Global 1-km Coastal Bathymetry from Sentinel-2 Wave Inversion using the Satellite-to-Shores (S2hores) Toolbox Authors: Rafael Almar, Erwin W. J. Bergsma, Grégoire Thoumyre, Alain Giros, Patrick Marchesiello, Solange Lemai-Chenevier, Stéphanie Artigues, Sophie Loyer, Jean-Marc Delvit Published: 2025 ABSTRACT Reliable and up-to-date coastal bathymetry remains scarce for many of the world’s shorelines, limiting progress in coastal monitoring, hazard assessment, and numerical modeling. The Satellites to Shores (S2Shores) initiative addresses this gap by applying a wave-based inversion approach to more than one million Sentinel-2 images, producing a global coastal bathymetry dataset at 1-km spatial resolution. The open-source S2Shores pipeline integrates compositing, blending, and quality-control procedures optimized for large-scale, high-performance computing. Here we describe the dataset generation workflow, data curation steps, and global validation results. Comparisons against multibeam and chart-derived bathymetry show root mean square errors (RMSE) of 2–5 m and high consistency (R2 > 0.7) for depths between 10 and 40 m, where performance is most reliable. The effective retrieval depth varies regionally, reaching up to 60 m in energetic coastal environments. The dataset and accompanying source code are openly available, providing a reproducible global resource to support coastal research, climate studies, and operational forecasting systems. Background & Summary Coastal bathymetry is a key variable for coastal research, infrastructure planning, hazard mitigation, environmental management, and defense operations. However, traditional in situ surveys are costly, time-consuming, and often unavailable in remote or inaccessible regions. As a result, large portions of the world’s shallow coastal zones remain poorly mapped, with existing data frequently outdated or inconsistent across regions. With advances in satellite missions offering improved spatial resolution, temporal revisit, and spectral coverage, satellite remote sensing has emerged as an effective alternative for monitoring nearshore bathymetry globally1. Existing global bathymetric compilations, including the General Bathymetric Chart of the Oceans (GEBCO)2) and the NOAA ETOPO 2022 Global Relief Model3, integrate satellite-derived gravimetry with shipborne surveys and, where available, regional satellite-based estimates to achieve kilometer-scale global coverage. At the continental scale, the European Marine Observation and Data Network (EMODnet) provides a detailed, harmonized bathymetry for European seas4. In data-sparse regions, alternative digital nautical charts such as the C-MAP Global Electronic Sea Chart Database (CM-93;5 fuse multiple local surveys to provide detailed bottom features (e.g., channels and shoals), although these products are not formally recognized by hydrographic authorities. Several remote-sensing techniques have been used to derive nearshore bathymetry. Historically, satellite gravimetry has provided large-scale, deep-water estimates of seafloor topography6,7. More recently, the Sentinel constellation (ESA) has enabled optical and radar-based retrievals of shallow bathymetry. Radar missions have been employed to derive wave-based bathymetry8–10, a capability expected to improve further with the forthcoming high-resolution SWOT mission (CNES/NASA)11. Optical radiometric methods, based on color–depth inversion, are well established for clear-water environments12–14, while emerging lidar missions such as ICESat-2 offer promising results in very shallow, transparent waters15. Wave-based inversion approaches16–18 extend this capability to deeper, often turbid regions (30–80 m), depending on local wavelength conditions1. Each method presents specific advantages and limitations; their combination will likely underpin the next generation of global bathymetric products, potentially enhanced through machine learning19,20. At present, most satellite-derived bathymetric studies have been performed at regional or local scales, often requiring site-specific calibration with in situ data21,22. This limits methodological transferability and comparability across diverse coastal environments. As a first step toward a globally consistent framework, we developed a prototype based on Sentinel-2 imagery23, which forms the foundation of the present work. One of the main challenges is the management and processing of the large Sentinel-2 archive, now facilitated by high-performance datalake infrastructures such as Copernicus WEkEO and Google Earth Engine. In this study, we used the CNES TREX high-performance computing system, launched in 2023, to access and process imagery using the newly developed Satellites to Shores (S2Shores) pipeline24–26. Here, we present the first global coastal bathymetry dataset derived from