Abstract
Landslide susceptibility mapping is a crucial initial step in risk mitigation strategies. Landslide hazards are widely spread all over the world and, as such, mapping the relevant susceptibility levels is in constant research and development. As a result, numerous modelling techniques and approaches have been adopted by scholars, implementing these models at different scales and with different terrains, in search of the best-performing strategy. Nevertheless, a direct comparison is not possible unless the strategies are implemented under the same environmental conditions and scenarios. The aim of this work is to implement three statistical-based models (Statistical Index, Logistic Regression, and Random Forest) at the basin scale, using various scenarios for the input datasets (terrain variables), training samples and ratios, and validation metrics. A reassessment of the original input data was carried out to improve the model performance. In total, 79 maps were obtained using different combinations with some highly satisfactory outcomes and others that are barely acceptable. Random Forest achieved the highest scores in most of the cases, proving to be a reliable modelling approach. While Statistical Index passes the evaluation tests, most of the resulting maps were considered unreliable. This research highlighted the importance of a complete and up-to-date landslide inventory, the knowledge of local conditions, as well as the pre- and post-analysis evaluation of the input and output combinations.
Funding Information
  • Politecnico di Milano

This publication has 43 references indexed in Scilit: