Beyond turbidity: embedded modelling of suspended sediment concentration and distributed sensing for morphodynamic observation

Suspended sediment concentration (SSC) plays a central role in sediment transport, river morphodynamics, ecosystem functioning, and the impact of human activities on fluvial systems. Despite its importance, long-term and spatially distributed SSC monitoring remains limited. Most operational monitoring approaches rely on turbidity as a proxy for SSC, which requires frequent site-specific calibration and often performs poorly when sediment properties vary in time or between catchments. This limits the comparability of measurements across river networks and hydrological conditions.

Here, we present a new approach to in situ SSC monitoring that combines improved optical sensing with a distributed sensor network and embedded data-driven modelling. The monitoring network comprises approximately 20–30 autonomous sensors deployed in Swiss rivers across a range of climatic and geomorphic settings, operated in close collaboration with scientific partners. While the embedded model focuses on improving SSC estimation at the sensor level, the network design enables early network-scale observation of suspended sediment dynamics relevant for morphodynamic analyses.

We introduce an improved optical suspended sediment sensor designed for long-term field deployment. Compared to earlier sensor versions, the instrument shows increased signal stability and sensitivity under variable flow and concentration conditions. A key design feature is access to raw optical measurement signals rather than internally processed turbidity outputs, enabling SSC estimation approaches that are not constrained by traditional turbidity-based assumptions and extending the effective measurement range up to 20 g/L.

Building on this capability, we develop a lightweight embedded machine learning model that estimates SSC directly from raw sensor signals. Instead of using turbidity as an intermediate proxy, the model exploits multi-dimensional signal characteristics that better represent catchment sediment properties. The model is trained and evaluated using paired in situ measurements and reference samples collected across multiple deployment sites.

We assess model performance at selected field sites spanning contrasting hydrological regimes and sediment sources. Results show improved agreement with reference SSC measurements compared to conventional turbidity-based estimates, particularly during periods of rapidly changing sediment concentrations. The approach shows reduced sensitivity to short-term signal variability and sensor drift.

While the network is still at an early stage, these results demonstrate how improved SSC estimation at the sensor level, combined with distributed sensing, can support more transferable observations of sediment dynamics across river systems. The presented framework has implications for sediment transport studies, morphodynamic model calibration and validation, and the design of scalable monitoring networks.