Advancing the First Break Picking with Neural Networks
The very beginning of O&G exploration is seismic surveying, when the raw data that represents the geology of the region of interest is obtained. This data is used to construct a geological model that allows making accurate conclusions about oil reservoirs in the area. Precise geological modeling, in its turn, cannot be carried out without proper seismic interpretation, in particular Horizon Detection, which requires raw data to be denoised, cleaned, and structured — and that is the purpose of Seismic Processing stage.
The whole processing pipeline consists of lots of stages. In this article, we focus on one of the initial steps — the First Break Picking.
The First Break Picking
The task can be formulated as follows: given a seismic trace, determine arrival time of the wave traveling from source to receiver.
These times are later used for studying the near-surface velocities and subsequently performing static corrections, the procedure that eliminates the effect of varying elevations of receivers and sources. It allows modeling the seismic survey as if it took place not on the earth's surface, but instead on some flat datum plane.
Static corrections directly affect resulting seismic image: the well-performed procedure can reveal, extend, and amplify the horizons! On the other hand, inaccurate first break picking dramatically decreases coherence of traces after corrections.
How is it being traditionally solved?
There are several basic approaches to solve the first break picking task. They are usually based on the concepts of energy in moving windows, cross-correlations, or calculating higher-order statistics of the trace. We focus on the windows-based method, as its variations are most commonly used in the Oil&Gas software tools for Seismic Processing.
The STA/LTA (short term average — long term average) algorithm is based on the ratio of amplitudes in two moving windows: the sudden increase in the ratio indicates the arrival time. See the visualization of the method below: the first subplot shows a seismic trace with a low level of noise before the first break, the second — STA/LTA ratio function, the third — the derivative of this function. The moment when the derivative exceeds the manually chosen threshold for the first time (or ultimately when the derivative reaches its maxima) is declared as the first break.
This method works quite well on simple data, but it is not robust to noisy traces! See the performance of the algorithm on the “complex” trace: the significant level of noise before the first break leads to misprediction.
Unfortunately, the data usually contains a considerable amount of such traces, depending on the survey conditions. In our practice, we worked with datasets containing up to 10% of such traces, and that is not the limit! This makes picking time-consuming since it takes weeks for specialists to manually adjust method parameters (windows lengths, threshold) across the survey.
That's one of the motivations for using ML models here — to automate the stage and reduce human efforts.
Neural Network for First Break Picking
In terms of Machine Learning, it is intuitive to treat this task as Regression, when given a seismic trace, the model predicts a single value — time. Such an approach, however, did not show good results in our experiments. Instead, we treat the task as Semantic Segmentation — for each trace the model predicts a binary mask. Each sample of a trace before the first break is labeled as the no-signal one and all the samples after the first break — as signal ones.
To perform semantic segmentation we exploit ResUnet with Attention. This is basically the 4 stage UNet, where each stage of encoder and decoder consists of two ResBlocks, followed by SCSE (spatial and channel squeeze and excitation) block.
Since the output of a segmentation model is a mask, some post-processing is required in order to extract the time of the first arrival. Different heuristics can be used here, the most intuitive of them is to declare the first signal or the last no signal sample as the first break. We, however, define the first break as the beginning of the longest sequence of signal samples, as shown in the image below. Such an approach proved itself to be more robust when dealing with complex data.
Comparing Neural Networks with STA/LTA method.
See the overall comparison of the performance of the STA/LTA method, vanilla UNet, and Attention ResUNet. ResUNet performs 10x times better than STA/LTA, and almost 2x times better than the UNet in terms of MAE metric, measured in samples.
In addition to being more accurate, Neural Networks also work significantly faster: it takes an hour to automatically pick the survey with millions of traces, while the experts spend days and weeks manually adjusting parameters of unsupervised algorithms in order to properly process the whole survey.
Estimating picking quality on the whole survey.
The MAE metric can only be calculated when the ground truth is present. However, when the new survey is carried out, there is no ground truth. Therefore it's important to be able to estimate the model performance on the entire survey having only predictions. Thus, we need unsupervised metrics.
Since for each trace we have the distance between source and receiver and predicted time, we can calculate the velocity of the signal in the upper layer. There is the basic assumption during seismic processing that these velocities should not significantly vary among neighboring traces. We find the regions with a significant ratio of traces with extreme (very high or very low) velocities. Most of them correspond to areas with bad pickings, but some can represent areas with complex survey conditions. This allows a specialist to inspect the problem regions and find out how well the overall picking was performed, no matter whether it’s Neural Network or another algorithm.
Conclusion
This example of First Break Picking shows that neural networks have great potential in seismic processing. The model completes the task faster and more accurately compared to the traditional method used in the industry. It also automates the process, leaving experts to deal with more interesting and complex tasks.
If you are interested in more technical details and other applications of neural networks in Oil&Gas, visit SeismicPro and seismiqb — our frameworks for machine learning on seismic data.
More articles about Machine Learning in Oil&Gas are coming. Stay tuned!
