Method for Meteorological Early Warning of Precipitation-Induced Landslides Based on Deep Neural Network

摘要

The meteorological early warning model of precipitation-induced landslides is a significant yet challenging task, due to the complexity and uncertainty of various influence factors. Generally, the existing machine learning methods have the drawbacks of poor learning ability and weak capability of feature extraction. Inspired by deep learning technology, we propose a deep belief network (DBN) approach with Softmax classifier and Dropout mechanism for meteorological early warning of precipitation-induced landslides to overcome these problems. With the powerful nonlinear mapping ability of DBN when training a large number of sample data, we use the greedy unsupervised learning algorithm of DBN to extract the intrinsic characteristics of landslide factors. Then, to further improve prediction accuracy of landslides, the Softmax classifier is added to the top layer of DBN neural network. Moreover, the Dropout mechanism is introduced in the training process to reduce the prediction error caused by the over-fitting phenomena. Taking Wenchuan earthquake affected area for example, after analysis of the factors influencing landslide disasters, the meteorological early warning model of landslides based on Dropout DBN-Softmax is established. Compared with the existing BP neural network algorithm and BP algorithm based on Particle Swarm Optimizer (PSO-BP) algorithm, the experimental results show that the new approach proposed has the advantages of higher accuracy and better technological performances than the former algorithms.