Robust seismic data denoising via self-supervised deep learning

We investigated the estimation of the ensemble channelized Hotelling observer (CHO) signal-to-noise ratio (SNR) for ordered-subset (OS) image reconstruction using noisy projection data. Previously, we computed the ensemble CHO SNR using a method for approximating the channelized covariance of OS reconstruction, which requires knowledge of the noise-free projection data. Here, we use a “plug-in” approach, in which noisy data is used in place of the noise-free data in the aforementioned channelized covariance approximation. Additionally, we evaluated the use of smoothing of the noisy projections before use in the covariance approximation. Additionally, we evaluated the use of smoothing of the noisy projections before use in the covariance calculation. The task was detection of a 10% contrast Gaussian signal within a slice of the MCAT phantom. Simulated projections of the MCAT phantom were scaled and Poisson noise was added to create 100 noisy signal-absent data sets. Simulated projections of the scaled signal were then added to the noisy background projections to create 100 noisy signal-present data set. These noisy data sets were then used to generate 100 estimates of the ensemble CHO SNR for reconstructions at various iterates. For comparison purposes, the same calculation was repeated with the noise-free data. The results, reported as plots of the average CHO SNR generated in this fashion, along with 95% confidence intervals, demonstrate that this approach works very well, and would allow optimization of imaging systems and reconstruction methods using a more accurate object model (i.e., real patient data).

The aim of this article is to present a dynamic record linkage solution using high level machine learning to solve the problems associated with noisy, inconsistent and dynamic data sets. Classic deterministic and probabilistic models fail to work in such environments because they are static and based on assumptions. Incorporating supervised learning algorithm, active learning, and ensemble techniques such as random forests and boosting allows the proposed structure to change in response to different data types and thus become more accurate and scaleable. Some of its highlights are feature selection to ensure match accuracy, clustering to support noisy inputs, and active learning to minimize reliance on large labeled datasets. Simulations with real data, such as government or healthcare data, show that compared with the traditional approach, linkage is significantly more accurate and efficient. The flexible model led to 15% higher F1-scores on noisy datasets and was scalable across large data sets. This work demonstrates how adaptive machine learning can revolutionize the modern record linkage tasks and provides a powerful and effective solution to ever-changing data conditions.

Seismic erratic noise, characterized by large isolated events following non-Gaussian distributions, significantly degrades seismic data quality by masking useful signals. Methods based on conventional priors remain essential but face inherent challenges as they struggle to balance noise attenuation and signal preservation. Supervised deep learning approaches are constrained by the scarcity of high-quality labeled training pairs, while existing unsupervised techniques often suffer from suboptimal accuracy and high computational cost. To address these limitations, we propose an unsupervised deep learning framework based on implicit neural representation (INR) for erratic noise suppression in seismic data. The proposed method employs Fourier feature mapping to encode the spatial coordinates of noisy seismic data, which are then processed by a lightweight multilayer perceptron (MLP). The MLP is optimized using a robust Huber loss function to learn a continuous representation of the underlying seismic wavefield, effectively attenuating erratic noise while preserving valuable signal components. The Fourier feature mapping enhances the MLP’s ability to capture high-frequency signal details, while the Huber loss adaptively weights residuals based on amplitude, enabling precise noise suppression. Experimental results on synthetic and field datasets demonstrate its superior performance in suppressing noise while preserving signal fidelity.

