Pith. sign in

REVIEW 4 major objections 4 minor 21 references

Spatially Resolved Meteorological and Ancillary Data in Central Europe for Rainfall Streamflow Modeling

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Five central European river basins are now covered by a 9 km gridded dataset for rainfall-streamflow modeling.

desk verdict A useful gridded hydrology dataset for central Europe, but a mislabeled land-cover class and missing data links need fixing before it is usable. read the letter →

arxiv 2506.03819 v1 pith:23RJ4MWH submitted 2025-06-04 stat.ML cs.LG

Marc Aurel Vischer , Noelia Otero , Jackie Ma This is my paper
classification stat.MLcs.LG
keywords spatiallyresolvedhydrologyrainfallstreamflowmodelingmachinelearningERA5-LandcentralEuroperiverbasinsgriddedmeteorologicaldatastaticcatchmentattributesneuralnetworks
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper compiles a fully spatially resolved dataset for rainfall-streamflow modeling covering the upper Danube, Elbe, Oder, Rhine, and Weser basins, and argues that such gridded inputs can move neural network hydrology beyond lumped catchment averages. Daily meteorological forcings (temperature, potential evapotranspiration, precipitation) from October 1981 to September 2011 are combined with 46 static ancillary features on soil, rock, land cover, and orography, all harmonized to a common $0.1^{\circ} \times 0.1^{\circ}$ (roughly 9 km) grid. The authors' central claim is that this ready-to-train, gridded format lets a model exploit spatial covariance among variables, such as how soil sealing or slope interacts with a torrential rainfall, rather than discarding that information in a catchment mean. A sympathetic reader would care because the dataset is positioned as the step between lumped large-sample hydrology datasets and the spatially distributed inputs of operational physics-based models.

What carries the argument

The load-bearing object is the regular $0.1^{\circ} \times 0.1^{\circ}$ ERA5-Land grid used as the common denominator for all sources. Each source is re-projected and sub-sampled at the nodes of this grid; categorical spatial data (rock productivity classes, land cover classes) are converted into distributional vectors or prevalence ratios per grid cell, while continuous data (soil properties, elevation) are averaged per cell, and daily curves of temperature, evapotranspiration, and precipitation are derived from three-hourly forcings. This machinery turns heterogeneous vector and raster inputs into one tidy two-dimensional grid that can be fed coordinate-wise to neural networks and can be extended to new variables, areas, or temporal resolutions.

What would settle it

Run a fixed neural architecture on this gridded dataset and on the same forcings lumped to catchment means across the five basins; if the gridded inputs consistently fail to match or outperform the lumped baseline on held-out discharge, the central motivation collapses, and a direct check of 9 km precipitation against a dense independent gauge network would reveal whether storm-scale structure survives sub-sampling.

Watch

Extended reading notes

Core claim

The paper's central discovery is a reusable, end-to-end data product: six dynamic meteorological features and 46 static ancillary features covering five central European basins, delivered in two NetCDF files on the ERA5-Land grid. The meteorological forcings are produced by aggregating three-hourly ERA5-Land fields to daily minima, means, maxima, sums, and variance; the static features are obtained by re-projecting vector and raster sources and sampling them at grid nodes, encoding hydrogeological productivity, land cover class proportions, soil property averages, and orographic elevation and gradient statistics. The authors claim that because every grid cell carries a complete set of dynamic and static variables, the locations can be processed independently and in parallel by neural networks, making spatially explicit rainfall-streamflow modeling inside large basins practically feasible. They also provide code that combines the dataset with river discharge data from the original provider, so the package supports end-to-end training of data-driven rainfall-streamflow models.

Load-bearing premise

The original data sources are valid, and re-projecting and sub-sampling them onto the 9 km grid preserves hydrologically relevant spatial variability.

Editorial extensions

If this is right

  • Rainfall-streamflow models can be trained end-to-end on spatially resolved daily forcings and static attributes at roughly 9 km resolution, without reducing each catchment to a single lumped value.
  • Models can in principle capture spatial covariance among variables, such as the interaction of bedrock productivity, land cover, and local rainfall intensity, inside large basins.
  • Because the grid matches the ERA5-Land grid, users can extend the dataset with extra meteorological variables, larger areas, or finer temporal resolution without changing its spatial skeleton.
  • The contiguous 570,581 square kilometer study area with dense stream gauge coverage supports benchmark comparisons of gridded neural models across contrasting central European hydroclimates.
  • The released processing scripts and discharge-combining code make the pipeline reproducible and applicable to new basins.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the same modular pipeline could be extended to other basins or to newer reanalysis products, turning this five-basin resource into a template for gridded continental-scale hydrology benchmarks.
  • A direct ablation the data invite but the paper does not perform is whether the 46 static ancillary features add predictive signal beyond the dynamic forcings once inputs are already spatially resolved.
  • Because precipitation is aggregated daily to a sum and a variance, the dataset captures rain intensity only coarsely; coupling the grid with sub-daily precipitation would test whether the approach extends to flash-flood-prone catchments.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This manuscript describes a new gridded hydro-meteorological dataset for five Central European river basins (upper Danube, Elbe, Oder, Rhine, Weser). The dataset combines daily meteorological forcings from ERA5-Land with static ancillary features derived from IHME hydrogeology, Corine Land Cover, European Soil Database Derived Data, and EU-DEM, all harmonized to the ERA5-Land 0.1° grid. The authors argue that this is the first rainfall-streamflow dataset for the region that avoids catchment-level lumping and provides spatially resolved inputs for neural network models. The paper also states that processing scripts, testing code, and a PyTorch wrapper are provided.

