Interpretable machine learning algorithms to predict leaf senescence date of deciduous trees

Abstract

Predicting tree phenology accurately is essential for assessing the impact of climate change on ecosystems. However, the current process-based models are still hard to fully explain and quantify the effects of multiple biotic and environmental factors (e.g., spring phenology, local adaptation, productivity, climatic variables) on autumn phenology. Using leaf senescence data (1980-2015) of 232766 site-year records (3020 sites and 6 deciduous tree species) in Europe, we first calibrated and evaluated 4 process-based models developed by previous studies. Subsequently, by using different machine learning (ML) algorithms (RF, EBM, and GAMI-Net), we developed 3 ML-based models for predicting the leaf senescence date of the same species and quantifying the importance and response function of 63 biotic or environmental variables. We found that the root mean square error (RMSE) of process-based models (averaged from all species) for the test dataset ranged from 11.97 to 12.91 days. The ML-based models outperformed process-based models for all species, with RMSE ranging from 10.01 to 10.58 days. For most species, the recently developed ML algorithms (EBM and GAMI-Net) are more effective than the classical RF algorithm developed in the early 21th century. Besides the temperature and photoperiod in autumn, the geographic factors (especially elevation, longitude, and latitude) were identified as the most important variables in the ML-based models, implying that leaf senescence date is an adaptive trait. Furthermore, for most species investigated, earlier leaf-out dates tended to advance the leaf senescence date. Our results highlight that the ML algorithms not only could effectively improve the performance of the process-based models for predicting the leaf senescence date, but also help to understand the nonlinear and interactive effects of multiple driving factors on autumn phenology.