To effectively utilize forest point cloud data for estimating forest carbon sinks, it is crucial to achieve accurate and efficient separation of leaf and branch point clouds. Forest point clouds are characterized by strong overlap, complex features, and massive data volume. Existing research methods face challenges such as high computational costs, reliance on tree geometric and spectral features, and weak cross-species generalization capabilities. To address these challenges, this paper proposes the use of the PoinTramba method, which employs Transformer and Mamba to model intra-group and inter-group relationships, respectively. Based on the PoinTramba framework, we further improve its grouping strategy and adjust the BIO strategy to adapt to the characteristics of forest point clouds, enabling it to handle the heterogeneous data of forest point clouds while enhancing cross-species generalization. Experimental results show that the proposed method achieves 94.20% OA (overall accuracy), 85.53% mIoU (mean intersection over union), and 87.37% mAcc (mean accuracy) on the test set, demonstrating significant improvements compared to Transformer-based methods. Additionally, the practical effects of the grouping strategy and BIO strategy adopted in this paper are analyzed, showing certain advantages over the basic PoinTramba method and other improvement measures.
树木的结构特征(如树高、胸径、冠幅)是评估其碳储量的关键。快速准确地获取这些特征,对于森林碳储量测算至关重要[1-4]。传统人工测量方式费时费力、效率低下而地面三维激光扫描技术可通过非破坏性方式,从点云中提取树干材积并估测生物量,为树木结构定量分析与三维重建提供了新方法[5]。基于点云的结构分析首先需要实现枝叶与枝干的精确分割。传统方法多依赖几何与辐射特征。Vicari等[6]结合无监督几何分类与最短路径分析实现枝叶分离;Ferrara等[7]基于密度的含噪声空间聚类算法(density-based spatial clustering of applications with noise,DBSCAN)算法,通过体素密度差异分割树冠;Ma等[8]则融合区域生长、形态分割与k-means进行细化。然而,这些方法受传感器差异与参数设置影响较大,在准确性与效率上存在局限。在单木骨架重构方面,赵永辉等[9]采用Laplace收缩与拓扑细化方法,提升了枝干重建的精度与效率;赵钧坤等[10]基于关键路径探测策略,实现了分层递进的骨架提取;杨军等[11]基于背包式激光扫描(backpack laser scanning,BLS)点云成功构建了蒙古栎胸径-树高模型,证实了背包激光雷达在天然林单木因子提取中的可行性与精度,为林业数字化管理提供了数据基础。
现有方法面临三大挑战:首先,因为Transformer在点云分析任务中,其注意力机制的复杂性是二次方的,所以基于Transformer模型的方法因其二次复杂度导致计算成本很高;其次,现有方法过度依赖几何或光谱特征,限制了方法的跨树种泛化能力;最后,林木点云重叠性强,点云特征驳杂,数量庞大,这些导致处理效率低下。随着状态空间模型(state space models,SSM)的提出,通过动态建模实现高效的训练和推理,显著降低内存使用量[15],为解决Transformer方法的固有问题提供了可能。本研究提出了一种基于状态空间特征增强(PoinTramba)[16]的树木点云枝干叶分割方法。PoinTramba是一种用于点云分析的新型混合框架,其利用了Transformer的强大分析能力和Mamba的效率。为了减少对于树木几何和光谱等特征的依赖,适应性提取不同树种的多样性特征,同时提高点云分割的效率,并减少树木点云参数量大的问题带来的影响,本研究进行了如下改进。
1)分组策略改进
采用基于自适应最近邻密度峰值聚类算法(adaptive nearest neighbor density peak clustering,ANN-DPC)[17]的自适应聚类分组,通过集成ANN-DPC的自适应邻域搜索与动态密度阈值,PoinTramba的分组策略在保持计算效率的同时,显著提升了组内一致性与跨树种泛化能力。试验验证其在复杂森林点云中的优越性,为后续Transformer层的长程依赖建模提供了更高质量的特征输入。
采用ANN-DPC策略对PoinTramba的分组方式进行了优化,在分组策略的对比试验中,设置了基于排序识别聚类结构算法(ordering points to identify the clustering structure,OPTICS)分组和最远点采样法(farthest point sampling,FPS)方法与本方法进行对比。通过OA、mIoU来判断对分割性能的影响,见表3。
CHENX G, ZHANGX Q, ZHANGY P,et al.Changes of carbon stocks in bamboo stands in China during 100 years[J].Forest Ecology and Management,2009,258(7):1489-1496.
