From acoustic Doppler backscatter to near-real-time suspended sediment monitoring in Finnish boreal rivers using a Bayesian-optimized data-driven model

In Finland, the monitoring of river water quality is crucial for measuring nutrient loads and assessing ecosystem condition; however, the dynamics of suspended sediment are poorly characterized due to a lack of temporal information from routine sampling. This prevents the further development of digital water systems that demand continuous or near-real-time indications of sediment transport. This study proposes a framework that transforms acoustic Doppler current profiler (ADCP) information into quantitative, data-driven knowledge on suspended sediment concentration across multiple survey campaigns in some Finnish boreal rivers. It leverages a Bayesian-optimized support vector regression to predict suspended sediment concentration by integrating multiple combinations of hydrological parameters, acoustic backscatter features, and a robust backscatter magnitude (RBM). The method is applied to ADCP surveys from four Finnish boreal rivers, and model performance is assessed using leave-one-campaign-out cross-validation to evaluate out-of-survey predictive performance under new hydrological and sedimentary conditions. Moreover, probabilistic modeling is utilized in order to evaluate the level of uncertainty associated with all predictions. Despite the results of log-transformed suspended sediment concentration demonstrating consistent and comparable performance across all the rivers, substantial differences in the accuracy of model predictions and uncertainty were identified within various study sites. Prediction intervals exhibited strong calibration, characterized by narrow uncertainty in more stable systems and much greater uncertainty in rivers influenced by episodic, event-driven sediment movement. Feature-scenario studies indicated that the main factors influencing suspended sediment concentration estimates varied among rivers; in some systems, hydrological variables mostly explained suspended sediment concentration variability, but in others, acoustic structure provided the key predictive insights. Furthermore, in sediment-rich rivers, the incorporation of RBM was essential for more accurate suspended sediment concentration predictions. For the purpose of providing a scalable foundation for near-real-time digital monitoring of suspended sediment and nutrient transport in river networks, the examined framework proposed valuable links between limited acoustic surveys with continuous hydrological data and provides physically meaningful uncertainty estimates.