In recent years, a substantial number of rainfall streamflow datasets were released that follow the example of the popular CAMELS dataset . They cover Chile , Great Britain , Brazil , Australia , the upper Danube basin , France , Switzerland , Denmark and Germany . There is also a world-wide version of CAMELS, CARAVAN . These datasets bundle a range of data sources and harmonize them to a readily ingestible, common spatio-temporal data format. Besides meteorological variables, they contain additional static information such as land cover, soil and bedrock type and orographic features. While suitable for a range of hydrological modeling approaches, these publications specifically facilitated a surge in popularity of neural network models for rainfall streamflow modeling. For example, have shown that neural network models are particularly suited to learn from such multi-variate, large-scale data.

A common downside of the above-mentioned datasets however is that they aggregate (“lump”) each variable within a catchment to a single value. By doing so, all information about spatial variability is lost: A pattern of soil types might be reduced to the most prevalent one, or a range of different amounts of precipitation over a large area might be averaged to a single, unexpressive average value. This reduction of information is unnecessary and counter-intuitive, especially for large catchments or catchments with high spatial variability. The principle advantage of spatially resolved inputs is that they enable the model to capture spatial covariance among different variables, e.g. the interacting effects of soil sealing or steepness of terrain and a torrential rainfall. Physical models, still the standard model type in active operation, resolve their equations on such a grid for exactly this reason, but neural network training also benefit from vast amounts of data. Additionally, as each grid cell contains a complete, self-contained set of meteorological and ancillary variables, they can be processed independently. Recent advances in memory parallelism can leverage this property and make large-scale processing of such data without prior aggregation practically feasible: in , we show that a neural network model is capable of efficiently handling this large amount of data. However, parallelization is just as applicable for physical or conceptual modeling approaches. Recently, demonstrated this by using a combined neural network, conceptual and physical approach on spatially resolved data from Switzerland to significantly outperform a lumped baseline. report similar results with a neural network model trained on two snow-driven basins in Kazakhstan and the USA. Both studies highlight the importance of representing spatial detail. and discussed general advantages and applicability of semi-distributed models for future world-wide flood-prediction models at recent conferences, further testifying to the relevance of providing datasets suitable for this research direction.

The study area of the dataset covers 5 entire basins in central Europe, namely the upper reaches of the Danube (until Bratislava), Elbe, Oder, Rhine and Weser. It is contiguous, 570.592 km² large and spans 10 countries. The temporal coverage ranges from 1 January 1980 to 31 December 2024. We bundle 6 spatiotemporal (“dynamic”) meteorological features with 46 static (“ancillary”) features: 3 hydro-geological features, 16 land cover features, 19 soil features and 8 orographical features. We based our choice of which kind of dynamic and ancillary information to include on the work of and to allow for maximum comparability with recent hydrological literature in general and neural network-based literature in particular. The dataset consists of data derived from a variety of publicly available sources – no new data was recorded. Our contribution consists in collecting the data and harmonizing it to a common grid: We provide a comprehensive, peer-reviewed, publicly licensed dataset that integrates well with already existing datasets while supporting novel research questions regarding spatial heterogeneity. Figure  provides an overview of the study area, common grid and types of variables. The remainder of this section first describes each data modality. Original data sources are listed in Table , detailed lists of all dynamic and ancillary features that we derived to compile this dataset can be found in Tables  and . We then describe how the different spatial data sources were harmonized to a common grid. Along with the data, we release all scripts for processing the raw source data into the dataset. This allows users to both verify and adapt our data aggregation pipeline. We also provide an additional script that combines the dataset presented here with river discharge data, after manual download from the original provider, the Global Runoff Data Center (GRDC) (https://grdc.bafg.de/data/data_portal/, last access: 11 September 2025). In the preprocessing code linked below, we also show that our study area is covered densely and uniformly with river gauging stations. As there are much fewer stations in the lower Danube basin, we decided to only include the upper part in order to eliminate this source of sampling bias, which can be a major concern not only for data-driven approaches. The discharge time series come at daily resolution, which is the reason that we provide our temporal features in daily resolution as well. This data can serve as targets for end-to-end training in data-driven rainfall streamflow modeling, such as in our study .

The dataset is available at https://doi.org/10.4211/hs.d7f2cbb587ab4a75ac7987854e8f62ca, . Dynamic meteorological forcing data and static ancillary data are stored in two separate NetCDF4 files, “ancillary_pub.nc” and “dynamic_pub.nc”. This format allows for labeled coordinates such as latitude, longitude and date for convenient selection on spatial and temporal domains, respectively. All variables are named in a self-explanatory manner and we provide labeled metadata. See Tables  and for a detailed reference of the included variables.

The data was processed in several Python Jupyter Notebooks that can be found in this repository: https://gitlab.hhi.fraunhofer.de/vischer/spatial_streamflow_dataprep, last access: 11 September 2025. The code requires Python 3.11 and is licensed under the Clear BSD licence. Additional dependencies are specified in an Anaconda environment specification contained in the repository. The scripts are stand-alone and do not require further input parameters. Along with the code to process the data, we provide a script that loads the data, selects subsets and visualizes them. This can serve as a starting point for the user to interact with the data. Furthermore, we provide code to wrap all the data in a PyTorch Dataset class for further processing in a machine learning context. Since dense arrays are required for this, we provide an alternative format version of our features in the files “ancillary_paper.nc” and “dynamic_paper.nc”, where dimensions were transformed from latitude and longitude to a unique grid cell index. In this version, all features were standardnormalized in order to suit better the requirements of neural networks.

All data sources from which we obtained the original data have been widely used across various scientific fields for years, so we assume the original data to be valid. In order to technically validate our processing steps, we feature a testing script in our repository with extensive tests and visualizations of the compiled data. We also successfully employed this dataset in training a neural network model for rainfall streamflow modeling .

Combining a variety of data sources, we provide the first spatially resolved dataset for multivariate rainfall-streamflow modeling. It covers five entire river basins and it thus particularly suited for large-scale modeling of hydrological processes. With the publication of this dataset, we hope to stimulate further development of spatially resolved, high-resolution hydrological modeling beyond the scope of lumped catchments. Suitable for neural network models as well as conceptual and physical modeling approaches, we hope this dataset will facilitate model comparison and stimulate future development in the spatially resolved domain. Using the same spatial grid as ERA5 as well as daily resolution limits its expressiveness of small scale, e.g. convective events, where higher spatial and temporal resolution would be preferable. If users decide to spatially aggregate our data and want to use derived variables like hydrological or climate signatures as input for their models, they would have to manually compute them from the raw data contained in our dataset. This is especially the case for snow-related variables, as particularly the Southeast of our study area is dominated by snowmelt dynamics. Lastly, the dataset is of only limited use for training training models that are to be deployed world-wide. We focus on a contiguous area in central Europe, which means in turn that the dataset contains only a particular subset of all hydrological dynamics that can be observed.