基于改进YOLOv5模型的城市樟树单木树冠检测方法

doi:10.13466/j.cnki.lczyyj.2026.02.009

摘要/Abstract

摘要：

单木检测在城市生态系统管理和绿化监测中具有关键作用。尽管深度学习算法凭借在计算机视觉领域的优异表现,被广泛应用于单木检测,但大多数方法在复杂的城市环境中难以实现精准检测,尤其在树冠大小不一、树种多样且郁闭度高的城市复杂环境中表现不佳。为了克服这些局限性,提出一种基于YOLOv5的改进模型—YOLOv5-CSC,用于检测无人机红绿蓝(RGB)图像中的樟树。该模型首先通过改进型卷积神经网络(ConvNeXt)模块中的多样化卷积核实现细粒度的局部特征提取,并通过增加快速空间金字塔池化(SPPF)模块和C3TR模块提升感受野与特征提取、融合的效率,有效提高了模型在复杂城市环境中树冠检测的能力。实验结果显示,YOLOv5-CSC模型与现有的Faster R-CNN、SSD和YOLOv7模型相比,其平均精度(mAP@0.5)分别提高了5.3%、8.8%和8.2%;相较于YOLOv5模型,其平均精度(mAP)从80.3%提升至85.5%。改进后的模型在处理复杂背景和树冠重叠的场景中表现出更高的精度和鲁棒性,并在不同场景下展现出卓越的适应性,有效支持了城市绿化管理和生态监测的现代化需求。

关键词: 单木检测, 深度学习, 无人机, ConvNeXt, YOLOv5

Abstract:

Individual tree detection plays a critical role in urban ecosystem management and greening monitoring.While deep learning algorithms are widely applied in tree detection due to their strong performance in computer vision,most methods struggle to achieve precise detection in complex urban environments,especially with varying crown sizes,diverse species,and high canopy density.To address these challenges,this paper proposes an improved YOLOv5-based algorithm,YOLOv5-CSC,for detecting camphor trees in unmanned aerial vehicle(UAV RGB)images.The model uses diverse convolution kernels in the ConvNeXt module for fine-grained local feature extraction,and incorporates SPPF and C3TR modules to enhance the receptive field and feature fusion efficiency,significantly improving crown detection in complex urban settings.Experimental results show that the YOLOv5-CSC model achieves mean average precision(mAP@0.5)improvements of 5.3%,8.8%,and 8.2% compared to existing models such as Faster R-CNN,SSD,and YOLOv7.Compared to the baseline YOLOv5 model,the mAP increased from 80.3% to 85.5%.The improved model demonstrates higher accuracy and robustness in handling complex backgrounds and overlapping crowns,and excels in adaptability across various scenarios,effectively supporting urban greening management and ecological monitoring.

Key words: individual tree detection, deep learning, unmanned aerial vehicle(UAV), ConvNeXt, YOLOv5

中图分类号:

S771.8

郑哲, 夏凯. 基于改进YOLOv5模型的城市樟树单木树冠检测方法[J]. 林草资源研究, 2026,(2): 93-101.

ZHENG Zhe, XIA Kai. Urban single-tree crown detection method for camphor trees based on an improved YOLOv5 model[J]. Forest and Grassland Resources Research, 2026,(2): 93-101.

