Ukraine Crop Yield Estimation Using Machine Learning

---

Project Overview

This project focuses on estimating wheat crop yield in Ukraine using MODIS-based variables and weather data, comparing the performance of various machine learning algorithms. The methodology is adapted from the following research paper:

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.
DOI: 10.1016/j.agrformet.2021.108530

While the original paper focuses on county-level crop yield prediction in the United States, this project targets Ukraine, using wheat as the primary crop. The methodology, data preprocessing, and feature selection have been adjusted accordingly.

---

Research Objective

• Predict wheat yield at a regional level in Ukraine.
• Compare the performance of six machine learning algorithms for yield estimation:
  1. Support Vector Regression (SVR)
  2. Random Forest Regressor (RF)
  3. Gradient Boosting Regressor (GB)
  4. eXtreme Gradient Boosting (XGB)
  5. Decision Tree Regressor (DT)
  6. K-Nearest Neighbors Regressor (KNN)
• Analyze and identify the most suitable machine learning algorithm for yield prediction in Ukraine.

---

Data Description

1. MODIS Data

The following MODIS-based variables were used as features:

• NDVI: Normalized Difference Vegetation Index
• EVI: Enhanced Vegetation Index
• LST (Day & Night): Land Surface Temperature
• FPAR: Fraction of Photosynthetically Active Radiation
• LAI: Leaf Area Index

These variables were preprocessed using weighted monthly averages, computed over different resolutions (250m, 500m, and 1000m), depending on the variable.

2. Weather Data (ERA5 Land)

The following weather variables were extracted from ERA5 Land data:

• 2m Temperature (t2m)
• Soil Temperature (stl1, stl2)
• Surface Solar Radiation (ssr)
• Total Precipitation (tp)
• Volumetric Soil Water (swvl1, swvl2)

---

Methodology

Step 1: Data Preprocessing

1. MODIS Data Preprocessing

   • Monthly weighted averages of MODIS variables were computed.
   • The data was organized into regional-level time series, covering the growing season from March to October (2010–2023).

2. ERA5 Land Data Preprocessing

   • Weather data was extracted and aggregated on a monthly basis for each region.

Step 2: Model Training and Evaluation

• Six different machine learning algorithms were trained on the preprocessed data:

  1. SVR
  2. Random Forest
  3. Gradient Boosting
  4. XGBoost
  5. Decision Tree
  6. KNN

• The models were evaluated based on Mean Absolute Error (MAE) and R-squared (R²) metrics.

---

Results

• The comparison of model performance showed that the Random Forest Regressor (RF) and XGBoost (XGB) performed the best for wheat yield prediction in Ukraine.
• Detailed results can be found in the results folder.

---