The manuscript addresses the variability of atmospheric mercury concentration in a coastal city in Southeast China. The manuscript aims to report the main factors driving GEM variability by deploying the regression analysis method. The scientific question is relevant to the scientific community. However, many issues can be highlighted in the manuscript.

The main concern in the manuscript is its design, how the Generalized Additive Model was used, and the premises assumed for the pattern recognition of the factors driving GEM variability. The authors lack knowledge of the used method. The signal extracted from the matrix of trace gases, PM, and meteorological data used to reconstruct GEM, is not explicitly linked to GEM sources, transport, or processes. The factorization was constrained by a minimum concentration covariance that led to the meteorologic factor as the main cluster. I am afraid that the authors were misled by a spurious correlation in the propagation of the eigenvector, where the main factor explaining the GEM was seasonality.  The main disadvantage of the unsupervised learning technique as the one used by the authors is the fact that the possible solution is no-unique.

Specific comments:

Line 236: The authors call data from two months “trend over 2012”; however, it corresponds only to teen months of data for a period of nine years. The terminology “trend” is incorrect throughout the manuscript and should be revised. After all, it is not clear why the authors used only January and July data.

Line 239-249: The emission data should be presented, and regression with observation should be discussed.

Line 243: “aggressive” what does it mean?

Line 252: Would it be possible to show the coal consumption in Fujian and China?

Line 259: Probably, the authors mean inter-annual variation rather than an inter-annual trend. I am afraid that the data exploitation presented by the authors does not allow a proper evaluation of the trend.

Section 3.1.2

I am afraid that using only two months is inappropriate for seasonality evaluation. In addition, one month represents only 1/3 of the season.

Perhaps it would be more appropriate to call the section January/July comparison rather than “seasonal”.

Line 271-282: The polar plot does not support the statement of dominant wind from the North or a higher concentration of GEM on this wind. If the plots are correct, the predominant source of GEM in January is in the west, and long transport does not play a major role in the level of GEM at Xiamen. Actually, the plot shows only a low level of GEM at wind from the sea.

Line 283-288: It seems confusing; the authors should consider rewording it.

Line 289: The diurnal pattern observed for July can be potentially constrained by sea/land breeze since it is a coastal place.

Line 297 – 298: For kinetic reasons, photo-oxidation cannot be the explanation for the observed reduction of GEM in the daytime. It is most like related to GEM fluxes. The authors speculate about the diurnal variation of GEM without a solid clue about the processes driving it.

Polar plots are quite limited in providing emission locations. Concentration-Weighted Trajectory could improve this section; it would map GEM, allowing hotspot concentration identification.

Figure 5 does not bring insight into the mercury source location. A different kind of plot should be presented. In addition, a clearer CWT method should be presented.

Line 365-370: It seems a last-minute explanation; since only fluxes can explain variation in the atmospheric mercury concentration, the authors should look into Hg emission to address a more convincing explanation.

Section 3.3.2

This section has major concerns

Unsupervised learning techniques are power statistic methods applied successfully to extract signal and meaning information from high-dimensional data. Deploying nonnegative matrix factorization, we can more than explain covariance; we can extract the pattern of source and transport of atmospheric trace gases. However, I am afraid that the authors did not design the factorization properly. The species considered in the matrix were chosen without criterium. It was only convenient for the authors to have those species there. What is the sense of having PM in the matrix? Considering species that do not bring retrieval signals will not provide insight into mercury processes/source/fade. It only increases uncertainty and chances of spurious correlation, misleading the eigenvector's propagation.

The major problem in this study was the correlations extracted from the species inserted in the factorization. The differences in the GEM concentration through the season, which are dependent on the seasonality of the emissions, were correlated with the seasonality of the meteorological parameters, which were extracted as the causes of GEM reduction in July.

The direct incorporation of meteorological parameters into the factorization misled the eigenvector propagation. The seasonal differences created cluster minimizing the variance but do not sign origin/source/or fluxes of GEM..

The factor obtained by the authors does not provide any insight into the GEM reducibility (computationally speaking) since it does not bring information on the source/or fluxes of GEM. Seems that the authors did not plan the species to be considered in the calculation.

Moreover, the meteorological variables cannot be included directly in the factorization matrix. In order to evaluate transport, the authors should use an inversion accoupled with a transport model.

