Preprint review of amt-2024-130

General Comments

This work demonstrates an approach to calibration of a multi-gas sensor (NO, NO2, CO, O3) across three distinct locations, and by varying two driving factors (averaging time and concentration range) in improving calibration coefficients.  There is also some discussion on a technique that purports to improve sensitivity and precision by a process that apparently removes water vapor, which is a known interfering compound for many of these electrochemical approaches.  This work was performed at three different locations which were described by authors as being significantly different in climatology and in their pollution mixtures.

The work is generally sound, but viewed as somewhat incremental.  It appears there are two distinct methods combined in this manuscript: the use of a water-removing improvement for sensitivity, and a simple linear modelling approach to assess adequacy of calibrations.  Aside from using the same instrument platform, it is unclear how these two are closely related.

There is a significant amount of this manuscript that discusses the employment of their ‘dynamic baseline tracking method’ to reduce water interference.  However, aside from some clearly improved RMSE and R2 values when this approach was used, very little data are presented.  Furthermore, this is approach simply improves sensitivity of an existing measurement method, which by definition, is incremental. This may indeed be a significant technological advancement, but as presented, the conclusions are not adequately supported by the data.

More interesting are the results of their statistical methods for calibrating across different time domains.  The general conclusion, that collocations should be on the order of 5-7 days, is generally coherent with other findings, and is additional evidence that supports a broader convergence of calibration approaches in the lower cost sensor paradigm.  The data are not entirely persuasive, in part because there was limited discussion (aside from the introduction) in how these approaches varied across diverse airsheds of Macau, Hong Kong, and Shanghai.  Assessing the dynamic range of concentrations needed is certainly important, but so is assessing sensor averaging time performance in the presence (or absence) of co-pollutants that vary across space. It would seem reasonable that these are important interferents that may affect sensor performance.  This work would be much stronger if separate analysis for different environmental conditions were presented rather than combined together – if the findings (of 5-7 day collocations) were robust across different airsheds, this would be a very important finding.  I would assume that this is more nuanced, and one might find significant averaging time differences in locations with substantially different composition, just as the authors found with concentration loadings.  But, unfortunately, one cannot gain this insight from the work as presented.

Specific Comments

L14, 77, 138.  The authors routinely refer to this technology as ‘patented’ but it is not clear what the purpose of this statement is.

L35-45: much of this is well established science, and could be reduced with appropriate referencing.

L93-95: This is relatively well known methodology with Alphasense sensors; this could be refined and reduced.

L139: the authors refer to the approach as only allowing ‘water molecules’ to pass through a filter.  Are all other gases excluded?  This would be a very unusual method to assess water vapor interference.

L171: The authors use a Python function to compute random numbers, but then note that this is to ‘simulate real world sensor calibrations practices, and ensure randomness…’  While this reviewer would certainly agree that random.choice() indeed chooses data randomly, it does not simulate anything.

L180: What does ‘with superior validation performance..’ mean?

L188: What do the authors mean by ‘showcasing’?  Reference monitoring sites are not normally a showcase, but focus on high quality empirical measurements.  Consider revising this language.

Figure 2: The conceptual diagram does not add much to this paper, unless the focus of the paper were on method development for water interference signal removal.

Line 304:What do the authors mean by ‘as determined in the just-obtained results’’?

Line 369:  This sentence does not make sense.  Isn’t this always a plausible explanation for the failure of calibration models?

Editorial/Minor comments

L66: typo on ‘more easily to be standardized’, and needs clarification.

While there are few specific editorial comments to address, the manuscript has a substantial amount of indirect language, including many unnecessary linguistic flourishes.  The writing is far too verbose, and makes the work laborious to read.  There are periods in which a number of sentences begin with unneeded adverbs (e.g. Line 20, 99, 307 all start with ‘moreover; L 304, 397 begin with ‘notably’, L47, 283, 306 all begin with “consequently’).  In fact, in the paragraph that begins at line 304, every sentence begins with this unneeded language (Additionally, Notably, Consequently, Moreover, Overall, However).

This study is a thorough analysis of several factors influencing the calibration of low-cost sensors: concentration range, calibration duration, and time averaging, and provides recommendations for each. The study utilizes the MAS sensors, which have a built-in in ‘dynamic baseline tracking’ feature that promises to eliminate (reduce?) the effects of environmental conditions such as temperature and humidity.

The dynamic baseline tracking method described in section 2.2 is fascinating, and more detail in the explanation would help the reader understand it better. How does the PDF work, and can any information be shared on its accuracy at filtering out gases? Is it more or less accurate for any specific gases? Figure 2 is helpful for understanding this, but in the upper right plot, is there really zero difference between ORG and PDF in the lab? If not, a similar figure showing real data from the lab is important for readers understand how perfect or imperfect the method is, even if in the supplemental. It is not clear from the figure what “laboratory conditions” in the right panel means – were temperature, pressure, or humidity held constant, or were all fluctuating?

The authors later state that, “the influence of temperature and RH on sensor signals has been eliminated”. Can you prove to the reader with real data that this is entirely eliminated, or to a certain extent eliminated? In theory, Figure 3 could help answer this for the ambient data, but it is hard to read. Figure 3 panel F is the only one I can somewhat make out the difference between solid and dotted lines for. Moving the black reference data to the back of these plots might help make the other colors and lines more visible, but additional edits might be necessary for readability.

The analyses of concentration range, calibration period, and time averaging are thoughtful and well-explained. The R2 and RMSE of the validation data are reported. Would these results change if you applied the best model from one location and applied it to another location? As the authors state, it is important to replicate the conditions of the deployment to the best extent possible during the colocation, but lack of availability of reference instruments in certain locations can make this difficult. The analysis shown here could be made more practical by showing examples of how these trends may deviate as low-cost sensor users frequently have to adhere to non-ideal constraints. The recommendations provided for each of these are well thought out for the best case scenario of being able to co-locate exactly where the deployment will take place, but I am left wondering if these recommendations would still apply in a real-world scenario.

The limitations and future works could be expanded into their own section instead of lumped into the conclusion, since these topics have not been discussed earlier in the paper. Acknowledgement of the practicalities of variation in sensor co-location vs deployment locations could also be expanded upon here.