FirstBreaksPicking

This project is devoted to pick waves that are the first to be detected on a seismogram (first breaks, first arrivals). Traditionally, this procedure is performed manually. When processing field data, the number of picks reaches hundreds of thousands. Existing analytical methods allow you to automate picking only on high-quality data with a high signal / noise ratio.

As a more robust algorithm, it is proposed to use a neural network to pick the first breaks. Since the data on adjacent seismic traces have similarities in the features of the wave field, we pick first breaks on 2D seismic gather, not individual traces.

Examples

In the drop-down menus below you will find examples of our model picking on good quality data as well as noisy ones.

It is difficult to validate the quality of the model on noisy data, but they show the robustness of the model to various types of noise.

How model was trained?

Model training is described in detail here in the previous version of the project README.

The latest model was trained similarly, but: 1) Real land seismic data is used instead of synthetic data. 2) UNet3+ model is used, instead of the vanilla UNet.

Installation

Library is available in PyPI:

pip install -U first-breaks-picking

GPU support

You can use the capabilities of GPU (discrete, not integrated with CPU) to significantly reduce picking time. Before started, check here that your GPU is CUDA compatible.

Install GPU supported version of library:

pip install -U first-breaks-picking-gpu

The following steps are operating system dependent and must be performed manually:

• Install latest NVIDIA drivers.
• Install CUDA toolkit. The version must be between 11.x, starting with 11.6. Version 12 also may work, but versions >=11.6 are recommended.
• Install ZLib and CuDNN: Windows and Linux.

Compiled desktop application

Compiled executable files (.exe in .zip archives), as well as installers (.msi), can be downloaded in the Releases section. If you want to use the capabilities of your dedicated GPU, also follow the steps in the previous step to install the driver, CUDA, CuDNN, and ZLib.

You can run the desktop application on other operating systems (Linux, MacOS) as well, but it must be run using the Python environment.

Extra data

• To pick first breaks you need to download model.

• If you have no seismic data, you can also download small SGY file.

It's also possible to download them with Python using the following snippet:

from first_breaks.utils.utils import (download_demo_sgy,
                                      download_model_onnx)

sgy_filename = 'data.sgy'
model_filename = 'model.onnx'

download_demo_sgy(sgy_filename)
download_model_onnx(model_filename)

How to use it

The library can be used in Python, or you can use the desktop application.

Python

Programmatic way has more flexibility for building your own picking scenario and processing multiple files.

Minimal example

The following snippet implements the picking process of the demo file. As a result, you can get an image from the project preview.

from first_breaks.utils.utils import download_demo_sgy
from first_breaks.sgy.reader import SGY
from first_breaks.picking.task import Task
from first_breaks.picking.picker_onnx import PickerONNX
from first_breaks.desktop.graph import export_image

sgy_filename = 'data.sgy'
download_demo_sgy(fname=sgy_filename)
sgy = SGY(sgy_filename)

task = Task(source=sgy,
            traces_per_gather=12,
            maximum_time=100,
            gain=2)
picker = PickerONNX()
task = picker.process_task(task)
task.picks.color = (255, 0, 0)

# create an image from the project preview
image_filename = 'project_preview.png'
export_image(task, image_filename,
             time_window=(0, 60),
             traces_window=(79.5, 90.5),
             show_processing_region=False,
             headers_total_pixels=80,
             height=500,
             width=700,
             hide_traces_axis=True
             )

For a better understanding of the steps taken, expand and read the next section.

Detailed examples

Desktop application

Application under development

Desktop application allows you to work interactively with only one file and has better performance in visualization. You can use application as SGY viewer, as well as visually evaluate the optimal values of the picking parameters for your data.

Launch app

Enter command to launch the application

first-breaks-picking app

or

first-breaks-picking desktop

Select and view SGY file

Click on 2 button to select SGY. After successful reading you can analyze SGY file.

The following mouse interactions are available:

• Left button drag / Middle button drag: Pan the scene.
• Right button drag: Scales the scene. Dragging left/right scales horizontally; dragging up/down scales vertically.
• Right button click: Open dialog with extra options, such as limit by X/Y axes and export.
• Wheel spin: Zooms the scene in and out.
• Left click: You can manually change values of active picks.

