Image classification is a process whereby the spectral information of an image, based on its digital numbers, attempts to classify individual pixels to a theme or specific object (e.g. vegetation, water, vehicles, people, etc.). The output is generally an image map or mosaic of pixels, each of which belong to a particular theme or identification to produce an independent overlay of the original image. This overlay can be used to provide a post analysis regarding changes that are occurring in a sequence of images or, for example, identify a potential hazard that can trigger an action for human intervention. The accuracy of image classification is based on having enough information to train a model to identify the theme or object of interest. This paper presents the results of a supervised machine learning technique whereby target objects were identified and models run to train the classification algorithm to identify changes in supernatant pond size, rates of rise, detection of inflows of water to an area and presence of mobile equipment. Training images were acquired from site-based static time-lapse cameras that have been taking images since early 2017 of different areas of a tailings storage facility in the north of Chile.
Geomechanical risk in mining is universally understood to depend on many apparently disparate factors acting together such as stress, stiffness, mine geometry, rock mass character, rock type, structure, excavation rate and volume, blasting, and seismicity. We have worked on many case studies over the years in both underground and open pit mines with the objective of discovering and documenting the correlation of such factors with the experience of geomechanical failure. Whether that failure is slope failure, strainbursting, fault slip-induced rockbursting, roof fall, or any other of many possible failure types, statistical correlations among the different classes of data can be found, and predictive rules for understanding geohazard based on their quantitative combination can be established and deployed in day-to-day operations. This data-driven approach requires application of methods and avoidance of pitfalls that can be standardised into a universally applicable workflow. We discuss the workflow and the pitfalls in analysis to be avoided through case study examples.