As floods are the most significant and costly natural hazard, efforts are underway around the world to improve how we assess, forecast, and map floods and their impacts (Barendrecht et al., 2017). Hydraulic modelling is the most widely used method for flood mapping to facilitate risk assessment (Teng et al., 2017). A hydraulic model allows the simulation of the spatio-temporal evolution of a river's hydraulic variables such as water surface elevation (WSE) and flow velocity, as well as the flooded area, through the application of mass and momentum conservation equations. It requires information on discharge, boundary conditions of the study domain, and characteristics of the area under study (river shape, roughness). The validation of a hydraulic model mainly depends on the quantity, quality, and type of observed WSE. However, hydraulic models require extensive data, which poses significant challenges (Teng et al., 2017). Given the decline in hydrometric stations (Grimaldi et al., 2016; Mishra and Coulibaly, 2009), their limited spatial and temporal coverage and the limited accessibility of measurement sites, especially during flood periods, observations are not always available or sufficient, which complicates the validation of hydraulic models and affects the reliability of simulations.

In this context, the integration of remotely sensed variables such as WSE (Domeneghetti et al., 2021) and flood extents (Wood et al., 2016) into flood hydraulic modelling could potentially provide valuable information for model validation. Flood extents are usually derived from optical (Huang et al., 2018) or synthetic aperture radar (SAR) sensors (Landuyt et al., 2018). Radar can measure regardless of illumination or weather conditions, whereas optical sensors cannot. This can significantly influence the ability to continuously monitor water, particularly during floods. By overlaying the water extent extracted onto a digital elevation model (DEM), it is possible to retrieve the WSE (Grimaldi et al., 2016), although the accuracy of the WSE value is affected by the resolution of the DEM. Laser and radar altimeters have also been used to measure river WSEs directly (Cretaux et al., 2018). Nielsen et al. (2022) worked with data from several altimetric missions in rivers with widths ranging from a few hundred meters to around 3 km, whereas Li et al. (2023) evaluated how well ICESat-2 laser altimetry could measure WSE in rivers ranging from medium-width (around 1000 m wide) to narrower ones (under 50 m). These findings revealed that, despite advances, getting reliable measurements for rivers narrower than 30 m and sometimes even those up to 90 m wide, remains challenging (Li et al., 2023).

The Surface Water and Ocean Topography (SWOT) mission, led by the National Aeronautics and Space Administration (NASA) and the Centre National d'Études Spatiales (CNES), in collaboration with the Canadian Space Agency (CSA) and the UK Space Agency (UKSA), was launched in December 2022 to address these gaps (Biancamaria et al., 2016; Fu et al., 2024). The Ka-band Radar Interferometer (KaRIn), combined with SWOT's low near-nadir incidence angles, enables the satellite to observe at least 90 % of the world's rivers wider than 50–100 m, as well as lakes, reservoirs, and wetlands larger than 250 m × 250 m (Biancamaria et al., 2016). The expected vertical accuracy of WSE is 10 cm when pixels are averaged over 1 km² (Peral et al., 2024). The satellite was initially deployed on a 1-day calibration/validation (CAL/VAL) orbit covering only a certain portion of the Earth between April and July 2023. It was then moved to its nominal 21-day orbit.

To study the potential of the SWOT mission to monitor floods with its 21-day orbit, some studies have been conducted using synthetic SWOT data generated by the CNES large-scale hydrology simulator (SWOT simulator) prior to the mission launch. Frasson et al. (2019) estimated that SWOT would have provided at least one measurement of 55 % of the floods recorded between 1985 and 2018 by the Dartmouth Flood Observatory. According to Frasson et al. (2019), SWOT's ability to observe floods mainly depends on the site's latitude and the duration of the event. Indeed, long-duration floods are more likely to be observed more than once. Sites located between 20° S and 20° N will be observed once or twice per 21-day cycle, while higher latitudes will generally be observed two or more times per cycle. Following the satellite's launch, Laipelt et al. (2025) confirm the usefulness of SWOT data for flood studies with a very strong relationship (R2= 0.99) between actual SWOT observations of WSE variations and in situ measurements during an extreme flood that occurred in southern Brazil in 2024.