S2Shores, produced at a spatial resolution of 1 km and optimized for regional-to-global coastal modeling. The dataset is generated from a spatio-temporal composite of Sentinel-2 wave-based estimates, blended with GEBCO to ensure seamless coverage from the shoreline to deeper waters. This initial release establishes a reproducible framework for global, satellite-based coastal bathymetry that will be progressively refined as new data and methodological advances become available. Data and Methods The S2Shores dataset was developed through a collaboration between CNES, LEGOS, and SHOM to produce a consistent, global coastal bathymetry product derived from Sentinel-2 imagery27. The initiative builds on earlier regional prototypes23,28, which demonstrated the feasibility of estimating shallow-water depths from satellite wave observations but also revealed limitations in data consistency, spatial coverage, and usability. These lessons informed the design of the present global processing workflow, optimized for high performance, reproducibility, and systematic quality control29. Overview of data processing In 2024, the S2Shores system was deployed at global scale on the CNES TREX high-performance computing platform30. Over 2.6 million Sentinel-2 Level-1C images (2019–2023) were processed, limited to scenes with less than 10% cloud cover and with at least 10% of pixels within 50 km of a coastline. The complete dataset generation process consists of three major stages: (i) extraction of wave characteristics from Sentinel-2 imagery, (ii) conversion of these properties into depth estimates using the linear wave-dispersion relationship, and (iii) multi-temporal compositing and blending with external reference bathymetry. Bathymetry estimation from Sentinel-2 imagery Wave patterns visible in Sentinel-2 bands B2 (green) and B4 (red) were analyzed to determine dominant wavelength, orientation, and phase celerity. A combination of Radon and Fourier transforms was applied to identify the prevailing wave direction and spectral peaks, from which the dominant wavelength and celerity were retrieved31. For each 800-m computation window, the linear dispersion relation was solved to infer depth, assuming negligible currents and uniform wave conditions. Estimates exceeding the deep-water or highly non-linear limits were excluded using a quality filter based on normalized celerity ratios. The method is applicable in coastal zones where the wave field remains influenced by the seabed (typically kh ≲ 1 − − 2, or depths less than about one-half of the local wavelength)24. Depths are referenced to Mean Sea Level by correcting for astronomical tide using the FES2022 global tidal atlas32 publicly accessible33. Selection and filtering Since the reliability of individual satellite scenes varies with sea state and image quality, a two-stage filtering process was applied. An initial broad selection generated preliminary depth estimates on a 1-km output grid consistent with GEBCO 2020 resolution. Each tile (approximately 12,000 km2) was then evaluated for spatial consistency, assuming a generally seaward-decreasing depth gradient. A Wald test on the cross-shore depth–distance relation was used to reject inconsistent results. The ten most coherent depth fields per tile were retained for further compositing (Fig. 1).Fig. 1Number of “best” S2Shores bathymetry estimates, derived from the systematic processing of all Sentinel-2 coastal images available on the new datalake. Number of “best” S2Shores bathymetry estimates, derived from the systematic processing of all Sentinel-2 coastal images available on the new datalake. Compositing and quality control Spatial and temporal compositing were performed to enhance coverage and reduce noise (Fig. 2). For each acquisition date, adjacent tiles within a common UTM grid zone were merged and filtered to remove isolated artifacts and high-frequency noise using variance and morphological operators. Missing values were filled by interpolation, constrained by GEBCO bathymetry where available. Temporal aggregation was achieved by computing the pixel-wise median across all valid dates (typically 5–30 per location), which suppresses transient effects from clouds, turbidity, or sensor noise. The resulting composite bathymetry was resampled to a global 30-arc-second grid (≈1 km) to ensure compatibility with global ocean and climate models. The entire processing chain, from image ingestion to compositing, is implemented in open-source code available at https://github.com/CNES/S2Shores.Fig. 2Description of the different processing steps from images spectral bands images (A) to a final bathymetry, including individual bathymetry estimation (B), through selection of best estimates (C) and spatio-temporal compositing (D) and blending with GEBCO (E). Description of the different processing steps from images