Semantic and Geometric Fusion for Object-Based 3D Change Detection in LiDAR Point Clouds

Accurate three-dimensional change detection is essential for monitoring dynamic environments such as urban areas, infrastructure, and natural landscapes. Point-based methods are sensitive to noise and lack spatial coherence, while object-based approaches rely on clustering, which can miss fine-scale changes. To address these limitations, we introduce an object-based change detection framework integrating semantic segmentation and geometric change indicators. The proposed method first classifies bi-temporal point clouds into ground, vegetation, buildings, and moving objects. A cut-pursuit clustering algorithm then segments the data into spatially coherent objects, which are matched across epochs using a nearest-neighbor search based on centroid distance. Changes are characterized by a combination of geometric features—including verticality, sphericity, omnivariance, and surface variation—and semantic information. These features are processed by a random forest classifier to assign change labels. The model is evaluated on the Urb3DCD-v2 dataset, with feature importance analysis to identify important features. Results show an 81.83% mean intersection over union. An additional ablation study without clustering reached 83.43% but was more noise-sensitive, leading to fragmented detections. The proposed method improves the efficiency, interpretability, and spatial coherence of change classification, making it well suited for large-scale monitoring applications.