Prevenzione di Potenziali Catastrofi Basata sulla Tecnica Radar Interferometrica e sull'Intelligenza Artificiale

This dissertation presents a comprehensive exploration of ground displacement monitoring as a means to identify potential geohazards early and prevent critical damage. Utilizing Interferometric Synthetic Aperture Radar (InSAR) for its sub-millimeter accuracy in monitoring displacements, the research addresses the technique's inherent challenges, including the need for high-level expertise, management of massive data sets, and other complexities. A significant contribution of this work is the development of an automated methodology to detect ground displacements directly from wrapped interferograms and coherence maps, a process that traditionally demands extensive manual intervention. The study embarks on a comparative analysis of various machine learning algorithms, employing selected pixels from filtered-wrapped interferograms of Sentinel-1, based on a coherence threshold. These pixels are labeled according to their displacement velocity, categorizing them into distinct movement classes. The creation of interferograms, coherence, and deformation velocity maps was facilitated using the European Space Agency's Parallel Small Baseline Subset service, followed by the application of a high pass filter to isolate displacement signals from atmospheric errors. The methodology successfully identified patterns corresponding to slow and fast movements, focusing on case studies from Italy, Portugal, and the United States due to their landslide sensitivity. Among the algorithms tested, the Cosine k-Nearest Neighbour model exhibited superior accuracy, including in adjacent areas beyond the training set, demonstrating robust generalizability. Enhancing the model's performance, Pseudo-labeling was employed to further assess the model's applicability beyond its training environment. The lowest test accuracy achieved was an impressive 80.1%. Additionally, the integration of ArcGIS facilitated a more nuanced visualization of the results, particularly in assessing displacement impacts on main roads in the study areas. Building upon these findings, the dissertation's subsequent objective involved using Long Short-Term Memory (LSTM) models to predict displacement time series, addressing both regular and irregular patterns over single and multiple time steps. The effectiveness of these predictions was assessed through an analysis of learning curves, homoscedasticity, and the autocorrelation function. An additional evaluation of the prediction outcomes involved the computation of the percentage of residuals surpassing a risk displacement threshold of ±0.35 cm. To address the irregularities inherent in the time series data, two preprocessing methods were employed: missing values imputation and feature engineering. A comparative analysis between these methods revealed that missing values imputation yielded more favorable results for the datasets in question. The study further extended its scope by implementing a Time Gated LSTM (TG-LSTM) model for forecasting multiple time steps in irregular time series. A comparative evaluation of this advanced model against standard LSTM models demonstrated the superior efficacy of TG-LSTM in all examined aspects for this specific predictive task. The final objective of this research culminated in the development of an ArcGIS Pro-based toolbox, incorporating the predictive model methodology established in the second objective. This innovative toolbox is designed for interactive prediction of both single and multiple displacement steps in regular and irregular time series. A key feature of the toolbox is its flexibility, allowing users to modify essential hyperparameters of the model to accommodate different datasets. The functionality, application, and constraints of the toolbox are thoroughly examined and discussed in the corresponding section. The primary aim of developing this toolbox is to underscore the significance of integrating InSAR, Geographic Information Systems, and Artificial Intelligence.