A novel data-driven phenology model reveals how and why seasonal timing shifts under climate change

Biosphere atmosphere coupling is an uncertain but important process for the terrestrial carbon sink and therefore for climate projections. The coupling depends on the phenological state, in particular for deciduous plants which only lead to a strong coupling during the growing season. However, plant phenology dynamically itself adapts to the changing climate. Therefore understanding how and why phenology shifts is an important task.

The main challenge in modeling phenology is that it mixes vastly different time scales from daily meteorology that determines the timing of phenological events to the century long lifespan of trees that influence phenology through the different behavior of different species. Mechanistic models (e.g. GDD, GSI) cannot capture this and instead only use the recent past to estimate the phenological state. A promising alternative is data driven models but existing models also use either only the recent past (Liu et al., 2024) or only very long time scales (Reimers et al., 2023) and fail to represent the full complexity.

In this work we propose a novel neural network architecture that in two steps calculates, first, the daily phenological potentials for each day and, second, combines these into a phenological time series using transformer models. We train this model on the PhenoCam V2 and Daymet datasets
for seven different plant functional types across North America.

We demonstrate that our model is a plausible mechanistic representation of plant phenology. It has strong prediction performance on a hold out test set, it emits the same latent relationships as the observations and the sensitivities of the model agree with sensitivities reported in the
literature.

We find that under warming the start of the season moves forward (3.27d ^◦C⁻¹ for deciduous broadleaf forests (DBF)) and the end of the season moves backward (2.30d ^◦C⁻¹  for DBF) but the level of the growing season does not increase. Further we find that current meteorology is less important than long term memory effects through, for example, species. In fact current meteorology is only during the start and end of the season more important than the memory effect.

Additionally we investigate the effect of frost at the beginning of the growing season. We find that such a frost event delays the development of the plant such that it stays delayed until the end of the season. Under warming the vulnerability as well as the timing of the strongest vulnerability changes, but these changes vary between plant functional types. Such data-driven insights in phenological behavior in response to environmental change are key to inform the next generation of Earth system models.

 

_{Liu et al., 2024 Liu, Guohua, et al. “DeepPhenoMem V1.0: deep learning modelling of canopy greenness dynamics accounting for multi-variate meteorological memory effects on vegetation phenology.”}

_{GeoscientificModel Development 17.17 (2024): 6683-6701.}

_{Reimers et al., 2024 Reimers, Christian, et al. “Comparing Data-Driven and Mechanistic Models for Predicting Phenology in Deciduous Broadleaf Forests.” arXiv preprint arXiv:2401.03960 (2024).}