the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
A globally scalable, light data framework for flood-hazard mapping using open geospatial services
Abstract. Flood-risk screening is often limited by the lack of globally consistent hazard layers, because detailed hydraulic models require local calibration, boundary conditions, and substantial computation. This study presents a light data framework that uses application programming interfaces to assemble global elevation, hydrography, and road-network data for flood-hazard susceptibility mapping. Height Above Nearest Drainage is derived with drainage calibration constrained by OpenStreetMap hydrography, while the Multiresolution Index of Valley-Bottom Flatness adds complementary information on valley planarity. An interpretable monotonic gradient-boosted regression model is trained on return-period inundation inventories and converted into a common five-class ordinal hazard scale using threshold sets suited to planar and incised river settings. Applications to three independent river reaches show that the upper hazard classes consistently capture the 100-year flood footprint. The ordinal ranking remains physically coherent across return periods: low-susceptibility terrain is largely insensitive to increasing flood severity, whereas higher classes show progressively greater inundation likelihood. The model distinguishes flooded from non-flooded terrain in 78–83 % of pairwise comparisons and remains effective under strict false-alarm constraints. The framework delivers screening-grade hazard layers for prioritisation and for integration with exposure and vulnerability analyses. Future work should test broader climatic and geomorphic settings and refine transferability across regions.
- Preprint
(5131 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-3884', Anonymous Referee #1, 19 Oct 2026
-
AC1: 'Reply on RC1', Igor Moreira, 02 Nov 2026
We thank the referee for the comments. We understand the claim for revision, but we respectfully disagree that the study lacks a test on excluded stretches or that the classification assigns one fifth of every new reach to its highest class.
1.
Dottori et al. (2016) mapped inundation for specific return periods. Nardi et al. (2019) delineated geomorphic floodplains at 250 m. We agree that both papers can be cited to enhance the literature review, but our model addresses a significantly different, yet related question: can a model trained on return-period extents use HAND and MRVBF to rank sites along a new river corridor without local inundation maps? About the reported AUCs, they measure agreement with Metroplan and it would be wrong to compared them with existing methods. But to make it clear that the proposed model is better than HAND alone, we propose comparing continuous HAND and our model on the same held-out pixels, selecting any HAND threshold on the calibration reaches alone.
2.
Classes 1–5 denote levels of relative propensity to river flooding. Outside the mapped channel, each class represents a band of the terrain-based score learned. Thuss, the labels express an ordering without assigning a return period or annual probability to any cell.
Quintiles define the cutoffs in each calibration reach. For a new reach, five score percentiles determine how the two stored sets of cutoffs are combined. Two windows with the same five percentiles receive the same cutoffs regardless of area. No class has to occupy 20% of the new window. A different boundary can change the terrain sampled and the resulting scores, but the difference in calibration area does not by itself establish a window artefact.
The evidence for regional screening comes from the withheld flood maps. All three validation reaches met the trend criterion for increasing inundated fraction across classes at all five return periods (Table 6). In Areia, 87.2% of Class 5 lay inside the modelled 100-year footprint, compared with 46.8% of Class 4 (Fig. 8). The cutoff rule does not impose this contrast on a withheld footprint.3.
Eq. (1) measures the proportion of extracted channel cells within 1 km of OSM hydrography. Gravataí’s 100% score therefore establishes proximity alone. Removing cells can raise or lower the score, and the code retains the densest network among tied maxima. We understand that an omitted tributary could bias HAND locally, but 8 of the 15 predictors are computed independently of the extracted network. For example, MRVBF comes directly from the DEM, and channel distances are measured from OSM.
4.
