基于深度学习的树木根系探地雷达多目标参数反演识别

李爽; 张潇巍; 谭旭; 徐凌飞; 吕生华; 文剑

doi:10.12171/j.1000-1522.20230259

基于深度学习的树木根系探地雷达多目标参数反演识别

北京林业大学工学院，北京 100083

基金项目: 国家自然科学基金项目（32071679），北京市自然科学基金项目（6202023）。

详细信息

作者简介:
李爽。主要研究方向：根系无损检测。Email：beilinls@163.com　地址：100083 北京市海淀区清华东路35号北京林业大学工学院

责任作者:
文剑，博士，教授。主要研究方向：树木无损检测。Email：wenjian@bjfu.edu.cn　地址：同上。

中图分类号: S771.8
计量
- 文章访问数: 472
- HTML全文浏览量: 76
- PDF下载量: 58
出版历程
- 收稿日期: 2023-09-26
- 修回日期: 2023-12-30
- 网络出版日期: 2024-04-14
- 刊出日期: 2024-04-24

Deep learning-based inverse identification of multi-target parameters for tree rooting ground-penetrating radar

School of Technology, Beijing Forestry University, Beijing 100083, China

摘要

摘要:
目的
使用深度学习方法实现对根系雷达图像的多目标检测和多参数估计。
方法
构建了一种以YOLOv5s和CNN-LSTM为主要框架的网络模型实现对根系雷达扫描图像的多目标检测和多参数估计。首先，通过仿真模拟和预埋试验获取试验所需的根系雷达剖面图数据，同时为了增加数据的多样性，使用CycleGAN风格迁移网络获取了一批具有真实雷达图像特征的仿真数据；然后，使用YOLOv5s目标检测网络识别并提取根系响应区域；接着引入频域变换，获取频域特征，并将根系雷达图像的时域特征和频域特征融合；最后，利用卷积神经网络、卷积注意力机制以及长短期记忆网络强调和提取与根系参数相关的信息特征，并使用多任务学习的方法实现对根系半径、深度、相对介电常数以及水平倾角的预测。
结果
（1）仿真试验中，根系半径估计的最大误差是4.3 mm，R²为0.980，均方根误差为1.32，深度估计的最大误差是35.1 mm，R²为0.962，均方根误差为17.68，相对介电常数估计的最大误差是3.1，R²为0.960，均方根误差为1.10，水平倾角估计的最大误差是10.2°，R²为0.821，均方根误差是4.96。（2）在实测数据上对根系半径估计的平均相对误差是9.112%，深度估计的平均相对误差是5.772%，水平倾角估计的平均相对误差是11.25%。
结论
本文提出的基于深度学习与探地雷达的多目标检测方法可以为根系检测和根系参数估计提供便利。
- 探地雷达 /
- 多参数估计 /
- 无损检测 /
- 多任务学习
Abstract:
Objective
This paper takes multi-target detection and multi-parameter estimation of root radar images using deep learning methods.
Method
In this study, a network model with YOLOv5s and CNN-LSTM as the main framework was constructed to achieve multi-target detection and multi-parameter estimation of root radar scanning images. Firstly, the root system radar profile data required for the experiments were obtained through simulation and pre-embedding experiments, and at the same time, in order to increase the diversity of the data, a batch of simulation data with the characteristics of real radar images were obtained using CycleGAN style migration network; then, the root system response region was identified and extracted using the YOLOv5s target detection network; then, the frequency domain transform was introduced to obtain the frequency domain features and fuse the time domain features and frequency domain features of the root system radar image; finally, a convolutional neural network (CNN), a convolutional attention mechanism, and a long short-term memory network (LSTM) were used to emphasize and extract the information features related to the root system parameters, and a multi-task learning approach was used to achieve the prediction of the root system radius, depth, relative permittivity, and horizontal angle.
Result
(1) In the simulation experiments, the maximum error in the estimation of root radius was 4.3 mm with R² of 0.980 and a root mean square error of 1.32. The maximum error in the estimation of depth was 35.1 mm with R² of 0.962 and a root mean square error of 17.68, and the maximum error in the estimation of relative permittivity was 3.1 with R² of 0.960 and a root mean square error of 1.10, and the maximum error in the estimation of the horizontal angle was 10.2°, with R² of 0.821 and the root mean square error was 4.96. (2) The average relative error of root radius estimation on measured data was 9.112%, the average relative error of depth estimation was 5.772%, and the average relative error of horizontal angle estimation was 11.25%.
Conclusion
The experimental data show that the method proposed in this paper can facilitate root detection and root parameter estimation.
- ground penetrating radar /
- multi-parameter estimation /
- nondestructive testing /
- multi-task learning

