Abstract: Objectives： Urban surface change detection is vital for disaster early warning and urban planning. Compared to traditional 2D methods, 3D point cloud change detection (3DPCsCD) can capture vertical changes in ground features and offers greater robustness against variations in illumination and viewing angles. This study aims to address the challenge of balancing computational efficiency and fine-grained recognition in current 3DPCsCD tasks. Methods： This paper proposes a novel network that directly processes unstructured point clouds using a single backbone architecture. First, an efficient temporal-aware embedding strategy is designed to maintain continuous cross-temporal feature interaction during the downsampling process. Then, a hybrid spatial feature association mechanism is introduced to selectively enhance feature responses in changed areas while suppressing background interference from unchanged regions. Finally, a class-optimized multi-task output head and a comprehensive loss function are constructed to further refine the detection accuracy. Results： Experimental validation on a publicly available multi-source dataset demonstrates the following: (1) On airborne point cloud data, the proposed method achieves a mean Intersection over Union (mIoU) of 84.62%, significantly outperforming mainstream methods. (2) On multi-sensor point cloud data, the method exhibits robust recognition and generalization capabilities. (3) An efficiency analysis confirms that the proposed approach achieves an optimal balance between computational cost and detection accuracy. Conclusions： The experimental results demonstrate that the proposed method provides excellent recognition performance for 3D urban change detection. Its efficient design and balanced performance make it highly suitable for meeting the practical demands of complex urban scenarios.