Road collapse is a frequent and damaging disaster in cities. The complexity and uncertainty inherent in urban environments pose significant challenges to mitigating road collapses. This paper presents a novel framework integrating machine learning-based susceptibility assessment and geophysical detection validation for urban road collapse risk reduction. Three oversampling techniques, random oversampling, synthetic minority oversampling technique for nominal and continuous features (SMOTENC), and adaptive synthetic sampling (ADASYN), are first utilized to implement data augmentation on urban road collapse accident samples. Subsequently, three machine learning models, support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost), are developed to evaluate road collapse susceptibility by extracting collapse-inducing patterns from historical accident data. Particularly, on-site geophysical hazard detection is conducted to validate the assessment results. The results demonstrate that XGBoost with SMOTENC achieves satisfactory performance in identifying road collapses with accuracy (0.9608) and AUC (0.9796). The spatial distribution of road collapse susceptibility in Shanghai central area follows a high-moderate-low pattern from northwest to southeast. The geophysical detection reveals a correlation between higher road collapse susceptibility and increased severity of underground diseases, validating the generalization capacity of XGBoost in actual operational environments. Additionally, the structural problems of underground pipelines are identified as the most influential factors for urban road collapse. This research offers valuable insights for urban road collapse mitigation and resilience improvement of transportation infrastructure.