Dense Supervised Dual-Aware Contrastive Learning for Airborne Laser Scanning Weakly Supervised Semantic Segmentation

Airborne laser scanning (ALS) point clouds, characterized by low point density due to extensive ground coverage, pose significant challenges in labeling work. In addition, the ALS scanning pattern can lead to occlusions, resulting in data gaps and further increasing the difficulty of accurately labeling points within obstructed areas. In this context, fully supervised deep learning methods, which rely on a large volume of precisely annotated point-level data, are not well-suited due to the time-consuming and costly nature of obtaining extensive labeled datasets. Addressing these challenges, we propose an innovative weakly supervised learning strategy: a dense supervised dual-aware contrastive learning (DSDCL) approach. We introduce a context-aware contrastive constraint to enhance the model’s ability to differentiate similar data points by leveraging subtle contextual differences, improving feature learning robustness in environments with incomplete objects. In addition, we apply a scene-aware dense contrastive constraint to enable multiscale point cloud representations while enforcing class-aware feature space constraints through gradient-based learning across multiple layers. We employ this contrastive constraint to enhance category discrepancies in prototype features predicted by pseudo-labels from dual-branch data, improving model robustness to categorical distribution differences by focusing on class prototype features and mitigating the impact of erroneous pseudo-labels. Extensive experiments on three benchmark ALS datasets demonstrate that DSDCL achieves performance comparable to advanced fully supervised methods, even when using only 0.1% of labeled samples.