Fast and accurate learned multiresolution dynamical downscaling for precipitation

Abstract

Abstract. This study develops a neural-network-based approach for emulating high-resolution modeled precipitation data with comparable statistical properties but at greatly reduced computational cost. The key idea is to use combination of low- and high-resolution simulations (that differ not only in spatial resolution but also in geospatial patterns) to train a neural network to map from the former to the latter. Specifically, we define two types of CNNs, one that stacks variables directly and one that encodes each variable before stacking, and we train each CNN type both with a conventional loss function, such as mean square error (MSE), and with a conditional generative adversarial network (CGAN), for a total of four CNN variants. We compare the four new CNN-derived high-resolution precipitation results with precipitation generated from original high-resolution simulations, a bilinear interpolater and the state-of-the-art CNN-based super-resolution (SR) technique. Results show that the SR technique produces results similar to those of the bilinear interpolator with smoother spatial and temporal distributions and smaller data variabilities and extremes than the original high-resolution simulations. While the new CNNs trained by MSE generate better results over some regions than the interpolator and SR technique do, their predictions are still biased from the original high-resolution simulations. The CNNs trained by CGAN generate more realistic and physically reasonable results, better capturing not only data variability in time and space but also extremes such as intense and long-lasting storms. The new proposed CNN-based downscaling approach can downscale precipitation from 50 to 12 km in 14 min for 30 years once the network is trained (training takes 4 h using 1 GPU), while the conventional dynamical downscaling would take 1 month using 600 CPU cores to generate simulations at the resolution of 12 km over the contiguous United States.

Regional climate model emulator based on deep learning: concept and first evaluation of a novel hybrid downscaling approach

“…In this context, the perfect model framework allows us to focus on the downscaling function of the RCM, specifically. This approach is similar to the "super-resolution downscaling" mentioned by Wang et al (2021) and deployed by Vandal et al (2017) using observational data in a empirical statistical downscaling framework.…”

“…In this context, the RCM-emulator proposed here is based on a different fully convolutional neural network architecture known as UNet (Ronneberger et al 2015). Wang et al (2021) propose a different emulator following a different strategy than ours as they train it using a low and high-resolution version of the same RCM and another type of neural network (namely CGAN).…”

Construction of surface air temperature over the Tibetan Plateau based on generative adversarial networks

“…The convolution operation can extract the relationship of adjacent pixels, which makes the convolution neural networks (CNNs) adaptable in the construction and even prediction for meteorological variables (Hofer et al ., 2015; Ham et al ., 2019; Kadow et al ., 2020; Kumar et al ., 2021; Wang et al ., 2021). The CNNs can learn the key spatial and temporal characteristics and obtain good results of construction and prediction.…”

“…Wang et al . (2021) compared the precipitation generated by different types of CNN networks with an interpolator. The result shows that the interpolator was less accurate than any of the CNNs.…”

Using Machine Learning to Cut the Cost of Dynamical Downscaling

“…Traditional statistical downscaling techniques include perfect prognosis (Schubert, 1998), model output statistics (MOS) with bias correction (Eden & Widmann, 2014), and stochastic weather generators (Gutiérrez et al, 2019;Wilks & Wilby, 1999). Compared to DD, the implementation of SD is fast and far less computationally intensive (Wang, Liu, et al, 2021). However, SD models can perform poorly under extrapolation to future climates as few methods account for nonstationary relationships between predictors and predictands under climate change (Hernanz et al, 2022;Hewitson et al, 2014;Lanzante et al, 2018;Salvi et al, 2016;Schoof, 2013), although there are some exceptions (Baño-Medina et al, 2022;Pichuka & Maity, 2018).…”