Du, Jing2026-01-232026-01-232026-01-232026-01-12https://hdl.handle.net/10012/22901Modern urban environments undergo continuous transformation as emerging infrastructure appears worldwide. Traditional semantic segmentation methods for 3D point clouds operate on fixed taxonomies, producing static representations that cannot adapt to novel categories. This dissertation addresses novel class discovery (NCD) in large-scale 3D point cloud segmentation through geometry-aware mechanisms, adaptive multi-source fusion, and hybrid supervision frameworks. The first study establishes geometric foundations through voxel-geometry integration with region-centric organization, termed CHNCD. The framework couples voxel representations with original spatial coordinates via index mapping, identifies semantically informative points within clusters, accelerates neighbor retrieval through proximity hash mapping, and consolidates localized features with global context via spatial attention. Experiments on S3DIS, Toronto-3D, SemanticSTF, and SemanticPOSS demonstrate consistent improvements over discovery baselines. The second study deepens representation through adaptive geometric sequence modeling, dynamic Gaussian embeddings, and gated multi-source fusion, termed AGDNet. Adaptive geometric sequence modeling employs learnable dimension weighting and dynamic grouping adjusted to local point density. Dynamic Gaussian embeddings represent point clouds as 3D Gaussians and compute Mahalanobis distances to generate multi-scale spatial embeddings. Gated multi-source fusion intelligently weights features through context-aware mechanisms. Three knowledge-transfer objectives operate at category, instance, and distribution levels to bridge semantic gaps. Evaluation on Toronto-3D, SemanticSTF, and SemanticPOSS demonstrates substantial improvements. The third study integrates discovery with operational land cover mapping, termed 3DLCDM. The framework processes features through a supervised head for established categories and a dual unsupervised head comprising a primary branch with fixed prototypes and an over-segmentation branch with progressive scheduling. Temporal Sinkhorn-Knopp normalization with adaptive temperature scheduling stabilizes pseudo-labels, while dynamic weighting combines per-batch and global frequency statistics to address class imbalance. Evaluation on DALES and H3D datasets demonstrates substantial improvements for continuous land cover discovery mapping. Taken together, the three studies advance a progressive research agenda unifying discovery and 3D segmentation for large-scale point cloud scenes. The dissertation demonstrates consistent gains across six benchmark datasets, exhibits generalization across sensors and acquisition geometries, and provides a principled route to maintain updated urban maps as new structures emerge.en3d point cloud semantic segmentationnovel class discoverygeometry aware representationmulti source feature fusionhybrid supervisionland cover discovery mappingDoctoral Thesis