A novel semi-supervised method for airborne LiDAR point cloud classification

Existing supervised-learning-based methods have achieved remarkable performance for Airborne LiDAR point cloud classification with the help of large-scale human-annotated datasets. However, human annotations are usually labor-intensive and time-consuming, especially for dense segmentation labels, and an inadequate dataset may lead to poor generalization ability of the learned models. In this paper, to relieve the need for large-scale dense annotations, we introduce a semi-supervised method for airborne LiDAR point cloud classification which requires only a small fraction of points to be labeled during model training. To achieve this goal, we introduce a masked supervision module that provides supervision signals from only a few labeled points. Then, three unsupervised supervision modules are introduced to encourage global context consistency, transformation consistency, and spatial smoothness of the learned features. Experiments are conducted on the two benchmark datasets and the results demonstrate the effectiveness of the proposed method for semi-supervised airborne LiDAR point cloud classification. Specifically, the proposed methods can obtain a classification performance comparable to its fully supervised counterpart with only 10% and 1% labeled points for ISPRS 3D Labeling Vaihingen dataset and 2019 IEEE GRSS Data Fusion Contest 3D Point Cloud Classification dataset respectively.