Enhancing landslide management with hyper-tuned machine learning and deep learning models: Predicting susceptibility and analyzing sensitivity and uncertainty

Abstract

Introduction: Natural hazards such as landslides and floods have caused significant damage to properties, natural resources, and human lives. The increased anthropogenic activities in weak geological areas have led to a rise in the frequency of landslides, making landslide management an urgent task to minimize the negative impact. This study aimed to use hyper-tuned machine learning and deep learning algorithms to predict landslide susceptibility model (LSM) and provide sensitivity and uncertainty analysis in Aqabat Al-Sulbat Asir region of Saudi Arabia. Methods: Random forest (RF) was used as the machine learning model, while deep neural network (DNN) was used as the deep learning model. The models were hyper-tuned using the grid search technique, and the best hypertuned models were used for predicting LSM. The generated models were validated using receiver operating characteristics (ROC), F1 and F2 scores, gini value, and precision and recall curve. The DNN based sensitivity and uncertainty analysis was conducted to analyze the influence and uncertainty of the parameters to the landslide. Results: Results showed that the RF and DNN models predicted 35.1–41.32 and 15.14–16.2 km2 areas as high and very high landslide susceptibility zones, respectively. The area under the curve (AUC) of ROC curve showed that the LSM by the DNN model achieved 0.96 of AUC, while the LSM by RF model achieved 0.93 of AUC. The sensitivity analysis results showed that rainfall had the highest sensitivity to the landslide, followed by Topographic Wetness Index (TWI), curvature, slope, soil texture, and lineament density. Discussion: Road density and geology map had the highest uncertainty to the landslide prediction. This study may be helpful to the authorities and stakeholders in proposing management plans for landslides by considering potential areas for landslide and sensitive parameters.