Development of land subsidence susceptibility maps based on PS InSAR data using machine learning models: The case of Aksaray province

This study aims to evaluate land subsidence susceptibility in the Aksaray province using machine learning (ML) methods. The creation of land subsidence susceptibility maps requires a comprehensive inventory of land subsidence occurrences, which is challenging to obtain through fieldwork for large areas. Therefore, this study utilized the PS InSAR (Persistent Scatterer Interferometric Synthetic Aperture Radar) technique to generate the subsidence inventory. Over a five-year observation period, 80 Sentinel-1A satellite images were processed for two distinct frames, resulting in the analysis of 63,437 PS points. To assess land subsidence susceptibility, 15 conditioning factors were analyzed, including lithology, curvature, slope, aspect, elevation, groundwater level, drainage density, well density, Topographic Wetness Index (TWI), Stream Power Index (SPI), land use, proximity to settlements, proximity to fault lines, proximity to roads, and NDVI. These factors were incorporated into machine learning algorithms such as Random Forest, XGBoost, CatBoost, LightGBM, and AdaBoost. Among these models, Random Forest demonstrated the highest predictive performance with an accuracy of 95%. Furthermore, feature importance analysis revealed that lithology, elevation, groundwater level change, and land use had the most significant impact on land subsidence. The findings indicate that Sultanhanı, Eskil, and the central districts of Aksaray are at high and very high risk of land subsidence. This study provides valuable insights for local governments to better understand land subsidence risks and develop effective risk management strategies. Additionally, it offers critical data for advancing sustainable agricultural practices and efficient water resource management.