UFNet：基于移动激光扫描（MLS）点云的景观森林高大乔木林下干扰物滤除方法

UFNet: an intelligent filtering approach for understory disturbance of tall trees in landscape forests using Mobile Laser Scanning (MLS) point clouds

摘要

【目的】高大乔木的结构稳定性是景观森林可持续管理的核心，其精准监测与信息提取已成为城市绿地管理的关键任务。然而，林下支撑物与灌木等干扰物常遮挡或混淆主干点云，制约林分信息获取精度。针对这一问题，本研究提出一种林下干扰物自动滤除方法，以获取清晰、准确的高大乔木主干结构信息，支撑城市森林数字化监测与精细化管理。【方法】本文提出一种基于移动激光扫描（MLS）点云的智能滤除方法 UFNet。为准确提取体素内的局部几何特征，该方法首先通过体素特征编码模块实现原始点云的特征增强；随后，利用子流形稀疏卷积模块在保持数据稀疏性的同时高效提取深层语义信息，避免特征在复杂林区场景下的无效扩散；最后，通过层次化的编码器−解码器结构实现多尺度特征融合，以增强对细小干扰物与乔木主干边界的感知能力，实现干扰物的精准自动滤除。【结果】UFNet 在复杂林下环境中展现出稳定的语义分割性能，能够有效区分干扰物与树干。在上海植物园林木样地的实验中，UFNet 整体表现为精度 88.72%、召回率 88.65%、 F₁ 88.69%、交并比 79.67%。与 5 种深度学习方法相比，UFNet 的精度、召回率、 F₁、交并比分别提升 3.48 ~ 10.21、5.04 ~ 11.37、4.27 ~ 10.80 和 6.63 ~ 15.88 个百分点。干扰物滤除后，点云结构显著清晰，单株分割的 F₁ 分数提升至 94.64%，交并比提升至 89.83%。【结论】UFNet 具备出色的林下干扰物滤除能力，可有效提升高大乔木结构信息的获取精度，可为以乔木为主体的城市景观森林开展三维建模与精细化管理提供关键技术支撑。

Abstract

[Objective] The structural stability of tall trees is central to the sustainable management of landscape forests, and their precise monitoring and information extraction have become key tasks in urban green space management. However, understory clutter, such as supports and shrubs, often obscures or confuses trunk point clouds, limiting the accuracy of forest information acquisition. To address this issue, this study proposes an automatic understory clutter filtering method to obtain clear and accurate structural information of tall tree trunks, thereby supporting digital monitoring and refined management of urban forests. [Method] This paper presents UFNet, an intelligent filtering method based on MLS point clouds. To accurately extract local geometric features within voxels, the method first enhances the original point cloud features through a voxel feature encoding module. Subsequently, a submanifold sparse convolution module is employed to efficiently extract deep semantic information while preserving data sparsity, thus avoiding ineffective feature diffusion in complex forest scenes. Finally, a hierarchical encoder-decoder architecture is used to achieve multi-scale feature fusion, enhancing the perception of small clutter objects and tree trunk boundaries, thereby enabling precise automatic filtering of understory clutter. [Result] UFNet demonstrates robust semantic segmentation performance in complex understory environments, effectively distinguishing clutter from trunks. In experiments conducted in forest plots at the Shanghai Botanical Garden, UFNet achieved an overall precision of 88.72%, recall of 88.65%, F₁-score of 88.69%, and Intersection over Union (IoU) of 79.67%. Compared with five other deep learning methods, UFNet improved these four metrics by 3.48–10.21, 5.04–11.37, 4.27–10.80, and 6.63–15.88 percentage points, respectively. Following clutter filtering, the point cloud structure became markedly clearer, with the F₁-score for individual tree segmentation increasing to 94.64% and the IoU rising to 89.83%. [Conclusion] UFNet exhibits effective capability in understory clutter filtering and can effectively enhance the accuracy of structural information acquisition for tall trees. It provides critical technical support for 3D modeling and refined management of tree-dominated urban landscape forests.

关键词

城市森林 / 移动激光扫描（MLS） / 林下干扰物滤除 / 点云语义分割 / 子流形稀疏卷积 / 层次化编码器−解码器 / 深度学习 / 单木分割

Key words

urban forest / Mobile Laser Scanning (MLS) / understory clutter filtering / point cloud semantic segmentation / submanifold sparse convolution / hierarchical encoder-decoder / deep learning / individual tree segmentation

参考文献

[1] Zhang B, Li X, Du H, et al. Estimation of urban forest characteristic parameters using UAV-lidar coupled with canopy volume[J].Remote Sensing, 2022, 14(24): 6375.

