Landslide Susceptibility Mapping using Ensemble based Machine Learning Models in the Southeast of Gyeonggi-do, Korea

In Korea, there are many landslides occurred in rainy season and the landslide have damaged to people and facility frequently. To prevent and response the damage, the landslide susceptibility map should be prepared. So, to make landslide susceptibility map, the ensemble based machine learning models such as AdaBoost and Bagging models have applied and validated to the Southeast of Gyeonggi-do, Korea. A landslide inventory map including a total of 230 landslides was compiled on the basis of earlier reports and interpretation of aerial photographs. All landslides were randomly separated into two data sets: 50% of landslides were selected for establishing the model and the remaining 50% landslides were used for validation purposes. The 15 factors were used as controlling factor of landslide occurrences. Topographic factors such as slope, aspect, curvature, effective air flow heights, terrain surface convexity, terrain surface texture, wind exposition index and melton ruggedness number, hydrologic factors such as valley depth and topographic wetness index, land surface factor such as land-cover, forest type, forest age and forest density and geologic factor such as lithology, were used for landslide susceptibility mapping. Then, landslide susceptibility maps were calculated and drawn using the AdaBoost and Bagging models. Finally, the maps were validated using Receiver Operating Characteristic (ROC) curve and Area Under the Curve (AUC) methods. As the results, the Bagging (84.79%) and AdaBoost (86.42%) showed high and satisfactory accuracy. respectively. Overall, all landslide models adopted in this study had a higher accuracy more than 80%. These models could be used for landslide susceptibility assessment in landslide prone areas. The produced landslide susceptibility map will help to the decision makers during site selection and site planning processes. The map may also be accepted as a basis for the landslide risk-management studies. The machine learning modeling approach, combined with GIS spatial data, yields a reasonable accuracy for the landslide prediction.