High-Coverage Satellite-Based Coastal Bathymetry through a Fusion of Physical and Learning Methods

Abstract

An up-to-date knowledge of water depth is essential for a wide range of coastal activities, such as navigation, fishing, study of coastal erosion, or the observation of the rise of water levels due to climate change. This paper presents a coastal bathymetry estimation method that takes a single satellite acquisition as input, aimed at scenarios where in situ data are not available or would be too costly to obtain. The method uses free multispectral images that are easy to obtain for any region of the globe from sources such as the Sentinel-2 or Landsat-8 satellites. In order to address the shortcomings of existing image-only approaches (low resolution, scarce spatial coverage especially in the shallow water zones, dependence on specific physical conditions) we derive a new bathymetry estimation approach that combines a physical wave model with a statistical method based on Gaussian Process Regression learned in an unsupervised way. The resulting system is able to provide a nearly complete coverage of the 2–12-m-depth zone at a resolution of 80 m. Evaluated on three sites around the Hawaiian Islands, our method obtained estimates with a correlation coefficient in the range of 0.7–0.9. Furthermore, the trained models provide equally good results in nearby zones that lack exploitable waves, extending the scope of applicability of the method.

Keywords

• coastal bathymetry
• machine learning
• unsupervised learning
• linear wave model
• Gaussian Process regression
• multispectral images
• Hawaiian Islands