Critical threshold mining of landslide deformation and intelligent early-warning methods based on multi-factor fusion

Deformation and development of landslides are highly complex processes and relying solely on a single monitoring factor is insufficient to accurately assess the entire evolutionary trend of landslide deformation. Typically, multiple sensors should be deployed on a landslide mass to obtain multidimensional monitoring data and comprehensively analyze the landslide development process. Monitoring data obtained through multiple sensors exhibit certain randomness and redundancy. Effective processing of these data is the key to landslide warning systems. However, the deployment of various sensors on landslide masses incurs significant costs, which usually limits their application to key landslide-prone points rather than enabling landslide warning systems over large areas. To construct a low-cost and widely applicable landslide warning model, this study installed two types of conventional monitoring devices on a landslide mass (displacement meters and rain gauges). First, the Saito method was applied to identify the macroscopic deformation stages of the landslide and to calculate the daily average deformation rates at each stage. Subsequently, a five-level warning pattern based on deformation rates was established and the critical thresholds for each warning level was determined. Finally, using daily displacement and rainfall as well as bedrock hardness and slope as the factors, a feature vector set was constructed by associating the warning levels corresponding to the daily average deformation rates at each stage. An integrated machine learning network was employed to develop an intelligent landslide warning model, which enables the intelligent identification of landslide deformation stages and assists in decision-making for landslide warning over large areas.