文献类型：期刊文献

中文题名：融合注意力和残差网络的无人机影像树种识别

英文题名：Tree Species Recognition Using Combined Attention and ResNet for Unmanned Aerial Vehicle Images

作者：徐志扬[1,2,3] 陈巧[1,2] 陈永富[1,2]

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]国家林业和草原局林业遥感与信息技术重点实验室,北京100091;[3]国家林业和草原局华东调查规划院,浙江杭州310019

年份：2023

卷号：60

期号：2

起止页码：80-88

中文期刊名：激光与光电子学进展

外文期刊名：Laser & Optoelectronics Progress

收录：CSTPCD;;Scopus;WOS:【ESCI(收录号:WOS:001154750000008)】；北大核心:【北大核心2020】；CSCD:【CSCD2023_2024】；

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;残差网络;有效通道注意力;无人机可见光图像;单木树冠影像块

外文关键词：tree species recognition;residual network(ResNet);efficient channel attention(ECA);UAV visible image;single tree crown image patch

分类号：S771

摘要：为探索采用无人机(UAV)遥感影像进行亚热带树种识别的应用潜力,提出一种结合残差模块和有效通道注意力的网络(ECA-ResNet)对单木树冠影像数据集进行模型训练和识别。首先,利用单木分割算法提取单木树冠,构建不同尺度的UAV可见光影像单株树冠影像块样本数据集,并将其划分为训练数据、验证数据和独立测试集;其次,以ResNet50为主干网络,在瓶颈层插入有效通道注意力并调整网络结构,构建ECA-ResNet;最后,将数据集载入预训练的ECA-ResNet模型,进行参数迭代训练和验证并进行独立测试,择优确定单木树冠的合适窗口大小。结果表明:ECAResNet对64×64像素的单木树冠影像数据集中树种的识别效果更为理想,训练精度和验证精度分别达98.98%和96.60%,独立测试识别精度、Kappa系数分别达85.61%、0.8140;ECA-ResNet模型的训练、验证、独立测试精度分别高于ResNet50网络2.63个百分点、1.80个百分点、5.31个百分点。该研究结果证明卷积神经网络(CNN)能够充分提取可见光图像的空间特征,有效通道注意力能够有效提升CNN的单木树种识别能力。
To explore the application potential of unmanned aerial Vehicle(UAV)remote sensing images for subtropical tree species recognition,the ECA-ResNet with residual module and effective channel attention is proposed to train and recognize single tree crown images.First,the single tree crown was extracted by single-tree segmentation algorithm.The single-tree crown image patch dataset of UAV visible image was constructed by means of clipping images with different window sizes,and they were divided into training data,validating data,and independent test dataset respectively.Second,with ResNet50 as a backbone network,by inserting effective channel attention into ResNet bottleneck and adjusting network structure,the ECA-ResNet was constructed.Then,the datasets were inputted into pretrained ECAResNet model for parameter training and validation iteratively,and independent test.After that,the optimum window size of single-tree crown image was determined.The results show that the ECA-ResNet gets a better recognition result for tree species in single-tree crown image patch dataset with window size of 64×64 pixel,the accuracy of training and validation of the proposed network reaches 98.98%and 96.60%,respectively.The recognition accuracy and Kappa coefficient of independent test reach 85.61%and 0.8140.The training,validation,and independent test accuracy of ECA-ResNet in this paper are 2.63 percentage points,1.80 percentage points,and 5.31 percentage points higher than that of the ResNet50 respectively.It is proved that,convolutional neural network(CNN)can fully extract the spatial features of UAV visible images for tree species recognition,effective channel attention can effectively improve CNN'single tree species recognition capability.