the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 23 Sep 2026
Multi-Resolution Flood Hazard and Exposure Assessment in the Upper Narmada Basin
Abstract. Spatial resolution strongly influences how flood hazards and exposed assets are represented, yet this source of uncertainty is often overlooked in impact-oriented flood studies. This study examines how hydraulic model resolution and exposure-data aggregation affect flood hazard classification and estimates of exposed buildings and population in the downstream reach of the Upper Narmada Basin, India. Two-dimensional flood simulations were performed for the August 2022 and 2020 events and evaluated using radar-derived inundation extent and observed discharge data. A 30 m hydraulic model was used as the reference and compared with coarser 140 m and 280 m resolutions. Building exposure was derived from combined OpenStreetMap and Open Buildings datasets, while population exposure was estimated using WorldPop data. Coarser resolutions systematically overestimated flood extent, hazard zones, and exposed elements. The overpredicted inundation area increased from 51.30 km² at 30 m to 58.45 km² at 140 m and 73.36 km² at 280 m. Similarly, the estimated number of affected buildings increased from 894 (2.5 times) at 30 m to 3051(8.8 times) at 140 m and 11 537 (33 times) at 280 m, while the affected population increased from 1612 to 5982 and 24 733, respectively. These results show that coarse-resolution hazard and exposure data can substantially inflate exposure estimates and may distort emergency planning, infrastructure prioritization, and resource allocation. Spatial resolution should therefore be treated as a key source of uncertainty in flood hazard and exposure assessment, and model/data resolution should be aligned with the research question and intended decision scale.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Natural Hazards and Earth System Sciences.
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.- Preprint
(1665 KB) - Metadata XML
-
Supplement
(108 KB) - BibTeX
- EndNote
Status: open (until 03 Dec 2026)
- RC1: 'Comment on egusphere-2026-4699', Dimaghi Schwamback, 21 Oct 2026 reply
-
RC2: 'Comment on egusphere-2026-4699', Anonymous Referee #2, 28 Oct 2026
reply
This study presents some very interesting findings that, although observed before, have, to my knowledge, not been systematically studied until now. I therefore believe this article could be a good addition to the NHESS journal, as it raises some very important concerns for flood risk modellers. However, I do have serious concerns about two methodological choices the authors have made, and I think these should at least be acknowledged in the text as potentially influencing the conclusions of this study. Since I do not think these methodological choices are sufficiently harmful to invalidate the results, I believe the manuscript can be accepted with minor revisions.
More important points:
Why did you choose to use SRTM as the DEM? SRTM is a surface map rather than a terrain map. It therefore captures the height of trees and building roofs rather than the elevation of the terrain over which water flows. Why didn't you use MERIT, FABDEM, or another terrain map specifically designed to correct this issue? My concern is that this could affect the results, as a surface map may have sharper elevation changes and may therefore be more sensitive to resolution differences than a terrain map. You should probably discuss this concern in the discussion section. In general, I would have been more convinced if this study had been conducted using a LiDAR map to ensure that spatial elevation differences reflect real terrain elevation differences rather than artifacts from trees and rooftops or artifacts introduced when correcting radar maps. In any case, I don't think these choices invalidate the results, but I think readers should be made aware of these limitations, and these choices should definitely not be repeated in follow-up studies.
The study uses satellite-based flood data to calibrate the flood model and compares the number of people affected according to satellite data with the modelled number of people affected. In doing so, it assumes that the satellite data is more reliable than the modelled data. However, older satellite-based flood maps are known to be inaccurate around buildings. That is, they may misclassify flooded areas with dense buildings as dry land because building roofs look different from open water. This means that the difference between your 30 m resolution results and the satellite data may be entirely due to artifacts in the satellite data. I think you should either devote at least a paragraph to firmly establishing that the satellite data is more reliable than the model results, directly addressing this concern with good evidence that such misclassification did not occur, or treat the satellite data as just another potential result rather than as validation data. In fact, it could be another interesting finding that some satellite products potentially underestimate exposure.