图/表 10

图1

表1

图2

图3

图4

表2

表3

图5

表4

图6

参考文献 18

[1]	Jabbar M, Yusoff M M, Shafie A. Assessing the role of urban green spaces for human well-being:A systematic review[J]. GeoJournal, 2022, 87:4405-4423. doi: 10.1007/s10708-021-10474-7
[2]	Semeraro T, Scarano A, Buccolieri R, et al. Planning of urban green spaces:An ecological perspective on human benefits[J]. Land, 2021, 10(2):105. doi: 10.3390/land10020105
[3]	Sishodia R P, Ray R L, Singh S K. Applications of remote sensing in precision agriculture:A review[J]. Remote Sensing, 2020, 12(19):3136. doi: 10.3390/rs12193136
[4]	曹明兰, 李亚昆, 张力小, 等. 深度学习在无人机遥感城市行道树提取中的应用[J]. 中南林业科技大学学报, 2023, 43(6):79-84.
[5]	Qin Haiming, Zhou Weiqi, Yao Yang, et al. Individual tree segmentation and tree species classification in subtropical broadleaf forests using UAV-based LiDAR,hyperspectral,and ultrahigh-resolution RGB data[J]. Remote Sensing of Environment, 2022, 280:113143. doi: 10.1016/j.rse.2022.113143
[6]	杨军, 苏显辉, 王彦博, 等. 无人机遥感数据的森林病虫害变色立木提取分析[J]. 林业勘查设计, 2024, 53(2):67-71.
[7]	Monakhova K, Yanny K, Aggarwal N, et al. Spectral DiffuserCam:Lensless snapshot hyperspectral imaging with a spectral filter array[J]. Optica, 2020, 7(10):1298-1307. doi: 10.1364/OPTICA.397214
[8]	Plesoianu A I, Stupariu M S, ŞAndric I, et al. Individual tree-crown detection and species classification in very high-resolution remote sensing imagery using a deep learning ensemble model[J]. Remote Sensing, 2020, 12(15):2426. doi: 10.3390/rs12152426
[9]	Deng Susu, Xu Qi, Yue Yuanzheng, et al. Individual tree detection and segmentation from unmanned aerial vehicle-LiDAR data based on a trunk point distribution indicator[J]. Computers and Electronics in Agriculture, 2024, 218:108717. doi: 10.1016/j.compag.2024.108717
[10]	Vahidi H, Klinkenberg B, Johnson B A, et al. Mapping the individual trees in urban orchards by incorporating volunteered geographic information and very high resolution optical remotely sensed data:A template matching-based approach[J]. Remote Sensing, 2018, 10(7):1134. doi: 10.3390/rs10071134
[11]	Fu Hancong, Zhao Hengqian, Jiang Jinbao, et al. Automatic detection tree crown and height using Mask R-CNN based on unmanned aerial vehicles images for biomass mapping[J]. Forest Ecology and Management, 2024, 555:121712. doi: 10.1016/j.foreco.2024.121712
[12]	赵子琪, 李丹丹, 赵鼎, 等. 基于深度学习的树冠分割及生物量估算[J]. 森林工程, 2024, 40(5):145-155.
[13]	Wang Jiansen, Zhang Huaiqing, Liu Yang, et al. Tree-Level Chinese fir detection using UAV RGB imagery and YOLO-DCAM[J]. Remote Sensing, 2024, 16(2):335. doi: 10.3390/rs16020335
[14]	Touvron H, Cornet V, Jouvet T, et al. Resmlp:Feedforward networks for image classification with data-efficient training[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 45(4):5314-5321. doi: 10.1109/TPAMI.2022.3206148
[15]	Wu Weichao, Fan Xijian, Qu Hongyu, et al. TCDNet:Tree crown detection from UAV optical images using uncertainty-aware one-stage network[J]. IEEE Geoscience and Remote Sensing Letters, 2022, 19:1-5.
[16]	Zhuang Fuzhen, Qi Zhiyuan, Duan Keyu, et al. A comprehensive survey on transfer learning[J]. Proceedings of the IEEE, 2020, 109(1):43-76. doi: 10.1109/PROC.5
[17]	Wang Xiaofeng, Wang Yi, Zhou Chaowei, et al. Urban forest monitoring based on multiple features at the single tree scale by UAV[J]. Urban Forestry & Urban Greening, 2021, 58:126958.
[18]	Lucena F, Breunig F M, Kux H. The combined use of UAV-based RGB and DEM images for the detection and delineation of orange tree crowns with mask R-CNN:An approach of labeling and unified framework[J]. Future Internet, 2022, 14(10):275. doi: 10.3390/fi14100275

环境	参数
系统环境	Windows 11
CPU	Intel Xeon Platinum 8358P @2.60 GHz
GPU	NVIDIA GeForce GTX 3090
内存	48 GB
深度学习框架	PyTorch 1.10.0,CUDA11.3和CUDNN8.2.4

YOLOv5	mAP/%	R/%	P/%
使用预训练权重	80.3	76.0	78.2
未使用预训练权重	78.9	72.5	84.5

组别	ConvNeXt	SPPF	C3TR	mAP/%	参数量 /(×10⁶)	GFLOPs/ (×10⁹)
1				80.3	46.1	10.3
2		√		82.1	54.8	14.5
3		√	√	84.2	58.5	15.8
4	√			83.3	60.5	16.3
5	√	√	√	85.5	65.8	16.6

模型	mAP/%	R/%	P/%	F₁
SSD	76.7	57.1	86.8	0.69
Faster R-CNN	80.2	80.0	57.1	0.67
YOLOv5	80.3	76.0	78.2	0.77
YOLOv7	77.3	66.5	81.6	0.74
YOLOv5-CSC	85.5	82.6	77.2	0.80