“Globally scalable” describes how the susceptibility model can be applied reach by reach. Given a new study window, our model retrieves Copernicus elevation and OpenStreetMap hydrography, selects the local map projection, derives the terrain predictors and applies the fitted model. It sets class limits from the new reach’s predicted scores and the two stored prototypes. Wherever the input coverage is adequate, adding a reach requires neither local inundation maps nor a new supervised fit. The two calibration reaches supply the anchors for this procedure. Each reach-level descriptor takes one value per reach, while HAND, valley-bottom flatness and channel proximity vary from cell to cell.
The study also includes a test on withheld reaches. Metroplan provided the only return-period simulations available to us for these reaches. Using one provider keeps the reference source consistent, while withholding the target maps from model fitting and class assignment leaves their flood boundaries unseen by the model. Paranhana is a tributary of the Sinos, as the referee notes, but even when we set Paranhana aside, the model discriminated terrain inside and outside the modelled extents on the Jacuí (AUC 0.827) and Areia (0.794).5.
The excluded zones outlined in Fig. (4) were a deliberate choice about what our machine learning model should learn. A low-lying floodplain may not be captured by a modelled inundation extent because a dike or road embankment interrupts its connection to the river. If that area were used as an ordinary dry training example, the model could attribute the absence of inundation to the shape of the terrain and carry that inference over to a similar, unprotected floodplain. Excluding those examples from fitting preserves the distinction between terrain susceptibility and engineered protection.
That distinction also explains the map’s use in developed areas. For example, if a mapped dike borders land that our model ranks highly and Metroplan’s footprint stops at the dike, planners can identify a specific corridor where the role of the defence merits closer examination. The ranking describes the terrain as represented in the DEM. Establishing whether the dike is well sited, what it protects, or how flooding would change if it overtopped requires local hydraulic analysis. We would make this sequence explicit in the applications section. Our model can identify where to prioritise such assessments before zoning, evacuation or resource-deployment decisions are made.On all three reaches withheld from calibration, our model produced susceptibility rankings that distinguished terrain inside and outside the modelled flood extents without using local inundation maps to assign the classes. We respectfully ask the editor to invite a major revision so that we can explain more clearly how the classification transfers to new reaches and how its results can guide the selection of sites for detailed hydraulic assessment.
References:
Dottori et al. (2016). Development and evaluation of a framework for global flood hazard mapping. https://doi.org/10.1016/j.advwatres.2016.05.002
Nardi et al. (2019). GFPLAIN250m, a global high-resolution dataset of Earth's floodplains. https://doi.org/10.1038/sdata.2018.309
PRIORF v2.1.0. https://doi.org/10.5281/zenodo.18990241Citation: https://doi.org/10.5194/egusphere-2026-3884-AC1
-
AC1: 'Reply on RC1', Igor Moreira, 02 Nov 2026
-
RC2: 'Comment on egusphere-2026-3884', Anonymous Referee #2, 30 Oct 2026
The manuscript presents methodology for developing basic flood maps in an area with little data availability. The paper is well written and I didn’t identifiy notable grammatical errors. The method uses only globally available spatial data and thus should be useful in less developed regions. However, certain further explanation and background is needed to make publishable manuscript. In particular, the need for calibration to existing flood maps should be expanded upon further to better explain the global relevance that the manuscript suggests for this methodology. Overall I recommend major revisions to the manuscript before publication. In particular, the following should be addressed:
-The backgroung review should include more discussion of other global scale modelling alternatives. There are other approaches in this field that have been developed, but they werent described here.
-More context of the area that the method was developed should be included. Floodplain maps are very dependent on their geography. An overall map of the test site locations should be included. In addition, the manuscript should include a description of the sites (geomorphology, landcover, climate) and how they are similar or different to eachother.
-More care should be used to describe the two sites. The site map/figures says reach 1/2 whereas the text refers to them by name. It is difficult to tell which is which (gravatai, sinos)
-More information should be included for the validation reaches. How far are the validation reaches from the calibration reaches? It is unclear where they are.
