These parameters determined the channel capacity of the rivers and thus indicate the moment after which inundation processes started. Calculating the bankfull discharge is explained in Sect. 2.2.3 and Appendix A1.

We chose an automated event-based approach to assess the model's robustness and performance under varying conditions. Our simulations were structured event by event and assumed each observed flood event as a distinct and independent occurrence. The timings of the input hydrographs and model domain extent were derived from moderate resolution imaging spectroradiometer (MODIS) satellite observations, compiled into the Cloud to Street database (Sect. 2.4.1). This database includes a subset of 499 riverine flood events spanning five (sub)continents and 96 countries. To simulate floods on a global scale for each event, discharge data was obtained from global hydrological model LISFLOOD v4.0, as described in Sect. 2.2.2. Each event was processed and run separately, with all associated flooded basins (MERIT-BASINS) simulated in one model run. Notably, a single flood event can occur across multiple basins. We refer to these combined basins as the “model domain”. This setup allowed us to handle large-scale flood modelling tasks efficiently. By applying the model across many events, we validated the model across a wide range of flood magnitudes in different geographical (e.g. catchment sizes) and climatological settings.

Discharge source points were set up in the SFINCS model to simulate water coming in from rivers. These source points can be river inflow points, which simulated water entering the model domain from upstream rivers (e.g. rivers that cross the boundary of the model). Water can also enter the model coming from headwater rivers, which are streams that originate inside the model domain. The start or origin of these headwater rivers is called the headwater point. Accurate placement of both headwater and inflow points is essential for reliable flood modelling since the misalignment of these points can lead to the significant over- or underestimation of downstream flooding, particularly during catchment-wide events where multiple tributaries converge into a main river channel. Therefore, to ensure hydrological consistency between the hydrodynamic model and the hydrological input (which is at a much coarser resolution), headwater and inflow points are snapped to the grid cell that best matches the upstream contributing area of the hydrological model used. This linkage between hydrological discharge and the hydrodynamic river network is based on upstream drainage area matching, a method commonly used in large-scale flood modelling frameworks (e.g. Eilander et al., 2023a). The use of a tolerance in upstream area ensures correct placement of forcing points, and the influence of hydrological input resolution on model performance is further evaluated in Sect. 2.5 Sensitivity Analysis.This snapping was implemented using the HydroMT-SFINCS plugin (see Sect. 2.3.1: Data handling and pre-processing). Discharge input for inflow and headwater points was sourced from multiple datasets, as described in Sect. 2.2.2. To account for differences in resolution and small mismatches in the upstream area, we applied a tolerance of 5 % deviation when snapping. This approach helped to ensure that discharge was applied at hydrologically consistent locations, which reduced the risk of errors in flood extent prediction due to poor spatial alignment between the hydrological and hydrodynamic models.

Static input data provided essential environmental and topographical information to the SFINCS model. The following datasets were used for the global automated setup which include:

The FABDEM V1-2 dataset (Hawker et al., 2022) was used to provide elevation data at a global scale. This dataset has a 30 m resolution and is proven to be more accurate than existing global elevation datasets (Hawker et al., 2022).

Land-use patterns were derived from the ESA Worldcover 2021 dataset on a 10 m resolution (Zanaga et al., 2022). The land-use data was used to define Manning's roughness coefficient for different land-use types.

River geometries were obtained from the MERIT-SWORD dataset v0.4 (10.5281/zenodo.14675925, Wade et al., 2025), which is a dataset created by combining the SWOT River Database (SWORD; Altenau et al., 2021) and MERIT-BASINS (Lin et al., 2019). The MERIT-SWORD dataset transfers data, such as river width from SWORD rivers (30 m wide and greater), to corresponding MERIT-BASINS rivers by generating bidirectional links. Drainage areas smaller than 25 km2, including non-channelised areas along the coast or certain endorheic regions (e.g. incomplete basins or hillslopes), were excluded to maintain focus on larger river systems in this dataset.

The MERIT-BASINS dataset (Lin et al., 2019) was used to define the boundaries of individual basins. Each individual basin in MERIT-BASINS carried a unique basin identifier and an upstream area attribute that directly corresponded to a specific river reach in the MERIT-SWORD network.

