We develop a high-resolution bathymetric dataset, Satellite Derived Bathymetry Elevation Dataset (SDBED), covering the nearshore region of Belize. The SDBED is developed by integrating satellite-derived bathymetry dataset for clear shallow waters using ICESat-2 and Landsat imagery with data from nautical charts for the turbid lagoon, and previously published reef profile measurements (Kobara & Heyman, 2010) to capture sharp slopes in the fore-reef zones.

Satellite-Derived Bathymetry (SDBED) for the nearshore regions of Belize

We develop a Nautical and Slope Augmented Satellite-Derived Bathymetry dataset (SDBED) up to the 200 m depth contour for the Belize nearshore region that combines information from nautical charts, satellite-derived bathymetry, field-based slope contours and GEBCO data. As a first step, we delineate the coastal margins of Belize into four distinct groups based on spectral characteristics; Land, Shallow Water, Intermediate Water, and Deep Water. To generate this layer, we created a 20th-percentile composite of Sentinel-2 images acquired between June 2020 and June 2023 which reduces the influence of clouds and transient turbidity (Traganos et al., 2018). We then apply a Random Forest (RF) classification model trained to the four separate categories based on visual interpretation of the Sentinel-2 20th percentile composite. The coastal zone classification output defines where ICESat-2 bathymetric photons can be reliably collected and identifies regions where optical satellite-derived bathymetry is limited by turbidity and depth. This coastal margin layer provides the spatial mask for integrated satellite-derived bathymetry within the shallow and intermediate classes, but nautical charts or GEBCO within the deep-water class.

We access ICESat-2 data for the study regions from the National Snow and Ice Data Center (NSIDC) using the icepyx Python software (The icepyx Developers, 2023). We categorized bathymetric photons using C-SHELPh, a python algorithm developed by (Thomas et al., 2022) to efficiently extract bathymetric photons from noise. Following the automated processing, we perform a final visual inspection of the data to remove erroneous points such as shallow returns in deep water from wave troughs or noise from the water column. In Belize, we observe that ICESat-2 can penetrate to a maximum depth of ~23 m due to water quality and turbidity, which can scatter and absorb the laser light, reducing the effectiveness of the sensor in capturing accurate bathymetric measurements

We acquired Landsat 8 OLI atmospherically corrected surface reflectance (Collection 2/Tier 1) scenes (over a one-year period (2022-01-01 to 2022-12-01) for the study region via the Google Earth Engine (GEE) catalog and preprocessed these data for satellite-derived bathymetry modeling. For these analyses, we focused on the visible spectrum bands including the blue (SR_B2), green (SR_B3), red (SR_B4), and a calculated ratio of blue/green bands. Each image was pre-processed to remove clouds and cloud-shadows using standard cloud masking procedures (Coluzzi et al., 2018). These spectral variables were then paired with the ICESat-2 bathymetric photon depths to train Random Forest (RF) regression models, enabling the statistical relationship between water surface reflectance and seafloor depth. The ICESat-2 water depth data was randomly split into training and validation data to training and test model performance. We applied two separate training/validation in order to have an equitable distribution of points across shallow and intermediate water depths, as a large majority of points fall within the shallow water zone. For the ICESat-2 water depth classified photons that intersected with the shallow coastal zone we randomly selected 5% for training and 95% for validation following similar procedures in Thomas et al, 2022, but we do not ground-truth the predicted depths with sonar data. As a modification and to ensure model training with deeper water depths, ICESat-2 data that intersected with the intermediate water depth class were split into 70% for training and 30% for validation. The water depth training data from ICESat-2 was then combined with the corresponding pixel-wise Landsat 8 spectral data for each available cloud-masked image over the 1-year period, resulting in multiple water depth maps. Finally, reduce the collection of water depth maps to a single 2022 bathymetry map using descriptive statistics (mean, median, and standard deviation) following (Thomas et al., 2022). The reduction to a single composite allows for the generation of a robust bathymetric model with high accuracy and low root mean square error (RMSE) values. It also enables the spatial distribution of uncertainty to be mapped across the study region at high resolution (Thomas et al., 2022). We use publicly available nautical chart data (www.gpsnauticalcharts.com/) within the 0 – 200 m contour to fill in gaps in the bathymetry that occur outside the shallow and intermediate coastal zones used in the Satellite-Derived Bathymetry. Using the nautical charts, we hand-digitized sounding measurements across the lagoon and tidal channels. These individual points were imported into ArcGIS Pro and used in a surface contour map to create an interpolated seafloor map. For additional gaps along the reef edges, we use slope profiles for Turneffe Atoll, and Lighthouse Reef from (Kobara & Heyman, 2010) who conducted field measurements of profile elevations along multiple reef edges of the central Belize shoreline. We use the median profile for each reef atoll and linearly extend this up to 200 m. Beyond the 200 m depth contour we merge our dataset with GEBCO. The SDBED, nautical chart, profile and offshore GEBCO datasets are merged into a seamless bathymetry dataset in ArcGIS Pro. In short, SDBED integrates ICESat-2-calibrated Landsat-derived depths (0-15.7 m) with nautical charts (turbid and deeper areas) and reef profiles (sharp slopes), merged seamlessly in ArcGIS Pro using kriging interpolation at the 0 m contour.