HTML全文

图 1 系统框架

Figure 1. System framework

下载: 全尺寸图片幻灯片

图 2 YOLOv5s网络结构

Backbone表示特征提取部分，Neck表示特征融合部分，Prediction表示预测部分。Focus、CSP1_1、CSP1_3、SPP、FPN和PAN表示YOLOv5s中特定的网络结构。Backbone represents the feature extraction part, Neck represents the feature fusion part, and Prediction represents the prediction part. Focus, CSP1_1, CSP1_3, SPP, FPN, and PAN represent specific network structures in YOLOv5s.

Figure 2. YOLOv5s network structure

下载: 全尺寸图片幻灯片

图 3 参数估计模块

CA表示通道注意力模块，SA表示空间注意力模块，CBAM表示卷积注意力模块，C表示特征向量的通道数，H表示特征向量的高，W表示特征向量的宽，LSTM表示长短时记忆网络。CA represents channel attention module, SA represents spatial attention module, CBAM represents convolutional block attention module, C represents the number of channels in the feature vector, H represents the height of the feature vector, W represents the width of feature vector, and LSTM represents long short-term memory network.

Figure 3. Parameter estimation module

下载: 全尺寸图片幻灯片

图 4 CycleGAN 扩充数据集

G1、G2表示生成器，D1、D2表示鉴别器，GAN loss表示生成对抗损失，Cycle loss表示循环一致性损失。G1 and G2 represent generators, D1 and D2 represent discriminators, GAN loss represents generative adversarial network loss, and Cycle loss represents cycle consistency loss.

Figure 4. CycleGAN expanded dataset

下载: 全尺寸图片幻灯片

图 5 数据预处理

Figure 5. Data preprocessing

下载: 全尺寸图片幻灯片

图 6 学习率和损失变化曲线

Figure 6. Learning rate and loss variation curves

下载: 全尺寸图片幻灯片

图 7 仿真模型

Figure 7. Simulation model

下载: 全尺寸图片幻灯片

图 8 指标曲线

Figure 8. Index curves

下载: 全尺寸图片幻灯片

图 9 预测结果与误差分布

Figure 9. Prediction results and error distribution

下载: 全尺寸图片幻灯片

图 10 预埋试验

α表示探地雷达前进方向和根之间的夹角。α represents the angle between the direction of ground penetrating radar movement and the roots.

Figure 10. Pre-embedding experiment

下载: 全尺寸图片幻灯片

表 1 不同模型效果展示

Table 1 Display of different model performances

模型 Model	平均精度 Mean precision	每秒帧数 Frame per second
YOLOv5s	0.788	134
YOLOv5m	0.793	81
YOLOv5l	0.797	34

下载: 导出CSV

表 2 仿真参数设置

Table 2 Simulation parameter settings

参数 Parameter	数值 Numerical value
模型尺寸 Model size	2.500 m × 0.600 m × 0.002 m
空间离散步长 Spatial discrete step length/m	0.002
时间窗口 Time window/ns	15
天线中心频率 Antenna center frequency/MHz	900
土壤相对介电常数 Relative permittivity of soil	6
根系相对介电常数 Relative permittivity of root	7 ~ 30
根系半径 Root radius/cm	0.5 ~ 5.0
深度 Depth/cm	0 ~ 60

