Configuration and Intercomparison of Deep Learning Neural Models for Statistical Downscaling

Abstract

Abstract. Deep learning techniques (in particular convolutional neural networks, CNNs) have recently emerged as a promising approach for statistical downscaling due to their ability to learn spatial features from huge spatio-temporal datasets. However, existing studies are based on complex models, applied to particular case studies and using simple validation frameworks, which makes difficult a proper assessment of the (possible) added value offered by these techniques. As a result, these models are usually seen as black-boxes generating distrust among the climate community, particularly in climate change problems. In this paper we undertake a comprehensive assessment of deep learning techniques for continental-scale statistical downscaling, building on the VALUE validation framework. In particular, different CNN models of increasing complexity are applied for downscaling temperature and precipitation over Europe, comparing them with a few standard benchmark methods from VALUE (linear and generalized linear models) which have been traditionally used for this purpose. Besides analyzing the adequacy of different components and topologies, we also focus on their extrapolation capability, a critical point for their possible application in climate change studies. To do this, we use a warm test period as surrogate of possible future climate conditions. Our results show that, whilst the added value of CNNs is mostly limited to the reproduction of extremes for temperature, these techniques do outperform the classic ones for the case of precipitation for most aspects considered. This overall good performance, together with the fact that they can be suitably applied to large regions (e.g. continents) without worrying about the spatial features being considered as predictors, can foster the use of statistical approaches in international initiatives such as CORDEX.

Bias correction for spatially interpolated daily mean air temperature during winter in eastern Hokkaido using multimodal machine learning

“…A limited amount of machine learning studies has been reported in terms of downscaling of air temperatures Baño-Medina et al, 2019; Sachindra and Kanae, 2019 , whereas considerable success has been achieved in downscaling of precipitation e.g., Wilby et al, 1998;Vandal et al, 2017;Misra et al, 2018;Baño-Medina et al, 2019;Pan et al, 2019 . Gaps between model formulas and actual events of rainfall processes may be greater than those in temperature, and hence, there may be still room for the contribution of machine learning.…”

“…Another important moiety to be addressed is spatial estimation, namely, downscaling. Several studies have attempted to apply these techniques to meteorological downscaling and demonstrated their effectiveness in terms of air temperatures Baño-Medina et al, 2019;Sachindra andKanae, 2019 , precipitation e.g., Wilby et al, 1998;Vandal et al, 2017;Misra et al, 2018;Baño-Medina et al, 2019;Pan et al, 2019 , andwind speeds Li, 2019 . Most of these earlier studies focused on obtaining variables at a grid spacing of approximately 10 -50 km by downscaling large-scale variables simulated by general circulation models.…”

Spatio-Temporal Downscaling of Climate Data Using Convolutional and Error-Predicting Neural Networks

“…In climate science, deep learning has recently been applied to a number of different problems, including microphysics (Seifert and Rasp, 2020), radiative transfer (Min et al, 2020), convection (O'Gorman and Dwyer, 2018), forecasting (Roesch and Günther, 2019;Selbesoglu, 2019;Weyn et al, 2019), and empirical-statistical downscaling (Baño-Medina et al, 2020). For example, Yuval et al (2021) have applied deep learning for parametrization of subgrid scale atmospheric processes like convection.…”

Statistical downscaling with the downscaleR package (v3.1.0): contribution to the VALUE intercomparison experiment

“…Notable efforts have been made in order to assess the credibility of regional climate change scenarios. In the particular case of SDS, a plethora of methods exists nowadays, and a thorough assessment of their intrinsic merits and limitations is required to guide practitioners and decision makers with credible climate information (Barsugli et al, 2013). In response to this challenge, the COST Action VALUE (Maraun et al, 2015) is an open collaboration that has established a European network to develop and validate downscaling methods, fostering collaboration and knowledge exchange between dispersed research communities and groups, with the engagement of relevant stakeholders (Rössler et al, 2019).…”