<p>Landslides are a significant geohazard that poses a major threat to lives and infrastructure in complex mountainous terrain, particularly in regions with intricate topography, such as the Chitral district in Pakistan. Landslide susceptibility mapping (LSM) is crucial for effective risk management, but traditional models often struggle with the multifaceted nature of landslide occurrence. In this study, an extensive landslide inventory comprising 329 landslides and fourteen conditioning factors was compiled. Afterwards, an integrated machine-learning (Gradient Boosting Machine, Extreme Gradient Boosting) and deep-learning framework (Convolutional Neural Network (1D, 2D), Long Short-Term Memory) for LSM and subsequent propagation analysis. Furthermore, a hybrid model was developed that couples LSTM temporal dependency capability with XGBoost robust feature handling quality for binary landslide prediction. To cross-validate the LSM, PS-InSAR was applied over three years to analyze ground deformation. The predictive capacity of conditioning factors were determined utilizing Shapley Additive Explanations (SHAP), the Information Gain Ratio (IGR), and chi-square, which identified elevation and lithology as the major contributing factors. The models were then assessed using seven evaluation metrics. The results show that the hybrid model achieved the best performance, with an AUC of 97.6%, accuracy of 92.3%, and precision of 88.7%. Landslide propagation was subsequently simulated using the empirical Flow-R model, where the hybrid-based LSM produced the largest and most dispersed runout, classifying 19.7% of the study area as high-risk. PS-InSAR surface deformation analysis was utilized as independent cross-validation and showed spatial agreement with modelled high-susceptibility zones. The integrated approach offers a practical workflow for improving landslide risk assessment and guiding resilient infrastructure planning.</p>