CSA-PointNet:一种面向针阔混交林的树种分类模型  

CSA-PointNet:a tree species classification model for coniferous and broad-leaved mixed forests

作　　者：霍燕平 冷亮[1] 王民水 纪雪 王明常[1] HUO Yanping;LENG Liang;WANG Minshui;JI Xue;WANG Mingchang(College of Geo-exploration Science and Technology,Jilin University,Changchun 130026,China)

机构地区：[1]吉林大学地球探测科学与技术学院,长春130026

出　　处：《世界地质》2024年第4期551-556,573,共7页World Geology

基　　金：国家自然科学基金面上项目(42171407);吉林省教育厅科学研究项目(JJKH20241288KJ)。

摘　　要：LiDAR技术能够快速获取树木的三维结构信息,结合深度学习算法可以实现单株级别的树种分类。为了解决将点云转换为二维图像或三维体素时损失树木的垂直结构、树冠形状和空间分布等信息的问题,笔者在PointNet的基础上提出了CSA-PointNet,该模型可以直接将树木的单株点云作为输入,通过引入通道注意力机制和空间注意力机制,增强对单体植株的细节特征和空间分布特征的提取能力,从而提高树种分类精度。应用制作的净月林场数据集分别对CSA-PointNet与主流模型(VoxNet和PointNet)进行训练和测试,结果表明CSA-PointNet的整体分类精度为74.69%,kappa系数为0.51,均高于VoxNet和PointNet。LiDAR technology is able to quickly acquire 3D structural information of trees,and when combined with deep learning algorithms,it can achieve tree species classification at the single plant level.In order to solve the problem of losing the vertical structure,canopy shape and spatial distribution information of trees when converting point clouds into 2D images or 3D voxels,the authors propose CSA-PointNet based on PointNet,which can directly take the point cloud of an individual tree as input.By introducing both the channel-attention mechanism and the spatial-attention mechanism,it enhances the extraction of detailed features and spatial distribution characteristics,thereby improving the accuracy of tree classification.The Jingyue forest dataset is used to train and test CSA-PointNet and mainstream models(VoxNet and PointNet),respectively.The results show that the overall classification accuracy of CSA-PointNet is 74.69%,and the kappa coefficient is 0.51,both of which are higher than those of VoxNet and PointNet.

关 键 词：LIDAR 单木分割 深度学习 注意力机制 树种分类 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]