Semantic supported urban change detection using ALS point clouds

Abstract

Airborne laser scanning (ALS) has experienced significant development and boosted widespread applications over the recent years. With increasing data quality and acquisition convenience, multi-temporal ALS point clouds have been used for change detection (CD) in earth observation. Data registration or geometric difference calculations can effectively obtain geometric changes using point clouds acquired from various epochs. However, the lack of semantic labeling means that traditional CD methods are unable to distinguish the changes in semantic information or remove the interference in some categories. Therefore, we propose a semantic supported change detection (SSCD) method. First, a point cloud semantic segmentation network is developed under the classic PointNet++ frame to provide semantic labeling for change detection, including an attention-based local feature embedding module that uses the spatially sensitive features of 3D points and an integrated multi-scale framework with local spatial attention convolution. Secondly, the point cloud is voxelized based on an Octree, where the label of each voxel is determined by the result of semantic segmentation. Finally, the change detection result is obtained by comparing the changes in voxel coordinates and semantic information of the two point clouds. We tested the proposed method using experiments conducted on the AHN datasets. In the two selected experimental areas, the accuracy of change detection rate reached 0.9099 and 0.9029, respectively. The F1 rate reached 0.4571 and 0.3455. Evaluations using the object-based method show that changes were correctly identified in both of the tested areas.