"Geophysical Log Data Classification In Crystalline Rocks Using Pattern Recognition Methods"

Well log interpretation is one of the prime sources of information for the deep lithology in drilling research.In fact, it is difficult to interpret geophysical well log data in crystalline rocks due to their complex geological features and the difficulty in understanding and using the complex and intensive information content in these data. Besides, to date, there is no systematic formulated interpretation/classification methods available for crystalline rocks in geophysics; thus posing challenges in accurate identification of lithology using geophysical log data. Automating well log interpretation provides two immediate benefits:firstly, it makes it possible to process large amounts of logs rapidly, thus supporting the creation of large knowledge bases. Secondly, automation produces uniform results. Since the early days of the introduction of computers to geosciences, algorithms combined with geophysicist reasoning have made a significant contribution to the field of geophysics. The most important of these are pattern recognition methods. Motived by the successful prediction abilities of pattern recognition methods to solve different problems in geophysics, this thesis explore the applicability of using pattern recognition methods to process and classify geophysical log data in the context of crystalline metamorphic rocks. These pattern recognition methods are calibrated on Chinese Continental Scientific Drilling Main Hole (CCSD-MH) data, which provides core data and 13 geophysical well logs (CNL [Compensated Neutron], DEN [Compensated Bulk Density], PE [photoelectric absorption capture cross section], VP [P-wave velocity], GR [gamma ray], K[potassium content], KTH [Potassium Plus Thorium], TH[Thorium], U[Uranium content], RD [resistivity of laterolog deep], RSFL[Spherically Focused Resistivity], RS [Shallow resistivity], GRSL[Gamma Ray from Series 1329 Spectrum]).By some criteria of relevance, the log specialist/interpreter always considerably reduces the dimensionality of the log set to be made. Such reduction may be based on (1) exclusion of logs by specific judgment (e.g. visual inspection); (2) statistical techniques. The former approach tend to a reduction of logs in that the number of log to be observed is reduced. However, this method has proven to be time consuming and error prone. Based on this, there has been a move towards the use of computer aided statistical techniques for extracting/selecting meaningful logs from the original log set. This latter approach, in general, can reduce logs by showing that a subset of the logs is "important" for discriminating rocks type. This thesis is concerned with the application of this latter approach. Dimensionality reduction such as Principal Component Analysis (PCA), Factor Analysis (FA) and Linear Discriminant Analysis (LDA) are used to reduce the dimensionality of the original log set of CCSD-MH to a convenient size, holding as much of the original information as possible, and then feed the reduced-log sets into the classifiers. Five classifiers is addressed, namely, Support Vector Machines (SVM), k-Nearest Neighbor(k-NN), Back propagation neural network (BPNN), Radial Basis Function neural network(RBFN) and Self-Organizing Map(SOM) in the classification of metamorphic rocks. The strategy of combining dimensionality reduction and classifiers is demonstrated and discussed. The problems which has to be solved is to find the best suitable log set and classifier which can be used to characterize metamorphic rocks. Another area of concern that is addressed in the study is the use of cross validation technique as a way to take a better advantage of the available training data. Additionally, this thesis increase knowledge and greater ability of the log specialists using inference statistic such as Student paired t test to guide in classification decision.From the experiments results in this study, the reduced log sets found from dimensionality reduction (DR) can separate the metamorphic rocks types better or almost as well as the original log set. LDA was found to outperform PCA and FA with better performance because it deals directly with class discrimination. The results of geophysical log data classification on 5 groups of rocks; orthogneiss, paragneiss, amphibolite, eclogite, and ultramafic rocks showed that all classifiers are helpful for the lithology classification using geophysical well logs from crystalline metamorphic rocks. However, BPNN was the best among all the supervised classifers investigated. Since it show better performance in low and high dimension log spaces. Comparison of unsupervised SOM to BPNN showed that the performance of SOM and BPNN are reasonably comparable. In the sense that the magnitude of difference between SOM and BPNN is not statistically significant. The idea behind this study is not to remove the expertise and interpretive experience of a qualified geophysicist but to show how the task can be simplified and made more effective. In this way geophysicists are able to focus on the important information. This study will provide a better understanding of pattern recognition systems and their relevance in the analysis of CCSD-MH data. Last, but not least, it will add a new dimension to the existing knowledge and will be useful to the geoscientist community.