Discharge data required to force the SFINCS model at predefined discharge points (headwater and inflow points) was obtained from three sources, two hydrological models and a database with observed discharges. The primary global dataset was GloFAS (Grimaldi et al., 2022), which provided long-term global coverage discharge data suitable for simulating floods at a global scale. The second source was the GEB model (De Bruijn et al., 2023), which generates higher-resolution behaviourally informed discharge estimates that are particularly useful in basins where human interventions significantly influence flood dynamics. Finally, observed discharge records from the USGS database (USGS, 2023) were used as a benchmark to evaluate the performance of the model predictions in selected US basins.

The GloFAS dataset (Alfieri et al., 2013; Grimaldi et al., 2022) is generated using the global hydrological model LISFLOOD-OS, which is a distributed, physically based rainfall-runoff model that uses the ERA5 reanalysis data as input to simulate global river discharge at ∼5.5 km resolution at the equator. GloFAS discharge is primarily designed for river basins larger than ∼500 km2 (Grimaldi et al., 2022). In this study, it is applied across a wider range of basin sizes to assess how model performance varies with basin scale when using globally available discharge datasets, including smaller basins. This allows us to evaluate the behaviour of the modelling framework under conditions where the input data may be less reliable. Because the timing of the peak discharges in GloFAS hydrographs can differ from the timing of the floods observed in the satellite-derived extent maps (Sect. 2.4), we captured GloFAS data 10 d before and 10 d after the timing of the flood event in the observed data. In this way, we are sure the flood event (i.e. discharge peak) was in the GloFAS dataset. The GloFAS dataset was used as the default dataset for discharge in the global automated setup. The GloFAS dataset was selected because it offers long-term, global coverage of discharge data, which made it suitable for simulating floods on a large scale. The following mentioned datasets were used in the sensitivity analysis for this study (Sect. 2.5):

The GEB model (De Bruijn et al., 2023) integrates an agent-based model (ABM) and a hydrological model to simulate the flood and drought management decisions of farmers and urban households interacting with the hydrological system. Within GEB, the ABM is dynamically linked with the spatially distributed grid-based hydrological model CWatM at ∼1 km resolution at the equator (Burek et al., 2020). The GEB model incorporates not only surface and groundwater hydrological processes but also the impacts of human activities like water consumption and reservoir operations. We used the daily discharge estimates from the GEB model as input to SFINCS during the sensitivity analysis (Sect. 2.5). The GEB model was selected because it captures socio-hydrological interactions and provides high-resolution, behaviourally informed discharge estimates, which are particularly useful in basins where human interventions significantly influence flood dynamics, such as the Krishna Basin in India, where it was applied in this study (see Sect. 2.5).

The USGS (2023) provides real-time streamflow data, which we used as a benchmark to evaluate how well the global hydrological model's predicted discharge aligned with actual observations in a selection of US basins (Sect. 2.5). We limited the benchmark to US basins due to the availability of high-quality, long-term, and consistent discharge records from the USGS, which have been globally recognised for their reliability.

Bankfull discharge represents the level at which a river fills its channel without overflowing and is a key threshold in determining the channel dimensions of a river. We followed the approach outlined by Sampson et al. (2015), using bankfull discharge as the primary indicator for estimating the river's cross-sectional shape. To calculate the bankfull discharge, we used the 2-year return period calculated from the GloFAS hydrological model discharge (Sect. 2.2.2). The 2-year return period was selected because it closely corresponds to the bankfull flow for many rivers, representing a typical flow regime that occurs frequently enough to shape the channel's morphology over time (e.g. Edwards et al., 2019). However, recent studies have shown that this threshold can vary significantly across different (sub)basins (Roy and Sinha, 2016). Therefore, we included a sensitivity analysis to assess the impact of this assumption (Sect. 2.5). This allows for modelling of channel dimensions based on recurrent hydrological conditions. We then assumed a rectangular cross-sectional shape for the river channels, which was burned into the DEM (for more detailed information, see Appendix A1).

To estimate the bankfull width, we primarily used values from the MERIT-SWORD database, which integrates observations from multiple global rivers and satellite-related datasets (see Sect. 2.2.1). For the river sections whose values were not represented in MERIT-SWORD, we applied a power-law relationship between bankfull discharge and channel width, as proposed by Leopold and Maddock (1953). This empirical relationship reflects how larger discharges are accommodated by wider channels, while smaller discharges are carried by narrower channels. This approach ensured consistency in width estimation where observational data was lacking and serves as a gap-filling strategy to ensure complete channel representation across the river network.

