Testing deep learning methods for downscaling climate change projections: The DeepESD multi-model dataset

<p>Deep Learning (DL) has recently emerged as a powerful approach to downscale climate variables from low-resolution GCM fields, showing promising capabilities to reproduce the local scale in present conditions [1]. There have also been some prospects assessing the potential of DL techniques to downscale climate change projections, in particular convolutional neural networks (CNNs) [2]. However, it is still an open question whether they are able to properly generalize to climate change conditions which have been never seen before and produce plausible results.&#160;</p><p>Following the &#8220;perfect-prognosis&#8221; approach, we use in this study the CNNs assessed in [2] to downscale precipitation and temperature for the historical (1975-2005) and RCP8.5 (2006-2100) scenarios of&#160; an ensemble of eight Global Climate Models (GCMs) over Europe. The resulting future projections, which are gathered in a new dataset called DeepESD, are compared with 1) those derived from benchmark statistical models (linear and generalized linear models), and 2) a set of state-of-the-art regional climate models (RCM) which are considered the &#8220;ground-truth&#8221;. Overall, CNNs lead to climate change signals that are in good agreement with those obtained from RCMs (especially for precipitation), which indicates their potential ability to generalize to future climates. Nevertheless, for some GCMs we find&#160; that there are considerable regional differences between the &#8220;raw&#8221; and the downscaled climate change signals, an important aspect which was unnoticed in a previous work that focused exclusively on one single GCM [2]. This highlights the importance of considering&#160; muti-model ensembles of downscaled projections (such as the one presented here) to conduct a comprehensive analysis of the suitability of DL techniques for climate change applications. Indeed, understanding the nature of the mentioned differences is necessary and future work towards this aim would imply carefully analyzing some of the assumptions made in&#8220;perfect-prognosis&#8221; downscaling (e.g., stationarity of the predictor-predictand link, adaptation of the statistical function to the climate model space). Therefore, following the FAIR (Findability, Accessibility, Interoperability and Reuse) principles we have made publicly available DeepESD through the Earth System Grid Federation (ESGF), which will allow the scientific community to continue exploring the benefits and shortcomings of DL techniques for statistical downscaling of climate change projections.&#160;</p><p>References:</p><p>[1] Ba&#241;o-Medina, J., Manzanas, R., and Guti&#233;rrez, J. M.: Configuration and intercomparison of deep learning neural models for statistical downscaling, Geoscientific Model Development, 13, 2109&#8211;2124, 2020.</p><p>[2] Ba&#241;o-Medina, J., Manzanas, R., and Guti&#233;rrez, J. M.: On the suitability of deep convolutional neural networks for continental-wide downscaling of climate change projections, Climate Dynamics, pp. 1&#8211;11, 2021</p><p>&#160;</p><p>Acknowledgements</p><p>The authors would like to acknowledge projects ATLAS (PID2019-111481RB-I00) and CORDyS (PID2020-116595RB-I00), funded by MCIN/AEI (doi:10.13039/501100011033). We also acknowledge support from Universidad de Cantabria and Consejer&#237;a de Universidades, Igualdad, Cultura y Deporte del Gobierno de Cantabria via the &#8220;instrumentaci&#243;n y ciencia de datos para sondear la naturaleza del universo&#8221; project for funding this work. A.S.C and E.C. acknowledge project IS-ENES3 funded by the EU H2020 (#824084).</p><p><br><br></p>