Sequential spatial decoupling for street-level crime: An explainable GeoAI framework

Street-level crime is spatially concentrated, yet separating local environmental associations from structured spatial dependence remains difficult in empirical analysis. This study applies a sequential spatial decoupling framework to address this difficulty in central Shanghai, China. Using 8732 street segments and multi-source data, including crime records, street-view imagery, points of interest, and socioeconomic layers, we implement a two-stage approach: (1) ZINB-BYM2 under INLA to estimate structured spatial effects, and (2) XGBoost with SHAP, using the spatial component as a fixed log-offset, to model residual non-linear environmental associations. Across five crime types, the results distinguish structured spatial dependence, conditional local environmental associations, and non-linear interactive patterns. Ambient population displays a saturation-type contribution pattern, while the Visual Diversity Score exhibits a threshold-like transition across its observed range. The Risk-Environment Mismatch Matrix identifies distinct street profiles, including High-Low segments where elevated spatial risk coincides with a negative net environmental contribution, thereby differentiating locations that are otherwise grouped together by a spatial-risk classification alone. These comparisons provide a differentiated empirical examination of place-based opportunity conditions and assist street-level diagnostics under an explicit spatial baseline.