Determination of Block Sizes of Jointed Rock Masses Using Close-Range Photogrammetry

Abstract

The identification of rock mass discontinuities is of critical importance not only for the construction of infrastructure such as tunnels, highways, slope stabilization systems, etc., but also for the early prediction of natural hazards such as rockfalls. Traditional methods, mainly those including the manual measurement techniques, lead to significant challanges in detecting discontinuities and may involve vital risks. The developments in photogrammetry and remote sensing have facilitated studies on the detection of discontinuities with laser scanners and optical images acquired from remotely piloted aircraft systems (RPAS). However, discontinuity detection using point clouds sourced from either of these systems is time-consuming and imposes a significant hardware cost. In this thesis, a novel end-to-end framework has been developed to detect discontinuities from terrestrial and RPAS-based images in three main steps, without the requirement of point cloud generation. In the first step, a deep learning (DL) approach was employed to detect discontinuities, which yield to an accuracy of 91.7% expressed in F1-score. A multi-image data augmentation method was also proposed here to improve the performance of the DL model training with a small amout of data, which is often the case in rockfall sites. In the second step, based on the detected discontinuities, scanline measurements, which are widely used in geological studies, were performed with an image-based approach. In the third step, as an additional contribution to the literature, rock blocks were defined by calculating the volumetric joint count (Jv) from the detected discontinuities. Thus, this thesis follows a fully image-based approach, reducing both hardware and time costs. It also demonstrates the usability of image-based methods in detecting discontinuities, with an orientation deviation of approximately 6 degrees based on discontinuity orientation, and a discontinuity spacing error of 8 mm relative to an average discontinuity spacing of 33 cm. However, as a major limitation, in cases such as excessive illumination and shadows in the images under unfavourable weather conditions, the prediction performance of the DL model may decrease. The framework has been implemented as in-house software in most parts, especially the process of detecting discontinuities, calculating their orientations, and defining blocks based on scanline drawings.