Unlocking km-scale European temperature fields with crowdsourced observations

High-resolution gridded climate datasets are increasingly needed to analyse local climate variability, assess weather related impacts, and support climate services at scales relevant to society. Yet official surface observation networks alone cannot resolve the fine-scale spatial patterns associated with urban climates, land use transitions, or small-scale extremes. At the same time, the growing availability of crowdsourced and IoT weather observations offers unprecedented spatial density, but their integration into spatial climatologies requires robust statistical methods capable of handling heterogeneous data sources, variable quality, and evolving observational coverage.

To address these challenges, the Royal Netherlands Meteorological Institute (KNMI) has developed an operational workflow for producing high-resolution spatial climatologies by integrating official surface observations with crowdsourced data. The approach builds on multi-fidelity Gaussian Process (GP) regression, in which official observations and personal weather station (PWS) data are combined while explicitly learning their relative error scales and biases. The workflow includes near real-time acquisition of crowdsourced data, basic quality control to remove implausible values, and interpolation using covariates such as land use fractions, terrain descriptors, and distance-to-coast. Scaling GP models to continental domains poses substantial computational challenges, as naïve implementations rely on full covariance matrices with prohibitive computational cost. This extension was made possible through collaboration with the Dutch National Research Institute for Mathematics and Computer Science (CWI), who developed scalable refinements including sparse covariance structures, Nyström-based low-rank approximations, Wendland-induced sparsity, and adaptive landmark selection. These innovations reduce the computational burden sufficiently to enable near real-time 1-km European fields.

Although designed for high-resolution mapping, the posterior fields produced by this system can be aggregated to hourly, daily, or longer periods, enabling the construction of spatially consistent climatological datasets with quantified uncertainty. From these aggregated fields, a wide range of standard climatological indices can be derived, including fixed-threshold metrics such as frost days, heatwave duration, or growing season length, while preserving spatial coherence across heterogeneous landscapes. The increased level of detail also improves the representation of urban climates and local extremes, offering finer-scale baselines for applications such as drought monitoring, energy planning, and hydrological modelling.

Taken together, these developments show that crowdsourced weather observations can play a meaningful role in modern spatial climatology when combined with robust statistical methods and explicit uncertainty quantification. The approach demonstrates that we can now produce high-resolution maps across the European domain, a capability that is essential for representing local extremes, urban climates, and fine-scale land‑use transitions. In doing so, it supports the development of next-generation European climate products with enhanced spatial detail and traceable uncertainty.