In this study, the frequency of number of light rain days was calculated based on the dataset of CPC Global Unified Gauge-Based Analysis of Daily Precipitation (CPC-Global) (https://psl.noaa.gov/data/gridded/data.cpc.globalprecip.html, last access: 30 May 2025). The CPC-Global dataset is with a spatial resolution of 0.5° × 0.5° and produced based on assimilation and interpolation of observations from 30 000 stations by National Oceanic and Atmospheric Administration (NOAA) through the CPC Unified Precipitation Project (Xie et al., 2007). The daily ERA5 reanalysis data from 2000 to 2022 provided by the European Centre for Medium-range Weather Forecasts (ECMWF) was also used in this study (https://cds.climate.copernicus.eu/, last access: 30 May 2025). These data include the relative humidity (RH) at 850 hPa, temperature (T) at 2 m height, the wind speed (WS) at 850 hPa, total column cloud liquid water (TCLW), low cloud cover (LCC), convective available potential energy (CAPE) and evaporation (E). The horizontal resolution for the meteorological data is 0.25° × 0.25°. The PM_2.5 data are obtained from the China High Air Pollutants dataset (CHAP) at a spatial resolution of 1 km × 1 km from 2000 to 2022 (Wei et al., 2021, 2019; Wei and Li, 2024).

In this study We defined the light rain event as the daily precipitation is between 0.1 and 10 mm (Qian et al., 2009; Huang and Wen, 2013; Wu et al., 2015) which is referenced from the standards of the CMA. To better demonstrate the effect of changes in PM_2.5 and other meteorological parameters on changes of the number of light rain days, we conducted the analysis using data based on two periods from 2000 to 2013 and from 2013 to 2022. The year 2013 was selected as the breakpoint because it marks a definitive shift in China's air pollution policy with the implementation of the national “Air Pollution Prevention and Control Action Plan”, which led to a proven reversal in the long-term trend of PM_2.5 concentrations nationwide (Wei et al., 2021). Moreover, the frequency of light rain in most regions of China, especially those heavily affected by human activities, also showed a change pattern of first decreasing and then increasing before and after 2013 (Figs. 1, 2). The XGBoost model was applied separately to the two periods (2000–2013 and 2013–2022) rather than to the entire dataset. This approach was chosen because the underlying physical relationships between the predictors and light rain are expected to differ significantly between the pollution-accumulating and pollution-abatement regimes. Training separate models prevents the estimation of a misleading “average” relationship and allows for a more accurate quantification of the distinct drivers operative in each period. Least squares technique (linear regression) is firstly employed to estimate the annual trends of light rain days and the examined factors. Additionally, the extreme gradient boosting (XGBoost) model is applied to examine the impact of these factors, including PM_2.5, RH, WS, T, E, TCLW, CAPE, and LCC, on the number of light rain days. The selection of these specific factors is grounded in their well-established physical linkages to light rain processes, as extensively documented in prior studies (Qian et al., 2009; Huang and Wen, 2013; Li et al., 2017). Briefly, PM_2.5 is included to quantify aerosol impacts on cloud microphysics, while the meteorological variables collectively represent the thermodynamic, moisture, dynamic, and cloud-related conditions that are fundamental to light rain formation. This approach ensures our model captures the key mechanistic drivers identified in the literature.

The importance of each parameter affecting the accuracy of the light rain days prediction is assessed. The XGBoost model, which is a powerful and reliable machine learning technique, optimizes the gradient boosted decision tree algorithm to solve regression and classification problems (Chen and Guestrin, 2016). It offers several advantages, such as handling missing values, preventing overfitting, enhancing computing speed and accuracy, and has been widely applied in predicting air pollutants and emissions (Gui et al., 2020; Si and Du, 2020, 2020). Previous studies have shown that XGBoost performs well with relatively small-scale datasets (Cheng et al., 2021, 2023). This work used 5-fold cross-validation of the training set to test the performance of the XGBoost model and defined the parameter search space to determine the optimal model parameters, finally generating the prediction results. The predicted frequency of the number of light rain days by the XGBoost method exhibit high consistency with the observed values, with mean R2 of 0.83 and 0.90 (Fig. S3 in the Supplement). Through K-fold cross-validation evaluation, each model achieved a mean coefficient of determination (R2) of 0.80 on the cross-regional dataset, demonstrating robust generalization performance across different geospatial scales. Then, the drivers of the long-term variations of the light rain days over the two periods were further explored. For this, the annual trend (depicted by slope obtained based on linear fitting) of daily data for each individual factor as well as the frequency of light rain days were firstly calculated as the input variables and target parameter respectively. The result shows good performance with mean R2 of 0.90 (Fig. S6), The K-fold cross-validation (R2 > 0.88) results also demonstrate robust performance across different datasets.

