Machine Learning Algorithms for Landslide Susceptibility Mapping applied on Portofino Promontory (Italy)

Rainfall-triggered landslide susceptibility maps (LSM) identify landslide-prone areas, providing valuable support to decision-makers in defining risk reduction strategies. They provide relevant data for spatial planning, preventing urban development in high-risk areas, and enhancing early warning systems. The present research is part of the RAISE project, funded by the Ministry of University and Research (MUR) under the National Recovery and Resilience Plan (PNRR). The main objective is to identify the most reliable machine learning approach to build landslide susceptibility maps in the Liguria Region (Italy); in particular the present research aims at comparing the landslide susceptibility map based on a heuristic approach with ones elaborated using different machine learning algorithms. The methodology is applied to the Portofino Promontory (Liguria Italy) where widespread shallow landslides are triggered due to the occurrence of intense and short-duration rainfall events. In the investigated area, the susceptibility map elaborated by other authors by means of the semi-quantitative analytical hierarchy process (AHP) method is assumed as benchmark. Such heuristic approach utilizes nine natural and anthropogenic landslides conditioning factors: lithology, slope aspect and acclivity, land use, terraced landscape, distances to hydrographical networks, distances to man-made cuts elements, distance to man-made structures and existing gravitational processes. A 110-year inventory of rainfall-triggered landslides, including 102 landslide events, identified by previous studies, have been considered. The proposed methodology assumes the same conditioning factors as well as the same landslide inventory. Since landslides involved limited areas, the number of landslide samples represents the minority class being significantly lower than the number of non-landslide sites (majority class) thus determining highly imbalanced dataset that affect the performance of machine learning algorithms adversely. In this framework, different techniques of under sampling/oversampling (e.g. Random Under Sampling, SMOTE) have been tested and compared to build a balanced dataset. In the balanced dataset, 70% of the samples are used for training and 30% for testing the classification algorithms. Several well-known metrics, such as recall, precision, accuracy are used to evaluate the performance of landslide susceptibility algorithms. The calibrated algorithms have been applied to the investigated area providing for each pixel a probability of landslides occurrence, then categorized into 5 susceptibility classes (very low, low, medium, high, very high). Consistently with the AHP method, natural breaks classification method is used to classify the susceptibility. One limitation of the use of machine learning algorithms for decision support tools, lies in their interpretability, that is the ability to understand the functioning of the model. Therefore, in the present research, the feature importance analysis is performed to quantify the relevance of each conditioning factor to the landslide susceptibility assessment, contributing to enhance machine learning model interpretability. The adopted machine learning algorithms reveal suitable to build landslide susceptibility map for the area of concern; further applications to different sites affected by rainfall-trigged landslides need to be carried on to define a reliable landslide susceptibility map at the regional scale.