From Model-Driven to Data-Driven Seismic High-Resolution Processing

Abstract

Summary The classical model-driven seismic high-resolution processing method using (time-variant) deconvolution or reflectivity inversion is derived to be a special case of the data-driven high-resolution processing method using artificial neural networks (ANNs). To obtain high-resolution subsurface images, we propose an interpretable gated recurrent encoder-decoder networks (IGREDN), an advanced ANN-type method, to process the observed band-limited seismic data. The developed networks consist of an encoding network and a decoding network, which are both mainly composed of bidirectional gated recurrent unit networks (Bi-GRU). The Bi-GRU is well-suited to process sequential signals including well-log data and seismic data. In addition to using the observed band-limited seismic data to supervise the seismic data generated by the decoding network, the wide-band reflectivity data derived by well logs is used to supervise the output of the encoding network. These two collected supervisors are utilized to train the network parameters, instead of estimating the wavelet(s) and solving the inverse of the matrix. Furthermore, the IGREDN-based data-driven high-resolution processing method gets rid of the fixed forward model with various assumptions, in contrast to the model-driven method. Examples are adopted to illustrate the effectiveness of the proposed IGREDN-based data-driven high-resolution processing method.