Significance. If the dataset is correct and accessible, it would fill a practical gap in ML-driven hydrology: existing CAMELS-style datasets for Europe aggregate static and dynamic attributes to the catchment level, whereas this product supplies per-grid-cell inputs that can be used to train spatially distributed neural network models. The authors are careful to separate provenance from processing, and they explicitly note that no new data were recorded and that no parameters were fitted, so the central utility claim is not entangled with any fitted model. The stated strengths are reproducible processing scripts, a testing suite, and a PyTorch interface; these are appropriate for a data descriptor and should be retained. However, the paper currently contains several unresolved issues that affect the artifact itself: an inconsistency in the temporal coverage, an apparent mislabeling of a static land-cover feature, and the absence of persistent data/code identifiers. These issues must be corrected before the dataset can be used as described.

major comments (4)
  1. [Abstract and Methods] The abstract states that the daily values span from October 1981 to September 2011, while the Methods section states that the temporal coverage is from 1 October 1980 to 30 September 2011, i.e., the 31 water years 1981 to 2011. This is a one-year discrepancy in the delivered temporal extent, and it directly affects the number of training samples and any evaluation split a user might construct. The authors must decide which interval is actually encoded in the NetCDF files and make the abstract, Methods, and any metadata consistent.
  2. [Table 3, row 17] The feature is labeled CLC_51_Water_bodies_Marine_waters, but in the Corine Land Cover nomenclature marine waters are class 52, not class 51. If the extraction used class 52, the label is wrong and misleads any user who relies on the documented names. If the extraction used class 51, the feature duplicates inland waters and marine waters are absent, producing an incorrect 16-channel land-cover vector for every grid cell. Either way, the static ancillary data cannot be trusted without fixing this label and verifying the underlying class code against the delivered data. The manuscript currently provides no data link or code link that would let a reader resolve this ambiguity.
  3. [Data Records / Code availability] The manuscript says the data are stored in a HydroShare repository and that the processing scripts can be found 'here', but no URLs, DOIs, or repository identifiers are provided. For a dataset paper, the data and code are not supplementary materials; they are the primary artifact under review. Without persistent identifiers, neither the reviewers nor future users can inspect the actual NetCDF variable names, check the CLC class issue raised above, or verify that the processing scripts match the described workflow. The authors should provide complete repository links and, ideally, a DOI for the dataset before this paper can be accepted.
  4. [Technical Validation] The Technical Validation section states that all original data sources are widely used and therefore assumed valid, but it does not include any quantitative check of the harmonization step itself. The central contribution is the claim that re-projecting and sub-sampling these sources onto the 9 km ERA5-Land grid produces a reliable spatially resolved product. The paper would be substantially stronger if it reported, for example, a comparison of basin-averaged precipitation or temperature against independent station-based data, or a check that land-cover and soil feature values reproduce known spatial patterns at a few sample locations. Without such a check, the reader has no evidence that the spatial harmonization preserves hydrologically relevant variability.
minor comments (4)
  1. [Introduction] There are several typos, including 'In recent years years' and 'from the the original provider'; these should be corrected.
  2. [Methods] The description says the data were harmonized by 're-projecting and sub-sampling at the nodes of this grid', but for land cover and soil the text also describes averaging or computing proportions within each grid cell. Please clarify whether the procedure uses point sampling at nodes or zonal aggregation over grid cells, because the two operations can produce different values.
  3. [Usage Notes] The statement that the scripts are 'stand-alone and do not require further input parameters' is confusing, because the accompanying discharge script requires a manual download from the original provider. Please clarify which scripts are fully automatic and which depend on external downloads.
  4. [Abstract and throughout] The phrase 'fully spatially resolved' is used repeatedly, but the product is on a 0.1° grid with many static features derived from coarser or vector sources. I suggest replacing 'fully spatially resolved' with 'gridded at 9 km resolution' or defining exactly what 'fully' means in relation to the source data resolutions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the dataset paper compiles public data with no fitted parameters or derived predictions, so the central claim is fully self-contained.

