Landslide susceptibility evaluation based on rough set and back-propagation neural network

Landslide susceptibility evaluation is critical for landslide medium and long-term prevention and territorial planning. In order to improve the precision and accuracy of evaluation result, the study was carried out in Longfeng Town of Enshi City. Firstly, the Geographic Information System (GIS) was suggested to use as the basic tool for spatial data management, 13 initial evaluation factors were selected including lithology, slope angle, distance to geological structures etc. Then the rough set theory based on genetic algorithm was used to reduce the redundant information of 13 initial factors in the decision table and determine the kernel including 8 representative factors, namely, lithology, altitude, curvature, distance to roads, distance to river, slope angle, aspect, stream power index. After that, the kernel factors were used to train a BP neural network model, and landslide susceptibility index (LSI) and landslide susceptibility classification map were achieved. The highest susceptibility zone is about 12.82% of the total area, including 78.11% landslide-prone area. The ROC curve test result shows that the prediction accuracy of the RS-BPNN model is about 90.9%, proving that the RS-BPNN model has advantages of excellent prediction performance and efficiency and has higher practical value.