-
oa Influence of Noise from Passive Seismic Reservoir Detection
- Publisher: European Association of Geoscientists & Engineers
- Source: Conference Proceedings, GEO 2010, Mar 2010, cp-248-00403
Abstract
We present a numerical modeling study of the impact of surface noise on the ability to correctly detect<br>low-frequency micro-tremors originating from the subsurface. The analysis is motivated by the need to<br>assess the extent by which the empirically observed hydrocarbon (HC) micro-tremor can be masked by surface noise.<br>Subsurface micro-tremors and surface noise sources (hereon referred to as signal and noise) are<br>placed within a p-wave velocity model of a producing field. Placement of the virtual signal and noise<br>sources is based on known reservoir and noise locations. Signals are modeled by vertically polarized<br>Ricker wavelets with 3Hz center frequency randomly distributed in time. Noises are modeled by either<br>vertically or horizontally polarized white noise filtered between 1-18Hz. The sources are propagated<br>using a staggered grid finite difference solver, and the particle velocities are recorded at ground level<br>by virtual receivers. A series of forward simulations is run with different noise strengths to achieve<br>various signal-to-noise ratios (SNRs). Placement and signature of the signal and noise sources are kept constant.<br>A spectral ratio attribute known to be indicative of hydrocarbons is computed for the synthetic results.<br>From this attribute, the virtual receivers are classified into two groups, above HC and away from HC,<br>using two classification methods: Jenks' natural breaks and neural networks (NNs). The classification<br>results from the two methods are consistent. The best performing NNs, trained on synthetic data for<br>each noise level, are then used to generate hydrocarbon probability maps based on real passive<br>seismic data acquired on the same producing field. These maps are compared with the actual oil-water<br>contact (OWC). Results show that the predictions were most reliable for NNs trained on a SNR of 0.5<br>(or -3 dB). Predictions from NNs trained on higher or lower SNRs give inferior predictions for this<br>dataset. A summary of the influence of noise on the ability to predict reservoir presence is provided.<br>This study shows that even though noise impairs the ability to correctly identify hydrocarbon related<br>tremors, more reliable results can be achieved by applying processing methods that accommodate for variations from noise.