Deep Learning for spatio-temporal analysis of anthropogenic ground deformation in the North Adriatic coasts of Italy recorded by GNSS time series

Detecting and understanding ground deformation caused by human activities in areas with multiple hazards, such as natural subsidence, flooding, and earthquakes, is challenging, especially in coastal regions. This thesis addresses this challenge by using Global Navigation Satellite System (GNSS) displacement time series and advanced Deep Learning methods, including Convolutional Neural Network (CNN) and Autoencoder frameworks, to automatically identify ground deformation signals linked to both natural subsidence and anthropogenic activities. The study focuses on the North Adriatic coasts of Italy, where gas and oil storage and production have occurred between 2010 and 2023. In these areas, hydrocarbons are extracted with some changes in the volumes during the history of the sites, while gas storage involves seasonal injection (April-October) and extraction (November-March). Our approach involves simulating gas and oil reservoir behavior using a simple Mogi model and handling data gaps in the GNSS time series with the weighted principal component analysis (WPCA) technique. To generate synthetic training data for the CNN-Autoencoder model, we used 45 GNSS stations and randomly simulated gas/oil fields by varying locations (i.e., longitude and latitude), depths, and volume changes over time. The volume changes were modeled using different functions, including seasonal, exponential, multi-linear, bell-shaped, and real volume shapes corresponding to known gas storage and production sites. Since CNN-Autoencoder operates with image datasets, the Kriging interpolation method, known as Gaussian process regression, was applied to generate 2D spatial representation of daily displacements in three directions (east, north, and vertical). After calibrating the CNN-Autoencoder with these synthetic data, the model was tested on real GNSS data. Our results show the ability to detect significant subsidence in hydrocarbon production areas (-18 mm) and ground uplift in storage facilities (+2.7 mm) over the 14-year period (2010-2023), highlighting the method's effectiveness for analyzing anthropogenic deformation patterns in dynamic coastal environments.