Sentinel-2 and ICESat-2 Data Fusion for Mapping Coastal Bathymetry and Ecosystem Structure

Kurzfassung

Ocean color in remotely sensed imagery is indicative of water depth and can be combined with in situ data to estimate benthic depth and model bathymetric surfaces. However, this has yet to be fully exploited at the regional scale due to the limitations faced in collecting sufficient in situ data for model calibration and validation. Here, we provide a method for combining Sentinel-2 optical imagery and ICESat-2 spaceborne lidar for overcoming this challenge. Discrete benthic depths are extracted from ICESat-2 data and combined with Sentinel-2 composite imagery to create spatially continuous bathymetric models for three study locations, comprised of the Bay of Biscayne (Florida), Gulf of Chania (Crete) and island nation of Bermuda. Image composites were created using the Google Earth Engine, using the 20th percentile reflectance per-pixel value from an image stack, in order to reduce image artifacts such as sun glint and turbidity. Both Top-Of-Atmosphere and Surface Reflectance imagery were compared in order to determine the costs and benefits of advanced optical image correction on model performance. We tested the two primaryreflectance-depth algorithms of Stumpf and Lyzenga, as well as Support Vector Machine (SVM) regression to determine the strongest performing algorithm. Across all of the study sites we successfully created 10 m spatial resolution spatially continuous nearshore ocean bathymetric surface models from Sentinel-2 imagery, trained with ICESat-2 derived depths. We achieved depths of up to 26 m with a low RMSE of 10%. We determined that Surface Reflectance data did not outperform TOA data, making advanced time and computationally expensive correction unnecessary to achieve accurate models. Across all study sites, the Lyzenga method outperformed other algorithms with the SVM algorithm consistently performing the most poorly. Independent NOAA Digital Elevation Models (DEMs) and in situ single sonar beam data were used to validate the models in Biscayne Bay/Bermuda and Gulf of Chania, respectively. The independent model at Bermuda was determined to be of inferior quality to the Sentinel-2 derived model, thus a purely spaceborne approach was developed which used ICESat-2 derived depths as both calibration and validation data. This was an important step to achieving a purely spaceborne estimate of nearshore coastal bathymetry and ecosystem structure. Coastal seascapes, composed of mangroves, seagrasses, coral reefs and tidal flats support a range of critical ecosystems. They support billions of livelihoods and generate billions of dollars in revenue. They provide 25% of the oceanic carbon pool and support as much as 25% of global biodiversity. However, our current lack of knowledge on coastal ecosystem structure prevents them from being adequately accounted and monitored. We provide a purely spaceborne method for the wall-to-wall mapping of sub-aquatic structure at 10 m spatial resolution, in order to overcome this knowledge gap. Moving forward, we also provide results from a preliminary automated workflow that can automatically detect surface depths from ICESat-2 data and create bathymetric maps on a per-scene basis, thus creating spatially continuous maps of uncertainty. Additional advances include the use of advanced bootstrapped machine learning models that performed with high accuracy. Our goal is to use spaceborne derived maps of nearshore coastal bathymetry to improve current estimates of sub-aquatic topography and facilitate important accurate submerged ecosystem accounting.