Abstract: Chengguan District of Lanzhou City, located at the western margin of the Longxi Loess Plateau, is characterized by extensive surface deposits of self-weight collapsible loess. Frequent ground collapses triggered by extreme precipitation and urban pipeline leakage have posed significant threats to urban safety in recent years. Supported by the Lanzhou Urban Geological Borehole Database Construction Project, this study analyzed 1,076 boreholes to delineate the distribution of collapsible loess within the 67.92 km² built-up area. Five key influencing factors were identified: loess thickness, collapsibility grade, collapsible depth, collapsibility type, and loess origin. An Analytic Hierarchy Process (AHP)-GIS spatial overlay model was developed to assess water-induced collapse susceptibility, classifying the study area into high-, moderate-, and low-susceptibility zones. Results indicate that high-susceptibility zones (12.86 km², 15.37%) are concentrated in Class II or higher terraces along the northern and southern mountain fronts (e.g., Dashaping and Fulongping), featuring thick loess layers (＞10 m) and high collapsibility grades (III–IV). Moderate-susceptibility zones (40.87 km², 48.86%) surround high-risk areas with collapsibility grades II–III, while low-susceptibility zones (29.92 km², 35.77%) occupy the Yellow River valley and Class I terraces with weaker collapsibility. The study reveals a "hidden seepage-progressive collapse" mechanism and proposes a refined zoning methodology integrating borehole data and GIS, overcoming limitations of small-sample analyses and achieving meter-level spatial characterization of collapse risks. These findings provide critical insights for disaster mitigation planning and pipeline network optimization in Chengguan District, offering a replicable framework for loess plateau cities. Immediate measures include large-scale geological surveys and pipeline leakage detection, while long-term strategies advocate smart utility tunnel systems to address collapse risks at their source.