下载: 导出CSV

表 3 不同结构网络模型的性能比较

Table 3 Comparison of accuracy of different structural network models

模型 Model	参数 Parameter	CA	FPN	CBAM	最大绝对值误差 Maximum absolute value error	R²	均方根误差 Root mean square error
Model-1	半径Radius/mm				5.6	0.952	2.32
	深度 Depth/mm				459.0	−	266.30
	相对介电常数 Relative permittivity				29.4	−	16.94
	水平倾角 Horizontal inclination/（°）				19.5	0.701	6.28
Model-2	半径 Radius/mm	√			10.3	0.930	2.65
	深度 Depth/mm				58.2	0.952	19.04
	相对介电常数 Relative permittivity				4.2	0.957	1.19
	水平倾角 Horizontal inclination/（°）				13.0	0.746	5.76
Model-3	半径 Radius/mm	√		√	7.7	0.950	2.41
	深度 Depth/mm				78.5	0.947	19.63
	相对介电常数 Relative permittivity				4.3	0.948	1.31
	水平倾角 Horizontal inclination/（°）				13.3	0.773	5.46
Model-4	半径 Radius/mm	√	√		10.9	0.921	2.93
	深度 Depth/mm				73.7	0.936	21.52
	相对介电常数 Relative permittivity				8.4	0.899	1.82
	水平倾角 Horizontal inclination/（°）				12.1	0.793	5.21
Model-N	半径 Radius/mm	√	√	√	4.3	0.980	1.32
	深度 Depth/mm				35.1	0.962	17.68
	相对介电常数 Relative permittivity				3.1	0.960	1.10
	水平倾角 Horizontal inclination/（°）				10.2	0.821	4.96
注：表中–表示无明显相关性，√表示网络中使用了该模块，空白表示模型中未使用此模块。Notes: ”–” indicates a negative coefficient and √ indicates that the module is used in the network，blank indicates that this module is not used in the model.

下载: 导出CSV

表 4 样本根系预埋试验参数预测结果

Table 4 Predicted results of sample root parameters of pre-embedded experiment

序号 No.	图像 Image	根系半径 Root radius				根系深度 Root depth				水平倾角 Horizontal inclination
序号 No.	图像 Image	真实值 Real value/mm	估计值 Estimated value/mm	绝对误差 Absolute error/mm	相对误差 Relative error/%	真实值 Real value/mm	估计值 Estimated value/mm	绝对误差 Absolute error/mm	相对误差 Relative error/%	真实值 Real value/(°)	估计值 Estimated value/(°)	绝对误差 Absolute error/(°)	相对误差 Relative error/%
R1		22.0	20.7	1.3	5.91	278	287	9	3.24	90	84	6	6.7
R2		23.1	23.7	0.6	2.60	273	271	2	0.73	75	63	12	16.0
R3		27.6	28.9	1.3	4.71	245	269	24	9.80	60	64	4	6.7
R4		31.5	23.5	8.0	25.40	268	296	28	10.45	45	52	7	15.6
R5		21.6	20.1	1.5	6.94	302	288	14	4.64

下载: 导出CSV

参考文献(33)

