Hybrid Phenology Modeling for Predicting Temperature Effects on Tree Dormancy

TL;DR

This work tackles structural bias in biophysical tree dormancy models by hybridizing them with a neural module. It replaces the chill function $c(oldsymbol{x}_t)$ with a learnable $ ilde{c}(oldsymbol{x}_t; heta)$ while preserving the cumulative structure $C_t= extstyleigl(orall au ext{ to }tigr) c(oldsymbol{x}_ au)$ and $F_t= extstyleigl(orall au ext{ to }tigr) f(oldsymbol{x}_ au) r^{(c)}_ au$, and optimizes both model parts jointly. Evaluated on cherry blooming dates across Japan, Switzerland, and South Korea with $S=274$ and MERRA-2 temperatures, the hybrid model yields lower mean absolute error than standard biophysical models and the LSTM, and demonstrates transfer to unseen sites without per-site recalibration. The learned chill function provides interpretable insights into temperature effects on dormancy while preserving biophysical constraints, offering a practical and data-efficient route to improved climate-phenology projections in diverse environments.

Abstract

Biophysical models offer valuable insights into climate-phenology relationships in both natural and agricultural settings. However, there are substantial structural discrepancies across models which require site-specific recalibration, often yielding inconsistent predictions under similar climate scenarios. Machine learning methods offer data-driven solutions, but often lack interpretability and alignment with existing knowledge. We present a phenology model describing dormancy in fruit trees, integrating conventional biophysical models with a neural network to address their structural disparities. We evaluate our hybrid model in an extensive case study predicting cherry tree phenology in Japan, South Korea and Switzerland. Our approach consistently outperforms both traditional biophysical and machine learning models in predicting blooming dates across years. Additionally, the neural network's adaptability facilitates parameter learning for specific tree varieties, enabling robust generalization to new sites without site-specific recalibration. This hybrid model leverages both biophysical constraints and data-driven flexibility, offering a promising avenue for accurate and interpretable phenology modeling.

Figures (4)

• Figure 1: Schematic overview of two approaches modeling the phenological stages between dormancy initiation and flowering. Biophysical model candidates (Top) show structural discrepancies in how temperature contributes towards endodormancy progression. In this work (Bottom), we replace this module with an MLP and jointly optimize it with the biophysical model parameters, obtaining a new temperature response function. We then evaluate its ability to generalize to unobserved seasons and reflect on its biophysical plausibility.
• Figure 2: Spatial distribution of locations that were available in the dataset for each country (not to scale), colored by (partially assumed) tree variety occurrence. Varieties include: Prunus ×Yedoensis (Red circles), PrunusSargentii (Blue crosses), PrunusCampanulata Maxim (Green stars), PrunusNipponica Matsum (Purple squares), PrunusJamasakura (Orange diamonds), PrunusAvium (Brown triangles).
• Figure 3: Scatter plots of blooming date predictions for all locations observing the Prunus ×Yedoensis tree variety in Japan. Plots are shown for the Utah model (Top) and the model with learned chilling function (Bottom) in the three evaluation settings. It can be seen that, once consistent tree variety parameters are enforced in all locations the mechanistic model underfits, whereas the temperature response function learned by the hybrid model transfers to many locations.
• Figure 4: (Top) Visualization of the chill response function learned by the MLP in Japan for the first three random seeds used for evaluation, where the seed controls MLP weight initialization as well as the train/test split. Plots showing the effect of each source of randomness are included in the attached repository. The figure shows the intensity of the response plotted against the mean of the function input (i.e. mean temperature in ° C of the 24 hourly measurements of the respective day) for all days occurring in the test dataset. Darker colors highlight the more frequently observed output at each temperature level by showing the density of each column (temperature level discretized to intervals of 0.5 ° C). (Bottom) Same visualization when made using the three biophysical models used as baselines. The biophysical models show striped artifacts due to their weighting of discrete temperature intervals.