• English 
  • Spanish 

On the suitability of convolutional neural networks for climate downscaling

Machine learning approaches are being increasingly used by the climate community due to their ability to efficiently treat the vast amount of data (e.g., satellites, in-situ observations, climate models) typically involved in meteorology/climate problems. For instance, deep learning algorithms have been recently applied for hurricane tracking, detection of extreme events and model parametrization emulators with promising results. In this context, and given their capacity to learn complex spatio-temporal patterns, convolutional neural networks (CNN) may provide an outstanding alternative to deal with other climate-related problems.
In this work, we analyze the suitability of CNN for statistical downscaling of temperature and precipitation. To do this, we frame our experiment under the VALUE intercomparison project, which provides the ideal scenario for a rigorous comparison of these techniques against standard (benchmark) methods which have been traditionally used for the same task (e.g., generalized linear models, analogs). To be as comprehensive as possible, we explore the adequacy of different network architectures and assess not only their performance to reproduce the observed local climate in present conditions, but also their potential suitability to generate robust, high-resolution climate change scenarios.
Our results show that CNN outperform the standard methods top ranked in VALUE, providing better spatial and temporal representation of the local climate (especially for precipitation) whilst leading to plausible —i.e., compatible with the coarse outputs given by the global climate models— regional-to-local climate change signals. Moreover, unlike traditional downscaling techniques, CNN can be efficiently applied to continental-sized domains. This could foster their comprehensive use in international downscaling initiatives such as CORDEX, which has mainly relied on computationally-expensive dynamical models to-date.