Assessing the causal impact of the Chinese Spring Festival on PM<sub>2.5</sub> air quality in Beijing-Tianjin-Hebei and surrounding region using a machine learning counterfactual modeling approach&nbsp;

Li, Yuan; Dai, Qili; Zhu, Wubin; Liu, Xuan; Shen, Jiandong; Yan, Renchang; Li, Yunshan; Ding, Jing; Lee, Young Su; Zhang, Yufen; Feng, Yinchang

doi:10.22541/essoar.174559329.93866726/v2

Preprints

https://doi.org/10.22541/essoar.174559329.93866726/v2

Preprints

13 Oct 2025

| 13 Oct 2025

Status: this preprint is open for discussion and under review for Atmospheric Chemistry and Physics (ACP).

Assessing the causal impact of the Chinese Spring Festival on PM_2.5 air quality in Beijing-Tianjin-Hebei and surrounding region using a machine learning counterfactual modeling approach

Yuan Li, Qili Dai, Wubin Zhu, Xuan Liu, Jiandong Shen, Renchang Yan, Yunshan Li, Jing Ding, Young Su Lee, Yufen Zhang, and Yinchang Feng

Abstract. Acute short-term exposure to extremely high PM_2.5 levels posed serious health risks. Human culture-based festival activities can significantly alter emission patterns, often leading to sharp yet understudied fluctuations in air quality. The Chinese Spring Festival (CSF), marked by large-scale family reunions and widespread use of fireworks, raises air pollution concerns. Commonly, this effect is quantified using receptor models or chemical transport models, but the relevant chemical component data and emission inventories are often lacking. This study presents a machine learning counterfactual approach to causally quantify PM_2.5 changes associated with holiday activities. The results align well with traditional chemical composition-based estimates of fireworks contributions, highlighting the strong potential of using widely accessible routine monitoring data to quantify source contributions driven by specific interventions. Applied to the twenty-eight major cities in Beijing-Tianjin-Hebei and surrounding area, one of the most polluted regions in China, the approach revealed an average PM_2.5 reduction of 19.0 ± 17.5 μg/m³ during the CSF holiday period in 2025, with fireworks accounting for ≥35 % of first-day severe deteriorated PM_2.5 air quality and up to 89 % in Baoding. The approach offers a robust tool for evaluating holiday emissions and guiding air quality interventions.

Received: 16 Sep 2025 – Discussion started: 13 Oct 2025

Yuan Li, Qili Dai, Wubin Zhu, Xuan Liu, Jiandong Shen, Renchang Yan, Yunshan Li, Jing Ding, Young Su Lee, Yufen Zhang, and Yinchang Feng

Status: open (until 22 Dec 2025)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2025-4562', Anonymous Referee #1, 29 Oct 2025 reply
This manuscript "Assessing the causal impact of the Chinese Spring Festival on PM_2.5 air quality in Beijing-Tianjin-Hebei and surrounding region using a machine learning counterfactual modeling approach" by Yuan Li and team, addresses an important and interesting topic: the influence of the Chinese Spring Festival (CSF) on regional PM_2.5concentrations, particularly the attribution of emissions to fireworks. The use of a machine learning counterfactual model is an interesting approach to isolate the festival's effect. However, the core conclusions regarding the high contribution of fireworks, especially at the regional scale, are based on data and methodological interpretations that lack sufficient resolution and rigor to justify the claim. Specifically, the analysis appears to conflate highly local, transient firework plumes with persistent regional emissions from industrial and urban sources. This weakness must be addressed before the manuscript can be considered for publication.
Major Comments:
The manuscript interprets elevated specific chemical tracers (e.g., Al, Ba, Cu and K) as fireworks and then equates the PMF-resolved “fireworks” factor with a very large fraction to city PM_2.5 (e.g., fireworks reaching 137 µg m^-3 and up to 81.8% at the peak) and then treats that as validation of the ML holiday-attributable signal. But trace metals from fireworks can adsorb or be scavenged onto pre-existing aerosol mass-a single atom/molecule of a tracer attaching to many aerosol mass units can cause PMF to label existing background aerosol as “firework-contaminated” even if the mass contribution from fireworks is small. This overlooks the critical aspect of aerosol mixing and coating. The presence of a fireworks marker on an aerosol particle only indicates mixing or coating has occurred, not that the particle's mass is primarily sourced from fireworks. For instance, a persistent industrial Black Carbon (BC) or organic carbon particle, when mixed in the atmosphere with even a single molecule of Al/Na/Cu or K from fireworks, will be registered a 'firework' marker. But in reality, the particle's overall mass and emission may still be dominated by the pre-existing industrial or alternative sources of BC core. Would this not induce a overestimation of mass contribution of fire works and biases the source apportionment. I think the author’s revise their interpretation to differentiate between a source signature (the tracer) and mass contribution (the overall particle composition). One approach to distinguish this is maybe to examine the ratios of Ba/Al, K/Ba or K/EC or K/BC and their co-variation with EC or BC to see whether the tracers appear as spikes superimposed on otherwise continuous mass.