[1]	Liang H, Xing L, Lin J. Application and algorithm of ground-penetrating radar for plant root detection: a review[J/OL]. Sensors, 2020, 20(10): 2836[2022−5−16]. https://doi.org/10.3390/s20102836.
[2]	王哲旭. 基于探地雷达的树木根系检测算法研究[D]. 无锡: 江南大学, 2023. Wang Z X. Reasearch on tree root detection algorithm using ground penetrating radar[D]. Wuxi: Jiangnan University, 2023.
[3]	Liu L B, Zhong Q L, Ni, J. Allometric function-based root biomass estimate of woody plants in a karst evergreen and deciduous broadleaf and mixed forest in central Guizhou Province, southwestern China[J]. Acta Ecologica Sinica, 2018, 38(24): 8726−8732.
[4]	Riedell W E, Osborne S L. Monolith root sampling elucidates western corn rootworm larval feeding injury in maize[J]. Crop Science, 2017, 57(6): 3170−3178. doi: 10.2135/cropsci2017.04.0218
[5]	Chirol C, Carr S J, Spencer K L, et al. Pore, live root and necromass quantification in complex heterogeneous wetland soils using X-ray computed Tomography[J/OL]. Geoderma, 2021, 387: 114898[2021−10−31]. https://doi.org/10.3390/electronics9111804.
[6]	Sharma N, Khajuria Y, Sharma J, et al. Microscopic, elemental and molecular spectroscopic investigations of root-knot nematode infested okra plant Roots[J]. Vacuum, 2018, 158: 126−135. doi: 10.1016/j.vacuum.2018.09.039
[7]	Lei W, Hou F, Xi J, et al. Automatic hyperbola detection and fitting in GPR B-scan Image[J/OL]. Automation in Construction, 2019, 106: 102839[2021−06−25]. https://doi.org/10.1016/j.autcon.2019.102839.
[8]	Luo W, Lee Y H, Ow L F, et al. Accurate tree roots positioning and sizing over undulated ground surfaces by common offset GPR measurements[J]. IEEE Transactions on Instrumentation and Measurement, 2022, 71: 1−11.
[9]	Barton C V M, Montagu K D. Detection of tree roots and determination of root diameters by ground penetrating radar under optimal Conditions[J]. Tree Physiology, 2004, 24(12): 1323−1331. doi: 10.1093/treephys/24.12.1323
[10]	Cui X, Chen J, Shen J, et al. Modeling tree root diameter and biomass by ground-penetrating radar[J]. Science China Earth Sciences, 2011, 54(5): 711−719. doi: 10.1007/s11430-010-4103-z
[11]	Tanikawa T, Ikeno H, Yamase K, et al. Can Ground-penetrating radar detect adjacent roots and rock fragments in forest soil?[J]. Plant and Soil, 2021, 468(1−2): 239−257. doi: 10.1007/s11104-021-05116-3
[12]	Zhang X H, Derival M, Albrecht U, et al. Evaluation of a ground penetrating radar to map the root architecture of HLB-infected citrus trees[J/OL]. Agronomy, 2019, 9(7): 354[2021−07−03]. https://doi.org/10.3390/agronomy9070354.
[13]	Lei W, Luo J, Hou F, et al. Underground cylindrical objects detection and diameter identification in GPR B-scans via the CNN-LSTM framework[J/OL]. Electronics, 2020, 9(11): 1804[2021−10−31]. https://doi.org/10.3390/electronics9111804.
[14]	王泽鹏, 张潇巍, 薛芳秀, 等. 探地雷达树木根系定位与直径估算[J]. 农业工程学报, 2021, 37(8): 160−168. Wang Z P, Zhang X W, Xue F X, et al. Estimating the location and diameter of tree roots using ground penetrating radar[J]. Transactions of the Chinese Society of Agricultural Engineering, 2021, 37(8): 160−168.
[15]	李光辉, 王哲旭, 徐汇, 等. 基于探地雷达和深度学习的果树根径预测方法[J]. 农业机械学报, 2022, 53(11): 306−313, 348. Li G H, Wang Z X, Xu H, et al. Root diameter prediction method of fruit trees based on ground penetrating radar and deep learning[J]. Transactions of the Chinese Society of Agricultural Machinery, 2022, 53(11): 306−313, 348.
[16]	肖维颖, 王健, 李文顺. 松树株数识别的 YOLOv5 轻量化算法研究[J]. 森林工程, 2023, 39(4): 126−133. Xiao W Y, Wang J, Li W S. Research on YOLOv5 lightweight algorithm for pine tree strain identification[J]. Forest Engineering, 2023, 39(4): 126−133.
[17]	王辉, 欧阳缮, 刘庆华, 等. 基于深度学习的探地雷达二维剖面图像结构特征检测方法[J]. 电子与信息学报, 2022, 44(4): 1284−1294. Wang H, Ouyang S, Liu Q H, et al. Structure feature detection method for ground penetrating radar two-dimensional profile image based on deep learning[J]. Journal of Electronics & Information Technology, 2022, 44(4): 1284−1294.
