To highlight the performance of the DL-based method, the results were compared with the results obtained by a baseline method, quantile mapping.

The merged data show a strong average improvement in mean error (~ 47%), correlation coefficient (~ 37%), and root mean square error (~22%).

The proposed merging framework was applied to mismatches in precipitation and surface pressure representing more and less chaotic variables, respectively. The COSMO-REA6 reanalysis data were used as a reference. Here, the climate simulations over Europe from the Terrestrial Systems Modeling Platform (TSMP) were used as input to the CNN. This study focused on using UNet encoder–decoder convolutional neural networks (CNNs) for extracting spatiotemporal features from model simulations for predicting the actual mismatches (errors) between the simulation results and a reference data set. Machine learning (ML) and/or deep learning (DL) methods have been shown to be useful tools in closing the gap between models and observations due to the capacity in the representation of the non-linear space–time correlation structure.

Merging model-based data with observations yield promising results benefiting simultaneously from the information content of the model results and observations. However, due to model uncertainties arising from, e.g., initial value and model errors, the simulation results do not match the in situ or remotely sensed observations to arbitrary accuracy. Physically based numerical weather prediction and climate models provide useful information for a large number of end users, such as flood forecasters, water resource managers, and farmers.

Hidden deep g2a