NEAR LOSSLESS SEISMIC DATA COMPRESSION USING SIGNAL PROJECTION TECHNIQUE
Seismic data files in SEGY format can be of substantial size as these contain generally hundreds of traces collected from multiple shots. The data is usually transmitted through direct wires to different processing centers. It is important to preserve the integrity of the data in transmission and for storage, however, with very little loss, it is possible to compress the data with large compression factors. To this extend we propose a computationally efficient and robust technique for compressing multiple traces from multiple shots using Principle Component Analysis (PCA). Here, we use PCA to reduce the dimensionality of the data. The basic concept relies on finding a linear transformation that could project the original data over a set of orthogonal basis. The transformation is obtained from Singular Value Decomposition (SVD) of the estimated autocorrelation matrix. The autocorrelation matrix represents the dependencies across the traces from the same sensors and across different sensors. In our experiments we used publically available data from the Texas Seismic database. The data is gathered from 18 shots recorded by 33 sensors. The shots are generated by dynamite in 80-100ft depth holes. Each sensor is located 220 feet from another sensor. The data consists of 18 traces per sensor each trace contains 1501 data points sampled with 2 ms of sampling period and are filtered by 8-64Hz bandpass filter. The experiment covered 18 shots. Using the developed algorithm we reduce the dimensionality of the data from 18 to 3. The dimension of the output feature vector (3) was obtained by preserving more than 95% of the total energy in the original data which can be seen as almost a lossless compression.
