Simulating compound weather extremes responsible for critical crop failure with stochastic weather generators

Abstract. In 2016, northern France experienced an unprecedented wheat crop loss. This extreme event was likely due to a sequence of particular meteorological conditions, i.e. too few cold days in late autumn-winter and an abnormally high precipitation during the spring season. The cause of this event is not fully understood yet and none of the most used crop forecast models were able to predict the event (Ben-Ari et al., 2018). Here we focus on a compound meteorological hazard (warm winter and wet spring) that could lead to a crop loss. This work is motivated by two main questions: were the 2016 meteorological conditions the most extreme under current climate? and what would be the worst case meteorological scenario that would lead to the worst crop loss? To answer these questions, instead of relying on computationally intensive climate model simulations, we use an analogue-based importance sampling algorithm that was recently introduced into this field of research (Yiou and Jezequel, 2020). This algorithm is a modification of a stochastic weather generator (SWG) that gives more weight to trajectories with more extreme meteorological conditions (here temperature and precipitation). This approach is inspired from importance sampling of complex systems (Ragone et al., 2017). This data-driven technique constructs artificial weather events by combining daily observations in a dynamically realistic manner and in a relatively fast way. This paper explains how a SWG for extreme winter temperature and spring precipitation can be constructed in order to generate large samples of such extremes. We show that, with some adjustments, both types of weather events can be adequately simulated with SWGs, highlighting the wide applicability of the method. We find that the number of cold days in late autumn 2015 was close to the plausible maximum. But our simulations of extreme spring precipitation show that considerably wetter springs than what was observed in 2016 are possible. Although the crop loss of 2016 relation to climate variability is not fully understood yet, these results indicate that similar events with higher 20 impacts could be possible in present-day climate conditions.