文献类型：期刊文献

中文题名：LiDAR单木分割辅助的无人机影像CNN+EL树种识别

英文题名：Tree Species Recognition Based on Unmanned Aerial Vehicle Image with LiDAR Individual Tree Segmentation Aided

作者：徐志扬[1,2] 陈巧[1,3] 陈永富[1,3]

第一作者：徐志扬

机构：[1]中国林业科学研究院资源信息研究所,北京100091;[2]国家林业和草原局华东调查规划院,杭州310019;[3]国家林业和草原局林业遥感与信息技术重点实验室,北京100091

年份：2022

卷号：53

期号：3

起止页码：197-205

中文期刊名：农业机械学报

外文期刊名：Transactions of the Chinese Society for Agricultural Machinery

收录：CSTPCD;;EI(收录号：20221511962148);Scopus;北大核心:【北大核心2020】；CSCD:【CSCD2021_2022】；

基金：中央级公益性科研院所基本科研业务费专项资金项目(CAFYBB2018SZ008)。

语种：中文

中文关键词：树种识别;无人机影像;无人机激光雷达;卷积神经网络;集成学习

外文关键词：tree species recognition;unmanned aerial vehicle image;UAVLiDAR;convolutional neural network;ensemble learning

分类号：S771

摘要：为研究激光雷达单木分割辅助条件下无人机可见光图像树种识别应用潜力,提出联合卷积神经网络(CNN)和集成学习(EL)的树种识别方法。首先利用同期无人机激光雷达数据和可见光影像数据进行单木树冠探测并制作单木树冠影像数据集;其次引入ResNet50网络并结合引入有效通道注意力机制、替换膨胀卷积、调整卷积模块层数搭建出4个卷积神经网络,使用ImageNet大型数据集进行模型预训练,加载预训练参数进行模型初始化并利用制作的单木树冠影像数据集训练出5个不同的分类模型;最后通过相对多数投票法建立集成模型。实验结果表明,单木探测总体精度达到83.80%,集成学习的训练精度、验证精度、独立测试精度分别达到了99.15%、98.34%和90.15%,较ResNet50网络提高了4.23、3.04、9.09个百分点,独立测试精度较随机森林分类最优结果高32.31个百分点。激光雷达单木分割辅助条件下利用卷积神经网络和集成学习策略能够充分提取无人机图像特征用于树种识别。
In order to study the application potential of tree species recognition based on unmanned aerial vehicle(UAV)visible image with LiDAR individual tree segmentation aided,a tree species recognition method combined with convolutional neural network and ensemble learning was proposed.Firstly,individual trees were detected by means of individual tree segmentation of simultaneous UAVLiDAR point clouds and multiscale segmentation of UAV visible image,and then individual tree canopy image datasets was sliced from UAV visible image.Secondly,ResNet50 convolutional neural network was introduced,meanwhile,a ECAResNet50 network was bulit by using ResNet50 as the backbone network framework and inserting the effective channel attention(ECA)mechanism model to residual bottleneck module,and then a ECA-ResNet50Dialate network was bulit by replacing normal 3×3 convolution of residual module with dilated convolution,and ECA-ResNetmini and ECA-ResNetminiDialate network were bulit by adjusting the convolution layer number of convolutional modules in the end.The pre-trained model parameters,which were pre-trained by using ImageNet datasets,were loaded to initialize the five network models,after that five recognition models were trained by using the individual tree canopy image datasets.Finally,the five convolutional neural network models were ensembled by the relative majority voting method.The experimental results showed that the overall accuracy of individual tree detection was 83.80%,and the training,verification and independent test accuracy of ensemble learning reached 99.15%,98.34%and 90.15%,respectively,which were 4.23,3.04 and 9.09 percentage points higher than that of ResNet50 network,and the independent test accuracy was still 32.31 percentage points higher than the traditional optimal result of random forest classification.The combination of convolutional neural network and ensemble learning strategy could fully extract UAV visible image features for tree species recognition with LiDAR individual tree segmentation aided.