First-break Research Articles

A Comparative Analysis of Machine Learning Models for First-break Arrival Picking

First-break (FB) picking is an important and necessary step in seismic data processing and there is a need to develop precise and accurate auto-picking solutions. Our investigation in this study includes eight machine learning models. We use 1195 raw traces to extract several features and train for accurate picking and monitoring the performance of each model using well-defined evaluation metrics. Careful investigation of the scores shows that a single metric alone is not sufficient to evaluate the arrival picking models in real-time. Correlation analysis of predicted probabilities and predicted classes of machine learning models confirm that the performance metrics that use predicted probabilities have higher score value than those that use predicted classes. Our study which incorporates comparisons of different machine learning models based on different performance metrics, training time, and feature importance indicates that the approach we developed in this study is helpful and provides an opportunity to determine the real-time suitability of different methodologies for automatic FB arrival picking with clear deep insight. Based on performance scores, we bench-marked the Extra Tree classifier as the most efficient model for FB arrival picking with accuracy and F1-score above 95%.

Multi-pattern algorithm for first-break picking employing open-source machine learning libraries

Accurate first-break (FB) onsets are crucial in processing seismic data, e.g. for deriving statics solution or building velocity models via tomography. Ever-increasing volumes of seismic data require automatic algorithms for FB picking. Most of the existing techniques are based either on simple triggering algorithms that rely on basic properties of seismic traces or on neural network architectures that are too complex to be easily trained and re-used in another survey. Here we propose a solution that addresses the issue of multi-level analysis using a time-sequence pattern-recognition process implemented in the newest open-source machine learning library like Keras or Scikit-Learn. We use well-established methods such as STA/LTA, entropy, fractal dimension or higher-order statistics to provide patterns required for generative model training with artificial neural networks (ANN), Support Vector Regression (SVR) and an implementation of gradient boosted decision trees – XGBoost (Extreme Gradient Boosting). FB picking is cast as the binary classification that requires a model to differentiate FB sample from all other samples in a seismic trace. Our workflow (provided freely as a Jupyter Notebook) is robust, easily adaptable and flexible in a way of adding new pattern generators that might contribute to even better performance, while already-trained models can be saved and re-used for another dataset collected with similar acquisition parameters (e.g., in multi-line surveys). Application to real seismic data showed that the models trained on 1000 and 340,100 manually-picked FB onsets are able to predict FB on the rest of 470,000 of traces with the success rate of nearly 90% and 95%, respectively.

Seismic Waveform Classification and First-Break Picking Using Convolution Neural Networks

Regardless of successful applications of the convolutional neural networks (CNNs) in different fields, its application to seismic waveform classification and first-break (FB) picking has not been explored yet. This letter investigates the application of CNNs for classifying time-space waveforms from seismic shot gathers and picking FBs of both direct wave and refracted wave. We use representative subimage samples with two types of labeled waveform classification to supervise CNNs training. The goal is to obtain the optimal weights and biases in CNNs, which are solved by minimizing the error between predicted and target label classification. The trained CNNs can be utilized to automatically extract a set of time-space attributes or features from any subimage in shot gathers. These attributes are subsequently inputted to the trained fully connected layer of CNNs to output two values between 0 and 1. Based on the two-element outputs, a discriminant score function is defined to provide a single indication for classifying input waveforms. The FB is then located from the calculated score maps by sequentially using a threshold, the first local minimum rule of every trace and a median filter. Finally, we adopt synthetic and real shot data examples to demonstrate the effectiveness of CNNs-based waveform classification and FB picking. The results illustrate that CNN is an efficient automatic data-driven classifier and picker.

PQC: A novel approach for robust automatic near-surface analysis in low-relief geology

A novel approach, named pQC, was developed to automatically derive near-surface velocities and statics by applying surface-consistent analysis to refraction data, where the data consist of first-break (FB) traveltimes and early-arrival waveforms. This approach is facilitated by the introduction of a new sorting domain in which FB traveltimes and traces are organized in a multidimensional space consisting of a common-midpoint (CMP) location, offset, and azimuth, called the XYOA domain. The FB traveltime data sorted in this domain can be analyzed efficiently with statistical methods to obtain localized average depth domain velocity trends for calculation of long-wavelength statics. Short-wavelength (residual) refraction statics are then calculated based on the cross correlation of early-arrival waveforms in the XYOA gathers over multiple offset classes and inverted for surface-consistent source and receiver corrections. The application of the combined statics solution to the traces enhances the continuity and power of the stacked reflected events without the need to perform tomography. Surface-consistent analysis of refraction data in the XYOA domain is fully automatic, provides robust solutions for noisy data, and enhances subsequent reflection residual statics calculations. This method can also be exploited for quality-control analysis of FB picks and for automating, in an iterative procedure, the picking of large 3D seismic data sets. The method is demonstrated on a 3D land survey over complex geology where the novel refraction surface-consistent solution enhances the imaging results and employs a fraction of the time needed by a typical tomographic workflow.

Fully automated near-surface analysis by surface-consistent refraction method

Industry practices for near-surface analysis indicate difficulties in coping with the increased number of channels in seismic acquisition systems, and new approaches are needed to fully exploit the resolution embedded in modern seismic data sets. To achieve this goal, we have developed a novel surface-consistent refraction analysis method for low-relief geology to automatically derive near-surface corrections for seismic data processing. The method uses concepts from surface-consistent analysis applied to refracted arrivals. The key aspects of the method consist of the use of common midpoint (CMP)-offset-azimuth binning, evaluation of mean traveltime and standard deviation for each bin, rejection of anomalous first-break (FB) picks, derivation of CMP-based traveltime-offset functions, conversion to velocity-depth functions, evaluation of long-wavelength statics, and calculation of surface-consistent residual statics through waveform crosscorrelation. Residual time lags are evaluated in multiple CMP-offset-azimuth bins by crosscorrelating a pilot trace with all the other traces in the gather in which the correlation window is centered at the refracted arrival. The residuals are then used to build a system of linear equations that is simultaneously inverted for surface-consistent shot and receiver time shift corrections plus a possible subsurface residual term. All the steps are completely automated and require a fraction of the time needed for conventional near-surface analysis. The developed methodology was successfully performed on a complex 3D land data set from Central Saudi Arabia where it was benchmarked against a conventional tomographic work flow. The results indicate that the new surface-consistent refraction statics method enhances seismic imaging especially in portions of the survey dominated by noise.