[18]	魏锦山, 陈争光, 焦峰. 基于不同卷积尺度融合与近红外光谱的土地分类模型研究[J]. 光谱学与光谱分析, 2023, 43(2): 460−467. Wei J S, Chen Z G, Jiao F. Research on land classification model based on fusion of different convolution scales and near-infrared spectroscopy[J]. Spectroscopy and Spectral Analysis, 2023, 43(2): 460−467.
[19]	Liu T, Bao J, Wang J, et al. A hybrid CNN–LSTM algorithm for online defect recognition of CO₂ welding[J/OL]. Sensors, 2018, 18(2): 4369[2021−12−10]. https://doi.org/10.3390/s18124369.
[20]	Jeong M, Park M, Nam J, et al. Light-weight student LSTM for real-time wildfire smoke detection[J/OL]. Sensors, 2020, 20(19): 5508[2020−09−25]. https://doi.org/10.3390/s20195508.
[21]	Agethen S, Hsu W H. Deep multi-kernel convolutional LSTM networks and an attention-based mechanism for videos[J]. IEEE Transactions on Multimedia, 2020, 22(3): 819−829. doi: 10.1109/TMM.2019.2932564
[22]	Lantini L, Massimi F, Tosti F, et al. A deep learning approach for tree root detection using GPR spectrogram imagery[C]// 45th International Conference on Telecommunications and Signal Processing (TSP). Prague: IEEE, 2022: 391−394.
[23]	盛同杰, 赵惊涛, 彭苏萍. 地震绕射波弱信号U-net网络提取方法[J]. 地球物理学报, 2023, 66(3): 1192−1204. Sheng T J, Zhao J T, Peng S P. Seismic diffraction extraction method using U-net for weak signals[J]. Chinese Journal of Geophysics, 2023, 66(3): 1192−1204.
[24]	张省, 李山山, 魏国芳, 等. 面向精细化多尺度特征的遥感图像目标检测[J]. 遥感学报, 2022, 26(12): 2616−2628. doi: 10.11834/jrs.20221801 Zhang S, Li S S, Wei G F, et al. Refined multi-scale feature-oriented object detection of remote sensing images[J]. National Remote Sensing Bulletin, 2022, 26(12): 2616−2628. doi: 10.11834/jrs.20221801
[25]	Paknezhad M, Rengarajan H, Yuan C, et al. Improving transparency and representational generalizability through parallel continual Learning[J]. Neural Networks, 2023, 161: 449−465. doi: 10.1016/j.neunet.2023.02.007
[26]	Gu J, Ye J C. AdaIN-based tunable CycleGAN for efficient unsupervised low-dose CT denoising[J]. IEEE Transactions on Computational Imaging, 2021, 7: 73−85. doi: 10.1109/TCI.2021.3050266
[27]	Dou H, Chen C, Hu X, et al. Asymmetric CycleGAN for image-to-image translations with uneven complexities[J]. Neurocomputing, 2020, 415: 114−122. doi: 10.1016/j.neucom.2020.07.044
[28]	Qin Y, Wang Z, Xi D. Tree CycleGAN with maximum diversity loss for image augmentation and its application into gear pitting Detection[J/OL]. Applied Soft Computing, 2022, 114: 108130[2022−11−27]. https://doi.org/10.1016/j.asoc.2021.108130.
[29]	Pan J, Sun M, Wang Y, et al. A time-delay estimation approach for coherent GPR signals by taking into account the noise pattern and radar Pulse[J/OL]. Signal Processing, 2020, 176: 107654[2021−06−05]. https://doi.org/10.1016/j.sigpro.2020.107654.
[30]	Wang X, Liu S. Noise suppressing and direct wave arrivals removal in GPR data based on Shearlet Transform[J]. Signal Processing, 2017, 132: 227−242. doi: 10.1016/j.sigpro.2016.05.007
[31]	王明凯, 李文彬, 文剑. 基于探地雷达对粗根的识别技术研究[J]. 森林工程, 2020, 36(3): 21−27. Wang M K, Li W B, Wen J. Study on recognition technology of coarse roots using ground-penetrating radar[J]. Forest Engineering, 2020, 36(3): 21−27.
[32]	邱啟璜, 牛健植, 王迪, 等. 基于探地雷达识别林地粗根和石砾[J]. 北京林业大学学报, 2023, 45(7): 99−109. Qiu Q H, Niu J Z, Wang D, et al. Identification of coarse roots and rock fragments in woodland based on ground penetrating radar[J]. Journal of Beijing Forestry University, 2023, 45(7): 99−109.
[33]	朱安宁, 吉丽青, 张佳宝, 等. 不同类型土壤介电常数与体积含水量经验关系研究[J]. 土壤学报, 2011, 48(2): 263−268. Zhu A N, Ji L Q, Zhang J B, et al. Empirical relationship between soil dielectric constant and volumetric water content in various soils[J]. Acta Pedologica Sinica, 2011, 48(2): 263−268.