This study developed and demonstrated a point sampling method to automatically measure emissions from a large-scale of individual vehicles. In this works, the authors present their system that can be used for particulate matter (PM) and gas emissions measurements, which is notably independent of vehicle type. They find that when using their peak detection algorithm (TUG-PDA), they can separate vehicle-specific emissions down to a spacing of just a few seconds between vehicles. In this study, they present initial findings from the use of this method that collected ~100,000 vehicle records from several measurement locations, mainly in urban areas. Their findings include a detailed evaluation of the main influencing factors on point sampling measurements, specifically for carbon dioxide (CO2) and black carbon (BC) measurements. When compared to equivalent remote sensing measurements, the authors found good agreement even with the newest standards which are harder to capture due to their lower emissions and the current remote sensing abilities. This paper is well written and organized. However, the novelty of such work needs to be explored further. While the authors specify that this point sampling method is novel and/or surpasses the ability of many other (cited) studies on roadside emission measurement, this reviewer questions that assumption with the only notable differences coming in the use on light duty vehicles and the automation. It needs to be clear how this work contributes to the scientific knowledge on vehicle emissions measurements. Further exploration of the thousands of measurements made could help to enhance the novelty of this work by highlighting new or potential trends in vehicular emission such as emission control technology deterioration as mentioned in the introduction of this study. Additional findings, revisions and additional review would need to be completed prior to acceptance and publication. 

Additional comments:

Introduction, L 36.

“…by making wrong measurements.”

Can this be further explained or cited?

Methods, 2.1 Measurement Setup, L93.

“Using light barriers restricts the measurement location to single-lane roads or roads with islands between the lanes. Alternatively, vehicle detection can be performed with radar, video, or LiDAR systems.”

What does this limitation have on the type of vehicles able to be measured with this system?

Do the other sampling options for vehicle tracking listed have the same capabilities but better capture for more road types. If so, why were they not used? Please sure explain the impacts of this sampling method especially with regards to vehicle population and potential bias.

Methods, 2.1 Measurement Setup, L110.

“In general, the closer the sample inlet is to the emission source (tailpipe) the smaller the dilution and the higher the capture rate are.”

Also discussed in 3.2.1 Sampling Position and eventually mentioned on L463.

Though true, in the schematic, diagram, and later sections, the sampling inlet is located near the ground, what was the capture rate for vertically oriented tailpipes? How did that influence your sample population? Is there potential for this method to be adjusted to capture all tailpipe orientations?

Methods, 2.2.2 Emission event processing, L187.

“At the same time, care must be taken to ensure that the CO2 plume detected of the passing vehicle is related to the pollutant emission detected. Therefore, checks are implemented which compare the duration of the integrated CO2 and pollutant data and verify if the areas overlap appropriately.”

Generally, the procedure of peak identification and peak alignment with passing vehicles needs to be further clarified. Can you explain more on how you know that CO2 has returned to baseline to meet the conditions outlined? Does the end of the pollutant peak only rely on another vehicle pass being detected? If so, what does it mean with regards to truly capturing the extent of a CO2 peak? Other works cited have specific quantitative assumptions for the rise above baseline for pollutants and CO2 as well as for the return to baseline after a vehicle passes, can more quantitative information like this be provided?

Conclusion, L559.

“The core of this software is the TUG-PDA, which determines and separates vehicle emissions down to a distance of 3 s between the vehicles, if appropriate instruments are used.”

Does this software have potential to be adapted fit the sampling behaviors of a range of instrumentation? The following bullet point is helpful to understand instrument requirements but if others were to try to adapt or replicate this work, is it fully instrument limited? Thinking back to the example provided comparing the AE33 and the BCK. How could one use the AE33 with this method/software? If this is outside the scope of this work, that is fine but please acknowledge.

General Comments:

This manuscript describes a sampling system that can be used to capture and evaluate in-use emissions by thousands of vehicles. The authors include detailed data that quantify the impact of important sampling location and configuration features and environmental conditions on the success rate of their point sampling method, and a first look at the emission trends that have been captured by this system. Such an automated platform for capturing on-road vehicle emissions and determining emission factors would be extremely useful for both regulators and researchers for tracking fleet trends and identifying high emitters.

