基于3D深度胶囊网络的高光谱和LiDAR数据融合分类

针对高光谱和LiDAR成像优势,通过构建三维深度胶囊网络(3D-deep capsule network,3D-DCN)探索了这两种遥感数据源在城市地物分类上的应用.该网络通过使用两层3D-CNN结构实现融合后数据的非线性特征映射,然后紧跟胶囊网络生成代表特征的矢量并实现卷积、封装和分类;针对胶囊网络层间的非线性激活函数提出一种称为e-squash的非线性激活函数用于特征学习.在城市数据集上的分类实验表明,使用LiDAR高程特征数据极大地改善了高光谱图像分类精度,采用提出激活函数的3D-DCN在城市数据分类方面比经典分类方法和未采用e-squash的胶囊网络具有更大的应用潜力.

3D-deep capsule network classification for fused hyperspectral and LiDAR data

As the advantages of both hyperspectral and LiDAR imaging, this paper explored the two source remote sensing data for classification of a urban area by deploying a 3D-deep capsule network(3D-DCN). 3D-DCN used a two-layer 3D-CNN structure at the beginning to realize non-linear feature mapping for the fused data. And it followed a capsule network to gene-rate vectors to represent features and realize encapsulation, convolution and classification. In addition, this paper proposed a nonlinear activation function named e-squash function to be used in the capsule convolution layer. The classification experiments on the urban datasets show that the LiDAR data can improve greatly the hyperspectral image classification result and the 3D-DCN with proposed activation function is of more potential than the state-in-art methods in the classification of the urban data.