SEISMIC DATA INTERPOLATION USING ATTENTION-BASED DEEP LEARNING

Abstract

Seismic data suffer from insufficient spatial data sampling due to many restrictions. Therefore, seismic data interpolation is demanded as pre-processing on data processing to enhance data quality. Despite of wide usage of conventional seismic data interpolation methods, it is still challenging to interpolate where the data has consecutive missing. In order to improve interpolation performance at consecutive large gap, we propose a convolutional neural network equipped attention mechanism, called CUNet. We test the performance of proposed method using seismic field data. We also compare the results of CUNet with results of MWNI and UNet methods to verify the reliability of the proposed method. The test results show that CUNet is robust to restore the amplitude at consecutive traces missing. It indicates that the proposed method is feasible to interpolate even where the traces gap is relatively large.