A combined method was developed to forecast the spatial and the temporal probability of occurrence of rainfall-induced shallow landslides over large areas. The method also allowed to estimate the dynamic change of this probability during a rainfall event. The model, developed through a data-driven approach basing on Multivariate Adaptive Regression Splines technique, was based on a joint probability between the spatial probability of occurrence (susceptibility) and the temporal one. The former was estimated on the basis of geological, geomorphological, and hydrological predictors. The latter was assessed considering short-term cumulative rainfall, antecedent rainfall, soil hydrological conditions, expressed as soil saturation degree, and bedrock geology. The predictive capability of the methodology was tested for past triggering events of shallow landslides occurred in representative catchments of Oltrepò Pavese, in northern Italian Apennines. The method provided excellently to outstanding performance for both the really unstable hillslopes (area under ROC curve until 0.92, true positives until 98.8%, true negatives higher than 80%) and the identification of the triggering time (area under ROC curve of 0.98, true positives of 96.2%, true negatives of 94.6%). The developed methodology allowed us to obtain feasible results using satellite-based rainfall products and data acquired by field rain gauges. Advantages and weak points of the method, in comparison also with traditional approaches for the forecast of shallow landslides, were also provided.

Rainfall thresholds define the conditions leading to the triggering of shallow landslides over wide areas. They can be empirical, which exploit past rainfall data and landslide inventories, or physicallybased, which integrate slope physical–hydrological modeling and stability analyses. In this work, a comparison between these two types of thresholds was performed, using data acquired in Oltrepò Pavese (Northern Italian Apennines), to evaluate their reliability. Empirical thresholds were reconstructed based on rainfalls and landslides triggering events collected from 2000 to 2018. The same rainfall events were implemented in a physicallybased model of a representative testsite, considering different antecedent pore-water pressures, chosen according to the analysis of hydrological monitoring data. Thresholds validation was performed, using an external dataset (August 1992–August 1997). Soil hydrological conditions have a primary role on predisposing or preventing slope failures. In Oltrepò Pavese area, cold and wet months are the most susceptible periods, due to the permanence of saturated or close-to-saturation soil conditions. The lower the pore-water pressure is at the beginning of an event, the higher the amount of rain required to trigger shallow failures is. physicallybased thresholds provide a better reliability in discriminating the events which could or could not trigger slope failures than empirical thresholds. The latter provide a significant number of false positives, due to neglecting the antecedent soil hydrological conditions. These results represent a fundamental basis for the choice of the best thresholds to be implemented in a reliable earlywarning system.

Landslides cause severe damage to the road network of the hit zone, in terms of both direct (partial or complete destruction of a road or blockages) and indirect (traffic restriction or the cut-off of a certain area) costs. Thus, the identification of the parts of the road network that are more susceptible to landslides is fundamental to reduce the risk to the population potentially exposed and the financial expense caused by the damage. For these reasons, this paper aimed to develop and test a data-driven model for the identification of road sectors that are susceptible to being hit by shallow landslides triggered in slopes upstream from the infrastructure. This model was based on the Generalized Additive Method, where the function relating predictors and response variable is an empirically fitted smooth function that allows fitting the data in the more likely functional form, considering also non-linear relations. This work also analyzed the importance, on the estimation of the susceptibility, of considering or not the sediment connectivity, which influences the path and the travel distance of the materials mobilized by a slope failure until hitting a potential barrier such as a road. The study was carried out in a catchment of northeastern Oltrepò Pavese (northern Italy), where several shallow landslides affected roads in the last 8 years. The most significant explanatory variables were selected by a random partition of the available dataset in two parts (training and test subsets), 100 times according to a bootstrap procedure. These variables (selected 80 times by the bootstrap procedure) were used to build the final susceptibility model, the accuracy of which was estimated through a 100-fold repetition of the holdout method for regression, based on the training and test sets created through the 100 bootstrap model selection. The presented methodology allows the identification, in a robust and reliable way, of the most susceptible road sectors that could be hit by sediments delivered by landslides. The best predictive capability was obtained using a model in which the index of connectivity was also calculated according to a linear relationship, was considered. Most susceptible road traits resulted to be located below steep slopes with a limited height (lower than 50 m), where sediment connectivity is high. Different land use scenarios were considered in order to estimate possible changes in road susceptibility. Land use classes of the study area were characterized by similar connectivity features. As a consequence, variations on the susceptibility of the road network according to different scenarios of distribution of land cover were limited. The results of this research demonstrate the ability of the developed methodology in the assessment of susceptible roads. This could give the managers of infrastructure information about the criticality of the different road traits, thereby allowing attention and economic budgets to be shifted towards the most critical assets, where structural and non-structural mitigation measures could be implemented.

Rainfall-induced shallow landslides are common phenomena in many parts of the world, affecting cultivation and infrastructure and sometimes causing human losses. Assessing the triggering zones of shallow landslides is fundamental for land planning at different scales. This work defines a reliable methodology to extend a slope stability analysis from the site-specific to local scale by using a well-established physically based model (TRIGRS-unsaturated). The model is initially applied to a sample slope and then to the surrounding 13.4 km2 area in Oltrepò Pavese (northern Italy). To obtain more reliable input data for the model, long-term hydro-meteorological monitoring has been carried out at the sample slope, which has been assumed to be representative of the study area. Field measurements identified the triggering mechanism of shallow failures and were used to verify the reliability of the model to obtain pore water pressure trends consistent with those measured during the monitoring activity. In this way, more reliable trends have been modelled for past landslide events, such as the April 2009 event that was assumed as a benchmark. The assessment of shallow landslide triggering zones obtained using TRIGRS-unsaturated for the benchmark event appears good for both the monitored slope and the whole study area, with better results when a pedological instead of geological zoning is considered at the regional scale. The sensitivity analyses of the influence of the soil input data show that the mean values of the soil properties give the best results in terms of the ratio between the true positive and false positive rates. The scheme followed in this work allows us to obtain better results in the assessment of shallow landslide triggering areas in terms of the reduction in the overestimation of unstable zones with respect to other distributed models applied in the past.