Surface AQI & HCHO Hotspot Analysis over India

ISRO Hackathon - Problem Statement 3

Predict surface-level Air Quality Index from satellite data using deep learning, and detect formaldehyde (HCHO) hotspots during biomass burning seasons over India.

---

What This Project Does

Objective	Input	Method	Output
Surface AQI Prediction	TROPOMI + MODIS AOD + ERA5 + CPCB	CNN-LSTM Deep Learning	Spatial AQI maps over India
HCHO Hotspot Detection	TROPOMI HCHO + MODIS Fire + ERA5 Wind	Statistical Threshold + DBSCAN	Hotspot maps + Source regions

---

Key Results

Model Performance (CNN-LSTM)

Pollutant	Pearson R	R² Score	RMSE
PM2.5	0.949	0.895	20.0
PM10	0.963	0.924	27.0
NO2	0.888	0.785	9.5
SO2	0.840	0.696	4.2
CO	0.892	0.772	0.4
O3	0.603	0.330	10.5

HCHO Analysis

• Fire-HCHO Correlation: R = 0.81 (1-day transport lag)
• Hotspot Regions: Indo-Gangetic Plain, Eastern UP, NE India, Central India
• Study Period: January 2023 - June 2025 (covers 2 burning seasons)

---

Dashboard Preview

The interactive Streamlit dashboard has 4 tabs:

Tab	What It Shows
How It Works	Project methodology, data pipeline flowcharts, data source table
Objective 1: Surface AQI	Live city AQI checker, station map, gridded AQI, year-over-year comparison
Objective 2: HCHO Hotspots	HCHO heatmap with fire overlay, temporal evolution, lag correlation
Analysis & Validation	Model accuracy, predicted vs actual plots, feature importance

---

Project Architecture

Satellite Data                      Ground Data               Meteorological Data
(TROPOMI S5P)                       (CPCB Stations)           (ERA5/IMDAA)
NO2, SO2, CO, O3, HCHO             PM2.5, PM10, NO2,         Temperature, Wind,
+ MODIS/INSAT-3D AOD               SO2, CO, O3               Pressure, Rainfall
        |                                |                          |
        +----------------+---------------+--------------------------+
                         |
                   ┌─────▼──────┐
                   │  CNN-LSTM  │  30-day sliding window
                   │   Model    │  11 features → 6 outputs
                   └─────┬──────┘
                         |
              ┌──────────┼──────────┐
              ▼          ▼          ▼
        Surface AQI   Gridded    Validation
        per station   AQI Map    Metrics

HCHO Pipeline

TROPOMI HCHO ──→ Mean + 2σ Threshold ──→ DBSCAN Clustering ──→ Source Regions
MODIS Fire   ──→ Fire Count Extraction ──→ Pearson Correlation ──→ Lag Analysis
ERA5 Wind    ──→ Transport Vectors ──→ Wind Field Maps

---

Data Sources

Source	Variables	Access
TROPOMI Sentinel-5P	NO2, SO2, CO, O3, HCHO column densities	Google Earth Engine
INSAT-3D	Aerosol Optical Depth (AOD)	MOSDAC
MODIS (MOD08_D3)	AOD (proxy for INSAT-3D when unavailable)	Google Earth Engine
ERA5 / IMDAA	Temperature, Wind (u,v), Pressure, Precipitation	CDS / NCMRWF
MODIS/VIIRS FIRMS	Active fire locations & brightness temperature	NASA FIRMS
CPCB	PM2.5, PM10, NO2, SO2, CO, O3 (ground-level)	CPCB Data

---

Quick Start

1. Clone & Setup

git clone https://github.com/Adityakumarrama/ISRO.git
cd ISRO/ISRO_Hackathon

# Create virtual environment (Python 3.12 required for TensorFlow)
py -3.12 -m venv venv
venv\Scripts\activate        # Windows
# source venv/bin/activate   # Linux/Mac

pip install -r requirements.txt

2. Run the Pipeline

# Step 1: Download satellite data (requires GEE authentication)
# python 1_download_satellite.py    # Optional - needs earthengine authenticate
# python 2_download_fire.py         # Optional - needs earthengine authenticate

# Step 2: Preprocess & build features (auto-generates synthetic data if GEE unavailable)
python 3_preprocess.py
python 4_feature_matrix.py

# Step 3: Train CNN-LSTM model
python 5_train_model.py

# Step 4: Generate predictions
python 6_predict_aqi.py

# Step 5: HCHO hotspot analysis
python 7_hcho_hotspots.py

# Step 6: Generate static maps & plots
python 8_visualize.py

3. Launch Dashboard

streamlit run dashboard.py

Open http://localhost:8501 in your browser.

