Stormwater runoff is a major vector for the pollution transport. Monitoring its quality is necessary for informing effective management strategies. This thesis focuses on an analytical tool increasingly utilized by both field practitioners and researchers: sensor technology. The workflows surrounding on-site sensor deployment, data handling, and interpretation involve multiple often-overlooked nuances and decisions that are not yet common practice, motivating this systematic effort.

The work began with a critical literature review that situated sensor technology within the context of urban stormwater monitoring. The review showed that sensors are frequently treated as turn-key solutions, revealing a mismatch between their perceived maturity and the actual level of methodological development in the field. It further demonstrated that the limited set of water quality parameters that can be measured directly and continuously in-situ has limited standalone value, and that the primary analytical value of sensor data emerges when it is coupled with modelling approaches, which are commonly used to derive pollutant concentration time series.

Drawing on both the reviewed literature and a multi-year field monitoring campaign incorporating continuous sensor measurements alongside sample-based laboratory analyses, this work systematically investigated the problems and limitations of in-situ water quality sensors. Two principal types of adverse effects were distinguished: loss of data and the introduction of bias and uncertainty. The results show that several of the most frequently encountered problems are amenable to post-validation correction. A comparative evaluation of simple interpolation methods and machine-learning–based reconstruction techniques indicates that interpolation is generally sufficient under moderately dynamic conditions, while machine-learning approaches offer only limited advantages for highly dynamic segments. Comparison of field sensor measurements with laboratory reference analyses revealed parameter-specific responses, with strong agreement observed for electrical conductivity, minor field-induced effects for pH, and substantial, condition-dependent bias for turbidity related to seasonal processes. The literature review indicates that uncertainties associated with analytical context are seldom systematically investigated or quantitatively reported.

Finally, this work quantified how adverse effects propagated into pollutant concentration modelling by analysing the influence of data completeness and field-induced uncertainty. Both conceptual and regression-based models were evaluated, including simple statistical and machine-learning regression models. Model performance was strongly influenced by dataset completeness and diversity, with predictive accuracy deteriorating proportionally to the magnitude of uncertainty in the data. Conceptual buildup-washoff and washoff-only models showed poor performance, whereas higher regression model performance depended primarily on the choice and combination of explanatory variables.