Cascade Hybrid-Deep Learning Driven Multi-Factor Sensitive Landslide Prediction System

Abstract

In recent years, the world has witnessed exceedingly high dynamism in nature’s behavior including tsunami, landslide, flood and fire. More specifically, the events of landslide have become frequent globally, thus claiming thousands of innocents lives every year. Landslide prediction requires understanding varied factors like geological, geomorphometric, soil, environmental and precipitation features for the specific geography, eventually becomes a complex task. The increasing severity of landslide has alarmed academia-industry to develop robust and reliable landslide susceptibility mapping and allied prediction solution. Though, advanced software computing, satellite technology and data analytics has broadened the horizon for landslide prediction; however, it requires optimality of features as well as the learning environment. In the past, numerous approaches are developed by using precipitation features like rainfall for landslide prediction; however, merely applying standalone feature can’t ensure reliability of the system, especially when global climate pattern is undergoing decisive change due to global warming. Considering these factors, this paper proposes a robust hybrid-deep learning driven multi-factor sensitive landslide prediction system. The proposed model applies multiple landslides influencing factors like geological, geomorphometric, soil, environmental and precipitation to train a strategically designed hybrid deep learning model for landslide prediction. To ensure computational efficacy, the landslide inventory factors were processed for significant predictor test, which helped retaining the optimal set of influencing factors collected from the GIS benchmarks. To train the selected features optimally for learning and prediction, a cascade architecture was used, with Long Short-Term Memory (LSTM) networks retrieving local features and Bidirectional-LSTM (Bi-LSTM) networks retaining long-term dependency features. The suggested model achieves landslide prediction accuracy of 96.38%, precision of 95.19%, recall of 95.33%, and F-Measure of 95.56%, according to the simulation data. The suggested model is robust towards real-time applications, as confirmed by the mean average error of 0.1109 and the root mean square error value of 0.1860.