A Point-Based Fully Convolutional Neural Network for Airborne LiDAR Ground Point Filtering in Forested Environments
A Point-Based Fully Convolutional Neural Network for Airborne LiDAR Ground Point Filtering in Forested Environments
复制标题
用于森林环境中机载激光雷达地面点过滤的基于点的全卷积神经网络
DOI：
10.1109/jstars.2020.3008477
复制
发表时间：
2020-01-01
影响因子：
5.5
通讯作者：
Guo, Qinghua
中科院分区：
文献类型：
--
作者：
Jin, Shichao;Sun, Yanjun;Guo, Qinghua
机载激光扫描（ALS）数据是地形产品生成中最常用的数据之一。滤波接地点是ALS数据处理的前提步骤。传统滤波方法主要采用手工特征或预定义分类规则，并进行预处理/后处理操作，迭代滤波接地点，具有经验性和繁琐性。深度学习以其自学习特征的能力为解决分类和分割问题提供了一种新的方法，在许多领域，特别是遥感领域受到青睐。本文提出了一种基于点的全卷积神经网络（PFCN），该网络直接消耗仅具有几何信息的点，并同时提取逐点和逐块特征对每个点进行分类。对网络进行了来自14个站点的37449157个点的训练，并对不同森林环境下的6个站点进行了评价。此外，将该方法与五种广泛使用的过滤方法和最佳的基于点的深度学习方法（PointNet++）进行了比较。结果表明，PFCN在平均遗漏误差（T1 = 1.10%）、总误差（Te = 1.73%）和Kappa系数（93.88%）方面取得了最好的结果，但在由最严重的委托误差引起的数字地形模型均方根误差方面排名第二。此外，我们的方法在准确性上与PointNet++相当甚至更好。此外，该方法只消耗1 / 3的计算资源和1 / 7的训练时间。我们认为PFCN是一种简单灵活的方法，可以广泛应用于地点滤波。
Airborne laser scanning (ALS) data is one of the most commonly used data for terrain products generation. Filtering ground points is a prerequisite step for ALS data processing. Traditional filtering methods mainly use handcrafted features or predefined classification rules with preprocessing/post-processing operations to filter ground points iteratively, which is empirical and cumbersome. Deep learning provides a new approach to solve classification and segmentation problems because of its ability to self-learn features, which has been favored in many fields, particularly remote sensing. In this article, we proposed a point-based fully convolutional neural network (PFCN) which directly consumed points with only geometric information and extracted both point-wise and tile-wise features to classify each point. The network was trained with 37449157 points from 14 sites and evaluated on 6 sites in various forested environments. Additionally, the method was compared with five widely used filtering methods and one of the best point-based deep learning methods (PointNet++). Results showed that the PFCN achieved the best results in terms of mean omission error (T1 = 1.10%), total error (Te = 1.73%), and Kappa coefficient (93.88%), but ranked second for the root mean square error of the digital Terrain model caused by the worst commission error. Additionally, our method was on par with or even better than PointNet++ in accuracy. Moreover, the method consumes one-third of the computational resource and one-seventh of the training time. We believe that PFCN is a simple and flexible method that can be widely applied for ground point filtering.