Improving Seismic Resolution by a Sequential Convolutional Neural Network

Abstract

Summary Seismic data with high resolution provides better insights for extracting geological information. Due to the earth filtering, gathered seismic data are band limited. As a routine operation in seismic processing, high resolution processing involves multiple techniques. Some commonly-used methods are theory-based, demonstrating both advantages and disadvantages due to the assumptions and theories governing their issues. Besides, some data-driven methods have been introduced to perform high-resolution processing. Inspired by the encoder-decoder network and residual network, a sequential convolutional neural network was developed to improve seismic resolution. The encoder-decoder architecture works well for sequence to sequence transformation and contributes to extracting high-hierarchy features. The usage of residual learning is good for accelerating convergence and improving generalization. Taking 1D low-resolution and high-resolution time series pairs as training data, the proposed sequential convolutional neural network is trained by deep learning. Considering the limitation of acquiring actual high-resolution seismic, synthetic seismic traces are generated with separate low-resolution and high-resolution wavelets to prepare realistic and correlative input feature and output label pairs. Experiments on raw seismic data demonstrate the efficiency and generalization of the proposed method.