Neural-network based multi-azimuth processing

This paper describes the results of a series of experiments with neural networks, dip-steered noise reduction filters and other techniques aimed at combining multi-azimuth data. The seismic data was first pre-processed by applying dip-steered noise reduction filters, amplitude correction and inter-volume trace matching for dynamic shift corrections. Then the individual azimuthal stacks were combined using first unsupervised - and then supervised neural networks using custom-made semi-automated workflows.<br>The main conclusions drawn from this study are that incremental improvements were achieved after consecutively: aligning the azimuth volumes, unsupervised stacking and supervised stacking. Alignment proved to be a mandatory step. Unsupervised segmentation provided a useful segment volume that highlights the area affected by stacking issues, while the same segmentation was also used for re-stacking the seismic data. Main improvements were achieved by selecting the relative weight to use for stacking. Supervised neural network stacking were further used to smoothen the transition between segment. The “MLP weighted” output is considered better than the input mazstack. The “MLP weighted” stack is perfectly fit for interpretation since no processing related artifacts were accepted. The workflow was adapted to the pre-stack domain but no additional gains were obtained.<br>