The bankfull depth was then calculated using Manning's open channel flow equation. This equation accounts for both riverbed roughness and slope. Importantly, the equation also incorporates the bankfull width either observed (e.g. from MERIT-SWORD) or estimated using the power-law relationship ensuring that depth estimates were consistent with the defined channel geometry.

In the sensitivity analysis, we also use the GEB model outputs and the USGS discharge time series to estimate the 2-year return period and the related bathymetry (Sect. 2.5).

We used the HydroMT package (Eilander et al., 2023a) to set up a separate model for each basin cluster across all events. HydroMT handled the automatic data processing and configuration of the SFINCS hydrodynamic model while integrating multiple data sources (e.g. hydrographs, DEMs, and land-use maps). This automation reduced potential errors in manual data processing and enabled the rapid testing of different scenarios. The modular setup of the model allows for further iterative enhancements to the model, ensuring it can effectively adapt and improve as new data is integrated over time. To run both HydroMT and SFINCS for all clusters and events, we further automated our approach using Snakemake (Mölder et al., 2020), a Python-based workflow management system to manage parallel simulations efficiently. Appendix Sect. A2 explains the use of Snakemake and the data handling of our study in more detail.

A validation procedure was designed to validate the simulated flood extents predicted by the SFINCS model. We evaluated whether these simulated flood extents aligned with satellite-derived flood events provided by the Cloud to Street (2022), also known as the Global Flood Database. This database was selected because it offered several key advantages for this type of analysis. Firstly, the database provided global coverage, which ensured that flood events from diverse regions around the world were represented. Secondly, it organised all satellitederived flood events into clear categories of flood drivers. Thirdly, the data was openly accessible and downloadable, which streamlines integration into research studies. Importantly, all the flood maps were consolidated under a single dataset, which eliminated the need to combine multiple data sources.

The database was developed using data from NASA's MODIS satellite which captures daily flood extents documented by the Dartmouth Flood Observatory (DFO) with a spatial resolution of 250 m. MODIS provides consistent global coverage with a daily revisit cycle, making it well-suited for large-scale flood monitoring. The full dataset included 913 observed flood events from 2000 to 2018 (Tellman et al., 2021), which represented the maximum observed surface water extent over the event duration. However, it is important to note that these maximum flood extents may be underestimated due to limitations such as cloud cover, vegetation, or narrow channels that may obscure floodwaters in the satellite imagery (Tellman et al., 2021). To prepare the dataset for validation, we selected events (n=499) from the full database by applying the following filters:

Flood type in observations (heavy rain only). The global flood database lists the main flood drivers, which we filtered only to include floods caused by “Heavy Rain”, which excluded other flood drivers such as tropical storm surge, dam breaks.

We computed the maximum flood extent across all time steps to generate a single inundation map representing the peak of the event for every SFINCS simulation. The observation data only showed if a pixel was flooded or dry (without depth data), so we also changed the SFINCS output to a pixel with only two states: either flooded or dry. In a postprocessing procedure, a minimum depth of 0.05 m was used to classify a pixel as flooded (Wing et al., 2024). Note that we resampled the high-resolution SFINCS outputs of 30 to 250 m using mode resampling to match the resolution of observed flood extents. This approach avoids downscaling the MODIS observations, which would require disaggregation of coarse resolution pixels and could create observed flooded pixels outside the actual flooded areas. By aggregating the model outputs to the resolution of the observations using mode resampling, a consistent comparison between simulated and observed flood extents is ensured. This approach meant that a 250 m SFINCS output pixel was classified as flooded if the majority (i.e. more than one-half) of the underlying SFINCS 30 m cells were flooded. This method ensured consistency with the binary “flooded or dry” classification of MODIS data while reflecting the majority condition in each area. Notably, the permanent water bodies (already 250 m), identified from band 5 of the satellite data, were also masked from the SFINCS simulated flood event after it has been resampled to 250 m. Importantly, the additional downstream basin included in the model domain to avoid boundary effects was not considered when applying the validation criteria. This approach ensured that the validation only assessed the basins directly affected by flooding to provide a fair comparison between the simulated and observed flood extents.

