GRIT-ADB: A Global Attribute Database for the GRIT Hydrography

Zhang, Boen; Wortmann, Michel; Liu, Yinxue; Kratzert, Frederik; Moulds, Simon; Slater, Louise

doi:10.5194/essd-2026-279

Preprints

https://doi.org/10.5194/essd-2026-279

Preprints

26 Jun 2026

| 26 Jun 2026

Status: this preprint is currently under review for the journal ESSD.

GRIT-ADB: A Global Attribute Database for the GRIT Hydrography

Boen Zhang, Michel Wortmann, Yinxue Liu, Frederik Kratzert, Simon Moulds, and Louise Slater

Abstract. Global hydro-environmental databases provide essential information for large-scale hydrological, ecological, geomorphological, and Earth system analyses. Most existing global databases are built upon convergent river representations that do not explicitly capture bifurcating, multi-channel river systems. In addition, these databases primarily characterise long-term climatological means or static representations of environmental conditions derived from earlier-generation global datasets, limiting their applicability for time-varying analyses of hydroclimatic and geomorphological processes. Here we present GRIT-ADB, a new attribute database for the vectorised Global River Topology (GRIT), a new river hydrography dataset created from a 30 m resolution river mask and terrain data that provides a topology-explicit and physically realistic representation of river networks including divergent flow pathways. GRIT-ADB provides standardised hydro-environmental information for 19.6 million km of rivers and streams. It currently comprises 64 time-varying (multi-dimensional embeddings) and 35 static variables (>300 attributes), spanning five categories: hydrology, physiography, climate, land cover and use, and soils and geology. Attributes are derived by aggregating and harmonising data from state-of-the-art global datasets and are accumulated along the river network from headwaters to basin outlets, while preserving the topology of divergent and complex flow pathways. The attributes are linked to multiple GRIT scales, including hierarchically-nested subbasins, individual river reaches of up to 1 km long, and coarser-scale river segments of several kilometres long, providing a flexible framework that can accommodate future extensions of GRIT and additional attributes. By combining a standardised attribute framework with explicit representation of bifurcating river hydrography, GRIT-ADB enables improved large-scale yet high-resolution analyses of river connectivity, hydrological extremes, hydro-ecological processes, and climate impacts in complex river systems, supporting a wide range of global hydrological and environmental applications.

Received: 13 Apr 2026 – Discussion started: 26 Jun 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 15937 KB)

Supplement (3183 KB)

Download & links

Preprint (15937 KB)
Metadata XML
Supplement (3183 KB)
BibTeX
EndNote

Boen Zhang, Michel Wortmann, Yinxue Liu, Frederik Kratzert, Simon Moulds, and Louise Slater

Status: final response (author comments only)

RC1:
'Comment on essd-2026-279', Anonymous Referee #1, 13 Jul 2026
Zhang and colleagues provide an extremely exciting and useful new dataset for the large-sample hydrology community. Overall the paper is well-written, structured and following a nice flow, which is greatly appreciated from a reviewer’s perspective. I do have some minor comments that should be addressed/answered before the manuscript could be accepted for publication.
What do the authors mean by “>300 attributes”? This expression is also used elsewhere in the text (e.g., L211 and L9–10). Does it mean that summing all the attributes listed in Table 1 results in more than 300 attributes? Is this because the dataset provides yearly values for some of them? Do the authors also provide the 12 monthly mean values, as is done in HydroATLAS? I feel that this is still somewhat unclear based on the current wording. A table in the Appendix or Supporting Information, similar to those provided for Caravan (Kratzert et al., Tables 1 to 5), for example, would make this easier for readers to understand.

Does ADB stand for something? Is it an abbreviation, or is it simply the name of the product?

Is the upstream accumulated value provided for each reach or for each segment?

L52: This may be a matter of preference, but I feel that the text would flow more naturally if a connector were added here, for example, “Therefore, to address…”. This impression may arise because “To address this” has already been used in L46.

Table 1: Are the authors citing the underlying dataset, the aggregation methodology, or both? Woods (2009) introduces the methodology for calculating precipitation seasonality, correct? I appreciate and agree with citing the methodology used, but it would also be useful to cite the underlying dataset.

L60: GRIT has already been defined. The authors could simply use “GRIT hydrography” here.

L101: Has this point not already been mentioned earlier in the same paragraph? I am not necessarily against the repetition, but, as currently written, it sounds like new information and confused me briefly. Could the authors rephrase the sentence if this is indeed a repetition? The sentence I am referring to is: “The accumulation statistic depends on the…” For example: “The details of the accumulation statistic, which depend on the variable, are…”. This is only a preliminary wording suggestion, so please feel free to find a more suitable formulation.

