The study area was observed during two flood events in December 2019 and early February 2021 by the Sentinel-1 C-band Synthetic Aperture Radar (SAR) instrument at 5.405 GHz. The extent of the study area and the satellite image acquisitions are visualized in Fig. . Sentinel-1 “Ground Range Detected” products were downloaded from ASF Data Search Vertex (https://search.asf.alaska.edu/, last access: 26 May 2026) in VH polarization. The role of polarization is beyond the scope of this study. Most flood mapping methods rely on the analysis of a single polarization image. Accordingly, the VH polarization was used since, in the literature, VH polarization better distinguishes flooded from dry areas than VV polarization . Two additional Sentinel-1 images, acquired under non-flooded conditions, were used as references for one of the flood mapping methods. For the 2019 event, the reference image was acquired on 10 December 2019. For the 2021 event, the reference observation was acquired on 28 January 2021. Each image consists of an Nx×Ny grid of pixels, with a spatial resolution of 10×10 m. The images are projected onto a common grid that covers the entire study area. Here, the image's dimensions are Nx=2644 and Ny=2312. All times reported below are given in Coordinated Universal Time (UTC).

For comparing the observed flood maps and water depth fields derived from satellite images, observed water marks and stage gauging stations are available. These observations were also compared to simulated flood maps and water depth fields. The simulations are not the ground truth, but serve as a reference for comparison. The main goal of the study is to compare the variability of the outputs due to preprocessing, method choices, and hyperparameters, so the validation dataset is not the most important. We describe these datasets below.

Watermarks are visible traces left on buildings, trees, or other infrastructures during a flood event at the peak water level. For both the 2019 and 2021 flood events, watermarks were collected and made available on the French collaborative platform “Repères de Crues” (https://www.reperesdecrues.developpement-durable.gouv.fr/, last access: 26 May 2026). For the 2019 and 2021 flood events, 121 and 178 are available, respectively. For both events, satellite images were acquired near the flood peak (17 December 2019 and 3 February 2021), so the watermarks should coincide with the water depths extracted from the satellite images.

Three stage-gauging stations are available in the study area (Tonneins, Marmande, and La Réole). Discharge and water level data at these locations during the flood events are available on Vigicrues (https://www.vigicrues.gouv.fr/, last access: 26 May 2026, a flood-monitoring service that collects watermarks during floods in France). The measured discharges at the three stations for both flood events are shown in Fig. .

For flood simulations, a numerical solution of the Shallow Water Equations (SWEs) was used and solved with TELEMAC-2D, part of the openTELEMAC open-source hydrodynamic modeling system (https://www.opentelemac.org/, last access: 26 May 2026) . The SWEs are expressed in Cartesian coordinates where gravity acts uniformly in the vertical direction as -gez: 1∂h∂t+∂(hu)∂x+∂(hv)∂y=0, 2∂(hu)∂t+∂(hu2)∂x+∂(huv)∂y=-gh∂η∂x+∇⋅hνe∇u-1ρτbx, 3∂(hv)∂t+∂(hv2)∂y+∂(huv)∂x=-gh∂η∂y+∇⋅hνe∇v-1ρτby. The system unknowns are the water depth h≥0 and the depth-averaged velocity uex+vey both functions defined as functions of spatial coordinates x, y and time t∈[0,T). The water surface is denoted by η=h+zb, where zb(x,y) is the prescribed bottom elevation. A constant viscosity νe>0 is assumed, and the bed shear stress is expressed as τbxex+τbyey depending on the variables h, u, and v. The bed shear stress is computed using Manning-Strickler formulation : 4τbx=ρ⋅g⋅uh1/3⋅Ks2u2+v2τby=ρ⋅g⋅vh1/3⋅Ks2u2+v2, where Ks is the Strickler coefficient which varies spatially with x, and y .

Upstream discharge was retrieved from the Tonneins stations for 11–21 December 2019 and 25 January–10 February 2021 and reported in Fig. . A rating curve was imposed at the downstream boundary. The simulations were initialized at t=0 with a base flow of 800 m³ s⁻¹ for the 2019 event, and 2300 m³ s⁻¹ for the 2021 event. One simulation was conducted for each flood event. The Strickler values were calibrated in previous studies for the river channel , and are fixed in the floodplains based on the land use . The spatial distribution of Strickler values used in the simulation is reported in Fig.  and their values are described in Table  taking the median value of the Strickler value ranges of Design of Experiment 1. Bathymetry in the channel was reconstructed from 70 cross-sectional elevation profiles, and floodplain topography was obtained using data from IGN (the French National Institute of Geographic and Forest Information) combined with aerial imagery. The reader can refer to and for more information on the construction and parameterization of the numerical model.

