【作者】 李爽；

【导师】 文剑； 林晨；

【作者基本信息】 北京林业大学 ， 电子信息硕士（专业学位）， 2024， 硕士

【摘要】 根系不仅是树木重要的营养器官,还起到固定和支撑植株的作用。相较于地上部分,根系难以观察和取样,导致根系研究相对落后。传统根系检测方法需要花费大量时间和精力,甚至会对根系土壤环境造成破坏。探地雷达(Ground Penetrating Radar,GPR)技术因其高效、无损和便捷性受到林业专家的广泛关注,但是由于土壤环境以及根系分布复杂导致雷达波解译困难,给根系检测和参数识别带来了困难。为了解决上述问题,本文研究了基于YOLOv5目标检测算法的根系反射双曲线区域自动识别和提取以及基于CNN-LSTM网络的根系参数估计网络,为根系雷达检测和参数估计提供了支撑。本文主要研究内容和成果如下:(1)基于YOLOv5实现了根系雷达图像目标检测。针对仿真雷达数据和真实雷达数据风格差异大且获取大量可用于网络训练的真实数据困难的问题,研究了基于Cycle GAN的图像风格迁移网络,该网络能够实现不同域之间图像风格的转换,通过该网络能够快速获取大量具有真实数据特征的仿真数据;针对根系雷达图像中因双曲线特征不明显以及叠加问题导致的检测困难和漏检的问题,对YOLOv5网络的特征提取和特征融合部分的网络结构以及确定目标检测框时的NMS算法进行改进,有效提升了网络对根系雷达图像的特征提取能力以及对双曲线重叠区域的检测能力,提升网络对根系雷达图像的检测效果。(2)构建了基于CNN-LSTM的根系多参数估计网络。根系雷达成像受多个参数的共同影响,不同根系雷达图像的时域以及频域特征均能反映根系参数的变化。本文构建的CNN-LSTM网络以根系雷达图像的时频域融合特征作为输入,从中提取根系双曲线区域的空间位置特征和时序特征,实现对根系多个参数的估计。仿真和实测数据的实验结果表明,本文提出的根系参数估计方案能够实现根系多个参数的估计。(3)研制了根系探地雷达数据处理专用软件,并进行了现场实验。结合前述研究中使用的根系识别算法、根系参数估计算法以及根系雷达数据预处理算法,本文研发了基于Py Qt5的根系雷达数据处理软件。为了验证该软件的稳定性和有效性,进行了现场实验,现场试验结果表明该软件能够用来处理根系雷达数据。更多还原

【Abstract】 Roots are not only important nutrient organs for trees but also play a role in anchoring and supporting plants.Compared to the aboveground parts,roots are difficult to observe and sample,leading to relatively lagging research in root systems.Traditional root detection methods require a significant amount of time and effort,and may even cause damage to the root-soil environment.Ground Penetrating Radar(GPR)technology has attracted widespread attention from forestry experts due to its efficiency,non-destructiveness,and convenience.However,the complex soil environment and root distribution make radar wave interpretation difficult,posing challenges to root detection and parameter identification.To address these issues,this study investigated the automatic recognition and extraction of root reflection hyperbolic regions based on the YOLOv5 object detection algorithm,as well as the root parameter estimation network based on the CNN-LSTM network,providing support for root radar detection and parameter estimation.The main research content and achievements of this study are as follows:(1)Based on YOLOv5,root radar image object detection has been achieved.To address the challenge of significant stylistic differences between simulated radar data and real radar data,coupled with the difficulty of obtaining a large amount of real data suitable for network training,a Cycle GAN-based image style transfer network was studied.This network can achieve style transfer between different domains,enabling the rapid acquisition of a large amount of simulated data with realistic data characteristics.Furthermore,to address the detection difficulties and missed detections in root radar images caused by unclear hyperbolic features and overlay issues,improvements were made to the feature extraction and fusion parts of the YOLOv5 network,as well as to the Non-Maximum Suppression(NMS)algorithm used for determining target detection boxes.These enhancements effectively enhanced the network’s ability to extract features from root radar images and detect hyperbolic overlapping regions,thereby improving the detection performance of the network on root radar images.(2)A CNN-LSTM-based root multi-parameter estimation network has been constructed.Root radar imaging is influenced by multiple parameters,and the temporal and frequency domain features of different root radar images can reflect changes in root parameters.The CNN-LSTM network constructed in this study takes the fusion features of time-frequency domain of root radar images as input,extracts spatial positional features and temporal features of root hyperbolic regions,and achieves estimation of multiple root parameters.Experimental results using both simulated and real-world data demonstrate that the proposed root parameter estimation approach can effectively estimate multiple root parameters.(3)A specialized software for processing root ground penetrating radar(GPR)data has been developed and field experiments have been conducted.Combining the root identification algorithm,root parameter estimation algorithm,and root radar data preprocessing algorithm used in the previous research,a root radar data processing software based on Py Qt5 has been developed.To validate the stability and effectiveness of this software,field experiments were conducted.The on-site experimental results demonstrate that the software is capable of processing root-zone GPR data.更多还原