Reprocessed 3D seismic-reflection data and neural-network fault cube, offshore of Point Sal, central California, from 2012-08-12 to 2012-10-05 (USGS field activity P-04-11-CC)

Seismic data were acquired using a triple-plate boomer source operated at 1.75 kJ and the P-Cable streamer system. The 3D data show a frequency range of approximately 100-700 Hz, with a dominant frequency of 200-225 Hz (PG&E, 2014). The P-Cable system consisted of 14, 8-channel digital streamers with 6.25 m hydrophone group spacing (Nishenko and others, 2012; Ebuna and others, 2013). Data were sampled at 0.25 ms with a record length of 0.75 s two-way travel-time. Seismic processing steps conducted by Fugro Consultants included quality control, tidal corrections, velocity analysis, NMO correction, stacking, surface related multiple elimination, deconvolution, crossline statics, and pre-stack migration (Fugro Consultants Inc., 2012; PG&E, 2014). However, upon public release and inspection by the USGS, significant high-frequency noise was observed, limiting the interpretation of imaged fault structures (Kluesner and Brothers, 2016).

The publically released 3D seismic data were downloaded and analyzed by the USGS and a post-stack data-conditioning workflow was designed to reduce noise, short-period multiples, and enhance the overall quality. The workflow followed the methods described in Kluesner and Brothers, 2016. Data were imported into the OpendTect software package, and a dip-steering volume was calculated using a 1x1x1 (inline, crossline, sample interval) calculation window. Following dip-steering calculation a median filter was applied with a 1x1x1 step-out, resulting in a smoothed 3D volume of dip and azimuth information for all seismic events. The dip-steered median filter was then applied to the data with a 1x1 (inline, crossline) stepout, with no time window. Filtered results were then subtracted from the original data to produce a volume with significant noise suppression. Both random and apparent acquisition noise were suppressed, while enhancing laterally continuous events. Following filtering and noise suppression, a post-stack 3D predictive deconvolution filter was designed in order to suppress apparent ringing within the data. Throughout the 3D volume, the seafloor reflection suffers from a long wavelength, masking geology below. The filter effectively suppressed the ringing/short-period multiples and also sharpened reflections at depth. For enhanced fault detection, a neural-network multi-attribute workflow was designed that used 32 different seismic attributes. The 32 attributes, which consisted of calculations such as similarity, curvature, and noise, were used as input nodes in the neural-network. The hidden layer of the neural-network consisted of 16 nodes, and user picks of faults and non-faults were used to supervise the network training. Training was stopped when the normalized root-mean-square (RMS) error reached a minimum value (< 0.5), along with misclassification percentage of picks. Post training, a 3D volume was generated (fault cube) that contains fault probability measurements scaled between 0 (non-fault) and 1 (fault). This information was then projected onto vertical and horizontal seismic slices to help identify and interpret probable faults imaged within the dataset.

Edited metadata to add keywords section with USGS persistent identifier as theme keyword. No data were changed.

Data were moved to a new repository, and metadata was edited to reflect the new location. Other edits were performed as needed to bring the metadata up to PCMSC standards, including a slight rewording of the title. No data were changed.