The paper uses ICESat data for a training data set for a linear model (multiple linear regression) for SDB from Sentinel-2. The paper shows good results for estimating SDB for several areas in the South China sea. Demonstrations of methods for applying ICESat data for tuning, and for execution are a useful addition to the SDB field.

The methods section has insufficient information to reproduce the application to the satellite. That is necessary, given this is a methods paper. The most noteworthy is how the apparently multiple h_0 and h_i coefficients developed for each ICESat trackline were applied around an island.  This should come right after line 340. This is non-trivial, there are 3 h_i coefficients (3 bands).

Also, the limitations of some of the statistical validation need to be identified.  

Methods:

The Sentinel-2 reflectance source and atmospheric correction are not explained.

Was one image used for each location?

The selection of areas for the regression is not clear. Perhaps because text is split up between lines 170 and 320-335.

Line 170  What does this mean?  “We used the GEBCO_2023 model to identify and remove deep-water effects (>100 m) in SDB estimation”. Weren’t the NDWI and ICESat depths used to do this?

Line 320-335. This is not clear. Dividing the track into segments “based on water depth variation trend (from ascend to descend)”.  Divide how?  This is critical to  how the correlation coefficients will be determined.  The red band will disappear much sooner than green or blue.  “Ascend to descend” should be changed, they are actions so it doesn’t make sense.  “Shallow to deep”?

After that section, how is SDB determined for the whole island? H_0 and h_i were determined for each track “segment”. Then what? Were they interpolated or averaged?

 And were the correlation coefficients determined for all locations on the track within the shallow water mask. The shallow water mask was determined by the intersection of the NDWI and ICESat?. And how was ICESat screened, line 174? 

The split of data was “80% training and 20% validation”.  What does this mean?  Was this random, were non-overlapping ICESat transects left out of training? If not, and a random split was used, the validation is not independent. It fails to consider spatial autocorrelation (there are a lot of papers on this topic), which would bias in favor of the results. The study “validation”  does not need to be redone, but this problem needs to be clearly identified, and text calling it a validation should be changed.  Perhaps saying that “Model consistency was evaluated. “

Figure 12 and Table 3 do provide one independent validation, as the lidar was not used for training.

On statistisics. In spite of the popularity of Rsquared (R2) as a validation metric, it is both a poor error metric and it is redundant to RMSE (and so unnecessary).  And R2 cannot be compared for samples with different ranges (variance in X).  Many statisticians have reported this; King 1986 (https://www.jstor.org/stable/2111095) is a good example.  There are several descriptions of the problem on the web (R2 is the fit of the line against the variance in the data, so a wider range of data will have a higher R2). Figure 9 shows the problem. Compare 9e to 9b. Occurring to R2, 9e (0.938) outperforms 9b (0.878). However, 9e has twice the error, 1.631 vs  0.802 for 9b. R2 does not provide useful information. Why? 9e has twice the range of depths, so the squared variance is much greater.  It’s ok to leave the R2 in the figures, because there are people who are desperate to see it, but leave any comparisons of R2 out of the text.  Remove R2 reference from 362-374, 410-425, 529.  This problem should be stated at line 341:  e.g., “R2 is actually redundant with RMSE. However, R2 also varies with data range, so unlike RMSE, R2 values cannot be meaningfully compared between different samples.  R2 values are included because they are familiar.“ 

Line 335.  “The effects of deep-water areas were then removed to minimize the influence of bottom reflection on SDB estimation”.  What effects were removed from what?  Does this text belong before line 312?  (“average deep-water reflectance”).

Figure 12 and 13 captions are not clear, please include the letters in the caption.  It would be even better to label each column.

This manuscript integrated ICESat-2 data and multispectral imagery from Sentinel-2 to construct a high-resolution, high-accuracy shallow water depth model namely HHU24SWDSCS. In geopolitically sensitive areas, such as the South China Sea, where the seafloor topography is complex, existing water depth data primarily rely on sparse multibeam sounding technology and satellite altimetry-derived depths. The HHU24SWDSCS model developed in this study successfully filled the gap in shallow water areas and offered an alternative that can be applied to similar regions globally. The comparisons with existing bathymetry models demonstrate that the computed model offers significant advantages in both accuracy and details for shallow areas, highlighting its potential for high-quality shallow water depth measurement. The manuscript demonstrates notable innovation and scientific significance. Publication is recommended following revisions.

• The English writing should be further polished.
• Line 19: What is the full name of HHU24SWDSCS?
• Line 31: How to define the shallow water in the study?
• Line 105: It is seen that the authors only retrieved water depths over island areas, is it possible to perform SDB modeling in nearshore areas (e.g., estuarine region), and what is the SDB quality there?
• Table 1: More information should be further shown, like max depth, min depth, mean depth, etc.
• Table 1 and Table 2: These tables should follow the three-line table format.
• Line 145: The explanation regarding the selection of training and validation data is not sufficiently clear. Why was an 8:2 rule used for training and validation? How about the data distribution?
• Figure 3: How did the authors perform water mask in the imagery? Since the study focuses on shallow water areas, it is easy to cause confusion between land and water in Sentinel-2 imagery.
• Section 3.1: What is the bounding depth detected by IceSat-2? What factors can affect the ability of IceSat-2-based depth detection?
• Section 3.2: What is the bounding depth detected by Sentinel-2? The description of the methodology may be improved. For instance, the choice of modeling region and the reasons for data segmentation and Pearson correlation analysis are not adequately explained.
• Line 310: How generalizable is the LBM model trained by the authors? Can it be applied to other marine areas with insufficient ICESat-2 data for SDB modeling?
• Line 350: How deep of water depths can be detected from SDB (seems ~ 30 m in this study area), is it possible to apply SDB modeling over water areas with deeper depths than the one used in this study?
• Figure 8: The authors mentioned several islands and reefs in their model but omit other islands and reefs in the South China Sea, such as the ones in the Zhongsha Islands region. Please explain this issue.
• Figures 9 and 10: The point cloud density was uneven, with most points concentrated in the 1-3 meter depth range. While Figure 7 shows that the ICESat-2 data was predominantly concentrated in shallow water areas, this could lead to inconsistent fitting of the regression model across different depth ranges. I suggest the authors consider resampling the ICESat-2 data within the 1-3 meter range.
• Figure 11: Similar to the previous figures, Figure 11 showed uneven point cloud density distribution. Additionally, the origin of the XY axes should be at 0 m, rather than -10 m. Please redraw these figures.
• Line 440: There was no introduction of the DTU18BAT, topo_27.1, GEBCO_2023, or SRTM models earlier in the manuscript. What data these models use for construction? What are their spatial resolutions and accuracies? More information can be included.
• Line 495: The authors have used ICESat-2 data spanning over five years for SDB modeling. Have they considered the potential impact of temporal changes in seafloor topography due to human activities or ocean currents during this period?
• Line 520: The conclusion contained some repetitive expressions and redundant language. It can be streamlined for conciseness.
• Line 545: The appendix lists all the islands and reefs used for modeling, but does the author model each of these individually? How much data was used for each island or reef?
• The manuscript contains several expressions like "point clouds," where "point" should be replaced with "data".
• References: The format of literatures should be further normalized.