Changes in rainfall amounts and patterns have been observed and are expected to continue in the near future with potentially significant ecological and societal consequences. Modelling vegetation responses to changes in rainfall is thus crucial to project water and carbon cycles in the future. In this study, we present the results of a new model-data intercomparison project, where we tested the ability of 10 terrestrial biosphere models to reproduce the observed sensitivity of ecosystem productivity to rainfall changes at 10 sites across the globe, in nine of which, rainfall exclusion and/or irrigation experiments had been performed. The key results are as follows: (a) Intermodel variation is generally large and model agreement varies with timescales. In severely water-limited sites, models only agree on the interannual variability of evapotranspiration and to a smaller extent on gross primary productivity. In more mesic sites, model agreement for both water and carbon fluxes is typically higher on fine (daily-monthly) timescales and reduces on longer (seasonal-annual) scales. (b) Models on average overestimate the relationship between ecosystem productivity and mean rainfall amounts across sites (in space) and have a low capacity in reproducing the temporal (interannual) sensitivity of vegetation productivity to annual rainfall at a given site, even though observation uncertainty is comparable to inter-model variability. (c) Most models reproduced the sign of the observed patterns in productivity changes in rainfall manipulation experiments but had a low capacity in reproducing the observed magnitude of productivity changes. Models better reproduced the observed productivity responses due to rainfall exclusion than addition. (d) All models attribute ecosystem productivity changes to the intensity of vegetation stress and peak leaf area, whereas the impact of the change in growing season length is negligible. The relative contribution of the peak leaf area and vegetation stress intensity was highly variable among models. K E Y W O R D S drought, irrigation, rainfall manipulation experiment, terrestrial biosphere models

Abstract. Compound weather events may lead to extreme impacts that can affect many aspects of society including agriculture. Identifying the underlying mechanisms that cause extreme impacts, such as crop failure, is of crucial importance to improve their understanding and forecasting. In this study, we investigate whether key meteorological drivers of extreme impacts can be identified using the least absolute shrinkage and selection operator (LASSO) in a model environment, a method that allows for automated variable selection and is able to handle collinearity between variables. As an example of an extreme impact, we investigate crop failure using annual wheat yield as simulated by the Agricultural Production Systems sIMulator (APSIM) crop model driven by 1600 years of daily weather data from a global climate model (EC-Earth) under present-day conditions for the Northern Hemisphere. We then apply LASSO logistic regression to determine which weather conditions during the growing season lead to crop failure. We obtain good model performance in central Europe and the eastern half of the United States, while crop failure years in regions in Asia and the western half of the United States are less accurately predicted. Model performance correlates strongly with annual mean and variability of crop yields; that is, model performance is highest in regions with relatively large annual crop yield mean and variability. Overall, for nearly all grid points, the inclusion of temperature, precipitation and vapour pressure deficit is key to predict crop failure. In addition, meteorological predictors during all seasons are required for a good prediction. These results illustrate the omnipresence of compounding effects of both meteorological drivers and different periods of the growing season for creating crop failure events. Especially vapour pressure deficit and climate extreme indicators such as diurnal temperature range and the number of frost days are selected by the statistical model as relevant predictors for crop failure at most grid points, underlining their overarching relevance. We conclude that the LASSO regression model is a useful tool to automatically detect compound drivers of extreme impacts and could be applied to other weather impacts such as wildfires or floods. As the detected relationships are of purely correlative nature, more detailed analyses are required to establish the causal structure between drivers and impacts.

Abstract. Compound weather events may lead to extreme impacts that can affect many aspects of society including agriculture. Identifying the underlying mechanisms that cause extreme impacts, such as crop failure, is of crucial importance to improve understanding and forecasting. In this study we investigate whether key meteorological drivers of extreme impacts can be identified using Least Absolute Shrinkage and Selection Operator (Lasso) in a model environment, a method that allows for automated variable selection and is able to handle collinearity between variables. As an example of an extreme impact, we investigate crop failure using annual wheat yield as simulated by the APSIM crop model driven by 1600 years of daily weather data from a global climate model (EC-Earth) under present-day conditions for the Northern Hemisphere. We then apply the logistic Lasso regression to predict which weather conditions during the growing season lead to crop failure. We obtain good model performance in Central Europe and the eastern half of the United States, while crop failure years in regions in Asia and the western half of the United States are less accurately predicted. Model performance correlates strongly with annual mean and variability of crop yields, that is, model performance is highest in regions with relatively large annual crop yield mean and variability. Overall, for nearly all grid points the inclusion of temperature, precipitation and vapour pressure deficit is key to predict crop failure. In addition, meteorological predictors during all seasons are required for a good prediction. These results illustrate the omnipresence of compounding effects both between meteorological drivers and different periods of the growing season for creating crop failure events. Especially vapour pressure deficit and climate extreme indicators such as diurnal temperature range and the number of frost days are selected by the statistical model as relevant predictors for crop failure at most grid points, underlining their overarching relevance. We conclude that the Lasso regression model is a useful tool to automatically detect compound drivers of extreme impacts, and could be applied to other weather impacts such as wildfires or floods. As the detected relationships are of purely correlative nature, more detailed analyses are required to establish the causal structure between drivers and impacts.