The Median filter replaces the center pixel with the median value of all pixels within a local scanning window, such that: 6R^i,j=medianIm,n|i-k≤m≤i+k,j-k≤n≤j+k, with a window size of 2k+1 with k a non-negative integer. In this study, the window size is treated as a hyperparameter. The median filter is effective at removing isolated spot noise, but it tends to blur edges and erase thin linear features . To address this limitation, other approaches referred to as adaptive filters have been developed that incorporate local image statistics to better preserve structural details.

The Lee filter accounts for the local statistics of the image within a moving window. From the multiplicative noise definition in Eq. (), the mean and variance of R can be estimated, such as: 7R‾i,j=I‾i,j,Var(Ri,j)=Var(Ii,j)-I‾i,j2Var(Si,j)Var(Si,j)+S‾i,j2, where the quantities I‾i,j and Var(Ii,j) are the local mean and variance of pixel intensities within a window of size 2k+1 centered at (i,j). In this study, the window size is treated as a hyperparameter. The Lee Filter assumes a linear estimator and using the local mean and variance within each scanning window, the estimator is written as: 8R^i,j=R‾i,j+Var(Ri,j)Var(Ii,j)(Ii,j-S‾i,jR‾i,j). The Lee filter aims to reduce speckle in homogeneous areas while preserving image details, such as edges and fine structures, in areas of high variance.

The Lee filter effectively reduces speckle noise but can cause blurry edges and loss of detail in heterogeneous areas. To address this issue, the Lee Sigma filter was introduced. It calculates local statistics similarly to the Lee filter but using only the pixels whose intensities fall within a range to exclude outliers. The range is defined as [R‾i,j-2Var(Si,j)R‾i,j,R‾i,j+2Var(Si,j)R‾i,j]. Because the a priori mean, R‾i,j, is unknown, it is approximated by Ii,j, the value of the center pixel.

An improved version relaxes the usual range and introduces a new interval (I1,I2) that meets two conditions:

The interval captures a fixed cumulative probability ξ:9ξ=∫I1I2p(I)dI,where p(I) is the empirical probability distribution of the pixel intensity computed on the scanning window.

Similarly to the Lee filter, the Frost filter is based on the local statistics of the images and the multiplicative noise model. The Frost filter replaces a pixel value with a weighted sum of the values of its neighbors within a moving scanning window, such that: 11R^i,j=∑m=-kk∑n=-kkwm,n⋅Ii+m,j+n, where the weights wm,n decrease with distance from the pixel of interest with an exponential decay controlled by a damping factor α, such that: 12wm,n=e-αd(m,n)∑m=-kk∑n=-kke-αd(m,n), with d(m,n) the Euclidean distance of pixels at position m,n from the pixel at the center of the scanning window. For a pixel center located at i,j the distance is defined as d(m,n)=(m-i)2+(n-j)2. For the Frost filter, two hyperparameters are considered in this study, the window size and the damping factor α.

Due to the difficulty of removing speckle noise with empirical models based on image statistics, deep learning approaches are increasingly used. SAR2SAR is a deep learning-based despeckling method using a U-Net architecture which aims to predict speckle noise. The U-Net is trained with synthetic speckle realizations, then fine-tuned on real SAR image pairs from a time series. After training, weights are used to compute the denoised image as a function of input pixel intensities. In this study, we used a pre-trained model with weights available at https://gitlab.telecom-paris.fr/ring/sar2sar (last access: 26 May 2026). For SAR2SAR, no hyperparameters were tested.

Global thresholding of SAR images is a straightforward method to separate foreground (e.g., flooded pixels) from background (e.g., dry pixels) widely used in flood mapping for their simplicity . This approach assumes that the two-pixel classes (e.g., flooded and dry pixels) can be separated with a threshold value for the whole image. The threshold value is usually determined automatically .

In this study, we evaluated the impact on the flood mapping of two widely used global thresholding methods: Otsu's method and the Kittler and Illingworth (KI) method . Complementary details on both methods are provided in Appendix .

