1) General comments

The study of Bernet et al. (2022) provides total ozone trends at three high latitude stations for the recent period 2000-2020. While a start of the ozone recovery has been recently observed at mid-latitudes and Antarctic (e.g., WMO2018, Godin-Beekmann et al., 2022), it was still not the case for Northern high-latitudes. In the recent study of Weber et aL (2022), zonal mean 60-90°N total ozone trend in March from satellite measurements is still observed non-significant. It is therefore highly scientifically relevant to give an up-to-date ground-based perspective on ozone trend in the Arctic. Furthermore, the use of a multiple regression model to explain as much as possible the ozone variability makes also the paper well in scope with the ACP objectives. The paper is also very clear and well structured.

I therefore recommend the publication of this study in ACP, after some comments and questions are addressed (mainly on the combined data sets / drifts with satellites / predictors).

2) Specific comments:

Section 2.4: Combined ground-based total ozone data

It’s a good idea to combine the ground-based data sets to have a more complete temporal coverage. But I have a few comments/questions:

• Before combining, it would be good to show that the different ground-based measurements do not show significant bias among them, inter-comparing data in temporal coincidences.
• You use one technique as a baseline then successively fill the temporal gaps by other techniques. But why not using all measurements, also when different techniques are measuring at the same time, and then make a daily mean? Of course, in the way it is now (SAOZ at sunset/sunrise, the other instruments at +/-2h around local noon), the instruments are not co-located in time to make the average. But as you use anyway SAOZ in the combined time-series together with the +/-2h around noon averages, without any correction, I guess it would be valuable to make the “total” daily mean averages? If there is a reason to choose your approach of filling gaps instead of taking everything available, could you add an explanation?

Section 3: Time series comparisons

Drifts after 2015/Fig.4: Because the drifts/jumps between GBcomb and the satellites are different depending on the stations, and similar at each station looking at different satellites, it looks like the drifts are in the GBcomb products (for Andoya and Ny-Alesund), which could be an issue for interpreting the trends obtained in Sect.5 (especially for Ny-Alesund). It could be worth to evaluate quantitatively the drifts, and discussed them with the trends. In Fig. 2, comparisons between Brewen and ERA-5 at Andoya show a kind of “jump” between 2015-2019, which is related in the discussion and figure to some disturbance observed in SL measurements (that are corrected, but apparently not perfectly since the jump is still observed). Could it be that this impacts the comparisons with satellites as well (Fig.4h)? In Fig.4h, GBcomb are used, but the Brewer measurements are 84% of the GBcomp for Andoya. Maybe the comparisons between Brewer and GUV (suggested in the comment above on Sect. 2.4), can confirm if the drif/jump is also observed with local measurements (and not due to e.g. grid resolution of ERA and Satellites)? Is there a way to adjust the correction that has been made in order to remove the drift/jump at Andoya? For Ny-Alesund: how do the SL measurements look like there?

However, when looking at anomalies Fig4e-f, it looks like the trends obtained from the satellites would not be so different than the ones from GBcomb. Maybe it would be interesting to calculate those satellite trends (e.g from 2005) at the location of the stations and see if they agree with GBcomp’s trend  starting in 2005?

It looks like SBUV and Era-5 show a similar drift in 2000-2005 at all sites, suggesting that the drift is due to SBUV and Era-5 and not to the GBcomp in that period. Maybe it is worth saying in Sec. 2. 6 which satellites are assimilated in Era-5? Is it mainly SBUB in this period /latitudes?

