Linear passive source surface wave dispersion curve picking based on supervised deep learning and ambient noise tomography for the evolution of the internal structure in landslide area

Abstract. The complex structural system of landslide, influenced by interactive triggering factors, plays an important role in its stability. The early identification and continuous characterizing of internal geometry variation and failure mechanisms, constitutes a crucial step for hazard analysis and monitoring. Recent advances in non-invasive geophysical methods, particularly ambient noise tomography, have revolutionized landslide investigation by providing near-continuous view and rapid wide-area scanning for the landslide structure imaging. In this study, we used a seismic array in a landslide-prone area in Guizhou, China, aiming to characterize the spatial properties and determine the temporal variations in subsurface structure of the landslide. The extended spatial auto-correlation method (ESPAC) as a simple and robust seismic observational method for linear arrays was carried out to extract surface wave signals from ambient noise. Furthermore, in order to make the core but time-consuming process of dispersion curve picking more intelligent and reliable, this article proposed a deep learning-based method (lightweight U-net) regarding the dispersion curve extraction as an image classification problem for automatic process. Subsequently, the CPSO program was executed, combined with the hydrogeological data, to obtain the S-wave velocity structure of landslide area for observation periods. Data interpretation revealed the internal spatial structure characteristics of the landslide body, including two contrasting lithologies, namely the upper Gravelly clay deposit and a relatively dense weathered bedrock (limestone) at the bottom, and potential sliding surfaces. Besides, monitoring the temporal variations of velocity detected from long-term ambient seismic noise recordings can be attributed to structural evolutions in the very near surface, likely induced by surface erosion and shallow groundwater due to rainfall. The theoretical research and practical application in our work represent an efficient and collaborative comprehensive technical system to elucidate the triggering factors and enhance the ability of landslide identification and early warning, and furthermore to promote the development of landslide disaster monitoring towards intelligence in sight.