Multi-platform observations of persistent BC and other aerosols in the adjascent regions – such as Xuzhou and Suzhou, which share air-mass pathways with Shandong and Linyi- reveal substantial , continuous emissions from urban and small industrial sources (Tiwari et al., 2025). These background sources have far greater emission potential than transient firework events. The manuscript should reconcile its high firework attribution with these documented, regional emissions, placing CSF findings within the borader context of persistent regional aerosol loading. This also demands a clarification to the geographic scope of the conclusions. A localized short duration fire works plume from Shandong is unlikely to represent a regional-scale PM_2.5 driver across the entire Beijing-Tianjin-Hebei domain. The authors should specify whether their results describe a local fireworks effect or a regional influence and adjust the claims accordingly.

The ML counterfactual approach is trained and evaluated on city-averaged PM_2.5 (14 monitoring stations averaged for Hangzhou; Fig. 3 caption), but does the DN-PMF chemical composition used for validating the fireworks factor comes from a single site (Wolongqiao). Is the manuscript assuming the single-site composition being representative of the whole-city PM_2.5composition when assessing fireworks contribution. Please clarify and justify representativeness, if possible via spatial correlation of species across stations or multi-site composition if available. If the differences are large, the validation claim should be downweighted or additional PMF runs may be necessary for other sites.

The authors attribute ML > DN-PMF magnitude difference partly to “unproportionally amplification effect due to midnight affect due to the midnight unfavorable dilution conditions (extremely low VC)”. This is plausible, however, if low VC multiplies existing aerosol mass, then fireworks mass should be amplified only in concentration but not in emission strength if fireworks were purely local instantaneous injections. Thus, clarifying the distinction between concentration amplification vs inferred emission-strength amplification would further enhance the overall understanding of this section.

In Sect. 3.3-3.5 the authors compute fireworks contribution on the first day as “increase relative to counterfactual, under assumption that other sources remained unchanged.” This is a very strong assumption: other sources may decline (traffic reductions) or increase (localized cooking, residential heating). The manuscript needs to either: (a) quantify the extent of other source changes using available proxies (traffic counts, NO₂/CO changes, DN-PMF vehicle factor), or (b) present their fireworks fraction as an upper/lower bound with explicit caveats. Right now the wording implies more certainty than justified.

I think the claim firework explain 68.8% of instantaneous PM_2.5 needs more uncertainty reporting, to be deemed robust. The DN-PMF uses 8 factors and labels Factor 1 as firework because Al/Ba/K peaks. But PMF solutions can be non-unique and sensitive to input species and uncertainties. Shedding light on PMF diagnostics would make the results more robust.

Minor/Technical Comments:
The Methods state “No strict hyper-parameter tuning requirements were imposed — either Bayesian optimization or grid search could be used” (Sect. 2.2). That is ambiguous. Please state what the authors actually used (grid values, CV folds, early stopping) and include the tuned hyperparameters in a supplement. Report learning curves and variable importance (SHAP or partial dependence) so readers see what drivers the model uses.

Mixing gas-phase species (CO/NO_X) with PM speciation should be echoed in the manuscript as a caveat: gas tracers and PM tracers respond to different source processes and lifetimes, so combining them without accounting for atmospheric chemistry and differential lifetimes could mislead source attribution (Li et al., 2025). The authors briefly note NO₂/CO decline; they should deepen that discussion (e.g., use NO₂ as independent corroboration of traffic reduction).

Figure 1, Figure 2 and Figure 3 have missing references in the main text.

Reference:
Tiwari, P., Cohen, J.B., Lu, L. et al. Multi-platform observations and constraints reveal overlooked urban sources of black carbon in Xuzhou and Dhaka. Commun Earth Environ 6, 38 (2025). https://doi.org/10.1038/s43247-025-02012-x
Li, X., Cohen, J.B., Tiwari, P. et al. Space-based inversion reveals underestimated carbon monoxide emissions over Shanxi. Commun Earth Environ 6, 357 (2025). https://doi.org/10.1038/s43247-025-02301-5

Reply
Citation: https://doi.org/10.5194/egusphere-2025-4562-RC1
RC2: 'Comment on egusphere-2025-4562', Anonymous Referee #2, 05 Nov 2025 reply

The manuscript proposes a machine-learning counterfactual framework to estimate the impact of the Chinese Spring Festival (CSF) on PM2.5 in Hangzhou and the “2+26” cities. The study is timely and policy-relevant, with a clear intention to distinguish air-quality changes from emissions, and the manuscript is well organized and clearly presented. However, several aspects of causal ML practice, the temporal validation strategy, and issues of data representativeness need to be strengthened before publication.
Major Comments:
1. The manuscript frames its analysis within a causal framework, treating the Chinese Spring Festival (CSF) as a “treatment” and using the XGBoost model to predict a counterfactual business-as-usual (BAU) scenario. While this is a conceptually appropriate starting point, the current methodology does not yet meet a rigorous causal ML design. The CSF is a composite factor, bundling the effects of fireworks, altered traffic patterns, and changes in industrial/construction activity. This complexity challenges the core identification assumptions required for causal claims. 

