Satellite-derived bathymetry from WorldView-2 based on linear and machine learning regression in the optically complex shallow water of the coral reef ecosystem of Kemujan island

Abstract

Shallow coastal water bathymetry information is required for various navigation, management, conservation, and research purposes. Bathymetry data modeled through satellite remote sensing (SDB) is considered more effective and efficient because it is cheaper, can cover a large area, and can be done routinely. In this study, we assess the linear (LR) and machine learning regression performance of bathymetry mapping in the optically shallow water of a coral reef ecosystem. We selected random forest regression (RFR) and support vector machine regression (SVR) machine learning regression models. All three regression models were trained using field bathymetry data, and different tuning parameter scenarios were applied to each model to obtain the best bathymetry map. Independent field bathymetry data were used for the assessment, which included root mean squared error (RMSE), a plot between reference and predicted depths, and an underwater topographic profile in three unique areas. Each unique area represented different water depths, underwater topography, benthic cover variations, and water clarity. This highlights and provides a comprehensive insight into how each model performs under different underwater topographic conditions. Kemujan Island, which has various coral reef geomorphic classes and underwater topographic variations, was selected as the study site. The RMSE of the SDB mapping accuracy obtained in this research ranged from 0.82 to 0.87 m for RFR, 0.96–1.17 m for SVR, and 1.00–1.23 m for LR. Overall, RFR produced SDB with higher accuracy and precision compared to SVR and LR. RFR managed to estimate the depth that follows the underwater transverse profile accurately and precisely in three sites and does not require a band ratio to produce accurate SDB in areas with various benthic covers. LR competes well with RFR in predicting water depths of up to 6 m but cannot accurately predict depths beyond 6 m. The study results provide a comprehensive reference for developing robust automated mapping models for producing SDB, increasing the availability of optically shallow water bathymetry information worldwide while reducing the cost of conducting bathymetric surveys.