In Section 5.3 and the introduction, I think some of the statements are no longer entirely correct. Over the last few years, detailed and freely available building footprints have become available almost globally (from Google and Microsoft). The problem of coarse exposure data in flood risk modelling, which was significant 10 years ago, is therefore much less of a problem today. Uncertainties in structure and occupancy type are still a problem, but these are more closely related to vulnerability than to exposure and flood map resolution. Furthermore, there are methods to compute flood maps at a coarse resolution and then increase the resolution through post-processing (e.g. based on the assumption that water levels are locally flat). These approaches are, in fact, very computationally efficient and may eliminate most of the observed uncertainties for larger floodplains. That being said, I don't think the fact that this problem is avoidable invalidates this study. Many modellers are not aware of the importance of resolution, and highlighting this in a systematic study is useful to ensure that this error is made less frequently. Furthermore, many older studies that did not yet have access to high-resolution data are still being used, and it is good to raise awareness that these studies may overestimate the number of people affected. However, I do not think that framing this problem as "unavoidable" is entirely correct.
The first half of the introduction is excessively long, with many details that are irrelevant to the study. The second half of the introduction is more relevant but is written in such abstract terms that it becomes generic. It would be better to be clearer and more direct about the problems. See some examples in the detailed comments.
Minor points:
Lines 103–122: Some of the language is generic and obscures the key problems, which are well described in the DEM literature. Radar maps are widely available for free at resolutions down to 30 m, but they represent the surface and become highly uncertain when corrected to represent terrain elevation. LiDAR technology is a solution to vertical errors but is expensive and requires local access and flying permits. Hydrodynamic computational costs are a different and more serious problem that should be addressed more directly, especially when multiple scenarios need to be computed. Financial costs are a problem for radar data below 30 m resolution or LiDAR.
Lines 120–143: Here again, some practical points are not communicated well. It should be noted that the availability of free, high-resolution building data has greatly improved in recent years. That doesn't eliminate the concerns, but this is mostly about choices the modeller makes rather than data availability.
Section 2: It is unclear whether WorldPop data is disaggregated to building footprints from OpenStreetMap or used directly. You should also describe exactly what data you obtained from OpenStreetMap. Were these points or polygons? Furthermore, in this section, I find it difficult to form a clear picture of the type of river we are dealing with. This seems to be a rapidly responding, relatively upstream river section. Perhaps mention some historical peak discharges and clarify whether you are studying a relatively flat floodplain or a narrow, rapidly flowing mountain river. It would also help if you added a legend to the elevation map. Perhaps consider reversing the colours in the elevation map, because typically brown indicates high elevations and green indicates low elevations, whereas you seem to have chosen the opposite convention.
When you mention a uniform Manning coefficient in the floodplain, please clarify that you used a different coefficient in the river itself.
Citation: https://doi.org/10.5194/egusphere-2026-4699-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 303 | 141 | 58 | 502 | 98 | 65 | 62 |
- HTML: 303
- PDF: 141
- XML: 58
- Total: 502
- Supplement: 98
- BibTeX: 65
- EndNote: 62
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General assessment
Dear Author,
Congratulations on the paper. This manuscript addresses an important and practical question: how does spatial resolution affect flood hazard mapping and the estimation of exposed buildings and population? The comparison of three spatial resolutions and two flood events provides a useful basis for exploring this issue, and the results could be relevant for flood-risk assessment and emergency planning in data-scarce regions.
I believe the manuscript has potential, but some methodological aspects need to be clarified before the main conclusions can be fully supported. In particular, the hydraulic model resolution and the exposure-data resolution change together, making it difficult to distinguish their individual effects. I also encourage the authors to explain more clearly how the models were validated, how exposure was assigned to hazard classes, and what the reported differences mean in practical terms.
My comments follow the order of the manuscript and are intended to help strengthen the analysis and improve the clarity of the paper. I am looking forward to receive the corrected version!
Comment 1.
The abstract reports inundation areas of 51.30, 58.45, and 73.36 km² for the 30 m, 140 m, and 280 m resolutions, respectively, describing these values as overpredicted inundation area. It would be helpful to clarify exactly how these numbers were calculated. Do they represent the total simulated inundated area, the area of overprediction relative to the satellite-derived flood extent, or the spatial disagreement between the simulated and observed flood maps?
Comment 2.
The abstract states that coarser resolutions systematically overestimate the number of exposed buildings and people. However, the reported values depend on both the simulated hazard maps and the spatial representation of the exposure data. I suggest clarifying that the estimates increase substantially with coarser resolution and reserving the term overestimation for comparisons against an independent reference or validation dataset.
It would also help to briefly state the main practical contribution of the study without implying that the 30 m simulation necessarily represents the true exposure.
Comment 3.
The Introduction provides a useful overview of flood hazards, DEM resolution, and exposure mapping. However, the discussion could be more focused on the precise gap this study aims to address.