SHAP is a machine learning interpretability technique based on coalitional game theory (Lundberg and Lee, 2017), with its core mechanism lying in the mathematical quantification of feature contributions. Within the SHAP framework, SHAP values essentially represent a formalized mathematical deformation of Shapley values, which decompose model predictions to attribute the deviation of each sample's predicted outcome to the contributions of individual features. The explanation could be specified as: 1g(z′)=θ0+∑jMθjz′jg where g is the explanation model, z′∈{0,1}M is the simplified features, M is the maximum coalition size, and θj∈R is the weighted average of all marginal contributions for a predictor variable j, the Shapley values.

This method is particularly effective for identifying the relative importance of predictor variables in XGBoost models, providing feature attribution analysis that integrates global consistency with local interpretability. Compared to feature importance methods such as the gain method or split count, SHAP not only reveals key driving factors through global importance rankings but also visualizes feature contribution distributions at the individual sample level, enabling an interpretability analysis of the interaction between model prediction and characteristic effect.

SEM is a method that is usually used to test hypotheses on relationships of multi factors within a complex system (Lamb et al., 2014; Ganjurjav et al., 2021). These hypotheses are articulated in the form of a series of regression equations, known as structural equations. This model enables the simultaneous analysis of multiple variables with causal relationships and overcomes the limitations of traditional multivariate analytical techniques such as multivariate analysis of variance and correlation analysis. One of the most important advantages of SEM is its capacity to disentangle complex interdependencies among multiple parameters. This is achieved by estimating the independent (direct) effect of each variable while simultaneously accounting for its correlations with other variables in the model. When constructing an SEM, hypotheses about causal relationships between variables are first formulated based on theoretical and prior research findings, and then the result is adjusted according to whether the fit indices meet statistical criteria. The key indices for evaluating SEM model fit include the Comparative Fit Index (CFI), Root Mean Square Error of Approximation (RMSEA), and Standardized Root Mean Square Residual (SRMR). In general, when CFI approaches 1, 0 ≤ RMSEA ≤ 0.05, and 0 ≤ SRMR ≤ 0.08, the model indicates a very good fit (Schreiber et al., 2006; Ma et al., 2022). In addition, the chi-square test (χ2) and Adjusted Goodness of Fit Index (AGFI) are also commonly used for model evaluation. In this study, the final model fitting results showed that CFI was 0.998 and RMSEA was 0.033, which indicates that the SEM had a very good fit with the data and that the model fit was almost ideal.

Figure 1 illustrates the spatial distributions and long-term trends of light precipitation frequency and PM_2.5 mass concentrations across China during the studied periods (2000–2022). Statistically significant decreasing trends in light rain days (mean rate: 1.0 d yr⁻¹, p < 0.05) were observed nationwide during 2000–2013, with the most pronounced decline in southern China (2.3 d yr⁻¹, p < 0.01) (Fig. 1). Conversely, a reversal trend emerged during 2013–2022, showing continent-wide increases (1.9 d yr⁻¹, p < 0.01), particularly in southwestern China and the Yangtze River Basin (central-eastern regions), where the growth rate reached 2.6 d yr⁻¹. The data analysis also reveals an inverse correlation between light precipitation trends and PM_2.5 variations: a substantial significant PM_2.5 increase (0.39 µg m⁻³ yr⁻¹, p < 0.01) occurred during 2000–2013, contrasting with a significant decline (2.5 µg m⁻³ yr⁻¹, p < 0.01) post-2013. Actually, the rapid decrease in PM_2.5 mass concentration since 2013 have been widely observed due to the implementation of the national “Air Pollution Prevention and Control Action Plan” in China (Zhang et al., 2020; Wei et al., 2021; Bai and Liu, 2024; Zhang et al., 2024). Note that, before 2013, the national average annual light rain days showed significant interannual fluctuations, with the lowest number of light rain days recorded in 2007 (Fig. 1a). This makes it seem as if 2007 was a turning point where the trend of light rain frequency shifted from decreasing to increasing. However, when the analysis is focused on specific different regions, it is found that the turning point of light rain frequency still occurred in 2013 (Fig. 2). Considering the variation trend of PM_2.5 comprehensively, this study divides the research period into two phases of 2001–2013 and 2013–2022.

