An integrated framework for landslide risk management using digital twin, AI, and crowdsourcing technologies

Abstract

Landslides are among the most significant natural hazards and pose considerable risks to mountainous regions and environmentally sensitive areas. Therefore, predicting susceptible areas and designing monitoring systems capable of rapidly detecting changes and issuing timely warnings play an important role in reducing their impact. Despite the advances achieved through remote sensing and machine learning in identifying landslide hazards, several challenges remain, including selecting the most important factors affecting landslide occurrence, integrating diverse data sources into the modeling process, interpreting modeling and prediction results, managing uncertainty, and making use of geological and geotechnical field data. Accordingly, this research presents a framework for the assessment and dynamic modeling of landslide susceptibility and environmental scenarios for adaptive prediction. The first part of the research focuses on improving landslide susceptibility modeling using multi-source remote sensing data and feature selection. In this section, three algorithms, Random Forest, Convolutional Neural Network, and XGBoost, were compared within a three-stage framework. First, baseline models were trained using geographic and meteorological data. Then, soil-moisture-related data such as Sentinel-1 and land surface temperature were added to the models. Finally, feature selection was performed to identify the most important variables. The results showed that the Random Forest model achieved the best performance in all stages in terms of accuracy, F1-score, and Receiver Operating Characteristic-Area Under the Curve (ROC-AUC), and that incorporating remote sensing data and feature selection further improved its performance. In the second part, a geo-digital twin was developed to support landslide susceptibility assessment and evaluate environmental change scenarios. The results demonstrated that vegetation loss can substantially alter landslide susceptibility, but its effects were spatially heterogeneous and strongly dependent on initial terrain and environmental conditions. Consequently, combining landslide susceptibility modeling with scenario analysis provides more valuable insights than static susceptibility modeling alone. These findings highlight the strong potential of leveraging geo-digital twins as multi-purpose decision-support tools for investigating environmental changes in landslide-prone areas. In the third part, prediction uncertainty was incorporated, and a framework was established for crowdsourced field observations. The results showed that uncertainty information can be used to identify areas where additional observations are most valuable for improving model reliability. A priority-based field data collection strategy yielded 53 observations within one month, 21 of which were independently verified using spatial, imagery, and environmental data. The integration of these observations reduced prediction uncertainty in the monitored areas by an average of 4%. Furthermore, the developed framework demonstrated the potential to support near-real-time landslide warning by combining spatial susceptibility information with environmental threshold conditions. Overall, this dissertation demonstrates a transition from machine learning and remote sensing methods to a geo-digital twin and subsequently to the development of an uncertainty-based monitoring and early warning system. The main innovation of this research lies in integrating machine learning, scenario-based assessment, targeted crowdsourcing, and sensor-based monitoring within a single framework, thereby improving predictive accuracy, interpretability, and decision-making for landslide risk reduction.