In search of an optimal in-field calibration method of low-cost gas sensors for ambient air pollutants: Comparison of linear, multilinear and artificial neural network approaches

Abstract The current compliance networks of automatic air-quality monitoring stations in large urban environments are not sufficient to provide spatial and temporal measurement resolution for realistic assessment of personal exposure to pollutants. Small low-cost sensor platforms with greater mobility and expected lower maintenance costs, are increasingly being used as a supplement to compliance monitoring stations. However, low-cost sensor platforms usually provide data with uncertain precision. To improve the precision, these sensor platforms require in-field calibration. Our paper aims to demonstrate that data from each individual sensor system can be corrected using that sensor system's own data to achieve much improved data quality compared to a reference. However, in this procedure, there are practical difficulties such as individual sensor outputs from the multi-sensor system not being sufficiently available due to malfunctions for instance. We explore how this can be dealt with. In our opinion, this is a novel approach, of practical importance both to users and manufacturers. We present a detailed comparative analysis of Linear Regression (univariate), Multivariate Linear Regression and Artificial Neural Networks used with a specific aim of calibrating field-deployed low-cost CO and O 3 sensors. For Artificial Neural Network models, the performance of three common training algorithms was compared (Levenberg-Marquardt, Resilient back-propagation and Conjugate Gradient Powell-Beale algorithm). Data for this study were obtained from two campaigns conducted with 25 multi-sensor AQMESH v.3.5 platforms used within the activities of the CITI-SENSE project. The platforms were co-located to reference gas monitors at the Automatic Monitoring Station Stari Grad, in Belgrade, Serbia. This paper demonstrates that Multivariate Linear Regression and Artificial Neural Network calibration models can improve the output signal. This improvement can be measured by changes in the median and interquartile ranges of statistical parameters used for model evaluation. Artificial Neural Networks showed the best results compared to Linear Regression and Multivariate Linear Regression models. The best predictors for CO, in addition to CO low-cost sensor data, were PM 2.5 and NO 2 , while for O 3 , in addition to O 3 low-cost sensor data, the most suitable input predictors were NO and aH. Based on residual error analysis, we have shown that for CO and O 3 , a certain range of concentrations exists in which calibrated values differ by less than 10% from the reference method results. In addition, it was noted that for all models, CO sensors consistently showed lower variability between platforms compared to O 3 sensors.