Section 4: Multiple linear regression

• Error covariance matrix: you give higher uncertainty to monthly means that show higher standard deviation of the measurements, so with higher variability within the month. Could this also lead to less weight given for events with higher variability (e.g. ozone loss due to VPSC would give a high ozone variability within the month), therefore this could lead to minimizing the impact of proxies (which is not desired)? How this error covariance matrix impacts the obtained R2, trends,...?
• Tested predictors: a local proxy that has been proven to have significant influence on the ozone variability at high latitudes stations is the equivalent latitude (See Vigouroux et al., 2015, Fig.4). It can take into account the O3 short term variability due to the fact that the station is in/out of the polar vortex. Did you try this (or alternatively the potential vorticity)?
• Final choice of predictors: it is decided to keep using T50 although it correlates with TropP (0.51), and to not use VPSC because it correlates to EHF (-0.33). It looks quite arbitrary to have these opposite decisions. The use of T50 despite the correlation is motivated by the increased R2 (from 0.91 to 0.96 at Oslo). It is said that the use of VPSC improved the fit residuals at Ny-Alesund, but the R2 improvement is not given: how much would the VPSC inclusion increase R2 at Ny-Alesund? If negligible, then it should be fine to remove it from the model. If not, then the correlation motivation does not seem in line with the choice made for the T50 inclusion.
• L-265-267: comment: I find interesting and consolidating your results that the same behavior is observed with FTIR total ozone (Vigouroux et al., 2015): EHF significant at the polar sites, but not at mid-latitudes. Solar, QBO, and ENSO also insignificant or very small at polar sites.
• 6 and 7: the solar cycle parameter is found negative. To my knowledge the impact of solar cycle on total ozone is expected to be positive (solar maximum = increase of O3; see e.g Weber et al. 2022,…). This is not what is found at Oslo for annual trend, and even more for March trend. I don’t understand the discussion on February trend at Oslo (l.313-319): the explanation seems to say that O3 observed maximum coincide with the maximum of solar cycle, but the parameter is found negative (Fig.7). Could it be that the MLR gives a wrong interpretation of solar cycle influence? How would be the February trend without the solar cycle included? Maybe more realistic (now it looks like an outlier in Fig.9). The proxy with few cycles included (less than 2 for 11-years solar cycle 2000-2020), can be “by chance” interpreted to have an influence while “physically” they don’t. It was also the case for short time-series in Vigouroux et al. 2015. There is no “scientific reason” that the impact of solar cycle would be so stronger in February, is there? So it would be good to give the trend obtained without this solar cycle. Especially, because the time-series starts at a solar maximum and ends at a minimum, the inclusion or not of this proxy will modify the final trends (if the parameter is large, as it is the case in Feb in Oslo).
• Lower R2 at Andoya and Ny-Alesund (l.282-284): could you try with VPSC and equivalent latitude included? From FTIR stations (Vigouroux et al. 2015), the R2 was larger at polar sites (including Ny-Alesund) compared to mid-latitudes because of this larger variability (less dominated by noise in agreement with what you discussed l.279-280). When the variability is well explained, R2 is more easily higher.

Section 5: Trend results:

- 5.1: You obtain slightly increasing trends with latitude. Could you give the obtained trends with SBUV (only satellite you use covering the 2000-2020 period) at the 3 sites to check if this is also observed by the satellite? Because of the observed drifts in Fig.4, it might be good to consolidate your results.

- 5.2: February in Oslo: see comments above on the solar cycle impact.

- 5.2: October positive trends: note that zonal means trends around 80°N in October were still found negative in Morgenstern et al. (2021). So interesting indeed that you don’t find this over Scandinavia. It would be interesting to see what is observed at other longitudes. You could add this reference to motivate your work on regional trend.

3) Minor or technical comments:

Abstract:

- l. 10-11: specify “positive annual trends at Andoya,…”; and give uncertainties

- l.11: “no significant annual trend at Oslo (0.1%/dec + uncertainty)

Introduction:

- l.30-31: I guess drifts can also be observed in ground-based measurements.

- l. 48: “in the best possible way’. Maybe too assertive. Can be moderated by “… and define a state-of-the-art set of predictors that explains the natural ozone variability at the three stations”; or something similar.

Section 2:

• the GI method is applied to Oslo and Andoya. Why not to Ny-Alesund ?
• you give the available months for GUV and SAOZ; maybe give them also in Sect. 2.1.1 (DS) and 2.1.2 (GI); or summarize all this in a Table instead, as you wish?
• you give the uncertainties of SAOZ measurements (l.112): give also the uncertainties of DS, GI, GUV (It could be in the same Table as for period of measurements suggested in the previous comment?)

Section 4:

• 259: “full” trend: I guess you mean “annual trend”? I would use annual, it is more common (same for title 5.1,…)
• To my opinion, it is easier for the reader to include Fig B1 and B2 in Fig. 7.

Section 5:

• 1: provide the uncertainties on the trends also in the text.
• 296-298: I would give the numbers in the two reported satellite studies to better compare with your results.
• 299-302: I would give the trends and uncertainties from Svendby et al., to help the reader to see quickly the improvements on uncertainties obtained using the MLR.

Conclusions:

• Typo l.358: “In conclusion,”
• 358: “the urgency”: too strong to my opinion. Maybe “the need”?

This is the first review of the manuscript “Total ozone trends at three northern high-latitude stations” submitted by Bernet et al. to the ACPD.