-All the map figures (3,4,7) should be much bigger, should use more contrasting colors (eg hydrography and floodplain are both blue) and should strive to give a better visual representation of what the performance statistics mean in real world terms for flood hazard. To help promote utlization of these approaches we need to have a better understanding of what this level of performance means for real flood hazard planning. Including real physical landmarks (roads, towns, etc) would be helpful in this regard. The inclusion/exclusion of physical features in the experimental maps is more relevant than a performance statistic for real world use.
-Most importantly, the mapping framework should be described in more detail. It is unclear how often or for how similar an area would a new calibration be needed. For what spatial area is the calibration to the "real" flood maps valid? Considering that the training results (table 1) are substantially different for the 2 reaches, I presume recalibration/application would be necessary many different times if this were applied over a large area. If new calibration maps are needed for every application, the global usability of the method is greatly hampered, but I can't tell if that is the case from the description here. Adding a validation case on somewhere far away from the test site would help to show the wider applicability of the method.
Citation: https://doi.org/10.5194/egusphere-2026-3884-RC2 -
AC2: 'Reply on RC2', Igor Moreira, 02 Nov 2026
We thank the referee for the comments. We understand that the background review should be enhanced, and we want to do that by starting from the references we cited in our response to Referee #1. About the other comments:
1.
All five reaches are in Rio Grande do Sul, southern Brazil. In Figures 3 and 4, (i) is Sinos, the broader, less confined calibration window, and (ii) is Gravataí, the more confined window. In Figure 7, (i) is Paranhana, (ii) Jacuí and (iii) Areia. We thank you for pointing out the need to make this clearer in our manuscript. In the revised paper, we are going to make sure that we label each panel by river name, add a locator map showing all five windows and their catchments, and report the distances from each validation-window centre to both calibration-window centres.2.
Again, we thank you for pointing this out. In the revised manuscript we will enlarge Figures 3, 4 and 7, distinguish river channels from the modelled 100-year outline, and label towns and road crossings.3.
We thank you for asking what must be recalibrated when our model is applied to a new basin. Flood inventories were essential for training our model, but once fitted, the model can generate a susceptibility map without an inundation map for the new reach. Local reference data are only needed to assess how well that prediction transfers.
The percentages in Table 1 (23.1% for Gravataí and 6.9% for Sinos) are thresholds for extracting drainage from the DEM. Both are relative to the maximum flow accumulation in their respective windows. Our model repeats this step in each new area, using OpenStreetMap hydrography to select the drainage network for HAND.
The same model was fitted jointly on Sinos and Gravataí using Metroplan’s modelled return-period extents. In a new reach, we retain the fitted model and use the distribution of its predicted scores to combine the two stored sets of class limits. In the reported evaluation, the Paranhana, Jacuí and Areia reference maps were reserved for testing. The different Table 1 values therefore do not imply that the model must be refitted for each reach.Citation: https://doi.org/10.5194/egusphere-2026-3884-AC2
-
AC2: 'Reply on RC2', Igor Moreira, 02 Nov 2026
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 296 | 174 | 60 | 530 | 43 | 40 |
- HTML: 296
- PDF: 174
- XML: 60
- Total: 530
- BibTeX: 43
- EndNote: 40
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Overall
The paper builds an automated workflow that downloads a 30 m DEM and OpenStreetMap rivers, computes HAND, MRVBF and 14 derived variables, fits a gradient-boosted model to Metroplan return-period flood maps, and cuts the score into five "hazard" classes. It is trained on two reaches and tested on three others, with AUC of 0.784–0.827 (Table 6). Open data and code are welcome, but the problems are in the foundations, not the details. The premise ignores existing global flood hazard layers. The five classes are quantile ranks within the study window and have no fixed meaning. The drainage calibration, presented as the first methodological contribution, favours a network reduced to the main stem. The model learns from two reaches, and all testing uses maps of one agency in one region, while the title claims global scalability. These issues cannot be fixed by revision, and I recommend rejection.
Major comments
References mentioned in this report