spectral bands images (A) to a final bathymetry, including individual bathymetry estimation (B), through selection of best estimates (C) and spatio-temporal compositing (D) and blending with GEBCO (E). Blending with reference bathymetry To provide seamless transition from satellite-derived coastal data to deep-water regions, S2Shores composites were blended with the GEBCO 2020 bathymetric grid2. GEBCO was smoothed with a 10-km Gaussian kernel to match its effective resolution, and S2Shores data replaced GEBCO values in shallow areas (0–30 m). In a 1–2 km transition zone, the two datasets were linearly weighted using a cosine taper: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${D}_{blend}(x)=w(x){D}_{S2S}(x)+[1\,-\,w(x)]{D}_{GEBCO}(x)$$\end{document}Dblend(x)=w(x)DS2S(x)+[1-w(x)]DGEBCO(x) where w(x) varies smoothly from 1 (S2Shores) to 0 (GEBCO). A final low-pass filter ensured continuity across boundaries. This approach preserves high-resolution nearshore features while maintaining consistent regional depth structure. Preliminary validation against EMODnet Digital Bathymetry 20244 and multibeam surveys indicates improved accuracy (RMSE 2–5 m for 10–40 m depths) relative to GEBCO alone. Data Records The dataset is available at the Hydroweb data portal34 for visualization and download. The S2Shores global bathymetry dataset is distributed as a NetCDF-4 dataset containing S2Shores satellite-derived global nearshore bathymetry data blended with GEBCO 2020. This dataset provides global satellite-derived coastal bathymetry at 1 km resolution (30 arc-seconds), derived from multi-date Sentinel-2 wave inversion using the S2Shores framework. The gridded data spans a geographic extent for latitudes from –90° to +90° (21600 grid points) and longitudes from –180° to +180° (43200 grid points). Product Overview The main variables are Depth_S2Shores, the bathymetry derived from wave inversion using Sentinel-2 imagery in meters, and Depth_S2Shores_blended, the depth S2Shores blended with GEBCO to fill gaps and to extend coverage. Both stand for sea floor depth below mean sea level, positive values indicate depth below mean sea level. lat and lon are the latitude and latitude coordinates (regular spacing, 30 arc-sec), in degrees north and east, respectively. spatial_ref describe the coordinate reference system (WGS 84, EPSG:4326, WKT format). The vertical reference for depth is mean sea level. Positive values indicate depth below mean sea level (positive down). Global attributes include the dataset title (S2Shores Spatio-Temporal Global Bathymetry Median Composite from Sentinel-2 images), the producing institution (CNES–LEGOS), the creation date (2025-02-15T12:00:00), geographical sytem (epsg: EPSG:4326), the resolution in lon/lat (30 arc-second), the method used (SPATIAL DFT), and the version of the pipeline chain. The associated processing pipeline, installation and use notebooks are available as open source at https://github.com/CNES/S2Shores. Technical Validation Previous studies have validated its estimates against multibeam sounder baseline surveys in several regions, including the French Mediterranean coast (fetch-limited waves), the mud-dominated coast of Guyane, Guadeloupe in the Caribbean, and the Gironde estuary along the French Atlantic coast23. Additional validation was performed along the Moroccan coast35 and in the Gulf of Guinea28,36 using indicative digital charts (C-MAP CM93 v.2), although these lack standard quality controls (S-57 or International Hydrographic Organisation S-100 chart). The S2Shores estimates were also compared to the EMODnet reference bathymetric survey26, with correlations ranging from 0.45 to 0.71. Intermediate waters (10–40 m) yielded the most reliable depth estimates. In shallower areas, depths were generally overestimated, likely due to the influence of strong currents, turbidity, irregular wave fields, and amplitude dispersion effects37. At greater depths, environmental biases became more pronounced as the signal-to-noise ratio decreased and wave celerity approached its deep-water asymptotic limit. Despite successfully capturing important features such as the main channels of the Gironde estuary, high noise levels (errors up to 4-5 m) prevent Category of Zone of Confidence CATZOC-C qualification, with only 20% of the estimates meeting the classification criteria for depths below 30 m. Additionally, we present here a large-scale validation in the Bay of Bengal, a region with limited available data, where a regional bathymetric map was recently produced by combining various echosounder-derived datasets across the wide and shallow shelf38. The shelf has a median depth of approximately 10 m (Fig. 4) and includes morphological features such as channels that influence hydrodynamics and coastal hazard modelling39. In this context, the comparison between