Pit slope failures at open pit mines are dangerous, regularly causing multiple fatalities and mine closures across the world
Soil cracks running parallel to the pit edge are a critical indicator or precursor of slope instability in open-pit mining operations. Detecting these types of cracks early can help a mine operator take appropriate action before a dangerous and costly collapse.
Keen AI and engineering experts AMOG have developed an automated pipeline for detecting and monitoring potentially dangerous cracks
- Early detection
- Crack prioritisation
- Enhance regulatory compliance
- Intuitive reporting and actionable insights
- Reduce manual reviews
- Change monitoring
Over the past year, researchers at Keen AI and AMOG have been exploring reliable methods for processing drone footage and identifying cracks of concern.
Today, we’re announcing SlopeWise™ our automated service for doing this. It allows site operators to quickly identify areas of concern, prioritise them and take appropriate action.
Significantly, it also monitors the evolution of cracks over time enabling better operational decisions.
Figure 1. Screenshot from SlopeWise™ Mine Safety Monitoring Solution
Reliably detecting cracks of concern whilst minimising false positives and false alarms was a challenge
The problem poses a number of challenges in order to be able to reliably detect a crack of concern whilst at the same time minimising false positives and not swamping a mines operations or safety team with false alarms. The main challenges are:
- Key attributes such as soil colour vary across sites
- A pit has lots of cracks or things that look like cracks which are of no concern
- Shadows can hide cracks or confuse ML pipelines.
- Objects like pipes or hoses can mimic cracks, leading to potential misclassification.
- Sharp edges formed by transitions from land to water or abrupt soil colour changes can also lead to misclassification.
We have developed a four-step pipeline to reliably detect cracks of concern (figure 2.) The system searches for cracks running parallel to the pit edge indicating a potential slip precursor. A deep learning model detects soil cracks from high-resolution aerial imagery, followed by a computer vision pipeline that analyses their orientation relative to the pit edge to identify high-risk cracks associated with slope failure.
Figure 2. Computer vision detection of mine cracks as early predictors of slope failure (by Keen AI)
1.
Crack Detection Model
A deep learning based model was trained to detect cracks across various surfaces and fine-tuned for soils in open-pit mines. The model generalises well across sites. It identifies crack locations and segments their extent, allowing for monitoring crack extent changes over time.
2.
Crack Orientation
Every individual crack is isolated to measure its orientation. Computer vision techniques are applied to detect crack edges, and statistical methods determine the dominant orientation based on the detected features.
3.
Pit-Edge Orientation
Model for generating contours of the pit edge. A segment of the contour line is extracted, and its orientation is calculated using the trigonometric relationship between its start and end points.
Figure 4. Local pit edge orientation estimation (by Keen AI)
4.
Crack Scoring Index (CSI)
This index quantifies the risk associated with detected cracks in open-pit mines by assessing their alignment relative to the pit edge and providing a rating between 0 to 1. This scoring system provides an objective measure to prioritise cracks for further monitoring and mitigation.
Figure 5. Crack detection, segmentation, and crack risk assessment with the implemented pipeline
We can assess the pipeline’s ability to accurately identify and segment cracks based on their risk of slope instability
While hundreds of earth cracks can be seen only a handful were flagged as high priority cracks.
By visually inspecting multiple examples of detected cracks, we can assess the pipeline’s ability to accurately identify and segment cracks based on their risk of slope instability. The consistency of these results across various conditions indicates that the pipeline performs effectively for its intended application. Figure 6 presents an example of a high-risk crack (risk score: 0.99), where the crack angle aligns with the local pit edge orientation. In contrast, Figures 7 and 8 show low-risk cracks with scores of 0.13 and 0.3, respectively, where the cracks are approximately perpendicular to the local pit edge.
Crack location and local pit orientation
RGB Image
Detected Crack and Orientation
Detected Crack and Risk Level
Figure 6. High risk crack – Crack orientation aligned with the local pit edge. Crack location and estimated edge angle, RGB image, crack detected and estimated angle, Estimated crack risk.
Crack location and local pit orientation
RGB Image
Detected Crack and Orientation
Detected Crack and Risk Level
Figure 7. Low risk crack – Crack orientation perpendicular to local pit edge. Crack location and estimated edge angle, RGB image, crack detected and estimated angle, Estimated crack risk.
Crack location and local pit orientation
RGB Image
Detected Crack and Orientation
Detected Crack and Risk Level
Figure 8. Low risk crack – Crack orientation perpendicular to local pit edge. Crack location and estimated edge angle, RGB image, crack detected and estimated angle, Estimated crack risk.
Early detection of potential mining slope failures has numerous benefits
Faster Detection
Regular drone flights capture detailed imagery as model inputs. These models are trained to identify and classify signs of slope instability. This allows for a larger area to be inspected quickly immediately after a potential triggering event e.g. seismic activity, heavy rainfall or blasting.
Enhanced Regulatory Compliance and Actionable Insights
Our solution helps mine sites meet regulatory requirements for detecting, monitoring, and managing slope instability risks. An intuitive dashboard ranks risk areas, enabling users to prioritise actions for improved planning, safety, and risk mitigation.
Change Monitoring
Once a site’s pits are assessed and areas of interest detected, subsequent flight data is used to detect any change. This is especially useful if a crack which was previously considered low risk evolves into something bigger.
About AMOG
About Keen AI
Credits
We acknowledge Element Geospatial for providing the aerial data used in this work. Their data contribution was essential for the development and validation of our approach.
Footnotes
[1] Goldstein et al, “Framework to predict open pit mine failure runout”, Australian Centre for Geomechanics, Perth, Western Australia, 2023 (DOI https://doi.org/10.36487/ACG_repo/2335_41)