The paper is focused on the assessment of regional ozone trends as measured by ground-based instruments in the Arctic. Multiple publications state that the Arctic environment is rapidly changing due to the impacts of climate change.  The changes in the global circulation and stratospheric temperatures have been reported in the literature, however, the full extent of the processes contributing to regional ozone variability in the Arctic is not well established. Moreover, the impact of the halogens on the spring-time stratospheric ozone depletion is intermittent in the Arctic due to the typically high instability of the polar vortex that influences the formation of the Polar Stratospheric Clouds that must last into the spring (or transported to the lower latitudes) when sunlight can activate the rapid ozone destruction mechanism. The intensive ozone depletion events are still hard to predict; thus, the analyses of the past records are of great importance.

Several papers published in ACP in 2022 found that total column ozone trends derived from the combined satellite records are not the same across the Arctic. Therefore, it is important to compare these results with the trends from the long-term ground-based records that are known for their stability and regular quality assurance of observations.

The authors present the study of long-term records from three stations in Norway that collect observations in one section of the Arctic. They also represent different latitudes that might be impacted by different processes, including being inside or outside the polar vortex and downwind of different pollution sources.  The attribution of the long-term changes in the ozone column is performed statistically using multiple proxies that represent chemical and dynamical processes that impact stratospheric ozone. A thorough evaluation of the correlations between proxies and their respective ability to explain observed variability is needed to determine statistically significant trends. The annual and monthly trends are derived to address the regional and seasonal variability in total column ozone in the Arctic and to uncover the signals of the ozone recovery under the Montreal protocol guidance.

The paper is well written, the figures are used well to illustrate the main points of the discussion. However, there may be the need to have additional figures in the Supplemental material since there are several occasions when the analyses are mentioned but the figures or summary of the results (i.e. in the table format) are not available. The published literature is properly referenced. This paper can be published after the authors address the following comments.

Here are several comments

General comment: 

1)    It would be important to discuss the location of the stations with respect to the vortex position before and after it breaks. Are stations typically within or outside of the vortex during spring season observations?

2)    I did not find any discussion about the potential impact of the long-range pollution transport to the Arctic with respect to the total ozone variability. Is there any evidence for tropospheric ozone increases in the Arctic?

3)    It would be of interest to the reader to have information about the magnitude of the trends in the proxies, or at least if the trends are positive or negative.

Technical comments

Line 79. The authors refer to the SL abnormalities being “corrected therein”.  What does it mean? Are data corrected daily, monthly, or yearly? Did the authors have to correct the data for this paper? How large were the corrections and if any were applied with a step change? Please explain.

Line 82 (and Appendix A). It is not clear if the GI method is applied to the Brewer observations. I am guessing it is as the discussion is under section 2.1 Brewers. Was the GI processing applied to the Brewer data for the first time and you are presenting these new data in this paper? If this is true, please also spend more time discussing the validation of the Brewer GI data (it can be done in the appendix/supplemental material). I would move most of the discussion of the GI data (including Figure 2) to the Appendix.  Also, please include plots for other stations (Oslo?) to demonstrate the stability of the GI method in comparison to the Brewer DS measurements. 

Otherwise, please provide the reference to the paper that describes the application of the GI method for the Brewer data processing and demonstrates its validation. 

Line 85. Does the reference to the “Ground-based UV radiometers” include both Brewers and GUV?

Line 87. Figure 2 comments. Please add a plot of the difference between ZS and GI data and reduce the y axes range to see the details. To make it more clear, the data could be averaged into monthly means to reduce the daily noise.

Line 90. Please quantify “slightly” by providing %bias for ZS/ERA5 comparisons for the reader’s reference. In my opinion, the ERA 5 should not be used as truth to evaluate the quality of observations, but it can help to investigate sudden changes in the record that can be associated with the instrumental artifacts (or the changes in the assimilated data).

Lines 91-92. Please provide information about the cause of the degradation of the instrument

Line 118. Are secondary data corrected for biases from the primary data? Was a comparison between the primary and secondary records done for each station during the overlap periods? Alternatively, the references for the publications describing the comparisons can be provided.

Line 124. How do sunset and sunrise data compare to the noon observations when all three are available? Was the standard deviation of the mean used to screen the data?

Line 117. What are the dates for change in satellite record assimilation over the period of observations? 

Line 144. Please rewrite the sentence (i.e. “seasonal cycle in ground-based observations”).

Line 150. Was there a change in 2014 in the selection of datasets assimilated into ERA5?

Line 157. What is the distance between satellite overpasses and Ny-Ålesund station location in early spring?

Line 182. Should there be a reference to the previous publications that discussed the meaning of using detrended vs. not detrended proxies for trend analyses? 

