Deep Capsule Network Classification For Fused LiDAR And Hyperspectral Data

In recent years,due to the continuous innovation and development of remote sensing sensor technology,earth observations through space or aviation have become more and more comprehensive.Among them,the application of the most representative hyperspectral passive remote sensing technology and lidar active remote sensing technology has become important.Research hotspots.Among them,hyperspectral data has a wealth of image information and spectral information,which can not only describe the appearance attributes of features,but also characterize the physical attributes of features.Due to the large number of dimensions,large amount of data,and complex data structure of hyperspectral data,it is very important for classification networks.High requirements are put forward,and it is difficult to achieve accurate feature classification using only a single hyperspectral data.Lidar data not only contains accurate spatial information of features,but also has height information describing features,but it is also difficult to accurately classify features with similar geometric structures.Therefore,through the fusion of hyperspectral and lidar data,the advantages of the two data sources can be complemented and information sharing can be achieved,forming a unique feature space,which can provide a data basis for effective ground object classification,and is conducive to the improvement of classification accuracy.In view of the respective imaging advantages of hyperspectral and LiDAR,this paper explores the application of these two remote sensing data sources in the classification of urban features by constructing a Double Channel Deep Capsule Network(DCDCN).Through the analysis of hyperspectral data,a two-layer three-dimensional convolution structure is used to extract the image information and spectral information of the hyperspectral data,and a two-layer two-dimensional convolution structure is specifically used to extract the elevation information of the lidar data to realize different data sources.The classification and hierarchical learning of the feature information are finally fused and classified by using the extracted feature information based on the capsule network that can learn the relative position relationship of the feature space.Experiments show that the performance is greatly improved by performing feature learning adapted to the characteristics of the data on each branch of the two data sources and then fusing classification,and the classification accuracy is improved about 2% compared with using a single data source.Aiming at the squash nonlinear activation function between the capsule network layers,this paper proposes a nonlinear activation function called e-squash for feature compression learning.Experiments were performed on three urban feature datasets containing hyperspectral and LiDAR data.The results showed that when the squash activation was changed to the e-squash activation function,DCDCN improved the classification accuracy by more than 1% on each dataset,especially in the combined use of hyperspectral data and LiDAR elevation feature data fusion classification,the classification accuracy is further improved.It shows that the DCDCN network that uses the activation function proposed in this paper has greater application potential in urban data classification than the classic classification method and the capsule network that does not use e-squash.