Review

Spectral Analysis Approach for Assessing Accuracy of a Low-Cost Air Quality Sensor Network Data

The study aims to investigate the accuracy and reliability of data from low-cost air quality sensor networks, which have emerged as a promising tool for high spatio-temporal monitoring of air quality. The authors employed a frequency analysis to identify the contributions of individual periodic sources to local air quality in Chicago using a two-year data record of PM2.5 from both the EPA and Purple Air networks. Their findings highlight the source dependence of low-cost sensor network measurements and emphasize the need for exceptional care in the analysis of ambient data from these networks, particularly when used to evaluate and implement air quality policies.

The manuscript is well-written and clearly structured. Furthermore, the spectral analysis method used to evaluate the data is novel in the context of low-cost sensors. To criticize, the study does not yield scientifically significant new findings. The main conclusion is that the sensor response is source dependent and that without proper calibration, there is a high risk of data misinterpretation. This is the same conclusion that has been made in most, if not all, studies investigating low-cost sensors.

I recommend publication of this study because I consider the approach used to evaluate sensor data valuable. Furthermore, I encourage the authors to consider the following points to strengthen the impact of the research.

1. The manuscript lacks a clear statement of the limitations of the study. This would be useful for readers to contextualize the findings.
2. The authors suggest that the results of their analysis will provide guidance in devising new approaches to calibrate data from low-cost sensors, but it is unclear what specific recommendations are being made. A more explicit discussion of the implications of the study's findings for future research and policy decisions would have strengthened the overall impact of the article.
3. Line 89 foe;d typo?
4. 1 Consider adding a scale for the map and units for the population density.
5. Local correction model; justify the use of both temperature and relative humidity in multiple linear regression. These variables are correlated with each other which can be problematic as the independent variables in MLR should be independent.
6. Line 294 “Our analysis clearly demonstrates for the first time that the PA network’s very different sensitivity to different sources.” I suggest you remove the “first time” part here.

This paper uses spectral analysis to assess the accuracy of low cost air quality sensor networks. This approach is sensible because as the authors note “Sources with short time periods, relative to the calibration period, are averaged out and inadequately accounted for in the calibration. Thus long time scale events are completely lost in the calibration process.” The new calibration approach is interesting and appears to have potential.

The paper highlights that both the regulatory and low cost networks show distinct peaks in the power spectrum at 4, 8, 12 and 24 hours. It is stated that the 24 hour peak “likely represents harmonics of the 8 hour and 12 hour signals, and hence represents a combination of both (8 and 12 hours) sources”. Isn’t it more likely that the same sources occur at the same times daily, i.e. morning rush hour, evening rush hour, etc. all start at the same time each day and so you would expect 24 hours to turn up in the power spectrum very strongly?  The 8 hour peak being traffic seems very plausible and the analysis on the days of the week is good evidence due to the preponderance of 8 hour working days. The argument put forward for the 12 hour peak would seem to make more sense to me to be at 24 hour peak, i.e. peak sunlight follows a 24 hour period not a 12 hour period.  Could the 12 hour period be representative of 12 hour work shifts, in addition to the 8 hour shifts? Whereas 8 hour shifts represents a standardised “9-5 working day through Mon-Fri”, a 12 hour shift is more likely for shift work where workers have a certain number of days on and a certain number of days off, which would not show so significantly in the day of the week effect? The peaks at 4 and 6 hours are also intriguing, what could be causing these?

The lack of concrete understanding of the periods should be stated more clearly in the conclusions and abstract. The results are interesting, but far from being definitive as yet.

With the rationale of calibration periods missing short term signals through the averaging out of signals, it is unclearly why outliers (L100) were removed.  These signals could very possible be true, for example due to construction dust etc.

It’s not clear why a linear RH correction is used (L118) when the RH effect has been shown to be due particle hygroscopicity, which is non-linear with respect to RH.  The k-Kohler approximation well in many low cost sensors, e.g. Crilley et al. 2018 - https://amt.copernicus.org/articles/11/709/2018/ The lack of diurnal RH effect in the correction might explain some of the difference seen in the power spectra of the EPA and low cost data.

The authors should provide some rationale why they think the low cost sensors are missing the high frequency outputs. For example, why would the low cost sensors be blind to traffic data for example?  Are the low cost sensors not measuring the smallest particles that the EPA sites are measuring?  If so, then the low cost measurements are blind to a subset of particles. What is the implication of this? 

Minor comments

L23-24 the literature around links between air pollution and COVID-19 outcomes are still contested. I would insert a ‘maybe’ within “…exposure to PM2.5 [might] also impact responses to acute diseases such as COVID-19”.

L29 – define FRM and FEM and indicate how they differ.

L44 “Using two sensors… allows for the robustness of data collection”. Spell out why two sensors improve robustness.

L80 provide some rationale why Chicago PM2.5 have near doubled from 2017 to 2019.

L89 typo ‘foe;d’

L110 the thermal conditioning within EPA measurements can impact upon particle mass through the removal of volatiles and this should be highlighted.

L162 define ‘stationarity’

Figure 1 – provide units on population density

Figure 2 – define the coloured lines on top of the E2 and P6 time plots.