Subsequent validation studies have shown that the reported accuracy of SWOT WSE measurements varies with the study region, river width, and product type (PIXC, RiverSP node, and RiverSP reach), with errors decreasing because of spatial averaging. At the global scale, based on a worldwide evaluation including rivers of diverse width and hydrological conditions and using satellite-based WSE reference datasets such as Hydroweb (https://www.theia-land.fr/blog/product/hauteur-des-lacs-et-rivieres/, last access: 4 June 2026) and G-REALM (https://ipad.fas.usda.gov/cropexplorer/global_reservoir/, last access: 4 June 2026), the average measurement error is estimated to be below 0.15 m (Yu et al., 2024), which is consistent with the expected order of magnitude of the mission's performance. At the subcontinental scale, validation over India based on more than 14 000 observations across 419 hydrometrics stations showed relative errors of about 0.18 m for rivers wider than 100 m and about 0.26 m for narrower rivers, based on node-level evaluations, highlighting a clear dependence of SWOT WSE accuracy on river width (Patidar et al., 2025). These results consistently indicate that node-level SWOT WSE accuracy decreases for narrower rivers, while remaining within the range of typical in situ measurement uncertainties. At finer spatial scales, a watershed-scale assessment evaluated satellite performance using node-level WSE data. The SWOT Science Requirements are defined for reach-scale products, whereas the present assessment relies on node-level observations. Comparison with the mission requirements is therefore not appropriate. During the CAL/VAL orbit, validation against United States Geological Survey (USGS) gauging records, based on 39 stations, showed a median unbiased Root Mean Square Error (ubRMSE) of about 0.25 m, even for rivers narrower than 100 m and in some cases below 50 m (Jiang et al., 2025). This level of agreement is consistent with typical uncertainties in ground-based WSE measurements used in flood applications in the United States, which are commonly on the order of 0.15 to 0.30 m according to USGS guidelines (NOAA, 2011). In contrast to global-scale evaluations based on reach-averaged products, the present analysis relies on node-level WSEs, for which error metrics are expected to be larger due to the averaging over a smaller number of pixels. Nevertheless, studies specifically addressing SWOT performance in narrow rivers remain limited, particularly during a flood event. Moreover, most existing assessments rely on earlier versions of the SWOT products (Version C), while a more recent Version D is now available. These assessments primarily evaluate relative errors, in which potential systematic biases between SWOT WSE and reference observations are removed.

In this context, the present study investigates the ability of SWOT observations to support the analysis of hydraulic consistency and the calibration and validation of hydraulic models in a narrow river, using the most recent RiverSP Version D product. The analysis focuses on the Du Gouffre River (Quebec, Canada; ≈ 40 m wide), which experienced a major flood in May 2023 with an estimated return period of 60 years, based on three-hour averaged flows (Gouvernement du Québec, 2023), and for which a LISFLOOD-FP model (Bates and De Roo, 2000) was available. This rare and extreme event offers a valuable opportunity to assess SWOT performance under flood conditions in a narrow river through a concrete hydraulic case study. The Du Gouffre River sector was covered by the SWOT CAL/VAL orbit, enabling daily observations of the flood event. Unlike most previous studies, the evaluation is based on the smoothed WSE variable (wse_sm) from Version D and relies on a direct comparison of absolute WSEs, ensured through a consistent geoid conversion between SWOT observations and model outputs.

The extracted WSEs from SWOT products (PGD0) were first compared with measurements (at one-minute intervals) from tide gauge station 03057 (Saint-Joseph-de-la-Rive, Fig. 1A) for all satellite passes between April 2023 and August 2025, during ice-free period. Figure 3a shows the probability that the absolute WSE error is less than or equal to a given value (Likelihood of occurrence), indicating that the 68th percentile (corresponding to the 1-sigma metric defined in the SWOT requirements) is 0.24 m, with an RMSE of 0.25 m. The boxplot (Fig. 3b) complements this probabilistic analysis by illustrating the distribution of absolute WSE errors (m), with a median of 0.16 m. It shows a low median and a relatively limited interquartile range, indicating that the errors are well centred and not dominated by systematic bias or by extreme deviations in the error distribution. The SWOT observations used are of good quality (node_q < 3 et xovr_cal_q < 2). The obtained values meet the WSE accuracy requirement defined by the SWOT mission, which specifies a 1-sigma accuracy of 0.25 m or better for water bodies whose surface area is between 250 m² and 1 km² (Desai, 2018). At the spatial scale relevant to the Du Gouffre River, the 68th-percentile absolute error of 0.24 m therefore falls within the applicable 0.25 m threshold.

