Research map / Probabilistic and scientific modeling / Molecules and proteins
Hybrid Phenology Modeling for Predicting Temperature Effects on Tree Dormancy
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Replaces the chilling function of biophysical dormancy models with a learnable MLP trained jointly with the model's parameters, predicting cherry blooming dates in Japan, South Korea and Switzerland.
Problem
Biophysical phenology models for tree dormancy disagree structurally in how temperature contributes to endodormancy release and need site-specific recalibration, while pure ML models lack interpretability and overfit with limited data.
Approach
Make the standard chill/forcing threshold model differentiable with generalized logistic approximations and substitute the chilling function with a small MLP, jointly optimizing MLP and biophysical parameters by backpropagation under a flowering-day negative log-likelihood.
Key results
- Hybrid model achieves lowest MAE in temporal setting: 2.17 days in Japan vs 2.80 for LSTM and 3.07 for Utah Chill, 5.36 in Switzerland, 4.86 in South Korea.
- In the variety-level setting, the hybrid model reaches 3.28 days in Japan (matching LSTM at 3.24), 10.2 in Switzerland and 5.41 in South Korea, outperforming all biophysical models.
- In spatiotemporal generalization to held-out locations, the hybrid model's MAE increases roughly half a day per country, while biophysical models degrade more.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the Utah chilling function with a learned MLP chilling function (ablation) | Hybrid · Japan — temporal | MAE (days): 3.17 \textpm 0.07 -> 2.17 \textpm 0.03 | Helped |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · Japan — temporal | MAE (days): 3.07 \textpm 0.06 -> 2.17 \textpm 0.03 | Helped |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · Switzerland — temporal | MAE (days): 7.11 \textpm 0.09 -> 5.36 \textpm 0.15 | Helped |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · South Korea — temporal | MAE (days): 6.12 \textpm 0.27 -> 4.86 \textpm 0.06 | Helped |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · Japan — temporal (variety) | MAE (days): 3.24 \textpm 0.05 -> 3.28 \textpm 0.04 | Hurt |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · Switzerland — temporal (variety) | MAE (days): 12.21 \textpm 0.31 -> 10.2 \textpm 0.15 | Helped |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · South Korea — temporal (variety) | MAE (days): 6.97 \textpm 0.17 -> 5.41 \textpm 0.08 | Helped |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · Japan — spatiotemporal (variety) | MAE (days): 3.94 \textpm 0.18 -> 3.64 \textpm 0.09 | Helped |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · Switzerland — spatiotemporal (variety) | MAE (days): 18.13 \textpm 0.59 -> 10.6 \textpm 0.37 | Helped |
| Use the hybrid model instead of the strongest baseline | Cherry blooming prediction · South Korea — spatiotemporal (variety) | MAE (days): 7.31 \textpm 0.55 -> 6.07 \textpm 0.38 | Helped |
Surprising findings
- Learned chill response functions differ across random seeds, yet predictive error varies little, suggesting the model is under-constrained.
- The learned chill function shows non-zero contribution above 12.5°C, which no biophysical model includes, i.e. biologically implausible behavior.
Limitations
- Learned chill function is under-constrained and not guaranteed to be biophysically plausible.
- Training data lacks extreme weather events and non-flowering observations.
- Results rely on a single competition dataset with noisy Swiss and limited South Korean data.
Experimental setup
Datasets: George Mason University International Cherry Blossom Prediction Competition dataset MERRA-2 hourly surface temperature
Benchmarks: cherry blossom bloom dates in Japan, Switzerland, South Korea (held-out test years and locations)
Metrics: mean absolute error (days) standard error
Baselines: Chill Hour Utah Chill Chill Days CNN LSTM
Builds on: Chill Hours model Utah chill model Chill Days model Growing Degree Days
Closest papers by meaning
- GreenHyperSpectra: A multi-source hyperspectral dataset for global vegetation trait prediction (NeurIPS 2025 · similarity 0.82)
- DroughtSet: Understanding Drought Through Spatial-Temporal Learning (AAAI 2025 · similarity 0.81)
- Multi-modal Learning for Geospatial Vegetation Forecasting (CVPR 2024 · similarity 0.80)
- Prediction-Augmented Trees for Reliable Statistical Inference (NeurIPS 2026 · similarity 0.80)
- Deep Generative Models for Phylogenetic Inference with Complex Evolutionary Processes (NeurIPS 2026 · similarity 0.80)
- Specialized Foundation Models Struggle to Beat Supervised Baselines (ICLR 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.