Hydro-Dataset: A Multi-Modal Open-Source Pipeline for Nordic Hydrological Research and Foundation Model Benchmarking

Getnet Demil, Md Moinul Islam, Mourad OussalaharXiv 2026 · arXiv

TL;DR

Researchers have released Hydro-Dataset v1.1, an open-source pipeline and dataset that transforms raw meteorological station data from Sweden into 1.8 million "ML-ready" station-days. By providing high-quality, observation-anchored labels for snow water equivalent (SWE) and five other key parameters, this work provides the foundational infrastructure needed to train and benchmark the next generation of hydrological AI foundation models in sub-arctic environments.

What Problem Does This Paper Address?

In the sub-arctic Nordic region, snowpack dynamics govern over 70% of the annual water cycle, making accurate snow monitoring vital for hydropower, flood management, and climate modeling. However, researchers face a "ground truth" crisis.

Current satellite and reanalysis products (like ERA5-Land) provide broad coverage but are notorious for systematically overestimating Snow Water Equivalent (SWE) in boreal forests. While dense station networks exist in countries like Sweden and Finland, their data isn't "plug-and-play" for machine learning. It requires extensive cleaning, harmonization, and the derivation of secondary variables (like melt rates or freeze-thaw cycles) that are critical for training robust AI models. This paper closes that gap by providing a reproducible pipeline that turns raw sensor readings into a standardized, multi-parameter benchmark.

Methods

The authors developed a strict four-phase Python-based pipeline: Extraction → Interim → Processing → Derivation.

Fig. 1. Hydro-Dataset v1.1 pipeline: Extraction (SMHI/FMI APIs) →Interim storage →Processing (alignment, QC) → Derivation (SWE, phenology, melt, ROS, FT, cold content).

Key technical choices include:

  • SWE Derivation: Instead of using generic models, they adopted the Sturm et al. (2010) "taiga-class" density model, which is specifically calibrated for the cold continental boreal conditions found in the Nordics.
  • Spatial Interpolation: They used Ordinary Kriging to map point-station data onto a 2.5 km grid. Unlike simpler methods, Kriging provides a variance field that explicitly quantifies how uncertain the interpolation is at any given point.
  • Multi-Task Labels: The pipeline automatically calculates six derived parameters, including snow phenology (onset/melt-out dates), melt/accumulation rates, and "Rain-on-Snow" (ROS) event flags—rare but high-impact events that are crucial for flood forecasting.

Datasets

The primary output is a CF-1.9 compliant NetCDF product containing:

  • Source Data: 582 stations from the Swedish Meteorological and Hydrological Institute (SMHI) spanning nine winters (2015–2024).
  • Scale: 1.82 million annotated station-days.
  • Validation Set: ERA5-Land reanalysis data was used as a comparison baseline to highlight the necessity of station-anchored data.
Table 1. Dataset variable characterization (582 SMHI sta- tions, 2015/16–2023/24, 1.82 M station-days). Tier 0: raw SMHI observations; Tier 1: derived parameters. Key values and ranges are 9-season means or min–max across qualifying stations.

Key Insights and Ideas

  • Reanalysis is not Ground Truth: The study confirms that ERA5-Land overestimates SWE by a staggering 262 mm on average in this region. Training AI models on reanalysis data essentially "bakes in" these systematic errors.
  • The "Taiga" Difference: Snow density dynamics in the Nordic boreal forest differ significantly from the Alpine regions where many standard models are calibrated. Using a region-specific density model is non-negotiable for accuracy.
  • Foundation Model Readiness: By co-registering six different snow parameters in a single file, the authors eliminate the "alignment overhead" that usually plagues multi-modal research, making it much easier to experiment with multi-task learning.

Key Results

The dataset successfully captures the high interannual variability of the region. For instance, the 2017-18 season was the heaviest (+36.8 mm SWE anomaly), while 2019-20 was the lightest.

The cross-validation against ERA5-Land (shown below) reveals that while both products track the same seasonal timing (Pearson $r=0.78$), they diverge wildly in magnitude. The negative KGE score for ERA5-Land indicates that the reanalysis product is a poor substitute for actual station observations.

Table 3. SWE validation: Sturm (2010, taiga) vs. ERA5- Land, Dalarna 2023-24 (N=2,770).

Limitations and Future Work

  • Geographic Scope: Version 1.1 is limited to Sweden. However, the authors have a clear roadmap (the "AquaNord" project) to integrate Finnish (FMI), Norwegian (met.no), and Danish (DMI) stations in upcoming releases.
  • Model Assumptions: The "taiga" density class is currently applied uniformly. Future versions may need to use "maritime" classes for coastal areas where snow behaves differently.
  • Validation: While validated against reanalysis, the authors plan further validation against manual "snow course" surveys (physical measurements taken along a path) to further refine the SWE estimates.

Summary updated

Abstract

Despite world-class snow monitoring by the Swedish Meteorological and Hydrological Institute (SMHI) and Finnish Meteorological Institute (FMI), no station-anchored, multi-parameter snow benchmark dataset exists for the Nordic boreal domain. We present Hydro-Dataset v1.1, a reproducible open-source pipeline that harmonizes 582 SMHI meteorological observation stations across nine hydrological winters (2015/16-2023/24) into a CF-1.9 NetCDF product with 1.8~million annotated station-days. Six derived snow parameters, such as Snow Water Equivalent (SWE), snow phenology, melt rates, degree-day melt factor, rain-on-snow flags, and freeze-thaw cycles, supply benchmark-ready labels for hydrological foundation model training. ERA5-Land cross-validation (Pearson R=0.78, bias=-262mm) confirms observation-anchored SWE provides a more reliable benchmark than reanalysis for machine learning training.
BibTeX
@misc{demil2026hydro,
  author = {Getnet Demil and Md Moinul Islam and Mourad Oussalah},
  title = {{Hydro-Dataset: A Multi-Modal Open-Source Pipeline for Nordic Hydrological Research and Foundation Model Benchmarking}},
  year = {2026},
  eprint = {2611.10992},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10992},
}
Figure