Hierarchical network design for nitrogen dioxide measurement in urban environments, part 2: network-based sensor calibration

TL;DR

This paper introduces a network-based calibration framework for low-cost NO2 sensors, using data distribution comparisons and co-location data to improve accuracy and detect sensor failures in urban environments.

Contribution

It extends previous linear calibration methods to account for complex sensor responses influenced by environmental factors, demonstrating a scalable calibration approach for large sensor networks.

Findings

Effective estimation of sensor response parameters using distribution divergence minimization.

Successful application of the calibration framework to a 56-sensor network in urban regions.

Identification of sensor and proxy failures through parameter variation analysis.

Abstract

We present a management and data correction framework for low-cost electrochemical sensors for nitrogen dioxide (NO2) deployed within a hierarchical network of low-cost and regulatory-grade instruments. The framework is founded on the idea that it is possible in a suitably configured network to identify a source of reliable proxy data for each sensor site that has a similar probability distribution of measurement values over a suitable time period. Previous work successfully applied these ideas to a sensor system with a simple linear 2-parameter (slope and offset) response. Applying these ideas to electrochemical sensors for NO2 presents significant additional difficulties for which we demonstrate solutions. The three NO2 sensor response parameters (offset, ozone (O3) response slope, and NO2 response slope) are known to vary significantly as a consequence of ambient humidity and temperature variations. Here we demonstrate that these response parameters can be estimated by minimising the Kullback-Leibler divergence between sensor-estimated and proxy NO2 distributions over a 3-day window. We then estimate an additional offset term by using co-location data. This offset term is dependent on climate and spatially correlated and can thus be projected across the network. Co-location data also estimates the time-, space- and concentration-dependent error distribution between sensors and regulatory-grade instruments. We show how the parameter variations can be used to indicate both sensor failure and failure of the proxy assumption. We apply the procedures to a network of 56 sensors distributed across the Inland Empire and Los Angeles County regions, demonstrating the need for reliable data from dense networks of monitors to supplement the existing regulatory networks.