Abstract. Low-cost sensors offer an attractive solution to the challenge of establishing affordable and dense spatio-temporal air quality monitoring networks with greater mobility and lower maintenance costs. These low-cost sensors offer reasonably consistent measurements but require in-field calibration to improve agreement with regulatory instruments. In this paper, we report the results of a deployment and calibration study on a network of six air quality monitoring devices built using the Alphasense O3 (OX-B431) and NO2 (NO2-B43F) electrochemical gas sensors. The sensors were deployed in two phases over a period of 3 months at sites situated within two megacities with diverse geographical, meteorological and air quality parameters. A unique feature of our deployment is a swap-out experiment wherein three of these sensors were relocated to different sites in the two phases. This gives us a unique opportunity to study the effect of seasonal, as well as geographical, variations on calibration performance. We report an extensive study of more than a dozen parametric and non-parametric calibration algorithms. We propose a novel local non-parametric calibration algorithm based on metric learning that offers, across deployment sites and phases, an R2 coefficient of up to 0.923 with respect to reference values for O3 calibration and up to 0.819 for NO2 calibration. This represents a 4–20 percentage point increase in terms of R2 values offered by classical non-parametric methods. We also offer a critical analysis of the effect of various data preparation and model design choices on calibration performance. The key recommendations emerging out of this study include (1) incorporating ambient relative humidity and temperature into calibration models; (2) assessing the relative importance of various features with respect to the calibration task at hand, by using an appropriate feature-weighing or metric-learning technique; (3) using local calibration techniques such as k nearest neighbors (KNN); (4) performing temporal smoothing over raw time series data but being careful not to do so too aggressively; and (5) making all efforts to ensure that data with enough diversity are demonstrated in the calibration algorithm while training to ensure good generalization. These results offer insights into the strengths and limitations of these sensors and offer an encouraging opportunity to use them to supplement and densify compliance regulatory monitoring networks.

Abstract. Rising awareness of the health risks posed by elevated levels of ground-level O3 and NO2 have led to an increased demand for affordable and dense spatio-temporal air quality monitoring networks. Low-cost sensors used as a part of Internet of Things (IoT) platforms offer an attractive solution with greater mobility and lower maintenance costs, and can supplement compliance regulatory monitoring stations. These commodity low-cost sensors have reasonably high accuracy but require in-field calibration to improve precision. In this paper, we report the results of a deployment and calibration study on a network of seven air quality monitoring devices built using the Alphasense O3 (OX-B431) and NO2 (NO2-B43F) electrochemical gas sensors. The sensors were deployed at sites situated within two mega-cities with diverse geographical, meteorological and air quality parameters – Faridabad (Delhi National Capital Region) and Mumbai, India. The deployment was done in two phases over a period of three months. A unique feature of our deployment is a swap-out experiment wherein four of these sensors were relocated to different sites in the two deployment phases. Such a diverse deployment with sensors switching places gives us a unique opportunity to ablate the effect of seasonal, as well as geographical variations on calibration performance. We perform an extensive study of more than a dozen parametric as well as non-parametric calibration algorithms and find local calibration methods to offer the best performance. We propose a novel local calibration algorithm based on metric-learning that offers, across deployment sites and phases, an average R2 coefficient of 0.873 with respect to reference values for O3 calibration and 0.886 for NO2 calibration. This represents an upto 9 % increase over R2 values offered by classical local calibration methods. In particular, our proposed model far outperforms the default calibration models offered by the gas sensor manufacturer. We also offer a critical analysis of the effect of various data preparation and model design choices on calibration performance. The key recommendations emerging out of this study include (1) incorporating ambient relative humidity and temperature as free parameters (or features) into all calibration models, (2) assessing the relative importance of various features with respect to the calibration task at hand, by using an appropriate feature weighing or metric learning technique, (3) the use of local (or even hyper-local) calibration techniques such as k-NN that seem to offer the best performance in high variability conditions such as those encountered in field deployments, (4) performing temporal smoothing over raw time series data, say by averaging sensor signals over small windows, but being careful to not do so too aggressively, and (5) making all efforts at ensuring that data with enough diversity is demonstrated to the calibration algorithm while training to ensure good generalization. These results offer insights into the strengths and limitations of these sensors, and offer an encouraging opportunity at using them to supplement and densify compliance regulatory monitoring networks.

No abstract

This document presents supporting material to the main paper, including descriptions of various baseline parametric and non-parametric baseline calibration algorithms, as well as details about the calibration performance offered by various algorithms that are not included in the main paper. Appendix A: A Recapitulation of Notation and Error Metrics For sake of convenience to the reader, we revisit the definitions of the error metrics, a discussion on how to interpret the 5 statistical tests, and the notation. A1 Error Metrics and Statistical Hypothesis Testing Error Metrics: calibration performance was measured using four popular metrics, mean averaged error (MAE), mean absolute percentage error (MAPE), root mean squared error (RMSE), and the coefficient of determination (R 2) (see below). Here n denotes the number of test points for a given dataset and split thereof, the variable t runs over all time-stamps in the testing set, 10 y t denotes the reference value (either O 3 or NO 2) at the t-th time-stamp,ŷ t denotes the corresponding value predicted by the