结合地面LiDAR和无人机影像点云的森林场景三维建模框架

doi:10.11707/j.1001-7488.LYKX20250278

摘要/Abstract

摘要：

目的: 针对复杂森林场景重建精度低、训练耗时长和细节丢失等问题，融合LiDAR数据与无人机影像点云，提出一种基于多源数据协同与3D高斯溅射优化的三维建模框架，以提升重建精度和效率，为森林资源精准管理提供技术支撑。方法: 运用多尺度迭代最近点（ICP）算法实现多源数据配准，融合光谱与垂直结构信息构建建模基础；引入3D高斯溅射优化网络（3DGS）进行点云初始化，分析多源数据的几何、辐射和语义特征，实现高斯点参数化表示；同时结合变形多层感知机（MLP）增强建模，并采用显式光线?高斯体渲染生成高质量森林场景三维模型。结果: 基于自建Forest数据集验证，并与NeRF、MVSNeRF、3DGS、GaussianPro对比，结果显示，倒角距离（CD）最低达8.2 cm，较NeRF降低73.4%；坡度较大区域树高反演均方根误差（RMSE）为0.92 m，冠幅提取准确率达93.04%；消融试验验证3DGS优化网络与变形MLP的协同有效性，训练时间显著缩短。结论: 多源数据融合3DGS框架能够实现森林场景的三维重建，不仅可提高重建精度，还可有效缩短训练时间，为森林资源的数字化管理和可持续经营提供关键技术支撑，可广泛应用于森林资源清查和生态监测等领域。

关键词: LiDAR, 无人机, 3D高斯溅射优化网络, 森林场景重建, 单木参数提取

Abstract:

Objective: To address the problems of low reconstruction accuracy, long training time, and loss of details in complex forest scenes, this study integrates LiDAR data with unmanned aerial vehicle (UAV) image-derived point clouds and proposes a three-dimensional modeling framework based on multi-source data collaboration and 3D Gaussian splatting optimization. The proposed framework aims to improve reconstruction accuracy and efficiency, providing technical support for precise forest resource management. Method: Multi-scaleiterative closest point (ICP) algorithm was employed to achieve multi-source data registration, and spectrum was integrated with vertical structural information to establish the modeling foundation. A 3D Gaussian splatting (3DGS) optimization network was subsequently introduced to initialize point clouds, analyze geometric, radiometric, and semantic features of multi-source data, and realize parametric representation of Gaussian points. At the same time, deformable multilayer perceptron (MLP) was integrated to enhance modeling, and explicit ray-Gaussian volume rendering was used to generate high-quality 3D forest scene models. Result: Validation was conducted on a self-constructed Forest dataset, with comparative analysis against NeRF, MVSNeRF, 3DGS, and GaussianPro. The results showed that the chamfer distance (CD) reached a minimum of 8.2 cm, representing a 73.4% reduction compared to neural radiance fields (NeRF). For tree height inversion in steep-slope areas, the root mean square error (RMSE) was 0.92 m, and the accuracy of crown width extraction was 93.04%. Ablation experiment confirmed the synergistic effectiveness of the 3DGS optimization network and deformable MLP, with significant reduction in training time. Conclusion: The proposed multi-source data fusion 3DGS framework effectively achieves high-precision, high-efficiency 3D reconstruction of forest scenes, not only improves reconstruction accuracy, but also substantially reduces training time, providing critical technical support for digital management and sustainable operation of forest resources, with broad applicability in forest resource inventory and ecological monitoring.

Key words: LiDAR, unmanned aerial vehicle (UAV), 3D Gaussian splatting (3DGS), forest scene reconstruction, individual tree parameter extraction

中图分类号:

S717.19

赵紫琦,雷令婷,柴国奇,张晓丽. 结合地面LiDAR和无人机影像点云的森林场景三维建模框架[J]. 林业科学, 2026, 62(9): 96-109.

Ziqi Zhao,Lingting Lei,Guoqi Chai,Xiaoli Zhang. A 3DGS 3D Modeling Framework for Forest Scenes Based on the Combination of LiDAR Data and UAV Image Point Clouds[J]. Scientia Silvae Sinicae, 2026, 62(9): 96-109.

