Accelerated forest modeling from tree canopy point clouds via deep learning

Rapid generation of tree models from point clouds of tree canopies holds wide-ranging applications in the field of earth sciences, including forest ecology research, environmental monitoring, and forest management. Traditional tree modeling methods rely on procedural models to simulate tree growth, which are timeconsuming due to their extensive manual parameterization. Furthermore, existing deep learning methods struggle to generate visually realistic tree models because of the complex branch structures and specific natural patterns of trees. To address these challenges, this paper proposes a novel deep learning-based method for rapidly generating tree models that align with the shape of the tree canopy. Different from traditional methods, we use deep neural networks to build branch graphs for generating tree models. Our method consists of two main steps: i) the 3D coordinates of each tree node are generated from the canopy point cloud by the designed node coordinate generation network; ii) a graph neural network is proposed to predict node attributes and the adjacency relationship between nodes. To form the tree structure, the discrete nodes are connected by using the minimum spanning tree algorithm combined with the adjacency relationship. The attributes of the node include width, whether it is a leaf node, and leaf node size, which are used for subsequent construction of the tree's mesh. To validate the effectiveness of our proposed method, a large-scale dataset containing 10 forests with 3216 tree canopies is constructed and open sourced for the study of generating tree models from point clouds of tree canopies. Experimental results demonstrate our method's efficiency in generating tree models quickly (reducing the average canopy-to-tree reconstruction time from 7 min to less than 0.5 s) while preserving visual authenticity and accurately matching tree canopy shapes, making it suitable for a wide range of forest reconstructions.