• Cohen's kappa index
• data-driven methods
• generalized additive model
• Liguria
• likelihood ratio model
• Lombardy
• northern Apennines
• physically based models
• prediction rate curve
• predisposing factors
• rainfall-induced shallow landslides
• receiver operating characteristic curve
• shallow landslide database
• shallow landslide susceptibility
• SHALSTAB

Rainfall-induced shallow landslides are one of the most serious natural hazards in hilly and mountain regions worldwide because of fast movement, difficulty of spatial-temporal prediction, lack of knowledge about triggering mechanisms and long travel distance. In the last decades, extreme meteorological events frequently occurred in the northern Apennines, provoking hundreds of shallow landslides that led to loss of lives, damages to infrastructures and buildings. In some cases, the inadequate settlement development planning, the intense urbanisation and the lack of in-depth knowledge about geological-morphological features of many territories can make such areas more vulnerable to shallow landslide triggering. Since the last 30 years, the scientific community is continuously being developed methods and techniques improvements to predict shallow landslides. However, a comprehensive evaluation of the applicability, reliability and predictive capability of different methods and validation techniques for landslide susceptibility is fairly undeveloped in the specialized scientific literature. Indeed, existing works only deal with some different methods and post-processing operations techniques for landslide susceptibility assessment that are often focused on single study area. A poor attention was often paid in the overall evaluation of the effectiveness of the tools under examination, testing their effectiveness and predictive capability in different areas, in comparison with different methods and validations techniques and also changing some input criteria (e.g. different landslide inventories, different landslide types).
The present Thesis aims to fill this gap not only by focusing on ongoing challenges in methods improvements, but mainly providing a detailed and comprehensive review of some methodological approaches and validation techniques for landslide susceptibility modelling, through their application in different areas. Thus, different study cases concerning some aspects of landslide susceptibility assessment were presented inside the Thesis, whose main objectives consists in: i) the evaluation of role played by some morphological, geological and land use predisposing factors in shallow landslide source areas distribution; ii) the development of a data-driven methodology based on GAM (Generalized Additive Model) easily applicable in various environmental contexts; iii) the implementation a of a bivariate methodology (LR; Likelihood Ratio) from a proprietary software into a free software developing a detailed and reliable procedure; iv) the comparison of the statistical procedures developed and reviewed with a physically based method (SHALSTAB; Shallow Landslide Stability Model), as well as the comparison between different validation techniques (PRCs, Prediction Rate Curves; ROCs, Receiver Operating curves). All these analyses carried out referred to some basins of the northern Apennines recently involved by rainfall-induced shallow landslides. The research was not planned with the only goal to broaden knowledge about landslide susceptibility of northern-west Italy settings, but also to provide a detailed evaluation of some aspects connected to landslide susceptibility assess from the methodological standpoint. Study cases were mainly exploited with the aim to promote a review of methods and techniques for pure research purpose, where the computer processing have been the main tools to reach this goal.
The main results arising from the study cases were that geology, curvature and land use are ones of most significant predisposing variables in landslide occurrence. In particular, the health condition of woodlands and the degree of maintenance of agricultural areas have a great influence on the landslide distribution. The statistical procedure based on the GAM method was characterized by a good predictive capability and reliability in the investigated areas. Overall, statistical methods are more effective to predict future shallow landslides, as they showed a higher predictive capability and a greater reliability of outcomes than the SHALSTAB method. However, slight better results were found for the LR method in the Pogliaschina T. basin, thus suggesting that this more simple method can reach high performance as well as the more complex GAM method. Statistical models built with fewer number of predisposing factors are adequate to modelling landslide susceptibly, since the adding of more variables was not capable to raise significantly the model performance. The PRC and ROC curves validation techniques provide very similar results, indicating that both can be equally used in model validation operations.
This Thesis can represent a step forward in the evaluation of some aspects related to landslide susceptibility analyses, both for furnishing significant outcomes that can be used by decision-makers for a more appropriate planning strategies of the investigated areas and for providing valid and reliable statistical procedures that can be applied everywhere to predict shallow landslide-prone areas.