Adaptive Seismic Denoising Based On Expected Patch Log-likelihood And Reinforcement Learning

Seismic exploration utilizes the propagation and reflection characteristics of seismic waves to explore subsurface structures and resource distribution.It is characterized by high resolution,extensive depth coverage,and adaptability to various environments.However,the seismic data are often contaminated by multi-type noise,which interferes with the subsequent accurate interpretation of seismic data and imaging of subsurface geological structures.Therefore,seismic denoising plays importent research significance.It not only improves the signal-to-noise ratio and clarity of seismic data,but also plays a crucial role in enhancing the quality of seismic data and refining the analysis of seismic structures.Complex geological conditions and diverse acquisition environments,such as the heterogeneity of subsurface media,seismic source characteristics,near-surface and surface effects,as well as external environmental interferences,result in significant spatiotemporal variability of seismic signal and noise characteristics.Emerging distributed acoustic sensing(DAS)technology offers advantages such as low cost and high sampling density.However,its highly sensitive instruments and fiber-optic observation systems introduce more complex types of noise,leading to more pronounced spatiotemporal variations in the collected seismic data.This variability places higher demands on the adaptivity of seismic random noise suppression techniques.This dissertation focuses on spatiotemporal seismic noise suppression and proposes a serious of adaptive denoising schemes according to the spatiotemporal varying degrees of seismic signal and noise characteristics.These schemes include expected patch loglikelihood total variation algorithm(patch adaptive),reinforcement learning based seismic denoising model(signal structure adaptive),and task-aware denoising model(denoising task adaptive).The adaptive ways and filtering operations are formulated based on the noise characteristics.For seismic low-frequency seismic random noise that severely contaminates and overlaps with signals in certain regions,this dissertation proposes the expected patch log likelihoodtotal variation(EPLL-TV)algorithm,which combines the EPLL prior with the TV prior to achieve patch adaptive denoising.The EPLL-TV algorithm leverages a Gaussian mixture model(GMM)to establish the EPLL prior for signal patches.It selects Gaussian components from a pre-learned GMM that exhibit the highest likelihood of matching signal features.The TV prior is then incorporated to capture structural characteristics within the current data,enhancing the gradient sparsity of seismic data to suppress low-frequency random noise and curbing its impact for Gaussian component selection.Notably,the seismic events with complex structures can be faithfully recovered via accurate selection of gaussian components,thereby providing a more reliable structure for the TV prior constraint.The interaction between these dual priors facilitates effective denoising.The establishment of a GMM for signal patches allows EPLL-TV to automatically adjust parameters based on the spatiotemporal variations of signals and noise,resulting in excellent denoising performance and adaptability.The proposed dual-prior constrained model is solved by the alternating direction method of multipliers(ADMM)and reconstructed in the Fourier domain to reduce computational complexity.The proposed EPLL-TV algorithm has significant advantages in recovering seismic events with complex structures and suppressing low-frequency seismic random noise with similar waveforms to the signal.For non-stationary seismic random noise characterized by spatiotemporal variations in noise intensity,this dissertation proposes the reinforcement learning based seismic denoising(RLSD)model with an asynchronous advantage actor-critic(A3C)framework to achieve structure adaptive denoising.Within the A3 C framework,the RLSD agent learns denoising policies for samples by a policy network and selects appropriate filters from the predefined action space.The value network guides the adjustment of policy network using expected accumulated reward,progressively achieving the goal of finding policies that maximize accumulated rewards.The action space consists of multiple simple and efficient seismic filters with varying parameters.Region adaptive weighted reward function and local similarity function are employed to respectively evaluate the filtering performance of the policy on the dataset to be trained and the data to be processed.The curriculum learning is employed to enhance the RLSD model’s convergence performance on complex seismic data and its adaptability to the data to be processed.The training dataset transition from stationary data to non-stationary data during the training process,and then to the data to be processed during the denoising process to fine-tuning the model.The proposed RLSD model automatically adjust denoising policies based on complex signal structure and noise levels,exhibiting exceptional performance in preserving signals and suppressing non-stationary seismic random noise.For multi-type mixing noise characterized by significant spatiotemporal variations,this dissertation proposes the reinforcement learning based task aware denoising(RL-TAD)model,which autonomously identifies denoising tasks in distinct regions and provides intelligent denoising policies.The presence of complex multi-type mixing seismic noise causes more pronounced spatiotemporal variations,imposing higher demands on the adaptability of denoising methods.While deep network architectures significantly enhance suppression effectiveness against multi-type noise,the subsequent computational burden restricts widespread applicability.Practical experience indicates that partitioned denoising is comparatively easier,as smaller regions tend to exhibit relatively singular noise types and simpler characteristics.Hence,the proposed RL-TAD model combines multiple single-noise intelligent filtering actions and leverages reinforcement learning to automatically identify denoising tasks,deconstructing the intricate multi-type noise suppression task into more manageable single-noise suppression task.The RL-TAD model employs lightweight residual networks as filtering actions,with each action tailored to address the specific noise.The hidden layers of intelligent filtering actions employ dilated convolutions to expand the receptive field within limited network depths,resulting in improved filtering effect.Deep reinforcement learning agent is utilized to perceive distinct denoising tasks across various regions and generate action selection policies.The downsampling policies is employed to align with the seismic data’s slow temporal changes,which aids in reducing computational complexity while avoiding artifacts stemming from inconsistent action selection policies in adjacent sample.The proposed RL-TAD model automatically identifies different denoising tasks and offers denoising policies,exhibiting excellent filtering results on seismic data affected by multi-type noise.By investigating the advantages of the EPLL method and deep reinforcement learning in the context of adaptive processing,this dissertation proposes a series of adaptive denoising schemes for handling low-frequency random noise,non-stationary random noise,and multitype mixing noise with complex spatiotemporal variations characteristic.The proposed schemes are evaluated on synthetic seismic records and field seismic records obtained from different survey areas such as desert,forested,mountainous and different acquisition techniques such as conventional seismic detector,DAS seismic detector.The experiment results confirm the significant benefits of the proposed schemes in enhancing signal-to-noise ratio,suppression complex seismic noise,and preserving valuable seismic signals.This research establishes a solid foundation for subsequent tasks such as seismic inversion,highprecise imaging,and seismic attribute analysis by providing higher quality seismic data with improved signal-to-noise ratios.