With an abundance of data provided by satellite-based measurements, such as Synthetic Aperture Radar Interferometry (InSAR) or the Global Navigation Satellite System (GNSS), an interest has grown in training highly complex data-driven models for geophysical applications, including displacement modeling. These methods, including machine learning (ML) and deep learning (DL) algorithms, represent a new approach to forecasting ground surface displacements. Yet, the effectiveness of such methods, including their generalization capabilities and performance on non-linear data, remains underexplored. This paper examines the performance of various data-driven algorithms, including regression models and deep neural networks, in predicting mining-induced subsidence. Ground surface displacement data obtained from the Small Baseline Subset (SBAS) InSAR were used as time series samples for training and validation. ML and DL models were evaluated over varying forecast horizons. The results show that data-driven approaches can effectively model InSAR-derived ground subsidence in mining areas. Deep learning models outperform other ML-based models, indicating that increased model complexity can lead to better forecasting accuracy. Nevertheless, it is shown that careful examination of performance metrics and forecast errors in the spatial domain is essential for appropriate model evaluation. The findings demonstrate that combining SBAS-InSAR measurements with data-driven modeling offers a promising direction for developing automated systems for monitoring and forecasting mining-induced ground deformation.
Dariusz Głąbicki (Tue,) studied this question.