full rationale

The paper's central claim is the release of a spatially resolved dataset compiled by reprojecting and subsampling public products (ERA5-Land, IHME, CLC, ESDAC, EU-DEM) onto a common 9 km grid. The Methods section states 'no new data was recorded' and the Technical Validation section says 'All 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.' There is no equation to derive, no parameter fitted to a target, and no prediction that is statistically forced by construction. The only internal validation statement, that the dataset was 'successfully employed ... in training a neural network model (under review),' is a reported use case rather than a derivation of the dataset's content. The possible CLC class-51/52 label ambiguity raised in the skeptical reading is a data-quality and provenance-consistency concern, not a circularity. Self-citations are absent except normal neutral references to CAMELS and Kratzert et al., which are not load-bearing. Thus score 0.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central claim is the existence and usability of a compiled dataset. No numerical model is fitted. The main unverified inputs are the validity of the five source data products and the fidelity of the grid harmonization, plus one undocumented hand-chosen encoding for hydrogeological productivity.

free parameters (1)
  • IHME productivity ordinal scores = not specified
    The paper averages the six IHME hydrogeological productivity classes within each grid cell (Methods, Ancillary Data). The ordinal numeric values assigned to these classes are not given, so the resulting feature depends on an arbitrary hand-chosen encoding that is not documented.
assumptions (2)
  • domain assumption All original data sources (ERA5-Land, IHME, CLC, ESDAC, EU-DEM) are valid and suitable for hydrology.
    Technical Validation states: 'All 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.'
  • domain assumption Re-projecting and sub-sampling all inputs at 0.1 degree grid nodes preserves the spatial information needed for streamflow modeling.
    Methods describe harmonization via re-projecting and sub-sampling at ERA5 grid nodes without a validation against an external reference; if sub-grid variability is lost, the 'fully spatially resolved' claim is weakened.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatially Resolved Meteorological and Ancillary Data in Central Europe for Rainfall Streamflow Modeling." pith.science (2026). https://pith.science/paper/23RJ4MWH

@misc{pith2026250603819,
  author       = {Pith},
  title        = {Pith review of: Spatially Resolved Meteorological and Ancillary Data in Central Europe for Rainfall Streamflow Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23RJ4MWH}},
  note         = {Machine review of arXiv:2506.03819}
}
read the original abstract

We present a dataset for rainfall streamflow modeling that is fully spatially resolved with the aim of taking neural network-driven hydrological modeling beyond lumped catchments. To this end, we compiled data covering five river basins in central Europe: upper Danube, Elbe, Oder, Rhine, and Weser. The dataset contains meteorological forcings, as well as ancillary information on soil, rock, land cover, and orography. The data is harmonized to a regular 9km times 9km grid and contains daily values that span from October 1981 to September 2011. We also provide code to further combine our dataset with publicly available river discharge data for end-to-end rainfall streamflow modeling.

Figures

Figures reproduced from arXiv: 2506.03819 by the authors.

Figure 1
Figure 1. Overview of study area and visualizations for an example feature of each type. The basins are outlined in black, with the outline of Germany 1 in turquoise for geographic reference. The right hand panel shows additional river network connectivity information as yellow arrows that can be derived from the GRDC data with code from our repository. Meteorological Forcings The meteorological forcings in our study were der… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 11 canonical work pages

  1. [1]

    Newman, A. J. et al. Development of a large-sample watershed-scale hydrometeorological data set for the contiguous USA: Data set characteristics and assessment of regional variability in hydrologic model performance. Hydrol. Earth Syst. Sci. 19, 209–223, 10.5194/hess-19-209-2015 (2015)

    doi:10.5194/hess-19-209-2015 2015

  2. [2]

    J., Mizukami, N

    Addor, N., Newman, A. J., Mizukami, N. & Clark, M. P. The CAMELS data set: Catchment attributes and meteorology for large-sample studies. Hydrol. Earth Syst. Sci. 21 (2017)

    work page 2017

  3. [3]

    Alvarez-Garreton, C. et al. The CAMELS-CL dataset: Catchment attributes and meteorology for large sample studies – Chile dataset. Hydrol. Earth Syst. Sci. 22, 5817–5846, 10.5194/hess-22-5817-2018 (2018). 4Corine Land Cover Map, version 2012. Generated using European Union’s Copernicus Land Monitoring Service information; https://doi.org/10. 2909/916c0ee7-...

    doi:10.5194/hess-22-5817-2018 2018

  4. [4]

    Coxon, G. et al. CAMELS-GB: Hydrometeorological time series and landscape attributes for 671 catchments in Great Britain. Earth Syst. Sci. Data 12, 2459–2483, 10.5194/essd-12-2459-2020 (2020)

    doi:10.5194/essd-12-2459-2020 2020

  5. [5]

