Leveraging low-cost sensors to predict nitrogen dioxide for epidemiologic exposure assessment

Abstract

Background

Statistical models of air pollution enable intra-urban characterization of pollutant concentrations, benefiting exposure assessment for environmental epidemiology. The new generation of low-cost sensors facilitate the deployment of dense monitoring networks and can potentially be used to improve intra-urban models of air pollution.

Objective

Develop and evaluate a spatiotemporal model for nitrogen dioxide (NO₂) in the Puget Sound region of WA, USA for the Adult Changes in Thought Air Pollution (ACT-AP) study and assess the contribution of low-cost sensor data to the model’s performance through cross-validation.

Methods

We developed a spatiotemporal NO₂ model for the study region incorporating data from 11 agency locations, 364 supplementary monitoring locations, and 117 low-cost sensor (LCS) locations for the 1996–2020 time period. Model features included long-term time trends and dimension-reduced land use regression. We evaluated the contribution of LCS network data by comparing models fit with and without sensor data using cross-validated (CV) summary performance statistics.

Results

The best performing model had one time trend and geographic covariates summarized into three partial least squares components. The model, fit with LCS data, performed as well as other recent studies (agency cross-validation: CV- root mean square error (RMSE) = 2.5 ppb NO₂; CV- coefficient of determination (${R}^{2}$) = 0.85). Predictions of NO₂ concentrations developed with LCS were higher at residential locations compared to a model without LCS, especially in recent years. While LCS did not provide a strong performance gain at agency sites (CV-RMSE = 2.8 ppb NO₂; CV-${R}^{2}$ = 0.82 without LCS), at residential locations, the improvement was substantial, with RMSE = 3.8 ppb NO₂ and ${R}^{2}$ = 0.08 (without LCS), compared to CV-RMSE = 2.8 ppb NO₂ and CV-${R}^{2}$ = 0.51 (with LCS).

Impact

We developed a spatiotemporal model for nitrogen dioxide (NO₂) pollution in Washington’s Puget Sound region for epidemiologic exposure assessment for the Adult Changes in Thought Air Pollution study. We examined the impact of including low-cost sensor data in the NO₂ model and found the additional spatial information the sensors provided predicted NO₂ concentrations that were higher than without low-cost sensors, particularly in recent years. We did not observe a clear, substantial improvement in cross-validation performance over a similar model fit without low-cost sensor data; however, the prediction improvement with low-cost sensors at residential locations was substantial. The performance gains from low-cost sensors may have been attenuated due to spatial information provided by other supplementary monitoring data.