Hybrid Machine Learning Model for Predicting Cherry Tree Phenology

This repository contains the code used for combining biophysical process-based model of cherry tree phenology with machine learning methods for inter-comparing their dynamics.

The code can be used to reproduce the results of the manuscript titled "Hybrid Phenology Modeling for Predicting Temperature Effects on Tree Dormancy".

The image above shows plots of different temperature response curves learned in Japan under different random seeds controlling weight initialization and data split. 3x3 plots are shown that were generated using seeds 1,2,3 respectively (from left to right) for weight initialization and 1,2,3 respectively (from top to bottom) for the data split.

Prerequisites

• Python 3.11
• Used packages and versions are specified in requirements.txt.
• Optional: CUDA (version 11.6 was used)

Data

A cherry Tree Phenology data was compiled from various sources by George Mason’s Department of Statistics and can be found in this GitHub repository. The repository data is updated every year. To ensure reproducibility of the work we included a snapshot of the dataset that we used in this repository. Data sources are listed in the corresponding data/original/README.md.

Publicly available MERRA-2 temperature data was used in this research. The code used for collecting the data can be found here. The resulting file should be stored in the data/merra_v2 folder. Alternatively, the file containing all temperature data can be obtained here. The file should be extracted and stored in data/merra_v2/daily_temperatures.csv.

Usage

Running main.sh reproduces all results presented in the manuscript. Note that fitting all models for all seeds can take a long time (in the order of weeks).

The runs folder contains all configurable Python scripts used to run experiments.

The scripts folder contains all configured scripts that generated the results presented in the manuscript. The seeds.txt file lists all seeds used to generate the results.

Reproducing Paper Results

The main.sh script can be decomposed in the following sections. All results/plots are generated in the results_figures folder.

Train and evaluate models in setting 1 (As described in the paper):

Model	Script
Chill Hours	train_pb_hour.sh
Utah Chill	train_pb_utah.sh
Chill Days	train_pb_days.sh
Hybrid	train_df_mlp.sh
LSTM	train_lstm_baseline.sh

Train and evaluate models in setting 2 & 3 (As described in the paper):

Model	Script
Chill Hours	train_pb_hour_species.sh
Utah Chill	train_pb_utah_species.sh
Chill Days	train_pb_days_species.sh
Hybrid	train_df_mlp_species.sh (setting 2) train_df_mlp_species_holdout.sh (setting 3)
LSTM	train_lstm_baseline_species.sh

Ablation study using the Utah model in the differentiable model approximation:

train_df_utah.sh

Visualizations of the learned temperature response curves:

plot_learned_temperature_function.sh

Examples

To fit a biophysical model (with location-specific parameters) using grid search run:

python3 -m runs.fit_eval --model_cls LocalUtahChillModel --include_temperature --model_name LocalUtahChillModel_seed_123 --seed 123 --locations selection

To fit the hybrid model in Japan run the following:

python3 -m runs.fit_eval --model_cls NNChillModel --include_temperature --model_name NNChillModel_japan_seed_123 --seed 123 --num_epochs 20000 --optimizer adam --lr 1e-3 --weight_decay 1e-4 --scheduler_step_size 2000 --scheduler_decay 0.9 --loss_f nll --locations japan_yedoensis_sargentii