Furthermore, the analysis does not adequately address potential influence of these assumptions, such as the inconsistent overlap in covariate distributions between festival and non-festival periods. Some features, like the lunar calendar day, are inherently confounded with the treatment, violating conditional independence. The study could be characterized as a causally inspired counterfactual prediction for BAU rather than a causal estimator under verified identification conditions. Hence, the authors may wish to reconsider the title and tone down the causal claims to avoid overstatement.
2. The current modeling approach, which relies on instantaneous covariates, does not account for the temporal auto-correlation inherent in air pollution. The concentration at any given time is also influenced by the emissions and meteorological conditions of previous periods. The choice of a random 80/20 split for model validation may introduce data leakage when evaluating the model performance. A blocked or rolling time-based cross-validation would be more appropriate here. 

Separately, uncertainty quantification has been extensively discussed in ML-based atmospheric remote sensing, yet is not addressed in the present manuscript; providing calibrated predictive uncertainty would improve the interpretability of the results.
3. The abstract opens with acute short-term health risks from extremely high PM2.5, but the regional result emphasizes an average decrease of 19.0 ± 17.5 μg/m3 over the extended holiday period. These two statements are not contradictory but currently feel weakly connected.

Besides that, Section 3.2 (Hangzhou) explicitly reports large concurrent source changes (e.g., vehicles -31%; dust +2790%), yet Section 3.5 (“2+26”) estimates fireworks’ contribution “under the assumption that emissions from other sources remained unchanged.” The authors need to address this inconsistency or provide sensitivity analysis under alternative assumptions.
4. Section 2.1 requires several clarifications. First, key details for the ERA5 dataset, including its temporal/spatial resolution and a reference link, should be provided in the manuscript or SI (Text S1/Table S1). To address the potential for reanalysis data to smooth over urban-scale extremes, a brief comparison of ERA5 variables against ground-station data would strengthen the analysis. Additionally, the usage of total precipitation (TP) needs to be explained; since it is an accumulated value, please describe any transformation performed to make it suitable for an hourly model. The specific parameters or a reference for Emanuel’s saturated vapor pressure formula should be included.
5. There is the spatial representativeness mismatch between the machine learning model, which uses a 14-site city average, and the DN-PMF analysis, which uses chemical data from a single site. This difference could introduce a bias, particularly for localized sources like fireworks. The authors could discuss this limitation and its potential impact on their findings.  Given the team’s related work (e.g., Journal of Environmental Sciences), a brief comparison of the methodological advantages and efficiency gains relative to prior work would also help position the contribution.
Minor corrections:
- Line 22: twenty-eight -> 28
- Line 119: meterorology -> meteorology
- Line 164: A -> An
- Line 207: in the midnight of the New Year Eve -> at midnight on New Year’s Eve
- Line 244: Please add units for RMSE and MAE.
- Line 260: reliablity -> reliability
- Line 261: techique -> technique
- Line 323: deterioriation -> deterioration
- Table S1: Please use Pa (not pa) for pressure unit.

Reply

Citation: https://doi.org/10.5194/egusphere-2025-4562-RC2

Yuan Li, Qili Dai, Wubin Zhu, Xuan Liu, Jiandong Shen, Renchang Yan, Yunshan Li, Jing Ding, Young Su Lee, Yufen Zhang, and Yinchang Feng

Viewed

Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.

Total article views: 165 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
163	0	2	165	0	0

HTML: 163
PDF: 0
XML: 2
Total: 165
BibTeX: 0
EndNote: 0

Views and downloads (calculated since 13 Oct 2025)

Month	HTML	PDF	XML	Total
Oct 2025	83	1	84
Nov 2025	80	1	81

Cumulative views and downloads (calculated since 13 Oct 2025)

Month	HTML	PDF	XML	Total
Oct 2025	83	1	84
Nov 2025	80	1	81

Viewed (geographical distribution)

Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.

Total article views: 165 (including HTML, PDF, and XML) Thereof 165 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 27 Nov 2025


Total:	0
HTML:	0
PDF:	0
XML:	0

Assessing the causal impact of the Chinese Spring Festival on PM2.5 air quality in Beijing-Tianjin-Hebei and surrounding region using a machine learning counterfactual modeling approach

Viewed

Viewed (geographical distribution)

Assessing the causal impact of the Chinese Spring Festival on PM_2.5 air quality in Beijing-Tianjin-Hebei and surrounding region using a machine learning counterfactual modeling approach