| 25 Aug 2026
A Change in the Fine-to-Coarse Fraction in a Gobi Desert Town: Evidence for an Emerging Mixed Dust–Pollution Aerosol Pattern
Abstract. Mineral dust from the Gobi Desert is one of the largest natural contributors to the global aerosol burden, yet it is unknown whether urbanisation within the source region is altering the fine-to-coarse aerosol ratio. We analysed 13 years of PM2.5 and PM10 observations at four sites: Ulaanbaatar (UB) and three sites in the Gobi Desert (Dalanzadgad, DZ; Zamyn-Uud, ZU; Sainshand, SS), to investigate the relationship between winter PM2.5 and spring r-ratio. Winter PM2.5 in UB and DZ (112±45 and 80±22 μg m−3) substantially exceeded values in ZU and SS (19 ± 5 and <5 μg m−3), reflecting anthropogenic dominance, while spring PM10 reached 467–649 μg m−3 during dust peaks. In DZ, winter PM2.5 increased 1.6-fold (65 to 102 μg m−3) after 2014 and was significantly correlated with spring r-ratio (r = 0.69, p = 0.05). This association persisted under strong-wind conditions, with the elevated spring r-ratio retained between the pre- and post-2014 periods. These findings indicate an emerging mixed dust–pollution regime in DZ, coexisting with the anthropogenic regime at UB and natural dust at ZU and SS. Fine-particle enrichment in DZ suggests longer atmospheric lifetimes and transport distances and challenges the assumed net cooling effect of Gobi dust in urbanised areas, which warrants investigation through chemical composition measurements.
Overarching comments:
This manuscript addresses an important question: whether human settlement within a major mineral-dust source region changes the aerosol size distribution observed during the following spring. The long PM record and the contrast among Ulaanbaatar, Dalanzadgad, Zamyn Uud, and Sainshand could make a useful contribution. The observations also appear to establish that winter particulate pollution is substantially higher at Dalanzadgad than at the less populated Gobi sites.
I cannot, however, support the title-level conclusion that the present analysis demonstrates an emerging mixed dust-pollution regime caused by seasonal carryover of winter combustion aerosol. The measurements establish associations in bulk PM mass and meteorology, not the source, storage, or remobilisation pathway of the fine fraction. More importantly, the public workflow does not match several methods stated in the manuscript, and these differences affect the headline Dalanzadgad correlation and trend significance. An end-to-end reanalysis is required before the mechanistic and radiative interpretations can be evaluated.
Major Comments
1 The public workflow does not reproduce the stated raw data analysis
Section 2.2.2 states that the gap-filled data were used only for trend analysis and that the period comparisons and correlations used quality-controlled raw data. Appendix B3 further states that only seasons with more than 780 valid paired PM2.5 and PM10 observations were retained. The posted Figure 7 script instead loads df_processed_filled.csv, averages the de-normalised filled variables, and does not apply the stated 780-observation rule.
This difference changes the central result. Using the posted files and the manuscript's ratio-of-seasonal-means definition, the posted workflow gives the reported Dalanzadgad relationship of r = 0.689 and p = 0.040 with nine years. Applying the stated raw paired-data requirement leaves seven years and gives r = 0.607 and p = 0.149. The filled analysis includes the 2012 and 2013 Dalanzadgad winters even though the posted raw data contain only 219 and 346 paired hourly observations, respectively. The manuscript must identify the authoritative pipeline and make the text, tables, figures, data, and code agree.
The numerical coverage statement also needs correction: 780 hourly observations are approximately 36 percent of a 90-day season, not 64-65 percent. If another denominator was intended, it must be stated and justified.
2 The treatment of PM2.5 divided by PM10 is inconsistent and sometimes nonphysical
The manuscript defines r as PM2.5/PM10 and states that it ranges from zero to one. The posted data nevertheless contain many paired records with PM2.5 greater than PM10. In the filled file, this occurs in approximately 4.8 percent of Dalanzadgad records and 26.9 percent of Ulaanbaatar records. This is not a minor plotting issue because r is the primary response variable.
The public scripts handle these observations in several different and undocumented ways. The trend script caps seasonal and annual ratios above one at one. The pre/post and supplementary-table scripts multiply ratios above 1.2 by 0.8 and ratios between 1 and 1.2 by 0.85. The correlation script caps values above one and sets seasonal ratios below 0.4 to missing. The latter threshold removes the very dust-dominated values that the manuscript cites as the natural Gobi benchmark near 0.37. The strong-wind analysis uses hourly ratios without the same rule. These outcome-dependent transformations are not scientifically justified and make the results difficult to interpret.
All analyses should start from time-matched PM2.5 and PM10 observations and one pre-specified quality-control rule. The authors should report the frequency and cause of PM2.5 greater than PM10, propagate instrument uncertainty, and provide sensitivity analyses. A more transparent approach would analyse PM2.5 and the coarse mass PM10 minus PM2.5 separately, subject to the physical constraint that both are non-negative.