Table 1 shows the validation metrics used to compare SFINCS results with the observations. We validated our maps for each flood event by calculating the contingency table values (i.e. Hit, Miss, False Alarm, and Correct Negative) over the full extent (i.e. all basins) of the event.

The HR measures the proportion of the flood area observed that was successfully predicted by the model, balancing misses (Miss) and hits (Hit). The HR can range from 0 to 1. A higher HR indicates better model accuracy in terms of correctly predicting the flood extent. 1Hit Rate=Hits(Hits+Misses)

The FAR represents the proportion of predicted flood areas that did not flood in satellite-derived flood extents. The FAR can range from 0 to 1. A low FAR suggests that the model is conservative in its flood predictions. 2FAR=FA(FA+Hits)

The CSI balances the HR and FAR in Eqs. (1) and (2) to provide a comprehensive metric of the overall model performance. The CSI can range from 0 to 1, with 1 indicating a perfect model. The CSI accounts for correct predictions (i.e. Hits) and wrong predictions (i.e. Misses and FAs), offering a more nuanced view of model accuracy. 3CSI=Hits(Hits+FA+Misses)

E is a metric used to determine whether a model is overpredicting or underpredicting the occurrence of an event, such as flooding. A value of E=1 indicates no bias, meaning the model's predictions are balanced between overprediction and underprediction. If E falls between 0 and 1, the model tends to underpredict, while values greater than 1 suggest the model overpredicts the event. 4E=FAMisses

To test the effect of our bathymetry estimation method, we used different return periods (1.5-year, 2-year (default in global setup) and 2.5-year) following Andreadis et al. (2013) to estimate bankfull discharge (see Appendix A1) and assess their impact on flood dynamics. We used GloFAS discharge to estimate bankfull discharge for rivers by applying different return periods. For some rivers, bankfull discharge might correspond to a higher or lower return period, so using the same return period for all rivers might not give accurate estimates of river bathymetry and flood extents (Roy and Sinha, 2016; Rad et al., 2024). Therefore, testing multiple return periods provided a better understanding of the river's bathymetry.

Different DEM resolutions (FABDEM 30 m vs. 3D Elevation Product 3DEP 1 m) were compared to assess the impact of topographic resolution on flood modelling. Higher-resolution DEMs often provide more accurate flood predictions, particularly in areas with complex terrain (Jiang et al., 2022).

The model's sensitivity to discharge inputs was tested by forcing the model with USGS observation discharge instead of simulated discharge from GloFAS. Thus, the actual event forcing and river channel size estimates were updated based on the observed discharge. Accurate river discharge input is crucial for hydrodynamic modelling, as it directly influences flood predictions (Zhou et al., 2022). This comparison between modelled (GloFAS, default input) and observed (USGS) discharge allowed for disentangling the source of the modelling errors from either the inflow or the processes within SFINCS. USGS River discharge data were filtered for locations containing data for more than 30 years to ensure that return periods and thus bathymetry was more accurately predicted.

Furthermore, to ensure robustness and avoid reliance on a single hydrological model, we also forced SFINCS with the GEB model discharge (Sect. 2.2.2). However, GEB is currently not available as a global dataset and is still under development for large-scale applications. As a result, its use in this study was limited to a subset of case studies rather than the full global event set. This comparison was designed to test the sensitivity of SFINCS to hydrological input characteristics since GloFAS and the GEB model fundamentally differed in model resolution. The GEB model discharges were specifically tested in the Krishna Basin in India (11 riverine events), a large and significant river basin known for its significance due to its size being nearly 8 % of the total geographical area of India and extensive agricultural and hydrological use. The GEB discharge outputs were used to drive the flood simulations and derive river channel dimensions. The resulting inundation extents were compared against the Cloud to Street database to assess model accuracy. This dual-model comparison between GloFAS and GEB provided insights into variability and reliability across different hydrological sources with different resolutions, thereby strengthening the analysis of SFINCS flood simulations.

