AASPI INDEPENDENT COMPONENT ANALYSIS ICA David LuboRobles and
AASPI INDEPENDENT COMPONENT ANALYSIS (ICA) David Lubo-Robles and Kurt J. Marfurt Norman, 11/16/2017
AASPI Objective ICA vs. PCA Workflow Geology Results Conclusions Using ICA ü Sensor records different source signals at the same time. ü When the data are not Gaussian, linear and stationary. ü When data can be noisy. 1
AASPI Objective ICA vs. PCA Workflow Conclusions Results Geology Cocktail-party Problem ü Statistically independent Maximize nongaussianity of the data Negentropy 2
AASPI Results Conclusions Negentropy m 1 Geology IC 2 m 2 PCA Whitening Workflow ICA vs. PCA ICA PCw 2 Objective PCw 1 IC 1 Negentropy w+=E{xg(w. Tx)}-E{g’(w. Tx)}w Gaussian distributions are forbidden 3
AASPI Objective ICA vs. PCA Workflow Results Geology PCA ICA Orthogonal projections Linear projection (not necessarily orthogonal) Minimizes the variance of the data (2 nd order statistics) (Fit Gaussian Distribution) Minimizes higher order statistics (Non-Gaussian distributions) Compress the data Separate the data m 2 PCA m 2 Why ICA? Conclusions ICA m 1 m 2 m 1 Based on Brian Moore m 1 4
AASPI Objective ICA Workflow ICA vs. PCA Geology Results Conclusions 5
AASPI Objective ICA Workflow ICA vs. PCA Geology Results Conclusions Kurtosis: Measurement in terms of tail extremity (heavy-tailed or light-tailed). Existence of outliers (Westfall, 2014) Super-Gaussian (Leptokurtic) Gaussian (Mesokurtic) Sub-Gaussian (Platykurtic) -5 -4 -3 -2 -1 0 1 2 3 4 5 6
AASPI Objective ICA vs. PCA Workflow Geology Results Conclusions Taranaki Basin § The Taranaki Basin is one of New Zealand’s largest sedimentary basins and it is located in the western side of the North Island. (Palmer, 1985). It was the first sedimentary basin of the country and it is currently the only producing province (NZP&M, 2014). Taranaki Basin § The Taranaki Basin covers approximately 330, 000 Km 2 and several discoveries have been made (NZP&M, 2014). § The Tui-3 D Seismic Survey is located on the Western Platform in the Taranaki Basin. The main depositional environment becomes increasingly deeper water to the northwest. (Yagci, 2016). Modified from Google Maps, 2017 7
AASPI Objective ICA vs. PCA Workflow Results Geology Conclusions 1. 5 E 7 0 Zone 1 Shallow water system ~ 350 ms Zone 2 Deep water system ~ 350 ms 0 -1. 5 E 7 4 Km 8
AASPI Objective ICA vs. PCA ICA Workflow Geology Results Conclusions Internal Architecture Analysis using Spectral Components 25 Hz 30 Hz 35 Hz 80 Hz ICA 1 ICA 2 RGB Blending PCA ICA 3 PCA 1 PCA 2 PCA 3 RGB Blending 9
AASPI Objective ICA vs. PCA Workflow Geology Conclusions Results Zone 2 2136 ms ICA 1 PCA 3 Principal Component Analysis (PCA) PCA 2 ICA 3 ICA 2 Independent Component Analysis (ICA) 10
AASPI Objective ICA vs. PCA Workflow Geology Conclusions Results Zone 2 2188 ms ICA 1 PCA 3 Principal Component Analysis (PCA) PCA 2 ICA 3 ICA 2 Independent Component Analysis (ICA) 11
AASPI Objective ICA vs. PCA Workflow Geology Conclusions Results Zone 1 764 ms ICA 1 PCA 3 Principal Component Analysis (PCA) PCA 2 ICA 3 ICA 2 Independent Component Analysis (ICA) 12
AASPI Objective ICA Workflow ICA vs. PCA Geology Conclusions Results Zone 1 708 ms ICA 1 PCA 3 Principal Component Analysis (PCA) PCA 2 ICA 3 ICA 2 Independent Component Analysis (ICA) 13
AASPI Objective ICA vs. PCA Workflow Geology Results Conclusions ü Independent Components Analysis (ICA) proved to be a powerful technique to reduce dimensionality. ICA uses higher order statistics that found more interesting projections than Principal Component Analysis (PCA) on 12 input volumes. ü ICA provided better resolution than PCA at interpreting deep water and shallow water systems. Channels can be successfully delineated. Some channels are only highlighted using ICA. ü Excess kurtosis (K) was used to sort the Independent Components. However, a visual inspection of each seismic volume is recommended. ü For further research, ICA will be applied using other seismic attributes. Techniques such as mean variance from the reconstructed data will be used and compare to kurtosis in order to sort Independent Components. 14
AASPI Acknowledgements We thank New Zealand Petroleum & Minerals for providing the raw data. In addition, we would like to thank the sponsors of the Attribute Assisted Seismic Processing and Interpretation (AASPI) consortium for their support and to Schlumberger for the licenses in Petrel, provided to the University of Oklahoma. 15
AASPI References 16
- Slides: 17