Extracting Coastal Water Depths from Multi-Temporal Sentinel-2 Images Using Convolutional Neural Networks

Figures & data

Figure 1. Distribution of study areas. Source: Map data from OpenStreetMap.¹

Figure 2. Six AOIs used in this study: AOI-1 (a), AOI-2 (b), AOI-3 (c), AOI-4 (d), AOI-5 (e) and AOI-6 (f). The left pictures show Sentinel-2 Level-2A image, and others present the in-situ depth data from LiDAR or MBES measurements.

Table 1. An overview of coastal water conditions in each study area.

Area	Turbidity	Benthic habitat
AOI-1: Ponce	Low – medium	Sand, patch reefs, seagrass, algae, coral rock (Appeldoorn et al. 2016; National Oceanic and Atmospheric Administration (NOAA)) 2017)
AOI-2: Lajas	Low – medium	Sand, reefs with high coral diversity and density, seagrass, algae, rock/boulder (Bauer et al. 2012; Larsen and Webb 2009; National Oceanic and Atmospheric Administration (NOAA)) 2017)
AOI-3: San Juan	High	Sand, mud, pollutants, seagrass meadows, reefs and coral rocks outside the ports (Puerto Rico Environmental Quality Board (PREQB)) 2016)
AOI-4: Key West	Low – high	Sand, mud, seagrass, algae, live coral (National Centers for Coastal Ocean Science, n.d.)
AOI-5: Makua	Low	Sand, coral reef, rock/boulder (Hawaii State Office of Planning, n.d.; University of Hawaii 2013)
AOI-6: Honolulu	Low – medium	Sand, mud, algae, reef, pavement (University of Hawaii 2013)

Figure 3. Sub-images extraction of multispectral images and depth. The pixels illustrate the surface reflectance of Sentinel-2, the white boxes are the window size of 9 × 9, and the dots indicate the water depth values. Spectral bands of each window correspond to the water depth value at the sub-image center.

Table 2. Accuracy assessment of SDB using CNN trained in AOI-1, AOI-2, AOI-3, and AOI-4.

AOI	RMSE (m)	R^2
AOI-1	1.22	0.95
AOI-2	1.63	0.93
AOI-3	3.03	0.55
AOI-4	0.73	0.94

Table 3. RMSE values per 5 m depth increment using CNN trained in AOI-1, AOI-2, AOI-3, and AOI-4.

AOI	RMSE (m)
	0-5 m	5-10 m	10-15 m	15-20 m
AOI-1	0.71	0.99	1.60	1.35
AOI-2	0.66	2.03	2.22	1.77
AOI-3	1.84	2.5	5.31	3.03
AOI-4	0.57	0.89	–	–

Figure 4. SDB accuracy per 1 m depth range using CNN trained in AOI-1 (orange), AOI-2 (grey), AOI-3 (yellow), and AOI-4 (blue).

Figure 5. AOI-1 SDB model (left) and cross profiles (right). The cross-section compares the SDB model using CNN trained (blue) and in-situ depth (green).

Figure 6. AOI-2 SDB model (left) and cross profiles (right). The cross-section compares the SDB model using CNN trained (blue) and in-situ depth (green).

Figure 7. AOI-3 SDB model (left) and cross profiles (right). The cross-section compares the SDB model using CNN trained (blue) and in-situ depth (green).

Figure 8. AOI-4 SDB model (left) and cross profiles (right). The cross-section compares the SDB model using CNN trained (blue) and in-situ depth (green).

Figure 9. SDB accuracy comparison between results with a single image and multi-temporal images.

Table 4. SDB accuracy in AOI-5 and AOI-6 using the trained model, pre-trained model of different study areas, and bathymetry model from Allen Coral Atlas.

CNN model	AOI-5		AOI-6
	RMSE (m)	R^2	RMSE (m)	R^2
Trained	0.79	0.98	1.34	0.94
Pre-trained on AOI-1	2.61	0.79	3.16	0.73
Pre-trained on AOI-2	2.92	0.69	2.98	0.71
Pre-trained on AOI-3	7.42	0.13	6.42	0.16
Pre-trained on AOI-4	6.84	−7.56	5.56	−4.72
Allen Coral Atlas	3.19	0.73	2.99	0.79
Sagawa et al. (2019)	1.24	0.93	–	–
Lyzenga, Malinas, and Tanis (2006)	–	–	3.01	–

Figure 10. Prediction vs reference depth plots in AOI-5 using different CNN models: trained (a); pre-trained on AOI-1 (b); pre-trained on AOI-2 (c); pre-trained on AOI-3 (d); pre-trained on AOI-4 (e); pre-trained on the combined AOI-1, AOI-2, and AOI-3 (f); pre-trained on the combined AOI-1, AOI-2, and AOI-4 (g); pre-trained on the combined AOI-1, AOI-2, AOI-3, and AOI-4 (h).

Figure 11. LiDAR bathymetric depths (a) and SDB results in AOI-5 using different CNN models: trained (b); pre-trained on AOI-1 (c); pre-trained on AOI-2 (d); pre-trained on AOI-3 (e); pre-trained on AOI-4 (f); pre-trained on the combined AOI-1, AOI-2, and AOI-3 (g); pre-trained on the combined AOI-1, AOI-2, and AOI-4 (h); pre-trained on the combined AOI-1, AOI-2, AOI-3, and AOI-4 (i).

