EY AI & Data Challenge 2026 — Water Quality Prediction

Predicting water quality in South African rivers using Landsat satellite imagery and TerraClimate climate data.

Final score: 0.313 R² | Baseline: 0.20 | Certificate threshold: 0.40 | Leaderboard top: 0.8529

---

The Problem

Predict three water quality parameters at river monitoring sites:

Parameter	Unit	Healthy Range
Total Alkalinity	mg/L	20–200
Electrical Conductance	µS/cm	< 800
Dissolved Reactive Phosphorus	µg/L	< 100

The core challenge: Training and validation locations are in completely different geographic regions. Models that learn location-specific patterns fail. Models that learn general climate–water relationships generalize.

---

Approach

Winning feature set — Climate + Normalized Indices:

Feature	Source	Why it works
PET (evapotranspiration)	TerraClimate	Captures heat/aridity
Precipitation	TerraClimate	Dilution and runoff effects
Soil moisture	TerraClimate	Groundwater quality proxy
Water deficit, max temp, runoff	TerraClimate	Regional climate signal
NDMI, MNDWI	Landsat	Normalized ratios — location-independent
PET/precipitation ratio	Derived	Aridity index
Month (sin/cos encoded)	Date	Seasonal wet/dry cycles

What didn't work: Raw Landsat band values (NIR, green, SWIR). These are absolute brightness values tied to specific locations and don't transfer to new regions.

Model: Random Forest Regressor with conservative depth settings (max_depth=8, min_samples_leaf=15) to reduce overfitting.

---

Score History

Version	Date	Leaderboard R²	Key Change
v1	Mar 11	0.080	17 features, XGBoost — severe overfitting
v2	Mar 11	0.222	10 features, RF vs XGBoost model competition
v3	Mar 11	0.228	Added TerraClimate precipitation
v4	Mar 12	0.289	Strategy comparison — dropped raw Landsat bands, indices + climate only, 70/30 split
v5	Mar 12	0.219	Log-transformed targets + trained on all data (no split) — hurt generalization
v6	Mar 12	0.281	Added soil moisture from TerraClimate
v7	Mar 12	0.119	Strategy B with soil + water deficit (def) — overfit
v8	Mar 13	0.313	Added tmax (max temperature)
v9	Mar 13	0.304	Added runoff (q) — marginal negative effect

Leaderboard context:

	R²
Challenge baseline	0.200
My best (v8)	0.313
Certificate threshold	0.400
Top 10 range	0.765 – 0.853

Scores above 0.4 required more advanced approaches — likely spatial feature engineering, gradient boosting with tuned hyperparameters, or ensemble methods. The top score (0.853) suggests the geographic distribution shift is solvable with the right feature set.

Key learnings:

• Internal CV R² (~0.49) consistently overestimated leaderboard performance (~0.31) — geographic distribution shift between train and validation regions is the core problem
• v4 → v5 regression: log-transforming targets and removing the train/test split improved internal scores but hurt leaderboard performance
• TerraClimate variables helped incrementally (v3, v6, v8) but with diminishing returns — the ceiling for this feature set was ~0.31
• v7's collapse (0.119) shows that feature combinations can interact badly; adding more features is not always better

---

Data Sources

Dataset	Source	Variables Used
Water quality labels	UNEP via EY	Total Alkalinity, EC, DRP
Landsat 8/9	NASA via Microsoft Planetary Computer	NDMI, MNDWI
TerraClimate	Microsoft Planetary Computer	PET, ppt, soil, def, tmax, q

Note: EY-provided datasets (water quality labels, pre-extracted Landsat and TerraClimate features, submission template) are excluded from this repository per the competition Terms & Conditions. To reproduce results, register for the challenge and download from the EY Challenge portal.

---

Project Structure

├── data/
│   ├── terraclimate_training_multi.csv     # Self-extracted climate variables (ppt, soil, def, tmax, q)
│   ├── terraclimate_validation_multi.csv   # Self-extracted climate variables for validation
│   ├── water_quality_training_dataset.csv  # ⛔ Not included (EY confidential)
│   ├── landsat_features_training.csv       # ⛔ Not included (EY confidential)
│   ├── landsat_features_validation.csv     # ⛔ Not included (EY confidential)
│   ├── terraclimate_features_training.csv  # ⛔ Not included (EY confidential)
│   └── terraclimate_features_validation.csv# ⛔ Not included (EY confidential)
├── submissions/
│   ├── submission_final.csv                # Best predictions (v8 — 0.313 R²)
│   └── submission_template.csv             # ⛔ Not included (EY confidential)
├── .gitignore
├── extract_terraclimate.py                 # TerraClimate API extraction script
├── ey_water_quality_model.py               # Final prediction model
├── README.md
└── requirements.txt

---

Usage

# Install dependencies
pip install -r requirements.txt
 
# Run the prediction model (requires EY datasets in data/ folder)
python ey_water_quality_model.py
 
# (Optional) Re-extract TerraClimate data from public API
python extract_terraclimate.py

---

Tech Stack

Python 3.11 · Pandas · NumPy · Scikit-learn · Microsoft Planetary Computer API

---

License

MIT

EY AI & Data Challenge 2026 — submitted March 2026