图/表 18

图1

表1

图2

表2

表3

图3

表4

图4

图5

图6

图7

表5

图8

表6

图9

图10

图11

图12

参考文献 0

	耿　林, 李明泽, 范文义, 等. 基于机载LiDAR的单木结构参数及林分有效冠的提取. 林业科学, 2018, 54 (7): 62- 72. doi: 10.11707/j.1001-7488.20180707
	Geng L, Li M Z, Fan W Y, et al. Individual tree structure parameters and effective crown of the stand extraction base on airborne LiDAR data. Scientia Silvae Sinicae, 2018, 54 (7): 62- 72. doi: 10.11707/j.1001-7488.20180707
	李永宁, 刘利民, 崔立艳, 等. 林冠开阔度对GNSS RTK在森林中定位可用性及初始化时间的影响. 林业科学, 2014, 50 (2): 78- 84. doi: 10.11707/j.1001-7488.20140212
	Li Y N, Liu L M, Cui L Y, et al. Effects of canopy openness on positioning availability and initialization time of GNSS RTK in forests. Scientia Silvae Sinicae, 2014, 50 (2): 78- 84. doi: 10.11707/j.1001-7488.20140212
	朱俊峰, 刘清旺, 崔希民, 等. 地基与无人机LiDAR融合点云的人工林蓄积量估算. 遥感技术与应用, 2024, 39 (1): 45- 54. doi: 10.11873/j.issn.1004-0323.2024.1.0045
	Zhu J F, Liu Q W, Cui X M, et al. Estimation of volume based on fusion point cloud of terrestrial and UAV LiDAR. Remote Sensing Technology and Application, 2024, 39 (1): 45- 54. doi: 10.11873/j.issn.1004-0323.2024.1.0045
	Ardizzone F, Cardinali M, Galli M, et al. 2007. Identification and mapping of recent rainfall-induced landslides using elevation data collected by airborne Lidar. Natural Hazards and Earth System Sciences, 7(6): 637‒650.
	Bagdasarian M T, Knoll P, Li Y H, et al. 3DGS.zip: a survey on 3D gaussian splatting compression methods. Computer Graphics Forum, 2025, 44, e70078. doi: 10.1111/cgf.70078
	Chang D, Božič A, Zhang T, et al. 2022. RC-MVSNet: unsupervised multi-view stereo with neural rendering. Computer Vision–ECCV 2022. Cham: Springer Nature Switzerland, 665‒680.
	Chang J R, Chen Y S. 2018. Pyramid stereo matching network. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, USA. IEEE, 5410‒5418.
	Chen A P, Xu Z X, Zhao F Q, et al. 2021. MVSNeRF: fast generalizable radiance field reconstruction from multi-view stereo. 2021 IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, QC, Canada. IEEE, 14104‒14113.
	Cheng K, Long X X, Yang K Z, et al. 2024. GaussianPro: 3D Gaussian splatting with progressive propagation. arXiv: 2402.14650.
	Chirico G B, Bonavolontà F. Metrology for agriculture and forestry 2019. Sensors, 2020, 20 (12): 3498. doi: 10.3390/s20123498
	Ding Y K, Yuan W T, Zhu Q T, et al. 2022. TransMVSNet: global context-aware multi-view stereo network with transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, LA, USA. IEEE, 8575‒8584.
	Gong Y H, Yu L T, Yue G H. 2024. Isotropic Gaussian splatting for real-time radiance field rendering. arXiv: 2403.14244.
	Guan H, Smith W A. P. Structure-from-motion in spherical video using the von mises-fisher distribution. IEEE Transactions on Image Processing, 2017, 26 (2): 711- 723. doi: 10.1109/TIP.2016.2621662
	Guo J W, Xu S B, Yan D M, et al. Realistic procedural plant modeling from multiple view images. IEEE Transactions on Visualization and Computer Graphics, 2020, 26 (2): 1372- 1384. doi: 10.1109/TVCG.2018.2869784
	Held J, Vandeghen R, Hamdi A, et al. 2025.3D convex splatting: radiance field rendering with 3D smooth convexes. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, TN, USA. IEEE, 21360‒21369.
	Huang Y H, Sun Y T, Yang Z Y, et al. 2024. SC-GS: sparse-controlled Gaussian splatting for editable dynamic scenes. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle, WA, USA. IEEE, 4220‒4230.
	Huang Z X, Xu M, Perry S. 2025. StructGS: adaptive spherical harmonics and rendering enhancements for superior 3D Gaussian splatting. arXiv: 2503.06462.
	Li X X, Huang J X, Li X, et al. Review of PPP-RTK: achievements, challenges, and opportunities. Satellite Navigation, 2022, 3 (1): 28. doi: 10.1186/s43020-022-00089-9
	Mildenhall B, Srinivasan P P, Tancik M, et al. 2020. NeRF: representing scenes as neural radiance fields for view synthesis. Computer Vision – ECCV 2020. Cham: Springer International Publishing, 405‒421.
	Rublee E, Rabaud V, Konolige K, et al. 2011. ORB: an efficient alternative to SIFT or SURF. 2011 International Conference on Computer Vision. Barcelona, Spain. IEEE, 2564‒2571.
	Shao J, Yao W, Wan P, et al. Efficient co-registration of UAV and ground LiDAR forest point clouds based on canopy shapes. International Journal of Applied Earth Observation and Geoinformation, 2022, 114, 103067. doi: 10.1016/j.jag.2022.103067
	Szalai-Gindl J M, Varga D. FPFH revisited: histogram resolutions, improved features, and novel representation. IEEE Access, 2024, 12, 67325- 67354. doi: 10.1109/ACCESS.2024.3400591
	Wang Y B, Jiang L, Yang M H, et al. 2019. Eidetic 3D LSTM: A Model for Video Prediction and Beyond. 2019 International Conference on Learning Representations (ICLR). New Orleans, LA, USA.
	Wu M Y, Dai H Z, Yao K X, et al. 2025. BG-triangle: Bézier Gaussian triangle for 3D vectorization and rendering. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Nashville, TN, USA. IEEE, 16197‒16207.
	Xu C L, Wu C D, Qu D K, et al. Accurate and efficient stereo matching by log-angle and pyramid-tree. IEEE Transactions on Circuits and Systems for Video Technology, 2021, 31 (10): 4007- 4019. doi: 10.1109/TCSVT.2020.3044891
	Yang Y, Xu H R, Weng L F. 2022. A multi-view matching method based on PatchmatchNet with sparse point information. Proceedings of the 4th World Symposium on Software Engineering. Xiamen China. ACM, 9‒14.
	Yao Y, Luo Z X, Li S W, et al. 2018. MVSNet: depth inference for unstructured multi-view stereo. Computer Vision – ECCV 2018. Cham: Springer International Publishing, 785‒801.
	You H T, Liu Y, Lei P, et al. Segmentation of individual mangrove trees using UAV-based LiDAR data. Ecological Informatics, 2023, 77, 102200. doi: 10.1016/j.ecoinf.2023.102200
	Yu T Y, Ni W J, Liu J L, et al. Extraction of tree heights in mountainous natural forests from UAV leaf-on stereoscopic imagery based on approximation of ground surfaces. Remote Sensing of Environment, 2023, 293, 113613. doi: 10.1016/j.rse.2023.113613
	Yun T, Jiang K, Li G C, et al. Individual tree crown segmentation from airborne LiDAR data using a novel Gaussian filter and energy function minimization-based approach. Remote Sensing of Environment, 2021, 256, 112307. doi: 10.1016/j.rse.2021.112307
	Zhang K, Riegler G, Snavely N, et al. 2020. NeRF++: analyzing and improving neural radiance fields. arXiv: 2010.07492.

