Objectives Airborne point cloud serves as a crucial data source for the extraction of building roof edge points. However, current extraction methods generally suffer from low automation levels and insufficient accuracy.

Methods To address these issues, we innovatively propose a machine-learning-based extraction strategy for building roof edge points from airborne point clouds. The proposed method efficiently integrates the advantages of various point-cloud-based edge-point detection features, enabling the adaptive extraction of building roof edge points in complex scenarios. First, the proposed method performs dimensionality reduction on the point cloud of the building roof, generating multi-level features and determining scale-invariant features. Subsequently, multiple point-cloud feature-detection algorithms are utilized to calculate highly discriminative features of building roof edges at multiple scales. These features form a feature matrix, which is then input into the categorical boosting machine-learning algorithm for training and prediction, ultimately achieving the accurate extraction of building roof edge points.

Results To verify the proposed method, the airborne point-cloud dataset of Trondheim, Norway is selected for testing. The extraction accuracy of building roof edge points can reach 91.16%, and the time-consuming performance is good and stable.

Conclusions Experimental results demonstrate that the proposed method can converge effectively with only a small amount of labeled data, providing an efficient, accurate, and highly generalizable solution for the extraction of building roof edge points.