Predictive Modeling of Landslide Susceptibility in Soft Soil Canal Regions: A Focus on Early Warning Systems

Abstract

The Mekong Delta (MD) has suffered significant losses in land resources, economic damage, and human and property casualties due to recent landslides. An early warning system for landslides is a valuable tool for identifying the effectiveness and timely detection of changes in the soil to promptly determine solutions and minimize damage caused by landslides in an area. In this study, we apply a machine learning approach based on the Long Short-Term Memory (LSTM) algorithm to experiment with early warning of landslide events on soft soil in the MD. Horizontal pressure, the change in inclination angles of the sensor pile due to the soil mass sliding in both the x and y directions, and the warning levels are determined based on the deformation and displacement of the soil along the riverbank, considered candidate factors for inputs in the model. Data from the established sensor system is used to train the model, creating a training and testing dataset of 374,415 samples. The accuracy of the detection and classification threshold of the system is proposed to be measured using the average F1 score derived from precision and recall values. The optimal prediction results are gleaned from an observational window of 4 minutes and 30 seconds to project roughly 2 hours into the future. The validation process resulted in recall, precision, and F1-score stands at 0.8232 with a remarkably low standard deviation of about 1%. The successful application of this research can help identify abnormal events leading to riverbank landslides due to loading, thereby creating conditions for developing a reliable information system to provide managers with the ability to suggest timely solutions to protect the lives, property of residents and infrastructures.