<p>This paper examines the predictors of urban greenness across more than 10,000 urban centres worldwide using harmonised data from the Global Human Settlement Layer Urban Centres Database. Three complementary outcomes are examined: the share of population living in high green areas, the share of high green area within built-up area, and the overall share of green area within built-up area. Methodologically, the paper integrates XGBoost, SHAP, and Double Machine Learning (DML). XGBoost models identify the most important predictors and capture nonlinear relationships, SHAP values provide interpretable evidence on feature importance and effect direction, and DML estimates the conditional effect of a selected urban-form variable while adjusting for high-dimensional confounding. Across all three outcomes, average built-up height consistently emerges as the most important predictor in the models. The SHAP dependence plots indicate a strongly nonlinear and predominantly negative association: lower-rise urban forms are associated with higher greenness, whereas greater built-up height is associated with lower green-space provision and exposure. DML estimates suggest a negative conditional association between average built-up height and the greenness outcomes, under observed-confounding assumptions. Heterogeneity analyses further show that the estimated association varies across income groups and geographic contexts, with the largest point estimates in lower-middle-income cities. Overall, the findings suggest that urban greenness reflects layered processes linking physical constraints, spatial structure, and population exposure, and point to the need for greenness-sensitive densification policies.</p>