Figure 12. Absolute error of SDB per pixel when using different CNN models: trained (a), pre-trained on AOI-1 (b); pre-trained on AOI-2 (c); pre-trained on AOI-3 (d); pre-trained on AOI-4 (e); pre-trained on the combined AOI-1, AOI-2, and AOI-3 (f); pre-trained on the combined AOI-1, AOI-2, and AOI-4 (g); pre-trained on the combined AOI-1, AOI-2, AOI-3, and AOI-4 (h).

Figure 13. Prediction vs reference depth plots in AOI-6 using different CNN models: trained (a); pre-trained on AOI-1 (b); pre-trained on AOI-2 (c); pre-trained on AOI-3 (d); pre-trained on AOI-4 (e); pre-trained on the combined AOI-1, AOI-2, and AOI-3 (f); pre-trained on the combined AOI-1, AOI-2, and AOI-4 (g); pre-trained on the combined AOI-1, AOI-2, AOI-3, and AOI-4 (h).

Figure 14. LiDAR bathymetric depths (a) and SDB results in AOI-6 using different CNN models: trained (b); pre-trained on AOI-1 (c); pre-trained on AOI-2 (d); pre-trained on AOI-3 (e); pre-trained on AOI-4 (f); pre-trained on the combined AOI-1, AOI-2, and AOI-4 (g); pre-trained on the combined AOI-1, AOI-2, and AOI-4 (h); pre-trained on the combined AOI-1, AOI-2, AOI-3, and AOI-4 (i).

Figure 15. Absolute error of SDB per pixel when using different CNN models: trained (a), pre-trained on AOI-1 (b); pre-trained on AOI-2 (c); pre-trained on AOI-3 (d); pre-trained on AOI-4 (e); pre-trained on the combined AOI-1, AOI-2, and AOI-3 (f); pre-trained on the combined AOI-1, AOI-2, and AOI-4 (g); pre-trained on the combined AOI-1, AOI-2, AOI-3, and AOI-4 (h).

Appeldoorn, R., D. Ballantine, I. Bejarano, M. Carlo, M. Nemeth, E. Otero, F. Pagan, H. Ruiz, N. Schizas, C. Sherman, et al. 2016. Mesophotic coral ecosystems under anthropogenic stress: A case study at Ponce, Puerto Rico. Coral Reefs 35 (1):63–75.

 Web of Science ®Google Scholar

National Oceanic and Atmospheric Administration (NOAA). 2017. “Benthic Habitat Mapping in Puerto Rico and the U.S. Virgin Islands for a Baseline Inventory.” 2017. https://coastalscience.noaa.gov/project/benthic-habitat-mapping-puerto-rico-virgin-islands/.

 Google Scholar

Bauer, L. J., E. Kimberly, K. K. Roberson, M. S. Kendall, S. Tormey, and T. A. Battista. 2012. Shallow-water benthic habitats of southwest Puerto Rico. NOAA Technical Memorandum NOAA NOS NCCOS 155:37. http://aquaticcommons.org/14709/1/SWPR_mapping_final_report_tagged.pdf.

 Google Scholar

Larsen, M. C., and R. M. Webb. 2009. Potential effects of runoff, fluvial sediment, and nutrient discharges on the coral reefs of Puerto Rico. Journal of Coastal Research 251:189–208.

 Web of Science ®Google Scholar

Puerto Rico Environmental Quality Board (PREQB). 2016. “Puerto Rico 305(b)/303(d) Integrated Report,” no. November: 243. http://www2.pr.gov/agencias/jca/Documents/AreasProgramáticas/AreaCalidaddeAgua/PlanesyProyectosEpeciales/303-305/2014/PuertoRico2014305b303dIntegratedReportAugust2014.pdf.

 Google Scholar

University of Hawaii. 2013. “Ohau Coastal Geology.” 2013. http://www.soest.hawaii.edu/coasts/publications/hawaiiCoastline/oahu.html.

 Google Scholar

Sagawa, T., Y. Yamashita, T. Okumura, and T. Yamanokuchi. 2019. Satellite derived bathymetry using machine learning and multi-temporal satellite images. Remote Sensing 11 (10):1155.

 Web of Science ®Google Scholar

Lyzenga, D. R., N. P. Malinas, and F. J. Tanis. 2006. Multispectral bathymetry using a simple physically based algorithm. IEEE Transactions on Geoscience and Remote Sensing 44 (8):2251–9.

 Web of Science ®Google Scholar

Data availability statement

A complete set of codes, including sample data, is provided in the Github repository at https://github.com/yustisiardhitasari/sdbcnn. The data that support the findings of this study are openly available in the Copernicus Open Access Hub at https://scihub.copernicus.eu/, the National Oceanic and Atmospheric Administration (NOAA) Bathymetric Data Viewer at https://www.ncei.noaa.gov/maps/bathymetry/ and the NOAA Data Access Viewer at https://coast.noaa.gov/dataviewer/. Map data copyrighted OpenStreetMap contributors and available from https://www.openstreetmap.org