Satellite-Derived Bathymetry of Danish Coastal Waters Using Machine Learning with Sentinel-2 and ICESat-2 data

Abstract

Accurate shallow-water bathymetry is essential for navigation safety and coastal zone management, but conventional hydrographic surveys are costly and spatially limited, especially in dynamic nearshore environments. This thesis evaluates the newly released ICESat-2 ATL24 bathymetry product. It investigates its suitability as training data for satellite-derived bathymetry (SDB) using Sentinel-2 imagery. The workflow comprises two main steps (1) validation of ATL24 depths against in situ multibeam echosounder (MBES) measurements, and (2) development of SDB models using ATL24 as the target variable to estimate depth from Sentinel-2 reflectance. Four approaches are examined: the Stumpf Ratio Method (SRM) and three machine-learning (ML) models - Random Forest (RF), CatBoost (CB), and Multilayer Perceptron (MLP). The study focuses on two Danish coastal sites with contrasting seabed conditions: Jammerbugten (depths between 0-15m) and Nordfyn (0-8 m). Validation reveals that ATL24 depths have higher uncertainty at Jammerbugten with a Root Mean Square Error (RMSE) of 0.72 m but closer agreement at Nordfyn (RMSE of 0.27 m). Across both sites, the ML models consistently outperformed SRM: both CB and MLP achieve RMSE values of 1.02 m at Jammerbugten and MLP reaches 1.07 m at Nordfyn. Seabed characteristics strongly influence model performance: the uniform sandy seabed at Jammerbugten promotes a stable reflectance-depth relationship, whereas the spatially variable substrate at Nordfyn introduces uncertainty, particularly at shallow depths where SRM performs slightly better. This study demonstrates that combining ICESat-2 ATL24 with Sentinel-2 enables useful large-area coastal screening to detect zones of potential depth change where targeted surveys may be prioritized.

Popular Abstract

Mapping the seafloor in coastal waters is essential for safe navigation, coastal planning, and environmental management. Despite this importance, large parts of shallow coastal waters remain poorly mapped, because traditional surveys that measure water depth using ships are costly and time- consuming. In addition, coastal environments can change over time, which increases the value of having broader and more frequent mapping.
An alternative way to obtain information about water depth is through satellite observations. In shallow water, sunlight can reach the seabed and reflect back toward the satellite sensor, and the amount of reflected light depends partly on water depth. By analysing satellite images, it is therefore possible to estimate water depth over large coastal areas, an approach known as satellite-derived bathymetry.
In this project, satellite-derived bathymetry is used to map shallow coastal waters at two study sites in Denmark: Jammerbugten on the northwest coast of Jutland and Nordfyn along the northern coast of Funen. The analysis combines satellite images from Sentinel-2 with satellite-based depth information from the ICESat-2 mission. The depth data are obtained from a newly released product called ATL24, which provides water depth measurements along narrow satellite tracks. These ATL24 depths are used as reference information to train machine-learning models that estimate water depth from Sentinel-2 images, allowing the seafloor to be mapped continuously across larger coastal areas.
The results show that satellite-derived bathymetry can reach high accuracy under favourable conditions. At Jammerbugten, the best results show an average depth error of about 1.02 metres in waters between 0 and 15 metres deep. This means that water depth can be estimated with an accuracy close to one metre over large areas without the need for ships or aircraft. At Nordfyn, where seabed and water conditions are more complex, the depth estimates are generally less accurate, demonstrating how local conditions influence performance.
Although this approach cannot replace traditional measurements, it is particularly useful for screening large coastal areas and identifying where variations in water depth may have occurred. Such information can support coastal management, environmental monitoring, and the planning of future hydrographic surveys by helping to prioritise where more detailed measurements are needed along the Danish coast.

• author: Kristoffersen, Freja
• supervisor: Zheng Duan
• organization: Dept of Physical Geography and Ecosystem Science
• course: GISM01 20252
• year: 2026
• type: H2 - Master's Degree (Two Years)
• subject: Earth and Environmental Sciences
• keywords: Geography, GIS, Satellite-Derived Bathymetry, ICESat-2, ATL24, Machine Learning
• publication/series: Master Thesis in Geographic Information Science
• report number: 204
• language: English
• id: 9222699
• date added to LUP: 2026-02-12 16:50:55
• date last changed: 2026-02-12 16:50:55