Several planets in our solar system are mapped far better than our own planet. Unfortunately we are willing to spend more to map other planets than our own. Typical mars mission costs around $3B U.S which is enough to map our world oceans using existing sonar technology. The foundational ocean floor or Bathymetry data provides basic infrastructure for scientific, economic, educational, managerial and political work. More specifically bathymetric data had wide range of applications including the production of nautical charts for navigation, predicting sediment budget and dredging operations, communication cable and pipeline planning, tsunami hazard assessment, benetic habitat mapping and also on the territorial claims under the Law of the Sea.

Despite the importance of bathymetric data only a small fraction of ocean floor has been mapped. Ideally there should be the use of modern multibeam-singlebeam ecosounder to produce high resolution bathymetric maps. Similar like sonar ship/aircraft based Bathymeytric Lidar system can also be deployed to develop sea bed Bathymetry. However the use of such techniques are very time consuming and expensive. National hydrographic agencies widely using this techniques to develop bathymetric map. Alternatively radar altimeter mounted on an orbiting spacecraft can be used to map the deep ocean surface at global scale. In this technique bathymetry is produced by observing variations in ocean surface height which is proportional to the pull of gravity caused by seafloor topography. For the last four decade multi satellite missions such as Geosat, ERS-1, Cryosat-2, SARAL/AltiKa, Jason-2, ICESat, Sentinel-3, Sentinel-5P is deployed with radar altimeter instrument. Since the beginning of their deployment they are collecting data and improving the marine gravity models. For example Jason-2 altimeter data tested accuracy indicated 3-5 cm. However considering the spatial resolution of altimetry data, still it has coarse resolution (approximately single point every 500 meter interval) and had less consistency at the coastal areas. Altimeter data is also affetced by atmospheric conditions and need to apply complex atmospheric algorithm to process altimeter data.

Recognizing the availability of multi sourced data on the world ocean's, an international initiative has been made by the GEBCO (General Bathymetry Chart of the Oceans) and the Nippon Foundation to complete mapping of the world oceans by 2030. The latest compilation of global bathymetric data product released by GEBCO known as GEBCO_2020 Grid is freely available. This bathymetric data product is produced using combination of shipboard soundings and depths predicted using satellite altimetry on a 15 arc-second grid.

Another exciting approach of deriving batheymetry at shallow water environment. is to use optical remote sensing. Compared with traditional system either using ships or aircraft, optical RS requires only 5% cost to map vast areas of the near shore sea floor. With this techniques without mobilizing people and equipment's like aircraft or ships, vast areas can be mapped quickly. No legal permits are required to accomplish such process. This is also hazard free and carbon neutral operation compared with conventional techniques (using aircraft or ships). International boarders are another critical areas where its almost impossible to mobilize such equipment's due to its political sensitivity.

The science behind optical satellite bathymetry is the Intensity or light decreases exponentially while it penetrates water. This formula can be used in remote sensing to collect data over water. Not all part of the electromagnetic spectrum attenuates equally inside water. The red and infra-red light attenuates more rapidly than shorter wavelength Blue light. With the Increasing depth in water, there is high absorption of longer wavelength and bottom features becoming less visible. Under a favorable environmental setting water depth up to 25 m can be estimated accurately. In such technique water penetrable wavelengths 490 nm – 665 nm from visible light can be utilized to draw bathymetry. Satellite batheymetry using optical remote sensing also had proved to produce similar results like Lidar and recovered very similar resolution. The most suitable commercial satellite available for satellite bathymetry is the WorldView-2 image. The satellite has water penetrable coastal blue band can penetrate water depth at 25 to 30 m and its availability broadly revolutionized satellite bathymetry. This dedicated band is more penetrable in water column and bottom features are more visible compared with red and green band. Another important point to be considered for satellite Bathymetry is the atmosphere. The quality of the bathymetric product significantly depends on the applied atmospheric corrections. Corrections are also needed to apply on sun glint areas when there are sun glints on the ocean water. Deglinted image produces better quality bethymetry. Water quality is another factor affecting quality of bathymetric product. Depth estimation at clear water is more accurate than turbid water. Temperature and salinity can also be another factor for such assessment.

