Integrating machine learning and semiempirical methods for bathymetric estimation and sediment classification using Sentinel-2 imagery and multibeam backscatter: A case study in the Natuna Sea

Accurate and up-to-date bathymetric and seabed sediment data are vital for coastal zone management, marine navigation, ecosystem monitoring, and disaster risk mitigation. This study presents a hybrid framework that integrates machine learning (ML) and semiempirical methods for estimating shallow-water bathymetry and classifying sediments in the Natuna Sea, Indonesia. Sentinel-2 multispectral imagery, combined with in situ depth measurements and multibeam echosounder (MBES) backscatter data, was utilized to develop and validate the models. Sediment classification was performed via the support vector machine (SVM) algorithm and benchmarked against traditional techniques, such as angular response analysis (ARA) and the Seabed Analysis Tools (SAT). The SVM demonstrated superior accuracy when sediment type references from Palu Bay were used. For bathymetric estimation, two semiempirical models (Lyzenga and Stumpf) and three regression-based ML models—Random Forest (RF), Decision Tree (DT), and Gradient Boosted Trees (GBT)—were assessed. The spectral reflectance and band ratio combinations, particularly for bands 2, 3, 4, and 8, outperformed the conventional two-band approaches (bands 3 and 4). The Lyzenga method yielded better results than the Stumpf method did, with an RMSE of 1.098 m, an MAE of 0.617 m, an R 2 of 0.945, and an MPE of 7.528%. Among the ML models, RF achieved the highest accuracy (RMSE = 1.082 m; MAE = 0.584 m; R 2 = 0.947; MPE = 6.804%) for depths up to 20 m. For bathymetric monitoring in coastal areas, based on the Natuna Sea case study and optically clear conditions, RF shows superior performance for depths of 0–10 m, whereas Lyzenga–GBT integration performs better for depths of 10–20 m. Fine sand was identified as the dominant sediment type in the Natuna Sea. This study underscores the potential of integrating satellite remote sensing and data-driven approaches to provide scalable, cost-effective solutions for bathymetric mapping and sediment classification in data-limited maritime regions, supporting marine spatial planning and coastal resource management in Indonesia and similar archipelagic settings. • A hybrid framework integrating machine learning and semi-empirical models was developed for shallow-water bathymetry and seabed sediment classification in the Natuna Sea. • Random Forest using four Sentinel-2 bands (B2, B3, B4, B8) achieved the highest bathymetric accuracy ( RMSE = 0.683 m, R 2 = 0.979 ) for depths up to 20 m. • The Lyzenga semi-empirical model outperformed Stumpf across all metrics ( RMSE = 0.756 m ) and improved further when integrated with Gradient Boosted Trees (GBT). • Support Vector Machine (SVM) provided superior sediment classification accuracy compared to ARA and SAT methods, identifying fine sand as the dominant sediment type. • The integrated remote-sensing–acoustic approach demonstrates a scalable, cost-effective solution for bathymetric mapping and sediment analysis in data-limited archipelagic regions.