Abstract: Objective　PM2.5, defined as fine particulate matter with an aerodynamic diameter of ≤ 2.5 μm, is a key pollutant affecting air quality and public health. Currently, the sparse and discontinuous distribution of ground monitoring stations makes it difficult to obtain spatially continuous PM2.5 mass concentration data across regions. On the other hand, traditional satellitebased estimation models (e.g., linear regression, geographically weighted regression) relying on aerosol optical depth (AOD) suffer from limitations in capturing the complex nonlinear relationships between PM2.5 and AOD and in addressing spatiotemporal heterogeneity. Moreover, previous studies have mostly employed AOD products with relatively low spatial resolution (e.g., MODIS data with 3 km or 10 km spatial resolution), which restricts the spatial detail of the estimation. To address these gaps, this study utilizes higher-resolution (1 km) Multi-Angle Implementation of Atmospheric Correction (MAIAC) AOD products combined with multi-source geospatial data to construct and compare several high-performance machine learning models. The ultimate objectives are to achieve high-precision, high-resolution estimation of PM2.5 mass concentrations in the Guanzhong region and to systematically reveal their spatiotemporal evolution patterns, thereby providing a scientific basis for targeted air pollution prevention and control. Methods　The PM2.5 mass concentration estimation method proposed in this study consists of three main steps: data preparation, model construction, and spatiotemporal analysis. Firstly, the 1 km resolution MAIAC AOD data from the MCD19A2 product, ground-measured PM2.5 mass concentrations, meteorological factors, normalized difference vegetation index (NDVI) data, and digital elevation model (DEM) data of the Guanzhong region from 2019 to 2021 were integrated. And all datasets underwent spatiotemporal matching, quality control, and normalization processing. Then, to ensure an objective evaluation of the models' spatial predictive capability, a site-based cross-validation strategy was adopted to divide the data into training and test sets. Three machine learning models, namely Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost), were constructed and optimized through ten-fold crossvalidation and grid search. The performance of the three models was evaluated using the coefficient of determination, root mean square error (RMSE), and mean absolute error (MAE). Finally, the optimal model was applied to generate annual, seasonal, and monthly PM2.5 mass concentration distribution maps at 1 km resolution, enabling a detailed analysis of spatiotemporal variations. Results and Discussion　The experimental results indicate that: (1) compared with the SVM and MLP models, the XGBoost model performs the best, achieving the highest coefficient of determination and the lowest RMSE and MAE, demonstrating its strong nonlinear fitting capability; (2) the annual average PM2.5 mass concentrations for 2019, 2020, and 2021 were 23.79, 24.38, and 22.83 μg/m³, respectively, showing a declining trend with a gradual reduction in the spatial extent of high-concentration areas, which may be associated with ongoing emission control policies; (3) the seasonal variation of PM2.5 in this region is significant, with the most severe pollution occurring in winter and the lowest in summer, which is closely related to the local heating demand and meteorological diffusion conditions; (4) the monthly average mass concentrations of PM2.5 in this region peak in January (60.73 μg/m³) and reach the lowest in September (12.91 μg/m³), consistent with seasonal variaion patterns. These results confirm that integrating high-resolution MAIAC AOD data with the XGBoost model can effectively improve estimation accuracy and provide detailed insights into the spatiotemporal evolution of PM2.5. Conclusions　A high-resolution PM2.5 mass concentration estimation method suitable for the Guanzhong region is presented in this paper. The method is developed by combing MAIAC AOD data with the XGBoost model, effectively overcoming the limitations of traditional approaches and providing reliable, continuous PM2.5 distribution data. The experimental results not only clearly reveal the spatiotemporal patterns of PM2.5 pollution in the Guanzhong region but also offer a scientific basis for regional air quality management. Future work will focus on incorporating more real-time variables and enhancing model interpretability to further support the formulation of pollution control decisions.

Key words: PM2.5, multi-angle atmospheric correction algorithm, aerosol optical thickness, machine learning, PM2.5-AOD relationship model