遥感影像与LiDAR点云多尺度深度特征融合的地物分类方法  

A Method of Ground Object Classification Based on Multi-scale Deep Feature Fusion of Remote Sensing Image and LiDAR Point Cloud

作　　者：王朝莹 邢帅[1] 戴莫凡 WANG Zhaoying;XING Shuai;DAI Mofan(Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《测绘科学技术学报》2021年第6期604-610,617,共8页Journal of Geomatics Science and Technology

基　　金：国家自然科学基金项目(41876105,41371436)。

摘　　要：针对基于深度学习的多源数据融合地物分类方法存在的特征考虑不全面、多尺度结构不明显等问题,提出一种遥感影像与LiDAR点云多尺度深度特征融合MSDFF的地物分类方法。首先通过Vgg16网络分别提取影像和点云数据3个不同分辨率的深度特征,利用特征连接(Concat)与1×1卷积相结合的方式实现两种数据相同分辨率特征的融合;其次借助反卷积完成了多尺度融合特征之间的交互,还引入了迁移学习与加权交叉熵损失函数以进一步提高分类精度;最后利用ISPRS提供的Vaihingen数据集进行了地物分类实验。实验结果验证了该方法的正确性和有效性,相较于基于单源数据和浅层特征融合的地物分类方法,树木、建筑物和地面的分类精度均有提高,其树木的分类精度提高了约10%。Aimed at the problems of incomplete feature consideration and inconspicuous multi-scale structure in the ground object classification methods of multi-source data fusion and deep learning, a classification method with multi-scale depth feature fusion of remote sensing images and LiDAR point clouds named MSDFF is proposed in the paper. Firstly, the depth features of three different resolutions of image and point cloud data by Vgg16 network are extracted, and the fusion of the same resolution features of both data is achieved by using the combination of feature concatenation(Concat) and 1×1 convolution. Secondly, the interaction between the multi-scale fused features is completed with the help of deconvolution, and migration learning and weighted cross-entropy loss function are also introduced to further improve classification accuracy. Finally, a classification experiment is conducted using the Vaihingen dataset provided by ISPRS. The result shows the correctness and effectiveness of the method, and the classification accuracy of trees, buildings and ground are improved compared with those of the method basing on single-source data and shallow feature fusion, in which the classification accuracy of trees is improved by 10%.

关 键 词：遥感影像 LIDAR点云 多源数据融合 地物分类 全卷积神经网络 深度特征 

分 类 号：P237[天文地球—摄影测量与遥感]