Global thresholding assumes that the histogram is bimodal, which may not hold when the number of flooded pixels is low or due to spatial variability in backscatter caused by terrain or surface conditions across the image. To address this limitation, local-based thresholding was introduced . Images are divided into smaller tiles to identify bimodal histograms representing flooded and dry classes. From these identified sub-tiles, a threshold is computed with global thresholding. Following , we used quad-tree decomposition to recursively split the image until reaching a minimum tile size (to guarantee statistical representativeness). A tile is eligible for thresholding if its histogram is bimodal, the distributions are normally distributed, and both classes are sufficiently represented. These conditions are evaluated by computing three qualitative coefficients (Ashman D, surface ratio, Bhattacharyya) described in Appendix .

This study evaluated the impact on the flood mapping of the minimum tile size and of the three coefficients for different values reported in Table .

A classical active contour model is based on the Chan-Vese segmentation method , which is inspired by the Mumford–Shah model . This approach determines a contour C that segments the image into two regions so that the pixel intensities are approximately constant within each region. The segmentation problem is formulated as the following minimization: 14min⁡CαLength(C)+νArea(inside(C))+λ1∑(i,j)∈inside(C)R^i,j-c12+λ2∑(i,j)∈outside(C)R^i,j-c22, where α≥0 controls the smoothness of the contour, ν≥0 penalizes the area inside the contour, λ1,λ2>0 are weighting the data fidelity inside and outside the contour, respectively, and c1 and c2 are the average intensities inside and outside the contour C, respectively. The algorithm is adapted so that c1<c2, forcing flooded pixels with lower backscattering to be inside C.

In this study, we considered α as a hyperparameter, while the remaining parameters were fixed to λ1=2, λ2=1, and ν=0 for computational cost reasons. We chose to constrain the problem with λ1>λ2 to favor the identification of regions with lower mean intensities (typically associated with flooded areas in SAR imagery) as the foreground while assigning higher-intensity regions to the background. Active contour models require an initial contour as a starting condition for the segmentation process. suggests selecting the initial contour manually. In this study, the contour was initialized using a map of permanent water bodies from the Global Flood Monitoring Service with a 10 m resolution (https://global-flood.emergency.copernicus.eu/, last access: 26 May 2026) .

Change detection methods require at least two satellite images of the same area, one during the flood and the others before or after the event. The output of change detection methods depends on the reference image without floods, so the reference image should be chosen carefully . Although there are methods using a stack of images , we focused on a pair of images. First, we computed the logarithmic difference between the two images because of the multiplicative character of speckle noise (see Eq. ) . Flooded pixels are then classified using global thresholding described in Sect.  with consistent parameter sampling. Similarly to global thresholding, the hyperparameter in the change detection method is the method to find the threshold on the image (Otsu or KI).

Two supervised classification methods, widely used for flood mapping based on labeled training data are used in this work: Random Forests (RF) and Convolutional Neural Networks (CNN) . For the RF approach, the pixel intensity is used as a feature. With only one band (VH polarization), the prediction capability of RF is limited. We added textural information by providing the mean value of each pixel with window sizes of 3×3, 5×5, and 7×7 pixels, resulting in four features.

Both models are trained with the Sen1Floods11 dataset (including 11 flood events with various geographic conditions) using VH polarization band and hand-labeled data dichotomizing flooded from dry regions. The trained model weights were used to generate flood maps from satellite imagery based on pixels' backscattering intensity. For reproducibility, the weights of both models are available at: https://github.com/jtravert/sar-flood-evaluation-framework/tree/main/sources/1_FloodExtent/methods/MLweights (last access: 26 May 2026). The training procedure for the CNN was adapted from using a Fully-Convolutional Network model with a ResNet-50 backbone. The Random Forest classifier was trained for various numbers of estimators, maximum depth, and minimum leaf number with Bayesian optimization using the Optuna Python library (https://pypi.org/project/optuna/, last access: 26 May 2026). For supervised classification methods, no hyperparameters are considered, as their hyperparameters were optimized during the training procedure using Bayesian inference.

First, the images generated without applying morphological post-processing operations are analyzed. Figure shows the variability of the Accuracy, F1-score and flooded area metrics across the four satellite acquisitions. Each box plot represents the range of metric variations of a flood mapping method due to varying input preprocessing and hyperparameter settings. For each method, box plots are displayed in different colors for the four satellite images. These box plots should be compared with one another for the same acquisition date to assess method performance variability under consistent conditions.

