A map reduce seismic texture analysis and barricaded boundary minority LS-SVM framework for marine seismic exploration data -

Abstract

Oil and gas exploration involves different complex and costly procedures. Specially designed vehicles (trucks or ships) send sound waves and collect their reflections using a set of predesigned geometrically distributed sensors. Further analysis is performed to extract the different seismic attributes which help identify the different lithological formations such as oil and gas reservoirs. The analysis also helps identify suitable drilling sites and estimate oil or gas quantity for business men and economists to assess the drilling risks and costs which can reach up to 1Billion dollars. In short, seismic data analysis is a distributed big data analysis by excellence: it involves many complex and computationally expensive operations from massive data acquisition, to data processing and data analysis. In this thesis, seismic data acquisition, processing and analysis are described to highlight the complexity of the problem. The overall seismic data processing and analysis flow are migrated into a distributed design that uses the Map-Reduce Paradigm. A sample seismic texture analysis is carried out to identify target locations in an oil bearing site where slices of a 3D seismic block data are processed separately to extract window samples and their corresponding Haralick attributes using the Grey Level Co-occurrence Matrix (GLCM). We propose the Barricaded Boundary Minority Oversampling Method (BBMO) which is based on a modification of the least square support vector machine (LS-SVM) since it can be easily distributed due to its equivalent incremental form. BBMO oversamples the minority samples at the boundary in the direction of its closest majority samples to fix the problem of data imbalance caused by the fact that oil bearing sites in a specific field are usually less than the non-bearing sites resulting in imbalance in the seismic exploration data. All operations are described and profiled to find the computationally most expensive in our proposed framework. Experimental results on BBMO performance and comp