Landslide susceptibility mapping based on landslide classification and improved convolutional neural networks

Based on landslide survey data and geological conditions of the research area, landslide susceptibility mapping is to analyze the impact of combination characteristics of hazard-inducing factors on the occurrence probability and divide the area into different susceptible areas. The engineering rock formation and geological hazards survey were conducted and established a landslide database of Boshan District. The 99 landslides in Boshan District were classified into 42 natural landslides and 57 engineering landslides, whose accuracy was validated by K-means cluster. The hazard-inducing factors of all landslides, natural landslides and engineering landslides were graded by the information value method and established a seven-layer improved convolutional neural networks. The model for the susceptibility assessment was trained and verified for all landslides, natural landslides and engineering landslides, whose accuracy had been validated by the AUC method. Based on ArcGIS12.0, the landslide susceptibility probability in Boshan District was calculated to have drawn a landslide susceptibility map. The results show that the landslide susceptibility probability had a minimum value of 0.136 and a maximum value of 0.841. Extreme, high, moderate, minor and minimal dangerous areas, respectively, accounted for 8.08% (56.4 km(2)), 17.62% (123.0 km(2)), 25.33% (176.8 km(2)), 32.87% (229.4 km(2)) and 16.10% (112.4 km(2)) of the total areas of Boshan District.