Advances for statistical downscaling of climate change precipitation scenarios based on machine learning techniques

Résumé

Even though they are the main tool to study climate change, global climate models (GCMs) still have a limited spatial resolution and exhibit considerable systematic errors with respect to the observed climate. Statistical downscaling aims to solve this issue by learning empirical relationships between large-scale variables, well reproduced by GCMs (such as synoptic winds or geopotential), and local observations of the target surface variable, such as precipitation, the focus of this thesis. We propose a series of novel developments which allow for improving the consistency of the downscaled fields and producing plausible local-to-regional climate change scenarios. The results of this thesis have important implications for the different sectors in need of reliable precipitation information to undertake their impact assessments.