Dynamic landslide susceptibility for extreme rainfall events using an optimized convolutional neural network approach

The development of machine learning (ML) and deep learning (DL) models for landslide susceptibility remains challenging in part due to the limited availability of comprehensive slope failure records, shortage of real-time soil moisture, and saturation measurements, uncertainties in precipitation data accuracy, and reliable information on landscape characteristics. In the aftermath of the impact of Hurricane Maria on Puerto Rico in 2017, a comprehensive landslide database was compiled, cataloging more than 70,000 slope failures along with high-resolution landscape information, remotely sensed real-time soil moisture and soil saturation data, and accurate precipitation forecasts. These elements make Puerto Rico an ideal location to advance DL approaches for landslide susceptibility. This study introduces an optimized U-shaped convolutional neural network (CNN) with encoder-decoder architecture, incorporating attention gates, and fully connected layers to generate high-resolution landslide susceptibility maps. The parameters and shape of the model architecture, including the number of convolutional filters, kernel size, depth, activation functions, dropout layers among other parameters, were fine-tuned using a Bayesian Optimization method. The findings indicate that the proposed CNN outperforms traditional ML models, including Random Forest, Support Vector Machine, and Logistic Regression, as well as prior CNN architectures, achieving an accuracy of 0.848 and an Area Under Receiver Characteristic Curve (AUC) of 0.922. The resulting proposed trained model can be utilized to predict landslide susceptibility by using forecasted precipitation, real-time soil moisture, soil saturation, and landscape data for future extreme weather events.