Custom Segmentation

Overview

This document details Datarock’s Custom Segmentation products.

Dependent Models

The outputs of the following models are used to determine any Custom Segmentation:

Model Name	Model Type
Image Preparation	Object Detection
Depth Registration	Semantic or Instance Segmentation
[Custom Segmentation] e.g. Vein Segmentation	Semantic or Instance Segmentation

Data Processing

The outputs of the Datarock Custom Segmentation model when applied over depth registered rows allow the area, percentage, count and polygon axis ratio of these segmented features to be calculated.

Below is an image illustrating the difference between a classification model and its potential outputs against an object detection, semantic segmentation (single-class) and instance segmentation (multi-class) model.

Detection of Custom Segmentation classes

Several segmentation classes can be detected by the Custom Segmentation model based on the training data collected during onboarding. An example of manually labelled examples of segmentation classes can be seen in the following image.

The following images show an example of segmented outputs for the vein classes as trained using the above manually labelled image. In the first row, the raw row image has been identified and cropped, the second row contains the predicted mask polygons for each of the 5 segmentation classes, and the third row includes these masks overlain across the raw cropped row with area statistics in cm².

Product Configuration Options

There are no configuration aspects to this product.

Output Intervals

Default interval length: one row of core (~0.8m)

Customisable interval available: Yes, via uploading an assay or geology logging table to Datarock Customer Success Team

User Data

User data can not be uploaded to the Platform via CSV at the current time.

The following data is required for customisable intervals (assay or geology logs) to be sent to Datarock Customer Success Team:

Hole_ID
Depth_From
Depth_To

Data Output

Results from this class of models are delivered in a batch nature and can be obtained from the Datarock Customer Success Team. The available CSV files include the following:

ProjectID_HoleID_segmentation_raw.csv
ProjectID_HoleID_segmentation_user_intervals.csv

These two CSV files contain the following headers:

File Header	Description	Raw CSV	User Intervals CSV
filename	Row image filename as defined by box and row	Yes	No
hole_id	Customer’s Hole ID	Yes	Yes
box_id	Platform assigned box number	Yes	No
row_id	Core row number	Yes	No
tray_id	Actual core tray number	Yes	No
depth_from	Start of interval	Yes	Yes
depth_to	End of interval	Yes	Yes
coherent_area_cm2	Total area of row containing coherent rock pixels	Yes	Yes
coherent_area_%	Percentage of row containing coherent rock pixels	Yes	Yes
incoherent_area_cm2	Total area of row containing incoherent rock pixels	Yes	Yes
incoherent_area_%	Percentage of row containing incoherent rock pixels	Yes	Yes
total_rock_area_cm2	Total area of row containing coherent and incoherent rock pixels	Yes	Yes
total_rock_area_%	Percentage of row containing coherent and incoherent rock pixels	Yes	Yes
[custom segmentation]_area_cm2	Total area of row containing [custom segmentation class] pixels	Yes	Yes
[custom segmentation]_area_%	Percentage of row containing [custom segmentation class] pixels	Yes	Yes
[custom segmentation]_count	Number count of [custom segmentation class] polygons identified within interval	Yes	Yes
[custom segmentation]_avg_axis_ratio	Average ratio of [custom segmentation class] polygon height to width across an interval	Yes	Yes

Product Limitations

Limitations	Comments
Reliance of row detection and depth registration	The Custom Segmentation model is based on predicting and masking geological features within a row of drill core. The dependency on the depth registered rows being identified means if a row is missed by the row model during Image Preparation, this row will not have the segmentation modelled applied.
Training is dependant on what can be seen within a row image	Some fine segmentation classes are very difficult to classify from a photo, in particular if resolution is poor. Therefore, Datarock’s Custom Segmentation model relies on these classes being trained using visually identifiable features. If segmentation classes are not identified, the resulting segmentation model will generally be lower than expectations based on any available site logging.
Training data must be representative of whole area segmentation is to be applied	If new imagery or segmentation class is introduced to the model, the performance may decline as these examples were not trained during onboarding. An initial model evaluation will need to be undertaken to see the suitability of the model in particular against any new imagery. Ideally, a new model version is trained to incorporate the new untrained drill core or segmentation class.

Document Version

Version	Date	Author	Rationale
1	29 June 2023	N Pittaway	Initial release