The authors present a novel study using Pandora MAX-DOAS measurements and the HeiPro retrieval algorithm to produce a three-year dataset of NO2 profiles and partial columns in Toronto, Canada. While the paper provides valuable insights into the spatial and temporal distribution of NO2. However, several areas require clarification, reorganization, and deeper analysis to strengthen the validity and comprehensibility of the study. My detailed comments are below.

General Comments

Stratospheric-Tropospheric Separation (STS) Method

• The authors use a complicated approach involving a box model and OMI observations to separate stratospheric and tropospheric columns for Pandora direct-sun (DS) measurements. The STS method is questionable due to multiple layers of assumptions and models, creating uncertainties.
• Why not employ stratospheric-tropospheric column ratios from established models, such as CAMS, TM5, or GEM-MACH, to simplify and enhance the accuracy?
• A direct comparison of Pandora-DS total columns with TROPOMI total columns would provide additional insights into discrepancies between ground-based and satellite observations.

MAX-DOAS Retrieval and Atmospheric Profiles

• The description of the MAX-DOAS profile retrieval lacks essential details:
• Are the ERA5 atmospheric profiles daily averages or spatio-temporally interpolated to the measurement times?
• What assumptions are made about NO2 above the retrieval height? Are these values based on standard atmospheric profiles or other sources?

Data Consistency and Filtering

• The differences in coinciding data points in Figure 3(a) (direct sun) and Figure 3(c) (model comparisons) require clarification:
• If Pandora DS and MAX-DOAS data originate from the same source, why is there a discrepancy in the data points?
• Did the authors apply data filtering, such as cloud filtering, before using Pandora data for MAX-DOAS retrievals?
• Similarly, in Figure 4, the differences between MAX-DOAS (HeiPro) results in panels (a) vs. (c) and (b) vs. (d) need to be explained.

Organization and Logical Flow

• Figure 1 is currently located in the introduction but would fit better in Section 2, "Instrument Description," to align with the discussion of the Pandora instrument and measurement conditions.
• Section 2 should follow a logical sequence: instrument description, retrieval algorithm, and then the models/data used for comparison and validation. This reorganization would enhance the clarity of the methods section.

Wavelength Range and Spectral Retrieval

• The wavelength ranges for direct-sun and MAX-DOAS spectral fits are unclear:
• Did the authors use the same wavelength range for both retrievals? If not, provide justification for the differences.
• Details of the spectral retrieval process should be included.
• Line 217-218 mentions NO2 retrieval in both UV and VIS bands—did the authors compare results from these bands? Specify which band was used for comparisons with direct sun, satellite, and model data.

Quality Filtering and Averaging Kernels

• Section 2.2.2 discusses filtering criteria (e.g., DOFS < 1), but the rationale behind these criteria is not well explained:
• Provide theoretical or empirical justification for the chosen threshold.
• What is the typical DOFS value in the retrieval? Including an averaging kernel plot would help visualize the retrieval sensitivity.
• Demonstrating the impact of data filtering on the results would also improve transparency.

Comparison of Data Sources

• In the results section, the biases between HeiPro, Pandora-DS, TROPOMI, and GEM-MACH are analyzed, but the spatial, temporal, and observational characteristics of these datasets are not sufficiently discussed:
• Summarize the resolutions, error characteristics, and limitations of each dataset in a table for clarity.
• Quantify the contributions of individual factors (e.g., PBL height, SAA, and seasonality) to observed biases rather than relying solely on trend descriptions.

Seasonal and Diurnal Trends

• Provide more detailed explanations for observed trends:
• How do lower PBL heights in the early morning contribute to NO2 accumulation?
• Why are concentrations higher in winter?
• Correlate these trends with emission sources and meteorological conditions at the observation site.

Extrapolation and Vertical Profiles

• Discuss the limitations of linear extrapolation methods used in vertical NO2 profile retrievals. Explore non-linear methods for more accurate surface concentration estimates.
• Analyze the changes in GEM-MACH profiles before and after smoothing across different height ranges, identifying layers with the most significant changes.

