Extended Data Fig. 1: Procedure for identifying patterns of memory effects in extreme events.

a, Schematic of the data-driven framework. The Entity-Aware Long Short-Term Memory (EA-LSTM) model is trained with meteorological variables over the past 5 years (week t-259 to t) and static variables to predict the GPP of the current week (week t). GPP predictions from multiple trained models are validated against site GPP data and TROPOMI SIF retrievals, and a final model from the model ensemble was selected based on the consistency with SIF. A peak-finding algorithm is applied to the model output to locate global GPP extreme events. The Integrated Gradients (IG) explainer is used to quantify the time-series contributions of meteorological variables during GPP extreme events, thereby identifying distinct patterns of memory effects that indicate divergent vegetation responses to climatic variations during and before these extreme events. b, Illustration of the temporal attribution of extreme events. The y-axis tick labels denote precipitation (Prep), radiation (Rad), temperature (T), vapour pressure deficit (VPD), and wind speed (Ws), respectively, while the x-axis represents the weeks before prediction (Time), covering a five-year range. Each grid represents the contribution of a specific meteorological variable aggregated over a 26-week period during each extreme event. c, Global distribution of the 200 flux tower sites used in this study and their corresponding plant functional types. Basemap data from Natural Earth (https://www.naturalearthdata.com).