Acquisition of incomplete data, i.e., blended, sparsely sampled, and narrowband data, allows for cost-effective and efficient field seismic operations. This strategy becomes technically acceptable, provided that a satisfactory recovery of the complete data, i.e., deblended, well-sampled, and broadband data, is attainable. Hence, we explore a machine-learning approach that simultaneously performs suppression of blending noise, reconstruction of missing traces, and extrapolation of low frequencies. We have applied a deep convolutional neural network in the framework of supervised learning in which we train a network using pairs of incomplete-complete data sets. Incomplete data, which are never used for training and use different subsurface properties and acquisition scenarios, are subsequently fed into the trained network to predict complete data. We develop matrix representations indicating the contributions of different acquisition strategies to reducing the field operational effort. We also determine that the simultaneous implementation of source blending, sparse geometry, and band limitation leads to a significant data compression where the size of the incomplete data in the frequency-space domain is much smaller than the size of the complete data. This reduction is indicative of survey cost and duration that our acquisition strategy can save. Synthetic and field data examples demonstrate the applicability of the proposed approach. Despite the reduced amount of information available in the incomplete data, the results obtained from the numerical and field data cases clearly show that the machine-learning scheme effectively performs deblending, trace reconstruction, and low-frequency extrapolation in a simultaneous fashion. It is noteworthy that no discernible difference in prediction errors between extrapolated frequencies and preexisting frequencies is observed. The approach potentially allows seismic data to be acquired in a significantly compressed manner while subsequently recovering data of satisfactory quality.

Acquisition of complete data, i.e., unblended, well-sampled and broadband data, is technically desirable. Obviously, such a scenario is prohibitively expensive to realize. To deal with economic considerations in a seismic survey without seriously compromising data quality, we propose a machine-learning approach that offers an opportunity to acquire incomplete data, i.e., blended, sparsely-sampled and narrowband data, while still benefitting from being able to process complete data. In this study, we utilize a deep convolutional neural network. The incomplete data are fed into the applied network that simultaneously performs suppression of blending noise, reconstruction of missing traces and extrapolation of low frequencies such that prediction of the complete data is attainable. We validate the performance of the proposed method using both synthetic and field datasets. Acquisition scenarios implemented to generate incomplete datasets impose a significant reduction of data size in the frequency-space domain. Despite the limited information available in the input data, the prediction results obtained from both numerical and field data examples clearly confirm that the proposed machine-learning approach is capable of dealing with deficiencies in the incomplete data and subsequently deriving the complete data of sufficient quality. In addition to suppression of blending noise and reconstruction of missing traces, no discernible difference in prediction errors between preexisting and extrapolated frequencies is observed, which is hardly realizable with existing geophysics-based approaches. As a consequence, the proposed scheme allows for optimal data enhancement even when seismic acquisition is performed in a blended, sparsely-sampled and narrowband fashion.