Comparison of Physical and Deep Learning Weather Forecast Models for Agricultural Decision-Making in Belgium

Abstract

Weather forecasts are crucial in agricultural decision-making. The first objective was to assess the performance of eight weather forecast models for predicting five key meteorological variables in agriculture: air temperature, humidity, wind speed, global radiation, and precipitation. The tested models were ICON-D2, AROME, HARMONIE, MAR, GFS, and three models from the ECMWF: the deterministic model (HRES), the ensemble model (ENS), and the deep learning-based model (AIFS). The forecasts were compared over a six-month period with observations from 22 weather stations located in Wallonia (Belgium). AIFS achieved the lowest RMSE for predicting global radiation. For lead times up to 48 h, ICON-D2 had the lowest RMSE for wind speed; ICON-D2 and AIFS showed the lowest RMSE for air temperature and humidity, and AIFS and HRES performed best at predicting rainy versus non-rainy days. For lead times beyond 48 h, AIFS had the lowest RMSE for predicting air temperature and humidity, and ENS had the lowest RMSE for predicting wind speed. The second objective was to evaluate the suitability of these forecasts to feed a simple agricultural decision-support tool, helping farmers identify the optimal time windows for spraying plant protection products. ICON-D2 and AIFS yielded the most reliable recommendations.