In particular, this inverse relationship is most evident in six anthropogenically influenced regions (Fig. 2): North China (NC), South China (SC), East China (EC), Southwest China (SWC), Central-South China (CSC), and the Fenwei Plain (FW). This regional division is based on established frameworks that consider distinct physiographic (e.g., topographic basins, climate zones) and anthropogenic (e.g., population density, industrial activity) characteristics (Chen et al., 2024). For instance, the Fenwei Plain (FW) is treated separately from North China (NC) due to its unique enclosed topography that fosters pollution accumulation, despite some similarities in overall trends.

Since previous studies attribute light rain suppression in polluted regions to aerosol-cloud microphysical interactions (Qian et al., 2009; Wang et al., 2016; Shao et al., 2022), our analysis suggests that the observed decadal shifts in light precipitation (2000–2013 decline vs. 2013–2022 increase) are likely driven by long-term aerosol concentration variability. Notably, regions with minimal anthropogenic activities (e.g., Inner Mongolia and northwestern China) exhibited no significant PM_2.5-light rain correlations but with good correlations with meteorological factors (e.g., RH) (Figs. 1,  S2), implying contributions from non-aerosol factors.

Figure 3 depicts the long-term trends of meteorological factors (T, RH, CAPE, E, LCC, TCLW, and WS) potentially associated with light rain variability in China during 2000–2013 and 2013–2022, while the statistical significance (p < 0.05) of these trends is shown in Fig. S1. A key observation is that, in contrast to the strong and statistically significant nationwide trends seen in PM_2.5, the trends of these meteorological parameters are characterized by pronounced spatial heterogeneity and largely insignificant changes over large portions of China. For the E, from 2000 to 2013, approximately 60 % of regions exhibited insignificant trends, only with the significant decline (p < 0.05) observed in southwestern China and the Qinghai-Tibet Plateau (Fig. S2). During 2013–2022, E just decreased significantly in very few areas of northeastern China and northern Xinjiang, areas concurrent with RH increases (Cong et al., 2009). Conversely, large areas of southeastern China experienced nonsignificant E reductions. Notably, E-enhanced regions seems expanded in recent decades, likely attributable to global warming effects (IPCC, 2023). However, the enhancement is not statistically significant (Fig. S2). Analysis reveals that 56 % of Chinese regions showed T increases slightly during 2000–2013, with accelerated warming rates post–2013. The CAPE demonstrated weakened trends across southeastern China during 2000–2013, while remaining stable in western/northwestern regions. This spatial pattern aligns with anthropogenic aerosol impacts. The aggravated aerosol pollution over 2000–2013 likely suppressed convection in densely populated eastern China via microphysical mechanisms (Zhao et al., 2006). However, PM_2.5 reductions since 2013 moderated CAPE declines, with some regions (e.g., middle-lower Yangtze River) even exhibiting strengthening trends, indicating meteorological sensitivity to aerosol loading changes. The TCLW decreased only in the Yangtze River Basin and Tibetan Plateau during 2000–2013, but reversed to increasing trends post-2013. There are no evident variations (statistically insignificant) in TCLW elsewhere during the two periods (Fig. S2). Nationwide WS reductions (statistically nonsignificant) were observed during both periods, though southeastern China experienced accelerated declines post-2013, potentially linked to persistent particulate pollution and urbanization processes (Zhang and Wang, 2021). The RH declines affected less than 30 % of China during 2000–2013, most prominently in the Tibetan Plateau, northeastern/southwestern China. During 2013–2022, it is observed with nonsignificant RH variations becoming dominant nationwide. The LCC trends spatially mirrored RH patterns, likely underscoring their coupled responses to anthropogenic and climatic drivers. While these meteorological parameters exhibited spatially variable trends, their magnitudes were generally much smaller than PM_2.5 variations. Subsequent sections will quantitatively evaluate their combined impacts on light rain frequency.

