Modeling landslide susceptibility using data mining techniques of kernel logistic regression, fuzzy unordered rule induction algorithm, SysFor and random forest

This paper introduces four advanced intelligent algorithms, namely kernel logistic regression, fuzzy unordered rule induction algorithm, systematically developed forest of multiple decision trees and random forest (RF), to perform the landslide susceptibility mapping in Jian’ge County, China, as well as well study of the connection between landslide occurrence and regional geo-environment characteristics. To start with, 262 landslide events were determined, and the proportion of randomly generated training data is 70%, while the proportion of randomly generated validation data is 30%, respectively. Then, through the comprehensive consideration of local geo-environment characteristics and relevant studies, fifteen conditioning factors were prepared, such as slope angle, slope aspect, altitude, profile curvature, plan curvature, sediment transport index, topographic wetness index, stream power index, distance to rivers, distance to roads, distance to lineaments, soil, land use, lithology and NDVI. Next, frequency ratio model was utilized to identify the corresponding relations for conditioning factors and landslides distribution. In addition, four data mining techniques were conducted to implement the landslide susceptibility research and generated landslide susceptibility maps. In order to examine and compare model performance, receiver operating characteristic curve was brought for judging accuracy of those four models. Finally, the results indicated that a traditional model, namely RF model, acquired the highest AUC value (0.859). Last but gained a lot of attention, the results can provide references for land use management and landslide prevention.