---

Project Structure

ISRO_Hackathon/
├── config.py                  # All constants, paths, model hyperparameters
├── 1_download_satellite.py    # GEE download: TROPOMI + MODIS AOD + ERA5
├── 2_download_fire.py         # MODIS/VIIRS fire count download
├── 3_preprocess.py            # Clean, merge, collocate all datasets
├── 4_feature_matrix.py        # Build 30-day sliding window sequences
├── 5_train_model.py           # Train CNN-LSTM (Conv1D→LSTM→Dense)
├── 6_predict_aqi.py           # Generate AQI predictions + gridded map
├── 7_hcho_hotspots.py         # HCHO hotspot detection + fire correlation
├── 8_visualize.py             # Cartopy maps + Matplotlib plots
├── dashboard.py               # Streamlit interactive dashboard
├── requirements.txt           # Python dependencies
├── utils/
│   ├── aqi_calculator.py      # CPCB AQI sub-index formula
│   ├── gee_helpers.py         # Google Earth Engine utilities
│   └── plot_helpers.py        # Cartopy India map helpers
├── data/
│   ├── raw/                   # Downloaded satellite data
│   ├── processed/             # Merged & cleaned datasets
│   └── cpcb/                  # Ground station CSVs
├── models/saved/              # Trained model weights & scalers
└── outputs/
    ├── maps/                  # Spatial maps (AQI, HCHO, clusters)
    └── plots/                 # Time series, scatter, correlation plots

---

CNN-LSTM Model Architecture

Input (30 days × 11 features)
    │
    ├── Conv1D(64, kernel=3, ReLU)     ← Extract spatial patterns
    ├── Conv1D(128, kernel=3, ReLU)    ← Deeper feature extraction
    ├── MaxPooling1D(2)                ← Downsample
    │
    ├── LSTM(128, return_sequences)    ← Learn temporal dependencies
    ├── LSTM(64)                       ← Compress temporal features
    │
    ├── Dropout(0.3)                   ← Prevent overfitting
    ├── Dense(32, ReLU)                ← Final feature combination
    └── Dense(6, Linear)               ← Output: PM2.5, PM10, NO2, SO2, CO, O3

Training Configuration:

• Loss: MSE | Optimizer: Adam (lr=0.001)
• Split: 70% train / 15% validation / 15% test (time-based, not random)
• Early stopping: patience=10 | LR reduction on plateau

---

AQI Calculation (CPCB Standard)

AQI Range	Category	Color	Health Impact
0-50	Good	🟢 Green	Minimal impact
51-100	Satisfactory	🟡 Light Green	Minor discomfort to sensitive people
101-200	Moderate	🟠 Yellow	Breathing discomfort for lung/heart patients
201-300	Poor	🔴 Orange	Breathing discomfort on prolonged exposure
301-400	Very Poor	🟣 Red	Respiratory illness on prolonged exposure
401-500	Severe	⚫ Maroon	Health emergency

Formula: AQI = max(Sub-Index₁, Sub-Index₂, ..., Sub-Indexₙ)

---

HCHO Hotspot Detection Method

1. Threshold: Pixels where HCHO > mean + 2σ flagged as hotspots
2. Clustering: DBSCAN (eps=0.5°, min_samples=5) groups hotspots into source regions
3. Correlation: Pearson R between daily fire count and mean HCHO (+ lag analysis up to 7 days)
4. Transport: ERA5 wind vectors show smoke transport direction from burning regions

---

Tech Stack

Category	Tools
Language	Python 3.12
Deep Learning	TensorFlow 2.21, Keras
Satellite Data	Google Earth Engine API
Data Processing	Pandas, NumPy, Xarray, Scikit-learn
Clustering	DBSCAN (Scikit-learn)
Visualization	Plotly, Matplotlib, Cartopy
Dashboard	Streamlit
Explainability	SHAP / Permutation Importance

---

Relevance to National Priorities

Initiative	Connection
NCAP	Track 20-40% pollution reduction target using satellite-derived AQI
SDG 11.6	Reduce adverse urban environmental impact through air quality monitoring
Chintan Shivir 2.0	Urban climate, air quality, and sustainable development targets
INSAT-3D	Demonstrates use of Indian satellite AOD data for pollution monitoring

---

Team

Built for ISRO Hackathon 2024 | Problem Statement 3

---

Satellite data making air quality information accessible to every Indian citizen.