3 Data coverage and gap filling require a complete audit
The paper repeatedly describes a 13-year record from 2008 to 2020 at four sites. In the posted data, Dalanzadgad ends on 20 April 2018 and contains no 2019 or 2020 rows. Ulaanbaatar and Sainshand also have partial 2020 coverage, while Zamyn Uud extends beyond the stated period. Because the filled file does not create a complete hourly date grid, entire missing intervals cannot be imputed by the posted workflow. The statement that 2019-2020 Dalanzadgad outages were gap-filled therefore needs reconciliation with the deposited data.
The counts and missing fractions in Table 1 do not match the posted processed file. For example, the posted file contains far more missing PM2.5 observations at Sainshand and Ulaanbaatar than Table 1 reports. The authors should provide a site-by-variable coverage table for every year and season, distinguishing expected hours, recorded hours, paired valid hours, short-gap imputations, long-gap imputations, and wholly absent periods.
The validation in Appendix B is not sufficient. Correlating filled values with raw values at observed timestamps will approach one when the filling procedure preserves observed values; it does not test predictions inside gaps. Use blocked out-of-sample validation in which observed episodes of realistic length are masked and reconstructed. Report error and bias by site, season, concentration range, and gap length. A single deterministic fill also understates uncertainty. Multiple imputation or a raw-data primary analysis with imputation as sensitivity analysis would be more defensible.
The imputation should preserve PM2.5 less than or equal to PM10 and should treat wind direction as a circular variable, for example through sine and cosine components. The current normality check for Ulaanbaatar meteorology does not validate the PM variables, the other sites, or a linear treatment of wind direction.
4 The statistical design does not establish a seasonally lagged effect
The 2014 breakpoint is central to the argument but is not justified by a documented intervention at Dalanzadgad or by an objective change-point analysis. The result should be tested as a continuous trend and with segmented regression or change-point methods, with sensitivity to nearby break years. If 2014 was selected after inspecting the series, the resulting p values are exploratory.
The p values in Table A3 and the text should be recalculated from the authoritative data. Using the posted filled seasonal means and the stated Welch test, I obtain approximately p = 6.1 x 10^-4 for the winter PM2.5 contrast at Dalanzadgad and p = 7.4 x 10^-4 for its spring r contrast, rather than 0.06 and 0.07. The Dalanzadgad table script uniquely formats p values as percentages, which may explain the discrepancy. Please audit every reported test, state n and degrees of freedom, and report effect sizes with confidence intervals.
Winter PM2.5 and spring r both change with calendar year. Their raw cross-year correlation can therefore arise from a shared trend rather than a physical lag. In the posted filled series, removing linear year trends reduces the Dalanzadgad association from r = 0.689, p = 0.040 to r = 0.391, p = 0.297. A time-aware model should estimate whether winter PM2.5 adds explanatory power after controlling for year, spring wind, precipitation or snow, soil moisture, heating demand, and other plausible covariates. First-difference and leave-one-year-out results should be reported.
Figure 7 evaluates six correlations at each of two sites, while Table A2 reports many site-by-season-by-variable trend tests. No multiplicity control is described. The nominal p = 0.040 headline correlation becomes approximately 0.106 after Benjamini-Hochberg adjustment across the six Dalanzadgad correlations alone. The authors should define a primary hypothesis before testing and control or clearly disclose the family-wise exploratory analyses.
Figure 6 treats hundreds of serially correlated hourly observations from the same days and dust events as independent Wilcoxon samples. This is pseudoreplication and will make p values too small. Aggregate by independent event, day, or year, or use a hierarchical time-series model with event or date clustering and residual autocorrelation. The figure tests a pre/post contrast within wind bins; it does not show that the winter-to-spring correlation itself persists under strong winds. The wording in the abstract and conclusion should be corrected.
5 The proposed deposition and resuspension mechanism remains a hypothesis
High winter PM2.5, a higher spring PM2.5/PM10 fraction, and a cross-season correlation do not identify the source or history of the spring particles. Night-time winter maxima are consistent with residential heating under a shallow boundary layer, but they do not by themselves confirm coal combustion. Likewise, a spring fine fraction does not distinguish combustion aerosol, secondary aerosol, traffic or construction emissions, or fine mineral dust.
The phrase that winter aerosol persists into spring is particularly problematic. The proposed pathway is not atmospheric persistence but deposition to a surface reservoir followed by remobilisation. Demonstrating that pathway requires at least a quantitative mass-balance argument and preferably source-specific evidence. Carbonaceous fractions, black carbon, levoglucosan, combustion-related metals, or stable and radiocarbon source fingerprints would provide direct tests. If such measurements are unavailable, the deposition-resuspension pathway should be presented as one testable explanation for future work, not as the mechanism established here.
