Open Cut » Rock Mechanics
Accurate geotechnical characterisation of mine spoils is critical to enhance dump stability which eventually leads to safe mining operations and optimal environmental rehabilitation. Conventionally, classification of spoils in relevant CoalSpoil (previously known as BMA) categories relies on the expertise of geologists/geotechnical engineers. The manual practice, based on visual/textural inspections, often requires vigorous field efforts and time. Besides exposing personnel to field hazards, the assessment results are at times biased depending on the expertise of an individual. This project explored the utilisation of image-based techniques to characterise spoil materials and apply machine learning to automate the process aiming to reduce the involved subjectivity.
Multifaceted analyses were conducted using both, the drone based approach and the close range image based approach. The data acquisition process involved eight aerial surveys across six distinct mine sites covering both the Sydney Basin and the Bowen Basin. Concurrently completed project C29044 Baseline data for the development of automated characterisation of waste materials, systematically compiled a total of 686 ground validation points. These ground samples of spoils were meticulously labelled into relevant spoil categories by an experienced geologist/geotechnical engineer. Drone images of relevant areas were acquired in high resolution to train machine learning models. These models can generate several features from optical and textural properties of the images which could be uniquely attributed to the consecutive labels. Statistical analyses indicated a strong correlation between lithology obtained from imagery and geotechnical ground parameters, providing valuable insights.
The study showcased two main automated characterisation approaches:
- Pixel based, wherein per pixel in an image was classified into a unique spoil category.
- Object based, wherein spoil was assessed as a whole and classified based on more dominant lithology.
Although most of the evaluation was performed on drone based images, the approach and model can be customised for close range photographs captured either from a smartphone or tablet.
The repetitive drone image acquisitions were also evaluated for its capability to generate 3D profile of the dump as it builds. The built 3D profile was consequently used to analyse the dump stability by investigating 3D slip surfaces. If sufficient time series data exist, the development of dump in 3D can be traced and variation in lithologies could be observed. This project also illustrates practical use cases of 3D data to enhance our understanding of dump stability. This approach of 3D profile building and stability analysis presents a promising avenue for future research and application in the field.
This project presents a significant advancement in spoil pile classification methodologies through the application of long/close range image based analysis. The integration of spatial and temporal dimensions, the use of advanced machine learning algorithms, and the incorporation of innovative processing techniques collectively transform our comprehension of spoil pile characteristics. This project not only promotes mine automation but also lays the groundwork for future research which would enhance operational efficiency in years to come.