Figure 5a shows that in regions with a good agreement between the SFINCS simulation and the satellite observations (Event DFO_4075; CSI=0.55), the resampling of a higher-resolution SFINCS output (30 m) to match the observations (250 m) introduced rough edges where the simulation could not match the boundaries of the observed inundation extent. These rough edges produced misses (red areas in Fig. 5a) around the boundaries, even when the true flood extent lay just inside the 250 m cell.

Figure 5b (Event DFO_1996; CSI=0.48) showed inconsistencies (i.e. FAs) occurring in the headwater rivers. These inconsistencies could have arisen from several factors, such as the snapping procedure (Sect. 2.1.3). For example, in headwater locations, the coarse GloFAS grid (∼5 km) often misaligned with the finer SFINCS grid, so even with a 5 %–10 % snapping tolerance (see Sect. 2.1), small discharge mismatches were introduced into headwater tributaries, which in these narrow channels could produce localised FAs (yellow pixels in Fig. 5b) in areas that did not flood. Furthermore, we did not have width observations, so headwater river widths were estimated from the power-law equation (Sect. 2.2.3). This approach can sometimes under- or overestimate channel widths in smaller headwater rivers. If the equation gives too-narrow widths, conveyance capacity is underrepresented, which forces excess water onto adjacent floodplains (causing false alarms) (Dey et al., 2022). Conversely, if the equation gives too-wide widths, the simulated flood may be unrealistically confined to the channel and lead to misses.

Figure 5c depicts a region (Event DFO_4140; CSI=0.178) where permanent water was not well represented in the underlying observation layer of the Cloud to Street dataset (i.e. band 5). Although both the model and the satellite-derived flood extent relied on the JRC Global Surface Water dataset as a permanent water mask, the Cloud to Street product provided a 250 m resampled version (Tellman et al., 2021). When rivers meander or curve and are narrower than a MODIS pixel, the resampling process causes permanent open-water bodies to disappear. As a result, channels that are correctly simulated as inundated by SFINCS are not flagged as water in the observation layer and are instead counted as FAs during validation. This effect was particularly visible in low-order rivers where the channel was below 250 m wide. In these cases, the issue was not that the model overpredicted flooding but that the validation mask failed to represent permanent water bodies.

In Fig. 5d, the region's (Event DFO_4451; CSI=0.177) lower performance was due to only one discharge point (i.e. inflow point) carrying water into the model domain. The other discharge points (i.e. headwater points) remained dry (see Appendix A, Fig. A1), which was likely caused by the coarse resolution (∼5.5 km2) of the GloFAS input discharge data. Thus, the model failed to capture peak flows in smaller upstream tributaries (Alfieri et al., 2013; Grimaldi et al., 2022), which created a significant gap in the simulated flood where the satellite observations clearly showed inundation.

We conducted a two-part sensitivity analysis in two global regions (i.e. India and the US) to evaluate the role of hydrological forcing in our global flood modelling framework. We investigated (a) the differences in model performance when using two different global hydrological models (i.e. GloFAS and GEB) for 11 Events in India and (b) the influence of modelled GloFAS discharges compared to observed discharges for ten basins in the US.

To compare the influence of using two different global hydrological models (GLOFAS and GEB) as input to SFINCS, we focused on the Krishna Basin in India. Here, we simulated floods for 11 discrete events with both models. Bathymetric calculations were done separately, using both GloFAS and GEB long-term discharges (>40 years) to compare all aspects of the two models. Using the GloFAS model discharge as forcing for SFINCS resulted in an average CSI of 0.34, while forcing with the GEB model led to an average CSI of 0.38.

One critical distinction between the two models lies in their spatial resolution. The GloFAS operates at a coarser grid resolution (∼5.5 km), while the GEB provides discharge data at a finer grid resolution (30′′; ∼1 km). An example basin cluster within the Krishna basin is shown in Fig. 6 (Event 3551). The hydrographs on the right (Fig. 6) show the discharge forcings from both hydrological models at six main discharge points of the SFINCS simulation.

Notably, the GEB hydrographs exhibited a more complete flood hydrograph shape, especially toward the later stages of the event (Fig. 6). In contrast, the GloFAS hydrographs of some of the inflow points did not capture a high peak flow, suggesting that peak flows could be underrepresented. This issue directly affected the resulting flood maps (Fig. 6a and b, left panels), where simulations using the GEB input captured more hits than those driven by the GloFAS.

