the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 31 Aug 2026
An Explainable Machine Learning Perspective on Anthropogenic Emission and Meteorology drivers of long-term PM2.5 Trends over Northeast China
Abstract. Northeast China is a cold industrial and agricultural region where PM2.5 pollution is influenced by emission changes, winter heating, agricultural burning, and meteorological variability, but long-term city-level evidence of these drivers remains limited. Here we combined hourly PM2.5 observations from 36 cities during 2015–2025 with ERA5 meteorology, LightGBM models interpreted using Shapley Additive Explanations, and weather normalization. Observed PM2.5 was separated into a weather-normalized component (PMemi), representing non-meteorological variability in the pollution baseline, and a meteorological modulation term (PMmet). Regional annual mean PM2.5 decreased by 39.7 %, from 48.4 μg m−3 in 2015 to 29.2 μg m−3 in 2025. The PMemi trend accounted for 95.5 % of the observed linear decrease and showed consistent decreases with independent CEDS emission indices, supporting an important role of anthropogenic emission reductions. However, regional annual mean PM2.5 changed little after 2022 and remained above 25 μg m−3. Regional exceedance days decreased from 52 to 12 d yr−1, but PMmet was positive on all 229 exceedance days, indicating meteorological amplification during high-pollution episodes. Model interpretation further linked exceedance-day PM2.5enhancement to thermal, pressure, moisture, wind, and boundary-layer conditions. VIIRS fire detections and CO observations indicated that agricultural biomass burning and other combustion sources were associated with short-term PM2.5 variability during spring and autumn burning seasons. These results show how emission reductions, residual combustion sources, and unfavorable meteorology jointly control PM2.5 variability in cold industrial and agricultural regions.
- Preprint
(3241 KB) - Metadata XML
-
Supplement
(588 KB) - BibTeX
- EndNote
Status: open (until 31 Oct 2026)
-
CC1: 'Comment on egusphere-2026-4127', Xuhui Gao, 21 Sep 2026
reply
-
CC2: 'Reply on CC1', Zhongfeng Pan, 22 Sep 2026
reply
We thank the commenter for raising the issue of temporal autocorrelation in hourly model validation. We agree that this point should be considered when interpreting random cross-validation results for time-resolved air-quality data.
In this study, the hourly random 5-fold cross-validation was used to evaluate the model's reconstruction capability across different meteorological and PM2.5 conditions, rather than as a strict future-year forecasting test. To further examine temporal robustness at a scale more relevant to the attribution analysis, we conducted an independent-year validation at the monthly scale. The model was trained using 2015–2024 data and tested on the independent year 2025, avoiding random splitting of neighboring hourly samples between training and testing sets.
The independent 2025 monthly test achieved R = 0.842, R² = 0.659, RMSE = 8.58 μg m⁻³, and MAE = 6.13 μg m⁻³. These results indicate that the model retains reasonable predictive skill under a temporally independent monthly validation setting. Therefore, the hourly random 5-fold results should be interpreted as a reconstruction diagnostic, while the independent monthly validation supports the robustness of the model at the temporal scale relevant to the subsequent trend-attribution analysis.
Citation: https://doi.org/10.5194/egusphere-2026-4127-CC2
-
CC2: 'Reply on CC1', Zhongfeng Pan, 22 Sep 2026
reply
-
RC1: 'Comment on egusphere-2026-4127', Anonymous Referee #1, 26 Sep 2026
reply
This manuscript investigates the long-term PM2.5 trends across 36 cities in Northeast China from 2015 to 2025 using a LightGBM model integrated with the SHAP explainability framework. By employing a meteorological normalization approach, the authors successfully disentangle the contributions of anthropogenic emissions and meteorological conditions to PM2.5 concentrations. The study is further strengthened by validations using independent indicators such as CEDS emission inventories, VIIRS fire counts, and CO observations. The topic is of significant regional interest, and the methodology is robust. The findings provide valuable insights for air quality management in cold, industrial, and agricultural regions. Minor revision is recommended. I also have some suggestions about this manuscript.
Major suggestions:
1.The results indicate that while regional observed PM2.5 levels remained essentially stable after 2022, PMemi increased, suggesting that meteorological factors partially offset the rise in emissions. This finding carries significant policy implications. It is recommended to state more explicitly in the abstract, discussion, and conclusion that the stability of observed PM2.5 does not imply a continued decline in the weather-normalized pollution baseline, and that further emission reductions remain challenging.
2. PMemi shows a high correlation with the CEDS index; however, this does not mean that PMemi is an emissions inventory. In the comparison section, it should be emphasized that PMemi is not an emissions inventory, but rather the non-meteorological component derived after weather normalization.
