چكيده لاتين
The Earth’s surface is continuously deformed by natural and anthropogenic processes, and such deformations can potentially evolve into severe geohazards. Interferometric Synthetic Aperture Radar (InSAR) has emerged as a key geodetic technique for precise and frequent monitoring of ground displacements over large spatial scales, particularly with the advent of spaceborne constellations such as Sentinel-1. Nevertheless, the automatic and robust detection of change points (CPs) in InSAR displacement time series remains highly challenging due to the presence of severe atmospheric noise, non-linear deformation patterns, data gaps, and irregular sampling intervals. These limitations substantially constrain the performance of conventional statistical and model-based approaches. The primary objective of this dissertation is to develop a comprehensive deep learning–based framework for automatic change point detection (CPD) in InSAR time series. To this end, a novel synthetic data generation pipeline was designed that realistically models’ displacement trends (linear and Mogi-type), seasonal components (semi-annual and annual sinusoidal signals), and multiple types of noise (Gaussian, Poisson, and white). To ensure realism and avoid manual intervention, an automated filtering procedure was implemented using a Multi-Layer Perceptron (MLP), Dynamic Time Warping (DTW), and DBSCAN clustering. Building upon this foundation, four advanced deep neural architectures were developed: MLP, providing computational simplicity and efficiency; MALkCNN, a large-kernel convolutional neural network optimized for extracting long-term trends and reducing local fluctuations; ATGLSTM, an Attention–Time-Gated LSTM designed to maintain robustness under noisy and gap-filled conditions; and CNN–LSTM, a hybrid architecture that leverages convolutional layers for spatial feature extraction and LSTM units for temporal dependency modeling. The proposed models were validated on both synthetic datasets and real-world InSAR time series acquired over geodynamically active regions in Europe, including Germany, Italy (Campi Flegrei caldera), and Iceland. Quantitative evaluations demonstrated outstanding performance: MLP achieved ~97% accuracy while reducing computational cost by nearly an order of magnitude compared to deeper models. MALkCNN reached an F1-score of ≈0.98 on synthetic data and ≈83% on real data, operating ~7× faster than the TG-LSTM baseline. ATGLSTM maintained high stability in gap-affected and noisy datasets, yielding F1 ≈72% on Icelandic time series. CNN–LSTM attained >95% accuracy on synthetic benchmarks and strong correlations with GNSS records at Campi Flegrei (Pearson’s ρ = 0.88 for ascending and ρ = 0.85 for descending tracks). These results confirm that the proposed deep learning framework substantially outperforms classical approaches in terms of detection accuracy, robustness to data irregularities, and computational scalability. Moreover, the models produced spatio-temporal change maps consistent with known geophysical processes such as magma migration, hydrothermal pressurization, and isostatic adjustment, thereby bridging computational advances with geophysical interpretation. Overall, this dissertation demonstrates that advanced deep learning architectures, when trained on realistic synthetic datasets and validated against independent geodetic observations, constitute a powerful and scalable solution for automatic, reliable, and near-real-time CPD in InSAR time series. The framework not only advances methodological research in radar remote sensing but also provides a practical foundation for the development of operational early-warning systems for volcanic unrest, subsidence, and other geohazards.