The reanalysis must also separate numerator and denominator effects. An increase in PM2.5/PM10 can result from more fine mass, less coarse mass, or both. The manuscript already notes declining wind and PM10, so a ratio-only analysis cannot attribute the change to residual winter particles. Plot and model absolute spring PM2.5 and PM10 minus PM2.5 alongside the ratio.
6 Instrument comparability and the meteorological source classes need validation
The methods do not provide enough information to assess a 13-year trend from light-scattering PM measurements across aerosol types as different as mineral dust and combustion aerosol. Please document inlet and size-cut configuration, instrument changes, calibrations and co-locations, maintenance, detection limits, flow checks, relative-humidity treatment, and the conversion from optical response to mass. Composition, density, shape, and humidity can affect optical mass estimates and therefore the PM2.5/PM10 fraction. These effects could change with season and site.
Meteorological measurements are also not directly comparable as described: wind and visibility were measured at 15 m in Ulaanbaatar and 3 m at the Gobi sites. Wind thresholds should be standardised to a common height or translated to friction velocity with roughness information. Dust mobilisation also depends on soil moisture, snow, surface condition, and precipitation.
The visibility and wind-speed framework in Figure 8 is a meteorological classification, not source apportionment. Visibility below 10 km and wind above or below selected thresholds cannot uniquely identify dust, anthropogenic aerosol, or a mixture. Validate the classes against independent chemical or optical source indicators, provide threshold-sensitivity tests, and otherwise rename the categories as meteorological regimes.
7 The radiative and regional climate claims exceed the measurements
PM2.5/PM10 alone does not determine single-scattering albedo, absorption, mixing state, aerosol optical depth, or radiative forcing. The manuscript acknowledges that the spring fine fraction could be clay or combustion aerosol, but the abstract, discussion, and conclusion still imply that the observations challenge the net cooling effect of Gobi dust and reveal a model bias. Those claims require composition and optical measurements or model calculations. They should be removed from the main conclusion or stated explicitly as hypotheses that motivate future work.
The strongest conclusion supported by the present measurements is descriptive: Dalanzadgad has elevated winter PM and a higher spring PM2.5/PM10 fraction than the less polluted Gobi sites. Whether winter deposition causes that spring fraction, and whether the radiative effect changes sign, remain unresolved.
Minor Comments
Use the correct quantity name throughout. PM2.5/PM10 is the PM2.5 mass fraction of PM10, not a fine-to-coarse ratio. A fine-to-coarse ratio would be PM2.5 divided by PM10 minus PM2.5. The title, axis labels, keywords, and text should be revised.
Figure 2 reports a long-term mean PM2.5 of 69.3 µg m-3 and PM10 of 61.3 µg m-3 at Ulaanbaatar, which is physically impossible for matched observations. The plotting script averages the two variables over different sets of complete days. Recalculate all summaries from paired timestamps and report the common n.
The Dalanzadgad longitude is 104.42 degrees E in the text but 104.22 degrees E in Table 1. Verify all coordinates.
Table A1 labels visibility as kilometres but reports values such as 10,331 and 19,288, which appear to be metres. Correct the unit or rescale the values.
Section 3.2.2 and the text surrounding Figure 7 conflate two different Ulaanbaatar results. The approximately 0.70 correlation is winter PM2.5 versus spring PM2.5; winter PM2.5 versus spring r is approximately 0.02 in the figure. Revise the caption and interpretation.
Line 261 states that an increase in wind speed supports the conclusion, whereas the study reports declining wind speed. Correct this sentence.
The manuscript uses DJF, MAM, JJA, and SON for the main analysis but November-February, March-June, and July-October in Figure 8. Explain the scientific reason and avoid comparing results across incompatible seasonal definitions without qualification.
Figure 4 is too dense to verify. Show annual points, data coverage, Sen slopes, confidence intervals, and the actual Mann-Kendall results. Do not display ordinary linear trend lines when the inferential method is Mann-Kendall and Sen slope unless the distinction is explicit.
Figure 5 should show annual observations and uncertainty rather than bars alone. The current panel encourages a binary pre/post reading of very small samples.
Figure 6 should report independent sample units, not only hourly n. The post-2014 bin above 13 m s-1 contains only six hourly observations and cannot support a general regime claim.
Reformat Table A3 with separate columns for difference, confidence interval, p value, and n. The present superscript formatting makes values such as 37 and 0.06 appear as a single number and may have contributed to the p-value error.
Clarify whether the Ulaanbaatar PM inlet was also at 15 m or whether only the wind and visibility sensors were at that height. Table 1 and Section 2.2.1 are ambiguous.
Replace causal wording such as confirm, demonstrate, residual effect, and first observational evidence with association-based wording unless the revised analysis supplies source-specific evidence.
Archive the exact data and scripts used for the accepted version in a versioned repository with a DOI, data dictionary, package versions, and a single executable workflow. The current repository lacks the code that creates the processed and gap-filled files, and the Data Availability statement does not reflect the files already posted.