Interpreting the influences of multiple factors on forcing requirements of leaf unfolding date by explainable machine learning algorithms

[Display omitted] •We used three machine learning (ML) algorithms to simulate leaf unfolding date.•Local adaptation and previous leaf senescence modulate spring phenology.•Temperature, water and light-related factors co-determine spring phenology.•ML algorithms could decipher the effects of multiple factors on phenology. Understanding how spring phenology of temperate trees responds to global change is essential for assessing the vegetation dynamics in the future. Besides chilling and forcing temperatures, the influences of other environmental and biotic factors on leaf unfolding date (LUD) are still unclear. Based on long-term records (1960–2015) of LUD for 6 typical temperate tree species at 4242 stations in Europe, we first used a classic process-based model (the Unified model) to describe the relationship between chilling and forcing requirements of LUD. Furthermore, we used 3 explainable machine learning (ML) algorithms (RF, EBM, and GAMI-Net) to quantify the influences of 45 biotic and environmental factors on the LUD. The root-mean-square error (RMSE) of the ML-based models averaged from all species (7.03 to 7.33 days) was lower than the Unified model (8.73 days). The ML algorithms detected 2 biotic (elevation and previous leaf senescence date) and 3 temperature-related variables (chilling accumulation, freezing days in March and annual temperature range) with high importance for most species. The trees adapted to a higher elevation or with later leaf senescence date in the previous year exhibited later LUD. The increase in chilling accumulation, freezing days in March and the annual temperature range could reduce the forcing requirement of LUD. Water-related and light-related variables were only important for one or two species. The increase in relative humidity in February could advance LUD, while soil moisture in March had the opposite effect. The increase in March day length and downward surface shortwave radiation exerted an advancing influence on the LUD of light-sensitive species. Our results suggest that incorporating multiple biotic and environmental factors into ML algorithms not only could effectively improve the accuracy of phenological prediction but also enhance our understanding of how spring phenology responds to each factor.