    Chagas, V . B. P.et al. CAMELS-BR: Hydrometeorological time series and landscape attributes for 897 catchments in Brazil. Earth Syst. Sci. Data 12, 2075–2096, 10.5194/essd-12-2075-2020 (2020)

    doi:10.5194/essd-12-2075-2020 2020

  6. [6]

    Fowler, K. J. A., Acharya, S. C., Addor, N., Chou, C. & Peel, M. C. CAMELS-AUS: Hydrometeorological time series and landscape attributes for 222 catchments in Australia. Earth Syst. Sci. Data 13, 3847–3867, 10.5194/essd-13-3847-2021 (2021)

    doi:10.5194/essd-13-3847-2021 2021

  7. [7]

    & Herrnegger, M

    Klingler, C., Schulz, K. & Herrnegger, M. LamaH-CE: LArge-SaMple DAta for Hydrology and Environmental Sciences for Central Europe. Earth Syst. Sci. Data 13, 4529–4565, 10.5194/essd-13-4529-2021 (2021)

    doi:10.5194/essd-13-4529-2021 2021

  8. [8]

    Delaigue, O. et al. CAMELS-FR: A large sample hydroclimatic dataset for France to explore hydrological diversity and support model benchmarking. Other, display (2022). 10.5194/iahs2022-521

    work page doi:10.5194/iahs2022-521 2022

Show all 21 references
  1. [9]

    Höge, M. et al. CAMELS-CH: Hydro-meteorological time series and landscape attributes for 331 catchments in hydrologic Switzerland. Earth Syst. Sci. Data 15, 5755–5784, 10.5194/essd-15-5755-2023 (2023)

    2023 doi

  2. [10]

    Liu, J. et al. CAMELS-DK: Hydrometeorological Time Series and Landscape Attributes for 3330 Catchments in Denmark. Earth Syst. Sci. Data Discuss. 1–30, 10.5194/essd-2024-292 (2024)

    2024 doi

  3. [11]

    Loritz, R. et al. CAMELS-DE: Hydro-meteorological time series and attributes for 1555 catchments in Germany. Earth Syst. Sci. Data Discuss. 1–30, 10.5194/essd-2024-318 (2024)

    2024 doi

  4. [12]

    Kratzert, F. et al. Toward Improved Predictions in Ungauged Basins: Exploiting the Power of Machine Learning. Water Resour. Res. 55, 11344–11354, 10.1029/2019WR026065 (2019)

    2019 doi

  5. [14]

    ERA5-Land hourly data from 1950 to present

    CopernicusClimateChangeService. ERA5-Land hourly data from 1950 to present. Copernicus Climate Change Service (C3S) Climate Data Store (CDS), DOI: 10.24381/cds.e2161bac (Accessed on 23-Oct-2021) (2022)

    2022 doi

  6. [15]

    & Duscher, K

    Günther, A. & Duscher, K. Extended vector data of the International Hydrogeological Map of Europe 1: 1,500,000 (Version IHME1500 v1. 2). Fed. Inst. for Geosci. Nat. Resour. (BGR), Hannover , Berlin, Ger

  7. [16]

    Mapping Soil Typologies: Spatial Decision Support Applied to the European Soil Database

    Hiederer, R. Mapping Soil Typologies: Spatial Decision Support Applied to the European Soil Database. (Publications Office of the European Union 127, 2013)

    2013

  8. [17]

    Mapping Soil Properties for Europe: Spatial Representation of Soil Database Attributes

    Hiederer, R. Mapping Soil Properties for Europe: Spatial Representation of Soil Database Attributes. (EUR26082EN scientific and technical research series 47, 2013)

    2013

  9. [18]

    & Caron, J

    Rew, R., Harnett, E. & Caron, J. NetCDF-4: Software implementing an enhanced data model for the geosciences. In 22nd International Conference on Interactive Information Processing Systems for Meteorology, Oceanograph, and Hydrology , vol. 6 (2006)

    2006

  10. [19]

    Ansel, J. et al. PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, V olume 2, 929–947, 10.11...

  11. [20]

    Granger, B. E. & Pérez, F. Jupyter: Thinking and Storytelling With Code and Data. Comput. Sci. & Eng. 23, 7–14, 10.1109/MCSE.2021.3059263 (2021)

    2021

  12. [21]

    & Drake Jr, F

    Van Rossum, G. & Drake Jr, F. L. Python 3 reference manual. Scotts V al. Creat.(2009)

    2009

  13. [22]

    Anaconda Documentation (2020)

    Anaconda Software Distribution. Anaconda Documentation (2020). Acknowledgements This work was supported by the Federal Ministry for Economic Affairs and Climate Action (BMWK) as grant DAKI-FWS (01MK21009A). Author contributions statement M.A.V . compiled the data with crucial ...

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.