Enhancing reservoir landslide displacement prediction with crack width data integration: A case study of the Daping landslide

Abstract

Existing studies on predicting reservoir landslide displacements primarily focus on rainfall and reservoir water level (RWL) as the main factors influencing landslide movement. However, these studies overlook the potential role of crack width, even though landslide cracks are critical indicators of landslide formation and movement. Currently, no predictive models in this domain have integrated crack width alongside rainfall and RWL. In response to this gap, this study investigates the predicative performance of models that combines crack width, rainfall and RWL as the set of input factors for predicting temporal variations in the displacements of the Daping landslide within the Three Gorges Reservoir Area. The multiple wavelet coherence (MWC) method is used to determine optimal time lags between the combined input factors (i.e., rainfall, RWL and/or crack width) and the output (i.e., displacement). The raw data of these input factors within these time lags are integrated as the inputs to displacement prediction models during both training and prediction phases. Commonly used deep learning models, such as the deep neural network, gated recurrent unit, bidirectional long short-term memory and transformer architectures, are adopted in our experiment. Experimental results show that incorporating crack width data improves the accuracy of transient landslide displacement predictions compared to models that exclude crack width data, for the adopted prediction models.