Analytics Ready Data (Image Preparation + Depth Registration)

This document details Datarock’s workflow to generate Analytics Ready Data that includes the Image Preparation and Depth Registration Pipelines and produces out analytics ready depth registered strip imagery.

Dependent Models

The outputs of the preprocessing rely on the outputs of the following machine learning models.

Dependencies

Data and metadata must be uploaded to the platform following the file name conventions described in these how-to articles in the Datarock knowledge base.

Uploading CSV Data

Uploading Images

Image Preparation

The first step of the Preprocessing workflow uses the Image Preparation pipeline to convert the raw core box image into a cropped box and strip image.

Core box cropping

The raw core image may contain several core boxes or background material that is not required for analysis. The box cropping model aims to predict a polygon around each core box instance and can crop up to 4 boxes per image. After cropping, a black null padding region is added around each image to ensure the active image is not right at the margins of the image.

Example: a typical core box image is cropped into two images with individual boxes and buffer zones.

Core row cropping

The box-cropped imagery is then further cropped via a segmentation model that is trained to predict a rectangle around each row of the core within the box. From these, individual images of the box channel can be cropped and created (red polygons).

Example: a cropped box imagery being cropped via the red cropping marks into individual row imagery

Configuration	Options	Description
Dewarp Imagery	Dewarp can be turned on or off	Dewarp turned off will not modify original shape of the core box. Dewarp turned on will warp the original box shape into a vertically sided rectangle.

Depth Registration

The aim of our depth registration process is to assign each pixel within rock material with an accurate depth position in the drill hole.

At a high level, the depth registration process aims to interpolate depth between known tie points such as core blocks, depth markers and box start and end depths, taking into account several factors including:

1. Adjustment and compaction of broken material

2. Core loss information

3. Optical character recognition to automatically read core blocks and depth marks

4. Measurement of detected rock within the tray (row models)

In this section, we will review the various inputs that feed into the depth registration process and explain at a high level how the final depth results are produced.

Compaction Factor

When rock material is broken, it can disperse in the box and occupy more length than it would have in situ in the ground. The depth registration process allows for zones of broken material to be compressed to better represent its original length in the ground.

Example: A compaction factor of 0.76 indicates that in order to optimise the depth interpolation, the incoherent material needed to have its measured length in the tray multiplied by 0.76. This will best match its actual downhole depth or length in the drill hole.

Example: compaction factor with a value of 0.76 between the two arrows is shown in the depth registration tab

Core Loss

By providing core loss data through a csv file (refer to Uploading CSV Data ), the depth registration process can eliminate these missing depth intervals from specific locations.

Example of 0.1m of located core loss being located next to a core block

Optical Character Recognition (OCR)

Using the machine learning technique called optical character recognition (OCR), depth marks and core blocks can be automatically extracted from the core. OCR reads all the text and numbers from the core image and then eliminates irrelevant information using specific rules. The OCR-identified depth markers appear in the depth registration tab and are used as depth tie points.

Example: two detected depths that were read from the core image using OCR.

Row Detection Models

During the depth registration process, the position, length, and classification of different materials in each row of the core tray are identified. This information, along with the measured lengths of the rock materials, is utilized in the depth registration process and the following Depth Health Check workflows.

Coherent Rock

The Coherent Rock model identifies all pieces of cylindrical pieces of drill core, and draws a polygon around each piece (see below). A coherent piece of rock is defined as cylindrical with top and bottom edges that are approximately parallel to the core box row.

Example coherent rock model polygon detections (light blue)

Incoherent Rock

The Incoherent Rock model predicts a polygon around regions that contain broken rock. Broken rock is defined as fragments of rock that are either non-cylindrical and irregular in shape or out of place.

Example of Incoherent rock model polygon prediction

Core block

The core block model predicts a polygon around each core block identified within the core tray.

Example core block model polygon detections (yellow)

Empty Tray

The Empty Tray model predicts a polygon around regions of core rows that do not contain rock material or core blocks. Empty tray is defined as a region where the base of the core tray is clearly visible. The Empty Tray class is also inserted in small (<1cm) spaces between the other row detection models (See bottom).

Example Empty Tray model polygon detections (purple)

Example of the Empty Tray class being inserted into a small gap between two coherent polygons (purple line)

Depth Interpolation

The "depth tie points" are fixed points that cannot be altered, and the final calculation of depth involves interpolating the depth between these points. The rock between each fixed point is detected and measured, and core loss files are used to remove depth where necessary. If there is an excess of rock length in an interval, then compaction of the zones with detected broken rock can occur.

1. Estimate depth between tie points

2. Use OCR to identify core blocks and depth markers that align with the initial depth estimation

3. Discard OCR detections that do not align with initial depths

4. Perform secondary interpolation of depth using new tie points

5. Apply core loss, remove irrelevant depths, perform a third interpolation of depth, and apply compaction to optimize interval length.

Depth Health Check

The length of a given interval of rock within a drill hole is calculated in two different ways:

1. From depths supplied as metadata (client provided) - Image filename depths, .csv depth information

2. From measurement of detected rock material (Datarock calculated) -

Datarock reports these rock length/depth deviations as part of the Depth Health Check found in the notifications page under the Depth Registration Tab (see below).

Some level of depth error will be expected due to the accuracy of the drilling process and subsequent depth marking. For this reason we allow the user to set a Depth error tolerance value of depth different/deviation below which a Depth Health Check task will not be reported. The minimum size of the depth interval that can be reported as part of a Depth Health Check can also be set by using the Interval length tolerance configuration (described below).

Example depth health checks within the Depth Registration Tab

Configuration	Options	Description
Dynamic row sampling	Dynamic row sampling can be turned on or off	Dynamic row sampling turned off uses a static sampling line that runs along the exact centre of the row image Dynamic row sampling turned on allows the sampling line to move up and down within the row image to best capture the predicted coherent rock polygons.	Example of row when dynamic sampling is turned off (line is along exact centre of the image) Example when dynamic row sampling is turned on (line is below the centreline of the box)
Minimum width of coherent rock	Input of a percentage between 0 and 100%	The minimum width of coherent rock allows the user to define how much of the vertical height of a row is required to contain rock to be classified as coherent rock	Example of coherent rock with a low minimum width (~30%) due material obscuring the core. This would be classified as incoherent rock if the threshold was set too high. Example of coherent rock with a high minimum width (~90%)
Depth error tolerance	Input of a number between 0 and n	Depth error tolerance value sets a threshold above which will trigger a Depth Health Check task. It is configured as a value of depth error per metre. i.e. if the error is set to 0.5m/m depth errors with less than 0.5m of length deviation per metre will not be reported.
Interval length tolerance	Input of a number between 0 and n	Interval length tolerance creates a threshold above which the user is alerted that the interval is above this threshold. This feature aims to alert the user to very long intervals with no pierce points that may not have a meaningful Depth Health Check.

Product Limitations

A number of factors can control the simplicity and overall accuracy of the Datarock depth registration process. These factors are related to the accuracy and prevalence of known depth pierce points such as box depths, metre/feet marks and core blocks as well as core loss information.