The authors mention previous work on DEM resolution, hydraulic model resolution, and exposure mapping. I suggest concluding this section with a clearer explanation of what remains unknown when these components are considered together. For example, is the main gap the effect of hydraulic grid size on hazard classification, the effect of aggregating buildings and population, or the combined influence of both? This distinction would make the study's contribution easier to identify.
Comment 4.
The manuscript asks how DEM resolution and hydraulic model grid size affect building-level and population exposure, and which resolution provides a practical balance between computational efficiency and reliable exposure mapping. These are relevant questions, but the second question requires an explicit assessment of computational performance as well as mapping reliability. I suggest stating which indicators will be used to evaluate that balance, such as model runtime, computational demand, flood-map agreement, and differences in exposure estimates.
Comment 5.
The Introduction correctly notes that higher resolution does not necessarily mean higher accuracy. This is an important point that could be developed further because the study uses a 30 m SRTM DEM as its starting dataset. I encourage the authors to explain that changing the resolution of an existing DEM is not the same as obtaining new, more detailed topographic information. This distinction is important for interpreting the results and avoiding the assumption that a finer grid automatically provides a more accurate representation of the terrain.
Comment 6.
The manuscript describes the Upper Narmada Basin and identifies Hoshangabad, Sehore, and Raisen as flood-prone districts. However, the hydraulic model covers only the downstream portion of the basin. I suggest explaining more clearly how the model domain was delineated and why this particular area was selected. A brief description of the main characteristics of the downstream floodplain, settlements, and river network would help readers understand the relevance of the selected domain.
Comment 7.
The manuscript uses observed discharge data and radar/MSS-derived inundation extent. Please provide sufficient information on the source, spatial resolution, acquisition date, and processing of the satellite-derived flood extent. It would also be useful to explain how the observed flood map was delineated and whether areas affected by cloud cover, vegetation, or other classification limitations were considered. Since the satellite-derived flood extent plays an important role in assessing model performance, readers need to understand its uncertainties and limitations.
Comment 8.
The building inventory combines OpenStreetMap and the Open Buildings dataset, while population exposure is estimated using WorldPop. These are appropriate sources for regional assessment, but their spatial and thematic uncertainties should be acknowledged. Could the authors clarify the reference year and spatial resolution of the WorldPop product, how overlapping building records were identified and removed, and how residential, commercial, and industrial uses were assigned? It would also be helpful to explain how population was allocated to buildings and whether the total population was preserved after aggregation. These details are particularly important because the exposure estimates increase substantially between resolutions.
Comment 9.
This is my main methodological concern. At 30 m, the study uses the hydraulic model grid and individual building footprints, whereas at 140 m and 280 m, both the hydraulic model and the exposure data are aggregated to coarser spatial units. Consequently, the observed differences cannot be attributed exclusively to the hydraulic model resolution. I strongly encourage the authors to consider an additional experiment in which the same building inventory and population dataset are used across all three hydraulic simulations. This would help isolate the influence of the hazard model resolution. A complementary analysis could then hold the hazard map constant while changing the exposure aggregation. If these additional experiments are not feasible, the limitation should be stated clearly, and the conclusions should refer to the combined effects of hydraulic resolution and exposure-data aggregation.
Comment 10.
The original 30 m SRTM DEM is resampled to 140 m and 280 m. Please specify the resampling method used and explain why it was selected. For example, were elevation values averaged, sampled from a representative cell, or processed using another method? This matters because different resampling approaches can produce different representations of channels, floodplain slopes, depressions, and local topographic barriers. A short sensitivity discussion would help readers assess how much the results may depend on this choice.
Comment 11.
The manuscript describes the exposure data using QuadTree tiles in EPSG:3857 and refers to approximate tile dimensions of 140 × 140 m and 280 × 280 m. Please clarify whether these dimensions are approximate ground dimensions or dimensions in the projected coordinate system. Since Web Mercator introduces spatial distortion, it would be useful to explain why this projection was retained for the analysis and whether the resulting tile dimensions were checked against the intended ground resolution.
4.2
Comment 12.
The authors test nine combinations of channel and floodplain Manning's roughness coefficients and select the best-performing configuration using the fitness index, NSE, and RRMSE. This is a reasonable starting point, but further clarification is needed. Please explain how the nine parameter combinations were selected, how the fitness index was calculated, and how the three performance metrics were used to select the final parameter set. Were all metrics given equal importance, or was one considered the primary criterion? I also encourage the authors to distinguish calibration from validation more explicitly. If the same observations were used to select parameters and evaluate model performance, the reported agreement may not represent an independent validation.
Comment 13.
