Scalable Urban Air Pollution Monitoring

Motivation

Urban air pollution poses a critical public health risk, particularly in densely populated cities of the Global South. However, fine-grained monitoring is hindered by the high cost and low coverage of industrial-grade air quality sensors. Without detailed spatiotemporal data, many pollution hotspots remain undetected, reducing the effectiveness of mitigation policies.

This project aims to improve air pollution monitoring resolution by two independent and complimentary approaches:

1. Alternative sensing using vision based ML algorithms running on top of mobile phone and CCTV images.

2. Interpolation based modeling of sparse sensor networks for pollution hotspot detection.

Contributions

Approach 1: Vision-based sensing from mobile and CCTV images

• We developed machine learning models to estimate PM2.5 levels from low-cost images, enabling high-resolution pollution sensing using mobile phones and cameras.

• We designed interpretable models that use haze as a proxy for particulate matter and deployed them on mobile devices for real-time, location-aware monitoring.

Approach 2: Interpolation modeling of sparse sensor networks

• We built spatial interpolation techniques, including Space-Time Kriging, to estimate pollution levels at unmonitored locations using sparse sensor measurements.

• We showed robustness to label noise and sensor failures, enabling accurate hotspot detection even with incomplete or low-fidelity data.

These two approaches can be unified in a sequential pipeline, where vision-based estimates enhance spatial resolution and interpolation models refine them into accurate, city-scale pollution maps.

Results

• Image-based model achieves MAE of 44 µg/m³ and improves to 35 µg/m³ using distributed averaging. In-interval accuracy reaches 74% despite noisy labels and limited data.

• The interpolation model achieves 95% precision and 88% recall for monthly hotspot prediction even with 50% sensor failure.

• Mechanistic model explains 65% of transient hotspots, providing policy-relevant insights.

• Augmented network reveals 189 previously undetected hotspots, affecting over 150,000 residents not covered by public monitors.

Papers

• A. Bhardwaj, S. Iyer, Y. Jalan, L. Subramanian. Learning Pollution Maps from Mobile Phone Images. IJCAI 2022, AI for Good Track.  PDF. Code.

• A. Bhardwaj, A. Balashankar, S. Iyer, N. Soans, A. Sudarshan, R. Pande, L. Subramanian. Comprehensive Monitoring of Air Pollution Hotspots Using Sparse Sensor Networks. ACM Journal on Computing and Sustainable Societies, 2025. PDF. Code.

People

• Ankit Bhardwaj (NYU)

• Shiva Iyer (NYU / Toyota ITC)

• Ananth Balashankar (NYU / Google)

• Anant Sudarshan (University of Warwick)

• Rohini Pande (Yale University)

• Lakshminarayanan Subramanian (Advisor, NYU)