Ukraine Crop Yield Estimation with Machine Learning

This repository contains code, notebooks, processed tabular data, and curated experiment outputs for estimating regional wheat yield in Ukraine using MODIS-derived vegetation variables, ERA5-Land weather variables, and machine learning regressors.

The workflow adapts the county-level crop yield prediction methodology from:

Ju, S., Lim, H., Ma, J. W., Kim, S., Lee, K., Zhao, S., & Heo, J. (2021). Optimal county-level crop yield prediction using MODIS-based variables and weather data: A comparative study on machine learning models. Agricultural and Forest Meteorology, 307, 108530. https://doi.org/10.1016/j.agrformet.2021.108530

What Is Included

• code/: reusable Python modules for data loading, Bayesian hyperparameter search, leave-one-year-out evaluation, scoring, and plotting
• notebooks/: supplementary preprocessing and modeling notebooks with outputs cleared
• data/processed/: compact processed CSV files used by the modeling workflow
• configs/search_spaces/: Bayesian optimization search spaces for each model
• results/: curated model metrics, predictions, and selected best hyperparameters
• DATA_POLICY.md: data provenance and artifact scope for public use

Large raw geospatial/weather files and intermediate extraction outputs are represented by the documented workflow rather than stored directly in this repository.

Models

The comparison covers six regression models:

• Support Vector Regression
• Random Forest Regressor
• Gradient Boosting Regressor
• XGBoost Regressor
• Decision Tree Regressor
• K-Nearest Neighbors Regressor

The default experiment uses 21 Ukrainian regions, March-to-October monthly features, and leave-one-year-out evaluation.

Data

The repository includes two compact processed CSV files:

• data/processed/combined_data.csv: regional monthly MODIS and ERA5-Land features for 2010-2023
• data/processed/crop_yield.csv: regional crop yield table for 2010-2023

Raw and intermediate files used to produce these tables can be placed locally under:

data/external/
data/raw/
data/intermediate/
outputs/

See DATA_POLICY.md for source and artifact details.

Modeling Workflow

The main modeling workflow is in:

notebooks/modeling/training.ipynb

To rerun the modeling notebook, install the dependencies and open Jupyter from the repository root:

pip install -r requirements.txt
jupyter lab notebooks/modeling/training.ipynb

The workflow loads:

data/processed/combined_data.csv
data/processed/crop_yield.csv
configs/search_spaces/*.yaml

and writes refreshed local outputs to results/ or outputs/, depending on whether you are reproducing the curated result tables or generating exploratory artifacts.

Results

Curated result tables are available in:

results/11_20/
results/12_20/

Each experiment folder contains:

• *_metrics.csv: yearly and aggregate metrics
• *_results.csv: true and predicted yield by region and year
• hyperparams/*_best_params.csv: selected model hyperparameters