Abstract
Climate non-stationarity and intensifying extreme rainfall challenge flood-risk assessment in cold-region old industrial cities, where renewal must work within an established built fabric. Conventional assessments apply fixed indicator weights and thresholds derived from a single design event, although the factors controlling inundation and their response positions can shift with rainfall intensity. This study examines 208 urban green-space plots within Shenyang’s Third Ring Road, combining plot-level field sampling, GIS-based spatial indicators, and maximum inundation depths simulated for 1-, 3-, 10-, 30-, and 50-year return periods. Four analyses were applied along a common scenario axis: indicator screening defined the factor set; multi-group structural equation modeling characterized shifts in latent-variable paths; random forests with permutation importance ranked observed factors; partial dependence analysis located response breakpoints; and XGBoost-SHAP described sample-level contributions. All predictive results are reported from 30 repeated train–test splits with uncertainty intervals. Screening reduced 59 candidate indicators to 13 key factors. The diagnostic evidence indicates a mechanism-shift interval between the 10- and 30-year return periods: the structural path from spatial-pattern resistance to inundation becomes significant from the 10-year scenario, the total effect of surface-cover pressure first exceeds that of vertical storage capacity in the same scenario group, and the response breakpoint of impervious surface ratio tightens from about 58% to about 33% within this interval. Predictive importance is dominated by grey infrastructure density, green infrastructure density, and catchment radius in all scenarios. Renewal priorities therefore vary with rainfall intensity: internal storage and infiltration under regular rainfall, joint surface-cover and catchment control within the shift interval, and coordinated grey–green system responses under extreme scenarios. These findings are diagnostic and predictive, and retrofit benefits require experimental or quasi-experimental validation. The framework offers a methodological reference for further local calibration in comparable cold-region old industrial cities.