Deep learning meets tree phenology modelling: PhenoFormer versus process‐based models

Abstract

Predicting phenology, that is the timing of seasonal events of plant life such as leaf emergence and colouration in relation to climate fluctuations, is essential for anticipating future changes in carbon sequestration and tree vitality in temperate forest ecosystems. Existing approaches typically rely on either hypothesis‐driven process models or data‐driven statistical methods. Several studies have shown that process models outperform statistical methods when predicting under climatic conditions that differ from those of the training data, such as for climate change scenarios. However, in terms of statistical methods, deep learning approaches remain underexplored for species‐level phenology modelling.

We present a deep neural architecture, PhenoFormer, for species‐level phenology prediction using meteorological time series. Our experiments utilise a country‐scale data set comprising 70 years of climate data and approximately 70,000 phenological observations of nine woody plant species, focussing on leaf emergence and colouration in Switzerland. We extensively compare PhenoFormer to 18 different process‐based models and traditional machine learning methods, including Random Forest (RF) and Gradient Boosted Machine (GBM).

Our results demonstrate that PhenoFormer outperforms traditional statistical methods in phenology prediction while achieving significant improvements or comparable performance to the best process‐based models. When predicting under climatic conditions similar to the training data, our model improved over the best process‐based models by 6% normalised root‐mean‐squared error (nRMSE) for spring phenology and 7% nRMSE for autumn phenology. Under conditions involving substantial climatic shifts between training and testing (+1.21°C), PhenoFormer reduced the nRMSE by an average of 8% across species compared to RF and GBM, and performed on par with the best process models.

These findings highlight the potential of deep learning for phenology modelling and call for further research in this direction, particularly for future climate projections. Meanwhile, the advancements achieved by PhenoFormer can provide valuable insights for anticipating species‐specific phenological responses to climate change.