基于改进的PointNet++模型的多光谱LiDAR数据分类方法

Multispectral LiDAR Data Classification Method Based on an Improved PointNet++ Model

摘要: 多光谱激光雷达(LiDAR)系统可同时并快速获取大范围空间目标地物的光谱强度信息和空间几何信息，为三维点云分类、语义分割、目标检测等研究提供新的数据源。然而，由于多光谱点云数据分布的不规则性以及数据量巨大等特性，使得地物特征的提取过程充满挑战。本文通过将通道注意力机制(SE-Block)和修正后的焦点损失函数嵌入至PointNet++网络中，提出了一种改进的PointNet++网络架构。PointNet++网络从不均匀采样的点中提取局部特征，并通过多尺度分组表示点之间的局部几何关系。将SE-Block嵌入至PointNet++网络中，通过显式地建模通道之间的相互依赖关系，自适应地重新校准通道方面的特征响应，从而强调重要通道并抑制不利于预测的无用通道，提高特征的显著性，以便更好地进行点云分类。另外，本文在改进的网络架构基础上利用修正后的焦点损失函数解决了多光谱LiDAR点云数据中类别不均匀分布的问题。本文提出的改进的PointNet++网络架构在托伯莫里港口数据集上进行了评估，获得的总体精度、mIoU、F1-score和Kappa系数分别为95.21%、62.59%、73.58%、0.918。与5个已建立的深度神经网络模型的比较实验证实，本文提出的改进的PointNet++网络架构在多光谱LiDAR点云分类任务中具有良好的性能。

Abstract: A multispectral light detection and ranging (LiDAR) system can simultaneously and quickly collect spectral intensity information and spatial geometric data of a large range of space objects, which provides a new data source for the research of 3D point cloud classification, semantic segmentation and object detection. However, due to the irregularly distributed property of multispectral point cloud data and massive data volume, the extraction process of land cover is full of challenges. In this paper, we propose an improved PointNet++ network architecture by embedding the Squeeze and Excitation Bock (SE-Block) and a modified focal loss function into the PointNet++ network. Point-Net++ network extracts local features from unevenly sampled points and represents local geomet-rical relationships among the points through multi-scale grouping. SE-Block is embedded into PointNet++ network, which explicitly models the interdependence between channels and adap-tively recalibrates the feature response in terms of channels, emphasizing important channels and suppressing useless channels that are not conducive to prediction, improving the saliency of fea-tures for better point cloud classification. In addition, based on the improved network architecture, this article utilizes the modified focal loss function to solve the problem of uneven distribution of categories in multispectral LiDAR point cloud data. The improved PointNet++ network architecture proposed in this paper has been evaluated on the Tobermory Port dataset and achieved an overall accuracy, a mean Intersection over Union (mIoU), an F1-score, and a Kappa coefficient of 95.21%, 62.59%, 73.58% and 0.918, respectively. Comparative studies with five established deep neural network models confirm that the improved PointNet++ network architecture proposed in this pa-per has good performance in multispectral LiDAR point cloud classification tasks.