[2]
LIW, NIUZ, GAOS,et al.Correlating the horizontal and vertical distribution of LiDAR point clouds with components of biomass in a Picea crassifolia forest[J].Forests,2014,5(8):1910-1930.
[3]
XIAS B, WANGC, PANF F,et al.Detecting stems in dense and homogeneous forest using single-scan TLS[J].Forests,2015,6(11):3923-3945.
[4]
YENT M, JIY J, LEEJ S.Estimating biomass production and carbon storage for a fast-growing makino bamboo (Phyllostachys makinoi) plant based on the diameter distribution model[J].Forest Ecology and Management,2010,260(3):339-344.
XIAM P, GUANF Y, FANS H,et al.Application status and prospect of TLS in forest resources inventory[J].Journal of Northwest Forestry University,2018,33(3):238-244.
[7]
VICARIM B, DISNEYM, WILKESP,et al.Leaf and wood classification framework for terrestrial LiDAR point clouds[J].Methods in Ecology and Evolution,2019,10(5):680-694.
[8]
FERRARAR, VIRDISS G, VENTURAA,et al.An automated approach for wood-leaf separation from terrestrial LIDAR point clouds using the density based clustering algorithm DBSCAN[J].Agricultural and Forest Meteorology,2018,262:434-444.
[9]
MAZ Y, PANGY, WANGD,et al.Individual tree crown segmentation of a larch plantation using airborne laser scanning data based on region growing and canopy morphology features[J].Remote Sensing,2020,12(7):1078.
YANGJ, WANGF, SONGZ Y.Development of natural forest Quercus mongolica diameter-height curve model based on BLS[J].Forest Engineering,2023,39(5):57-64.
[16]
CHENX X, JIANGK, ZHUY S,et al.Individual tree crown segmentation directly from UAV-borne LiDAR data using the PointNet of deep learning[J].Forests,2021,12(2):131.
[17]
WIELGOSZM, PULITIS, WILKESP,et al.Point2Tree (P2T)—Framework for parameter tuning of semantic and instance segmentation used with mobile laser scanning data in coniferous forest[J].Remote Sensing,2023,15(15):3737.
[18]
ZHANGS, CHENY P, WANGB,et al.SPTNet:Sparse convolution and transformer network for woody and foliage components separation from point clouds[J].IEEE Transactions on Geoscience and Remote Sensing,2024,62:5702718.
[19]
GUA,DAO T.Mamba:Linear-time sequence modeling with selective state spaces[J].arXiv preprint arXiv:2023.
[20]
WANGZ C, CHENZ H, WUY M,et al.Pointramba:A hybrid transformer-mamba framework for point cloud analysis[J].arXiv preprint arXiv:2024.
[21]
YANH, WANGM Z, XIEJ Y.ANN-DPC:Density peak clustering by finding the adaptive nearest neighbors[J].Knowledge-Based Systems,2024,294:111748.
[22]
MEAGHERD.Geometric modeling using octree encoding[J].Computer Graphics and Image Processing,1982,19(2):129-147.
[23]
WANGH P, LIUY, DONGZ,et al.You only hypothesize once:Point cloud registration with rotation-equivariant descriptors[C]//2022 30th ACM International Conference on Multimedia,October 10-14,2022,Lisbon,Portugal.2022.
[24]
MOMO TAKOUDJOUS, PLOTONP, SONKÉB,et al.Using terrestrial laser scanning data to estimate large tropical trees biomass and calibrate allometric models:A comparison with traditional destructive approach[J].Methods in Ecology and Evolution,2018,9(4):905-916.
[25]
XIZ X, HOPKINSONC, ROODS B,et al.See the forest and the trees:Effective machine and deep learning algorithms for wood filtering and tree species classification from terrestrial laser scanning[J].ISPRS Journal of Photogrammetry and Remote Sensing,2020,168:1-16.
[26]
MOP H, ZHANGY J, ZHAOZ Y,et al.High-speed and low-power molecular dynamics processing unit (MDPU) with ab initio accuracy[J].NPJ Computational Materials,2024,10(1):253.
[27]
CHOYC, PARKJ, KOLTUNV.Fully convolutional geometric features[C]//2019 IEEE/CVF international conference on computer vision.October 27- November 02,2019,Seoul,Korea (South):IEEE,2019.
[28]
QIANG, HAMMOUDH, LIG,et al.Assanet:An anisotropic separable set abstraction for efficient point cloud representation learning[J].Advances in Neural Information Processing Systems,2021,34:28119-28130.