The consistency of the SWOT WSEs with LISFLOOD-FP simulations was assessed by comparing the corresponding nodes, which represents the core of this study. SWOT observations were thus compared with the results of the LISFLOOD-FP hydraulic model for eleven observation cycles between 25 April and 7 May 2023 (Table 2). The differences between the two sources remain relatively small, with bias values ranging from -0.30 to 0.43 m and RMSE values between 0.22 and 0.54 m. Detailed bias and RMSE values are shown in Table 2.

The observed flood is covered by cycles 508, 509, and 510, with cycle 508 corresponding to the flood peak on 1 May. For cycle 508, a comparison between SWOT-derived WSEs and LISFLOOD-FP simulations prior to any adjustment of the Des Mares tributary inflow shows a bias of 1.30 m and an RMSE of -1.21 m. To investigate the origin of this discrepancy, a discharge adjustment was performed for the Des Mares tributary using SWOT-derived WSEs. Increasing the tributary inflow to 240 m³ s⁻¹ results in a bias of -0.30 m and an RMSE of 0.54 m (Fig. 4). The comparison before and after the adjustment highlights an improvement in model-SWOT consistency during the flood peak. Because SWOT observations are used to inform the tributary discharge adjustment for this cycle, this comparison is not interpreted as an independent validation of SWOT measurement accuracy, but rather as an assessment of consistency between the adjusted hydraulic simulation and SWOT-derived WSEs. For cycle 508, no discharge uncertainty is considered, as the value of 240 m³ s⁻¹ corresponds to the calibrated flow used in the simulation.

The following cycle, 509, shows a good agreement between the two sources. The SWOT points generally align well with the simulated WSEs (Fig. 5). The bias is -0.10 m and the RMSE is 0.22 m. Unlike cycle 508, the discharges for this overpass were well constrained and did not involve significant tributary inputs, which likely contributed to the good consistency between the two datasets.

SWOT data corresponding to cycle 510 could not be used. In this case, all nodes were associated with a xovr_cal_q value equal to 2, indicating a cross-over calibration that was considered unreliable. By applying the quality filters defined above, all these observations were discarded.

Cycles 502, 503, 506 and 507 precede the flood peak. The discrepancies between SWOT and the model are slightly more pronounced than those observed during the recession phase. Biases range from 0.28 to 0.43 m, and RMSEs from 0.47 to 0.53 m. During the recession phase, covered by cycles 511, 512, 513 and 514, the discrepancies are smaller; the biases do not exceed 0.33 m and the RMSE values range are from 0.29 to 0.40 m.

The interpretation of the results must be placed within the context of limited ground-based observations available along the studied reach, corresponding to the SWOT RiverSP reaches 7240900665 and 7240900285, which constitute the focus of this analysis. Aside from a downstream tidal gauge and a single upstream hydrometric station, no spatially distributed in situ WSE measurements were available during the flood event. Despite this limitation, preliminary comparisons provide useful independent benchmarks that help contextualize the analysis. First, the comparison between SWOT-derived WSEs and one-minute observations at the downstream tidal gauge shows good agreement, with a 68th percentile absolute error of 0.05 m and an RMSE of 0.24 m (Fig. 3), consistent with the SWOT mission accuracy requirements. In addition, an upstream comparison between LISFLOOD-FP simulations and in situ WSE at the Saint-Urbain hydrometric station yield a bias of 0.07 m and an RMSE of 0.12 m for all cycles except the flood peak, supporting the overall consistency of the hydraulic model outside peak-flow conditions.

Within this constrained observational context, the present study is therefore framed as a site-specific case study that explores the practical use of SWOT observations to support hydraulic analysis during a rare and extreme flood event, rather than as a rigorous assessment of SWOT measurement accuracy. The following discussion focuses on how SWOT-derived WSEs can complement sparse in situ data and contribute to the evaluation of hydraulic consistency and discharge dynamics in a narrow river under extreme flow conditions.