Minor Comments

Technical Issues

1. Line 225: The term "DOF" is inconsistently introduced. Ensure its full form and abbreviation are aligned throughout the text.
2. Standardize terminology for "Pandora direct-Sun" and "Pandora-DS" to avoid confusion.

Figures and Tables

• Figures 3 and 4: Clarify the differences in datasets and ensure consistent labeling.
• Add a table summarizing key attributes of the datasets (e.g., resolution, uncertainties).

Section 3.1.1

• The calculation of the 0–15 meter column concentration relies on surface NO2 measurements. Discuss the reliability of this assumption and its impact on comparisons. Quantify the average 0–15 meter column in absolute and relative terms.

The paper "Retrieval of NO2 profiles from three years of Pandora MAX-DOAS measurements in Toronto, Canada” by Ramina Alwarda et al., presents investigations on the NO2 profiles retrievals from the offaxis measurement scheme of the Toronto Pandora instrument over 3 years (2018-2020). The retrieval is done using Optimal Estimation HeiPro profiling algorithm and comparing the obtained profiles and partial columns to official Pandora direct-sun measurements, in situ observations, satellite data, and an air quality forecasting model.

The authors find that the HeiPro surface NO2 are close to the in-stiu NAPS measurements (small under-estimation of less than 10%) and underestmating the GEM-MACH model surface value (up to 40%). The HeiPro partial columns (up to 4km) are larger than the satellite S5p data (in agreement to other validation studies), while they are much larger than the direct-sun tropopsheric estimate and from the GEM-MACH model.

The paper is interesting, in the scope of AMT, and I would suggest its publication after some revision. 

In the current state of the manuscript, a comparison between different datasets is shown, but it is never clear if one of those datasets is considered as a reference, if it has already been validated elsewhere or if it is also prone to (large) uncertainties. The error are never mentionned, so it is not clear if the differences found between the datasets are within the combined uncertainties.

Another point is that a long investigation is performed to try to understand/quantify the causes of the differences between the Pandora MAX-DOAS and the Pandora direct-sun tropopsheric estimation, like investigating the influence of the PBL or the impact of different viewing angle due to the NO2 spatiotemporal heterogeneithy but the impact of some more basic assumptions are not really estimated and only very quickly discussed in the conclusion. In my opinion, two points are too briefly mentionned and not quantified enough:

- the choice of MAXDOAS long UV scans (oly discussed for the surface results)

- the quality of the Pandora direct-sun tropospheric estimation 

How would the comparisons for the partial columns be with the long VIS scans? How is the HeiPro MAX-DOAS comparing to the direct-sun total NO2 data? how much is removed from the orginal direct-sun total NO2 dataset to create the tropopsheric dataset? how good is the OMI stratospheric estimation? and its diurnal evolution estimation from the model? how do the orginal direct-sun total NO2 compare to S5p total NO2? 

These quantifications would allow to have more confidence in the Pandora retrievals and put the HeiPro data in relation to some reference data. 

The summary in my understanding:

1) HeiPRO VCD (0-4km) are larger than: a) PGN-DStropo (by a lot!), b) S5p (by a quantity similar to what other validation in similar context has found), c) the GEM-MACH model. For the latter case, it is clear in Fig4 (altough the larger differences are in winter months where maybe the number of points is not so representative?), but in Fig3c the comparison is more scattered and less clear. 

The first 2 cases imply clear-sky conditions (direct-sun measurements for one, and some cloud filtering for the satellite pixels in the second case). Are the comparisons wrt model done with some kind of filtering too (ie cloud filtering)? is not, would this improve the comparison? Maybe you could try to see if the comparisons with the model improves if you only select the same comparisons pairs that are selected when comparing HeiPRO to the direct-sun (ie the clear sky cases)?

2) Heipro surface concentration are smaller than: a) NAPS (by 4.4%/9%), b) model (by 40% and 36%)

When the model is smoothed with the MAX-DOAS AVK, the agreement seems better. We only see it in Fig.8. Please give numbers for this case too, as it seems imoprtant to me to report on the differences when taking into account the  vertical sensitivity of each technique.

It would be good to add a table with the summary differences for all the cases.

detailed comments:

--------------------

- line 88-89: "Comparisons of these PGN sky algorithm data products with other datasets at the measurement site in this study will be the subject of a future study" --> this is a pity that is is not included here, it would have brought an interesting additional comparison.