The improved performance of the GEB-SFINCS chain may be attributed to its higher spatial resolution, which allowed for a more accurate representation of smaller headwater rivers and their channel dimensions. In contrast, GloFAS lacks the ability to resolve these smaller streams directly, so bathymetric calculations relied on the nearest available grid point, which might not have aligned with the true upstream location. With the GEB, however, the finer resolution enabled matching bathymetry calculations to the correct upstream grid pixel, which resulted in more accurate low flows and a better representation of channel dimensions in smaller rivers. By resolving finer spatial scales, higher-resolution models like the GEB reduce uncertainties associated with inflow dynamics and provide more reliable input for hydrodynamic models like SFINCS. Moreover, the improved representation of human behaviour and reservoir management in the GEB can lead to more realistic streamflow estimates, which in turn enhances the accuracy of the flood simulations. This outcome underscores the importance of investing in higher-resolution global datasets to improve flood modelling accuracy in river systems.

This section describes our assessment of the performance of our modelling setup by comparing modelled discharge from GloFAS with observed discharge data from the USGS (2023). The analysis evaluated model performance using event-based hydrographs from the GloFAS and USGS. Because this analysis could only be applied in regions with long-term discharge records (e.g. for bathymetry estimation), only ten distinct basin clusters remained suitable.

The results revealed that the model using observed USGS discharge data considerably outperformed the one using GloFAS discharge data by achieving an average CSI of 0.67 compared to the average global mean of 0.39 (Table 2). This significant difference shows the importance of accurate discharge data for enhancing model accuracy while highlighting the limitations of global hydrological models. Figure 7a and b provide insights into one such high-performing basin forced with USGS observations (CSI=0.78), while the same basin forced with GloFAS discharge yielded a lower CSI of 0.56. Nonetheless, a considerable number of misses occurred, even when observed discharge data were used. These misses can be explained by the resampling causing the rough edges. Furthermore, in the northeastern part of the basin (Fig. 7b), a side stream showed underprediction (i.e. misses), which was likely due to the absence of a river gauge in that tributary. Thus, to better represent this basin, an additional discharge input should be provided to reflect the flood extent accurately. A disadvantage of working with observational river gauges is the potential for gaps in spatial coverage, such as the side stream shown in Fig. 7b, which can limit model performance.

Figure 7a shows the hydrograph during the flood event for both modelled (i.e. GloFAS) and observed (i.e. USGS) discharge data. The GloFAS hydrograph failed to capture the peak discharge during a flood event, which contributed to its lower performance (CSI=0.56) compared to the USGS-based forcing. The inability to capture discharge peaks decreased the precision of the simulated flood extents during high-flow conditions. Harrigan et al. (2020) showed that river discharge is negatively biased in 64 % of the basins globally in the GloFAS archive. Moreover, they reported that GloFAS-ERA5 discharge skill, measured by the Kling–Gupta efficiency (KGE), is strongly catchment-size-dependent: median KGE is only 0.21 for basins <10000 km2 (e.g. Fig. 7b basin has a catchment size of ∼500 km2), which rises to 0.56 for basins >50000 km2. This failure to capture discharge peaks in smaller basins has a more immediate and direct impact on flood extent simulations and can reduce the performance of the flood simulations.

A fixed 2-year return period for bathymetric calculations is commonly used as a proxy for bankfull discharge, but this approach may not be universally applicable (Andreadis et al., 2013). Adjusting this parameter based on the characteristics of specific river systems could improve the accuracy of bathymetric estimates and better reflect actual flood behaviour (Roy and Sinha, 2016). Thus, this section presents our evaluation of how model performance was affected when different return periods were used to derive bankfull discharge in the SFINCS fluvial setup. We compared GloFAS-derived discharge with observed USGS river gauge data for ten US basins to assess the accuracy of bathymetry estimates.

The bankfull discharge was calculated via long-term yearly maxima peaks extracted from the GloFAS hydrological model and was evaluated against observed data from USGS river gauges. The results showed that the modelled bankfull discharge (Qbf) was consistently underestimated across all tested return periods.