Figure 2: Given that the labels “a” and “b” are used elsewhere in the manuscript, please consider adding them to the corresponding panels.

Table 1: Could the authors add a description of each attribute in the Appendix or Supplementary Material? While reading Section 3.1, I started to wonder what exactly the authors mean by “sinuosity” and “stream gradient” — “river width” is completely clear, of course. This made me think that a table containing definitions and descriptions of the attributes would be very useful.

Such a table could also help clarify the citation of Woods (2009). For example, the authors could write: “Precipitation seasonality, representing XX, as defined by Woods (2009).” If the table were included in the Appendix, the citation would remain part of the paper and would appropriately acknowledge the original methodological contribution.

Another example is the aridity index. What definition do the authors use? Depending on the study, this may refer to P/PET, PET/P, or even normalized indices with values ranging from −1 to 1. A table containing explicit definitions would make these attributes much clearer.

L129: When the authors describe the values as heterogeneous, is this inferred visually from the figure, or did they also calculate a coefficient of variation or standard deviation? If the statement is based on the figure, perhaps add “(Figure 4a)” after “heterogeneous.” Although the claim may appear evident, this would make the basis for the statement clearer.

Figure 4: I am reviewing the manuscript using a printed copy, and in this format it is very difficult to distinguish catchments with values above 2%. Figure 2 uses a similar colour gradient for the variables but is considerably easier to read. Could the authors perhaps try using the same background colour as in Figure 2 and assess whether this improves readability?

I also have a question about the legend in Figure 4. Do the values range only from 0 to 2 because they have been normalized, or do they actually exceed 2? If values above 2 are included in the highest class, should the label not be “>2”? Apologies for being overly pedantic. The same comment applies to panel b.

Still regarding Figure 4, given that the panels are labelled “a” and “b,” the caption could simply refer to these letters, or use “Panel a” and “Panel b,” rather than “top panel” and “bottom panel.”

Section 3.3: Do the authors happen to have a global map showing these regions? Such a map could potentially be included in the Appendix or Supporting Information.

L144: While I agree that there are no clearly visible major differences in panel c, I have the impression that soil sand content differs considerably among the zones. For example, there appears to be, on average, more sand in arid areas than in the other zones, while the spread in the polar regions appears particularly large.

Perhaps Table A1 could be expanded to serve as the table describing each variable.

Figure 6: Out of curiosity, do the authors have any explanation for the concentrated area in the central-eastern part of the map, which appears particularly red in panel m? This does not necessarily need to be addressed in the manuscript, but it is a pattern that may also have caught the authors’ attention.

Table 2: Do the authors also have information on the upstream drainage-area threshold used for the other datasets?

What ordering is used for the variables in Table A1? Could the authors use the same order as in Table 1 to facilitate a more seamless comparison?
Citation: https://doi.org/10.5194/essd-2026-279-RC1
RC2:
'Comment on essd-2026-279', Anonymous Referee #2, 06 Aug 2026
General comments
GRIT-ADB is a useful and timely contribution. The community has needed an attribute layer on top of GRIT since the hydrography was published, and building it as a general, extensible framework is the right design decision. The scope of harmonised source data is respectable, the multi-scale structure (reach / segment / subbasin, local / upstream) is sensible, and the inclusion of AlphaEarth embeddings is forward-looking. I expect this dataset to be useful to the community.
The most immediate problem is that the global dataset described by the manuscript is not available for review. Section 6 provides only a Congo River basin example and states that the full product will be deposited before final publication. This prevents assessment of global completeness, internal consistency, regional quality variation, and practical usability. The ESSD data policy requires the data described by a submission to be accessible during review, and the journal asks reviewers to assess the dataset itself, not only the manuscript.
Beyond accessibility, there is no dedicated validation or quantified quality assessment of the derived product. The Discussion acknowledges that biases in the source datasets will propagate into GRIT-ADB, but provides no uncertainty characterisation, independent cross-check, internal consistency assessment, or sensitivity analysis on consequential processing choices. Figure 7 shows that GRIT-ADB differs from HydroATLAS; it does not show that GRIT-ADB is closer to reality.
Second, the manuscript repeatedly claims that bifurcation-aware accumulation materially improves attribute estimates, but never quantifies the magnitude or spatial distribution of the effect. This is the claimed advantage of the product and can be tested using the GRIT network and source data already available to the authors.
I therefore recommend major revision and re-review once the complete dataset is accessible and a substantive quality assessment has been added.
Major comments
1. Make the complete global dataset available for review
The Data Availability section provides only a Congo River basin example and promises the global product before final publication. That is not sufficient for a data-description paper whose subject is a global database. The complete release should be made accessible through a functional DOI or an anonymous repository review link before the review can be completed.
2. Add a data-quality, validation, and uncertainty assessment
The qualitative statement that source-data biases propagate into GRIT-ADB is not an assessment of the derived product. The ESSD criteria for data-description articles call for validation against independent reference data or, if that is unavailable, a structured plausibility assessment. Please add a dedicated section that addresses the following:
Completeness. Report valid, missing, and imputed observations by attribute and geographic domain, including the spatial pattern of failures and coastal filling.