I hope the authors do not feel disappointed or frustrated with my comments. I`m very enthusiastic about unsupervised learning methods for pattern recognition and estimation of fluxes and the implementation of nonnegative matrix factorization into inversion modelling. Indeed, it has great potential to bring new insight into atmospheric mercury reducibility. I hope the authors only feel motivated to learn and improve their research.

The manuscript by Shi et al. measured GEM concentrations in January and July in five individual years from 2012 to 2020, and the data were used to explore the potential factors controlling the inter-annual variations. Long-term measurements of GEM concentration are an useful tool for assessing of controls of regional anthropogenic emissions and global changes, and thus the data presented here are valuable. In the present study, the authors combined the multiple approaches including the analysis of GEM concentrations, criteria pollutants, backward trajectory and generalized additive model. I agree that it is practicable and relevant for using these kinds of methods to explore the controls in the change of atmospheric GEM. The manuscript is overall well organized, and can be read easily. I broadly agree with the discussions and findings of this manuscript. I therefore suggest a minor to moderate revision of this manuscript before final publication in ACP.

As mentioned at the beginning of the manuscript, the major objectives of long-term observations are to evaluate the changes in anthropogenic emissions, that is an important part for the implementation of the Minamata Convention on Hg. However, after a comprehensive analysis, the authors mostly highlight that the changes in meteorological conditions were the most important variable in controlling the long-term trend in GEM. This is valuable, but not very striking findings to me because it is well accepted that variations in GEM among different short periods (e.g., monthly) could be impacted by changing atmospheric transport (air transport would change with different periods and subsequently affect the source-receptor relationships). Thus, I would suggest the authors to focus on the impact of changing local and regional anthropogenic emissions and climate on the trends in GEM concentrations, which would better serve for their research objectives.

I am not clear why the meteorology is the major divers of changing GEM concentrations, and it also difficult to differentiate the impacts of meteorology, transmission, and emissions. I suspect that the transmission should be related to meteorology because the changes in local and regional meteorological conditions would further affect the transmission. Would the meteorology change land surface emissions and or atmospheric reactions that further affect the GEM? In addition, several previous studies reported declines in GEM concentrations in eastern China. Would this be an important cause for the changing contributions from transmission and meteorology? Thus, the authors may better define the three factors, which would help to better understand the real causes for the changes in GEM concentrations.  

Line 144-145: why did the authors only conduct a two-month observations at the sampling site? Why not conduct a year of continuous observations for the selected years? A two-month observations in one year are sometime not adequate for assessing the inter-annual variations because of many factors mentioned in the manuscript.

Line 166-167: the definition of local impact relating to air mass within a province might over-estimate the local effect. Why not define the local impact within the city?

Line 198-204: CO is mainly sourced from anthropogenic emissions but has a long atmospheric residence time, it may therefore a best proxy of local anthropogenic emissions. I would suggest the authors to consider using SO2, NO2, or PM10 to define the anthropogenic factor, although these parameters would have relative weaker correlations with GEM. Why use RH and SP to define meteorology? How could these two factors affect GEM concentrations? What are the 24h-latitude and -longitude? Are they referred as the air massed originated outside the city to define long-range transport?

Line223: the range of background GEM concentrations of 1.5-1.7 ng m-3 is somewhat higher to me. Better to use recent global observations.

 Figure 2: a statistic of the annual GEM concentrations should be added

 Line 274: the GEM lifetime here is not consistent with that in line 76

 Line 297-298: elevated O3 and decreasing GEM concentrations should be mainly related to subsidence of free troposphere and daytime production of O3. If the daytime declines in GEM is caused by oxidation, we would expect a much higher oxidation rate than experimental studies.

Line 317: a citation of references should be added here

As shown in Figure 5: a large fraction of air masses originated or passed over oceans, please add their weighted GEM concentrations in Table 2

Line 424-426: it is difficult to expect low regional anthropogenic emissions because the GEM measured are much higher the background levels in East Asia. I suspect that the other two factors of transmission and meteorology were also impacted by changing local and regional anthropogenic emissions. Actually, I do not know what are these three factors representing. Are the anthropogenic emission and transmissions representing local anthropogenic contributions and regional background? What is the meteorology? Is it representing natural emissions’?