Review of “Data fusion uncertainty-enabled methods to map street-scale hourly NO2 in Barcelona city: a case study with CALIOPE-Urban v1.0” by Criado et al.

Criado et al. present a data-fusion workflow that uses Universal Kriging (UK) to merge dispersion model output (from CALIOPE-Urban) to hourly observations and microscale land-use regression (LUR) models. The authors’ workflow is able to create high-resolution street-scale data of NO2 to compute exceedance probabilities, with uncertainty calculations based on the UK technique estimated error variance. This work is comprehensive and appears to have good improvement in correlation and error metrics. I will be happy to recommend this manuscript for publication after my (mostly minor) comments below are addressed.

Major comments:

1.    I want to note that the code isn’t available and thus cannot be reviewed in its current form. The authors state that “So, at this moment only reviewers can access these relevant sources under a previous mail in the request form.” But the request form on Zenodo requires the full name, e-mail address, and affiliation of the requester, compromising the anonymity of the referees. Thus, I was not able to review the code that is associated with this work. I would request that the authors provide the code used in this work for review, either through the Editor or the GMD portal. While I appreciate that the authors have archived the code in a repository with a DOI, access during review is important. Many other authors have publicly archived their code when submitting to GMD, despite the manuscript being under review.

2.    The authors have performed data-fusion using data from two LCS campaigns, one from 2018, and one from 2017. Is there an impact on the quality of the data-fusion technique if LCS data are provided in different years? Similarly, if only one LCS campaign data set is used, how would it impact the quality of the results? A brief assessment of how much data is necessary and the applicability of the methods shown in this work will help future readers to apply this technique in the future to other major cities where urban pollution is also a major health issue.

Specific comments:

1.    L57: “while the time-dependent LCS network explains the temporal behavior”. Is it implied that the temporal behavior is short-term here in contrast with the long-term spatial distribution provided by the urban model?

2.    L134: An adjustment factor is computed as the ratio between the observed 2017 annual mean and the average over the period of the experimental campaign. The LCS campaigns span only a few weeks (February 16th to March 15th, 2018; and February and March, 2017) – why is the 2017 annual mean used here instead of, for example, February-March mean?

3.    L136-137: The authors say that this processing adds some noise to experimental results but corrects the “environmental conditions influence”. What environmental conditions are referred to here? My impression is that this would mainly correct for bias in the low cost sensors’ instruments.

4.    Figure 3: Useful to label inset in the figure “Combined data” “CSIC” “xAire”.

5.    L247-248: The authors indicate that with the criteria (covariate slopes must be positive, less than four observations available in the hour) 14% and 2% of the hours in UK-DM, UK-DM-LUR are not corrected. How much percentage of these are due to negative covariate slopes? And how much are due to too few observations? If there is a significant percentage of nonphysical negative covariates, is there a common pattern to the conditions causing these?

6.    L273: “We attribute this behavior … also to the already poor predictive skills of CALIOPE-Urban in this concentration range.” A citation will be useful for CALIOPE-Urban’s underperformance in high-NOx conditions.

Technical corrections:

1.    L204: “back-transformed” -> “back-transformation”

2.    L210: “exceedance (P) a certain…” -> “exceedance (P) of a certain…”

3.    Lines 233, 234: middle dot -> cross sign for scientific notation.

4.    Figure 7: Units for MB, RMSE are missing the “^3” (shows as micrograms/m)

Review of EGUsphere-2022-1147: Data fusion uncertainty-enabled methods to map street-scale hourly NO2 in Barcelona city: a case study with CALIOPE-Urban v1.0, by Criado et al.

Urban NO₂ pollution shows strong gradients, and is almost always undersampled by reference networks. The authors present an interesting approach to assimilate complementary observational datasets, having different temporal sampling and accuracy, in a high-resolution urban dispersion model. Using Universal Kriging, they show that the predictive performance of the dispersion model improves if hourly measurements are included. Moreover, they show that the system further improves if a basemap is added in the data fusion, based on 840 Palmes tube measurements. The observations clearly resolve local spatial structures which are not properly described by the street model alone. The authors show that the error margin provided from the Kriging method is fairly realistic, which enables them to calculate maps with expected local exceedances of annual and hourly limit values of NO2 air pollution. 

The paper is well-referenced, positioning the study well in the current research efforts on this topic. I recommend publication after addressing the following comments:

General comments

1. The microscale LUR model is trained by Palmes observations done in end-February/begin-March. Although an annualization is applied (L132-L136), I assume the resulting basemap would look differently when evaluated for months with e.g. different typical NOx lifetime, boundary layer height, dominant wind direction. Does the usage of a February/March basemap throughout the year introduce a significant seasonal bias?

2. The study uses hourly NO2 measurements of 12 reference stations in the Barcelona area. To my knowledge, NO2 reference measurements are also done at the Observatori Fabra site, which is also within the considered domain. Is there any reason why this station is excluded?

3. Oftentimes, local authorities evaluate the air quality in their city based on measurements of the reference network only. This gives a distorted impression, as many local exceedances are not sampled. Data fusion methods, such as in this study, correct for this sampling bias. The authors show that a large part of the city does not meet the annual and hourly limit values for NO2 (L365-385). It would be interesting to see how this contrasts with an analysis based on station data alone. 

Specific comments

4. Section 3.1.2, Figure 6: The LUR basemap (6a) seems to be richer in detail than the mean of the dispersion model (6b). However, it misses the lower pollution levels in the mountainous area in the NW part of the domain. This introduces an unwanted bias in this area for UK-DM-LUR when compared to UK-DM, which escapes the validation statistics as there is no reference data available (or used) in this area. How could the microscale-LUR model be improved for non-built-up areas?

5. Section 3.2.2, Figure 8: Can part of the skewness of the distributions be explained from the back-transformation from the log-domain (L202-204)? 

6. Conclusions: I miss some general words about transferability of this method to other cities, referring to dependencies on databases, dispersion models, and local network configurations. Also, I recommend summarizing briefly the estimated exceedances in L265-385.

7. L394-395:  “The obtained microscale-LUR basemap (r=0.64, RMSE=11.87μg/m3) outperformed the raw annual-averaged dispersion model results (r=0.54, RMSE=13.68 μg/m3)”. This is not very surprising, as the annual-averaged model results is one of the inputs of the LUR model.

Technical corrections

L22: “obtaining high-resolution exposure to NO2 is crucial”→ “obtaining information on high-resolution exposure to NO2 is crucial”

L245: “if their slope is positive”. Confusing for me. I guess that a positive slope refers to a positive coefficient in the linear combination.