A Deep-Learning Based Framework for ESM Climate Downscaling and Its Application to the U.S. Northeast

Statistical downscaling is a computationally efficient approach in producing a large ensemble of fine-resolution future climate projections to support local climate adaptation efforts. Here we propose a deep-learning-based framework, named BC_DLD, to downscale the CMIP6 Earth System Models (ESMs) outputs from the coarse native resolution to a locally relevant fine-resolution at kilometer scale, and apply it to the Northeast United States, a region experiencing rapid increase of heavy precipitation and in great need for locally actionable climate information. Employing a two-phase multivariate approach by utilizing an Empirical-Quantile Delta Mapping (E-QDM) approach for bias correction at a coarse resolution and deep Convolutional Neural Networks (CNNs) for downscaling the corrected data to a finer spatial resolution, BC_DLD is designed to preserve the spatial characteristics from observations while retaining the temporal characteristics and climate change signals from ESMs. Here, we demonstrate the implementation of BC_DLD by downscaling the outputs of GFDL-ESM4 Historical run and Future SSP5-8.5 Scenario over the Northeast, from ~100 km to ~25 km (using ERA5 as reference) and then further down to ~6 km (using Livneh data as reference), and compare the downscaled data with an existing product, LOCA2. Despite using the same observational training data as LOCA2, BC_DLD produces more intense and rapid future increase of extreme precipitation than LOCA2, alleviating the underestimation of extremes common among statistical downscaling products. The emergent relationship between extreme precipitation intensity and temperature in BC_DLD more closely resembles the observations, further establishing its efficacy in producing useful and usable future climate projections.