Improving Highwall Monitoring Through Fracture Identification in Open-Pit Mines Using Image Segmentation
Loading...
Authors
Letshwiti, Tebogo Motsumi
Issue Date
2023
Type
Thesis
Language
Keywords
Alternative Title
Abstract
Open pit highwall monitoring is an important part of maintaining safe mine operations. The current monitoring practices are ideal for tracking mass slope movements through round-the-clock monitoring of ground acceleration, but they are not well suited for quantifying the extent of damage to the highwall by mining practices like blasting or in-situ conditions like faults and joints, which can lead to rockfall events that can harm people, damage equipment, and halt operations. The current monitoring practice to account for this gap is through in-person inspections by geotechnical engineers, which leaves the potential for large areas of the open pit highwall to go without coverage since there is only so much a human can do. There is current research focusing more on locating areas along the highwall where rockfall might have already happened, for example, using multispectral imaging, but the fracture prevalence has very few researchers looking into it. For this study, researchers utilized a U-Net model for image segmentation to identify cracks and fractures along the open pit mine highwall, aiming to enhance the current monitoring technique of visual inspections employed by geotechnical engineers. Unmanned aerial vehicles were used for data collection as they could access more of the highwall and capture high-quality imagery. Image annotation to label the cracks and fractures in the images was performed, developing the dataset needed to train a deep learning model such as U-Net. Several training schemes were followed to account for low amounts of data and to see which configuration would produce a good model for the problem at hand. Traditional edge detection using the canny edge detector was also used to illustrate the differences in prediction and workflow between deep learning methods and more traditional detection methods, such as edge detection. The model trained with a mix of original and augmented images gave the best performance at 97% accuracy and a relatively high intersection over union (IoU), as well as producing segmentations close to the GroundTruth segmentation mask.