结合物理与几何特性的机载LiDAR数据分类方法

Airborne LiDAR data classification method combining physical and geometric characteristics

机载LiDAR数据分类是根据数据特征为每个点指定类别标签。针对现有方法忽略全波形与点云在物理特性上的关联、缺乏对邻域几何和语义相关性的深入挖掘,从而导致捕获局部结构能力不足的问题,搭建了结合目标物理与几何特性的分类方法,实现了由全波形和点云组成的机载LiDAR数据端到端分类。首先,构建了特征融合模块,提取了全波形时序特征和点云几何特征,依据两种数据物理意义上的关联,通过双低秩矩阵实现了全波形与点云特征级融合。其次,构建了邻域特征增强模块,挖掘点对相关性,增强对局部几何结构的学习。最后基于层次化编解码结构搭建了分类网络。该网络在机载LiDAR数据集上测试,达到平均精度0.96、平均召回率0.90、平均F1分数0.92,证明了网络的有效性。

Objective Full-waveform airborne LiDAR can not only obtain 3D point clouds,but also obtain full-waveform of the target.The classification of full-waveform LiDAR data is to assign a category label to each point based on data characteristics.It is an important part of data post-processing and has significant application value in the fields such as remote sensing and topographic mapping.However,existing methods overlook the correlation between full-waveform and point clouds geometry in terms of physical characteristics,and lack in-depth exploration of the correlation in neighborhood geometry and semantic features between point pairs.Therefore,the existing methods lack the ability to capture local structural information,which affects the classification effect.To address the above issues,a classification method based on the physical and geometric characteristics of the target is proposed.This method mines the high-dimensional feature association between 3D point clouds and fullwaveforms on the high-dimensional physical characteristics of the target,and enhances the learning of local geometric structural information,thereby improving the classification ability of the model.Methods A classification method based on target physical and geometric features is proposed.Firstly,a highdimensional feature fusion module is proposed,which extracted the rich physical characteristics from the fullwaveform,and extracted the geometric features of the point clouds.Based on the correlation between fullwaveforms and point clouds on the physical characteristics of the target,the complex relationship between the two is learned through a dual low-rank matrix,and deeper physical features are excavated(Fig.2).Secondly,a local neighborhood feature enhancement module is designed to enhance the learning of local geometric structures by constructing a fully-connected neighborhood structure,mining the geometric and semantic correlations between neighboring point pairs(Fig.3).Finally,by using the hierarchical encoder-decoder structure,the characteristics of multiple receptive field can be conbined.A classification method based on the physical characteristics and geometric characteristics can be constructed to improve the classification ability of the model(Fig.4).Results and Discussions The proposed method is based on the correlation between point clouds and fullwaveforms on the physical characteristics of the target.By fusing the two,the physical meaning of the features is enriched and the learning of local geometric structures is enhanced.From Tab.2,it can be seen that the proposed method successfully generated correct labels for most points,achieving an average accuracy of 0.96,a recall of 0.90,and an F1 score of 0.92.Multiple methods such as FCN,GACNN,and FWNet2 are used for testing on the same dataset.Compared with FWNet2,which has the best performance among existing methods,our method has effective improvements in average accuracy,average recall,and average F1 score.Among them,compared with FWNet2,the accuracy,recall,and F1 score of our method have been improved by 0.02,0.01,and0.02 respectively in the ground category,and by 0.02,0.04,and 0.03 respectively in the street category.The test results are shown(Fig.6).Conclusions This article proposes a classification algorithm based on the physical and geometric characteristics of the target.This method focuses on the rich physical characteristics and vertical structural information of the target contained in the full-waveform,as well as the correlation between full-waveforms and geometric features of point clouds.The fusion of the two features is achieved using a dual low-rank matrix;And based on the local neighborhood fully-connected structure,the learning of local structural information are strengthened,thus constructing an end-to-end full-waveform airborne LiDAR data classification network.Multiple experiments show that the average accuracy and recall of our classification method are as high as 0.96 and 0.90,respectively,indicating the effectiveness of this method.At the same time,it provides some possibilities for exploration indepth feature extraction and fusion,based on the correlation in target physical characteristics between fullwaveforms and point clouds.