Evaluating multilinear and machine learning regression methods for Satellite-derived Bathymetry mapping using ICESat-2 and Sentinel-2 data on Google Earth Engine

Information of the water depth in coastal areas is important for navigation, risk assessment, disaster mitigation, coastal environmental management, marine habitat monitoring, and many other activities. Spaceborne shallow water bathymetry mapping has been an emerging practice to estimate the water depth where there are limited field data points. In this study, we evaluated five purely-spaceborne satellite-derived bathymetry (SDB) methods, namely clustered-based Lyzenga with 3 different regression functions (multilinear, robust multilinear, and ridge), Support Vector Machine regression, and Random Forest regression using multitemporal Sentinel-2 composites that were synthesized from over 800 images. We test these methods at the national-scale of Belize in optically shallow waters between the depth of 0 and 10 m and over 6,690 sq. km. We employed 641,562 bathymetry transect points derived from ICESat-2 ATL03 LiDAR (IS2) platform stratified within 2-m-interval to train the model and validated it with 3,100 field data acquired with CAMELS, Multibeam, and LiDAR sensors. We compared the performance of these five methods for SDB mapping by assessing the following error metrics: root mean square error (RMSE), mean absolute error (MAE), mean error (µ), and coefficient of determination (R2) over 2-m depth intervals as well. Our results show that the clustered-based Lyzenga method with multi-linear regression yields the best result, with error less than 10% of the maximum depth below the depth of 8 m. Beyond the depth of 8 m, the IS2-trained models tend to underestimate the depth when compared to the field data. This could show the lower sensitivity of the IS2 sensor on the deeper waters. We discuss preliminary benefits and challenges of big satellite data analytics, empirical and machine learning methods, and ICESat-2 data for national-scale satellite-derived bathymetry estimation