Strengthening or weakening? Multi-year quantification and comparison of urban green space cooling effects and their key 2D/3D drivers

• Multi-scale cooling indicators were developed to compare historical cooling effect. • Divergent cooling trends were found in Beijing (1985–2023) across cooling indicators. • Integrates multi-source data with GeoAI for robust cooling mechanism decoding. • NDVI outperforms landscape metrics and 3D features in explaining cooling effect. The cooling effect generated by urban green space (UGS) plays a vital role in mitigating heat waves and enhancing citizen well-being. However, long-term studies examining the evolution of cooling effect alongside rapid urbanization-induced UGS changes remain scarce. Focusing on Beijing, China, this study quantifies the spatiotemporal changes in UGS cooling effects and identifies their key two-dimensional (2D) and three-dimensional (3D) drivers over nearly four decades (1985–2023). We innovatively proposed three distinct cooling intensity (CI) indicators for spatially explicit analysis and employed both stepwise regression analysis and machine learning models. Key findings include: (1) Beijing’s UGS coverage exhibited an overall slight decline, yet its spatial distribution became more balanced. The proportion of grids with UGS coverage below 20 % decreased significantly, from 8 % in 1985 to less than 1 % in 2023. (2) The perceived trend in cooling effect strength (strengthening vs. weakening) critically depended on the CI indicator used. The traditional CI (temperature difference between UGS and surroundings) showed a continuous decrease from 2.65 °C in 1985 to 1.83 °C in 2023. Conversely, regression model slopes indicated that a 10 % increase in UGS coverage yielded a stronger cooling effect in recent years, despite declining model R 2 values. Additionally, nearly 32 % of the study area exhibited an increase in outside CI of at least 0.5 °C, indicating a strengthening trend in the cooling effect. (3) Both statistical and machine learning analyses consistently identified the Normalized Difference Vegetation Index (NDVI) as the dominant driver, explaining over 50 % of CI variations. Landscape metrics and 3D UGS features contributed 38 % and 10 %, respectively. Combining all 2D and 3D characteristics explained over 70 % of CI variations, with NDVI, mean vegetation height, and aggregation index (AI) being the top three positively influential features.