Research On Hyperspectral Data Classification Algorithm Based On Hybrid Convolutional Neural Network

With the continuous breakthrough of remote sensing technology,hyperspectral data classification technology has been gradually applied in various fields.For example,urban planning and construction,surface material analysis,complex terrain investigation,etc.In the development of remote sensing,hyperspectral data is an important branch of remote sensing.In particular,the classification technology for hyperspectral data is one of the more concerned technical directions in the industry.The main technical route for classifying hyperspectral data is to classify the pixels in hyperspectral data according to certain rules or algorithms according to the spectral chromaticity,spatial position arrangement or complementary information between different bands.However,due to the three-dimensional characteristics of hyperspectral data,scholars have carried out extensive research on its data processing and classification.Based on the characteristics of hyperspectral data structure,the main research contents of this paper are as follows:1.Firstly,according to the difference between hyperspectral data and traditional image data information,the corresponding classification methods are proposed for the problems of high redundancy and limited number of samples.Hyperspectral data not only contains the basic spatial information in the image,but also contains the unique spectral dimension information of hyperspectral data,so it is a three-dimensional cube data.Most of the previous methods use spectral joint spatial information to learn their image characteristics,so this method mostly uses three-dimensional convolutional networks to design classification structures.However,the three-dimensional convolution network has a large number of parameters,and the model complexity is often high when the relatively better accuracy effect is obtained.The two-dimensional convolutional network often learns feature information at the spatial level,so it is more suitable for processing spatial characteristics.Based on this,in order to fully learn the hyperspectral three-dimensional characteristic information to improve the accuracy of image classification and reduce the complexity of the model,this paper proposes a hyperspectral data classification algorithm based on single-branch multi-level hybrid attention network(DMCN).Firstly,a dense three-dimensional convolution structure is designed to efficiently learn spectral-spatial joint information.Secondly,a multi-level grouping residual two-dimensional convolutional network is designed to learn the spatial feature distribution.Finally,the collaborative attention mechanism is used to perceive the detailed features,enhance the joint information learning of the three-dimensional spatial spectrum and pay attention to the spatial feature position,so as to achieve the purpose of accurate classification.Comparative experiments were conducted on three public datasets,namely Indian Pines(IP),University of Pavia(UP),Salinas(SA).The results show that the model effectively improves the classification accuracy for all three datasets.2.Secondly,it is difficult to obtain hyperspectral data,which requires a large number of manual annotations for data expansion,and there may be mislabeling in the data samples.At the same time,in the process of data collection,external factors will lead to certain labeling category errors in the data.Based on this,in order to obtain a relatively good classification effect in the interference environment and improve the stability of the classification algorithm.In this paper,a hyperspectral data classification algorithm(DDAN)based on noise label double branch hybrid attention network is studied.Firstly,the algorithm designs two branches,which are spectral information feature extraction branch and spatial information feature extraction branch.In each branch,the three-dimensional dense convolution network structure is used to extract spectral information and spatial information respectively,and the information of different dimensions of the image is learned through two branches.Secondly,a two-dimensional convolution attention mechanism structure is designed to perceive the detailed information of spectrum and space,so as to achieve the purpose of perceiving and classifying things.Finally,the anti-noise loss function is used to make the model converge better and enhance the robustness of the model.In order to verify the effectiveness of the model,comparative experiments were conducted on three public datasets,namely Kennedy Space Center(KSC),University of Pavia(UP)and Salinas(SA).The results show that DDAN has better classification accuracy on KSC,UP and SA.