In this research we used relatively high resolution Multi Spectral Instrument(MSI) onboard Sentinel -2A and 2B to generate high resolution bathymetric maps. Complied data product from Multi mission satellite altimeter and batthymetric soundings used as an input to calibrate and validate the model. We used multi mission satellite altimetry data product, however we dident need to process such data product. Because complied synthesis of multimission data products had already done by GEBCO, and we only extracted indirect measures from GEBCO 2020 Grid which includes 'TID44-Bathymetric sounding' according to their definition “Depth value at this location is constrained by bathymetric sounding(s) within a gridded data set where interpolation between sounding points is guided by satellite-derived gravity data” and “TID-40 Prdiction based on satellite-derived gravity data - depth value is an interpolated value guided by satellite-derived gravity data” We had tested the both ratio and linear models on the Bay of Bangle over Saint Martin island. On this particular environmental setting at the test site it was observed that the performance of linear model is better than the ratio model and derived R Square value was around .80. We had found Sentinel-2 10 m MSI instrument was able to capture small-scale features, such as tidal channels, straits relevant to navigation or steep slopes.

The idea was to explore the potentials of optical remote sensing on ocean science. Therefore started with the challenge on the use of optical remote sensing to retrieve shallow water bathymetry at Saint Martin Island on the Bay of Bangle. The entire project was implemented using open sourced data and tools(ESA's SNAP SeNtinel's Application Platform, QGIS). The main idea is to retrieve high resolution bathymetry (10m) from Sentinel 2 satellite data using popular bathymetric models. The models were calibrated and validated using available open source ocean depth data. At the beginning explored available bathymetric data around Saint Martin Island. Bangladesh Navy's hydrographic charts or Bangladesh Water Development Board's Topographic maps were found to be most suitable, However they do not have open access. Therefore ocean depths surrounding the test site extracted from global compilation of ocean bathymetry (e.g., International Hydrographic Organization (IHO), Global Multi-Resolution Topography (GMRT), General Bathymetric Chart of the Oceans (GEBCO)). It was observed that ocean depths from GEBCO 2020 Grid produced under Seabed 2030 project had a cell-spacing of 15 arc sec composed of single-beam and multibeam echo sounders, superimposed on a base derived from satellite altimetry data, can be used this study. However on the test site GEBCO 2020 grid had only coarse resolution indirect (satellite altimetry derived bathymetry) measurement depths. Therefore it could also be mentioned that this research tries to derive shallow water bathymetry from multi mission based measures(satellite altimetry derived bathymetry, Bathymetric Sounding and Sentinel 2 Data) without using any ground truth or in situ data. Although there will be uncertainties on the both data sources because changing topography due to sedimentation, sun glint problems and temporal variations but this will definitely show an interrelationship between two data sources and open up possibilities to evaluate altimetry derived depth from other satellite sensors as well as performance of bathymetric models.

Initially Sentinel 2 data products from Multispectral instrument known Top of Atmosphere Resistances in cartographic geometry, L1c were atmospherically corrected and produced Bottom of Atmosphere Resistances in Cartographic Geometry, L2a data product. Then sunglint corrections were applied on the L2a data product. Finally bathymetric models simulated on sun glint corrected images and calibrated using randomly selected sample depths from GEBCO 2020 grid. A second sample set from GEBCO 2020 grid were also used to validate the results using correlative plot. Tested band ration Bathymetric Models known as Empirical Bathymetry and resulted R square value was more or less than .65. Later develop linear model using only single Blue band and applied low pass 3X3 AM filter as well as corrected the sample depth by eliminating outliers. This time the resulted r square value was around .80. Finally developed several cross section of the bathemety and compared results between derived bathymetry from two different models and GEBCO 2020 grid. We just realized that, It is possible to apply machine learning models and use national hydrographic agency's data to calibrate the model and produce more accurate result. A generic overview of the methodology we followed is presented in the following link. The methodology can be found inside UNESCO Remote Sensing handbook for Tropical Coastal Management.

