SPTNet: Sparse Convolution and Transformer Network for Woody and Foliage Components Separation From Point Clouds

The separation of woody and foliage components is beneficial in estimating the physical parameters of forests. However, many current methods incur high computational costs and rely on extensive prior knowledge. These methods display weak abilities in generalization for component separation from various LiDAR sensors and tree species. In this paper, a network that combines sparse convolution and transform blocks is proposed for the separation of woody and foliage components in tree point clouds called SPTNet. The sparse convolution block facilitates efficient and effective local feature extraction, while the transformer block offers a solution for the inadequate global feature extraction in sparse convolution blocks. Point feature extraction blocks, called Morphological Detection Coefficient (MDC) and Normal Difference Operator (NDO), were specifically developed to aid in the segmentation task. Distinct adaptive radius strategies are implemented for each geometric feature block to minimize the need for a priori knowledge. Eight different tree species datasets were used to improve methods, including a simulated larch dataset. The other datasets consist of actual trees and comprise seven distinct tree species along with a large tropical tree dataset. Our experimental results demonstrate that our method attains state-of-the-art performance across all datasets. It’s worth mentioning that SPTNet obtains an OA of 94.69% and 89.96% mIoU on the large tropical dataset, which encompasses 15 tree species. Moreover, SPTNet outperforms FWCNN, the current leading branch and leaf separation approach, by 0.43% OA and 0.72% mIoU.