Landslide Susceptibility Assessment by Ensemble-Based Machine Learning Models

Abstract

Slope failures are among the most hazardous natural disasters, causing severe damage to public and private properties. Casualties owing to landslides have been growing in many areas of the world, especially since the increase of climate changes and precipitations. To this, decision-makers need trustworthy information that may be employed to decide the spatial solution plans to protect people. Statistical landslide susceptibility mapping is facing a constant evolution, especially since the introduction of Machine Learning algorithms (ML). A new methodology is here presented, based on the ensemble of Artificial Neural Network, Generalized Boosting Model and Maximum Entropy ML algorithms. Such an approach has been used in Cinque Terre National Park (Northern Italy), severely affected over the years by landslides, following precipitation events, causing extensive damage in a World Heritage Site. Nine predisposing factors were selected and assessed according to the knowledge of the territory, including slope angle, aspect angle, planform curvature, profile curvature, distance to roads, distance to streams, agricultural terraces state of activity, land use and geological information, whilst a database made of ca. 400 landslides was used as input. Four different Ensemble techniques were applied, after the averaging of 150 stand-alone methods, each one providing validation scores such as ROC/AUC curve. Therefore, the results obtained through Ensemble modeling showed improved values, confirming the reliability and the suitability of the proposed approach for decision-makers in land management at local and regional scales.