Satellite-Derived Bathymetry for Shallow Waters, in Regions of High Optical Complexity, Using Empirical Methods and Convolutional Neural Network

Abstract

Recent advancements in remote sensing instrumentation and technics have positioned satellite-derived bathymetry as one of the most promising methods for monitoring coastal environments. This study compares two approaches for estimating shallow water depths from Sentinel-2 satellite imagery: (1) an empirical method based on linear models correlating water depth with the interaction of different bands of the electromagnetic spectrum within the water column, and (2) a machine learning approach employing convolutional neural networks to capture nonlinear relationships to address the complex influence of suspended particulate matter and colored dissolved organic matter in optically complex waters. The CNN-based model demonstrated significant accuracy in turbid coastal environments, achieving a mean absolute error of approximately 0.1 m in depths up to 30 m. These findings highlight the potential of combining empirical and machine learning methods to enhance bathymetric assessments in regions where traditional surveys are costly and challenging.