The CNN approach presented the highest median performance (between 0.80 and 0.94 for F1-score) with a narrow inter-quartile range (<0.005) in three out of the four acquisition dates (except for 17 December 2019 acquisition). The Random Forest also performed well (F1-score median between 0.76 and 0.93), but with slightly lower scores and higher variability. Both supervised classification methods exhibit low variance because of the absence of hyperparameter tuning, leaving speckle filtering as the main source of variation. In contrast, active contour models exhibited larger variability in Accuracy and F1-score across all dates. It underlined the sensitivity to either the smoothing parameter (α) or image preprocessing. While they occasionally achieved reasonable performance for specific configurations, their inconsistency (e.g., F1-score inter quartile range between 0.27 and 0.38 on the 16 December 2019) makes them less reliable in operational settings. This method can work only after careful manual tuning, limiting its operational appeal. Active contour models are usually better for local scales and simple flood patterns around the initial contour. Thresholding methods (global and local) showed moderate performance and low variability. Additional configurations with a larger hyperparameter space showed that the thresholding methods were not very sensitive to their hyperparameters. For some configurations, local thresholding resulted in comparable metric values to supervised classification while being easy to use. Finally, change detection methods provided similar moderate results regarding F1-scores and Accuracy. Their variance was relatively low but was higher for some acquisition dates (17 December 2019 and slightly 3 February 2021), likely due to differences in magnitude change between the pre- and during-flood images, with smaller flood extent on 17 December 2019. The supervised classification method provided the best trade-off between accuracy and robustness. When training data or model weights are unavailable, local thresholding or change detection methods remain attractive because they require almost no user input and result in similar scores with appropriate tuning.

Figure  also displays the flooded area for all flood mapping methods and satellite acquisitions. In the evaluation zone, containing the simulation domain without the exclusion zones and the minor bed representing a total of 100.2 km², the simulated flooded areas were 78.78 km² (16 December 2019), 52.32 km² (17 December 2019), 57.70 km² (2 February 2021) and 86.41 km² (3 February 2021). CNN produced the largest flooded area estimates, with the smaller discrepancy to the simulated flooded area. In contrast, the other methods highly underestimated the flooded areas systematically relative to the simulations. In general, outputs from hydrodynamic models provide smooth flooded surface with hydraulic connectivity. On the contrary, for the generated flood maps from SAR imagery, floods in vegetated or urban areas or between the dikes are often not well detected which can cause an underestimation of the flooded area, and the flood extent is more segmented and less smooth. On the one hand the simulations are conservative and overestimate the flooded area while the extracted flood maps underestimate the flooded areas. However, the main goal should be to match the geometric pattern and have a similar flood map geometry between the simulations and the observations even if there are small holes in the observed flooded area. The estimated flooded area can differ by 20 to 30 km² between methods, highlighting significant differences in flood map extraction approaches. Additionally, for a given method, the impact of hyperparameters or input images can be significant, with variations of 5 to 10 km² for active contour or change detection and CNN for 17 December 2019.

In most flood studies, the preprocessing of speckle noise is often assumed to be deterministic, with only one method (usually the Lee or Lee Sigma filter) ; however, the influence of speckle filtering choice on flood mapping can be significant. The preprocessing strategies should be evaluated for fixed configurations (flood mapping method and hyperparameter) to study the impact of preprocessing alone. For instance, the effect of preprocessing for CNN-generated flood maps is reported in Fig. . It highlighted the impact of preprocessing with a highly variable flooded area for Median filtering, depending on the window size of the filter (the only parameter for the Median filter). With filtering, the flooded area is larger than that of images without preprocessing. Between the different preprocessing methods, the flooded area has important variations showing the impact of the preprocessing method on the generated flood maps.

Figure  illustrates the impact on the flooded area of the different filtering methods with a single flood mapping configuration. In total, Table  lists 48 configurations (2 for global thresholding and change detection, 36 for local thresholding, 6 for active contour, and one for each supervised classification method). For the Median, Lee, Lee Sigma, and Frost filters, we analyzed the impacts of their hyperparameters on these 48 configurations. The difference between the minimum and maximum value for each metric was calculated for every filter configuration (e.g., for the CNN configuration on 2 February 2021, the range of variation between the min and max values for the Median filter was 23 km²). Figure  presents the min-max variation in F1-score and flooded area due to speckle filtering hyperparameters for 2 February 2021. Each flood mapping configuration is visualized with a point with distinct markers for the different methods. The results indicate that the sensitivity to speckle filter hyperparameters varies by flood mapping method. For instance, the active contour method exhibits high sensitivity to the Median filter’s window size, with variations ranging from 4 to 16 km². In contrast, global and local thresholding methods exhibit minimal sensitivity, with variations of less than 1 km². Overall, the Lee Sigma filter exhibited the lowest variability across all configurations. Local thresholding, change detection, and Random Forest configurations are more sensitive to Lee filter or Frost filter hyperparameters than Lee Sigma or Median filters.

