LAUSR

Abstract Phenological models can effectively infer historically missing phenological data, so as to investigate the long-term relationship between plants and climate change. Large numbers of ecophysiological and statistical models have been developed in the past few decades, but these models have been unable to make accurate predictions based on external data. Machine learning (ML) methods have an advantage over traditional statistical methods for natural science studies. However, only a few phenological models have been coupled with ML methods. In this study, using long-term leaf unfolding date (LUD) observations collected in Harbin, China, we adopted three popular ML algorithms for predicting plant LUD and compared the performances of 10 phenological models. We detected the temperature sensitivity (ST) of the LUD at the species level for the periods 1962–1987 and 1988–2016 (before and after the recent, sudden warming) and temporal changes in ST with a 15-year moving window for each period. The results show that the gradient boosting decision tree (GBDT) model performs obviously better than the other models for external validation data, while avoiding model overfitting. Most species showed an increase in ST during the 1988–2016 period, and the temporal changes in ST significantly decreased during both periods. The temporal changes in ST from the phenological data predicted by the GBDT model is significantly higher than that of other models, which indicates that the traditional phenological models may underestimate the response of LUD to climate warming. We found a prevalent decline in the magnitude of ST with increasing preseason temperature variance at the species level. Our research suggests that machine learning algorithms should be more widely used in future phenological model research, and temporal changes in ST should be investigated in order to broaden our understanding of plants’ ability to adapt to future climate change.