https://drive.google.com/file/d/17-YfsH_ej1cms2vmNy294uJnQeUf_sLA/view?usp=sharing

Please see the following steps presenting the way I used space agencies data for this project

Step 01 Data Selection and Download

We approached to the Copernicus Open Access Hub using the following weblink (https://scihub.copernicus.eu/) to visually inspect and select suitable image for this application. As mentioned we selected Sentinel -2 images for this application. First criteria for image selection was less cloud coverage which can be automatically selected from the platform. Our last two criteria was clear waster and sun glint free image. Using Initial visual inspection I had selected the following images.

https://drive.google.com/file/d/1P0vzWMYOJk1UZ7KLLXQztI_9zLRMtooo/view?usp=sharing

Images are presented with their captured date. However it was almost impossible to have sun glint and cloud free images on the test site. Based on the subjective judgement Sentinel -2 images captured over Saint Martin island on 8th May, 2020 were selected for the following procedure. Its important to mention that the downloaded Sentinel 2 data product from Multispectral Instrument was not atmospherically corrected and it is a Top of Atmosphere Reflectance L1c data product

Step 02:Atmospheric Correction and Image Subset

Various atmospheric correction modules are available. Somme of them are Sen2Cor, Acolite, OPERA, 6s/MODTRAN LUTs etc. In this research we used Sen2Cor processor to generate Level 2A data product. Sen2Cor performs atmospheric, terrain and cirrus correction for Top-Of-Atmosphere Level 1C input data and produce Bottom of Atmosphere Reflectance's L2a data product. The process is implemented on command line interface on SNAP software. After applying atmospheric correction a subset image over the test site is produced using Sentinel B2,B3,B4,B8 spectral band. A true color composition of atmospherically corrected Sentinel level 2a data product should look like the following image.

https://drive.google.com/file/d/1wiI8Z0EroGtnc83sKjulpUtiavNignlZ/view?usp=sharing

Step 03. Land Mask

Land mask layer was created using NIR band. Threshold value for land and water was determined to separate the land from sea. After applying the land mask to RGB bands the sea areas should look like the following image.

https://drive.google.com/file/d/1xZQVl5efGHIfWXxugfmFemPXqIMcgTSn/view?usp=sharing

Step 04. Sun Glint Correction

This type of phenomena occurred in the satellite images when there is an specular reflection of the sun on water surfaces. Corrections are applied to removes sun glint reflected waves by using NIR band. The process improve visual appearance of the data product. Its important to apply such correction as our image is affected by this type of contamination. There are several sun glint removal algorithms available, In this exercise I used the one developed by Headley et al.,(2005). Deglint is implement on spreadsheet using regression analysis.

The equation described by Hedley et al. (2005) for the deglinting of a multispectral image is: R’i = Ri–bi(RNIR–MinNIR)

Where R’i is the deglinted pixel in band i

Ri is the reflectance from visible band i

bi is the regression slope

RNIR is the NIR band value

MinNIR the minimum NIR value of the sample

Theoretically such procedure tries to scale the relationship between sun glint and NIR signals by using one more sample images of the affected area. Linear regression between visible and NIR is performed over deep water. “The slope of the regression then used to predict the brightness in the visible band that would be expected if those pixels had a NIR value of MinNIR”. A graphical presentation of this correction techniques (Headley et al,2005) are presented on the following link.

https://drive.google.com/file/d/1CiyJB4x3_XJ1VCzW5ULmwvAPkc_8FINB/view?usp=sharing