样地编号 Plot number	优势树种 Dominant tree species	坡度 Slope/ (°)	数量 Count	冠幅 Crown width/m	树高 Tree height/m
AB01	桉树 Eucalyptus spp.	42	53	4.40±1.00	15.08±2.36
AXZ01	桉树 Eucalyptus spp.	47	15	2.70±1.20	27.20±3.40
H06	红锥 Castanopsis hystrix	31	68	4.82±1.76	18.10±5.64
HXZ02	红锥 Castanopsis hystrix	34	48	4.46±1.11	17.63±4.39
S06	杉木 Cunninghamia lanceolata	30	16	4.10±1.03	19.21±4.53
S07	杉木 Cunninghamia lanceolata	27	30	4.50±1.02	18.73±3.11
GJ1-1	白桦树 Betula platyphylla	36	62	3.84±1.55	17.07±3.02
Q18	落叶松 Larix gmelinii	14	45	5.51±1.72	15.23±3.46
Q20	落叶松 Larix gmelinii	9	54	4.26±1.00	14.92±5.21

样地编号 Plot number	优势树种 Dominant tree species	平均点密度 Average point density/(pts·m^–2)	平均强度 Average intensity
AB01	桉树Eucalyptus spp.	94 357.214	11 706.795
AXZ01	桉树Eucalyptus spp.	92 014.623	10 211.335
H06	红锥Castanopsis hystrix	9 842.486	11 782.395
HXZ02	红锥Castanopsis hystrix	88 347.036	13 526.341
S06	杉木Cunninghamia lanceolata	68 698.334	13 307.698
S07	杉木Cunninghamia lanceolata	69 423.779	14 271.322
GJ1-1	白桦树Betula platyphylla	12 020.731	58.823
Q18	落叶松Larix gmelinii	45 296.338	66.352
Q20	落叶松Larix gmelinii	9 602.912	67.533

样地编号 Plot number	优势树种 Dominant tree species	平均点密度 Average point density/ (pts·m^–2)	平均强度 Average intensity	强度标准差 Intensity standard deviation
AB01	桉树Eucalyptus spp.	113 262.778	12 126.505	5 594.697
AXZ01	桉树Eucalyptus spp.	123 006.728	13 231.213	5 123.132
H06	红锥Castanopsis hystrix	172 85.203	11 232.123	2 394.632
HXZ02	红锥Castanopsis hystrix	118 352.624	13 001.441	4 671.397
S06	杉木 Cunninghamia lanceolata	72 698.335	12 307.221	2 345.343
S07	杉木 Cunninghamia lanceolata	71 158.774	14 507.308	5 499.795

样地编号 Plot number	优势树种 Dominant tree species	平均点密度 Average point density/ (pts·m^–2)	平均强度 Average intensity	强度标准差 Intensity standard deviation
GJ1-1	白桦树 Betula platyphylla	12 300.896	58.243	49.667
Q18	落叶松 Larix gmelinii	46 409.771	63.692	53.828
Q20	落叶松 Larix gmelinii	10 714.084	77.571	54.118

方法Methods	Time/min	CD/cm	$ \mathrm{\mathit{R}}_{\mathrm{q}} $	Cov (%)
A. w/o 3DGS	171.2	10.5	4.32	91.2
B. w/o MLP	119.52	9.8	4.55	92.5
C. w/o 3DGS & MLP	22.47	11.2	3.76	90.2
D. 本研究方法Ours	19.26	8.2	4.87	95.1