To elucidate the influence of aerosols and meteorological parameters on light precipitation events, variable importance metrics were derived using the machine learning – based XGBoost algorithm. Considering that the temporal variation trends of these factors are not identical, we presented and analyzed the conditions of the two research periods (2000–2013 and 2013–2022) (Fig. 4) – this was done to compare and explore whether the relative contribution or importance of each factor to precipitation has changed across different stages. The XGBoost-based model demonstrates robust predictive capability, achieving a correlation coefficient of 0.83 and 0.90 (Slope: 0.80, 0.85 p < 0.01) between predicted and observed light precipitation events (Fig. S3). This validation ensures the reliability of subsequent parameter importance evaluations for elucidating meteorological drivers of light rain. Figure 4 quantifies the relative contributions of the PM_2.5 and meteorological factors (RH, T, E, TCLW, CAPE, LCC, WS) to light rain occurrence during 2000–2013 and 2013–2022. Nationwide, RH exerted the dominant influence, accounting for 32 % of light rain variability across both periods. The PM_2.5, T, and E followed as key drivers, with mean contributions of 15.5 %/11.7 %/11.7 % in 2000–2013 and 12.1 %/10.5 %/11.5 % in 2013–2022 respectively. Notably, the contribution of PM_2.5 declined from 15.5 % to 12.1 % (a 3.4 % decrease, p < 0.05), coinciding with China's stringent emission controls (Shao et al., 2022; Wang et al., 2016). Conversely, the contributions of TCLW and CAPE increased significantly from 7.2 %/9.4 % in 2000–2013 to 13.6 %/10.6 % in 2013–2022 (p < 0.01). LCC and WS remained negligibly minor factors (< 6 % contribution). Moreover, spatial heterogeneity (Fig. 3) revealed no substantial differences in factor contributions between the two periods, except for PM_2.5 and TCLW. Conspicuously, the meteorological factors exhibited statistically insignificant temporal trends, implying stable physical mechanisms underlying their impacts on light rain.

Spatial heterogeneity characterizes the relative importance of meteorological factors (Fig. 5). For instance, the RH exerts critical influence in southeastern and northwestern China, peaking at 50 % contribution in southern regions. This spatial pattern aligns with established findings showing significant RH-light rain day correlations in southern China (Wu et al., 2015; Zhang et al., 2019a; Zhou et al., 2020). Conversely, the contribution of RH diminishes to ∼ 10 % in northern regions, likely attributable to lower atmospheric moisture content (Fig. 5b). In contrast, PM_2.5 emerges as the dominant driver in northern China, especially in North China Plain (NCP), where aerosol concentrations exceed those in pristine western regions (Fig. S4). This aerosol-mediated regulation aligns with cloud microphysical mechanisms suppressing light precipitation (Qian et al., 2009; Yang et al., 2016). The E demonstrates pronounced influence in arid Inner Mongolia and northwestern China, consistent with hydrological sensitivity in water-limited ecosystems (Chen et al., 2014; Yang et al., 2024). The E-light rain relationship weakens eastward, reflecting enhanced soil moisture buffering in humid regions. The TCLW assumes greater significance in western and northeastern China, where the water vapor content is relatively low, the warm cloud precipitation process, which relies on the growth of liquid water and the collision and coalescence of water droplets, makes the TCLW a crucial factor for the occurrence of light rain (Gao et al., 2016). Notably, importance of TCLW increased significantly (+ 18 %) in NCP and northeastern China post-2013, coinciding with PM_2.5 reductions that reduced cloud condensation nuclei (CCN) and elevated cloud droplet coalescence efficiency (Twomey, 1977; Guo et al., 2019). Consequently, even small changes in TCLW can directly reflect the occurrence of light rain. These findings underscore the complexity of regional meteorological controls, necessitating regionalized analyses for mechanistic correlation.

The analysis is further conducted in the six typical areas with elevated anthropogenic emissions. As shown in Fig. 5, there are large differences in the dominant factors affecting light rain frequency across regions. The RH plays a leading role in SC, EC, SWC, and CSC regions, with mean contributions of ∼ 40 %. In FW and NC regions, RH contributions diminish to 15 %–20 %, whereas PM_2.5 contributions increase markedly (20 %–30 %), exceeding the national average by approximately twofold. This anthropogenic imprint is amplified in other high-emission regions (CSC, SWC, EC), where PM_2.5 exhibits considerable influence. Other factors (CAPE, WS, T, E, LCC) show no substantial variations in contributions (< 10 %) across the six regions. These findings indicate the dominant role of anthropogenic emissions in light rain occurrence within heavily polluted areas. Notably, impact of PM_2.5 declined slightly in NC and FW during 2013–2022 compared to 2000–2013, aligning with China's pollution controls. The declines are consistent with decreases in the national average but showing steeper reductions. Concurrently, TCLW gained prominence in most regions (except CSC) post-2013, increasing its contribution to near 30 %. These results underscore aerosol-cloud microphysical effects as critical regulators of light precipitation in polluted regions.

