Forest point cloud registration using the tree top and the ground-level tree center

Unmanned aerial vehicle (UAV) and ground-based photogrammetry have been essential to the forest resource surveys, due to the low cost, high efficiency and scalability. However, the complex and heterogeneous nature of forest environments can often lead to the UAV platforms with the less understory information. While the ground platforms can frequently miss the canopy details. The information gaps can be caused by different observation perspectives from UAVs and ground platforms. Some challenges are also remained to efficiently register photogrammetric point clouds. In this study, a point cloud registration was proposed for forest images, according to three-dimensional (3D) point cloud from UAV and ground photogrammetry. Key points of feature were then extracted using coarse-to-fine registration, including the tree top and the ground-level tree center. The specific procedures were as follows: 1) Image Point Cloud Acquisition. In UAV platforms, DJI Phantom 4 RTK UAV were used to capture plot images via oblique photogrammetry. In ground platforms, a "simulated flight path" was employed to obtain the plot images. 3D reconstruction was implemented to generate 3D point cloud data. 2) Feature Point Extraction. Single tree segmentation was combined with AABB bounding boxes and Euclidean clustering. The highest points of individual UAV tree canopies were extracted as feature points. Euclidean clustering was used on ground image point clouds. Single trees were then segmented to extract the ground-level tree center as ground feature points. 3) Coarse Registration. Feature points were mapped from 3D to 2D. Transformation relationships were calculated using improved simulated annealing, partial/full point pair transformation and scale adjustment. The coarse registration was achieved in the original point cloud. 4) Fine Registration. Precise registration was achieved to reduced further error using iterative closest point. A series of tests were performed on six sample plots with different tree species. The results demonstrated the following: 1) The high accuracy and reliability were validated experimentally in the single tree segmentation with AABB bounding boxes and Euclidean clustering. 2) The better performance and stability were achieved in the forest image point cloud registration with improved simulated annealing and iterative closest point. The average errors of coarse and fine registration were 0.09 and 0.06 m, respectively, in the six sample plots. The coarse registration was closely matched the fine registration, indicating the refined registration of image point clouds from UAV and ground platforms. 3) Poplar trees shared a greater impact on coarse registration, compared with ginkgo and catalpa trees. While there was the minimal impact of different tree species on the fine registration. 4) Scale differences were emerged as the influencing factors on forest image point cloud registration. It was necessary to consider and correct during registration. Therefore, cross-platform registration was successfully implemented on the forest image point clouds using UAV and ground-based photogrammetry. A promising approach was presented to tackle the challenges of forest image point cloud registration. This finding can also provide the sound support to forest resource surveys, and 3D forest reconstruction in the wide application of image point clouds. Furthermore, the practical insights can be offered for the precision and intelligent forestry production.