Acoustic Televiewers (ATV) and Optical Televiewers (OTV) are both specialised tools used in geology and engineering to gather detailed information about the internal features of boreholes or wells. These tools play a crucial role in understanding the subsurface properties of rock formations, which can be valuable for various applications such as mineral exploration, oil and gas reservoir characterisation, geotechnical and groundwater studies, and geological mapping.
An ATV, also known as a borehole televiewer or acoustic borehole imager, is a tool that uses sound waves to create high-resolution images of the borehole wall. It operates on the principle of acoustic reflection, similar to how ultrasound imaging works in medical applications. The tool emits sound pulses that travel through the borehole fluid and into the surrounding rock formations. When these sound waves encounter changes in rock properties, such as fractures, bedding planes, or other geological features, they bounce back and are detected by the tool. As the sound pulses from the ATV must be passed through fluid, the ATV can only be run below the water table (or topped up) and not in dry holes.
An OTV, also known as an optical borehole imager or borehole camera, is a tool that employs optical technology to capture detailed visual images of the borehole walls and can be run either within water or in dry holes. It uses a combination of lighting and a camera to provide high-resolution photographs or videos of the borehole surfaces. These images can reveal intricate details of mineral properties, sedimentary deposition, and structural characteristics. Optical televiewers are particularly useful when detailed visual information is required, and they can complement the data obtained from other downhole tools.
For this blog post, we focus on how computer vision techniques can be utilised to extract structural information from OTV data, providing a framework for automating 3D structural modelling across a deposit scale.
Figure 1. Example Optical Televiewer data of Drillhole M14 from Bull Shoals Dam.
DATA
Televiewer data is reported as a continuous 2 dimensional array (or image) downhole that provides a full 360 degree representation of the inside of a borehole. In addition, the televiewer probe also measures the borehole diameter, dip and azimuth so that the data can then be georeferenced. Using this information a geoscientist can manually pick out geological features such as structures, bedding planes, lithology types etc and use this to build subsurface models. More information on the televiewer and how it collects data can be found here.
The dataset used in this blog is open source and supplied by the USGS (United States Geological Society) originating from Bull Shoals Dam, Arkansas (Figure 2). Both ATV and OTV were collected to improve the understanding of fractures and small bedding planes in the bedrock beneath Bull Shoals Dam with the intention to forecast models of dam sustainability. Nine holes were drilled in the dam but for this example we will look at only the OTV from hole M14, which was drilled to a depth of approximately 12m.
Figure 2. Map view of the Bull Shoals Dam drillholes.
As televiewer data can be viewed as an image, it is possible to perform computer vision and image analysis techniques to extract valuable information in a consistent and systematic manner. More specifically, we can split the continuous downhole image into tiles and perform analysis on these, gaining a more detailed understanding of the intersected geology (Figure 3). There is an abundance of geological information that can be extracted from images, however in-situ structural measurements are one of the most powerful for mineral resource definition and mine planning, as they are not always obvious or present when logging core data.
Figure 3. A visual representation of converting televiewer data into square image tiles.
EXTRACTING PLANAR DATA
To examine how we can use computer vision techniques to extract information from televiewer data, we will use a planar geological feature located in drillhole M14 at approximately 7.10m depth. Planar features in geology can include bedding, joints or fractures. It is likely that this specific feature in drillhole M14 is not structural but rather of sedimentary origin, however the following example is applicable to all planar features. Due to the difficulty of sourcing openly available televiewer data, ML driven identification and classification of different planar features is out of scope for this blog although it is possible.
There are a couple of things we know about planar features that can help guide us when building out a computer vision workflow. Firstly, planar features span the entire X extent of our image. As a televiewer image is a 360 degree representation of the drill hole interior, any planar feature that the drillhole intersects is captured by the entire image. Secondly, these features are often sinusoidal in appearance due to the warped representation and angle of intersection. This means that simple line detection algorithms may not be optimal for this problem and therefore more sophisticated approaches are required.
