High-resolution deep-learning and dynamical climate downscaling for impact modeling in Southeast South America

Abstract

This work introduces the experimental design and main results of a coordinated modeling study within the framework of a Flagship Pilot Study endorsed by the Coordinated Regional Climate Downscaling Experiment (CORDEX). The objective is to apply high-resolution climate downscaling for hydrological and agricultural modeling in Southeastern South America (SESA). To this end, targeted simulations using convection-permitting regional climate models (CPRCM at 4 km resolution) and deep learning-based empirical statistical downscaling were performed covering 3 consecutive years from June 2018 to May 2021. These simulations were used to drive the Variable Infiltration Capacity hydrologic model to simulate the streamflow of the Uruguay river and the Agricultural Crop Simulator to reproduce crop yields over Southern Brazil. An ensemble of six CPRCMs and a variety of statistical downscaling models based on convolutional neural networks (CNNs) contributed to this project. Overall, CPRCMs and CNNs show skill in reproducing daily precipitation over SESA, with different abilities in simulating the different aspects of precipitation extremes (dry and wet), although within the range of observational uncertainty. The intra-seasonal and inter-annual variability of extreme events and their frequency over the different sub-regions of SESA are very well captured by the simulations with most correlations between 0.63 and 0.88. This aspect also translates into the Uruguay River streamflow simulations (with time correlations above 0.42) whose response appears to be highly sensitive to precipitation intensity and location. The largest impacts on soybean yield simulation are related to low precipitation intensity and spatial variability, reaching to mean spatial biases up to -20%.