Orthogonal decomposition plays an important role in signal processing. We introduce a neural network learning algorithm for seismic data processing. We present an orthogonal decomposition approach for seismic data via adaptive principal-component extraction. In decomposed eigen-subspace, by means of eigen-extraction, noise is eliminated and the signal-to-noise ratio is enhanced. Compared with the conventional K-L method, the learning algorithm can selectively extract the required few principal components used in eigen-filtering, rather than all principal components of the input data. Therefore, computation costs can be reduced and the more traces there are, the more effective the approach. To demonstrate its effectiveness, we present two real data examples of seismic data processing.