A comparative study of random forests and multiple linear regression in the prediction of landslide velocity

Abstract

The monitoring of landslides has a practical application for the prevention of hazards, especially in the case of large deep-seated landslides. Monitoring data are necessary to understand the relationships between movement and triggers, to predict movement, and to establish an early warning system. This paper compares two phenomenological models for the prediction of the movement of the Kostanjek landslide, the largest landslide in the Republic of Croatia. The prediction models are based on a 4-year monitoring data series of landslide movement, groundwater level, and precipitation. The presented models for landslide movement prediction are divided into the model for the prediction of groundwater level from precipitation data and the model for the prediction of landslide velocity from groundwater level data. The statistical techniques used for prediction are multiple linear regression and random forests. For the prediction of groundwater level, 75 variables calculated from precipitation and evapotranspiration data were used, while for the prediction of landslide movement, 10 variables calculated from groundwater level data were used. The prediction results were mutually compared by k-fold cross-validation. The root mean square error analyses of k-fold cross-validation showed that the results obtained from random forests are just slightly better than those from multiple linear regression, in both, the groundwater level and the landslide velocity models, proofing that multiple linear regression has a potential for prediction of landslide movement.