The optimal floodplain Manning's roughness increases from 0.03 at 30 m and 140 m to 0.09 at 280 m, while channel roughness remains at 0.02. The authors interpret this as compensation for unresolved terrain and surface features.This is an interesting result and deserves further discussion. Could the authors explain whether these values fall within physically plausible ranges for the modelled channel and floodplain? More importantly, does the need for a substantially different floodplain roughness indicate that the coarser model is compensating for structural limitations in its representation of the terrain? I suggest discussing the implications for comparing model resolutions when each configuration uses a different optimal parameter set.
Comment 14.
The manuscript uses the 30 m simulation as the reference for subsequent comparisons. However, a finer grid does not necessarily guarantee a more accurate simulation, particularly when all simulations originate from the same DEM source. Could the authors report the performance of each resolution against the observed discharge and satellite-derived inundation extent using clearly defined, comparable metrics? For the flood maps, it would be useful to report the observed and simulated areas alongside a spatial agreement measure, such as intersection over union or a critical success index, as well as false-positive and false-negative areas. This would allow readers to assess whether the 30 m model is demonstrably more accurate, rather than simply more spatially detailed.
Comment 15.
The manuscript defines six hazard classes, H1–H6, based on flood depth and velocity thresholds, with increasing consequences for people, vehicles, and buildings. However, the procedure for assigning these classes to buildings and aggregated exposure units needs to be described in greater detail. For example, when a building footprint overlaps several hydraulic cells, is the hazard class determined by the maximum depth and velocity, the mean values, the building centroid, or another rule? Likewise, how are hazard values assigned to the 140 m and 280 m exposure tiles? This is particularly important because the reported increases in high-hazard exposure, especially H5 and H6, are among the study's most consequential findings.
Comment 16.
Since the classification depends on depth and velocity thresholds, small changes in these variables may cause an asset to move from one class to another. If feasible, I suggest a brief sensitivity analysis or discussion of how the chosen thresholds affect the reported hazard distribution. At minimum, the authors should explain whether the thresholds are applied consistently across all spatial resolutions and whether they represent the same physical conditions at each scale.
Comment 17.
The manuscript explains that buildings and population are available at the individual-building level and are subsequently aggregated into 140 m and 280 m units. However, the procedure used to combine the hydraulic outputs with the exposure inventory is not sufficiently clear. Please specify how an exposure unit is classified as flooded, how buildings spanning more than one grid cell are handled, and how population is counted when a tile contains both flooded and non-flooded areas. Is any minimum flooded fraction required for an asset to be considered exposed? A clear description of these rules would improve reproducibility and help readers understand why the number of exposed assets changes with resolution.
Comment 18.
The paper estimates population exposure using WorldPop and differentiates residential and commercial sectors. Please explain how the WorldPop population is distributed among buildings and how the aggregation process affects the total population represented by the exposure dataset. I also suggest clarifying how population associated with commercial and industrial buildings is estimated. Since the number of exposed people changes substantially between resolutions, it is important to demonstrate that the differences reflect changes in spatial exposure assignment rather than unintended duplication, redistribution, or changes in the population totals. Finally, please distinguish population classified as spatially exposed from people who were actually present or affected during the flood event.
Comment 19.
The calibration results are useful, particularly the differences in optimal Manning's roughness. However, I suggest presenting the performance metrics for all resolutions in a consistent table and clearly identifying the selected parameter combination for each case. It would also be useful to report the absolute and relative differences in performance between the 30 m, 140 m, and 280 m simulations. This would help readers assess whether the increase in computational efficiency at coarser resolutions comes with a small or substantial loss of model performance.
Comment 20.
The results report values of 51.30, 58.45, and 73.36 km² for the three resolutions. Please clarify whether these values represent total simulated inundation, the portion outside the observed flood extent, or another metric. If they represent total simulated area, I suggest avoiding the term overpredicted area unless the difference relative to the observed area is explicitly calculated. If they represent false-positive inundation, please provide the calculation and the corresponding observed and simulated areas. This clarification is essential for interpreting the differences as evidence of model bias.
Comment 21.
Figure 4 shows that the areas associated with the hazard classes increase at coarser resolutions, with a particularly strong increase in H5. This is an important result, but the physical interpretation could be developed further. Could the authors explain whether the increase in H5 and H6 is mainly caused by the expansion of the simulated flood footprint, changes in water depth and velocity, or the rules used to aggregate hazard values? These mechanisms have different implications for flood-risk applications. I suggest comparing the spatial distribution of depth and velocity between resolutions, rather than relying only on the changes in hazard-class area.
