Evaluation of image-based segmentation algorithms for discontinuity detection

Abstract

Accurate detection of discontinuities is critical to determine rock mass features such as block geometry, joint orientation, and potential failure surfaces, which govern structural stability in mining and geotechnical applications. Manual methods of detecting discontinuities are often time-consuming, expose personnel to hazardous situations, and are susceptible to biased judgement. In response, image processing techniques like image segmentation have been increasingly adapted to detect discontinuities from rock outcrop images. This study evaluates the performance of traditional and machine learning segmentation methods to identify a higher accuracy workflow for discontinuity detection and the following methodology was employed: RGB rock outcrop images were manually annotated to establish ground truth masks, pre-processed with noise-reduction filters and then processed using traditional Gradient-based operators, Canny edge detection, thresholding and machine learning approaches, U-net, Holistically-Nested Edge Detection (HED), and Segment Anything Model (SAM). The performance of these methods was quantified by evaluating the Boundary F1 score against the ground truth masks. Among discontinuity-based traditional segmentation methods, on a scale of 0 to 1, Canny edge detection with morphological gradient achieved F1 scores of 0.194 and 0.196, while among similarity-based segmentation methods, dilation of eroded threshold images achieved F1 scores of 0.264 and 0.202. For machine learning methods, SAM outperformed the other methods by achieving F1 scores of 0.752 and 0.632, but caused over-segmentation in highly discontinuous regions. The findings highlight the significance of combining computationally efficient traditional methods with targeted preprocessing for low-resource settings and underscore the trade-off between machine learning accuracy and its infrastructural demands.