Study On Division Method Of Seismic Facies In Three Seismic Data

In geological exploration,as seismic feature of different sediments in a sedimentary environment in a certain area,the seismic facies directly reflects the stock and patterns of minerals,such as oil,gas and coal.Accurate division of seismic facies can provide a detailed understanding of reservoir types,tectonic styles,transport systems and accumulation processes to achieve precise guidance for oil and gas exploration,reduce exploration risks,and generate enormous economic and social benefits.The division of seismic facies images is the core step of seismic facies division.At present,the division of seismic facies images mainly relies on geological experts to manually segment based on core,debris,drilling lithology and well logs.This method requires a lot of time and effort when performing geological exploration in large geological work areas or multiple small geological work areas.Deep learning is one of the latest trends in machine learning and artificial intelligence.It is good at extracting features from a large number of images and mimicking the human brain to interpret datas,like texts and images.It can also help geological experts divide seismic facies more efficiently and accurately.In this thesis,we introduces the deep learning algorithms into the division of seismic facies and replaces the traditional manual segmentation by autonomous learning features of the seismic facies images.The main work and innovations of this thesis are as follows:(1)A seismic facies image generation method based on Generative Adversarial Nets is proposed.In order to solve the problem that the traditional Generative Adversarial Nets can't work well on small datasets,we proposes a Boundary Equilibrium Generative Adversarial Networks for Limited Data,BeLiGAN.The model uses the Gaussian mixture model as the input,and the autoencodes as the discriminator.It utilizes the powerful fitting ability of the Gaussian mixture model and the reconstruction error of the autoencodes to accelerate the convergence of the generator.This makes BeLiGAN has the ability to generate seismic facies images on small datasets.(2)A U-Net-based seismic facies images Segmentation Method is proposed.in order to solve the problem that the traditional U-Net cannot accurately classify the boundary pixels with the same receptive field around the target pixels to be segmented,we add a Guided Superpixel Filter Layer into U-Net.The layer uses the result of superpixel enhancement as a guiding image to filter the the results of traditional U-Net semantic segmentation.This makes the semantic segmentation of U-Net have a more accurate result and a smoother boundary.