An Improved Robust Principle Component Analysis for The Denoising of Desert Seismic Data

Contamination of seismic data by background noise causes difficulties for the following inversion, imaging, stratigraphic interpretation, etc. Desert seismic records pose a particular problem because of the strong energy of desert random noise and its serious spectrum overlapping with effective signals. Thus, the robust principle component analysis (RPCA) is introduced into the denoising of desert seismic data. RPCA is a classical method of low-rank matrix recovery. By kernel norm optimization, it can decompose noisy data into optimal low-rank matrix (LM) and sparse matrix (SM) which contain most effective signals and noise, respectively. However, due to the low signal-to-noise ratio and serious spectrum overlapping of desert seismic record, there are still a lot of random noise in its optimal LM. Therefore, the convolutional neural network (CNN) is combined with RPCA to establish the optimal mapping relationship from noisy LM to desert random noise through the training of CNN, so as to accurately predict desert random noise from the LM of desert seismic record. Finally, the denoising result is obtained by subtracting the desert random noise predicted by CNN from the LM. Experiments show that this improved RPCA can suppress random noise and recover effective signal more effectively than the traditional methods.