U4SNet: Dynamic Class Balanced Hierarchical 3D Underground Semi-Supervised Semantic Segmentation

3D scene semantic segmentation is a key technology for intelligent perception in complex underground environments. Currently, issues such as multi-scale structural heterogeneity of targets and class imbalance due to long-tailed distributions persist in underground environments. In this study, we propose a semi-supervised semantic segmentation framework for underground environments, named U4SNet. To address the problem of multi-scale target representation, this framework employs a multi-level scene-adaptive perception mechanism. By hierarchically fusing low-dimensional local detail features with high-dimensional global topological features and combining them with a dynamic weight allocation strategy, it optimizes the capability of multi-scale target representation. To tackle the class imbalance issue, a class-balanced adaptive threshold algorithm is proposed. A differentiated threshold strategy is adopted—applying conservative thresholds to suppress noise in dominant classes and using lenient thresholds to retain effective information for minority classes. The thresholds are dynamically optimized through the combination of sample priors and model feedback. Validation based on a self-built coal mine dataset (MineSeg3D) and public datasets shows that the U4SNet model significantly outperforms existing methods in terms of mean Intersection over Union and Overall Accuracy metrics. It demonstrates particularly strong advantages in segmentation tasks for complex geological structures and small-scale targets, providing an effective technical solution for intelligent perception in underground spaces.