- Fig1: I would move this figure when explaining the differences in pointing between the MAX-DOAS and the direct-sun. Also add in the caption what time-period has been averaged to create the S5p NO2 map.

- line 194-> 197: the explanation of the stratospheric estimation substraction is presented here too quickly, and then again mentionned in Sect. 2.3.1. I would suggest to have one section explaining this part, with an illustration of the outcome and some quantification of the errors related to this step. When reading this part now, I have many questions coming up: why OMI and not S5p? and why that OMI product? is there any reference showing that it is a good stratospheric dataset? has it been validated?

For my own curiosity: are you performing zenith-sky twilight measurements with the pandora? if yes, you could also derive some stratospheric NO2 estimation from the Pandora itself...

- Pandora DS: I think it would be good to show its coherence wrt S5p total NO2 (eventually in the annex) - and discuss that this was done in the past, ie Zhao et al 2020, altough with a previous/different S5p product version.

- Sect. 2.2.2: explain a bit more the DOF from MAXDOAS, show an AVK and discuss more/show more results from both UV and VIS channel of the MAX-DOAS (not only decide because the long UV are in a closer agreement with the surface NAPS. By the way, I am not totally convinced by this statement. You base your analysis on the fact that the multiplicative factor in Fig A2a) is smaller than in A2b). But the regression slope and intercepts are smaller for A2b) thant for A2a)... 

Please also double check numbers. They are different in Fig A2a) and in in Fig 6a), while in my understanding, they are presenting the same datasets...

- lines 223-215: the choice of only using the long scans make sense to me (more information content with larger number of elevation angles) and is confirmed by the larger DOFs and RMS (not clear what RMS is meant in line 228), but then the choice of using the long UV scans instead of the long VIS scans is less convincing to me. They are more in agreement with the surface NO2 (with the theory that the horizontal extent of the line-of-sight is smaller, ok), but this reason is not so applicable to the direct-sun Pandora geometry or the satellite and model extent. 

I would show or at least comment the difference between results with both UV and VIS retrievals later on in the manuscript (at least as a form of error estimation?).

- line 268: GEM-MACH partial columns 0-5km. If you have the profiles, why you calculate the partial column up to 5km and not 4 km, as the MAX-DOAS?

- line 269: are the emissions in the model differenet for 2018, 2019 and 2020? you say in the discussion/conclusion that the emissions are outdated, and probably too high...

- line 281: why you use York linear fit for fig 3a and b) and then switch to ordinary linear regression in fig 3c)? it is a bit confusing 

- fig3: here or in the annex, I would add a plot of HeiPRO vs PGN total direct-sun and of PGN total direct-sun  vs S5p  total NO2.

why not stop the GEM-MACH partial columns 0-5km to 4km for fig 3 c)? as said above, you could integrate the GEM-MACH profiles only up to 4 km, to have the same vertical extent than the MAX-DOAS.

- line 399: " as well as the different horizontal sensitivities between the direct-Sun and multi-axis viewing geometries." --> you could refer here also to the HeiPRO data retrieved in the VIS long scans, which are sensitive to another (longer) horizontal path ?

- line 574: " this appears to skew the profile shape (see HeiPro profiles near 0 m in Fig. 8)," : I don't understand what your refer to.

- line 590: "do not account for free tropospheric NO2 sources while HeiPro measurements have some sensitivity to such layers," --> if you show the MAX-DOAS AVK, you can quantity/refer to its sensitivity in those layers

- line 593: "it is worthwhile to note that discrepancies between HeiPro and GEM-MACH profiles can be explained by the model inventories, " --> "can PROBABLY be explained by the.." (you don't show it is the case, you only assume it, right?!)

- line 598: "the smoothed GEM-MACH profile shows better agreement with HeiPro as it accounts for the MAX-DOAS measurement limitations and vertical sensitivity." --> please give also estimations for this case.

- line 700: "partial column RMS fitting residuals": what are those RMS? RMS between measured and modelles dDSCD in the OE inversion? please explain. DOFs are also never really explained, it is taken as granted that the reader know what they are...