Project Canary

Short-Term Forecasting of Ground-Level O₃ and NO₂ Using Multi-Source Data Fusion

Smart India Hackathon 2025 – Problem Statement SIH25178

---

🌍 Overview

Project Canary is a hybrid deep-learning model for forecasting NO₂ and O₃ concentrations by fusing satellite data, meteorological reanalysis, and ground-station observations. The system uses a CNN + BiLSTM + Attention pipeline to capture spatial pollutant behavior, temporal evolution, and weather-driven influences.

---

🚀 Key Highlights

• Multi-source fusion of Sentinel-5P, ERA5, and CPCB data.
• Hybrid deep learning architecture combining spatial and temporal modeling.
• Simultaneous forecasting of NO₂ and O₃.
• Spatial raster encoding via CNN.
• Temporal dynamics captured via BiLSTM.
• Interpretability through attention weights.

---

📡 Data Sources

Satellite (Sentinel-5P TROPOMI)

• NO₂ and O₃ tropospheric column densities
• Processed from GeoTIFF
• Resampled to 10 × 11 spatial grid aligned with Delhi AOI

Meteorological (ERA5)

• Hourly variables: Temp, RH, wind (u/v), precipitation, solar radiation, boundary-layer height, dewpoint, etc.
• Interpolated to station coordinates

Ground-Station (CPCB)

• Hourly NO₂ & O₃ measurements
• Used as supervised labels

---

🛠 Data Pipeline

1. Ingestion

   • Load Sentinel-5P rasters
   • Load ERA5 weather series
   • Load CPCB pollutant readings

2. Preprocessing

   • Temporal alignment
   • Imputation (previous-day fill, KNN)
   • Z-score normalization
   • Derived features (dewpoint, cyclical time features)

3. Spatial Grid Preparation

   • Crop satellite rasters to region
   • Resample to fixed 10×11 grid
   • Stack channels (NO₂, O₃, meteo grids)

4. Final Input Shapes

   • Satellite tensors: (T, 9, 10, 11)
   • Tabular station features: (T, 18)
   • Labels: (T, 2) for NO₂ & O₃

---

🧠 Model Architecture

Satellite Raster (T, C, 10, 11)
        ↓
     CNN Encoder
   (Conv → ReLU → MaxPool × 3)
        ↓
   64-Dim Spatial Embedding

Station Features (T, 18)
        ↓
        MLP
        ↓
   64-Dim Tabular Embedding

Fusion → BiLSTM → Attention → Dual Regression Heads

Components (compact summary)

• CNN Encoder: Learns spatial pollutant patterns from raster grids.

• MLP Block: Encodes meteorological + derived features.

• Fusion Layer: Concatenates spatial and tabular embeddings.

• BiLSTM: Models temporal evolution across past 24 hours.

• Attention: Highlights key timesteps (e.g., rush hours, wind changes).

• Output Heads:

  • NO₂ concentration
  • O₃ concentration

---

📈 Model Performance

NO₂

• MAE: 12.35
• RMSE: 17.89
• R²: 0.755

O₃

• MAE: 14.65
• RMSE: 20.91
• R²: 0.666

---

🔧 Tech Stack

• PyTorch – modeling
• scikit-learn – preprocessing & feature engineering
• NumPy / Pandas – time-series & data handling
• Rasterio – GeoTIFF reading & spatial ops

---

📚 References

Attention mechanism based CNN-LSTM hybrid deep learning model for atmospheric ozone concentration prediction