Abstract: Landslides have characteristics such as regionality, multipleness, and seriousness. The traditional area landslide spatial prediction method, under massive data condition, has poor real-time performance and strong subjectivity, and the evaluation performance is poor under multiple factors. A distributed regional landslide prediction method based on BP neural network is proposed in this paper. The algorithm is designed as a parallel computing environment of big data processing platform Spark, and the cost function of BP network is designed as two items of mean square error and L2 regularization, which improves generalization ability. Through statistics of the quantitative indicators of landslide factors and the definition of hazard index of monitoring profile, the influencing factors are selected. This approach is applied to massive data mining of 9 landslides in 11 years in Zhongxian area of Three Gorges Reservoir area, which achieves the combination of qualitative analysis and quantitative analysis. All the landslide monitoring sections in the study area were monthly evaluated to determine the risk level, and the spatial prediction of the monthly landslide risk degree was achieved. Experiments show that the fitting accuracy and efficiency obtained by the method are better than gradient-based decision trees and random forest algorithms, and the prediction area landslide risk assessment accuracy is good. This method can be used as a new approach for regional landslide spatial prediction.

Key words: BP neural network, Spark platform, regional landslide spatial prediction, monitoring profile