Spectrum analysis:

• Hold down the Shift key, the left mouse button, and select an area: a window with a spectrum will appear. You can create as many windows as you like.
• Move existing windows by dragging with the left mouse button.
• Delete the window with the spectrum by clicking on it and pressing the Del key.

Load model

To use picker in desktop app you have to download model. See the Installation section for instructions on how to download the model.

Click on 1 button and select file with model. After successfully loading the model, access to the pick with NN will open.

Settings and Processing

Click on 3 button to open window with different settings and picking parameters:

• The Processing section contains parameters for drawing seismograms; these parameters are also used during the picking process. Changing them in real time changes the rendering.
• The View section allows you to change the axes' names, content, and orientation.
• The NN Picking section allows you to select additional parameters for picking the first arrivals and a device for calculations. If you have CUDA compatible GPU and installed GPU supported version of library (see Installation section), you can select CUDA/GPU device to use GPU acceleration. It can drastically decrease computation time.

A detailed description of the parameters responsible for picking can be found in the Picking process chapter.

When you have chosen the optimal parameters, click on the Run picking button. After some time, a line will appear connecting the first arrivals. You can interrupt the process by clicking Stop button.

Processing grid

Click on 4 button to toggle the display of the processing grid on or off. Horizontal line shows Maximum time and vertical lines are drawn at intervals equal to Traces per gather. The neural network processes blocks independently, as separate images.

Picks manager

Click on 5 button to open picks manager.

When picking with a neural network, the result of the picking will appear here. You can also: add manual picks, load picks from headers, duplicate picks, aggregate several picks, or remove picks.

As control and navigation tools, you can choose to show picks or not (checkbox), select active picks (radio button), and change color (color label).

After selecting one of the picks, you can double-click or click the button with the “save” icon to show the export options.

Picking process

Neural network process file as series of images. There is why the traces should not be random, since we are using information about adjacent traces.

To obtain the first breaks we do the following steps: 1) Read all traces in the file. 2) Limit time range by Maximum time. 3) Split the sequence of traces into independent gathers of lengths Traces per gather each. 4) Apply trace modification on the gathers level if necessary (Gain, Clip, etc). 5) Calculate first breaks for individual gathers independently. 6) Join the first breaks of individual gathers.

To achieve the best result, you need to modify the picking parameters.

Traces per gather

Traces per gather is the most important parameter for picking. The parameter defines how we split the sequence of traces into individual gathers.

Suppose we need to process a file with 96 traces. Depending on the value of Traces per gather parameter, we will process it as follows:

• Traces per gather = 24. We will process 4 gathers with 24 traces each.
• Traces per gather = 48. We will process 2 gathers with 48 traces each.
• Traces per gather = 40. We will process 2 gathers with 40 traces each and 1 gather with the remaining 16 traces. The last gather will be interpolated from 16 to 40 traces.

Maximum time

You can localize the area for finding first breaks. Specify Maximum time if you have long records but the first breaks located at the start of the traces. Keep it 0 if you don't want to limit traces.

List of traces to inverse

Some receivers may have the wrong polarity, so you can specify which traces should be inversed. Note, that inversion will be applied on the gathers level. For example, if you have 96 traces, Traces per gather = 48 and List of traces to inverse = (2, 30, 48), then traces (2, 30, 48, 50, 78, 96) will be inversed.

Notes:

• Trace indexing starts at 1.
• Option is not available on desktop app.

Recommendations

You can receive predictions for any file with any parameters, but to get a better result, you should comply with the following guidelines:

• Your file should contain one or more gathers. By a gather, we mean that traces within a single gather can be geophysically interpreted. The traces within the same gather should not be random, since we are using information about adjacent traces.
• All gathers in the file must contain the same number of traces.
• The number of traces in gather must be equal to Traces per gather or divisible by it without a remainder. For example, if you have CSP gathers and the number of receivers is 48, then you can set the parameter value to 48, 24, or 12.
• We don't sort your file (CMP, CRP, CSP, etc), so you should send us files with traces sorted by yourself.
• You can process a file with independent seismograms obtained from different polygons, under different conditions, etc., but the requirements listed above must be met.

Acknowledgments

We would like to thank GEODEVICE for providing field data from land and borehole seismic surveys with annotated first breaks for model training.