Due to process explainability, a simple thresholding and pixel contouring methodology is used to identify and extract the planar feature in this example. All image processing steps are done using Python’s Open-CV library.
- As mentioned above, the first step is to tile the televiewer data and generate the square images.
- Secondly, the image tile is converted to greyscale (from 3 RGB bands to a single band) to allow for more universal single band image processing.
- A contrasting filter is then applied to the image to highlight locations of noticeable differences.
- A threshold is applied to the image to extract a binary mask of pixels. The thresholding approach used here is the Otsu method. This is an automated approach of thresholding that is based upon the histogram of greyscale values. More information on the Otsu method can be found here.
- A morphological transformation to ‘close’ any shapes and remove noise both internally and externally of shapes is then applied.
- Contours around the masked pixels are then calculated to determine the shape of objects
- Finally as mentioned above, we know the features we are looking for span the entire X extent of the image so we can set a filter to only select features that span this entire extent.
Figure 4. An outline of the example computer vision workflow for extracting sinusoidal objects in televiewer data.
From here we can calculate a line that best represents the shape of our structure by getting the average trend of the shape. Using the vertices of this line we can then transform this 2 dimensional representation into spherical coordinates (points). To do this we must first extract the caliper values at the features depth to determine borehole width, then use trigonometry to determine the extent of the points in geographical space. The below interactive plot displays the point locations converted into spherical coordinates. Notice that now that the points are in geographical space, they link back up to form the locations of the drillhole interior where the planar feature intersects.
Figure 5. The geographical representation of the planar object detected in televiewer data.
Once the outline of the feature has been identified and converted to points, it then becomes possible to fit a 2D plane to the points. This plane represents the geological feature geographically, and given downhole survey data, can provide a basis to calculate more information such as dip and dip direction. For example (assuming that the drilled hole is vertical), the dip and strike of this plane is calculated as 17/174. The following interactive plot displays this plane with respect to the OTV data in 3D, with information provided in the hover.
Figure 6. An interactive plot of the interpolated planar feature with respect to the televiewer data, displaying the dip and dip direction of the feature.
CONCLUSION
Compared to manually examining, identifying and selecting geological features in televiewer data, this methodology can provide a streamlined approach to extracting planar data in both OTV and ATV data. It provides the ability to consistently extract structural and bedding information, and in conjunction with additional datasets such as geochemistry or other downhole geophysics suites, it can provide a basis for geological modelling.
As mentioned, extracting planar data is only one possible application of computer vision in televiewer data. Lithology and alteration classification, clast/mineral grain analysis and sedimentary structure analysis are all examples of geological applications. In addition, non geological information such as potential casing failures/areas of weakness and drilling related issues can also be detected and flagged that may not be as obvious to the human interpreter.
It is also worth reiterating that the aforementioned object detection method is quite simple with respect to more modern mathematical practices. Advanced deep learning methodologies utilising convolutional neural networks can be used to detect and segment objects with higher accuracy and speed than ever before. Figure 7 displays the results of applying Meta’s open source Segment Anything Model (SAM) onto a different section of the televiewer data. The image displays the input image (A) and the resulting segmented objects (B). Using these segmentation shapes and knowing the borehole width from the caliper, a grain size analysis can be conducted providing quantitative information on the geology as seen in Table 1. This provides another example of how advanced computer vision techniques can be used with geological data.
Figure 7. An example of using the prebuilt Segment Anything Model to detect objects in televiewer data.
Table 1. The average grain size properties (in mm) from the grains detected in Figure 7.
The real power behind utilising computer vision with televiewer data is the ability to scale it to multiple drillholes and extract consistent data at deposit scale. This provides a tremendous benefit as a geoscientist, removing some of the non-technical clicking, selecting and typing tasks often associated with feature identification and data processing. This in turn allows more time for the important earth-science driven interpretation, understanding and geological modelling of the subsurface which is crucial for improving our knowledge of earth processes.
If you are interested in hearing more about how you can extract the most information out of any geological imagery data, please reach out to us at [email protected]
