Hybrid ensembled machine learning algorithm for landslide susceptibility assessment and data enhancement via Dual Discriminator Conditional Generative Adversarial Network

Landslide susceptibility mapping (LSM) is a critical component in geohazard assessment and risk management for disaster prevention and mitigation strategies. However, incomplete landslide datasets, landslide data scarcity and imbalance, and the accuracy of susceptibility mapping are challenging problems that were tackled via different novel techniques in this study. We aimed to develop a comprehensive landslide inventory, generate synthetic data, and employ machine learning for hybrid modeling. The methodology was implemented in three phases. The Small Baseline Subset (SBAS) and Persistent Scatterer Interferometry (PSInSAR) techniques were applied to design landslide inventory in the first phase. Secondly, the Dual Discriminator Conditional Generative Adversarial Network (DDCGAN) was trained on real landslide data to produce synthetic samples resembling the original data. In the final phase, a Hybrid Ensemble Machine Learning Model (HBEM) combining XGBoost, AdaBoost, and LogitBoost was trained on an enhanced dataset to forecast landslide susceptibility along the Gilgit-Skardu road in Northern Pakistan. The study compiled a spatial database with 12 predisposing parameters and 152 mapped landslides, correlatedly analyzed using a frequency ratio approach. The SBAS InSAR detected 76 landslides, 38 with PSInSAR, and mapped 38 additional rapid rockfalls and rainfall-induced landslides using high-resolution imagery, validated against historical data. The Random Forest (RF) model, with an accuracy of 0.986, was compared to HBEM. Our model, with 400 synthetic samples from DDCGAN, achieved a 0.996 AUC, laying the groundwork for future early warning systems to reduce risks and safeguard public safety. © 2025 Elsevier B.V., All rights reserved.