This is a very detailed and interesting study that compares different SAR based flood extent and water depth estimation techniques. The study is carried out for two flood events (December 2019 and February 2021) in the floodplains of the Garonne river in France. An impressive number of simulations were carried out using different SAR preprocessing approaches, models and model parameterizations and assessed using high quality hydraulic model outputs and observed watermarks. In my view this is much needed and highly valuable study investigating the strength and weaknesses of different algorithms in SAR data processing + flood mapping + water depth estimation. However, I have several major and minor comments.

MAJOR COMMENTS

1. My main comment is that the authors do not properly account for the effect that Sentinel-1 cannot map flooding in non-sensitive areas (forests, urban areas) and water-look-alike areas (streets, smooth fields, etc.). While they do mention these effects, it has no impact on their methodology of how the SAR derived flood and water depths maps are assessed. The point is that for Sentinel-1 the “ground truth” is not the actual flood area as simulated by the hydrological model or observed by the watermarks. Instead, for Sentinel-1 the “ground truth” is the real flood area minus the exclusion areas (e.g. all pixels where Sentinel-1 cannot detect flooding due to physical reasons). This has major implications for how the results are interpreted. At the moment the authors conclude that the CNN flood map outperforms the other results as the total flood area estimated by the CNN comes closest to the real flood extent. However, comparing the SAR backscatter images and the CNN maps shows that the CNN has a tendency to overestimate flood extent as seen by Sentinel-1. This problem then becomes apparent when estimating water depth from the CNN derived flood maps: The large RMSE values for the CNN shown in Figure 21 are evidence that some parts of the CCN maps are physically wrong.
2. My second major concern is that the comparisons are unfair. While the four non-ML based approaches (global threshold, local threshold, active contour models, change detection) only use VH data as input, both ML models use both VH and VV data as input. Apparently, this is an advantage for the two ML models, and should be considered in the discussion of the results or – even better – lead to a modification of the experimental setup!
3. Morphological filters may help to reduce underdetection in flood situations, but may also lead to overdetections in non-flood situations or to wrong flood pixels altogether. So while these filters may have improved the statistics with respect to the model simulations and watermarks, it is not always clear that the physical representation of observed flood areas is improved.
4. Most flood mapping algorithms are optimized for flood scenes, but may lead to wrong results during dry conditions. To judge the robustness of algorithms, also their performance for non-flood scenes should be tested.

MINOR COMMENTS

Section 1: What is the definition of a “hyperparameter”? What makes it different from a “normal” model parameter?

Section 1: There are some recent studies that investigated the effect of different model parameters on flood mapping accuracies (e.g. recent studies investigating different change detection algorithms). Please review the literature and relate this work to the existing publications (also come back to this point in the discussion section).

End of section 1 / beginning of section 2: Check for repetitions

Line 86: Why are so many configurations tested for the local threshold approach (36 versus 2 / 2 / 6 configurations for the other three methods). Does this mean that the threshold approach has advantages?

Figure 3: Show location of in situ sites

Line 163: Only in a narrow sense I would agree to this statement: “The main source of error in SAR imagery is speckle noise …”. In practice, there are many physical reasons for uncertainties in the SAR derived flood maps.

Line 248: Visually, the SAR2SAR filter looks indeed nicer than the other filters. But are there any quantitative indicators that can substantiate that SAR2SAR “outperforms the traditional methods”? How much of this filtered image is invented, how much of it is true?

Figure 6: These are the VH images, right?

Figure 7: Use the same y-axis for a direct comparison

Line 349: How many flood cases of the Sen1Flood11 cases show similar conditions as for the Garonne river flood.

Figure 9: I find the spread of the results for different algorithm / flood case combinations surprisingly low. What is the reason for this? Does this also reflect different pre-processing options?

Line 535: Some grasslands may cause “water-look-alike conditions”, but normally vegetation causes a loss of sensitivity of backscatter to flooding.

This study investigates the impacts of various preprocessing, mapping, and depth estimation methods on flood mapping and water depth estimation using Synthetic Aperture Radar (SAR) images. The results suggest that an ensemble approach, accounting for various uncertainty sources during the modeling process, should be preferred. Overall, this study is well-designed and comprehensive, and the findings are meaningful for flood risk management. However, I still have several comments and suggestions for improving the current work.

1) This paper is a little lengthy, especially for the method sections. If some of those methods are not new, it is suggested to make the description more concise, or put some of those in the appendix.

2) The overall study is like presenting the details of a calibration process for SAR-based flood modeling, which involves rigorous evaluation as well as subjective adjustment based on the model’s experience. Could you provide some general guidelines on how to make the model calibration more effective and efficient?

3) Figure 2: Does the longitude label "5-degree O" represent "W"? If yes, and to avoid confusion, it suggested to change it to "W". Also, it would be better to add a zoom-in view for the study area.

4) Lines 118-120: It is acknowledged that the simulated flood maps just served as a reference in this study. However, it should be noted that the uncertainty in the physics-based flood modeling process is not negligible. Furthermore, the limited flood events/areas selected in the study and the uncertainty in the “ground-truth” observations may play a role in the model evaluation process. Please refer to the two papers below.

References: 

"Uncertainty analysis and quantification in flood insurance rate maps using Bayesian model averaging and hierarchical BMA" (https://doi.org/10.1061/JHYEFF.HEENG-58)

"Beyond a fixed number: Investigating uncertainty in popular evaluation metrics of ensemble flood modeling using bootstrapping analysis" (https://doi.org/10.1111/jfr3.12982)

5) Figure 7 and Line 259: Why was the performance of SAR2SAR for the flooded area not as good as that for the dry area?

6) Lines 370 and 464: No hyperparameters were considered in the random forest and CNN approaches, but they are also critical in determining the model performance in both the training, testing, and application processes.

7) Line 382: There are 1222 generated flood maps, but 26*48=1248?

8) Lines 91 and 502: The DEM data was used in water depth estimation. Does it include bathymetry for the river channel?

9) Line 646: The word "First" does not make sense here since there is no "Second" in the following text.

10) Table 3: Please add units for the Strickler parameters.

11) Line 652: Is it "94%" or "95%"?