Preparation of sinkhole susceptibility maps with GIS and machine learning methods: The case of Konya Closed Basin

Konya Closed Basin (KKH) has a karst topography where lithologically easily soluble rocks are concentrated. Karstification products such as sinkholes are the most common karstification products seen in this region. In KKH, sinkholes threaten agricultural areas, infrastructures such as energy lines and highways and settlements. In order to reduce the harmful effects of sinkholes, it is necessary to know where the factors controlling these phenomena can occur. In this thesis, the question of under which conditions sinkholes recur and where they may occur spatially is sought to be answered with a sinkhole susceptibility map. Within the scope of the study, a database was created with 628 sinkhole inventory obtained by using orthophotos, field studies, satellite images and data obtained from AFAD Provincial Directorates. The 20 conditioning factors commonly used in the literature and 5 ensemble machine learning (ML) models, namely Random Forest (RF), Light Gradient Boost (LightGBM), Extreme Gradient Boost (XGBoost), Adaptive Gradient Boost (AdaBoost) and Categorical Boost (CatBoost), were used to produce susceptibility maps. In this context, prediction performance of 95% and above was obtained for all ML models and the highest accuracy score was the random forest algorithm (97.35%). SHAP, which is an explainable artificial intelligence approach, was used to determine factor importance ratios. Accordingly, the most contributing factors for the RF model were determined as the rate of change in groundwater level, proximity to volcanic forms, well density and elevation. The produced sinkhole susceptibility map shows that approximately 39 per cent of the basin soils are located in highly and very highly susceptible areas. It is thought that the basin-based sinkhole susceptibility map produced by this study fills an important gap in the literature and provides a base map for planning activities in the basin.