Study On 3D Residual Network Classification Of Hyperspectral Image Fused With Lidar Data

Hyperspectral Imaging(HSI)data can capture rich physical and chemical information based on continuous spectral bands.Even if the difference between substances is small,the identification of different types of data can be realized based on Hyperspectral Imaging.However,the problems of shadow effect,spectral variability,mixed pixels and multiple spectral bands in hyperspectral images limit the accurate interpretation of hyperspectral images.Although the current classification of hyperspectral images based on deep learning classification networks has achieved considerable results,some networks still have problems such as large number of parameters,network degradation and low boundary coincidence degree.Aiming at the above problems,this paper combines LIDAR data and deep residual network to deal with them.The main research of this paper includes the following two points.(1)Aiming at more spectral bands in hyperspectral image,and the problem of spectral variation in this paper a kind of main convex hull analysis data for band selection process,band selection method in the band selection while retain the original spectral data,in order to realize the important spectral characteristics of the mitigation of contracted and spectral variability,and ease the hyperspectral image and mixed pixels is the variation problem.(2)Aiming at the problems of shadow effect in hyperspectral images,network degradation and low boundary coincidence in common deep classification networks,this paper proposes a three-dimensional residual network classification algorithm for hyperspectral images that fuses LIDAR data.In this study,Li DAR data was used to make up for the lack of information in hyperspectral images,and the data were processed based on 3D residual network.The spectral residual module and the spatial residual module in the network realize the feature learning of spectral data and spatial data,and finally reestablish the similarity measurement rules by combining spatial features and spectral features to classify pixels and improve the classification accuracy.Experimental results show that the research of hyperspectral images band selection algorithm and based on the three dimensional residual network classification of hyperspectral image fusion algorithm in three public hyperspectral image data set has higher classification accuracy and classification of image boundary contour more accurately,feature recognition rate is higher,at the same time can reduce the shadow effect in hyperspectral image and spectral variability and solve network degradation phenomenon.