Authors: Ziwei Luo, Jun Jiang, Yuan Wu, Xiaoyun Yan, Yangjun Ou, Zhaojing Wang
Status: Final
Date of publication: 30 September 2026
Published in: ITU Journal on Future and Evolving Technologies, Volume 7 (2026), Issue 3, Pages 357-371
Article DOI : https://doi.org/10.52953/VARS8860
|
Abstract:
Airborne Laser Scanning (ALS) point clouds provide accurate three-dimensional geospatial information and are closely related to various remote sensing applications, such as real-scene 3D modeling, infrastructure informatization, and urban mapping. Deep learning has greatly improved ALS semantic segmentation, but fully supervised methods require dense point-wise manual labels, which are expensive to obtain for large-area airborne LiDAR surveys. Under this requirement, weakly supervised ALS segmentation has attracted increased attention because it can use limited manual labels to guide dense semantic interpretation of large area airborne LiDAR point clouds. This paper reviews weakly supervised semantic segmentation methods for ALS point clouds. We first summarize the development of deep learning-based ALS interpretation, commonly used datasets, and evaluation metrics. Existing methods are then reviewed from three main dimensions. First, weak annotation acquisition and sample selection are discussed, including sparse point labels, coarse subcloud-level annotations, click or scribble-based labels, and active learning strategies for selecting informative ALS samples. Second, reliability modeling of weak supervision is analyzed by focusing on pseudo-label confidence, entropy, uncertainty, consistency constraints, and feature-space separability, which are used to identify reliable supervision signals from incomplete or noisy labels. Third, supervision optimization is summarized from the perspectives of self-training, label propagation, pseudo-label refinement, curriculum learning, and robust loss design, aiming to improve the stability and effectiveness of weakly supervised ALS segmentation. Finally, we discuss future directions, including standardized weak-label benchmarks, cross-scene and cross-sensor generalization, promptable foundation models and multimodal remote sensing fusion. |