Date of Award

Fall 2025

Project Type

Thesis

Program or Major

Earth Sciences

Degree Name

Master of Science

First Advisor

Kim Lowell

Second Advisor

Yuri Rzhanov

Third Advisor

John G.W. Kelley

Abstract

Satellite-derived bathymetry (SDB) provides a cost-effective and scalable approach for large area bathymetric mapping in remote, shallow nearshore environments, where conventional techniques such as ship-based sonar or airborne LiDAR bathymetry (ALB) are often costly and logistically difficult. Since its launch in September 2018, NASA’s ICESat-2 satellite has emerged as a valuable source of cost-effective training data for SDB mapping. However, ICESat-2 photon event (PE) data exhibits significant noise, and existing literature has yet to thoroughly investigate whether algorithmic denoising methods can achieve comparable performance to the traditional manual extraction approach. Furthermore, ICESat-2’s limited depth penetration results in shallow-biased PE data, and it remains unclear whether depth weighted sampling improves SDB map accuracy compared to simply expanding the ICESat-2 data archive with more shallow-biased samples.

This study evaluated three ICESat-2 bathymetric data extraction techniques: manual identification (MAN), a progressive density-based filter (PDF), and a quality flag filter (QFF) based on ICESat-2’s internal PE classification. The denoised bathymetric datasets from each method were combined with Sentinel-2 multispectral imagery (MSI) and used in two types of bathymetric inversion models—linear regression (LR) and LightGBM—to produce SDB maps for an area centered on Key West, Florida. These maps were assessed for geographic consistency and depth accuracy against NOAA ALB reference depth data. Results from the LR model revealed minimal geographic variability in SDB map accuracy across all three denoising techniques, with consistent R² values ranging from 0.79 to 0.83 and RMSE values between 1.83 and 2.67 meters.

Additionally, this study examined how depth-weighted sampling affects SDB map accuracy by augmenting each of the three denoised datasets with simulated deep-water depths. While performance improvements were modest for the LR model (R² increases of 2–8%, RMSE reductions of 34–45%), the LightGBM model demonstrated significant gains in accuracy, with R² improvements of 33–88% and RMSE reductions of 71–82%.

In conclusion, the results indicated that algorithmic denoising methods (PDF and QFF) perform comparably to manual extraction, validating their use in scalable and efficient SDB workflows. Furthermore, the findings emphasized that improving SDB map accuracy requires more than increasing the volume of shallow-biased ICESat-2 PE data. Future research should focus on evaluating the accuracy of SDB maps across various coastal environments, including reefs, estuaries, and sediment-rich shorelines. Incorporating data from multiple seasons will also be crucial for understanding how variations in water quality and light conditions affect map accuracy. These efforts are essential for addressing global bathymetric data gaps and advancing coastal ocean mapping initiatives.

Download

Share

COinS