Compression of Seismic Signals via Recurrent Neural Networks: Lossy and Lossless Algorithms

In recent years, deep Recurrent Neural Networks (RNNs) have been successfully applied to the prediction and processing of time series data. Since there exists significant temporal correlation among samples of seismic traces, RNNs are also potentially suitable for the compression of seismic signsals. In this article, we propose two algorithms for lossy and lossless compression of seismic signals via deep RNNs. In both lossy and lossless cases, we show that the proposed compression algorithms outperform the current state of the art in two widely used seismic signal datasets. In particular, we show that seismic signals, depending on dataset, only need close to 16 bits per sample for lossless representation, rather than 32 bits per sample.