Tunnel surrounding rock joint recognition and rock mass integrity evaluation based on an improved YOLOv8-seg model

Abstract

Rock fractures significantly influence rock classification and engineering parameter determination. Traditional joint extraction methods are time-consuming, subjective, and error-prone, hindering intelligent tunnel engineering and compromising construction safety. This study proposes an automated rock integrity quantification method using the GAM-BiFPN-YOLOv8-seg model. A standardized process for collecting tunnel face images was established, creating a diverse joint image dataset. Incorporating global attention mechanisms and weighted bidirectional feature pyramid networks into the YOLOv8-seg framework improved detection of subtle joint features and multi-scale fusion, enhancing joint extraction accuracy. After binarization and applying the Zhang-Suen thinning algorithm, joint traces were combined with surrounding joint counts to automate rock integrity quantification. Integrating the relationship between surrounding rock joint counts and rock mass integrity indices, the method achieved a 6.34% improvement in mAP50 and an 8.42% improvement in mAP50:95 compared to the original YOLOv8-seg network, with marginal gains in detection speed. During engineering validation at the NEOM New City Tunnel project, the proposed method demonstrated consistent performance with human decision-making in quantifying tunnel face joints. The method provides an effective solution for the accurate assessment of rock mass integrity and has broad application prospects in the field of tunnel engineering and geological engineering.