Comment 22.
An increase in the area assigned to a hazard class does not necessarily mean that flood conditions have become more severe everywhere. It may also reflect the redistribution of water and the classification of larger spatial units. Please clarify this distinction in the Results and ensure that the figures and text consistently separate the total flooded area, the area in each hazard class, and changes in the spatial severity of flooding.
Comment 23.
Figure 5 is potentially one of the most informative figures in the manuscript because it illustrates how buildings and population move between hazard classes as the resolution changes. However, the interpretation would benefit from a clearer explanation of H0 (or H0*) and the exact population of assets included in each transition. Are the same buildings and population units tracked across all resolutions, or do the exposure units themselves change when the data are aggregated? Please also explain how the transitions are calculated when individual buildings at 30 m are compared with 140 m and 280 m tiles. This is particularly important for interpreting the movement of assets from the non-flooded class into H5 and H6.
Comment 24.
The Results report substantial numbers of newly exposed buildings and people at the coarser resolutions. I suggest adding a short explanation of how these numbers relate to the total exposure counts reported in Table 2. A compact table separating assets that remain in the same hazard class, move to a higher class, move to a lower class, or change from non-flooded to flooded would make the results easier to interpret and help readers distinguish changes in exposure from changes in hazard severity.
Comment 25.
Table 2 reports 346 buildings and 550 people in the satellite-derived reference, compared with 894 buildings and 1,612 people at 30 m, and substantially larger values at 140 m and 280 m. Could the authors explain exactly how the satellite-derived reference counts were obtained? In particular, how were buildings and population classified as affected using the satellite-derived flood extent, and were the same exposure inventories and spatial rules used for all comparisons? The satellite-derived reference is valuable, but differences between it and the modelled estimates may arise from flood-map detection limitations, exposure-data completeness, and spatial assignment rules, in addition to hydraulic model error. I therefore suggest describing these values as differences relative to the satellite-derived reference unless independent evidence establishes that all discrepancies represent overestimation.
Comment 26.
Figures 6 and 7 suggest that coarser resolutions increase exposure estimates and broaden the distribution of affected buildings and population across villages. Beyond reporting these increases, it would be useful to assess whether the same villages remain the main exposure hotspots across all resolutions. For example, the authors could compare village rankings or report the overlap among the villages with the highest estimated exposure. This would provide a more direct measure of the practical consequences of resolution choices. Even if total exposure estimates change, the implications for planning may be different if the same villages remain priorities.
Comment 27.
The histograms show strongly right-skewed exposure distributions, but the interpretation could be more quantitative. Please consider reporting a few summary statistics, such as the median, upper quantiles, and maximum village-level exposure for each resolution.
Comment 28.
The Discussion identifies the combined effects of an expanded flood footprint and aggregated exposure data as the main explanation for the increase in estimated exposure. This is an important interpretation, but the current experimental design does not appear to quantify the contribution of each component separately. I recommend either adding the complementary analysis described in Comment 9 or being more explicit that the study demonstrates the combined sensitivity of the hazard–exposure workflow to spatial resolution, rather than the isolated effect of hydraulic model resolution.
Comment 29.
The reported increases in exposure within H5 and H6 could have important implications for emergency planning. However, the movement of assets into these classes does not necessarily demonstrate that their actual physical risk has increased. Please distinguish between changes in modelled hazard classification, changes in estimated exposure, and changes in expected damage or risk. The manuscript does not appear to calculate actual damage or validate building failure, so claims about structural consequences should remain within the limits of the hazard-classification framework.
Comment 30.
The Discussion notes that coarser resolutions reduce computational demand but can substantially change flood and exposure estimates. However, the manuscript's stated objective includes identifying a practical balance between computational efficiency and reliable exposure mapping. I suggest reporting the actual computational cost of each simulation, including runtime, approximate number of computational cells, and the hardware used. If available, memory use would also be useful. These results would allow the authors to make a more evidence-based recommendation about when 140 m or 280 m simulations may be appropriate for regional screening and when finer-resolution modelling is needed.
Comment 31.
The distinction between regional screening and local-scale planning is useful. I encourage the authors to develop this point by explaining which applications may tolerate coarser resolution and which require more detailed hazard and exposure information. For example, a coarse model may help identify broad flood-prone corridors, whereas evacuation planning, building-level assessment, and the prioritization of individual settlements may require more spatially detailed information. The manuscript would be stronger if these recommendations were linked directly to the observed performance and exposure differences, rather than to resolution alone.