the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Towards a Universal Hygroscopic Growth Calibration for Low-Cost PM2.5 Sensors
Milan Y. Patel
Pietro F. Vannucci
Jinsol Kim
William M. Berelson
Abstract. Low-cost particulate matter (PM) sensors continue to grow in popularity, but issues such as aerosol size-dependent sensitivity drive the need for effective calibration schemes. Here we devise a time-evolving calibration method for the Plantower PMS5003 PM2.5 mass concentration measurements. We use 2 years of measurements from the Berkeley Environmental Air-quality and CO2 Network sensors deployed in San Francisco and Los Angeles in our analysis. The calibration uses a hygroscopic growth correction factor derived from k-Köhler Theory, where the calibration parameters are determined empirically using EPA AQS reference data at co-location sites during the period from 2021–2022. The parameters are found to vary cyclically through the seasons, and the seasonal cycles match changes in sulfate and elemental carbon PM composition fractions throughout the year. In both regions, the seasonal RH dependence calibration performs better than the uncalibrated data and data calibrated with the EPA’s national Plantower calibration algorithm. In the San Francisco Bay Area, the seasonal RH dependence calibration reduces the RMSE by ~40 % from the uncalibrated data and maintains a mean bias much smaller than the EPA National Calibration scheme (–0.90 vs –2.73 µg/m3). We also find that calibration parameters forecasted beyond those fit with the EPA reference data continue to outperform the uncalibrated data and EPA calibration data, enabling real-time application of the calibration scheme even in the absence of reference data. While the correction greatly improves the data accuracy, non-Gaussian distribution of the residuals suggests that other processes besides hygroscopic growth can be parameterized for future improvement of this calibration.
- Preprint
(1648 KB) - Metadata XML
-
Supplement
(284 KB) - BibTeX
- EndNote
Milan Y. Patel et al.
Status: open (extended)
-
CC1: 'Comment on egusphere-2023-1701', JAMES OUIMETTE, 10 Aug 2023
reply
Thank you for the interesting article. Could you please describe in your Methods section the source and time averaging for the RH data used in Figure 3 and your other analyses. Was it a reference RH instrument or was it the output from the PurpleAir Bosch Sensortec BME280?
Could you also describe the EPA PM2.5 reference instrument used for your PurpleAir comparison, such as MetOne BAM, Teledyne T640, GRIMM EDM Model 180, or TEOM?
Thanks.
Citation: https://doi.org/10.5194/egusphere-2023-1701-CC1 -
AC1: 'Reply on CC1', Milan Patel, 11 Aug 2023
reply
Thank you for your questions! These details will be added to the methods of the paper.
The data is averaged to hourly data points both for the PM2.5 and the RH measurements before any calibration or analysis is performed. The RH measurements are from an Adafruit BME280 sensor which we have located next to the Plantower PMS5003 in the BEACO2N node enclosure. All three of the EPA reference sites measured hourly PM with a Met One BAM-1020 Mass Monitor w/VSCC.
Citation: https://doi.org/10.5194/egusphere-2023-1701-AC1
-
AC1: 'Reply on CC1', Milan Patel, 11 Aug 2023
reply
-
CC2: 'Comment on egusphere-2023-1701', JAMES OUIMETTE, 15 Aug 2023
reply
It appears from your paper that you are attempting to develop a PM2.5 calibration scheme for the Plantower PMS5003 in the PurpleAir by assuming that it is a nephelometric sensor which measures the aerosol scattering coefficient. In this case an efficient and rigorous calibration scheme would therefore involve calculating how the PMS5003 mass scattering efficiency changes with relative humidity. A recent paper (reference below) shows that the PurpleAir does not act as a perfect nephelometer. As a result of its geometry and perpendicularly polarized laser it significantly truncates the scattering signal for particles as they grow hygroscopically. Instead of the PMS5003 PM2.5 growing exponentially at high RH like an unheated nephelometer would measure, the PMS5003 PM2.5 is muted significantly at high RH. While the PMS5003 PM2.5 is still greater at high RH than a heated low-RH regulatory PM2.5 instrument such as the BAM, the PMS5003 PM2.5 is significantly lower than one would calculate by ignoring the unique attributes of the PMS5003.