I applied the following statistics on visible bands to produce deglinted image

Blue Band

Reference band: NIR

Min NIR Computed:0.013600

MinNIRUsed:0.013600

Regression slope = 0.579346

Regression R - squared= 0.375671

Green band

Reference band: NIR

Min NIR Computed:0.013600

Regression slope = 0.506637

Regression R - squared= 0.176810

Red band

Reference band: NIR

Min NIR Computed:0.013600

Regression slope = 0.770151

Regression R - squared= 0.667050

https://drive.google.com/file/d/1atblv9IYQqy65Kii0qwu22KJBoTccllD/view?usp=sharing

Images on the left shows original RGB composition and deglinted image on the right

Step 05. Dark Object Substraction

Substracting the atmospherically offset from every pixel in the band can generate the corrected image. As this offset value is different in each band therefore the histogram cut-off point is applied separately on each band.  

Step 06. GEBCO 2020 Grid and Satellite Derived Bathymetry

As discussed earlier Multi mission compilation from GEBCO 2020 Grid is used for this study. Minor level stretching were applied on GEBCO 2020 Grid to present the depth variations at the study area. Initially 250 sample points were randmoly selected as shown on the following figure to simulate Ratio based bathymetry model. At later stage based on using visual inspection 50 spot heights were omitted from the analysis due to its inconsistency. Finally 200 sample points were used to drive linear based bathymetry model. Following figure shows GEBCO 2020 Grid with the extracted sample depths. Depth values labeled beside the samples.

https://drive.google.com/file/d/12EThqk7U3W0l-RqqEr4OMtPhK2n2l3aJ/view?usp=sharing

AS long as I had decent size ground truth data sets then its time to simulate the bathymetric model. Empirical model works on the principle that each band has a different absorption level over water and this diversity level theoretically will produce the ratio between band. (Stumpf et al. 2003). With this theory depth can be defined using the following formula.

𝑍= 𝑚1 [ln(𝑛𝑅𝑤(𝜆𝑖))/ln(𝑛𝑅𝑤(𝜆𝑗))] + 𝑚0

where n is a user-set constant which should be chosen to ensure the logarithms are positive and the relationship is maximally linear.

The Rw terms are the atmospherically corrected reflectance's in two bands i and j.

m1 and m0 are constant and calculated by calibrating dataset of points with known depths .

The ratio is applied for the pair of Blue-Green (B2-B3) bands and for n = 1000.

It is important to mention that from GEBCO 2020 Grid we only need to calculate the constant m1 and m0.

We also developed correlative plot using the predicted depth and observed depth.

https://drive.google.com/file/d/1F-bSBqmqOYqX3E_TRzY3x1In7JOku8v1/view?usp=sharing

As shown on the above figure values are more aligned to the X axis or observed depth than the predicted values or y axis. It is also apparent that the R square value is .62 which needed to be improved. It is also visible from the chart that prediction is not following the line after 32 meter since the model works for shallow water bathymetry. The calculated bathymetry using Empirical model should look like the following image.

https://drive.google.com/file/d/1evebNQrYkwnTGjH_l5AklIL-2vzD-uTY/view?usp=sharing

One reason why the ratio model couldn't perform well is because the bottom features are equally visible on both spectral bonds(B2, B3), We just noticed that the blue, greeen, red or even the NIR band can easily distinguish bottom features and water was very clear. Reflectance of NIR and Red band should be fully absorbed on certain water depth however due to shallow condition the bottom is becoming visible. The issue also affected while applying sun glint corrections, subsurface NIR reflectance overcorrect the shallow water areas and may appeared dark. Also observed that there are less variability on the bottom type. The entire bottom had unique type of bright sand. Another reason for low performance is because there are outliers in the calibrated database. A visual inspection was made on the 250 GEBCO sample depths at least 50 outliers found in the database and removed. Some of the outliers could easily be detected as they are having high depth value located close to the shoreline. Similar others extreme values can also be identified by making comparison with RGB composition. As discussed earlier subsurface features are clearly visible and can easily be predicted the high depth and low depth areas therefore any observations doesn't comply with such argument was removed form sample depths. Another way to detect the abnormalities on the input data is observing homogeneity of the depth values. In QGIS depth values were plotted and labeled, with this process regional outliers were easily visible.

