Integrating explainable machine learning and multi-scale spatial modelling to reveal spatiotemporal dynamics and drivers of aerosol optical depth in a complex-terrain city: a case study of Chengdu Plain region, 2001–2024

Investigating the drivers of aerosol optical depth (AOD) in basin cities can improve our understanding of atmospheric processes and their responses to complex terrain and environmental changes. However, the nonlinear relationships between AOD and its potential drivers, as well as their spatial variation across different scales, remain poorly understood. The Theil–Sen trend estimator and nonparametric Mann–Kendall test were used to estimate and assess temporal trends in AOD. XGBoost-SHAP was used to identify nonlinear AOD–driver relationships and quantify the relative importance of individual predictors, while multiscale geographically weighted regression (MGWR) was used to assess their spatial heterogeneity and characteristic spatial scales. AOD showed a distinct rise–decline pattern over the study period. The annual mean AOD decreased to 0.409 during 2015–2020, a 42.1% reduction from the pre-2015 level. Seasonally, AOD was highest in winter (0.844 ± 0.180), followed by spring (0.765 ± 0.152), while summer (0.702 ± 0.178) and autumn (0.700 ± 0.158) showed similar levels. Spatially, higher AOD values were concentrated mainly across the central and eastern plains, whereas lower values occurred over the western Longmen Mountains. XGBoost achieved an R 2 of 0.85 ± 0.03 in leave-one-year-out cross-validation and 0.91 on the independent 2024 test set. Elevation, relative humidity, and precipitation were the three most influential predictors, with normalized mean absolute SHAP contributions of 46.94%, 19.37%, and 13.54%, respectively, together accounting for 79.85% of the total. MGWR indicated distinct spatial scales among the drivers, with bandwidths of 44–47 km for topographic and meteorological variables, 38–41 km for socioeconomic and vegetation-related variables, and approximately 36 km for road-distance effects. A temporal comparison showed that the relative contribution of anthropogenic factors decreased from 73.6% in 2001–2013 to 48.2% in 2014–2024, indicating a change in the statistical structure of AOD drivers. Population distribution across AOD levels was uneven, with approximately 40% of residents living in areas where AOD exceeded 0.45 and nearly 20% in areas exceeding 0.55. These results show that AOD variability was associated with regional topographic and meteorological conditions as well as spatially varying anthropogenic factors, with their relative importance differing across time and spatial scales.