An Integrated Remote Sensing and Machine Learning Framework for Coastal Margin Classification and Shallow-Water Bathymetry

These three Google Earth Engine workflows form an integrated framework for mapping coastal zone structure and shallow-water bathymetry using satellite remote sensing and machine learning. The process begins with generating a Sentinel-2 image composite, followed by a coastal margins classification, in which a Sentinel-2 surface reflectance composite is used to delineate major coastal environments. User-defined training polygons representing deep water, intermediate water, shallow water, and land are used to extract spectral information and train a Random Forest classifier. The trained model is applied to the Sentinel-2 composite to generate a spatially continuous coastal margins map, which is then spatially smoothed to reduce noise and converted into vector polygons. These coastal zone vectors provide a standardized representation of coastal structure and are exported for reuse as spatial masks and contextual layers in subsequent analyses.

Building on this coastal zoning, the second workflow focuses on estimating shallow-water bathymetry by integrating ICESat-2 laser altimetry data with Landsat 8 multispectral imagery. ICESat-2 bathymetry point datasets are merged and filtered to retain observations within the shallow-water coastal margin zone, ensuring that model training is constrained to areas where optical depth retrieval is physically meaningful. Depth values are converted into binned classes and randomly split into training and validation subsets. A Landsat 8 Collection 2 Level-2 image time series is assembled over the study region, with quality assurance masking applied to remove clouds, shadows, and invalid pixels, and radiometric scaling applied to surface reflectance bands. Spectral predictors, including visible bands and a simple green–blue ratio, are sampled at ICESat-2 training locations to train a Random Forest model that predicts water depth.

The trained bathymetry model is applied to each Landsat image to generate per-scene depth estimates, which are then aggregated through time to produce summary bathymetry products, including mean, median, standard deviation, and count of valid predictions. These outputs provide both a best-estimate depth surface and quantitative measures of uncertainty and data support. Together, the coastal margins classification and bathymetry modeling workflows enable a hierarchical characterization of coastal environments, first defining major coastal zones and then resolving spatially explicit water depth within shallow coastal waters, producing consistent, reusable datasets for coastal process studies, ecosystem mapping, and coastal resilience applications. 

This is the initial workflow that will be revised to include more comments and user documentation.