The collection of a range of different seismic data types has greatly improved our understanding of the crustal architecture of Australia's Archaean Yilgarn Craton over the last few years. These seismic data include broadband seismic studies, seismic receiver functions, wide-angle recordings and mine-scale to deep seismic reflection transects. Each data set provides information on the three-dimensional (3D) tectonic model of the Yilgarn Craton from the craton scale through to the mine scale. This paper demonstrates that the integration and rationalisation of these different seismic data sets into a multi-scale 3D geological/seismic model, that can be visualised at once in a single software package, and incorporating all available data sets, significantly enhances this understanding. This enhanced understanding occurred because the integrated 3D model allowed easy and accurate comparison of one result against another, and facilitated the integrated questioning and interrogation across scales and seismic method. As a result, there are feedback questions regarding understanding of the individual seismic data sets themselves, as well as the Yilgarn Craton as a whole. The methodology used, including all the data sets in the model range, had to allow for the wide range of data sets, frequencies and seismic modes. At the craton scale, P-wave, S-wave and surface wave variations constrained the 3D lithospheric velocity model, revealing noticeable large-scale velocity variations within and across the craton. An interesting feature of the data, easily identified in 3D, is the presence of a fast S-wave velocity anomaly (> 4.8 km s − 1 ) within the upper mantle. This velocity anomaly dips east and has a series of step-down offsets that coincide approximately with province and terrane boundaries of the Yilgarn Craton. One-dimensional receiver function profiles show variations in their crustal velocity across the craton. These crustal velocity variations are consistent with the larger-scale geological subdivision of the craton, and provide characteristic profiles for provinces and terranes. The receiver function results and the deep seismic reflection data both agree on the depth to the Moho, and both indicate an increase in Moho depth to the east. The 2D seismic refraction results in the south-west of the craton provide crustal thickness information, an indication of middle and lower crustal compositions, and information regarding the broad-scale architectural framework. At the province- and terrane-scale, the deep seismic reflection data and the mine-scale seismic data provide geometric constraints on crustal architecture, in particular the orientation of the region's fault systems as well as variations in the thickness of the granite–greenstone succession. Integration of the results from wide-angle seismic refraction data coincident with the deep seismic reflection data provided additional constraints on likely upper crustal lithologies. The integrated 3D seismic model implies the dominant geodynamic process involved the development of an orogenic belt that developed with a series of contractional (folding and thrusting) events, separated by equally important extensional events. The seismic reflection data in particular suggests that extensional movement on many shear zones was more common than previously thought. The seismic reflection data suggest that the dominant mineral systems involved deeply sourced fluid flowing up crustal-penetrating shear zones. These deeply sourced fluids were further focussed into sites located above fault-breached domal regions in the upper crust.

Mutual calibration training: Training deep neural networks with noisy labels using dual-models

It can be challenging to pick high-quality first arrivals on noisy seismic data sets. The stability and smoothness criteria of the picked first arrival are not satisfied for data sets with shingles and interference from unexpected and backscattered events. To improve first-arrival picking, we have adopted an automatic first-arrival picking workflow using global path tracing (GPT) to find a global solution for first-arrival picking with the condition of smoothness of the traced path. Our methodology is composed of data preconditioning, GPT, and a final addition of traced and piloted traveltimes to compute the total picked traveltime. We propose several ways to precondition the data set, including the use of the amplitude and amplitude ratio with and without a pilot. A 2D GPT is composed of two steps, namely, accumulation of energy on the potential path and backtracking of the optimal path with a strain factor for smoothness. For higher dimensional data sets, two strategies were adopted. One was to split the higher dimension data into subdomains of two dimensions to which 2D GPT was applied. The alternative method was to smooth the preconditioned data set in directions except for the one used to trace the path before applying 2D GPT. We discuss the importance of choosing the proper parameters for data preconditioning and for constraining GPT. We demonstrate the robustness and stability of our automatic first-arrival picking via GPT using synthetic and field data examples.

Accurate blood glucose level prediction is crucial for effective diabetes management, enabling timely interventions and reducing complications. The limitations of traditional machine learning models like Random Forest and XGBOOST include the inability to deal with noisy data and datasets with imbalances. This study proposes a hybrid ensemble model that combines Random Forest for robust feature selection and initial predictions with XGBOOST to improve accuracy by focusing on instances that have been misclassified. The hybrid approach improves the handling of imbalanced and noisy datasets, achieving better accuracy and generalization. The model's superiority over standalone machine learning models is demonstrated by experimental results on publicly accessible datasets, highlighting its potential as a reliable method for predicting blood glucose levels and supporting diabetes management.

