A human-centered AI system for air pollution understanding, forecasting, and policy simulation. GETHER combines temporal deep learning, causal discovery, counterfactual reasoning, and explainable AI to help researchers, policymakers, and the public understand how emissions affect AQI (Air Quality Index) and what interventions could improve it.
Air pollution prediction alone is not enough. GETHER focuses on understanding pollution, not just predicting it.
The system answers questions like:
- What pollutants cause AQI spikes?
- What happens if emissions are reduced?
- Which environmental factors matter most?
- How reliable are model predictions?
The platform integrates data infrastructure, machine learning, explainability, and policy simulation into a single research pipeline.
1. Temporal AQI Prediction
Predict AQI using a hybrid system combining:
- Feature engineering
- Sequence modeling
- LSTM forecasting
2. Causal Discovery
Identify true causal relationships between pollutants and AQI using:
- Granger causality
- Dynamic causal graphs
- Spatiotemporal correlation analysis
3. Counterfactual Simulation
Simulate environmental policies such as:
- emission reductions
- traffic control
- industrial regulation
Example:
If PM2.5 emissions decrease by 20%, what happens to AQI over the next 30 days?
4. Explainable AI
Model explanations generated using:
- SHAP feature importance
- Local prediction explanation
- Feature contribution ranking
5. Self-Evaluation
The system evaluates its own predictions using:
- prediction confidence
- uncertainty estimation
- model reliability metrics
6. Interactive Dashboard
A Streamlit dashboard enables:
- real-time AQI prediction
- causal graph exploration
- counterfactual policy simulation
- downloadable analysis reports
The model predicts AQI trends based on pollutant history.
The graph compares:
- Actual AQI
- Predicted AQI
This helps evaluate model performance and temporal patterns.
The architecture integrates:
- Data acquisition
- Preprocessing pipeline
- Temporal ML models
- Explainable AI
- Policy simulation engine
- Dashboard interface
- Satellite data acquisition
- Weather API integration
- Raw data storage
- Preprocessing pipeline
- Feature engineering
- Train / validation / test split
- Data validation reports
Baseline models provide initial performance benchmarks.
Models used:
- Linear Regression
- Random Forest
- Basic LSTM
Outputs include:
- model comparison
- error analysis
- feature importance
The final forecasting system uses a deep temporal architecture.
Architecture:
Input Sequence
↓
LSTM (128)
↓
Dropout
↓
LSTM (64)
↓
Dense (32)
↓
AQI Prediction
Enhancements include:
- rolling statistical features
- lag features
- Bayesian hyperparameter tuning
- model averaging
Goal: identify cause-effect relationships in emissions data.
Techniques used:
- Granger causality testing
- spatial correlation analysis
- dynamic causal graph construction
Output:
PM2.5 → AQI
NO2 → AQI
SO2 → AQI
Explainability is provided through:
SHAP Analysis
- feature importance
- local explanations
- model transparency
Uncertainty Estimation
- confidence intervals
- prediction stability
- model reliability scoring
Automated reports are generated for researchers and policy analysts.
A simulation engine allows testing environmental interventions.
Example simulation:
Policy: reduce NO2 emissions by 25%
Baseline AQI forecast: 210
Counterfactual AQI forecast: 168
Average improvement: 20%
Policies can be ranked based on:
- emission reduction
- AQI improvement
- environmental impact
The final application provides:
- interactive AQI prediction
- causal graph visualization
- policy simulation interface
- explainability reports
Built with:
Streamlit
TensorFlow
SHAP
Scikit-learn
gas-emission-prediction
│
├── data
│ ├── raw
│ ├── processed
│
├── notebooks
│
├── src
│ ├── preprocessing.py
│ ├── feature_engineering.py
│ ├── run_preprocessing.py
│ ├── train_lstm.py
│ ├── visualize_predictions.py
│
│ ├── models
│ │ ├── baseline_models.py
│ │ └── advanced_lstm.py
│
│ ├── causal_discovery.py
│ ├── counterfactual_analysis.py
│ ├── explainability.py
│ └── self_evaluation.py
│
├── models
│ └── aqi_lstm_model.h5
│
└── README.md
Baseline experiments:
| Model | R² | MAE |
|---|---|---|
| Linear Regression | 0.13 | 0.066 |
| Random Forest | -0.64 | 0.102 |
| Basic LSTM | -0.37 | 0.092 |
Advanced LSTM:
Test Loss : 0.015
Test MAE : 0.094
Clone repository:
git clone https://github.com/yourusername/gas-emission-prediction.git
cd gas-emission-prediction
Install dependencies:
pip install -r requirements.txt
Preprocess data
python src/run_preprocessing.py
Train models
python src/train_lstm.py
Visualize predictions
python src/visualize_predictions.py
Researchers
- pollution trend analysis
- environmental modeling
- causal inference studies
Policy Makers
- policy intervention evaluation
- emission reduction planning
- air quality forecasting
Public
- AQI awareness
- pollution insights
- environmental impact visualization
Planned improvements include:
- Graph Neural Networks for spatial pollution modeling
- transformer-based temporal forecasting
- real-time satellite data ingestion
- multi-city causal networks
Ekaagra Gupta - B.Tech AI & ML