The role of hyperparameters used in flood mapping methods was evaluated, excluding the supervised classification methods (CNN and Random Forest), for which no hyperparameters were defined in this study. Similarly to the preprocessing analysis, we fixed the preprocessing configuration to isolate and evaluate the impact of flood mapping hyperparameters.

The range of variability due to the hyperparameters of flood mapping methods for all 26 speckle filtering configurations (3 configurations for Median and Lee filters, nine configurations for Lee Sigma and Frost Filters, one configuration for SAR2SAR and no preprocessing) is evaluated. The results for the acquisition on 2 February 2021 are presented in Fig. .

In this study, the active contour method showed the highest sensitivity to flood mapping hyperparameter variations, with F1-score differences reaching up to 0.4 and differences in flooded areas ranging between 15 and 25 km² for all speckle filtering configurations. In contrast, change detection and local thresholding methods demonstrated low sensitivity, with F1-score differences of less than 0.03 and flooded area variations limited to a few square kilometers. For active contour models, most of the variability can be attributed to the flood mapping hyperparameters. While thresholding methods showed limited variability, both active contour and change detection methods displayed greater sensitivity depending on the fixed speckle filtering configuration. For instance, variations in F1-score or flooded area due to hyperparameters for change detection were smaller for Lee Sigma configurations compared to Median filtering configurations.

Figure  shows the improvement or degradation in the F1-score for global thresholding, local thresholding, and change detection methods for one of the acquisitions (2 February 2021) when using morphological operations. The color intensity represents the average gain (or loss) relative to the configurations without morphological post-processing. The active contour and supervised classification approaches are not reported here, as morphological operations had negligible or no effect on the generated flood maps due to the minimal presence of holes or small unconnected flood elements in these flood maps.

We observed that the F1-score increased for the three methods, especially for large hole filling (50 to 100 px) and small patch removal (10 to 50 px). This is likely due to eliminating isolated false positives and filling small gaps, improving the spatial coherence of flood maps. However, with the increasing size of patch removal, the performance decreases, suggesting that the small-scale features are well-captured and should not be removed. While the morphological operations improved the statistics for the F1-score, the physical representation of the flood may not be improved, with for instance flooded regions on small hills when using filling operations.

Finally, Fig.  compares the flood maps with the highest F1-scores obtained for each method, including those enhanced with morphological operations for 2 February 2021. The CNN-based approach achieved the best overall performance, with an F1-score of 0.769, followed by the Change Detection and Random Forest methods. The flood map produced by the CNN is the most continuous but tends to overestimate the flooded area. The active contour model also exhibits a continuous flood extent, but it significantly underestimates inundation in both the upstream and downstream regions. The other methods result in less smoothed flood extent, capturing more localized details and irregularities. While some of these details may correspond to actual observations, their fragmented nature results in poorer alignment with the continuous structure of the simulated flood map, while having patterns that match the simulated flood map closer than for CNN.

The Flood Water Depth Estimation Tool (Fw-DET) quantifies water depth continuously in the domain using a flood extent polygon and a Digital Elevation Model. A schematic representation of Fw-DET principle is presented in Fig. II(1). Flood boundaries are derived from the flood map and converted into a line layer, while steep slope cells are filtered out using a slope threshold. This line layer is rasterized to align with the DEM grid. Elevation values are extracted for these boundary cells, and each cell within the flood extent polygon is assigned the elevation of the nearest boundary cell under the assumption of a flat water surface. Water depth is computed by subtracting the DEM elevation from the water surface elevation. To mitigate artifacts caused by mismatches between the DEM and the flood extent, a smoothing procedure is applied multiple times using a 3×3 window.

This study considered the number of smoothing iterations and the slope threshold as hyperparameters. The latest version of the code developed by is available at https://github.com/csdms-contrib/fwdet (last access: 26 May 2026). For the present study, the code was adapted to operate without the QGIS interface, using standalone Python scripts to enable batch processing.

The FLEXTH method presents an approach similar to the Fw-DET methodology but introduces improvements to mitigate unrealistic water depth estimates . Figure II(2), depicts the FLEXTH methodology schematically. The methodology was extended to account for additional inputs, such as exclusion or permanent/seasonal water body masks. The method expands the flooded area into adjacent no-data regions and estimates the water depths based on the DEM . As in Fw-DET, pixels along the flood boundaries with a slope exceeding a user-defined threshold are excluded from the boundary. Water depths within the flooded area are then estimated using a weighted average of the boundary cell elevations based on the closest cells up to a specified maximum number of neighbors.