To elucidate the drivers of long-term trends in light rain day frequency described in Sect. 3.1, we systematically investigated the relative contributions of multi-factors using an integrated machine learning (XGBoost), interpretability technique (SHAP)and SEM framework (Fig. 6). As stated previously, considering that the temporal variation trends of these factors are not identical, we presented and analysed the conditions of the two research periods (2000–2013 and 2013–2022) – this was done to compare and explore whether the driving factors of influencing the precipitation long-term trends has changed across different stages. Therefore, this dual-method approach quantifies both observed patterns (e.g., downward trends in 2000–2013 vs. upward trends post-2013) and underlying causal mechanisms. The results show good performance of our established model for predicting the trends of light rain frequency (with correlation coefficient R2= 0.90 between the measured and predicted trends of light rain frequency) (Fig. S5).

The results demonstrate aerosol dominant role in driving long-term trends in light precipitation frequency across both periods, contributing 59 %–63 % to the interannual variability of annual light rain days (Fig. 6). Specifically, the increasing trend in PM_2.5 concentrations in 2000–2013 has led to varying magnitudes of decreases in the number of light rain days in the six regions, 0.65 d yr⁻¹ (NC), 2.22 d yr⁻¹ (EC), 1.64 d yr⁻¹ (SC), 2.04 d yr⁻¹ (CSC), 3.0 d yr⁻¹ (SWC), and 1.0 d yr⁻¹ (FW). (Fig. 7). Conversely, the pronounced decrease in PM_2.5 levels from 2013 to 2022 has increased the light rain days of 0.7 d yr⁻¹ (NC), 2.42 d yr⁻¹ (EC), 1.66 d yr⁻¹ (SC), 1.85 d yr⁻¹ (CSC), 2.62 d yr⁻¹ (SWC), and 1.0 d yr⁻¹ (FW) (Fig. 7). Note that, compared to the period of 2000–2013, the relative importance of the variations in PM_2.5 in affecting the light rain in recent years of 2013–2022 becomes more prominent likely associated with the rapid nationwide decrease of aerosol concentrations since 2013. This aligns with Twomey's cloud albedo effect (Twomey, 1977), where elevated aerosol loadings suppress light precipitation via cloud microphysical feedbacks (Qian et al., 2009; Wang et al., 2016; Shao et al., 2022). Reduced aerosol concentrations post–2013 attenuated these suppression effects, enhancing light rain frequency through improved cloud droplet coalescence efficiency. Machine learning-based analysis further elucidates this underlying microphysical aerosol-mediated mechanism (Figs. 6, 7), revealing aerosols dominant influence on light precipitation. These findings reconcile national-scale aerosol-light rain described in Sect. 3.1 with microphysical process dynamics. Averaging the regional results from Fig. 7, the national mean contribution of PM_2.5 was -2.08 d yr⁻¹ during 2000–2013 and +1.97 d yr⁻¹ during 2013–2022.

Other meteorological factors, such as RH and LCC, demonstrated significant positive correlations with light rain day frequency in most regions (Figs. S7–S11). Our trend analysis in Sect. 3.1 revealed that they, along with the new insights from regional analysis (Fig. 8), exhibited statistically insignificant long-term changes over the study period. Consequently, their quantified contributions to interannual variability were negligible (mostly < 5 %), further supporting the conclusion that aerosols were the dominant driver of the observed trend reversal. Notably, in NC, the TCLW showed considerable importance in regulating the long-term variations of light rain frequency, explaining 26 % of light rain days decline (about -0.49 d yr⁻¹) over 2000–2013 and 25 % of the light rain increase (about +0.48 d yr⁻¹) during 2013–2022. Similar patterns were observed in FW, likely attributable to shared climatic and pollution regimes between these two northern regions.