our satellite-derived bathymetry and the reference data yields a root mean square error (RMSE) of 2.57 m and a coefficient of determination (R2) of 0.79, indicating an overall agreement for this type of environment. Usage Notes The S2Shores global coastal bathymetry dataset34 represents the current best implementation of the S2Shores framework, encompassing image selection, wave-based bathymetric inversion, high-performance computing (HPC) processing, and post-processing through spatial and temporal compositing (Figs. 2,3). Extensive validation has been conducted across a wide range of coastal environments using independent datasets, including EMODnet, SHOM surveys, and nautical charts, as reported in previous studies19,23,24,26,28,35,36,40–42. Despite these validations, users should be aware of the inherent limitations of optical wave-based remote sensing, which depend on wave field regularity, atmospheric and illumination conditions, sensor viewing geometry, and local turbidity. Data should therefore be interpreted with appropriate caution, particularly in areas where these factors are known to be challenging.Fig. 3S2Shores bathymetry inversion principle, with (a) important steps from input images (here Sentinel-2) to the generation of the netcdf output file, (b) bathymetry inversion from pairs of waves characteristics using wavelength L, period T and phase φ shift or celerity c. (c) How wave characteristics are captured from wave train propagation in spectral bands (here Sentinel 2 B2 and B4 green and red bands43) at different times using Radon polar and discrete Fourier transforms on 800 m computation windows, to produce directionnal wave spectrum showed in (d). From pairs of celerity and wavelength, the bathymetry is inversed using the linear wave disperion theory, and mapped over full Sentinel-2 tiles of 100 by 100 km. Figure adapted from26,31.Fig. 4Comparison of S2Shores satellite bathymetry with in situ surveyed bathymetry at regional scale in the Bay of Bengal. (a) Surveyed bathymetry and (b) S2Shores satellite bathymetry on top of ESRI world map true color composite. (c) comparison of the two bathymetry datasets. Color stands for the density of points within 1° radius, and the black solid line shows the 1:1 ratio. S2Shores bathymetry inversion principle, with (a) important steps from input images (here Sentinel-2) to the generation of the netcdf output file, (b) bathymetry inversion from pairs of waves characteristics using wavelength L, period T and phase φ shift or celerity c. (c) How wave characteristics are captured from wave train propagation in spectral bands (here Sentinel 2 B2 and B4 green and red bands43) at different times using Radon polar and discrete Fourier transforms on 800 m computation windows, to produce directionnal wave spectrum showed in (d). From pairs of celerity and wavelength, the bathymetry is inversed using the linear wave disperion theory, and mapped over full Sentinel-2 tiles of 100 by 100 km. Figure adapted from26,31. Comparison of S2Shores satellite bathymetry with in situ surveyed bathymetry at regional scale in the Bay of Bengal. (a) Surveyed bathymetry and (b) S2Shores satellite bathymetry on top of ESRI world map true color composite. (c) comparison of the two bathymetry datasets. Color stands for the density of points within 1° radius, and the black solid line shows the 1:1 ratio. The dataset performs best in regions characterized by moderate to high wave energy and regular wave patterns, where multiple cloud-free images are available across different seasons. Examples include portions of the Atlantic coast of West Africa and the Bay of Bengal, where reliable depth estimates were obtained in intermediate waters (10–40 m), with root mean square errors (RMSE) below 3 m and correlation coefficients (R2) exceeding 0.7. Users should exercise caution when applying the dataset in regions with complex hydrodynamics or limited observational coverage. These include areas with cross-wave or multi-directional wave patterns, zones influenced by strong tidal or residual currents, low-energy or fetch-limited basins (e.g., the Mediterranean Sea), high-latitude or persistently cloudy regions (e.g., equatorial rainforest coasts), and steep or rocky shorelines, where the wave-based inversion is less effective. The dataset is also less reliable in very shallow intertidal zones and in deeper offshore regions beyond the effective sensing depth of the method. Ongoing development focuses on improving compositing strategies, refining error flagging and confidence indicators (e.g., based on wave geometry, detector timing, and linearity conditions), and extending coverage in data-sparse regions. Users are encouraged to consult method’s documentation before use and, where possible, compare S2Shores bathymetry with local reference data to assess accuracy for specific