Figure 8 shows the simulated Qbf (2-year RP) against USGS-derived Qbf on a logarithmic scale. All points lay below the 1:1 line, which indicated that the GloFAS-based RP were systematically lower, particularly in the lower quartile of values, where smaller bankfull discharges were more strongly biased. This pattern indicates that GloFAS underrepresents the frequency and magnitude of the low-flow discharges that influence bankfull conditions. As a result, bathymetry derived from GloFAS tends to yield narrower and shallower channels, which reduces modelled conveyance capacity and potentially produces overestimated flood extents (i.e. higher FAs). These results imply that (i) using the 2-year default alone may not capture observed bankfull behaviour in the current fluvial setup and (ii) alternative approaches (e.g. gauge-based bankfull estimates, where available) can improve bathymetric realism (Zarrabi et al., 2025; Rad et al., 2024). Appendix Figs. A3 and A4 present results for the 1.5-year and 2.5-year return periods, respectively.

This section explores how different DEMs were found to influence the performance of the SFINCS model in simulating flood extents. All simulations in this section are forced using GloFAS discharge, while only the DEM input is varied. We focused on two open-source DEMs with different spatial resolutions: FABDEM (30 m; Hawker et al., 2022) and 3D Elevation Program (3DEP; 1 m; USGS, 2015). The resolution of the SFINCS model was kept the same between the two different inputs, although we increased the level of detail (i.e. hypsometry levels, n=20, default 10) stored in the subgrid tables when using the 3DEP DEM. Both DEM inputs (FABDEM at 30 m and 3DEP at 1 m) are internally reprojected to the same SFINCS model grid resolution using hydromt. Therefore, the key difference lies not in the model grid resolution itself, but in the richness of the subgrid representation: higher-resolution DEMs provide more detailed elevation information within each grid cell, which is captured through the subgrid hypsometry tables. The same ten sub-basins in the United States were selected for this analysis based on the availability of 3DEP. However, vertical inconsistencies were observed in the original 3DEP dataset, where merging adjacent tiles sometimes introduced elevation differences of up to 3 metres. As a result, only six sub-basins were used in the final performance comparison. When using FABDEM, the average CSI was 0.37 across the six sub-basins. The mean CSI increased to 0.57 when using 3DEP DEM, which highlights the positive effect of higher-resolution elevation data on model accuracy. This improvement likely resulted from the finer spatial resolution of 3DEP and higher vertical accuracy. Moreover, 3DEP's 1 m LiDAR-derived DEM achieved a root mean square error of 0.53 m (USGS, 2022) compared to FABDEM's mean absolute vertical error reductions from 1.61 to 1.12 m in built-up areas and from 5.15 to 2.88 m in forests (Hawker et al., 2022).

Figure 9 illustrates the model's performance with these two DEMs in one of the six basins. Figure 9b shows the FABDEM dataset results, where the CSI was 0.46 for this particular basin. In this case, the lower-resolution DEM resulted in poorly defined river channels, especially in side streams, which prevented water from flowing accurately to downstream areas of the basin. In the upstream part of this basin, FAs were seen where water accumulated but could not reach the downstream areas. In contrast, when the 3DEP DEM was used (Fig. 9a), the model's performance improved significantly. A CSI of 0.59 showed a much better flow representation (i.e. fewer FAs in the upstream part), as the higher-resolution DEM allowed water to reach the downstream parts of the catchment. The increased detail in the terrain allowed for a more accurate representation of floodplains and channels (Jiang et al., 2022), which improved the HR while reducing the number of misses (red pixels) and FAs (yellow pixels).

Our study is unique in the high number (n=499) of events covered, which is much higher than in existing studies covering approximately 2 to 35 events (Risling et al., 2024; Wing et al., 2021; Bernhofen et al., 2018). Moreover, our study includes small upstream basins (<50 km2) to address the critical need to understand local upstream floods better. Hence, our model's performance has been assessed across a more diverse hydrological and geographical setting. Notably, we simulated multiple events with different timings and intensities in the same basin, which can be quite challenging to explain performance (Wing et al., 2021).

