Research On Seismic Image Processing And Evaluation Methods Based On Deep Learning

Oil and gas resources are essential strategic resources for every country in the world in the industrial period.Seismic exploration is a necessary technology in the quest for subsurface resources and in supporting scientists in determining the geological structure.The clarity of seismic images is closely related to the outcomes of seismic interpretation when converting seismic survey data into image form.Seismic data,on the other hand,frequently contains a lot of noise,resulting in a seismic image provided to seismologists with a number of distracting variables.As a result,the focus of this article is on developing a new seismic data noise reduction approach as well as a seismic picture quality assessment method that does not rely on reference images.The following are the key responsibilities:(1)A tensor neural network-based unsupervised 3-D seismic data noise reduction approach is proposed.Popular deep learning approaches have shown to be effective in identifying the tight structure hidden behind tainted seismic data.However,most existing matrix-based deep learning degradation is due to the difficulty of obtaining real seismic data without noise in the real world.Because noise algorithms are unable to automatically describe such high-dimensional structures in an unsupervised manner,they may be unable to execute noise reduction tasks on 3D seismic data efficiently.In the absence of clean real seismic data,this research offers a tensor convolutional neural network-based data denoising strategy that employs Stein's unbiased risk estimation to discover the underlying high-dimensional structure.In this paper,following the nature of transform-based tensor-tensor product,after further derivation,the weight parameters of the tensor convolutional neural network are determined by implementing a matrix-based convolutional neural network on each front slice of the time-frequency domain(e.g.,wavelet domain)of the seismic data.Experiments on synthetic and real data show that the proposed model outperforms the current state-of-the-art methods.(2)The use of meta-learning to judge the quality of reference-free seismic images is proposed.Most existing deep learning-based image quality assessment metrics are based on pre-trained networks? however,because these pre-trained networks were not designed for seismic images and because there are many different types of seismic image distortions,these networks often have poor generalization ability when applied to seismic image quality assessment problems.At the same time,the quality of the dataset,particularly labeled data,is critical for training deep neural networks.This study employs synthetic data to compute the local similarity of images as labels by noise addition,which forms the training data pair,in order to overcome the aforesaid challenges.The dataset is divided into a support set and a query set in this research,and a second-level gradient optimization method is utilized to acquire the quality prior knowledge shared by these varied distortions in order to make the proposed model perform well on both support and query sets.Finally,the target dataset is fine-tuned in order to swiftly obtain a quality evaluation model.Extensive testing has revealed that the suggested metrics perform well on data sets with streaking,irregular,and Gaussian noise.