Advancing Shallow Water Bathymetry Estimation in Coral Reef Areas via Stacking Ensemble Machine Learning Approach

Abstract

Satellite-derived bathymetry technology plays a pivotal role in estimating shallow water depths. Although traditional machine learning (ML) models are extensively applied in water depth inversion, they frequently exhibit inconsistent performance across various environments, highlighting the challenge of constructing a model with high precision and robustness. This study proposed an innovative stacking ensemble ML (SEML) model, integrating the advantages of various mainstream ML algorithms to address this challenge. We evaluated the bathymetric performance of the SEML model by combining multitemporal Sentinel-2 imagery and sonar data from Houteng Reef and Wufang Reef in the Spratly Islands. The findings showed the performance rankings of these models at Houteng Reef were SEML, K-nearest neighbor (KNN), support vector machine (SVM), multilayer perceptron (MLP), and RF, while at Wufang Reef, they shifted to SEML, SVM, MLP, KNN, and RF. By contrast, the SEML outperformed traditional ML models in accuracy and robustness. At Houteng Reef, the SEML achieved an RMSE of 0.46 m, representing a 13.21% decrease compared to KNN. Similarly, at Wufang Reef, the RMSE of the SEML model was 0.75 m, achieving a 5.06% decrease compared to SVM. The SEML model significantly enhances the accuracy and robustness of water depth estimation, providing a new perspective for accurately mapping coral reef bathymetry.