Abstract: In response to the urban safety concerns caused by ground settlement in soft soil zones of Zhuhai City, this research focuses on the development and optimization of a multivariate prediction model. Recognizing the limitations of traditional prediction methods in modeling nonlinear settlement behavior, a novel multivariate regression LSTM prediction model is proposed, based on the characteristics of soft soils in the region. The model fully integrates InSAR monitoring data with various nonlinear influencing factors. Ten key influencing factors, including groundwater extraction intensity, soft soil layer thickness, and compression modulus, were systematically selected. Leveraging the LSTM’s gated structure, the model successfully eliminates the reliance on the time-sensitive physical parameters and the completeness of monitoring data typical of conventional methods. The results demonstrates strong predictive performance: over 88% of errors fall within ± 5mm, and the R² coefficient of the test set reaches as high as 0.91, indicating the model’s high accuracy and reliability. Further enhancement through intelligent optimization algorithms significantly improved hyperparameter tuning and feature selection, pushing the R² above 0.98. However, the model’s performance in geologically complex or highly heterogeneous regions still depends on the integration of diverse monitoring technologies to ensure data validity and model precision. Practical application suggests that the model can be effectively used in urban planning, disaster prevention and mitigation, providing reliable land subsidence data for government agencies and experts. Its adaptive learning mechanism holds significant potential for broader application in other similar soft soil regions across the Pearl River Delta.