Research On Collaborative Classification Of Hyperspectral And LiDAR Data

Hyperspectral and LiDAR remote sensing technology is the main technology to obtain surface information.With the rapid development of earth detection technology and aerospace technology,people have been able to obtain multi-source datasets about the same environment in real time,which has created a good technical foundation for the collaborative classification of hyperspectral and LiDAR data.Hyperspectral image can better reflect the spectral information of ground objects,showing their structure,texture and other characteristics,while LiDAR can efficiently and accurately obtain the elevation information of the ground.According to the difference characteristics of multi-source remote sensing data,combined with the diversity of data information,using multi-source remote sensing data to classify ground objects has become a new trend in remote sensing data classification research.Collaborative classification can effectively combine the complementary information of data,and also provides a new solution and effective means to solve the problem of surface feature classification.In order to combine the characteristics and advantages of two kinds of remote sensing data and improve the classification accuracy of ground objects,this paper proposes two cooperative classification methods of hyperspectral and LiDAR data.The main research contents include the following parts:First,the characteristics of hyperspectral and LiDAR data are analyzed,and the theoretical parts of depth separable convolution,cavity convolution and multi-head self-attention mechanism are analyzed in depth,laying the foundation for the subsequent chapters of this paper.Secondly,in order to fully exploit the characteristics of hyperspectral and LiDAR data,a collaborative classification method of hyperspectral and LiDAR data based on pyramid pooling in empty space and multi head attention mechanism is designed.The two branches use symmetrical feature extraction modules and pyramid pooling in empty space to enhance the discrimination of learning features.The separable convolution sharing mechanism is improved to extract hyperspectral and LiDAR features respectively,and at the same time,features are exchanged and shared with each other.Put the extracted features into the multi-head self-attention mechanism,further enhance the discrimination ability of the learning features through the feature fusion mechanism,make full use of the context information of the data and the feature information between the hyperspectral and LiDAR data,obtain a more comprehensive feature map,and improve the accuracy of ground object classification.Finally,the cross modal spatial enhancement module and spectral enhancement module are used to enhance the complementary attributes between different modes.The classification experiments were conducted on Houston2013 and Trento datasets respectively.Compared with the latest comparative experiment method,this method has better classification effect and improves the accuracy of ground object classification.