Individual Tree Species Classification using the Pointwise MLP-Based Point Cloud Deep Learning Method

In the practice of forest resource field sample surveys, tree species is an essential survey factor. With the continuous development of LiDAR technology, the use of ground-based LiDAR systems for sample plot scanning can quickly and accurately obtain the 3D structural parameters of forest trees. The accurate identification of tree species information from individual tree laser point clouds is a key focus and difficulty of current research. Traditional machine learning methods require manual extraction of a large amount of 3D structural information for modeling, and the accuracy of recognition is not high. In the field of computer vision, a breakthrough has been made in the classification of 3D object shapes using point cloud deep learning techniques, which provides a new practical direction for tree species classification. To explore the effectiveness of point cloud deep learning in classifying individual tree point cloud species, we use three point-by-point point cloud-based deep learning methods (PointNet, PointNet++, PointMLP) to identify individual tree point clouds of seven types of tree species. We downsampled the number of points in each individual tree point cloud to 1024 and 2048 using the farthest point sampling. We have achieved very exciting experimental results. The experiments using 2048 points involved can obtain higher classification accuracy. PointMLP, the current optimal pointwise MLP-based method, PointNet++ achieved the highest classification accuracy (0.90) in tree species classification. The tree species classification accuracies for the experiments using PointMLP and PointNet methods were 0.80 and 0.40, respectively. Our study illustrates that tree species information can be well identified from individual tree point clouds using pointwise MLP-based deep learning methods. As the current pointwise MLP-based approach of SOTA, PointMLP does not obtain the highest classification accuracy. We are analyzing the reasons for this phenomenon.