A CNN-based Downscaling Method of C3S Seasonal Forecast: Temperature and Precipitation

Copernicus Climate Change Service provides seasonal forecasts for meteorological outlooks several months in advance and can provide indications of future climate risks on a global scale. Using downscaling techniques, global variables can be transferred to the high-resolution regional scale, allowing the information to be elaborated for extreme events detection and further implementing and coupling with hydrological models for regional hazard prediction, thus serving agriculture and energy, improving planning for tourism and other sectors.

In this study, we applied a new CNN-based architecture for temperature and precipitation downscaling. Both variables are downscaled from 1 degree to 1 arcminute to fulfill the requirements as an input to the hydrological models. The architecture implements an auto-encoder/decoder structure to extract the data relations. The system is trained with seasonal forecast inputs and observation data to establish the relation between both scales. The model is then evaluated with the validation period from the observation data to achieve the best performance, changing network structures and tuning different network hyper-parameters. The results show a good fit for the observation data on the monthly scale, providing enough details in the downscaling product. Finally, the best-performing networks for downscaling temperature and precipitation are selected and could be extended for further utilization.