Closing the gaps: applicability of missing data handling techniques for urban runoff quality time series

ABSTRACT Comic-style panels show sensor failure causing data loss, selection of repair approaches, and context-driven choice leading to data recovery. Thus, graphical abstract illustrates that different tools suit different time series gaps, and context-aware intervention enables more accurate data recovery. Adoption of sensor-based on-site monitoring in urban-water engineering enables the collection of high-resolution data, but malfunctions and environmental disturbances can pose a serious challenge in uptake and use, as they result in data gaps and loss of information. One solution to handle these data gaps is the use of machine-learning (ML) algorithms. This study compared widely used deterministic interpolation methods and data-driven methods – particularly neural networks – for reconstructing extended data gaps in urban runoff quality time series. The analysis is based on runoff water quality data from a small and relatively homogeneous section of an urban road. While simpler interpolation methods achieved high overall accuracy, ML models outperformed the simpler methods during highly dynamic periods – those that are typically more challenging to capture and often of particular interest in stormwater management contexts. The findings highlight both the limitations and the potential of each approach, emphasizing the particular importance of understanding context and data conditions when selecting the most suitable method for the task.