Research And Application Of Seismic Random Noise Suppression Model Based On Semantics And Deep Learning

Seismic is one of the most effective methods for probing underground resources such as oil and gas.However,during the process of data collection,random noise is received by detectors alongside seismic signals.This severely diminishes the quality of seismic data,resulting in a low signal-to-noise ratio in seismic data and making subsequent interpretation work more challenging.The generation of random noise in seismic data is mainly associated with the underground propagation medium,collection environment,excitation method of artificial seismic waves,human activities,and other factors,leading to complex characteristics of random noise in seismic.In desert environments,non-Gaussian random noise with low-frequency tendencies and spatial directionality is prevalent.This type of noise is prone to spectral overlap with seismic signals,deviating from the assumptions of conventional denoising methods that are based on Gaussian noise.This mismatch can lead to inadvertent distortion of signals during the denoising process.This paper primarily addresses the challenge of suppressing strong low-frequency random noise in seismic records from desert environments with low signal-to-noise ratios.The study involves the development of noise suppression models based on semantics and deep learning,with the construction of three different approaches.The feasibility and effectiveness of these approaches were verified through theoretical derivations,simulation experiments,and practical processing of actual seismic data records.The results confirm that these methods demonstrate significant improvements in enhancing the signal-to-noise ratio of seismic records,suppressing random noise,and effectively recovering signal detail.In response to the challenges posed by low-frequency,non-Gaussian desert random noise that exhibits weak similarity with effective signals,making it difficult for denoising models to effectively distinguish between the characteristics of random noise and effective signals,the paper introduces a seismic data denoising network guided by semantic masks,referred to as the Mask Guided Model for Seismic Data Denoising(MGDNet).The MGDNet model employs a semantic convolutional neural network to predict masks containing semantic information related to seismic reflections from the noisy data.These masks,which encapsulate semantic information about seismic reflections,are used to differentiate between signals and background.Subsequently,the denoising network utilizes a Siamese network structure to extract similar features from both the masks and the noisy data.This process guides the model in learning differential features between random noise and effective signals,enabling the denoising network to more effectively filter out noise while preserving signal details.Experimental results conducted on simulated data,public datasets,and real desert seismic data demonstrate the significant effectiveness of MGDNet in suppressing desert random noise and restoring signal structures.Addressing the limitation of existing supervised deep learning seismic denoising models,which rely on simulated training-label data pairs and may not adequately capture real-world scenarios,this paper proposes a self-supervised mask-guided model for seismic data denoising,referred to as S-MGDNet.The S-MGDNet model leverages the inherent features of input data to construct constraints,eliminating the need for clean data labels in the denoising process.S-MGDNet adopts an unsupervised clustering approach to obtain semantic masks from the noisy data and introduces maximum mutual information to constrain the denoising direction of the model.By using semantic masks as prior guidance and employing maximum mutual information as a constraint,S-MGDNet enables the denoising model to learn the mapping from noisy to clean data using only the noisy data itself as training input,achieving selfsupervised denoising.Results from both simulated data and real desert seismic data demonstrate the effectiveness of the proposed method in suppressing desert random noise while preserving the content of effective signals.Addressing the challenge of pixel-level feature recovery in low signal-to-noise ratio seismic images within complex backgrounds,and aiming to completely suppress random noise while fully restoring effective signal content to avoid suboptimal denoising results,this paper proposes a novel Seismic Random Noise Suppression Model based on Downsampling and Super-resolution Reconstruction(DASRNet).The model utilizes a downsampling network with high fitting capabilities to extract effective signal features,producing low-resolution data containing only the semantic information of the effective signal.Employing a super-resolution network with this low-resolution data as a foundation,DASRNet reconstructs and restores the structure and content of the effective signal.DASRNet decouples the denoising task from the signal restoration task,reducing the difficulty of suppressing random noise and restoring effective signal content,while also lowering the likelihood of suboptimal outcomes.Evaluation on different types of seismic data demonstrates that DASRNet achieves excellent and reliable results in suppressing random noise and restoring effective signal content,with a maximum signal-to-noise ratio improvement of up to 34 d B on synthetic data.