Explaining spatially heterogeneous drivers of urban expansion with an XGBoost–SHAP–UGM framework

Urban expansion has significantly changed land use patterns and poses challenges to achieving Sustainable Development Goal (SDG) targets 11, 13, and 15. However, the diverse and phased mechanisms involved in this process have not yet been fully understood. This study proposes an integrated XGBoost–SHapley Additive exPlanations-Urban Growth Model (XGBoost-SHAP-UGM) framework to jointly simulate urban land conversion and interpret its driving forces at multiple scales. Using multi‑source data for Beijing, Wuhan and Zhaoqing from 2000 to 2020, we first train an XGBoost classifier to estimate the probability of conversion from non‑urban to urban land based on natural, accessibility and socio‑economic factors. SHAP are then applied to quantify the contribution of each factor, revealing nonlinear and threshold effects across different stages of urbanization. Exploiting the raster nature of the data, we compute local SHAP values for every grid cell and aggregate them into three driver scores (natural, accessibility, socio‑economic). K‑means clustering in this SHAP‑score space yields a mechanism‑oriented classification of driver regimes, which explicitly maps the spatial heterogeneity of urban growth mechanisms. The resulting regimes (balanced high‑potential growth zones, ecologically constrained barrier zones, accessibility‑constrained zones, and socioeconomically constrained low‑demand zones) show consistent patterns across the three regions while reflecting their different urbanization stages. The probabilistic outputs of XGBoost are further incorporated into a UGM to simulate urban growth trajectories, with high predictive skill. The proposed framework advances explainable urban growth modelling by transforming black‑box predictions into interpretable, cell‑level mechanism maps, and provides a transferable tool for stage‑specific and spatially differentiated urban growth planning in support of the SDGs.