Line 250. March makes sense for the co-linearity issue as this is the time for the vortex breakup and mixing of different air masses. However, results for September are not clear (unless there are tropical air intrusions to the high latitudes).  Please add a discussion here about the processes that the T50 tracer represents. Plotting residual for a model fit when including/excluding T50 proxy can potentially help with identifying years when it improves the fit.

Line 256. Please add an explanation of why you think it is “improved”. How much improvement did you find (use the R2 or explained variability) Also, the authors decided to keep the T50 proxy in the regression based on the improvement to one station fit. What is the “physical process” that underlines the preference for including  T50 vs VPSC proxy? 

The temperature and ozone are highly correlated in the stratosphere, and T50 has a seasonal cycle. Does T50 inclusion in the model reduce the seasonal terms?

Lines 260-264. Would other proxies have a larger contribution to the fit if T50 is not included?

Line 276. Would adding the “rejected” proxies improve the model fit for the summer months? The cumulative EFH is constant for May-August. What other proxy represents the dynamical variability of ozone in the summer months? Was the contribution of tropospheric ozone variability considered for explaining the increased noise in total column ozone in summer? Would the model fit improve if the seasonally (three or 4 months) averaged data are used instead of monthly records?

Line 280. Please explain what processes might be contributing to the noise. 

Line 283. I cannot see the higher variability in the provided plots. Please add information about the standard deviation to the plots or the Table.

Line 291. Instead of using “good” please provide the SE of the detrended data in %. This can help the reader with an understanding of the improvements in the model performance.

Line 305. I recall that the LOTUS model allows using of the entire dataset to retrieve the seasonal trends by adding the seasonal components to the proxies. This approach can reduce the uncertainties in the derived trends by including information from other seasons. Was this technique considered for this paper? Was the measurement error covariance matrix used to analyze the observed records? 

Lines 305-314. There seems to be a repeat of the information. Please consider rewriting to outline the results for all stations based on one season at a time. It might be useful to keep in mind the processes that affect ozone variability in spring vs. other seasons.

Line 318-319. The authors report that the solar cycle (SC) proxy is significant for model fit in February only. Some papers (i.e. Labitzke and Loon, 1988)  found positive correlations between polar stratospheric temperatures or geopotential height data and SC when QBO was in its western phase, and the opposite was found for the eastern phase of the QBO.  So, some of the SC attributions to ozone variability are probably represented by other proxies (i.e. T50 and TP) that you are using for model fit. Still, the contribution to ozone variability in February only is hard to justify because the same process should affect ozone variability in other months. Could the method of combining records for February in Oslo be impacting the time series? What is the percentage of DS vs GI observation is used in February for the combined record? Doe this percentage change from year to year?  Another possibility is that T50 or TP proxies are not representative of atmospheric variability in February in Oslo. Or there is a  sampling bias of ozone records in February.

Line 324 Please re-write this sentence (they mean “polar regions”?). Please provide the reference to the publications that discussed pre-2000 trends in the Arctic region.

Lines 330-332. Yes, it is remarkable that 3 stations have very similar trends in March. Was the PSC proxy tested to improve the trend model fit for analyzing data in March in an attempt to reduce the uncertainty? In section 4.3 it is not clear if the selection of final proxies was done based on the model fit to the entire dataset (an annual record) or if it was also assessed by using the fit to the monthly records.

Line 352. Please define “high” (for example, better than 0.9).

Line 365. Please replace the NDACC link with www.ndacc.org The “demo” link will be going away soon, but the “ndacc” link will be preserved for a long time.

Comments for Figures:

Figure 8. The trend results are provided in the Figure in % and DU, but Standard Errors (SD_res) are provided only in DU while the plots are using % in y-axes. It might also help the reader if all results can be also summarized in the Table.

Appendix A. 

General Comment. Thank you for using the color scheme in the plots that are friendly for vision-impaired people.

I would also consider reducing the range of X-axes in all panels (except for T50) in Figures 8, B1 and B2. Some boxes are impossible to discern. The range for the T50 panel can be different and the note can be added to the Figure caption.

A2. Since GI data are corrected for the SZA bias, it is possible that Brewer data also has a bias (i.e .in Oslo where Brewer is a single spectrometer). The change in the data processing is noted from the 4-wavelength to one wavelength ratio method at large SZAs. However, there is still a need to calibrate data by using the DS measurement. What is the range of the SZAs for the GI/DS comparisons? 

A3. It might be useful to show a plot (for both stations) where the GI/DS total ozone ratio is plotted as a function of the SZA before and after each correction (SZA and cloud correction).