Landslide susceptibility assessment using machine learning models is a popular and consolidated approach worldwide. The main constraint of susceptibility maps is that they are not adequate for temporal assessments: they are generated from static predisposing factors, allowing only a spatial prediction of landslides. Recently, some methodologies have been proposed to provide spatiotemporal landslides prediction starting from machine learning algorithms (e.g., combining susceptibility maps with rainfall thresholds), but the attempt to obtain a dynamic landslide probability map directly by applying machine learning models is still in the preliminary phase. This work provides a contribution to fix this gap, combining in a Random Forest (RF) algorithm a static indicator of the spatial probability of landslide occurrence (i.e., a classical susceptibility index) and a number of dynamic variables (i.e., seasonality and the rainfall amount cumulated over different reference periods). The RF implementation used in this work allows the calculation of the Out-of-Bag Error and depicts Partial Dependence Plots, two indices that were used to quantify the variables’ importance and to comprehend if the model outcomes are consistent with the triggering mechanism observed in the case of study (Metropolitan City of Florence, Italy). The goal of this research is not to set up a landslide probability map, but to 1) understand how to populate training and test datasets with observations sampled over space and time, 2) assess which rainfall variables are statistically more relevant for the identification of the time and location of landslides, and 3) test the dynamic application of RF in a forecasting model for the spatiotemporal prediction of landslides. The proposed dynamic methodology shows encouraging results, consistent with the actual knowledge of the physical mechanism of the triggering of shallow landslides (mainly influenced by short and intense rainfalls) and identifies some benchmark configurations that represents a promising starting point for future regional-scale applications of machine learning models to dynamic landslide probability assessment and early warning.

Central Asia is an area characterized by complex tectonics and active deformation; the related seismic activity controls the earthquake hazard level that, due to the occurrence of secondary and tertiary effects, also has direct implications for the hazard related to mass movements such as landslides, which are responsible for an extensive number of casualties every year. Climatically, this region is characterized by strong rainfall gradient contrasts due to the diversity of climate and vegetation zones. The region is drained by large, partly snow- and glacier-fed rivers that cross or terminate in arid forelands; therefore, it is also affected by a significant river flood hazard, mainly in spring and summer seasons. The challenge posed by the combination of different hazards can only be tackled by considering a multi-hazard approach harmonized among the different countries, in agreement with the requirements of the Sendai Framework for Disaster Risk Reduction. This work was carried out within the framework of the Strengthening Financial Resilience and Accelerating Risk Reduction in Central Asia (SFRARR) project as part of a multi-hazard approach and is focused on the first landslide susceptibility analysis at a regional scale for Central Asia. To this aim the most detailed landslide inventories, covering both national and transboundary territories, were implemented in a random forest model, together with several independent variables. The proposed approach represents an innovation in terms of resolution (from 30 to 70 m) and extension of the analyzed area with respect to previous regional landslide susceptibility and hazard zonation models applied in Central Asia. The final aim was to provide a useful tool for land use planning and risk reduction strategies for landslide scientists, practitioners, and administrators.

The use of machine learning models for landslide susceptibility mapping is widespread but limited to spatial prediction. The potential of employing these techniques in spatiotemporal landslide forecasting remains largely unexplored. To address this gap, this study introduces an innovative dynamic (i.e., space–time-dependent) application of the random forest algorithm for evaluating landslide hazard (i.e., spatiotemporal probability of landslide occurrence). An area in Norway has been chosen as the case study because of the availability of a comprehensive, spatially, and temporally explicit rainfall-induced landslide inventory. The applied methodology is based on the inclusion of dynamic variables, such as cumulative rainfall, snowmelt, and their seasonal variability, as model inputs, together with traditional static parameters such as lithology and morphologic attributes. In this study, the variables’ importance was assessed and used to interpret the model decisions and to verify that they align with the physical mechanism responsible for landslide triggering. The algorithm, once trained and tested against landslide and non-landslide data sampled over space and time, produced a model predictor that was subsequently applied to the entire study area at different times: before, during, and after specific landslide events. For each selected day, a specific and space–time-dependent landslide hazard map was generated, then validated against field data. This study overcomes the traditional static applications of machine learning and demonstrates the applicability of a novel model aimed at spatiotemporal landslide probability assessment, with perspectives of applications to early warning systems.