The key innovation presented in this paper is their automated peak detection algorithm. However, not enough information is given for a reader to replicate this methodology independently, and this reviewer also has several questions about how the algorithm functions (see below). If the algorithm will be made publicly available and if the below questions/comments are addressed, then I believe that the manuscript could pair well with it. But if the algorithm will not enter the public domain, then I think the manuscript requires major revision to be a more complete methods paper that could be independently duplicated by others. Alternatively, if the authors do not intend for this paper to be a methodology paper, then less focus should be spent on the results for their sampling platform/algorithm performance and more should be spent on the vehicle emissions that were sampled. They have a rich dataset for tens of thousands of in-use vehicles at multiple locations in Europe, which could easily be the focus of the paper. The manuscript in its current form seems split between the two narratives and incomplete for both, and it would be more compelling and impactful to focus on either the method or the fleet results. 

Additional general comments:

• Many acronyms are introduced but only used a couple of times (e.g., PC, PTI). This can be confusing for the reader, so this reviewer suggests only using those acronyms that are frequently used (e.g., HDVs, PS, etc.) and minimizing the introduction of others.
• I did not think enough information was given about many important details for the algorithm and have many specific questions that are listed below. Overall, I have questions about:
  • How the background concentrations are determined and applied
  • Peak separation and overlap
  • QA/QC steps and what is considered successful versus what is screened out of the analysis presented in Section 3.3
• This manuscript is well written, but is long and could be significantly shortened. In its current form, it’s difficult for the reader to pull out the key results and insights. These are well summarized in the Conclusions, but they’re otherwise not obvious with the current density of results, discussion, and figures. In many cases, results could be summarized more concisely with simpler statements like “results were comparable across all conditions” and the supporting figures could be moved to the Appendix, rather than describing them each in detail. In other cases, results and figures could be combined and presented together, rather than discussed in detail separately. For example, the results and discussion about sampling position and measurement location could be combined and streamlined to more directly and efficiently conclude that the capture rate is higher when you sample closer to the emission source.

Specific Comments:

• Introduction, Lines 20–24: Emission control system performance decline via tampering or malfunction is emphasized as the only source of high emissions of NO_x and PM, whereas these high emissions can also simply come from older engines without these newer after-treatment controls (e.g., non-DPF-equipped vehicles). In other words, skewed fleet emissions and the high emitter problem are not solely due to degrading DPFs or SCR systems that have been tampered with or are failing.
• Introduction, Lines 28 and 33: Is the interest really in PN concentrations (which vary with dilution) or emission rates?
• Methods, Figure 1 and Lines 108–109: How do you functionally position the sample inlet in the middle of the road where cars are driving? I’m assuming you put some sort of rigid protector over the tubing that vehicles drive over. But does that alter how they are operating (e.g., slow down while passing by, etc.), or is it small/inconspicuous enough that vehicles do not “see” it and do not change their driving patterns.
• Methods, Table 1: For vehicles without SCR or inactive SCR systems, NO_x concentrations can be even more elevated, up to ~10 ppm.
• Why is the algorithm called TUG-PDA? PDA is defined as peak detection algorithm, but TUG is not defined.
• Is TUG-PDA deployed in real-time, or is it used after data is collected? In other words, are peaks being detected and integrated live, or is this used as a post-processing step after data has been collected?
• Methods, Lines 180–181: If you smooth the data, does that not affect the peak area for the emission factor calculation?
• Methods, Lines 195–207:
  • I found the discussion and figures in Appendix C to be very helpful for better understanding how TUG-PDA operates, and suggest moving them (or a streamlined version of them) up to the main manuscript.
  • A 3-second delay seems really small between vehicles, when most peak events occur over 5–10 seconds. The only concern listed was misattribution of the captured pollutant peaks to the incorrect vehicle, rather than overlapping peak events. You detail peak separation in Appendix C, but not enough information is given here. How can your algorithm distinguish between vehicles if the CO₂ and pollutant concentrations do not return to background before starting a new peak integration? It’s unclear to me how you can get accurate integrations when peaks overlap with vehicles passing in rapid succession. Even if you assume a background concentration, the tails of the peak itself will be cut off. How are you certain that you have an accurate emission ratio under these circumstances? Have you characterized how different fractions of peaks captured and assumed baseline concentrations impact resulting emission factors, using the subset of peak events that were 100% isolated with all pollutants starting and ending at the background condition?
  • It’s unclear to me how TUG-PDA defines the start and stop time of peak events. Do the CO₂ start/stop times define the peak event, and then those are mapped onto each pollutant time series with the previously determined time adjustments due to different instrument responses? Or are the CO₂ and pollutant peaks handled independently by the algorithm, with CO₂ first for successful plume capture and then each pollutant if the CO₂ peak analysis was successful?
  • A background concentration based on the minimum value before the passing time is likely biased low, which would overestimate the true integrated area of each pollutant. For instance, background BC concentrations might bounce around –2 to 4 µg m^-3 on a secondly basis. A running average shortly before the start of the peak would more accurately capture the true baseline ~0–1 µg m^-3, rather than assuming a value of –2 µg m^-3, if that was the minimum concentration before the passing time. Similarly, for CO₂ concentrations under high traffic conditions, the background concentrations can vary by ± 50 ppm. The choice in background value can have large impacts on the resulting emission factor, and these questions need to be better addressed in the manuscript, especially when considering the limit of detection for this system when calculating near-zero emission rates with low emission events (i.e., good DPF and SCR performance) for those plumes with weak capture (i.e., small CO₂ peak area) events.
  • In Figure 4, you plot only positive values for BC, even though it looks like concentrations dip below zero for the background values before the peak events. Is this just a formatting choice for plotting this example, or does this mean that your algorithm ignores negative values? In this example, it doesn’t seem to matter for the peak integration, since there is a strong BC signal. But for the case where there is no BC peak (or other pollutant) that corresponds to a CO₂ peak, those near-zero concentrations can be positive or negative. The negative values are valid and should be included in an emission factor calculation.
  • How were the thresholds for positive concentration gradient (Line 199) and minimum CO₂ integrated peak area (Line 214) determined? Are these also dependent on sampling configuration, driving and/or engine load conditions, environmental conditions, etc.?
  • Line 216–217: If instruments have different response times, the pollutant peak could extend beyond the CO₂ If you’ve smoothed the data (Lines 180–181) to force this scenario to not happen, how have you verified that this does not affect the corresponding emission ratio? You extensively discuss time alignment in Appendix E, but not in terms of this question.
  • Lines 271–272: Do you determine if plumes can be separated and assigned clearly to a specific vehicle algorithmically via rules in TUG-PDA or with visual/manual inspection of TUG-PDA results? How do you QA/QC the data to verify that only valid emission factors that can be fully attributed to individual vehicles are included in the final dataset?
• Results: I suggest combining and streamlining sections 3.2.1 and 3.2.2, as the results and discussion are presented together, can be a little difficult to tease apart as they are currently discussed, and the existing text can be a little repetitive. I think choice of measurement location as described in Lines 314–320 is probably the most important factor in terms of successful point sampling, and it is best to describe those characteristics first. The sampling position details at a given location are more nuanced, and could be combined to better complement each other after establishing what a good sampling location requires in terms of road properties, traffic conditions, and vehicle operation. For instance, the discussion on Lines 354–361 about road width are very similar to the discussion in Section 3.2.1 about tailpipe and sampling direction and sampling heights.
• Results: Consider combining Figures 6, 7, and 8b to be side-by-side, since the three are very similar and discussed together in Lines 305–312. It’s difficult to synthesize all of the information presented in the current form while flipping back and forth between pages and plots.
• Results, Figure 6: The trend line seems like it might be showing the combined influence of sampling position and height. In particular, all of the middle sampling points occur at near-ground sample heights with high capture rates, compared to the left and right sampling results that span higher sampling heights and a broad range of capture rates. The combination of sampling position and height might be confounding this result/trend. What would these results look like if the left and right sampling configurations were also conducted at heights < 1cm, like the middle sampling results? How does the trend line shift if the middle results are excluded and only the >4 cm samples are included?
• Results, Lines 325–326: If measurements are often made after a crossroad or traffic light, could there be a bias in the emission profiles observed? How does that driving mode compare to “typical” operation?
• Results, Lines 382–383: Is the difference in capture rate noted for dry vs rainy conditions statistically significant? If not, I would suggest a slight re-wording of this paragraph that instead emphasizes that all of the results are comparable (like you do with the CO₂ and BC results), rather than pointing out minor differences in capture rate. Also, in this paragraph, you describe differences in average values for capture rate and CO₂ concentration, but median differences for BC emission ratios. Is there a reason for not reporting mean differences in BC emission ratios?
• Results, Figure 10b: Can you clarify what is meant by the y-axis label of “mean BC ratio”? I assumed that these are distributions of measured emission ratios from individual peak events, but please describe what has been averaged if they are instead distributions of mean ratios.
• Results, Line 477–478: If the emissions from previously passing vehicles are interfering with the measurement of the current vehicle, shouldn’t your algorithm and QA/QC process screen those results out as an unsuccessful capture? If there is interference from other vehicles, then you do not have an accurate measure of an individual plume that can be attributed to the target vehicle and it should not be included in your results. Or, you can consider fleet trend results from combined plumes like in Dallmann et al (2011), but not attribute any vehicle-specific information from license plate data to those emission factors.
• Appendix A, Line 620: “When plumes overlap or impacts from other sources occur, this concentration may be underestimated.” This is an important point that I think should be emphasized in the main manuscript when describing how your algorithm handles vehicles that pass by in rapid succession, especially if you do not exclude them from your results.
• Appendix C, Figure C1:
  • I’m confused by the shaded areas for CO₂ that extend all the way down to ~400 ppm CO2. From the time series, the background concentration looks to be more ~450 ppm, depending on the passing vehicle. Is your algorithm assuming a background concentration of ~400 ppm for all four vehicles? If so, this is an overstatement of the CO₂ peak areas. If not, and this is just a figure formatting choice, then I suggest instead either: (1) plotting background subtracted concentrations, (2) adjusting the secondary y-axis range so that the plot looks more like Figure 4, or (3) cutting off the shaded blue areas to only include the above-background portions of the peaks to represent the true peak areas included in the emission ratio calculations.
  • Does the weak capture for V2 pass the minimum requirements for calculating an emission factor? That rise in CO₂ (~10 ppm) does not look strong enough above normal noise in background concentrations to be a successful capture. Would this be flagged in QA/QC?
  • A4 is not shaded in or labeled, as noted in Lines 668–669.

