Landslide susceptibility mapping using extended random forest architecture with PS‐InSAR‐based optimisation

Abstract High‐elevation regions are prone to severe natural disasters, particularly landslides, which pose significant risks to both infrastructure and human life. Understanding and predicting landslide susceptibility is crucial for effective disaster management and mitigation efforts. This study aims to enhance landslide susceptibility mapping by developing and comparing the performance of an extended random forest (ERF) architecture with that of an artificial neural network (ANN). The primary goal is to enhance prediction accuracy and reliability by applying advanced machine learning (ML) techniques and integrating data. We employed a comprehensive dataset comprising 140 landslides and nine major landslide conditioning factors. The dataset was partitioned into training (70%) and testing (30%). Feature selection was conducted using recursive feature elimination (RFE) to identify the most influential variables. The ERF architecture integrated predictions from five robust ML algorithms using appropriate weighting to optimise ensemble performance. The proposed models demonstrated high robustness, with the ERF achieving a prediction accuracy of 0.97, significantly outperforming the ANN (0.91). The hybrid ensemble architecture showed superior performance because of its robust training phase and integration of multiple strong learners. Validation was performed using persistent scatterer interferometry (PS‐InSAR) from 2023 to assess deformation. PS‐InSAR validation revealed significant deformation velocities, confirming the model's ability to identify high‐susceptibility zones. These findings contribute to the field by providing a robust framework for landslide prediction and mitigation, particularly in high‐risk regions.