3. Using the method where grid center points fall within city polygons may introduce bias in areas with complex terrain; it is recommended to discuss this as a limitation.
4. Section 4, “Annual trend and fire-activity comparison” resembles an analysis methodology; it is suggested that this section be merged into Section 3 or renamed “Analysis methods”
5. Line 471，The statement “current annual standard of 25 μg/m³” should specify whether it refers to the Chinese national standard, WHO guidelines, or another standard. The current Chinese annual secondary standard for PM2.5 is 35 μg/m³, whereas the WHO 2021 interim target is 25 μg/m³. If the latter is intended, please state this explicitly.
6. CEDS typically provides NOx rather than NO. Please change “NO” to “NOx” and confirm that the selected species match the CEDS variable names.
7. Avoid presenting the correlation between PM_emi and CEDS as direct causal evidence of emission-control policy effects. Use more cautious wording such as “consistent with” or “associated with”
Minor suggestions:
1. Line 139-140, “For cities with multiple monitoring station” should be “For cities with multiple monitoring stations”.
2. Line 529, “meteorology more often acted as a cleaner under normal conditions” should be “meteorology more often favored PM2.5 removal under normal conditions”.
3. For title, “meteorology” should be “meteorological”.
4. Line 30, “thermal” should be replaced as “temperature”.
5. Ensure all abbreviations (e.g., BLH, SSRD) are defined at their first appearance. While most seem defined, a final check is recommended.
6. Please check the formatting consistency in the reference list, particularly regarding journal abbreviations and year placement.
7. I suggest that the authors should carefully review the grammar of this manuscript.
Citation: https://doi.org/10.5194/egusphere-2026-4127-RC1 -
RC2: 'Comment on egusphere-2026-4127', Anonymous Referee #2, 07 Oct 2026
reply
This manuscript combines hourly PM2.5 observations from 36 cities in Northeast China during 2015-2025 with LightGBM modeling, SHAP analysis, and weather normalization to investigate long-term pollution changes. The long-term, multi-city coverage provides a useful regional perspective, particularly given the combined influences of industry, winter heating, and agricultural burning. I have the following comments.
Comments
1. Definition of PMmet (Section 3.4). Equation (12) defines PMmet as observed PM2.5 minus PMemi. This difference also includes the model residual described earlier, although residuals are stated to be used only for diagnostics. Please clarify this point and avoid attributing the entire difference to meteorology, particularly when discussing exceedance days.
2. Data coverage and regional averaging (Sections 2.1 and 4.1). Please state whether minimum data-coverage requirements were applied to daily and annual means, and whether the set of monitoring stations remained consistent throughout the study. Regional means were calculated from all valid city-hour records, giving greater weight to cities with more valid hours. Please make this weighting explicit.
3. Meteorological resampling (Section 3.4).Weather conditions were sampled from all valid hours while the target time variables remained fixed. For example, a January record could be paired with weather observed in July, although its month variables still indicate January. Please explain this choice and discuss whether such unusual combinations could affect the weather-normalized estimates.
4. CEDS indices (Section 2.3 and Figure 4). Please specify the CEDS release and clarify the calculation of the summed-species and mean-species indices in Figure 4b–c. Because these indices combine primary particles and gaseous precursors, “PM2.5-related emission indices” would be more accurate than “PM2.5 emission indices.” Please also note that the high correlations with PMemi may partly reflect shared downward trends.
5. CO comparison (Section 2.1 and Figure 7c–d). Section 2.1 states that CO was used to assess PMemi. However, Figure 7c (all months) and Figure 7d (burning months) appear to show comparisons with observed PM2.5 rather than PMemi. Please confirm which series was used and make the stated purpose consistent with the analysis shown. A comparison with observed PM2.5 does not directly assess the weather-normalized component.
Citation: https://doi.org/10.5194/egusphere-2026-4127-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 297 | 132 | 57 | 486 | 121 | 57 | 62 |
- HTML: 297
- PDF: 132
- XML: 57
- Total: 486
- Supplement: 121
- BibTeX: 57
- EndNote: 62
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors used 5-fold random cross-validation to evaluate their hourly machine learning model. However, this setup likely leads to data leakage. Atmospheric conditions and PM2.5 levels are naturally very similar across nearby hours. With random splitting, test samples easily end up right next to training samples in time. This allows the model to perform well simply by interpolating nearby data, rather than learning the true physical processes. The authors need to explain how this temporal autocorrelation leakage was prevented. Without addressing this issue, the model's reported high performance and the subsequent attribution analyses remain questionable.