Abstract: Abstract: An accurate three-dimensional (3D) model has been one of the most practical significance to acquire the plant phenotypic traits without damage, particularly for the accurate planting, visual management, and intelligent control of crops. However, the existing plant 3D model cannot fully meet the rapid and accurate requirement of modern agriculture, due to its low accuracy of point cloud registration and a large amount of data during modeling. In this study, a classical iterative closest point (ICP) registration of point clouds was proposed for the plant 3D modeling using lightweight processing. The original data of point clouds was firstly sampled to determine the initial corresponding point set, where the wrong corresponding points were removed as well. The optimal coordinate transformation was then calculated by the least square method, indicating a better accuracy of stitching. Nevertheless, the operation speed and the convergence to the global optimum depended mainly on the given initial transformation and the generation relation. Therefore, the ICP was then optimized to construct a three-dimensional plant model with fewer data, in order to avoid the ICP falling into the local extremum. The specific procedure of optimization was as follows. Firstly, the background of the plant image from the point cloud was filtered by the human-computer interaction, where the outlier noise of the point cloud was identified to denoise using the efficient nearest neighbor search. The auxiliary feature coordinate parameters were also introduced to establish the feature points. Secondly, the initial transformation matrix of ICP was obtained by manual intervention. An initial solution was then provided to avoid the local optimal solution in the accurate registration for the multiple groups of point clouds. Finally, a three-dimensional voxel grid was created to approximately display the points in the voxel using the center of gravity of all points in each voxel. All points in the voxel were represented by a center of gravity, thereby effectively filtering out the redundant data points in the point clouds. The results show that better performance of registration was achieved for the plant point cloud, where the clear and outstanding phenotype was easy to be distinguished after ICP optimization. Both single and multiple plants presented an excellent registration, clear phenotype, and distinguishable details. The redundant data of the plant point cloud was also reduced by 96.90%-97.35% after simplification. The simplified point cloud of plant phenotype can be widely expected to effectively reconstruct the morphological traits of plants. Specifically, the plant height error of single plant was 0.20%-0.45%, and the crown width error was 0.17%-0.47%, whereas, the plant height error of multiple plants was 0.25%-0.60%, and the crown width error was 0.42%-0.80%. Consequently, the optimized ICP can achieve the precision fusion of point clouds in 124.3 s, which was 26.75% higher than the traditional in 169.7s, indicating an effective reduction in the redundancy data. The finding can also provide a strong reference for the lightweight processing on the 3D modeling of plant phenotype.