Internal invariants. Test conservation of extensive quantities across bifurcations and reconvergences.

Source-to-product checks. Re-extract a documented sample of raster and vector inputs and compare the independently calculated values with the released fields.

Independent or semi-independent plausibility checks. Examples could include GRIT drainage area against gauge metadata, known reservoir or lake totals in selected basins.

Sensitivity. Quantify sensitivity to bifurcation weights, midpoint versus catchment extraction and resampling choices.

Specific comments
l. 8-10 / Table 2 / repository metadata: 35 static variables plus 64 embedding dimensions equals 99, whereas Table 2 reports 98 and the example repository describes approximately 60 attributes. Reconcile all totals and define consistently what counts as a variable, embedding dimension, stored field, and expanded attribute.

Abstract l. 13-14 versus l. 63-64: The statements that reaches are up to 1 km long and that each segment is divided into equal-length reaches shorter than 1 km are compatible, but they do not uniquely specify the subdivision algorithm. State the exact rule, including treatment of short remainders and segment nodes, and provide the distribution of reach lengths.

l. 71-74: State the raster reprojection and resampling method for every data type. Nearest-neighbour, bilinear, and area-conservative resampling have materially different implications for categorical classes, continuous variables, and totals.

Table 1, "Year" column: "Most recent" is not a reproducible date. Give the exact source version, release date, and retrieval date used for GRanD v1.3, SoilGrids v2, FABDEM, GRIT, and any other rolling or versioned source.

Table 1, air-temperature seasonality: "Climatology" is capitalised and the citation is parenthesised inconsistently with adjacent rows.

Units across figures and Table A1: Sand fraction appears as percent in Figures 2 and 7 but as g/kg in Figure 5 and Table A1. Stream gradient appears as m/m in Figures 3 and 5 but m/km in Table A1. Harmonise units or explicitly annotate every conversion.

Figure 5 and l. 144-145: The statement that no visible differences occur in sand fraction across climate regions appears stronger than the boxplots support. Quantify the comparison or soften the statement.

Data availability and licensing: State the licence of the derived product and substantiate the assertion at l. 77-79 by listing the relevant licence and redistribution terms for each source dataset. Confirm that redistribution of the derived attributes is compatible with the entire source chain.
Citation: https://doi.org/10.5194/essd-2026-279-RC2
RC3: 'Comment on essd-2026-279', Anonymous Referee #3, 28 Aug 2026

============================================================

GENERAL ASSESSMENT AND SUMMARY

============================================================

The manuscript presents GRIT-ADB, a global hydro-environmental attribute database built on the GRIT bifurcating-river hydrography. For roughly 20 million river reaches (19.6 million km of rivers), it harmonises 35 static and a set of time-varying attributes from state-of-the-art global sources across five categories (hydrology, physiography, climate, land cover and use, soils and geology), and reports them as local and upstream-accumulated statistics at reach, segment and catchment scales. Its distinguishing feature is that the accumulation respects divergent, multi-channel flow, so attributes are propagated correctly through bifurcations and deltas that single-thread networks such as HydroSHEDS/HydroATLAS cannot represent.

The core idea is genuinely novel and, by ESSD's standards, unique: a topology-aware attribute database that follows bifurcating rivers does not exist elsewhere, and it is not something a user could reproduce on a routine basis. The choice to build on GRIT, to draw only from openly licensed, actively maintained sources, and to provide both reach- and segment-scale statistics is sensible and clearly useful for large-sample and network-based hydrology. Adding annual satellite embeddings as a time-varying layer is forward-looking. The topic sits squarely within ESSD's scope, and the underlying hydrography (Wortmann et al., 2025) is a strong foundation.

