Prediction of crustal movements with PS-InSAR and deep learning

Crustal movements, which are an increasing problem especially due to anthropogenic effects such as excessive groundwater use, are critically important to monitor and predict for sustainable resource management and disaster risk reduction. This doctoral thesis aimed to perform a long-term analysis of crustal movements in Afyonkarahisar-Bolvadin and their future prediction by integrating PS-InSAR and deep learning techniques. In this study, 240 Sentinel-1A SAR images covering the period 2015-2024 were processed using the PS-InSAR technique to produce annual LOS movement velocity time series. Analyses revealed subsidence-dominated movements in the region, ranging from +1 to -14 mm/year, particularly concentrated in the southwestern parts where groundwater extraction is intensive. These time series were modeled for the years 2025-2026 using LSTM, GRU, TCN, CNN, and Transformer deep learning models with a Recursive Multi-step Forecasting (RMF) strategy. The findings showed that all models exhibited high success (R² > 0.92) in single-step predictions, with the CNN architecture demonstrating the best performance. RMF projections predicted the continuation of the subsidence trend, and CNN and LSTM models were found to be more stable in multi-step predictions. This study has demonstrated the potential of PS-InSAR and deep learning integration for monitoring surface movements and making meaningful future predictions, providing a scientific basis for regional planning and management strategies. The inclusion of auxiliary parameters and ground-based validation are recommended for future studies.