Optimized BP Neural Network - Semiparametric Model in Landslide Forecasting

Landslide is a complex nonlinear process. Accurate forecasting of landslides is of great importance. As a very good data processing method, semiparametric model is closer to reality than other mathematical models. However, it's too simple to describe the true relationship between the observed data and reality when the parameters are linear. With self-learning and fast-optimization abilities, artificial neural network can approximate any nonlinear function and solve complex nonlinear problems. In this paper, a model called optimized BP neural network - semiparametric model, which combines BP neural network model and penalized least squares criterion based semiparametric model, is proposed. First, a forecast based on BP neural network is carried out. By taking the result of former forcast as the parametric component in semiparametric model, the non-parametric component was derived based on penalized least squares criterion. A forecast of a slope based on the proposed method is carried out with GPS data. The accuracy and feasibility of the proposed method in landslide forecasting is demonstrated by comparing with several currently widely used models.