A Refined Machine Learning Method for Coastal Bathymetry Retrieval Using Minimum Distance From Coastline and Geographical Features

Bathymetry is a critical marine parameter for applications, such as coral reef monitoring, erosion forecasting, coastal defense, and other related efforts. The predominant method for obtaining water depth data has been through satellite remote sensing. However, the utilization of multispectral processing in deeper and more complex benthic substrate environments poses significant challenges due to reduced light penetration at greater depths and the influence of benthic substrate on light reflection and absorption. In response to these challenges, a novel inversion method based on machine learning called the minimum distance from the coastline and geographical position-based machine learning bathymetry inversion model (MDCGP-ML) has been introduced in this research. The MDCGP-ML model leverages the spectral data from multispectral remote sensing images, alongside MDCGP features, with 70% of bathymetric data from the South China Sea region used for model training. The machine learning algorithms employed include random forest, support vector regression, backpropagation (BP) neural network, extreme gradient boosting, categorical boosting, and light gradient boosting machine (LightGBM). The 30% of in situ was utilized for validating model accuracy. Experimental results using WorldView-2 imagery indicate that the MDCGP-ML model delivers strong performance [root mean square error: 0.849–1.26]. Furthermore, compared to traditional log-ratio inversion models and the spectral information-based machine learning bathymetry inversion model, the MDCGP-ML approach exhibits distinct advantages. It effectively reduces the overreliance on spectral data in bathymetric inversion. This advancement offers a novel perspective for augmenting the bathymetry inversion capabilities of satellite remote sensing in complex nearshore aquatic regions.