TreePMNet: a deep matching network for registration of UAV and handheld LiDAR data in fine phenotyping of rubber plantations

Advanced phenotyping is essential for precise growth monitoring and reliable yield forecasting of economically significant rubber trees. Unmanned laser scanning (ULS) from an aerial perspective and handheld laser scanning (HLS) from the ground can efficiently capture 3D structural information of rubber plantations. Registering these two datasets can compensate for respective information gaps and achieve a more complete 3D representation. However, existing point cloud registration methodologies often suffer from low robustness or computational inefficiency when applied to forest scenes. This study proposes TreePMNet, a deep learning-based point cloud registration method. First, a Local Feature Aggregation (LFA) module is designed to generate salient keypoints and distinctive descriptors. Subsequently, a Progressive Transformation Estimation (PTE) module is used to achieve coarse-to-fine registration. The proposed method was evaluated using the public ULS-HLS-Rubber dataset. Experimental results demonstrate that TreePMNet facilitates both effective and efficient registration of ULS and HLS point clouds. The average RMSE, RRE, and RTE values reached 0.09 m, 0.17°, and 0.22 m, respectively. Furthermore, the average processing time for a 900 m 2 plot was only 0.98 s. The R 2 values for tree height, diameter at breast height (DBH), and crown diameter extracted from the registered point cloud data reached 0.989, 0.856, and 0.914, respectively. This study underscores the developmental potential of deep learning-based methods for ULS and HLS point cloud registration in complex forest environments. Our findings provide invaluable support for the accurate acquisition of 3D structural information and advanced phenotypic characterization in rubber plantations.