Technical Corrections:

• This might be a journal formatting requirement/preference, but it is sometimes hard to discern paragraph breaks without extra spaces between paragraphs or an indent at the start of a paragraph.
• Figure placement throughout the manuscript doesn’t always make sense. For instance, Figure 12 is discussed in Section 3.2.3 but appears halfway through Section 3.2.1, while Figure 13 is discussed in Section 3.2.3 but appears on the next page in the middle of Section 3.3. I realize this may be a journal formatting issue rather than one that the authors can address, but wanted to point it out in case it could be adjusted.
• Suggest replacing occurrences where “1000s” was used with the word “thousands” (e.g., Abstract Line 9 and Intro Line 75)
• Abstract, Line 13: define NO_x as “nitrogen oxides (NO_x)”
• Introduction, Line 19: define NO_x as “nitrogen oxides (NO_x)”
• Introduction, Line 23: suggest “malfunctioning”
• Introduction, Line 44: define emission factors as “(EFs)” since that is how it is used throughout the paper
• Results, Line 333: revise to “6,500” or “6500” depending on number format used throughout the manuscript; note that there are some inconsistencies throughout the manuscript that should be made uniform (e.g., “3000” on Line 334 versus “100,000” on Line 275)
• Line 480: should be “…150 mg (kg fuel)^-1”
• Line 587: consider word choice substitution for harsh, maybe something like “challenging”