Neural Networks for Slope Movement Prediction

Abstract

This article presents the use of neural networks for the prediction of movement of natural slopes. The aim is to predict velocity changes of a moving soil mass using climatological and physical data, such as rainfall and pore water pressure, which are used as input parameters in an artificial neural network (ANN). The network is designed to function as an alarm and is a decision‐making tool for persons in charge of landslide monitoring. The raw data were obtained from a continuously monitored landslide, located in Sallèdes, near Clermont‐Ferrand (France), and include daily precipitation, evaporation, pore water pressure, and landslide velocity values. The various networks used in this study are two layer perceptrons trained using the Levenberg‐Marquardt algorithm, based on backpropagation of error. The most sophisticated model presented in this article was developed by cascading two recurrent networks of the same type. This model permits a satisfactory 3‐day prediction of landslide velocity if quality data from continuous measurements are available. A simple example of the calculation of a safety factor for an unstable slope shows how neural techniques may be coupled to good advantage with purely mechanical models.