The accuracy of hydrodynamic flood models (e.g. SFINCS) depends on the quality and characteristics of the hydrological drivers used as input. Previous studies have shown that the choice of hydrological forcing can strongly influence simulated flood extents. For example, a study by Mester et al. (2021) assessed the sensitivity of hydrodynamic models to hydrological forcings using ten different global hydrological models and eight case study areas with observed flood extent values. These basins were selected based on their relatively large size, as it was assumed that the relatively coarse hydrological models would not perform well for smaller sub-basins. Mester et al. found that the agreement between simulated and observed flood extents varied significantly across models and climate forcings. Moreover, Wing et al. (2021) found a CSI of 0.87 but used only observed hydrological forcing for their hydrodynamic model. It is well-known that global hydrological models perform poorly in smaller upstream basins due to their coarse resolution (e.g. Salinas et al., 2013; Mester et al., 2021), which could partly explain the poorer performance of the SFINCS model in these areas. We saw higher performance with SFINCS when it was supplied with observed discharges (CSI=0.67). The inability of the GloFAS hydrological model to capture the discharge peaks during a flood can significantly impact performance (Figs. 5d and 7a). This uncertainty in event discharge significantly impacts flood extent and likely explains much of the reduced CSI (0.39 globally; 0.67 when using USGS observations).

In our current setup, we saw that the bathymetry and river channel dimensions were underestimated (Sect. 3.6.2), which resulted in higher FAs because channel conveyance capacity was not represented correctly. The SWOT satellite mission provides high-accuracy measurements of water surface elevation, river width, and slope for rivers greater than 100 m in width (Neal et al., 2021; Larnier et al., 2020). Incorporating this dataset may improve flood modelling accuracy, particularly for river systems that have no data. In addition, using the “gradually varying solver” method to estimate river channel capacity, can replace the traditional Manning's equation. This method can also significantly improve bathymetry and flood inundation simulations (Wing et al., 2024). Neal et al. (2021) demonstrated the benefits of this approach in localised studies, and its application at the global scale may yield more accurate flood predictions.

Another difference in our approach is the exclusion of permanent water bodies from the observed and simulated flood extent (250 m resolution), with no additional buffer applied around the rivers. This approach contrasts with methods used in other large-scale validation studies, such as Wing et al. (2017) and Dottori et al. (2022), which applied spatial buffers around the rivers to better capture floodplains and address benchmark data limitations. For example, Wing et al. (2017) applied a fixed buffer (i.e. ∼1 km around the rivers) to include areas potentially missed by benchmark data, particularly in small tributaries. Dottori et al. (2022) employed variable buffer zones (i.e. ∼5 and ∼10 km) tailored to account for diverse floodplain morphologies and the variable extent and density of the mapped river network. In our validation, we excluded permanent water using satellite-derived flood extent, so correctly simulated inundation in meandering and narrow channels (<250 m wide) was instead counted as an FA, even though these areas should be masked as permanent water. This misclassification inflated the FA ratio while decreasing the HR, which contributed to a lower average global CSI.

As Tellman et al. (2021) noted, MODIS often fails to capture floods in rapid flash-flood events or under dense canopy cover, which leads to underestimation of true flood extents and contributes to lower CSI values. Additionally, cloud and snow cover can further obscure floodwaters, which adds to observational uncertainty. Some northern latitudes can have errors greater than 65 % in the flood detection algorithm due to the low sun angle on dark soil, which causes low reflectance that mimics water.

The comprehensive validation dataset relies on 250 m MODIS pixels, which limits the detection of narrow or small-scale inundation features and can misclassify flood zone delineation (Landwehr et al., 2024). Moreover, most binary pattern matching metrics are sensitive to the proportion of flooded area, meaning that large-scale floods are favoured while smaller-scale floods are less accurately reflected in the validation scores (Landwehr et al., 2024). Binary flood extent masks and class-based metrics (e.g. the CSI) do not account for important factors such as flood depth and its influence on impacts and model accuracy (Stephens et al., 2014). As a result, some of the poorer performance in smaller basins likely reflects limitations in the reference dataset and validation framework rather than shortcomings in the hydrodynamic model.

Although studies on global flood protection standards exist (e.g. Scussolini et al., 2016), incorporating more accurate data on flood protection standards into future global flood models should be considered. Previous research, including studies by Mester et al. (2021), has shown that flood protection can influence model performance by increasing variability, though it does not necessarily change the maximum performance scores. Further exploration into this area may provide valuable insights into how flood protection measures impact flood risk across regions.