You may want to consider providing the readers a better physical understanding behind the statistics you present in your paper.
Thanks again for the excellent paper.
Ouimette, J. R., W. C. Malm, B. A. Schichtel, P. J. Sheridan, E. Andrews, J. A. Ogren, W. P. Arnott. 2022. Evaluating the PurpleAir monitor as an aerosol light scattering instrument. Atmospheric Measurement Techniques 15:655-676. doi.
Citation: https://doi.org/10.5194/egusphere-2023-1701-CC2 -
CC3: 'Comment on egusphere-2023-1701', JAMES OUIMETTE, 18 Aug 2023
reply
Could you please include a brief summary of the data you used - - the number of hours of the BAM, PMS5003, RH, etc. Also, please include the frequency distribution of those 1-hr average data values in your Main paper or the Supplement. This will help the readers decide if your results, which are limited to urban California, are extendable to their situation. Your paper would benefit from a discussion of how you handled the negative 1-hr ave BAM PM2.5 values. The figure below for a typical western US town shows that 12% of the 1-hr ave BAM PM2.5 are below zero, and that two-thirds of the values are at or below the EPA-published MDL for the BAM. How would you apply your Universal Hygroscopic Growth Calibration to their co-located PurpleAir data? Thanks.
Citation: https://doi.org/10.5194/egusphere-2023-1701-CC3 -
RC1: 'Comment on egusphere-2023-1701', Anonymous Referee #1, 25 Sep 2023
reply
This paper describes a new approach to calibrating low-cost PM sensors (specifically the Plantower PMS5003 nephelometer), based on a seasonal correction related to hygroscopic uptake. Sensors are colocated with more expensive PM instruments at air quality stations for 1-2 years, enabling the determination of seasonally-varying calibration coefficients using kappa-Kohler theory (eq 1). This is shown to provide a measurement of PM2.5 mass concentration that is improved over the “raw” Plantower output (from the factory calibration) as well as over an EPA-recommended RH correction.
The results of the study seem to be sound, and the topic – calibration of low-cost PM sensors – is certainly of interest to the readership of AMT (which has published many sensor-calibration papers in the past). However a major weakness of the paper is the lack of any provided context for the work – there is virtually no discussion of previous relevant studies, including those that use kappa-Kohler theory to correct for errors in PM mass, and little discussion of how this approach can be applied and what its limitations are. These represent major shortcomings of this work, and need to be addressed before this can be considered for publication
Specific comments are listed below.
1) Relationship to previous work. In describing their methods, and assessing results, the authors compare only against one other approach, the EPA Plantower correction (Barkjohn et al. 2020). But they make almost no mention of the many other studies in the literature that also correct mass concentration measurements by the Plantower (and other light-scattering-based low-cost PM sensors) for hygroscopic growth. There are at least four other studies that do this using the same general approach (kappa-Kohler theory) as the present paper: Di Antonio et al. 2018 (https://doi.org/10.3390/s18092790), Crilley et al. 2018 (https://doi.org/10.5194/amt-11-709-2018), Malings et al. 2020 (https://doi.org/10.1080/02786826.2019.1623863), and Hagan and Kroll 2021 (https://doi.org/10.5194/amt-13-6343-2020). None of these papers are discussed; only one is cited. By not discussing these previous studies, this text inaccurately represents the state of the art of the field, as well as the novelty of the present work.
The approaches used by these (and possibly other) studies need to be discussed and compared to that of the present study, and the parts of this study that are different/novel (e.g., the sinusoidal seasonal corrections based on colocation experiments) should be highlighted. Ideally the present results would be compared against results using the approaches from these other studies (in Figs 3-4), but I recognize this might be challenging.
2) Calibration factors unrelated to hygroscopic growth. Implicit in Equation 1 and the associated discussion (lines 63-74, 98-102) is that all errors in the “raw” Plantower output are from the neglect of hygroscopic growth (and its associated change to refractive index). However this is almost certainly not the case, as many other factors can contribute. These includes issues arising from differences in the aerosols that are used in the Plantower factory calibration and those measured in the atmosphere. Key properties that may differ include size distribution (number of modes, mode median diameter, mode width, fraction of particles outside the instrument size cutoffs), refractive index, and density. These are described and/or explored in detail in earlier papers (e.g., Malings et al. 2020, Hagan and Kroll 2021), and shouldn’t be simply ignored here. The authors eventually do show that other factors may be contributing to calibration error (lines 136-137), but given the previous work on the topic, these need to be mentioned much earlier in the paper. (It’s also possible that some of these potential errors are seasonally varying, so can be swept into the values of m or kappa, and therefore can be corrected for by eq. 1; this should be mentioned as well.)