[2] Zhang J, Wang J, Liu G. Vertical structure classification of a forest sample plot based on point cloud data[J].Journal of the Indian Society of Remote Sensing, 2020, 48(8): 1215-1222.

[3] 张建鹏, 王金亮, 刘广杰, 等. 基于地基雷达点云主方向的林下植被自动滤除[J].遥感技术与应用, 2023, 38(2): 405-412.

[4] Zhang J P, Wang J L, Liu G J, et al. Automatic filtering of understory vegetation based on point cloud main direction of terrestrial laser scanning[J].Remote Sensing Technology and Application, 2023, 38(2): 405-412.

[5] 聂辰辰, 程峰, 王成, 等. 地面激光雷达点云数据乔灌分离方法研究[J].测绘科学, 2024, 49(1): 106-116.

[6] Nie C C, Cheng F, Wang C, et al. Terrestrial laser scanning point cloud data tree-shrub separation method research[J].Science of Surveying and Mapping, 2024, 49(1): 106-116.

[7] Henrich J, van Delden J, Seidel D, et al. TreeLearn: a deep learning method for segmenting individual trees from ground-based LiDAR forest point clouds[J].Ecological Informatics, 2024, 84: 102888.

[8] Xia K, Li C, Yang Y, et al. Study on single-tree extraction method for complex RGB point cloud scenes[J].Remote Sensing, 2023, 15(10): 2644.

[9] Kim D H, Ko C U, Kim D G, et al. Automated segmentation of individual tree structures using deep learning over LiDAR point cloud data[J].Forests, 2023, 14(6): 1159.

[10] Zhou Y, Tuzel O. VoxelNet: End-to-end learning for point cloud based 3D object detection[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City. IEEE, 2018: 4490-4499.

[11] Graham B, Engelcke M, van der Maaten L. 3D semantic segmentation with submanifold sparse convolutional networks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City. IEEE, 2018: 9224-9232.

[12] Zhang G, Junnan C, Gao G, et al. Hednet: A hierarchical encoder-decoder network for 3d object detection in point clouds[J].Advances in Neural Information Processing Systems, 2023, 36: 53076-89.

[13] Liu K, Wang W, Tharmarasa R, et al. Ground surface filtering of 3D point clouds based on hybrid regression technique[J].IEEE Access, 2019, 7: 23270-23284.

[14] Cheng B, Girshick R, Dollar P, et al. Boundary IoU: improving object-centric image segmentation evaluation[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville. IEEE, 2021: 15329-15337.

[15] Martin D, Fowlkes C, Tal D, et al. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics[C]//Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001. Vancouver. IEEE, 2001: 416-423.

[16] Choy C, Gwak J, Savarese S. 4D spatio-temporal ConvNets: Minkowski convolutional neural networks[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach. IEEE, 2019: 3070-3079.

[17] Kingma D P, Ba J. Adam: a method for stochastic optimization [C]//International Conference on Learning Representations (ICLR). San Diego: ICLR. 2015.

[18] Ilya L, Frank H. SGDR: stochastic gradient descent with warm restarts [C]//International Conference on Learning Representations (ICLR). Toulon: ICLR. 2017.

[19] Qi C R, Yi L, Su H, et al. Pointnet++: deep hierarchical feature learning on point sets in a metric space [C]//Advances in Neural Information Processing Systems (NeurIPS). Long Beach: NeurIPS. 2017.

[20] Jiang L, Zhao H, Shi S, et al. PointGroup: dual-set point grouping for 3D instance segmentation[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle: IEEE, 2020: 4866-4875.

[21] Fan S, Dong Q, Zhu F, et al. SCF-net: Learning spatial contextual features for large-scale point cloud segmentation[C]//2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville: IEEE, 2021: 14499-14508.

[22] Park J, Kim C, Kim S, et al. PCSCNet: Fast 3D semantic segmentation of LiDAR point cloud for autonomous car using point convolution and sparse convolution network[J].Expert Systems with Applications, 2023, 212: 118815.

[23] 邢万里, 邢艳秋, 黄杨, 等. 基于体元逐层聚类的 TLS 点云数据单木分割算法[J].中南林业科技大学学报, 2017, 37(12): 58-64,71.

[24] Xing W L, Xing Y Q, Huang Y, et al. Individual tree segmentation of TLS point cloud data based on clustering of voxels layer by layer[J].Journal of Central South University of Forestry & Technology, 2017, 37(12): 58-64,71.

基金资助

国家自然基金面上项目(32371876)