Développement d'un outil opérationnel de prévision d'apparition de gel printanier dans un vignoble wallon, à l'aide du modèle MAR

Abstract

With climate change and the northward shift of isotherms, Belgium is becoming an increasingly suitable territory for vineyards. Indeed, Belgium possesses the bioclimatic indices necessary for vine growth. As a result, winegrowing activity is increasing in Belgium. However, Belgium is not immune to spring frosts, which cause a lot of damage to vineyards both in northern France and elsewhere in the world. To avoid losses, winemakers must deploy defense tools such as heating candles, wind towers, helicopters, aspersion, and many other tools. These tools cost time and money and require precise frost forecasts for their implementation. The purpose of this thesis is to develop an operational tool for spring frost prediction. This tool has been created from two databases, obtained from chips that recorded the temperature at the vine fruit level during the 2021-2022 year, as well as from the MAR (regional atmospheric model, forced by ERA5). Two methods were used to attempt frost prediction. These two methods were based on the years 2021-2022. A predictive model was first generated using an empirically modified formula: the MAR output temperature was transformed to be as close as possible to that of the vine fruit, estimated using the chips. Then, a second model was created through multivariate logistic regression, taking into account, in addition to the minimum temperature of the MAR, several variables that could impact the occurrence of frost. This resulted in an improvement in the prediction of frost nights compared to the raw outputs of the MAR. In fact, in 2022, from March onwards, out of 46 nights, the raw MAR (without modification) correctly predicted 7 frost nights, missed 9 frost nights, and correctly predicted 30 nights without frost, whereas the best model in this thesis (derived from logistic regression) predicted on the same 46 nights, 13 correct frost nights, 3 missed frost nights, and correctly predicted 30 nights without frost. The model still needs to be tested on future forecasts with the MAR forced by the global GFS model, for example.