AI-powered air quality projects: forecasting, sensor fusion and mapping

Air quality has been a recurring theme in my work, from a machine learning internship to hackathon demos. Each project below attacks a different part of the problem with machine learning: forecasting, public-health impact, hyperlocal mapping and low-cost sensing. The AI-based air quality forecasting and reporting system won 2nd place at Research Day 2025 at Sri Ramachandra Faculty of Engineering and Technology.

AeroSight fetches and forecasts air quality data using two models, Facebook Prophet and XGBoost. It correlates pollution levels with public health data such as mortality rates, uses Open-Meteo for weather, and automatically compiles the results into a multi-page PDF report for monitoring stations.

Pollumap V3 is a full-stack dashboard for Chennai with XGBoost 24-hour forecasts, an interactive Leaflet map and an A* healthy-routing feature that finds the least-polluted walking or cycling path. Pollumap V2 was the Smart India Hackathon 2025 demo for Delhi-NCR, and both have demo videos on my YouTube channel.

HyperlocalAQ is a high-resolution air quality modelling framework for Chennai. It integrates seven APIs (WAQI, NASA FIRMS, Open-Meteo, TomTom, OpenStreetMap, CAMS and OpenFDA) with hybrid inverse distance weighting, and adds fire-impact and traffic adjustments before drawing the result on an interactive Folium map.

PHYS-TRUST Lite estimates air quality from a single Arduino node with an MQ135 gas sensor and a DHT11 temperature and humidity sensor. A Kalman filter fuses the sensor physics, room mass-balance physics and an LSTM-ensemble prior, weighting each by a live trust score. It reports a 90% conformal uncertainty band and detects sensor faults, with a live React dashboard.

Forecasting tells people what is coming, mapping and routing help them avoid exposure, multi-source fusion fills gaps between monitoring stations, and sensor fusion makes cheap hardware trustworthy. All four projects are open source on GitHub.

• What will the air be like tomorrow? Pollumap V3 forecasts pollutants 24 hours ahead with XGBoost; AeroSight compares Prophet and XGBoost forecasts.
• How does pollution affect health? AeroSight correlates pollution with mortality records and writes a PDF report per station.
• What is the air like between monitoring stations? HyperlocalAQ blends ground stations with the CAMS satellite model, then adjusts for weather, fires, industry and traffic.
• Which route exposes me to less pollution? Pollumap V3 computes fastest, cleanest and balanced routes with A*.
• Can a cheap sensor be trusted? PHYS-TRUST Lite fuses a low-cost MQ135 sensor with physics and an LSTM prior, and reports its own uncertainty.

The projects share a set of open data sources: WAQI for ground station readings, Open-Meteo for weather, NASA FIRMS for satellite fire hotspots, the Copernicus CAMS model for satellite-based estimates, OpenStreetMap for roads and pollution sources such as industrial zones and landfills, and TomTom for live traffic.

Combining them matters because each one covers a gap in the others: ground stations are accurate but sparse, satellite models cover everywhere but less precisely, and weather and traffic explain why readings change hour to hour.