Temporal Downscaling of ICON Precipitation from Hourly to 10‑Minute Resolution Using a Physically Constrained U-NET

 The availability of high temporal resolution precipitation data is essential for understanding sub‑hourly hydrometeorological processes, extreme rainfall, and their impacts on hydrology and urban flooding. Especially with respect to climate change where precipitation extremes are expected to enlarge a profound data base is needed as an ensemble of downscaled climate scenarios. To store meteorological fields with high resolution in time and space is very resource demanding. The standard EURO-CORDEX dataset includes hourly precipitation data. For impact modellers however it is important to get data for the extreme events with higher resolution in time. In this study, we present a deep‑learning‑based framework to temporally downscale hourly ICON precipitation to 10‑minute resolution using a convolutional U‑Net architecture.

The source data consist of two input images corresponding to 1-hour accumulated precipitation fields. The target data are 10-minute precipitation fields derived from ICON simulations. The model is trained and evaluated over the following periods: 1980–1994 for training, 1995–1997 for validation, and 1998–1999 for testing. The model learns a mapping from the source data to the corresponding sequences of 10-minute precipitation. The U‑Net is trained to reconstruct the temporal distribution of rainfall within each hour while conserving the total hourly precipitation amount. We test the enforcement of conservation of total hourly precipitation with different techniques: a penalty term in the loss function, a constraint layer embedded into the architecture and conservation through a post-processing routine.

Model performance is evaluated using multiple statistical metrics to assess both the distribution and magnitude of precipitation. The histograms of predicted and target 10‑minute precipitation indicate that the model reproduces the marginal distribution well, while the scatter plot of total predicted versus total target precipitation summed over all grid cells and time steps shows that the model closely preserves the overall accumulated rainfall. Results also demonstrate that the U‑Net with the conservation enforcing constraint layer successfully reproduces sub‑hourly precipitation variability and captures the timing and intensity of short‑duration rainfall events more accurately than simple temporal disaggregation approaches.

This work highlights the potential of machine learning for efficient temporal downscaling of regional climate model outputs. The ultimate goal is to provide a tool for impact modelers to produce high-resolution precipitation data on their own demand . This framework has the potential to support applications in future warming scenarios. Since interested researchers can run the temporal downscaling model for their period of interest, there is no need for large memory resources to store precipitation datasets with a very high temporal resolution.