Step 07. Developing Linear Model

To get advantage of such environmental settings, We understand that it is more logical to apply linear model which could address the issue more efficiently. Next we developed our own version of bathymetric model using single band (B2). A linear regression is implemented between the reflectance of blue band and corrected GEBCO sample heights databases. This time we found strong correlation with the with the brightness of the blue band and the observed depths. https://drive.google.com/file/d/1J7szMNYFKLTogW0msF7izrksu-_DzvGd/view?usp=sharing

Initially R square value was .74, after applying a Log transformation on the sampled depth R square value increased to .78.

https://drive.google.com/file/d/1Dke97sD7gRu-RB4FEs4blcIGc3tFNyG_/view?usp=sharing

Using the slope and intercept from this regression line, brightness value from blue band is scaled ocean depth. With the linear transformation we produced our primary version of bathymetrioc product and finally a low pass filter with 3x3 AM applied on the bathymetric product and resulted the R square value was .78835. Following image shows GEBCO 220 Grid on the left and Sentinel -2 satellite derived bathymetry on the right.

https://drive.google.com/file/d/1gdWzh8yObONXl2gE2K4fVsmqKY3_Sncy/view?usp=sharing

https://drive.google.com/file/d/1uPwnv1TJOVgFfjbMDKFl_8Unf_d7G9xl/view?usp=sharing

The output for the project is a single raster file presenting ocean depths on each pixel. From SNAP tool It is possible to export the the file in multiple formats. In addition there will be certain other files like the final co-relative plot presenting prediction and observed value can also be added as an image file on the dashboard.

Profile plot is another interesting thing for bathymetric data can be integrated to the dashboard as a dynamic tool. Which will allow the user to draw a line anywhere inside the the bathymetric image and automatically produce cross section underneath the drawings. I had already seen similar tools available on the dashboard. With this limited timeframe I was not able to develop such tools but definitely be able to produce such tools If I had enough time for this project.

Stumpf, R. P., Holderied, K., & Sinclair, M. (2003). Determination of water depth with high-resolution satellite imagery over variable bottom types. Limnol. Oceanogr., 48, 547–556.

Hedley, J. D., Harborne, A. R. and Mumby, P. J., "Simple and robust removal of sun glint for mapping shallowwater benthos," Int. Journal of Remote Sensing, 26(10), 2107-2112 (2005).

Lyzenga, D. R. (1978). passive remote sensing techniques for mapping water depth and bottom features. Applied Optics, 17(3), 379–383.

Mayer, L. et al.,(2018)The Nippon Foundation—GEBCO Seabed 2030 Project: The Quest to See theWorld’s Oceans Completely Mapped by 2030, Geosciences

D.T. Sandwell et al.,(2006), Bathymetry from space: Rationale and requirements for a new, high-resolution altimetric mission,C. R. Geoscience 338 (2006) 1049–1062

Ryan, W. B. F., et al. (2009), Global Multi-Resolution Topography synthesis, Geochem. Geophys. Geosyst., 10, Q03014, doi:10.1029/2008GC002332.

D.T. Sandwell et al., (2019), Global bathymetry and topography at 15 arc sec: SRTM15+. Earth and Space Science 6 . https://doi.org/10.1029/2019EA000658

SEOM S2-4Sci Land & Water: Coral Reefs 2016, European Space Agency Accessed 27 June 2021

<https://sen2coral.argans.co.uk/>

The Copernicus Open Access Hub 2021, European Space Agency Accessed 27 June 2021<(https://scihub.copernicus.eu/)>

Data center for digital Bathymetry viewer 2021, International Hydrographic Organization, Accessed 27 June 2021<https://maps.ngdc.noaa.gov/viewers/iho_dcdb/>

This project has been submitted for consideration during the Judging process.