Simulating farmers’ weather station data for an extended evaluation of agricultural decision support

While weather data have always been important for agriculture, they have become even more crucial with the emergence of precision agriculture and data-driven crop production. Applications such as pest management, irrigation scheduling and yield prediction rely on weather data.

As farmers’ weather stations (FWS) are now more accessible and affordable, they are increasingly used to monitor local environmental conditions in real time and support agricultural decision-making (Rosillon et al., 2024).  However, FWS are prone to measurement errors, and the impact of these errors on agrometeorological model outputs remains unclear. Few studies exist, and their scope is limited by the time required for field trial data collection (Trilles et al., 2020; Mokhtarzadeh et al., 2025). Extended performance evaluations under diverse climatic conditions are nevertheless essential to test these stations (Kalaany et al., 2025).

This study aims to develop a FWS data simulator based on a three-year field experiment to simulate plausible FWS measurements beyond the observation period. These simulations are then used to model potato late blight (PLB) risk using a weather-based PLB model, and to assess the suitability of FWS across a wide range of climatic conditions.

Six FWS are installed alongside a reference weather station over three years to quantify sensor measurement errors. Modelling FWS data consists of modelling these errors and adding them to the reference weather data (Equations 1 and 2).

The age of the station (to account for sensor drift) and weather conditions are used as predictors. Three algorithms are tested: (1) a naïve model, based on random sampling of measurement errors from the marginal empirical distribution; (2) a multilinear regression combined with a first-order autoregressive model accounting for the temporal autocorrelation of model residuals; (3) a random forest combined with a random sampling of model residuals.

Algorithms are evaluated using cross-validation to assess their ability to appropriately simulate FWS data and FWS PLB risk for unsampled years. FWS data and corresponding PLB risk are then simulated over a historical period (e.g. 2006–2025), enabling extended performance evaluation of FWS in PLB modelling.

This analysis will first quantify the relative contribution of four key components (sensor drift, weather conditions, temporal autocorrelation, and random variability) to measurement errors. Then, it will evaluate algorithms performance in simulating FWS datasets. Finally, it will assess the ability of the FWS data simulator to accurately reproduce PLB risk.

Preliminary results show that random forest outperforms multilinear regression for simulating FWS data, with significant impact on PLB risk modelling. Detailed methods and results will be presented at the congress.