Landslide susceptibility analysis using machine learning: Insights from the bend Subcarpathians-Carpathians, Romania

The study presents a landslide susceptibility map using maximum entropy (MaxEnt) a Machine Learning algorithm. For this purpose, an area of 376 km² was selected in the Bend Subcarpathians and Carpathians (Romania). In this area, 966 landslides were mapped based on high-resolution aerial imagery and field surveys. A database containing geo-factors (lithological units, age of formations, slope, morpho-structural units, and land use) at medium resolution was built in a GIS environment. The results show that lithology and slope are the factors that best correlate with the spatial distribution of landslides. The landslide susceptibility map (probability of landslide occurrence) overlaps very well with the litho-structural units developed from southwest to northeast. The model's performance was evaluated using the Receiver Operating Characteristic (ROC) curve. The resulting value of 0.82 indicates a high level of predictive accuracy, comparable to values reported in other studies focused on landslide susceptibility assessment. Based on the results, four susceptibility zones (two high probability, two low probability) have been highlighted according to the litho-structural units.