Comparisons of numerical phenology models and machine learning methods on predicting the spring onset of natural vegetation across the Northern Hemisphere

The timing of vegetation spring onset is largely influenced by climate factors, making it sensitive to climate variation. Robust models that predict vegetation spring onset via the climate forcing data are needed in the land surface models for understanding the impacts of climate change on vegetation processes. In this study, we apply and assess both numerical phenology models and the machine learning models on predicting the timing of spring onset for different vegetation types, including deciduous vegetation, evergreen vegetation and stressed deciduous vegetation. We perform model calibration for numerical phenology models and machine learning models using both in-situ observations of spring onset dates from National Phenology Network in the United States and satellite-derived green-up dates in the Northern Hemisphere. Our experiment showed better performance of numerical models calibrated by ground phenology observations. Among all the numerical phenology models, the models developed based on Growing Season Index perform well on predicting the spring onsets of deciduous vegetation and stressed deciduous vegetation across the Northern Hemisphere. Machine learning models if trained appropriately could also capture the spatial variation of satellite-derived spring onset dates. Our study highlights the need of improvements on numerical phenology models for their uses in the land surface models. We also illustrate the benchmarking role of the machine learning models on predicting vegetation spring onsets via climate variables.

Automated summary

This study successfully predicted the timing of spring onset for different vegetation types using numerical phenology models and machine learning models with a combination of ground observations and satellite data. Numerical phenology models showed better performance after calibration with ground phenology observations, while machine learning models could also capture spatial variation of satellite data when appropriately trained.