Comparison of rotation forest based intelligence models for landslide susceptibility modeling

Abstract

The precision of landslide susceptibility assessment has always been the focus of landslide spatial prediction research. It can be considered as the possibility of landslide disaster under the action of human activities or natural factors, or both of them. For the further exploration of the mechanism of this process, Muchuan County was proposed as the study area, and four well-known machine learning models, namely rotation forest (RF), J48 decision tree (J48), alternating decision tree (ADTree) and random forest (RaF), and their ensembles (RF-J48, RF-ADTree and RF-RaF) were introduced to explore the mechanism. These models are established by twelve landslide conditioning factors, which are selected based to the local special geological environment conditions and previous related researches, including plan curvature, profile curvature, slope angle, slope aspect, elevation, topographic wetness index (TWI), land use, normalized difference vegetation index (NDVI), soil, lithology, distance to roads, and distance to rivers, as well as training (195 landslides) and validation (84 landslides) datasets were developed. The landslide prediction performance of the above conditioning factors was analyzed through the correlation attribute evaluation (CAE) model. Then, through the landslide susceptibility maps made by the above six different models, the Jenks natural breaks method is used to divide the landslide susceptibility into five grades, which are very low, low, moderate, high, and very high. In addition, the accuracy of the above six landslide susceptibility maps was verified by implementing the relative operating characteristic curve (ROC) and the area under the ROC (AUC). That is, the capabilities of the above six models are compared and verified in the landslide spatial prediction. Finally, the obtained results show that elevation, lithology and TWI are the three most principal landslide conditioning factors in this research. The RF-RaF and RaF models in the training dataset performed best, with the AUC value of 0.75, while the RF-ADTree model (0.74), RF-J48 model (0.74), ADTree model (0.71) and J48 model (0.70) performed poorly. Meanwhile, similar results also emerge from the validation dataset, in which the RF-RaF model acquired the best performance (0.82) and the rest are the RF-ADTree model (0.80), RaF model (0.79), RF-J48 model (0.77), ADTree model (0.76) and J48 model (0.71). Last but by no means the least, the results can provide scientific references for local natural resources departments.