My reservations are about the paper as a data paper rather than the idea. The main one is close to blocking until addressed: there is no validation or quality assessment of the attributes, so a reader cannot yet judge how far to trust them. A second concern is how the manuscript currently reads: the text is weighted toward motivation and illustrative science, and comparatively little of it does what a data descriptor should do, namely describe the dataset's contents, structure, reference periods and quality so that a reader can pick it up and use it. None of this is fatal; the dataset looks valuable and the fixes are mostly a matter of adding a validation section and rebalancing the writing. The comments below start with that quality gap, then the readability point, then more specific matters.

============================================================

MAJOR COMMENTS

============================================================

1. A data paper needs a quality assessment (Sect. 4).

A data descriptor of this kind needs validation against independent references or, where that is impossible, a plausibility assessment that characterises uncertainty and flags extrapolation zones and data-sparse regions. The manuscript currently offers neither. The nearest thing is a visual comparison with HydroATLAS in the Pearl River Delta and the Congo (Fig. 7), which shows that the two products differ. Since every attribute is inherited from an external source and then aggregated and accumulated along the network, two questions matter and go unanswered: how much uncertainty each source contributes (the SoilGrids spatial bias is noted at l. 200–202 but not quantified), and how the accumulation step propagates, and possibly amplifies, that uncertainty downstream. CAMELS (Addor et al., 2017) is a good model here, since much of its value came from documenting the limitations of each attribute. A dedicated validation and uncertainty section, even a plausibility assessment with a few independent cross-checks (gauged basins, RiverATLAS, in-situ soil or lake records), would move the paper from "here are some maps" to "here is a dataset you can trust."

2. Reads more like a research paper, less like a data description.

It is worth looking at how the text is distributed, since a data descriptor is consulted more often than it is read straight through. Three things stand out. The Introduction is the single longest section and spends much of its length re-motivating large-sample hydrology; it can lose a third. A reader would come to "Dataset overview" (Sect. 3) to learn what is in GRIT-ADB and how to use it. For that currently the reader has to go to the Appendix to find some of the reference material they actually need, namely variable definitions, file formats and naming, the keys that join attributes to GRIT geometries, per-attribute reference periods, and quality. The vignettes provided currently in the section are interesting, but not primary. Meanwhile the Conclusions largely restate the Abstract. A data descriptor should be concise and should say plainly what is in the dataset, where, when, and how. I would rebalance: tighten the Introduction, turn Sect. 3 into a genuine description of the dataset's contents and structure (moving the exploratory analyses into a short "example applications" subsection or the supplement), add the validation section from comment 1. The paper would end up both shorter and easier to use.

3. Clarify the attribute count.

The headline count, "64 time-varying (multi-dimensional embeddings) and 35 static variables (>300 attributes)," reads as if the database holds 64 independent environmental variables that change through time. In fact the 64 are the latent dimensions of a single product, the AlphaEarth satellite embedding (Brown et al., 2025); essentially the only other time-resolved quantity is lake area. Table 2 then lists GRIT-ADB as "98 (>300)" variables, which matches neither "35 + 64 = 99" nor the abstract. I would state plainly that the time-varying content is one embedding product of 64 dimensions and reconcile the counts. Also, the embeddings are abstract features whose value for hydrology is currently asserted in this manuscript (Fig. 6) rather than demonstrated or previous application cited. I would suggest to add few sentences on what a user can and cannot do with them would set expectations straight.

4. Engage the large-sample datasets the paper is built for.

GRIT-ADB is repeatedly motivated by large-sample and machine-learning hydrology (Abstract; l. 162–164), yet it does not cite the reference dataset of exactly that field: Caravan (Kratzert et al., 2023), which is a co-author's own work and the standard integration of CAMELS-style static and dynamic attributes for global ML hydrology. CAMELS itself (Addor et al., 2017) is also absent. A short comparison with Caravan, particularly on the time-varying-attribute angle where it is the obvious benchmark, would sharpen the novelty claim. Two further omissions bear on specific analyses: the lake-connectivity results (Sect. 3.2) use GLAKES but mention neither HydroLAKES (Messager et al., 2016), the standard global lake inventory, nor Lake-TopoCat (Sikder et al., 2023), an ESSD dataset that already provides global lake drainage topology and catchments and is the natural comparison for GRIT-ADB's lake-connectivity attributes.

