Di Muro, Chiara (2026) Artificial intelligence as a tool for ground instabilities hazard assessment. [Tesi di dottorato]

	Documento PDF DiMuro_Chiara_38_completa.pdf Visibile a [TBR] Amministratori dell'archivio Download (9MB) | Richiedi una copia
	Documento PDF DiMuro_Chiara_38_parziale.pdf Visibile a [TBR] Amministratori dell'archivio Download (4MB) | Richiedi una copia

Abstract

Ground instabilities are natural processes that profoundly alter the morphology of the territory and have a strong impact on urban settlements, anthropic activities and land use. For this reason, this issue has attracted considerable attention from the scientific community, as well as further disciplinary domains that offer a relevant contribution to the analysis of risk perception (Crozier and Glade, 2005; Glade et al., 2005). Thus, landslide study involves a wide range of methodologies for assessing susceptibility, hazard and consequently risk, and, over the years, several models, belonging to different analysis categories, have been developed. This study aims to estimate landslide susceptibility and hazard by means of distinct approaches and various input datasets in different geological and environmental contexts. The aim of the thesis is to apply statistical, multi-criteria decision making and machine learning methods, and subsequently conduct their comparative analysis, to capture the discrepancies and pinpoint which of them offers the most accurate landslide susceptibility analysis. The case study selected is Oregon (USA), therefore, an analysis was carried out on a national scale, using data at two spatial resolution (30 m and 10 m). This decision, moreover, derives from the intention of grasping the divergences between the outputs at different resolutions, to examine how the resolution itself conditions the results. Additionally, the study was contextualized within the NRRP RETURN (multi-Risk sciEnce for resilienT communities under a changiNg climate) project and VS2 Spoke, to which it pertains, to discern the relationship and make comparisons between the analyses conducted. The results revealed a solid level of reliability of the techniques adopted to predict landslides susceptibility and a satisfactory performance of the models, consistently with the Area under the Curve (AUC) values obtained in the validation phase. Likewise, they highlighted differences between the resulting maps and confirmed the role of spatial resolution and, therefore, the relevance of the selected input dataset. The second purpose of this study is to introduce a hybrid deep learning model, integrating Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks and attention mechanisms, to predict landslide-related ground deformations. The investigations incorporate both spatial and temporal data to assess landslide hazard. The investigated area regards some municipalities of Lazio, Marche, and Umbria regions of Central Italy, impacted by the multiple 2016-2017 seismic events. The analyses integrate interferometric data from the European Ground Motion Service (EGMS), relating to cumulative displacements for the period 2018-2022 and geospatial data which represents the landslide predisposing factors. The spatial and temporal Attention Mechanisms of the algorithms also optimize the model by attributing weights to critical regions and temporal intervals, highlighting areas with high levels of susceptibility. The results attest the ability of the model to reproduce the cumulative displacements observed in EGMS interferometric data, proven by the particularly low error values. Hence, the CNN-LSTM model is configured as an exhaustive and reliable methodology for hazard estimation in areas particularly subject to movements along the slopes. The ability to combine spatial, CNN-related, and temporal, LSTM-related, data ensures a comprehensive and wide-ranging analysis in the field of ground instabilities studies. The third objective concerns the evaluation of the seismically-induced landslide hazard through the development of a Virtual Environment and a Real Environment (“VERE” framework) and advanced data analysis, and machine learning features. The hazard is expressed in the form of maps of co-seismic and post-seismic displacements which take into account the seismic hazard of the investigated area. Indeed, the case study is the Upper Sele River Valley (Southern Italy), severely hit by a strong earthquake in 1980 with a magnitude of Mw 6.9. The VERE environment contains respectively sets of data and maps, ranging from morphometric and geotechnical data to seismic ones and, moreover, interferometric data, used to define the state of activity of landslides. Maps were generated through ML-based prediction, considering displacements in different initial scenarios of pore water pressure and seismicity. The analyses demonstrated the potential of the machine learning model, that enabled to undertake a study based on the integration of heterogeneous data, which, once combined, guarantee a detailed analysis of landslide hazard of the seismic-prone area, known for its recent seismic activity. In conclusion, the overall PhD work evidenced the significance of the employment of multimethodological approaches for landslides study from a spatial and temporal viewpoint, giving prominence to the dynamic analysis of ground instabilities, which transcends the traditional static perspective.

Downloads

Actions (login required)

Modifica documento