applications. Code architecture and availability The S2Shores toolbox is an open-source Python framework for estimating coastal bathymetry from satellite imagery. The full codebase, user documentation, and example workflows are publicly available at https://github.com/CNES/S2Shores, with additional installation and usage instructions provided at https://s2shores.readthedocs.io/en/latest/index.html and https://pypi.org/project/s2shores/#description. Step-by-step Jupyter notebooks are included to guide users through the processing chain and demonstrate typical applications. The processing framework consists of two main Python packages, illustrated in Fig. 5: S2Shores — a scientific library developed to estimate coastal bathymetry, shoreline morphology, and wave characteristics from Sentinel-2 and other optical satellite data. It applies advanced image-processing and inversion methods to extract wave parameters and compute nearshore water depth based on the linear wave-dispersion relation.BathyLauncher — a higher-level interface that automates the S2Shores workflow. It supports batch processing from various input formats (e.g., Sentinel-2 Level-1C or GeoTIFF), parameter management (e.g., gravity constant, wave period, shoreline distance), and seamless integration with high-performance computing (HPC) systems.Fig. 5Overall architecture of the S2Shores and BathyLauncher system for satellite-derived bathymetry. The diagram shows the interaction between input data (e.g., Regions of Interest, Delta_Times, distance-to-shore maps) and the main components of the framework. The workflow can be launched either via an Application Programming Interface (API) for automated communication with external software, or via a Command-Line Interface (CLI) for manual execution. Outputs are generated in NetCDF format, suitable for integration into geospatial or numerical modeling environments. Figure adapted from https://github.com/CNES/S2Shores. S2Shores — a scientific library developed to estimate coastal bathymetry, shoreline morphology, and wave characteristics from Sentinel-2 and other optical satellite data. It applies advanced image-processing and inversion methods to extract wave parameters and compute nearshore water depth based on the linear wave-dispersion relation. S2Shores — a scientific library developed to estimate coastal bathymetry, shoreline morphology, and wave characteristics from Sentinel-2 and other optical satellite data. It applies advanced image-processing and inversion methods to extract wave parameters and compute nearshore water depth based on the linear wave-dispersion relation. BathyLauncher — a higher-level interface that automates the S2Shores workflow. It supports batch processing from various input formats (e.g., Sentinel-2 Level-1C or GeoTIFF), parameter management (e.g., gravity constant, wave period, shoreline distance), and seamless integration with high-performance computing (HPC) systems. BathyLauncher — a higher-level interface that automates the S2Shores workflow. It supports batch processing from various input formats (e.g., Sentinel-2 Level-1C or GeoTIFF), parameter management (e.g., gravity constant, wave period, shoreline distance), and seamless integration with high-performance computing (HPC) systems. Overall architecture of the S2Shores and BathyLauncher system for satellite-derived bathymetry. The diagram shows the interaction between input data (e.g., Regions of Interest, Delta_Times, distance-to-shore maps) and the main components of the framework. The workflow can be launched either via an Application Programming Interface (API) for automated communication with external software, or via a Command-Line Interface (CLI) for manual execution. Outputs are generated in NetCDF format, suitable for integration into geospatial or numerical modeling environments. Figure adapted from https://github.com/CNES/S2Shores. Figure 5 summarizes the overall architecture of the toolbox. Input data (e.g., Regions of Interest, time windows defined by Delta_Times, and distance-to-shore maps) are processed via the BathyLauncher, which can be executed through an Application Programming Interface (API) for automated workflows or a Command-Line Interface (CLI) for direct terminal execution. The system produces bathymetric outputs in standard geospatial formats (e.g., NetCDF), ready for integration into hydrodynamic models or coastal GIS workflows. Both packages are written in Python using an object-oriented design to ensure modularity and extensibility. The implementation is optimized for both local and distributed HPC environments, with parallelization enabling efficient processing of large satellite datasets. This modular structure allows developers and users to integrate new inversion methods, input sources, or validation routines with minimal changes to the existing framework.