Predictive Modelling for Maize and Bean Yields in Kitui County, Kenya

Author: Mocraig Sande
Project Members: Walter Murimi · James Muteti · John Mutisya · Jeremy Kipkurui Programme: B.Sc. Applied Statistics with Computing, Moi University
Supervisor: Prof. Ann Mwangi
Academic Year: 2025/2026

---

Project Overview

This capstone project develops and validates predictive models for maize and bean yields among smallholder farmers in Kitui County, Kenya — one of Kenya's most food-insecure arid and semi-arid regions. Four modelling approaches are compared: Multiple Linear Regression (OLS), LASSO Regression, Random Forest, and XGBoost.

Key Results (Summary):

Crop	Best Model	RMSE (t/ha)	R²
Maize	LASSO	0.232	0.594
Beans	LASSO	0.155	0.633

---

Data Sources

Source	Variables	URL
Synthetic survey (n=200)	All agronomic & socioeconomic	Generated — see notebooks/02_synthetic_survey.ipynb
NASA POWER	Rainfall, temp, solar	power.larc.nasa.gov
SoilGrids 2.0	pH, SOC, clay, sand	rest.soilgrids.org

---

Repository Structure

kitui-yield-prediction/
├── data/
│   ├── synthetic/        # Synthetic farmer survey (n=200)
│   ├── raw/              # API downloads (gitignored, regenerable)
│   └── processed/        # Merged, clean, feature-engineered
├── notebooks/            # Colab notebooks (numbered, run in order)
├── src/                  # Reusable functions
├── outputs/
│   ├── figures/          # All plots (180–200 dpi PNG)
│   ├── tables/           # Model results, summaries
│   └── maps/             # Interactive HTML maps
├── docs/                 # Questionnaire, codebook, sources
└── tests/                # Pytest unit tests

Reproduce This Project

# 1. Clone
git clone git@github.com:crayglockes/kitui-yield-prediction.git
cd kitui-yield-prediction

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run notebooks in order (01 → 07) in Google Colab or Jupyter
# 4. Run tests
pytest tests/ -v

All random seeds are set to 42. Results are fully reproducible.

---

Limitations

1. Data are synthetic — calibrated against published distributions but not from an actual field survey. Findings should be validated with primary data before policy application.
2. SoilGrids and NASA POWER data represent 250m–0.5° resolution averages; farm-level soil and microclimate variation is not fully captured.
3. Models are cross-sectional (single season); temporal dynamics of multi-year yield trends are not modelled.

---

Ethical Statement

Synthetic data generation methodology is fully documented in notebooks/02_synthetic_survey.ipynb. No real farmer data was collected without consent. The questionnaire (docs/questionnaire.md) represents the intended instrument for future primary data collection under Moi University IREC ethics protocol.