Crossline interpolation with the traces-to-trace approach using machine learning

Abstract

Trace interpolation using machine learning (ML) has been actively studied recently. Especially, crossline interpolation in towed streamer system is an important task due to the sparsity of the crossline data compared to the dense inline data. The key to successful ML application in crossline trace interpolation is how similar the training data are to the target data, which are the crossline data to be interpolated. Considering the similarity, we use inline data for training the ML model, and then apply the trained model to the crossline data for the crossline interpolation. In this way, we can train the model using dense inline data, and fill the gaps on sparse crossline data using the same seismic data without additional datasets. For ML based crossline interpolation, we use traces-to-trace approach with LSTM (Long Short-term Memory) networks, which uses two traces as input and one trace between the input traces as an output. In addition, we design multiple networks to predict traces at different locations with different ratios between the input traces. A synthetic towed streamer data is used to demonstrate the effectiveness of the proposed crossline interpolation method using traces-to-trace approach with multiple networks.