Velocity model building via combining seismic slope tomography and supervised deep learning

Abstract

Seismic slope tomography is an effective method to build macro velocity model. In order to improve the accuracy and resolution of the slope tomography, we proposed a novel approach that combines slope tomography with supervised deep learning. First, the slope tomography is used to obtain the macro velocity model and the positions of reflection points. Subsequently, the slope tomographic model, positions of reflection points and the corresponding observed traveltimes are used as inputs simultaneously for a neural network, whereas the actual velocity models are used as the labels. Through training the neural network with sufficient samples, the mapping from the inputs to the real velocity model is established. The neural network learns the background velocity of the real model from the smooth tomographic model, the velocity details from the traveltimes and the formation interface information from the positions of reflection points. Consequently, a high-accuracy and high-resolution velocity model is obtained on the basis of the slope tomographic model. Both tests on synthetic seismic data and applications to field seismic data demonstrate the effectiveness of the proposed method.