Dr. Michael Langguth – Downscale-Bench : A benchmark dataset for statistical downscaling of meteorological data

This semester, we are excited to announce that the seminar series will be hosted by our collaborative partner ECMWF at their buildings in Bonn and it will be streamed online (via zoom).CESOC continues the seminar series “My Research” this Summer term 2024 with

Dr. Michael Langguth

from the Institute for Advanced Simulation (IAS) at the Research Centre Jülich, talking on their work

“Downscale-Bench : A benchmark dataset for statistical downscaling of meteorological data”

when: on Tuesday, 04 June 2024 at 11:00 am (CEST)
It is open to any interested person within the CESOC research disciplines (any Earth system sciences, mathematics or computer science).
Please contact info[@]cesoc.net, if you would like to participate.
Full Schedule could be found here!

Abstract:

Statistical downscaling with deep neural networks has recently gained a lot of momentum in the meteorological community. While several studies show promising results, direct comparison between different approaches is often impeded due to a large variety in the related downscaling task (target quantity), the deployed dataset and the evaluation framework. Furthermore, the performance of deep neural networks for statistical downscaling is barely compared to classical downscaling methods that have been established in the meteorological domain over the last decades. To enable fair comparison between different downscaling techniques, we suggest DownscaleBench, a benchmark dataset based on the coarse-grained ERA5 and the high-resolved COSMO REA6-dataset. DownscaleBench provides a standardized set of predictors for three (deterministic) downscaling tasks, that are downscaling of 2m temperature, 100m wind and solar irradiance. The ready-to-use datasets are complemented by a set of baseline models (standard deep neural networks and a classical method) whose performance is accessed in a task-specific, but standardized evaluation framework.