PointNet++ (MSG-DSLS): A Classification Framework With Dynamic Step-Size Loop-Sampling for High-Density Tree Point Clouds

PointNet++ (MSG), a PointNet++ module, can accurately identify tree species from 3-D point clouds. However, one drawback of the MSG method is its inability to classify high-density point clouds accurately. As the number of points in the input increases, the data quality improves but the classification ability of the model decreases. Compared with point clouds of other objects, tree point clouds possess more complex branching structures, requiring enhanced input density to better recover their detailed features. Therefore, PointNet++ (MSG-dynamic step-size loop sampling, MSG-DSLS) is proposed, and experiments are conducted on datasets from three different regions. The results demonstrate that the MSG-DSLS algorithm improves the classification accuracy by an average of 0.0242 across all the datasets, and the highest classification accuracy reaches 0.9886. In this study, the inability of the MSG method to classify high-density point clouds accurately is addressed, and the advantages of deep learning for tree species recognition are illustrated, providing an experimental reference for related studies.