I think Fig 1 and Fig 6 also provide some strong evidence that other non-water-uptake factors are at play. The sulfate and EC fractions are quite similar in the two cities (Bay Area and LA, respectively) – but the kappa and m values are not. (In fact, the LA site shows higher sulfate but a much lower fitted kappa). This would suggest that the fitted kappa and m values are not entirely driven by changes in hygroscopicity, and are controlled by other factors a well; this needs to be discussed.
An alternate explanation is that the other non-sulfate, non-EC components of the aerosol (organics, nitrate…) are controlling kappa at the two sites. But this is easy to test: one just needs to compare reconstructed kappa from the speciated measurements to the fitted kappa values. (This is worth doing regardless!)
3) Discussion of applicability and limitations of this approach. The paper describes this new approach nicely, and shows that it’s an improvement over the no-correction and EPA-correction cases. But there is almost no discussion of what conditions this approach can be used for, and what its limitations are.
For example, the approach is fundamentally based on the assumption that the regional aerosol (used for calibration) provides a good description of all the aerosols to be measured. But what happens if a sensor is measuring from some local sources? (For example, if a sensor is located near a factory, or a restaurant, etc…) Or what if there is a major air quality event, such as a wildfire? If such particles are substantially different in properties than the regional aerosol particles used for calibration – not only water uptake properties but also optical properties and size distribution – then the measurements could be subject to considerable error. This error would be systematic for a given aerosol type, so could lead to consistent over- or under-estimates of certain PM classes of interest. This PM sensor accuracy issue – that it depends critically on the representativeness of the calibration aerosol – is a well-known limitation of low-cost-PM sensors, but it isn’t mentioned here at all. It should be discussed, along with an examination of the sorts of errors that might arise for different aerosol types.
Another question related to applicability: just how much calibration is necessary? For example, for how long does the colocation calibration need to be done – six months, one year, or more than that? Also, over timescales of years, sensors can drift, and aerosol loadings and composition can change. Given that, for how long is the calibration accurate? It’s possible the authors don’t have enough data to address this at the time, but it should at least be mentioned as an important question and area for future research.
Other comments:
Line 51: Malings et al. 2020 applied a seasonal kappa-Kohler-based correction, so this study is not the first to do this.
In Fig 1 (and Fig 6), panels c and d are provided to show that compositional seasonal trends reasonably explain the seasonal trends in kappa. However this is done purely qualitatively only. How well do the values actually correlate? (While the overall seasonal behavior matches up reasonably well, the spikes in panel c do not seem to be reflected in panel a.) Better yet, as noted above, it would be useful to reconstruct kappa based on composition measurements, and compare this value to the fitted value of kappa.
Equation 1: this is a reasonably well-known equation, with some version applied to low-cost PM data in previous studies – it’s nearly the same as eq. 1 of Malings et al. 2020, and very similar to eq. 4 of Crilley et al. 2018 and eq. 7 of Di Antonio 2018. A proof of it in the SI is therefore unnecessary.
Fig 4 (and associated discussion): were there any smoke events, or other air quality events, during this time? Where do these fall on this plot?
Fig 4: RMSE and biases need units.
Lines 137-138: another useful test would be a comparison with a non-seasonal correction (i.e., a single average value of kappa and m from the entire colocation).
Fig 6a: y axis should be kappa, not k.
Citation: https://doi.org/10.5194/egusphere-2023-1701-RC1
Milan Y. Patel et al.
Data sets
Datasets Used in this Work Milan Y. Patel https://github.com/milan-y-patel/Plantower-Calibration-Paper
Milan Y. Patel et al.
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
217 | 97 | 20 | 334 | 29 | 5 | 6 |
- HTML: 217
- PDF: 97
- XML: 20
- Total: 334
- Supplement: 29
- BibTeX: 5
- EndNote: 6
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1