The results indicate a general consistency between SWOT observations and the WSEs simulated by the LISFLOOD-FP model across most of the analyzed cycles. The RMSE values, ranging from 0.22 to 0.54 m, with biases between -0.30 and 0.43 m were obtained, for river reaches with an average width of approximately 40 m. This width is significantly smaller than those investigated in most previous SWOT and radar altimetry validation studies (Table 3), which typically focused on rivers wider than 100 m and, in some cases, several kilometers wide. As reported by Domeneghetti et al. (2018), the performance of satellite-derived WSEs generally decreases as river width decreases, highlighting the challenging nature of the present case study.

When interpreted relative to the hydrological signal of the event, the magnitude of the errors remains moderate. During the flood, WSE varied by approximately 2 m along the studied reach. In this context, the maximum RMSE of 0.54 m represents about 27 % of the flood-stage amplitude, while the maximum bias of 0.42 m corresponds to approximately 22 %. These values therefore remain limited when compared to the overall flood dynamics.

Beyond this event-specific perspective, the obtained WSE differences are consistent with those commonly reported in the literature for satellite-based river WSE observations and their application in hydraulic modelling. A selection of relevant studies summarized in Table 3 shows that satellite-derived river WSE errors, as quantified by RMSE, generally remain hydraulically informative for model evaluation and calibration across a range of observational accuracies.

For instance, Villadsen et al. (2016) synthesized past satellite missions performance and proposed accuracy thresholds whereby errors below 0.30 m are considered “good” and values between 0.30 and 0.60 m considered “moderate”. These criteria were subsequently adopted by Kittel et al. (2021b), who relied on the same benchmarks to assess the usefulness of satellite-derived WSEs for large-scale hydraulic modelling. Shen et al. (2020) later applied this framework in a river case study to interpret RMSE values and demonstrated that satellite-based WSE observations with good to moderate accuracy can effectively support hydraulic model evaluation and calibration. More recently, Zhou et al. (2023) showed that ICESat-2 and Sentinel-2-derived WSEs with RMSE values ranging from 0.25 to 0.59 m enabled successful hydraulic model calibration, yielding validation RMSE values of approximately 0.36 m.

Importantly, the present results were obtained during an extreme flood event, a period during which uncertainties related to floodplain roughness (Pappenberger et al., 2005), inundation extent (Bates et al., 2014; Aitken et al., 2024), and channel-floodplain interactions (Croke et al., 2013) are expected to be highest. These challenges are further exacerbated in narrow rivers, where satellite altimetry measurements are inherently more difficult to interpret.

Achieving WSE differences comparable to those reported in the literature under such conditions, and for a narrow river under the nominal detection limits of SWOT, further supports the hydraulic relevance of SWOT observations in this context. Overall, these findings indicate that, for the present case study, SWOT-derived WSEs provide hydraulically meaningful information for the evaluation of flood dynamics and hydraulic model performance, even in a narrow river and under high-flow conditions.

During the flood peak on 1 May (cycle 508), the initial comparison between SWOT-derived WSEs and LISFLOOD-FP simulations showed a pronounced discrepancy. After adjusting the inflow of the Des Mares tributary, the agreement improved, yielding a bias of -0.30 m and an RMSE of 0.54 m. Compared with the pre-adjustment simulations, this result indicates an improved consistency between the simulated WSEs and the SWOT observations during peak-flow conditions. As the tributary inflow is adjusted using SWOT-derived WSEs for this cycle, the resulting improvement should not be interpreted as an independent assessment of SWOT measurement accuracy. Rather, within the context of this site-specific case study, it illustrates how SWOT observations can provide complementary information to diagnose discharge underestimation in a poorly gauged tributary and to enhance the internal consistency of hydraulic model inputs during an extreme flood event. Similar conclusions were drawn by Diouf et al. (2025), who showed that SWOT observations (using the L2_HR_PIXC product) can support hydraulic model calibration in poorly instrumented environments by constraining hydraulic parameters and improving the representation of WSE dynamics in areas lacking in situ measurements.

At the same time, the need for such adjustment underscores the inherent challenge of determining an accurate discharge representation in hydraulic models under conditions of spatial rainfall variability. Spatial variability in rainfall strongly influences flood runoff, they can alter the volume, timing, and peak flow of the hydrograph (Khakbaz et al., 2012). Discharge estimates that may become available in future versions of SWOT products (and already for some reaches in version D) will be particularly valuable in such situations, as they could help refine the discharge inputs used in hydraulic models.

