Display omitted •Models based on kriging with external drift on elevation were the optimal choices.•Additional RMI stations improved the accuracy but not the robustness of the models.•Meteorological reanalysis used as predictor did not improve spatial interpolation.•Spatial interpolation is a promising option to improve current agricultural DSSs.•Agromet.be is an application for spatial interpolation of weather data in Wallonia. Food production will have to increase in the future to face the growing world population. Agricultural decision support systems (DSSs) are part of the solution since they aim at protecting crops against fungal diseases, a significant contributor to yield losses, while minimising pesticide use. DSSs are mainly driven by weather data which, currently, are usually obtained from the nearest available weather station. Since the latter is sometimes located far away from a farmer’s field, this can lead to inaccurate recommendations. In order to provide better local weather data, spatial interpolation is a solution. However, since it must be delivered in near real-time, integrating a spatial interpolation process into an operational application necessitates addressing four constraints: Accuracy, Robustness, Reliability and Latency. This study aimed at developing an operational application for a near real-time spatial interpolation of air temperature and relative humidity at hourly and daily timescales. The first objective was to select the best spatial interpolation models among five algorithms: nearest neighbour, inverse distance weighting, multiple linear regression, ordinary kriging and kriging with external drift. The best models were based on kriging with elevation as external drift. They largely reduced the mean absolute error (MAE) compared to using the nearest station: for hourly air temperature MAE dropped from 0.93 °C to 0.59 °C. It performed also better than multiple linear regression (MAE = 0.68 °C). The second objective was to evaluate the impact of increasing station density by adding stations from the Belgian synoptic network. Additional stations improved Accuracy (MAE = 0.57 °C) but to a lesser extent than expected and had no clear impact on Robustness. The third objective was to assess the interest of using reanalyses (i.e. climate model outputs) as dynamic predictor variables. Reanalyses did not improve Accuracy (MAE = 0.62 °C) because, compared to elevation, they did not provide useful additional information that can be leveraged by the interpolation models. Using such a dynamical input also impacted Reliability negatively due to potential availability issues. Kriging models presented the highest computing times. However, Latency caused by the interpolation process itself was very small compared to the entirety of data processing. The selected models were implemented on an online application “Agromet.be”. Near real-time dissemination of interpolated weather data enables to produce local warnings helping farmers to take better decisions about spraying schedules. As a future improvement of spatial interpolation, integrating numerous personal weather stations owned by farmers seems promising.