🏆 PM2.5 Air Quality Prediction — Rank #1 Solution

---

📌 Competition Overview

The goal was to predict daily PM2.5 particulate matter concentration (µg/m³) for 179 cities across the globe — cities that were completely unseen during training. This made generalization the central challenge rather than raw predictive power.

Data came from three sources:

• 🛰️ Sentinel-5P Satellite — atmospheric pollutants (NO2, CO, HCHO, aerosol index, ozone, etc.)
• 🌦️ GFS Weather Data — temperature, humidity, wind components, precipitation
• 📡 Ground Sensors — actual PM2.5 readings (train only; target variable)

---

🎯 Key Results

Metric	Score
Public Leaderboard RMSE	31.755
Private Leaderboard RMSE	29.153
Final Rank	#1 / 93

---

💡 Winning Insight

The single most impactful decision was the target transformation.

Instead of the common log(1 + y) transform:

# ❌ log1p — over-compresses extreme values
y_train = np.log1p(train['target'])   # OOF raw RMSE: 27.12

# ✅ sqrt — the winner
y_train = np.sqrt(train['target'])    # OOF raw RMSE: 25.93

Why sqrt won: The test set contains high-pollution cities the model had never seen. log1p compresses large values too aggressively, causing the model to underpredict extreme PM2.5 concentrations. sqrt keeps predictions better calibrated across the full range.

---

🔧 Solution Pipeline

1. Feature Engineering

Dropped columns

• All 7 CH4 satellite columns (~81% missing, weak correlation)
• Satellite geometry angles (sensor/solar azimuth & zenith — near-zero correlation)

Engineered features

Feature	Description
wind_speed	√(u² + v²) — actual wind magnitude from GFS components
month, week_of_year	Captured COVID-19 lockdown emission drop from Week 5
day_of_week, day_of_month	Fine-grained temporal signal
CO_NO2_ratio	Pollution source type indicator
HCHO_NO2_ratio	Biomass burning vs traffic signal
CO_HCHO_ratio	Combustion source ratio
Rolling 3-day & 7-day means	Temporal momentum of key pollutants per city

2. City Clustering

Since train and test cities have zero overlap, test cities needed an identity. KMeans clustering (k=8) on median satellite + weather profiles grouped all cities into pollution archetypes — then each test city inherited the archetype's historical PM2.5 statistics.

km = KMeans(n_clusters=8, random_state=42, n_init=20)
train_city['city_cluster'] = km.fit_predict(tr_scaled)
test_city['city_cluster']  = km.predict(te_scaled)   # unseen cities get archetype

3. Model

LightGBM with 5-fold cross-validation:

model = LGBMRegressor(
    n_estimators=2000, learning_rate=0.02,
    max_depth=6,       num_leaves=63,
    subsample=0.75,    colsample_bytree=0.75,
    min_child_samples=20,
    reg_alpha=0.1,     reg_lambda=1.0,
    random_state=42
)

Predictions reversed with pred² → PM2.5.

---

📈 Experiment History

Version	Change	CV RMSE (raw)	LB RMSE
Baseline XGBoost	Raw features, log transform	31.24	34.92
v2 — LightGBM	City clusters + ratios + rolling	27.12	32.49
v3 — Over-engineered	Too many features (87 total)	24.39	32.65 ← overfit
Final — sqrt	Same v2 features, sqrt target	25.93	31.755 🏆

The v3 experiment taught a valuable lesson: aggressively improving CV on train cities can hurt leaderboard performance when test cities are completely different. Less can be more.

---

📂 Repository Structure

pm25-air-quality-prediction/
│
├── PM25_Rank1_Solution.ipynb   
├── submission_final.csv        
└── README.md                   

Note: Train.csv, Test.csv, and SampleSubmission.csv are not included due to Zindi's data sharing policy. Download them from https://zindi.africa/competitions/university-hackathon-by-olabisi-onabanjo-university/submissions

---

🛠️ Requirements

pandas
numpy
scikit-learn
lightgbm
matplotlib

Install with:

pip install pandas numpy scikit-learn lightgbm matplotlib

---

🚀 How to Reproduce

1. Clone the repo and download the data from Zindi
2. Place Train.csv, Test.csv, SampleSubmission.csv in the root directory
3. Open PM25_Rank1_Solution.ipynb in Jupyter or Google Colab
4. Run all cells — final predictions saved as submission_final.csv

---

📚 Data Sources

• Competition: Zindi — University Hackathon by Olabisi Onabanjo University
• Weather data: NOAA GFS 0.25°
• Satellite data: Copernicus Sentinel-5P via Google Earth Engine

---

Built as part of ongoing competitive ML practice and community contribution to African AI.