Ferriero, Flavia (2025) A full probabilistic approach to landslide forecasting. [Tesi di dottorato]

Documento PDF Ferriero_Flavia_37.pdf Visibile a [TBR] Amministratori dell'archivio Download (5MB) | Richiedi una copia

Abstract

Landslides are among the most destructive natural hazards, particularly in mountainous regions like southern Italy, where they frequently cause casualties and economic losses. In this context, landslide forecasting is essential to support early warning systems, operational actions, and land-use planning, especially as climate change and urban expansion increase exposure to rainfall-induced landslides, one of the most frequent and damaging types. From a scientific perspective, landslides are complex, nonlinear processes influenced by interacting geological, hydrological, and physical factors, with inherent uncertainties. Traditional deterministic models, based on threshold exceedance of triggering factors such as rainfall, provide binary predictions (landslide/no landslide) and cannot fully capture the probabilistic nature and variability of landslide occurrence, limiting their reliability for risk management. This thesis explores probabilistic approaches as a more realistic and informative framework for landslide forecasting, focusing on Bayesian methods that explicitly account for uncertainty and enable the combination of multiple information sources. The methodological strategies developed—including a Bayesian framework, the integration of information obtained from physically based models, and Bayesian logistic regression—contribute to the scientific understanding of landslide processes while developing flexible tools applicable to operational forecasting. To address these challenges, the work developed in this thesis is presented in five parts. The thesis is organized as follows. In the first part, I introduce the scientific and societal relevance of landslide forecasting, highlighting the destructive impacts of rainfall-induced landslides. Traditional deterministic approaches, such as empirical rainfall thresholds and physically based models, are limited by rigid binary predictions, strong site-specificity, and their inability to fully represent uncertainties in data and natural processes. Probabilistic approaches, particularly Bayesian methods, overcome these limitations by explicitly quantifying uncertainty, integrating diverse sources of information, and providing continuous probability estimates, offering a more realistic and informative framework for landslide forecasting. This probabilistic perspective underpins the methodological developments presented in the thesis, aimed at both improving scientific understanding and supporting better-informed decision making. The second Part begins by describing how I made a landslide catalogue for the selected study area, which includes the Sarno and Lattari mountain ranges and the Ischia Island in the Campania region, southern Italy. Through an extensive literature review and bibliographic research, I collect information of shallow, rainfall-induced landslides that have historically affected the area. The catalogue covers the period from 1950 to 2022 and reports, for each landslide, its geographical location (as an identification point), time of occurrence, and, when available, information on the landslide type. To have an insight about the completeness of the catalogue, I analyse the slope of the cumulative curve of landslide occurrences over time, using it as a loose proxy for completeness. Based on this analysis, I consider the period from 2000 to 2022 as complete, including 169 landslides. I collect daily rainfall measurements from the rain gauges of the Campania region. Each landslide is then associated with its nearest rain gauge within a planimetric distance of 5 km. Since the exact day of occurrence is known, I obtain for each rain gauge a 22-year daily time series including: the date, cumulative daily rainfall amount, and a binary variable indicating whether at least one landslide occurred on that day. Then, I use the 22-year dataset in the study area to apply a Bayesian probabilistic framework for predicting landslides, explicitly accounting for the sources of epistemic uncertainty that affect landslide occurrence. The method models the probability of landslide occurrence as a distribution, rather than a single value, allowing for a more realistic treatment of uncertainty arising from incomplete landslide inventories, variable measurements, and the inherent complexity of landslide processes. Posterior landslide probabilities are calculated for different daily rainfall thresholds using Bayes’ theorem, with prior and likelihood terms modelled as uniform and Beta distributions, respectively. Forecasts are then associated with Thiessen polygons representing the area of influence of each rain gauge. Results show that posterior probabilities increase progressively with cumulative rainfall, and no sharp physical threshold in the triggering rainfall emerges. The forecast skill improves with cumulative rainfall, as demonstrated by consistent gains in posterior over prior probabilities.This gradual trend supports the view of landslide triggering as a probabilistic process, challenging the use of deterministic rainfall thresholds in operational contexts. The proposed Bayesian probabilistic framework is generalizable to other triggers (e.g., earthquakes) and adaptable to different regions, assuming sufficient data are available. The method can provide uncertainty-informed forecasts that can support early warning systems and risk mitigation strategies. Future developments may include the incorporation of antecedent rainfall and geological conditioning factors across broader spatial and temporal scales. While the Bayesian probabilistic framework relies on observed landslide and rainfall records, its predictive capacity can be further enhanced by integrating information from physically based models. For this reason, in Part III, I present an application of the physically based model TRIGRS (Alvioli & Baum, 2016) to the Ischia Island, chosen as representative of the wider study area. I collect hourly rainfall measurements from the four rain gauges in the area. Rainfall events, identified using the CRTL-T software (Melillo et al., 2015, 2018), are used as model input, while for spatial analysis the island is partitioned into slope units (Alvioli et al., 2016) i.e., subdivisions of the terrain characterized by a certain internal geomorphological homogeneity, bounded by drainage and lines. After calibration, the model is applied to the northern flanks of Mount Vezzi and Mount Epomeo. For each rainfall event, the model identifies potentially unstable areas, which I then aggregate to define simulated landslides. I subsequently combine the catalogue of simulated landslides with the daily rainfall measurements presented in section 2.3.2, producing a time series of rainfall and landslides that can be used in the Bayesian framework in the same way as I did in section 2.4.1. Since physically based models tend to overestimate slope instability, the simulated landslides can be interpreted as all potential failures that might occur under given rainfall conditions. The use of the information obtained from a physically based model within the Bayesian probabilistic framework allows estimation of an “upper-bound” probability of landslide occurrence. Results show that probabilities derived from the simulated catalogue are about one order of magnitude higher than those obtained from the observed catalogue, supporting the range of prior distributions adopted in the probabilistic analyses presented in section 2.4.4.

Downloads

Actions (login required)

Modifica documento