Research On Anomalous Bodies Identification In 3D Seismic Images

Seismic exploration identifies the subsurface geological structure and lithology based on seismic reflection signals.To locate subsurface oil and gas reservoirs,many research works process,interpret and analyze 3D seismic data from the view of 3D seismic images.Subsurface anomalous bodies(such as faults,karst caves and channels)usually bring abnormal changes in seismic reflection signals,which will generate anomalies in3 D seismic images.The identification of anomalous bodies in 3D seismic images is an interdisciplinary problem that involves image understanding and oil and gas exploration.Anomalous bodies identification can delineate the shape,volume and spatial distribution of anomalous bodies,which provides the basis for reservoir prediction,reserve calculation and well location deployment,and can effectively improve the success rate of oil and gas exploration.There is a significant difference in the imaging principle between seismic images and nature images,and lots of challenges are raised in seismic image processing,such as limited image resolution,complex reflection patterns and strong uncertainty.Therefore,instead of directly applying theories and methods developed for natural image processing to identify anomalous bodies in 3D seismic images,we need to improve them to adapt to the specialty of the 3D seismic image.In this dissertation,we study anomalous bodies identification from two aspects of feature extraction(i.e.,seismic attribute extraction)and seismic image segmentation,and mainly discuss the edge enhancement of anomalous bodies,the fusion of seismic attributes and the introduction of prior information.A series of methods are proposed based on signal processing,image processing and deep learning,which has achieved significant improvements in the practical application.The main work and novelties of the dissertation mainly include the following four aspects:(1)Coherence is a traditional method for edge detection of the anomalous body.However,traditional coherence cannot clearly capture the edge of the anomalous body,especially lacks continuity of the edge.To solve this problem,this dissertation presents a method for edge enhancement based on principal component analysis(PCA).Firstly,the proposed method extracts windowed data along the targeted horizon,then applies dimensionality reduction to compress redundant information,which highlights the horizontal discontinuity and greatly improves the ability to delineate the edge of the anomalous body.Furthermore,theoretical analysis on the correlation between the proposed method and texture model regression(TMR)is given,which unveils the physical meaning behind the method.Experimental results demonstrate that the enhanced coherence based on PCA can greatly enhance the ability to detect horizontal discontinuities,and delineate the edges of faults and karst caves more accurately.(2)3D seismic image segmentation is an important step of anomalous bodies identification,which is important for the exploration of oil and gas reservoir and calculation of reserves.However,due to the complex seismic wavefield caused by the anomalous body,the lack of the natural frequency of seismic signals and the existence of noise,3D seismic image segmentation based on a single seismic attribute is not enough to accurately depict the anomalous body.Aiming at the above problem,this dissertation presents a segmentation method based on the fusion of seismic attributes.Firstly,according to the distribution characteristics of seismic attributes,seismic attributes are divided into two types: edgebased and region-based.Then,a segmentation model combining edge-based and regionbased features is proposed under the framework of the level set method(LSM).Based on the characteristics of each type of seismic attribute,this method builds an adaptive model,which highlights the most effective information in each type of seismic attribute.Experimental results demonstrate that compared with the segmentation method based on a single seismic attribute,the proposed method gives more accurate segmentation results.(3)Due to the effect of earth filtering and environmental noise,there are a lot of uncertainties in the characterization of the anomalous body in the 3D seismic image,which often leads to discontinuous segmentation results and breaks the integrity of the shape of the target.To solve this problem,based on the prior that the internal region of the anomalous body is usually connected,this dissertation proposes to introduce the constraint of connectivity in the segmentation process.The connectivity constraint is formulated to a linear constraint on the tree graph by constructing the shortest path tree.The optimal solution of the segmentation model is solved on the tree graph and gives segmentation results.However,false positives would easily deteriorate segmentation results when the connectivity constraint is included.This dissertation further proposes a pruning method based on the edge feature to improve the segmentation method with the connectivity constraint.Experimental results demonstrate that the introduction of the connectivity constraint can make full use of the morphological characteristics of the anomalous body and generate more accurate segmentation results.(4)Traditional methods of anomalous bodies identification heavily depend on handcrafted seismic attributes and segmentation methods,but deep learning can automatically learn more effective features from the massive training samples,which greatly improves the accuracy of anomalous bodies identification.However,there are very few effective training samples in the field of oil and gas exploration.Deep neural networks with few training samples would fail to capture intrinsic patterns of the data,which results in poor generalization ability.Aiming at the above problem,this dissertation proposes to introduce prior information into deep neural networks,which is quite difficult because existing deep learning-based methods lack transparency.Therefore,based on the interpretability of the model,this dissertation proposes to define a regularization term to introduce prior information.Experimental results demonstrate that interpretability based priors greatly improve the accuracy of anomalous bodies identification and the interpretability of the model,and prove the feasibility of introducing the prior information to deep neural networks based on the interpretability.