Two step calibration method for ozone low-cost sensor: Field experiences with the UrbanSense DCUs

Urban air pollution is a global concern impairing citizens' health, thus monitoring is a pressing need for city managers. City-wide networks for air pollution monitoring based on low-cost sensors are promising to provide real-time data with detail and scale never before possible. However, they still present limitations preventing their ubiquitous use. Thus, this study aimed to perform a post-deployment validation and calibration based on two step methods for ozone low-cost sensor of a city-wide network for air pollution and meteorology monitoring using low-cost sensors focusing on the main challenges. Four of the 23 data collection units (DCUs) of the UrbanSense network installed in Porto city (Portugal) with low-cost sensors for particulate matter (PM), carbon monoxide (CO), ozone (O-3), and meteorological variables (temperature, relative humidity, luminosity, precipitation, and wind speed and direction) were evaluated. This study identified post-deployment challenges related to their validation and calibration. The preliminary validation showed that PM, CO and precipitation sensors recorded only unreliable data, and other sensors (wind speed and direction) very few data. A multi-step calibration strategy was implemented: inter-DCU calibration (1st step, for O-3, temperature and relative humidity) and calibration with a reference-grade instrument (2nd step, for O-3). In the 1st step, multivariate linear regression (MLR) resulted in models with better performance than non-linear models such as artificial neural networks (errors almost zero and R-2 > 0.80). In the 2nd step, the calibration models using non-linear machine learning boosting algorithms, namely Stochastic Gradient Boosting Regressor (both with the default and posttuning hyper-parameters), performed better than artificial neural networks and linear regression approaches. The calibrated O-3 data resulted in a marginal improvement from the raw data, with error values close to zero, with low predictability (R-2 similar to 0.32). The lessons learned with the present study evidenced the need to redesign the calibration strategy. Thus, a novel multi-step calibration strategy is proposed, based on two steps (pre and post-deployment calibration). When performed cyclically and continuously, this strategy reduces the need for reference instruments, while probably minimising data drifts over time. More experimental campaigns are needed to collect more data and further improve calibration models