Deep Learning Deriving New Generation Geophysical Interpretation for Landslide by Microtremor Method

Deep Learning technique was proven to be benefit to geophysical survey for the growing numbers of landslide geohazards. Microtremor method has been widely applied in the surveying the landslides due to its convenience and low cost. As the key step for geophysical interpretation, we applied the DL technique to generate the new generation inversion method due to its strong adaptivity and mapping abilities on the big amount data. In this study, considering the ergodicity and ordering of the near surface wave velocity structure, we firstly build data set based on 106 wave velocity models according to the theory of constrained Markov Decision. Then, a hybrid neural network structure named CNN-LTSM, including local learning module and long-short term memory layer, was proposed to build the nonlinear mapping relationship between Ellipticity curves derived by forward models and S-wave velocity profile. Finally, predicted inversion was conducted through network training. According to the practical ellipticity curves (H/V curve) by microtremor method observed respectively in a huge ancient landslide deposit, the 3D wave velocity profiles were uncovered. Comparing with by traditional inversion method, the new method based on DL technique provides more “high-resolution” geophysical information for geological interpretation, although the results overall distribution tendency of the results is similar to those by traditional inversion method. Thus, the arbitrary precision arithmetic and excellent robustness of the DL inversion method may play a key role for us to understand the complex landslide.