Ambient fi ne particulate matter (PM2.5) has been ranked as the sixth most important risk factor for death and disability globally. To accurately prevent and control PM2.5 pollution, high-resolution spatiotemporal characterization of PM2.5 variations is imperative, prompting the development of modeling approaches based on limited monitoring stations. This study used PM2.5 concentration (2014—2023) from seven state—controlled stations in Hangzhou’s core area as the response variable, with urban land—use landscape patterns as explanatory variables. A Land Use Regression (LUR) model, the Extreme Gradient Boosting (XGBoost) algorithm, and Spearman correlation analysis were employed to develop the model, while data splitting, 10-fold cross-validation, and external data validation were used to evaluate its stability. This framework was designed to investigate the mechanistic impacts of explanatory variables on PM2.5 and analyze the spatiotemporal heterogeneity of PM2.5 pollution. The validation results showed that the hybrid model performed well, with R2 and adjusted R2 above 0.90; the cross-validated MSE, RMSE, and MAE were 1.32 μg/m3, 1.15 μg/m3, and 1.08 μg/m3, respectively. The shape complexity of the forests patches within 500 m buff er and the 1000 m impervious surface average patch area are signifi cantly correlated with PM2.5; predictive capacity rankings were: woodland (44%) > cropland (33%) > impervious surfaces (23%); the annual average concentration showed an overall fl uctuating downward trend as well as a spatial distribution pattern of high in the northwest, low in the southwest, and increasing from the south to the north; refl ecting the “U”—shaped seasonal change characteristic of “high in winter and low in summer”, the low value area was located in localized areas of Xihu and Gongshu districts, and the high value area was located in localized areas of Gongshu and Binjiang districts.