Seismic Data Compression Using Deep Neural Network Predictors

Seismic data compression is highly demanded to reduce the cost for transmission and storage due to an enormous volume of collected data. This paper presents a prediction based compression for seismic data using deep neural networks (DNN) and entropy encoding. First, a DNN with multiple hidden layers is pre-trained using restricted Boltzmann machines (RBMs) to obtain good initial weights. The DNN is then fine-tuned in a supervised fashion to achieve a better prediction precision. The residual between actual and predicted samples are quantized to achieve smaller data representation. The quantized residuals are further encoded using the Huffman coding. Our experiments with a real data set show that the DNN significantly outperforms the Linear Predictive Compression (LPC) in term of reconstruction quality.