Flood forecasting based on personal weather station rainfall data

An increasing number of personal weather stations (PWSs) is installed by citizens, resulting in a large amount of real-time available precipitation data. This study assesses the applicability of these data for flood forecasting. We focussed on 30 catchments (total area 2474 km²) located in the management area of Water Board Rijn and IJssel, a water authority in the Netherlands which uses PWS data as input for their operational flood forecasting system. We compared rainfall from a network of 869 Netatmo PWSs (after applying a quality filter) and the real-time radar product from the KNMI (Royal Netherlands Meteorological Institute). Next, we used both products as input for the rainfall-runoff model WALRUS and compared the simulated discharges. These two datasets with almost no latency were validated with the final reanalysis KNMI radar product and discharge observations, for a full year (2023).

For precipitation, the real-time radar was closer to the final reanalysis radar than the PWSs in terms of Kling-Gupta Efficiency, Pearson correlation coefficient and coefficient of variation, but had a stronger negative bias. However, discharge simulations based on PWSs were closer to observations and simulations with the final reanalysis radar than simulations based on the real-time radar. This contrasting result can be explained by the bias, which was stronger for the real-time radar than for the PWSs, and is amplified in the discharge simulations due to the memory in the hydrological system. We found no clear relation between catchment size, PWS density and PWS distribution and the performance of PWS rainfall product. Reducing the density of the PWS network only led to a small deterioration in performance. The results indicate the potential of these devices to be used in hydrological applications, especially when initial hydrological model conditions are improved with data assimilation in operational flood forecasting systems.