Modeling of Spring Phenology of Boreal Forest by Coupling Machine Learning and Diurnal Temperature Indicators

The roles of diurnal temperature in providing heat accumulation and chilling requirements for vegetation spring phenology differ. Although previous studies have established a stronger correlation between leaf onset and diurnal temperature than between leaf onset and average temperature, current research on modeling spring phenology based on diurnal temperature indicators remains limited. In this study, we confirmed the start of the growing season (SOS) sensitivity to diurnal temperature and average temperature in boreal forest. The estimation of SOS was carried out by employing K-Nearest Neighbor Regression (KNR-TDN) model, Random Forest Regression (RFR-TDN) model, eXtreme Gradient Boosting (XGB-TDN) model and Light Gradient Boosting Machine model (LightGBM-TDN) driven by diurnal temperature indicators during 1982–2015, and the SOS was projected from 2015 to 2100 based on the Coupled Model Intercomparison Project Phase 6 (CMIP6) climate scenario datasets. The sensitivity of boreal forest SOS to daytime temperature is greater than that to average temperature and nighttime temperature. The LightGBM-TDN model perform best across all vegetation types, exhibiting the lowest RMSE and bias compared to the KNR-TDN model, RFR-TDN model and XGB-TDN model. By incorporating diurnal temperature indicators instead of relying only on average temperature indicators to simulate spring phenology, an improvement in the accuracy of the model is achieved. Furthermore, the preseason accumulated daytime temperature, daytime temperature and snow cover end date emerged as significant drivers of the SOS simulation in the study area. The simulation results based on LightGBM-TDN model exhibit a trend of advancing SOS followed by stabilization under future climate scenarios. This study underscores the potential of diurnal temperature indicators as a viable alternative to average temperature indicators in driving spring phenology models, offering a promising new method for simulating spring phenology.