Development and Application of a United States wide correction for PM2.5 data collected with the PurpleAir sensor

Abstract

PurpleAir sensors, which measure particulate matter (PM), are widely used by individuals, community groups, and other organizations including state and local air monitoring agencies. PurpleAir sensors comprise a massive global network of more than 10,000 sensors. Previous performance evaluations have typically studied a limited number of PurpleAir sensors in small geographic areas or laboratory environments. While useful for determining sensor behavior and data normalization for these geographic areas, little work has been done to understand the broad applicability of these results outside these regions and conditions. Here, PurpleAir sensors operated by air quality monitoring agencies are evaluated in comparison to collocated ambient air quality regulatory instruments. In total, almost 12,000 24-hour averaged PM_2.5 measurements from collocated PurpleAir sensors and Federal Reference Method (FRM) or Federal Equivalent Method (FEM) PM_2.5 measurements were collected across diverse regions of the United States (U.S.), including 16 states. Consistent with previous evaluations, under typical ambient and smoke impacted conditions, the raw data from PurpleAir sensors overestimate PM_2.5 concentrations by about 40% in most parts of the U.S. A simple linear regression reduces much of this bias across most U.S. regions, but adding a relative humidity term further reduces the bias and improves consistency in the biases between different regions. More complex multiplicative models did not substantially improve results when tested on an independent dataset. The final PurpleAir correction reduces the root mean square error (RMSE) of the raw data from 8 μg m^-3 to 3 μg m^-3 with an average FRM or FEM concentration of 9 μg m^-3. This correction equation, along with proposed data cleaning criteria, has been applied to PurpleAir PM_2.5 measurements across the U.S. in the AirNow Fire and Smoke Map (fire.airnow.gov) and has the potential to be successfully used in other air quality and public health applications.

Figure 1.

Comparison of 24-hour averaged PM_2.5 data from the PurpleAir A and B channels. Excluded data (2.1%) are shown in red and represent data points where channels differed by more than 5 μg m⁻³ and 61%. AK3, CA7, WA5 were excluded from further analysis. Pearson correlation (r) is shown on each plot.

Figure 2.

State, local, and tribal (SLT) air monitoring sites with collocated PurpleAir sensors. Includes regions used for correction model evaluation.

Figure 3.

Comparison of the 24-hour raw PurpleAir (PA) cf_1 and cf_atm PM_2.5 outputs (A) and both outputs compared to the FEM or FRM PM_2.5 measurements (B and C) across all sites with the 1:1 line in red.

Figure 4.

Performance statistics including mean bias error (MBE) and mean absolute error (MAE) are shown by correction method (0–7), where each point in the boxplot is the performance for either a 12-week period excluded from correction building (“LOBD”), or a single state excluded from correction building (“LOSO”).

Figure 5.

Error and ratio between corrected PurpleAir (PA) and FRM or FEM measurements are shown along with corrected PurpleAir PM_2.5 data (corrected using Eq. 10) as influenced by temperature, RH, and FRM or FEM PM_2.5 concentration. Colors indicate states, and black points indicate averages in 10 bins.

Figure 6.

Scatterplot of the daily FEM or FRM PM_2.5 data with the PurpleAir data by U.S. region (see Figure 2) prior to any correction, after applying a linear correction, and after applying the final correction including RH. Data were corrected using the models built for the full dataset.

Figure 7.

24-hr AQI categories as measured by the corrected PurpleAir and the FEM or FRM for the full dataset generated with the models built using LOSO withholding.