"Come as you are": Reconsidering the need for complex quality control when gridding crowd-sourced weather data

Recording hourly-mean in situ observations of surface air temperature is a core task of most National Meteorological and Hydrological Services (NMHSs). Now, suppose that a user is looking for a best estimate of past or present air temperature at a user-specified location. Can crowd-sourced weather data help a NMHS to serve this need? In this work, we provide a quantitative answer to this question by computing the cross-validation root-mean-squared error for 1704 individual time slices of observed air-temperature data in the Netherlands over the year 2023. We consider two baseline services: (i) providing station data as a basis for nearest-neighbour approximation, a service widely used in legal disputes, insurance claim evaluations, etc., and (ii) providing a gridded dataset, a technically more advanced current service. Our results indicate that service (i) is not expected to be improved by including crowd-sourced data—for example, the observed relative success rate is 14%—while service (ii) is expected to be improved by multi-fidelity blending with crowd-sourced data—for example, the observed relative success rate is 91%. Furthermore, our observations indicate that state-of-the-art, well-tuned quality control of the crowd-sourced data (i.e., removing 85% of the data) can—in some of the metrics—lead to improvements of the expected performance of service (i), but—in all metrics—degrades the expected performance of service (ii) significantly. We then continue with a sensitivity study of 96 individual time slices of synthetic air-temperature data to further understand and nuance these conclusions. Zooming out to the larger perspective, when considering the need for quality control, our results indicate that, for service (i), fewer crowd-sourced data is better, while, for service (ii), more crowd-sourced data is better.