Abstract View
What is the Value of Information from Low Cost Sensor Networks? Balancing Sampling and Instrument Uncertainty
ROSE EILENBERG, R. Subramanian, Aliaksei Hauryliuk, Carl Malings, Albert Presto, Allen Robinson, Carnegie Mellon University
Abstract Number: 523
Working Group: Urban Aerosols
Abstract
It is hypothesized that the increased spatial density enabled by networks of lower-cost sensors (LCS) can be used to better capture the spatiotemporal variability of urban air pollution. However, the measurement uncertainty of LCSs is larger than regulatory-grade instruments.
To test that hypothesis, we modeled their performance under different conditions. First, we characterized the measurement uncertainty for the O3, PM2.5, CO, and NO2 sensors in the Real-time Affordable Multi-Pollutant monitor when collocated with regulatory-grade instruments. The overall uncertainties were concentration-dependent and ranged from 62%-133%; the slopes of concentration-dependent bias ranged from -0.22 to -0.66. Second, we performed Monte-Carlo simulations using highly spatially resolved data from a national land-use regression model as the ground truth to evaluate different hypothetical low-cost sensor networks. Using networks ranging from 1 to 100 nodes, we estimated population-weighted average concentrations, concentration disparities, and the influence of point sources.
Our analysis highlights that low-cost monitoring networks mainly provide advantages for pollutants with large spatial variability such as NO2, for which the estimate uncertainty of a single-monitor network with no measurement error can be 70%. However, a large sensor network is unnecessary for a relatively spatially invariant pollutant like ozone, for which the sampling error of an error-free single monitor can be 2.5%. Concentration-dependent biases can cause substantial challenges; for example, when two areas have average concentrations that vary by 20%, on average 4 error-free monitors are needed in each area. With a concentration-dependent bias slope of -1/4, five monitors are required. With a concentration-dependent bias of -1/2, 18 monitors are required on average. We found similar results when identifying the influence of a point source; in most cases, no difference was detectable when the concentration-dependent bias slope exceeded -1/4. When measurement error models are unbiased, the performance of LCSs can be similar to regulatory monitors over long averaging times.