This study proposes an innovative approach to develop a regional-scale landslide forecasting model based on rainfall thresholds optimized for operational early warning. In particular, it addresses two main issues that usually hinder the operational implementation of this kind of models: (i) the excessive number of false alarms, resulting in civil protection system activation without any real need, and (ii) the validation procedure, usually performed over periods too short to guarantee model reliability. To overcome these limitations, several techniques for reducing the number of false alarms were applied in this study, and a multiple validation phase was conducted using data from different sources. An intensity-duration threshold system for each of the five alert zones composing the Liguria region (Italy) was identified using a semiautomatic procedure called MaCumBA, considering three levels of criticality: low, moderate, and high. The thresholds were developed using a landslide inventory collected from online newspapers by a data mining technique called SECaGN. This method was chosen to account for only those events that echo on the Internet and therefore impact society, ignoring landslides occurred in remote areas, not of interest for civil protection intervention, which would adversely affect the model performance because they would result in false alarms. A calibration phase was performed to minimize the impact of false alarms, allowing at least one false alarm per year over the moderate criticality level. In addition, an innovative approach to include antecedent rainfall as the third dimension of the intensity-duration thresholds was applied, generating a consistent reduction in false alarms. The results were validated through an independent landslide inventory and were compared with (i) the alert issued by the regional civil protection agency to observe the improvements achieved with the proposed model and to evaluate to what extent the proposed model is consistent with the assessments of the civil protection and (ii) a dataset of the national states of emergency to verify the suitability of the developed thresholds for alerting citizens. The thresholds obtained showed high predictive capabilities, confirming their suitability for implementation in an operational landslide early warning system.

We conducted a systematic literature review of 105 landslide susceptibility studies in Italy from 1980 to 2023, retrieved from the Scopus database. We discovered that Italian researchers primarily focus on rainfall-induced landslides (86.67% of the articles), especially shallow and fast movements (60%), with 72% of studies conducted at the local scale, while regional and national-level studies are rare. The most common data sources include remote sensing images validated by field surveys and official data portals at the national or regional level. Data splitting usually follows a 70:30 ratio and 24 modelling techniques were identified, with logistic regression being historically prevalent, although machine learning methods have rapidly gained popularity. Italian studies used 97 predisposing factors, with slope angle (98.09%), lithology (89.52%), land use/land cover (78.09%), and aspect (77.14%) being the most employed. This review also identifies and discusses a few less-used factors, like soil sealing, rainfall, NDVI, and proximity to faults, which showed promising results in experimental studies. Predisposing factors are generally selected by expert judgment, but methods for forward factors selection and collinearity tests are becoming more common. This review synthesizes current knowledge, pinpointing gaps, highlighting emerging methodologies, and suggesting future research directions for better integration of susceptibility studies with landslide risk management.

Regional- and national-scale landslide warning systems are usually based on rainfall thresholds that forecast the possibility of landslide occurrence over wide spatial units called alert zones (AZs). This work proposes a substantial improvement of the state-of-the-art by combining the rainfall threshold outcomes with a set of spatially explicit risk indicators aggregated at the municipality level. The combination of these two different techniques is performed by means of a dynamic matrix, which was purposely calibrated to provide an output in the form of five possible levels of risk (from R0 to R4), which are connected to the growing intensity of expected impacts and a pre-defined confidence in issuing warnings without omitting alarms. Italy (about 300,000 km²) is used as a case study, producing a set of rainfall thresholds differentiated for 150 AZs and providing a specific calibration of the dynamic risk matrix for each of them. The verification of the matrix outputs was satisfactory as no AZs experienced landslides at the R0 level; only two of them had more than 10% of landslides at the R1 level, and most of the AZs had more than 90% of the landslides in the R2 to R4 risk classes. A comparison with a nation-wide dataset of very severe hydrogeological disasters further corroborated the consistency of the model outputs with real scenarios, as most part of the impacts occurred in places and times when the matrix outputs were at the highest levels. The proposed methodology represents a reliable improvement for state-of-the-art territorial warning systems, as it brings two main advances: the spatial resolution is greatly improved, as the basic spatial unit for warning is downscaled from AZs to municipalities (whose average extension, in Italy, is about 1770 and 38 km², respectively); second, the outputs can better address the needs of landslide emergency management, as the warning are specifically addressed to small areas based on the expected impacts (since risk indicators are used in the dynamic matrices), rather than on the mere probability of landslide occurrence.