Maintaining the channels and infrastructure for large commercial ports is costly and complex. Port authorities traditionally rely on bathymetry surveys to guide their dredging operations, which limits their ability to be proactive. This research aims to develop machine learning (ML) methods to predict sedimentation rates (SR) by analysing patterns in hydrological and meteorological (hydro-meteo) conditions and dredging data. By investigating the possibility of SR prediction models, this research can contribute to efficient maintenance operations.

The study focuses on the Botlek, a harbour in the Port of Rotterdam situated in the Rhine-Meuse estuary. Three data types are integrated to capture the dynamic interplay of saline and riverine factors within the estuary. The data consists of Multibeam bathymetry surveys, hydro-meteo variables (such as salinity, river discharge, and tidal variation), and dredging logs. The surveys provide the net sediment accumulation, which is assumed to result from a specific period of hydro-meteo conditions and dredging. Due to the limited availability of surveys, the number of features had to be managed. To avoid having more features than samples, the hydro-meteo variables are aggregated into daily, weekly, and
monthly means.

The ML algorithms evaluated in this research are Linear Regression (LR), Random Forest Regression (RFR), and Support Vector Regression (SVR). The feature importance scores from the RFR and the accuracy on small datasets of the SVR were decisive in this selection. All algorithms were tested and refined over multiple development phases to determine their predictive accuracy and robustness across different data configurations. It was found that the ML algorithms can reasonably predict SR. In addition to dredging data, incorporating hydro-meteo variables enhanced the predictive accuracy and consistency. Specifically, a feature set of dredging volumes, salinity, discharge, and tidal variation improved model performance. Training the model on dredging data only resulted in an inferior performance. The monthly aggregation of these hydro-meteo data was the most beneficial configuration. Among the tested algorithms, SVR consistently outperformed both LR and RFR, with its best configuration achieving a mean R2 score of 0.69 over four different dataset splits.

The ML models do not have to be able to predict sediment volumes with a small confidence interval to be practical for the port authorities. Instead, the models should be able to predict trends in sediment accumulation and provide actionable insights to enable proactive dredging. The research shows promising results, as most predictions fall within the acceptable range. However, predictive accuracy strongly varied across the different dredging areas in Botlek. Further development is required to provide real-time, reliable SR predictions before the models can be integrated into the maintenance operations of ports.