HYDRO-DATASET: A MULTI-MODAL OPEN-SOURCE PIPELINE FOR NORDIC HYDROLOGICAL RESEARCH AND FOUNDATION MODEL BENCHMARKING

Hydro-Dataset v1.1 is a research-grade, machine-learning-ready benchmark of Nordic snow hydrology, harmonising 582 SMHI MetObs station time series across nine hydrological winters (2015–2024) into a single self-contained CF-1.9 NetCDF product with 1.8 million annotated station-days. Contents The archive contains two NetCDF files and companion CSVs: nordic_tier0_2015_2024.nc (58 MB) - harmonised raw observations: snow depth, air temperature, precipitation, and discharge at daily resolution, reprojected to ETRS89/LAEA (EPSG:3035). nordic_tier1_2015_2024.nc (133 MB) - Tier 0 plus six derived snow parameters co-registered at each station: Snow Water Equivalent (Sturm et al. 2010 taiga density model), snow phenology (onset/melt-out DOY, duration, peak SWE day), daily melt and accumulation rates, degree-day melt factor, rain-on-snow binary flag, and freeze-thaw cycle count. benchmark_snow_params.csv - 9-season interannual statistics (peak SWE, DDM, ROS frequency) per station. era5_validation_metrics.csv - ERA5-Land cross-validation skill scores (RMSE, Bias, Pearson R, KGE) per station. Provenance Raw observations are sourced from the SMHI Open Data API (opendata.smhi.se) under CC-BY. ERA5-Land SWE is used solely for cross-validation. The full pipeline (extraction → processing → derivation) is openly available at the companion code repository (MIT licence) and fully reproducible via make enrich-netcdf. Intended use The dataset is designed as training and benchmarking infrastructure for Hydrological Foundation Models, multi-task snow process learning, and interannual SWE variability analysis in the Nordic boreal domain. All files are readable with xarray.open_dataset() with no additional dependencies. Cross-validation ERA5-Land SWE cross-validation at snow-present station-days yields Pearson R = 0.78, Bias = −262 mm (ERA5-Land systematically overestimates SWE in Scandinavian boreal terrain), confirming station-anchored observations are more credible ground truth than reanalysis for ML training.