Abstract
Transferable, quantitative frameworks for assessing freeze–thaw hazards (FTHs) along engineering corridors remain limited in warming marginal permafrost regions, particularly those integrating hydroclimatic forcing, terrain controls, and anthropogenic disturbance. Here, we developed and validated an integrated machine-learning (ML)–GeoDetector framework, supported by field surveys and electrical resistivity tomography (ERT), for the Jagdaqi–Mo'he section of National Highway G111 in the Xing'an Permafrost (XAP) region of Northeast China, a representative thermally transitional discontinuous permafrost environment. Field investigations identified six major FTH types, among which thermokarst lakes and uneven pavement settlement were the most widespread, with hazard clusters showing strong spatial heterogeneity concentrated where permafrost degradation and engineering disturbance converge. Among five ML classifiers, the random forest (RF) model achieved the highest area under the curve (AUC = 0.93) and overall accuracy (OA = 0.86). Risk mapping shows that 23.7% of predicted FTH points fall within high-risk zones, primarily where uneven pavement settlement coincides with pavement cracks development. Annual precipitation was identified as the dominant driver by both RF importance ranking (17.3%) and GeoDetector analysis (q = 0.54), and its interaction with slope angle yielded the highest explanatory power (q = 0.63), indicating that FTH hotspots are driven by hydroclimatic–topographic coupling rather than single-factor extremes. ERT profiles further show that similar near-surface thermal conditions can conceal markedly different subsurface permafrost structures, emphasizing the need for geophysical constraints in risk interpretation. The proposed RF−GeoDetector framework provides a transferable approach for corridor-scale FTH risk mapping and adaptive infrastructure management in warming marginal permafrost regions globally.