*4.3. Benthic Classification*

Classification algorithms performed very differently. It is difficult to assess the impact of the geomorphic predictors on the results. It seems that adding extra information did not impact the SVM and ML classifications, but could have degraded the NN (+1HL) classification. A large amount information could have undermined the results due to a redundancy in the information. The major change seems to be related to the use of the red spectral band. The results were sometimes better without the red spectral band, probably due to the fact that the corresponding maps were in the depth range of 3.7–15 m, for which benthic classes, such as sand and coral rubble and rocks and coral rubble, are not present. A reduced number of groups tends to enhance the algorithm performance.

The benthic classification is based on a visual recognition of general benthic classes based on experts' knowledge. Although, commonly done, identifying benthic classes on MS imagery is not as reliable as direct underwater observations. Living corals could have been confused for dead corals colonized by algae. As a matter of fact, the classification presented in Figure 11a presented very good results, while having a major bias in the classification of coral and algae in areas of deep water. Some regions selected both to train the algorithm and for further validation presented corals which were distinguishable but very dark, due to the depth. Those were confused with deep and dark water areas.

Moreover, Mayotte is a complex area. The tidal range reaches 4 m. Therefore, when the tide is the lowest, corals can be above the water surface and bleached. Dead corals are theoretically recognizable by their bright color, but they can get darker as they are often colonized by algae. The winds and the waves have the effect to break coral colonies and to create coral rubble areas which are difficult to identify, as they get mixed with sand and rocks and can be mixed up with areas of isolated corals.

#### **5. Conclusions**

This study aimed to evaluate the quality of VHR DDM and DAM generated from satellite data. A DDM calibrated with data from the satellite ICESat-2 presented a RMSE of 0.89 m along ICESat-2 ground track, i.e., around 6% of the maximum depth retrieved by ICESat-2. Bathymetric results were generally satisfying down to a depth of around 15 m, which is close to the maximum depth of the calibration data used. Marine habitat classification results were very heterogeneous, depending on the number of predictors used, the type of predictors, and the algorithm used. However, some combinations of parameters provided satisfactory results. The classification with the ML classification using Blue and Green spectral bands with the three geomorphic predictors provided an overall accuracy of 96.62% and a κ coefficient of 0.94. In addition, the SVM classification using Blue, Green, and Red spectral bands with the addition of the slope geomorphic predictor presented an overall accuracy of 96.50% and a κ coefficient of 0.95. This approach can be of strong interest to map coastal areas lacking bathymetry and marine habitat maps and for which field observations are difficult.

While the quality of the results obtained in this study can support coastal management and conservation, the accuracy of bathymetry predictions remains limited for applications, such as navigation, that require higher spatial accuracy. It would be interesting to pursue this research to get more accurate DDMs.

Further work, implementing this method on diverse study sites, would confirm the robustness of the method implemented. In the prospect of future studies, it would be relevant to consider several ICESat-2 ground tracks from the area of interest and even to add the points from other times that ICESat-2 surveyed the area. This would provide a better variability of depths and a better spatial distribution of the data for the calibration process. Moreover, this increase in the number of points opens prospects for the use of deep learning methods to generate DDMs.

Developing an algorithm dedicated to the processing of seafloor data generated from ICESat-2 datasets would be important. The correction for the refraction effect has proven necessary and reliable, but could be further enhanced. The water column properties are changing with depth and the refraction correction should also adapt according to the water column properties.

It would be relevant to also improve the seabed signal correction by considering the state of the sea (for instance, presence of waves on the water surface), in helping to develop a method that could be used in less sheltered areas [17].

In this study, the results presented were obtained using a MS imagery acquired by the Pleiades-1A sensor with four spectral bands and a VHR of 0.5 m using the panchromatic band. The correlation between spatial resolution and the quality of the resulting bathymetry has been demonstrated in this paper. Therefore, future studies could consider generating better quality DDMs using the WV3 sensor (eight spectral bands at 0.30 m using the panchromatic band) or even the sensor of the new Pleiades Neo constellation launched in early 2021 (six spectral bands at 0.3 m with the panchromatic band).

The ICESat-2 products produced by NASA are constantly enhanced and one can be very optimistic regarding the future quality of DDMs and by-products obtained using ICESat-2 measurements.

**Author Contributions:** Conceptualization, A.L.Q., A.C. and R.D.; methodology, A.L.Q., A.C. and M.F.J.; software, A.L.Q. and A.C.; validation, A.L.Q., A.C., M.F.J. and R.D.; formal analysis, A.L.Q., A.C., M.F.J. and R.D.; investigation, A.L.Q., A.C., M.F.J. and R.D.; resources, A.L.Q., A.C., M.F.J. and R.D.; data curation, A.C.; writing—original draft preparation, A.L.Q.; writing—review and editing, A.C., M.F.J. and R.D.; supervision, A.C. and R.D.; project administration, A.C. and R.D.; funding acquisition, R.D. All authors have read and agreed to the published version of the manuscript.

**Funding:** With the support of Montpellier Université d'Excellence (MUSE) KIM Sea and Coast program for funding the PhotonExplorer project.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** The Pleiades-1A imagery provided by CNES-AIRBUS through the DINAMIS platform is not publicly available. The ICESat-2 L2 ATL03 geolocated photons analyzed during this study are publicly available at https://search.earthdata.nasa.gov/search, last accessed: 28 December 2021. The Litto3D® dataset used for validation is publicly available from https://diffusion. shom.fr/, last accessed: 28 December 2021. The diffuse attenuation coefficient of 490 nm measured at 4 km resolution by MODIS is available publicly from https://oceancolor.gsfc.nasa.gov/l3/, last accessed: 28 December 2021. Altimetric information over Mayotte island are available from the SHOM website, https://data.shom.fr/, last accessed: 25 October 2021.

**Acknowledgments:** The authors gratefully thank the NASA for distributing the ICESat-2 data and for the support of the ICESat-2 Applied Users group, and the SHOM for providing us with the airborne LiDAR survey data. The authors would like to thank the DINAMIS platform for access to the Pleiades images and, the CNES "Pléiades © CNES 2020, Distribution Airbus DS". Pleiades-1 imagery is a courtesy from CNES-AIRBUS. Thank you to Thomas Claverie (UMR Marbec) and Aline Aubry (UMR Espace-Dev) from CUFR Mayotte for providing expert knowledge used for habitat mapping. The authors would also like to thank the reviewers for improving the manuscript with their comments and corrections.

**Conflicts of Interest:** The author declares no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

**Appendix A**

**Figure A1.** Different regression models tested to link the relative bathymetry points to ICESat-2 bathymetric measurements. (**a**) Second degree polynomial regression; (**b**) third degree polynomial regression; (**c**) linear regression; (**d**) logarithmic regression.

#### **Appendix B**

**Figure A2.** Evolution of the absolute error with depth on a transect along the ICESat-2 ground track.

**Figure A3.** Evolution of the absolute error with depth on a transect across the ICESat-2 ground track.

**Figure A4.** Evolution of the absolute error with depth on a transect far from the ICESat-2 ground track.