The slope threshold and the maximum number of neighbors used in the computation were considered hyperparameters. In , they highlighted that these two parameters are the most influential in the FLEXTH method. The FLEXTH method, initially developed by , is available as a Python code at https://code.europa.eu/floods/floods-river/flexth (last access: 26 May 2026) and was adapted for the needs of our study.

Figure  shows the RMSE of the estimated water depths compared to hydrodynamic simulations (see Fig. a) and watermarks (see Fig. ) for the Fw-DET and FLEXTH methods. Each box plot indicates the variability explained by the input flood map or water depth estimation hyperparameters.

FLEXTH and Fw-DET demonstrate comparable performance in terms of median RMSE. Both methods exhibit a median RMSE ranging from 0.8 to 1.9 m, depending on the satellite image acquisition. The variability of the RMSE is also comparable for both algorithms, with slightly more outliers for the FLEXTH method, which may be caused by the used hyperparameter values. These conclusions are similar either with the simulated water depth fields or watermarks as validation dataset. For the watermarks, the variability is more important. The observed variability and elevated RMSE values can be attributed to the flood map selection from previous processing steps, including high and low F1-score maps relative to the reference. This variability and the influence of water depth estimation method hyperparameters are further analyzed in Sect.  for the Fw-DET and FLEXTH methods.

Next, Fig.  presents the RMSE of water depth estimates on the two selected cross-sections (see Fig. ) for the cross-section approach, along with Fw-DET and FLEXTH methods projected onto these cross-sections.

The Fw-DET and FLEXTH methods exhibited similar RMSE performance for both cross-sections, though FLEXTH showed higher variability. The cross-section approach, while showing similar or slightly higher median RMSE values compared to FLEXTH and Fw-DET on the two cross-sections, exhibited more variability, particularly for the 2 February 2021 image, which was likely related to challenging flood contour detection conditions for that image. The Fw-DET and FLEXTH algorithms propagate the surface elevation from the edges of the flood to the inner flood. The cross-section approach, on the other hand, relies only on the identification of the right and left edges of the flood. Then, the cross-section approach is highly sensitive to identifying the border of the flood on the cross-section, while it is less the case for Fw-DET and FLEXTH, which rely on the whole flood extent borders. The variability of the RMSE for the cross-section approach was entirely due to the variability of the input flood maps, since the cross-section approach was considered without hyperparameters.

The influence of flood map inputs and method hyperparameters on water depth estimation performance was evaluated for the Fw-DET and FLEXTH methods. Figure  presents the distribution of RMSE values, computed against hydrodynamic simulations across four satellite acquisitions, grouped by flood mapping method used as input to the water depth estimation methodology.

The results indicated that the flood map used as input for the water depth estimation process significantly influenced the estimation accuracy. Flood maps derived from global and local thresholding yielded lower RMSE and variability. In contrast, flood maps produced with CNN and active contour models lead to greater variability and higher RMSE in water depth estimates. However, the median RMSE values for CNN and active contour flood mapping methods are close to the other algorithms, suggesting that they can achieve comparable or better results with water depth estimation method parameter tuning. The outputs from CNN and active contour models are smoother, but their ability to precisely delineate flood boundaries is reduced due to the smoothness. This can result in significant errors in water depth estimates, particularly in areas with steep terrain. Furthermore, the large RMSE for the CNN method suggests that the extracted flooded surface is physically wrong on these cross-sections. The tendency of the CNN to overestimate the flood extent size leads to water edges in high topography gradient leading to important errors in the water depth estimation.

To study the influence of hyperparameters in the FLEXTH and Fw-DET methodologies, we proceeded similarly as in Sect.  for flood mapping methods. For each hyperparameter configuration (9 each for FLEXTH and Fw-DET) of water depth estimation methods, we computed the range between the lowest and maximum RMSE for every flood map input. Figure  presents a statistical analysis of the variability of this range for all the flood map input configurations for the 2 February 2021 acquisition.

While Fw-DET showed minimal variability in RMSE due to its hyperparameters alone (except for some outliers beyond the interquartile range), the FLEXTH method displays a much larger interquartile range of RMSE values. In this study, the Fw-DET method demonstrates more consistent performance. In contrast, FLEXTH, while potentially more accurate in ideal cases, was more sensitive to hyperparameter choices and could produce larger errors depending on the configuration used in each test case.