Study of various machine learning approaches for Sentinel-2 derived bathymetry.

Area of Science:

• Remote Sensing
• Coastal Geomorphology
• Machine Learning Applications

Background:

• Accurate seabed depth (bathymetry) is crucial for coastal operations, but in-situ measurements are costly and time-consuming.
• Increasing availability of satellite imagery presents an opportunity for developing cost-effective and rapid bathymetry estimation methods.
• Existing methods include analytical models, physical models, and empirical techniques, with varying performance in optically complex waters.

Purpose of the Study:

• To compare the performance and precision of common analytical bathymetry estimation models against modern machine learning algorithms.
• To investigate the efficacy of shallow neural networks, decision trees, and Random Forest algorithms for satellite-derived bathymetry.
• To evaluate the impact of input data features (satellite bands, geographical weighting) on machine learning model accuracy.

Main Methods:

• Implementation and evaluation of machine learning models including shallow neural networks, decision trees, and Random Forest.
• Utilizing raw reflectance data from four satellite bands and specific band ratios (e.g., B2/B3 logarithm quotient) as input features.
• Incorporating geographical weighting into the machine learning training process to account for spatial variations.

Main Results:

• The regression tree algorithm demonstrated superior performance compared to other machine learning models and analytical approaches.
• Optimal performance was achieved using a regression tree model that integrated data localization, four satellite bands, and a quotient of logarithms of B2 and B3 bands.
• The study was conducted in the optically challenging and spatially variable waters of the South Baltic coastline near Szczecin and Hel Peninsula, Poland.

Conclusions:

• Machine learning, specifically regression trees, provides a highly effective method for satellite-derived bathymetry estimation in coastal areas.
• The selection of appropriate satellite data bands and the inclusion of spatial information significantly enhance model accuracy.
• This approach offers a viable alternative to traditional methods, enabling more frequent and cost-effective seabed depth assessments.