Air Pollution

Problem

Delhi's regulators decide where to act on air pollution using a network of 32 public monitors. That network leaves three questions open, and each depends on the one before it. The first is where pollution is high, since a sparse network can miss whole neighbourhoods. The second is which sources produced it, since a regulator can act on a source only if the data separate it from the others. The third is what a regulator should do when the attribution is too uncertain to justify a fine on its own.

Approach and findings

To see what the public network misses, we deployed 28 low-cost sensors alongside it from May 2018 to November 2020, at about 0.2% of the cost of equivalent reference-grade monitors. The combined network found 189 hotspots that the public monitors alone did not detect, and more than 150,000 people live in hotspot neighbourhoods that are absent from the city government's map. A space-time Kriging model predicts hotspots at locations with no sensor, with 95% precision and 88% recall when half of the sensors are removed. An earlier study asked whether phone photographs could stand in for sensors. Two government monitors at the same site disagree by about 55 µg/m³, so every label carries noise of that size, and we reported the fraction of predictions within that floor alongside the mean error.

Knowing where concentrations are high does not say which sources produced them, and two sources that leave similar traces at the monitors cannot be told apart by any fit. Our source-attribution method decides from the transport model, before fitting, which groups of sources the readings can separate, and it reports each group's contribution with an error bound. In controlled experiments on the Delhi domain, it flags an absorbed source and a nearly proportional pair of sources that standard methods do not flag. An earlier workshop paper framed the same question as causal inference, with the plume dispersion model as the mechanism linking emissions to readings.

When an attribution is too uncertain to justify a fine, a regulator can still use it to decide which sources to audit, because an audit measures a source's emission exactly. We study this choice as an online learning problem in which each source adjusts its emission to its audit probability. When emissions vary from period to period, the excess emission allowed by the best learning rule, compared with the best fixed audit rule chosen in hindsight, grows in proportion to T2/3 over T periods, up to logarithmic factors. This holds whether the regulator learns from audits, from sensor readings or from both.

Papers

Comprehensive Monitoring of Air Pollution Hotspots Using Sparse Sensor Networks
Ankit Bhardwaj, Ananth Balashankar, Shiva Iyer, Nita Soans, Anant Sudarshan, Rohini Pande, Lakshminarayanan Subramanian
ACM Journal on Computing and Sustainable Societies, 2025. Presented at ACM COMPASS 2025. Selected for ACM Showcase. [Paper] [Code] [Slides] [ACM Showcase]

Learning Pollution Maps from Mobile Phone Images
Ankit Bhardwaj, Shiva Iyer, Yash Jalan, Lakshminarayanan Subramanian
International Joint Conference on Artificial Intelligence (IJCAI), AI for Good track, 2022. Oral presentation. [Paper] [Code]

Identifiability-Aware Source Attribution of Air Pollution
Ankit Bhardwaj, Lakshminarayanan Subramanian
Under review.

Learning Where to Audit Polluters from Noisy Instruments
Ankit Bhardwaj, Lakshminarayanan Subramanian
Under review.

Towards Causal Understanding of Urban Air Pollution: Mechanistic Models under Sparse Sensing
Ankit Bhardwaj, Lakshminarayanan Subramanian
NeurIPS Workshop on Uncovering Causality in Science, 2025. [Paper]

Collaborators

• Shiva Iyer (NYU / Toyota ITC)

• Ananth Balashankar (Google DeepMind)

• Anant Sudarshan (University of Warwick)

• Rohini Pande (Yale University)

• Lakshminarayanan Subramanian (NYU, advisor)