• English 
  • Spanish 

Statistical downscaling with deep learning: A contribution to CORDEX-CORE

Machine learning is nowadays a very active research area in many disciplines and major breakthroughs have been recently obtained with deep convolutional neural networks in many complex problems. This is due to the ability of deep learning to efficiently treat high-dimensional spatiotemporal inputs extracting high-level feature representations with convolutional layers. Moreover, the technological advances boosted by data science applications provide efficient computational frameworks (e.g. TensorFlow) to transparently train these models on modern computing infrastructures using big datasets. As a result, deep learning provides an efficient alternative for statistical downscaling over wide domains, and some preliminary successful applications have been already reported. However,  the robustness and extrapolation capability of these models has yet to be tested for plausible applications in climate change problems.
In this work we analyze the potential of deep learning, and particularly convolutional neural networks, as suitable statistical downscaling techniques for climate change applications. In particular, we analyze cross-validation (using ERA-Interim predictors) and extrapolation (using GCM outputs) capabilities of different configurations of increasing complexity (starting from simple convolutional models) obtaining the best configurations to downscale precipitation and temperature. This validation study is first performed over the EURO-CORDEX domain, and it is then extended to other CORDEX domains to test transferability of the results. Finally, we describe the contribution to CORDEX-CORE and the resulting public dataset which is available to the downscaling community for intercomparison studies.