GEMS-GER: A Machine Learning Benchmark Dataset of Long-Term Groundwater Levels in Germany with Meteorological Forcings and Site-Specific Environmental Features

This repository provides the dataset accompanying the GEMS-GER (Groundwater Levels, Environment, Meteorology, Site Properties – Germany) benchmark for machine learning-based groundwater modeling. The dataset includes long-term groundwater level time series, meteorological and hydrological forcing data, site-specific environmental properties, and benchmark model evaluation results. All data originate from official public sources and have been harmonized across the 16 German federal states (Bundesländer). Contents of this repository: Groundwater level time series (GEMS-GER_data/dynamic/MW_*.csv): Weekly aggregated groundwater levels (GWL) from 3,207 monitoring wells (1991–2022), including: Daily temperature (mean, min, max) Precipitation and humidity (HYRAS/DWD) Real, potential, and reference evapotranspiration Soil moisture and soil temperature (at 5 cm depth) Snow water equivalent, snowmelt, and runoff (ERA5-Land) GWL_flag indicating observed vs. imputed values Site-specific static descriptors (GEMS-GER_data/static/static_features.csv): Hydrogeology and soil type Land use and climate classification Elevation and derived topographic parameters (e.g. slope, TWI) Benchmark model performance (GEMS-GER_data/model_performance/) Median NSE, RMSE, R², and Bias across 10 runs for each of the 3,207 wells: model_performance_single.csv – Single-well 1D-CNN models (trained individually) model_performance_global_dynonly.csv – Global LSTM model (trained jointly on all wells) using dynamic features model_performance_global_dynstat.csv – Global LSTM model (trained jointly on all wells) using both dynamic and static features Pre-generated time series plots (GEMS-GER_data_figures/DYN_Feat_MW_*.pdf): Visualizations of groundwater levels and selected forcing variables for all wells Provided separately to reduce the size of the main dataset download Directory structure: GEMS-GER_data/├── dynamic/ # 3,207 CSV files containing time series per monitoring well (MW_{ID}.csv)│ ├── MW_1.csv│ ├── MW_2.csv│ └── ...│├── static/ # Static site-specific descriptors│ └── static_features.csv # (e.g., geology, land use, climate statistics)│├── model_performance/ # Model performance metrics (median across 10 runs)│ ├── model_performance_single.csv # Single-well CNN models│ ├── model_performance_global_dynonly.csv # Global LSTM (dynamic inputs only)│ └── model_performance_global_dynstat.csv # Global LSTM (dynamic + static inputs) GEMS-GER_figures/├── DYN_Feat_MW_1.pdf # Input time series visualization per well├── DYN_Feat_MW_2.pdf└── ... The dataset is intended for research and benchmarking in hydrogeology, data-driven groundwater modeling, and environmental machine learning. It forms the basis of the GEMS-GER benchmark, as described in the associated preprint. All data originate from public sources and have been harmonized across administrative and institutional boundaries to enable consistent large-scale analysis. All associated code, documentation, and update announcements are maintained in the project's GitHub repository: https://github.com/KITHydrogeology/GEMS-GER ensuring transparency, traceability, and reproducibility.