5. Climatologies remain outdated (Table 1; l. 38–41).

A central argument of the Introduction is that existing databases rely on outdated climatologies (for example HydroATLAS discharge from 1971–2000). But GRIT-ADB's climate layer is itself a mix of periods: aridity and PET come from a 1970–2000 climatology (Global-AI_PET_v3), while temperature and precipitation use 1991–2020 (MSWX/MSWEP) and land cover is a single 2020 snapshot. So the gain in temporal currency is real for some attributes and absent for others. I would harmonise the reference periods where possible or, at least, state the mismatch openly, so that users are not misled by the "present-day" framing.

============================================================

MINOR AND TECHNICAL COMMENTS

============================================================

1. Broken DOI (l. 227).

The GRIT vector DOI is printed as "10.528 1/zenodo.17435232" with a space; correct to 10.5281/zenodo.17435232.

2. Dataset vs paper authorship.

The archived-dataset citation (Zhang et al., 2026; l. 225 and References) lists Zhang, Slater, Wortmann, Liu and Moulds, omitting Kratzert, who is a manuscript co-author. Please reconcile the authorship if this is unintended.

3. Table 1 citations.

Parenthesis typo on "Air temperature seasonality / (Beck et al., 2022)."

4. No Code availability statement.

The Methods describe an "efficient, automated workflow" (l. 203) and the Conclusions call the framework "reproducible" (l. 221). Do the authors intend to make the code open source? Please add a Code availability section and comment on this.

5. Replace "most recent" with a year (Table 1).

This is vague for GRanD v1.3, GRWL, GRIT and SoilGrids; give explicit year so users know how current each layer is.

6. Repetition.

Several sentences recur almost verbatim across the Abstract, Introduction and Conclusions; trimming them would cut length and aid readability.

============================================================

REFERENCES

============================================================

- Addor, N., Newman, A. J., Mizukami, N., & Clark, M. P. (2017). The CAMELS data set: catchment attributes and meteorology for large-sample studies. Hydrology and Earth System Sciences, 21, 5293–5313. doi:10.5194/hess-21-5293-2017.

- Kratzert, F., Nearing, G., Addor, N., et al. (2023). Caravan — A global community dataset for large-sample hydrology. Scientific Data, 10, 61. doi:10.1038/s41597-023-01975-w.

- Messager, M. L., Lehner, B., Grill, G., Nedeva, I., & Schmitt, O. (2016). Estimating the volume and age of water stored in global lakes using a geo-statistical approach (HydroLAKES). Nature Communications, 7, 13603. doi:10.1038/ncomms13603.

- Sikder, M. S., Wang, J., Allen, G. H., Sheng, Y., Yamazaki, D., Song, C., Ding, M., Crétaux, J.-F., & Pavelsky, T. M. (2023). Lake-TopoCat: a global lake drainage topology and catchment database. Earth System Science Data, 15, 3483–3511. doi:10.5194/essd-15-3483-2023.

Citation: https://doi.org/10.5194/essd-2026-279-RC3

Boen Zhang, Michel Wortmann, Yinxue Liu, Frederik Kratzert, Simon Moulds, and Louise Slater

Supplement

https://doi.org/10.5194/essd-2026-279-supplement

Data sets

GRIT-ADB: A Global Hydro-Environmental Attribute Database for the GRIT Hydrography B. Zhang et al. https://doi.org/10.5281/zenodo.19363178

Boen Zhang, Michel Wortmann, Yinxue Liu, Frederik Kratzert, Simon Moulds, and Louise Slater

Viewed

Total article views: 326 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
203	94	29	326	53	31	41

HTML: 203
PDF: 94
XML: 29
Total: 326
Supplement: 53
BibTeX: 31
EndNote: 41

Views and downloads (calculated since 26 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	92	38	6	136
Aug 2026	55	31	12	98
Sep 2026	56	25	11	92

Cumulative views and downloads (calculated since 26 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	92	38	6	136
Aug 2026	55	31	12	98
Sep 2026	56	25	11	92

Viewed (geographical distribution)

Total article views: 265 (including HTML, PDF, and XML) Thereof 265 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 16 Sep 2026

Download

Preprint (15937 KB)
Metadata XML

Short summary

We created a new global hydro-environmental database to better describe how rivers are connected and how surrounding environments vary across space and time. Most existing databases simplify river networks, miss complex flow paths, and are largely based on older global datasets. Our database provides more than 300 attributes for 19.6 million km of rivers and streams. It supports more accurate study of water movement, environmental change, and climate impacts worldwide.


Total:	0
HTML:	0
PDF:	0
XML:	0