3D Image Classification Method And Application Research

In order to exploit oil and gas more efficiently, especially to exploit unconventional reservoirs like shale reservoirs, seismic exploration represented by waveform classifica-tion is needed in many industrial fields for adequate analysis and interpretation for the reservoir areas. A more comprehensive and effective waveform classification technology can achieve more accurate seismic phase diagram of reservoir areas, which can further improve the veracity and reliability of the analyses of the underground strata and the lithology, and then save the production cost. This thesis first reviews the development status of waveform classification technique in seismic exploration and analysis of the deficiency of the existing technology. Due to the scarcity of labeled instances in the reservoir areas which is caused by the high cost of oil and gas well drilling, the current seismic waveform classification technique in seismic exploration can't handle those less mature reservoir areas efficiently. This thesis introduces the Transfer Learning ideology and proposes a new 3D seismic image classification algorithm to solve this problem. In this thesis, the main work and innovation are as follows:1.In view of the insufficiency of feature extraction method of the current 3D seis-mic image waveform classification technology, which can't express the relationship be-tween the geographic adjacent points or don't have lipschitz continuity, this thesis applies the Three-Dimensional Scattering Wavelet Transform into 3D seismic image processing. Scattering Wavelet Transform can make the signal feature expressions more complete with lipschitz continuity. Besides, the translation invariant characteristic of the feature expressions can solve the need of local invariance hypothesis in the following Transfer Learning Classification. This thesis expounds the principle of 3D Scattering Wavelet Transform and applies it into the feature engineering of the waveform classification, and achieves an ideal and excellent result in the tests with the actual reservoir areas data.2.1n view of the present situation of labeled instance scarcity in current seismic ex-ploration, this thesis proposes an application framework of the Transfer Learning in seis-mic waveform classification, which is all based on the ideas of Transfer Learning. At the same time, this thesis proposes the Modified Domain Adaption Support Vector Ma-chine algorithm. Besides, based on the Transfer Learning application framework of seis-mic waveform classification, this thesis successfully transfers the knowledge from some completely labeled seismic images (full well data collection of seismic signals) to the waveform classification of unlabeled seismic images (no well data collection of seismic signals) by the Modified Domain Adaption Support Vector Machine classification model, which achieves the supervised classification of unlabeled seismic image. Through test on actual reservoir areas, can successfully demonstrate the effectiveness of the proposed Transfer Learning model.Finally, this thesis prove its effectiveness by obtained a ideal and good result in the actual reservoir areas of 3D seismic images experiments.