Seismic swell noise, often observed in marine seismic data, is characterized by high amplitude and low frequencies. This noise significantly hides useful signals, underscoring the importance of attenuating it in the processing pipeline for marine seismic data. To date, most deep learning methods proposed for swell noise suppression have focused on supervised paradigms, which require a large dataset of paired noisy and clean data to gain insights into signal features and swell noise statistics at the training phase. In field data processing, however, obtaining generalizable training data often proves to be challenging. The inherent complexity and variability of the real world often make it difficult to acquire unbiased, noise-free datasets in the context of marine seismic data processing. To address this problem, we present a self-supervised deep learning method for suppressing swell noise even in the absence of access to clean seismic training data. First, we propose a strategy for the synthetic swell noise generation based on the amplitude, frequency, and coherence features of noisy traces. We implement the noisy-as-clean (NAC) strategy, wherein either the original noisy seismic data or its reorganized variant acts as the network’s target. Simultaneously, the network receives the observed noisy seismic data combined with simulated swell noise as its input. By leveraging these “noisy-noisy” pairs, we train a DnCNN network. Experimental evaluations conducted on both synthetic and field data demonstrate that, through the integration of swell noise simulation and the NAC strategy, the trained network consistently achieves superior denoising performance.

In cluster analysis, distinguishing abnormal behavior from noise is an important problem that has remained unresolved for many years. This paper presents a general approach for distinguishing abnormal behavior from noise and for identifying abnormal behavior in noisy data sets. The approach is an application of a new, complete linkage hierarchical clustering method for n·(n−1) over 2 + 1-level hierarchical sequences and a means for finding meaningful levels of such hierarchical sequences prior to performing a cluster analysis. These technologies were designed with small-n, large-m data sets in mind. Based on four broadly applicable assumptions, the approach uses higher dimensionalities to reveal inherent structure in noisy data sets and find meaningful levels in the corresponding hierarchical sequences. The new clustering method is used to construct only the cluster sets that correspond to these levels. Results from a first experiment show how the effects of noise are attenuated as the dimensionality of the data points increases. Results from a second experiment show how meaningful cluster sets can have real world meanings that are useful for identifying abnormal behavior.

Sparse domain thresholding filters operating in a sparse domain are highly effective in removing Gaussian random noise under Gaussian distribution assumption. Erratic noise, which designates non-Gaussian noise that consists of large isolated events with known or unknown distribution, also needs to be explicitly taken into account. However, conventional sparse domain thresholding filters based on the least-squares (LS) criterion are severely sensitive to data with high-amplitude and non-Gaussian noise, i.e., the erratic noise, which makes the suppression of this type of noise extremely challenging. In this paper, we present a robust sparsity-promoting denoising model, in which the LS criterion is replaced by the Huber criterion to weaken the effects of erratic noise. The random and erratic noise is distinguished by using a data-adaptive parameter in the presented method, where random noise is described by mean square, while the erratic noise is downweighted through a damped weight. Different from conventional sparse domain thresholding filters, definition of the misfit between noisy data and recovered signal via the Huber criterion results in a nonlinear optimization problem. With the help of theoretical pseudoseismic data, an iterative robust sparsity-promoting filter is proposed to transform the nonlinear optimization problem into a linear LS problem through an iterative procedure. The main advantage of this transformation is that the nonlinear denoising filter can be solved by conventional LS solvers. Tests with several data sets demonstrate that the proposed denoising filter can successfully attenuate the erratic noise without damaging useful signal when compared with conventional denoising approaches based on the LS criterion.