Additional qualitative validation elements confirm the reliability of SWOT observations during the flood peak. Drone imagery acquired on 1 May (cycle 508) (Fig. A1 in the Appendix) illustrates the extent of the flooding observed in the field. The extent of the flooded areas visible in these images is consistent with the WSEs measured by the satellite, thus providing an independent validation for the SWOT observations.

This agreement is further supported by the analysis of two internal quality indicators provided in the SWOT River SP product (Fig. 6). The first, wse_u (Fig. 6A), represents the total uncertainty (random and systematic) of the WSE at each node (in meters), while the second, dark_frac (Fig. 6B), corresponds to the fraction of “dark water” pixels used in the retrieval, expressed as a unitless ratio between 0 and 1. Dark water results from signal attenuation, which can be caused by rainfall or by a smooth water surface (e.g., in the absence of wind). This raises the question of whether dark water occurs more frequently during flood events. For cycle 508, most wse_u values are tightly clustered between 0.09 and 0.24 m, with a median around 0.13 m, indicating a generally low total uncertainty for most nodes. Only a few outliers exceed 0.3 m, reaching up to 0.76 m, but these remain exceptions. Regarding dark_frac, almost all nodes have values equal to 0, with only three points above zero (≤ 0.3 m). When compared with the remaining cycles (Fig. 6A), the distribution of wse_u values is similarly concentrated, with a median of 0.15 m. Similarly, dark_frac, values remain consistently close to 0 for all cycles, confirming the robustness of water detection throughout the study period. Taken together, these indicators show that the SWOT observations during the flood peak were of sufficient quality to ensure reliable WSE estimates. This case study demonstrates that the combination of SWOT observations with LISFLOOD-FP modelling provides valuable insights into discharge dynamics in poorly gauged reaches, particularly under extreme flood events where tributary contributions and hydraulic uncertainties are most pronounced.

During cycle 509, corresponding to the recession phase, differences between SWOT-derived WSEs and hydraulic model simulations remain limited. The bias remains low (-0.10 m) and the RMSE is 0.22 m, indicating a generally consistent behavior between the two datasets.

Cycle 510 was excluded from the analysis because all available observations had an xovr_cal_q value of 2, indicating low-quality cross-over calibration. This shows the value of using the quality indicators supplied with the data to assess their reliability prior to any comparison with a model. Beyond this automatic filtering based on quality flags, a limited manual consistency filter was applied for three cycles, and the sensitivity of the results to this additional step was explicitly assessed. For cycles affected by isolated outliers, manual filtering led to a substantial reduction in the error metrics. This marked improvement reflects the strong influence of a small number of inconsistent nodes on aggregate error statistics, highlighting the disproportionate impact that isolated outliers can have on RMSE and bias estimates. For instance, for cycle 502, the RMSE decreased from 1.29 to 0.53 m and the bias from 1.05 to 0.42 m after manual filtering. Similarly, for cycle 509, the RMSE decreased from 6.51 to 0.22 m and the bias from -5.5 to -0.10 m, while the number of retained nodes was reduced accordingly. These cases illustrate that the large pre-filtering errors were primarily driven by a few extreme outliers rather than by a systematic mismatch between SWOT observations and model simulations. In contrast, for cycle 506, where only a single node was removed, RMSE and bias remained nearly unchanged (RMSE from 0.52 to 0.50 m; bias from 0.35 to 0.33 m). These results show that the manual filtering step mainly removes a few clearly inconsistent observations that strongly inflate the error metrics, rather than artificially improving the agreement between SWOT observations and the model. The overall conclusions of the SWOT-model comparison therefore remain robust. Future work should therefore focus on establishing explicit links between SWOT quality flags and the automatic identification of outliers. Such an approach would help reduced satellite-related uncertainties, improve reproducibility, and streamline the selection of reliable WSE observations.

The pre- and post-flood cycles (503–507, and 511–514), as well as cycle 509 in the recession phase, show good agreement between the WSEs observed by SWOT and those simulated by the LISFLOOD-FP model. RMSEs are between 0.22 and 0.53 m. These findings are in line with the conclusions of Frasson et al. (2019), according to which SWOT data, despite their irregular spatio-temporal nature, can be useful for evaluating or adjusting hydraulic models, particularly when acquired after a flood event. In this sense, the study contributes to ongoing efforts to integrate satellite-derived river observations into hydraulic modelling frameworks by demonstrating their operational relevance under challenging hydrological conditions.