1. Introduction
Reference-grade air quality monitoring stations are expensive to deploy at the density needed to capture hyperlocal pollution variation, motivating dense low-cost sensor networks despite their well-documented drift and cross-sensitivity issues.
2. Methodology
Forty nodes equipped with low-cost optical PM2.5 sensors and electrochemical NO2 sensors were deployed across a 12 square kilometre urban area, publishing readings every 60 seconds over MQTT to an edge gateway. A random forest regression model, trained against a co-located reference-grade analyser, was applied to correct for temperature, humidity and sensor-age drift before data was pushed to a public dashboard.
3. Results
Following calibration, network-wide mean absolute error for PM2.5 dropped from 12.8 to 4.1 micrograms per cubic metre against reference measurements, a 68 percent improvement, while NO2 MAE improved from 9.3 to 3.7 parts per billion. Estimated deployment cost was approximately 3 percent of an equivalent-density reference-grade network.
4. Conclusion
Machine-learning calibration substantially closes the accuracy gap between low-cost and reference-grade air quality sensors, supporting dense urban deployment at a fraction of conventional cost. Future work will investigate transfer of calibration models across sensor batches without site-specific co-location.
References
[1] Castell N. et al., Can commercial low-cost sensor platforms contribute to air quality monitoring, Environment International, 2017. [2] Zheng Y. et al., Forecasting fine-grained air quality based on big data, KDD, 2015.