We present an approach to effectively use millions of images with noisy annotations in conjunction with a small subset of cleanly-annotated images to learn powerful image representations. One common approach to combine clean and noisy data is to first pre-train a network using the large noisy dataset and then fine-tune with the clean dataset. We show this approach does not fully leverage the information contained in the clean set. Thus, we demonstrate how to use the clean annotations to reduce the noise in the large dataset before fine-tuning the network using both the clean set and the full set with reduced noise. The approach comprises a multi-task network that jointly learns to clean noisy annotations and to accurately classify images. We evaluate our approach on the recently released Open Images dataset, containing ~9 million images, multiple annotations per image and over 6000 unique classes. For the small clean set of annotations we use a quarter of the validation set with ~40k images. Our results demonstrate that the proposed approach clearly outperforms direct fine-tuning across all major categories of classes in the Open Image dataset. Further, our approach is particularly effective for a large number of classes with wide range of noise in annotations (20-80% false positive annotations).

In seismic data processing, denoising is crucial for identifying the Earth’s subsurface structure. Erratic noise in seismic data is challenging because of its high amplitude and non-Gaussian distribution characteristics. Dictionary learning with KSVD requires a longer computational time due to its iterative nature at various stages. To address this, we propose a robust denoising DHHL method that utilises mini-batch dictionary learning, the quantum Harrow–Hassidim–Lloyd (HHL) algorithm, and quantum singular value transformation (QSVT) to attenuate random and erratic noise in seismic data. The HHL algorithm provides a significant estimate of the sparse coefficient. The HHL provides a theoretically efficient method for sparse coding with favorable asymptotic scaling, improving the potential scalability of the method. Mini-batch dictionary learning constructs hierarchical dictionaries, thereby enhancing the capture of signal features in seismic data. QSVT enables the direct transformation of singular values, leading to faster convergence and more accurate sparse representations for dictionary updating. Here, we use minimisation loss using ℓ<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2,1</sub> to provide group sparsity and preserve the signal features. The proposed DHHL method was applied to both synthetic and field datasets. The efficiency of the DHHL method was assessed using evaluation metrics such as the signal-to-noise ratio (SNR), mean squared error (MSE), and local similarity map. The quantitative and visual results indicate that the DHHL method achieved a higher SNR, minimised the MSE with less signal leakage, and reduced the computational time by accelerating sparse coding.

The fit of a sum of exponentials to noisy data

In this paper, a hidden Markov model (HMM)-based cue parameters estimation method for single-channel speech enhancement is proposed, in which the cue parameters of binaural cue coding (BCC) are applied to single-channel speech enhancement system successfully. First, the clean speech and noise signals are considered as the left and right channels of stereo signal, respectively; and the noisy speech is treated as the down-mixed mono signal of BCC method. According to the clean speech and noise data set and the corresponding noisy speech data set, the clean cue parameters and pre-enhanced cue parameters are extracted, respectively. Then the cue HMM is trained offline, which exploits the a priori information about the clean cue parameters and the pre-enhanced cue parameters for speech enhancement. Next, using the trained cue HMM, the clean cue parameters are estimated from noisy speech online. Finally, following the synthesis principle of BCC cue parameters, the speech estimator is constructed for enhancing noisy speech. The test results demonstrate that, for the segmental signal-noise-ratio (SNR), the log spectral distortion and PESQ measures, the proposed method performs better than the reference methods.

We present a novel, physically-motivated method for surface reconstruction that can recover smooth surfaces from noisy and sparse data sets, without using orientation information. A new volumetric technique based on regularized-membrane potentials for aggregating the input sample points is introduced, which manages improved noise tolerability and outlier removal, without sacrificing much with respect to detail (feature) recovery. In this method, sample points are first aggregated on a volumetric grid. A labeling algorithm that relies on intrinsic properties of the smooth scalar field emerging after aggregation is used to classify grid points as exterior or interior to the surface. We also introduce a mesh-smoothing paradigm based on a mass-spring system, enhanced with a bending-energy minimizing term to ensure that the final triangulated surface is smoother than piecewise linear. The method compares favorably with respect to previous approaches in terms of speed and flexibility.