STAR-ESDM: High-Resolution Station- and Grid-Based Climate Projections

We introduce a set of empirical-statistical downscaled climate projections produced by the new Seasonal Trends and Analysis of Residuals Framework (STAR). The downscaling component, STAR-ESDM, uses signal processing techniques to decompose the temperature or precipitation time series into three separate components: (1) the long-term trend, (2) static and dynamic climatologies, and (3) static and dynamic diurnal anomalies. The ESDM then downscales global climate model output to the spatial and temporal resolution of any observational dataset; here, station-based NOAA GHCN observations across North America and 1/16th degree gridded observations (Livneh et al. 2013) covering the contiguous U.S. The ESDM is trained for each individual high-resolution grid cell or weather station with each component of the signal being individually bias-corrected. In the case of daily anomalies, they are transformed using a nonparametric Kernel Density Estimation function into a probability distribution that closely resembles historical observations initially but that changes over time as the GCM's distributions change, yielding the downscaled and bias corrected future projections for the location of interest, whether station or grid cell. Evaluating this new method using the perfect model framework shows that it significantly improves on previous errors and biases at the tails of the distribution where extreme events are relatively rare but have a proportionally greater impact on infrastructure, agriculture, human health, etc. while retaining a high computational efficiency. STAR-ESDM output is currently available for daily values of minimum and maximum temperature as well as daily precipitation for CMIP5 and CMIP6 simulations corresponding to a historical total-forcing scenario and a lower and higher future scenario (RCP4.5 and RCP8.5 and SSP2-4.5 and SSP5-8.5) for the period 1950-2100.