Mapping Air Pollution Sources
with Sequential Transformer Chaining

A Comprehensive Case Study in South Asian Industrial Estates•NeurIPS 2024

AI-Powered Analysis
Environmental Impact
Health Correlations
Detection Coverage
99%

Percentage of irrelevant data filtered out

Manual Annotation Data
1%

Data manually annotated after transformer chaining

Pollutants Analyzed
3

SO₂, NO₂, and CO concentrations

Unique Dataset
Yes

Annotated chimney detection dataset

Chimney Index Developed
Yes

Novel multispectral chimney index

ARI Hotspots Identified
Yes

Acute Respiratory Infection hotspots

Transformer Models Used
2

ViT and Remote CLIP models applied

High Detection Accuracy
95%

Final accuracy after manual annotation

Detection Rate in Delhi
79.17%

Detection rate in Patparganj Industrial Area

Detection Rate in Lahore
68.42%

Detection rate in Sundar Industrial Estate

Detection Rate in Dhaka
70%

Detection rate in Tongi Industrial Estate

Methodology Overview
Advanced approach to mapping industrial pollution sources

1
Data Collection

  • •Sentinel-5P satellite data (SO₂, NO₂, CO)
  • •High-resolution imagery (0.5m/pixel)
  • •Health data from 15+ regional hospitals
  • •Meteorological patterns & wind data

2
AI Processing Chain

  • •Enhanced Vision Transformer (ViT)
  • •Dual Remote CLIP refinement
  • •Advanced GAN enhancement
  • •Expert verification system

3
Analysis Framework

  • •Chimney Index computation
  • •Density mapping algorithms
  • •Temporal emission analysis
  • •Health correlation metrics
Regional Detection Performance
Comparative analysis across major industrial zones
Transformer Chain Performance
Metrics across sequential processing stages
Health Impact Correlation
Monthly health incidents in studied regions
Industrial Pollutant Distribution Analysis
Breakdown of major pollutants detected across industrial zones
Conclusion and Future Work
Key findings and upcoming research directions

Our study presents a comprehensive framework for detecting pollution sources, specifically factory and brick kiln chimneys, in major South Asian cities using a combination of remote sensing data and advanced deep learning techniques. The sequential transformer chaining method effectively filters out 99% of irrelevant data from high-resolution imagery. This approach provides actionable insights for public health interventions and supports regulatory measures aimed at achieving the United Nations' Sustainable Development Goal 3 on health and well-being.

Key Achievements:

  • Developed a comprehensive framework for detecting pollution sources in South Asia
  • Sequential transformer chaining filters out 99% of irrelevant data
  • Introduced a unique annotated chimney detection dataset improving detection accuracy

Future Directions:

  • Plan to make the dataset and code publicly available after acceptance