of the art for statistical downscaling. Specifically, emerging methods that attempt to leverage the sparsity of data or covariate relations, as well as control for model complexity are examined in the context of dimensionality reduction or regression for mean and extreme values. The suite of methodologies are being evaluated with three sets of metrics which consider prediction skills on test data, model or covariate parsimony, as well as interpretability. Preliminary results suggest that while improving the state of the art for statistical downscaling in climate may be improved significantly, the choice of the methods may need to be specific to the nature of the problem, desired insights, as well as data quantity and quality.