Research On Band Selection And Classification Method For Hyperspectral Remote Sensing Image

Hyperspectral remote sensing image has dozens or even hundreds of continuous spectral channels and abundant spatial structure information,but the spectral information has high redundancy,and the discriminative features is extracted difficultly.To solve these two problems,this paper focuses on the band selection method and the method of using the deep network to learn the discriminative feature for classification.The main works are as follows:(1)For the problem of high redundancy of spectral information,a clustering algorithm based on local potential is proposed to remove the redundant information in the original bands and select the more representative bands.The local potential idea of the band considers the difference between bands to better describe the bands distribution,and a factor to evaluate the information of the band is introduced in the proposed new ranking rule which can help the algorithm select more representative bands,especially some boundary bands with abundant information can be found.The effectiveness of the algorithm is proved on three different data sets.(2)The previous deep networks based on spectral-spatial fusion less consider the spectral-spatial information fusion of different neighborhoods,and the classification accuracy of the network will decline with deeper layers.For this problem,a deep residual network based on multiscale spectral-spatial fusion is proposed.The network uses the idea of multiscale inputs and multiple residual blocks to fuse the spectral-spatial information of multiple neighborhoods to extract the discriminative features with spectral-spatial information of different neighborhoods.The idea of identity mapping in residual blocks ensures the network maintain high classification accuracy in deeper network.On three different data sets,the classification results of the network are better than other competitors.(3)To solve the problem that the previous deep networks proposed less consider the strong complementary and related information among different scale features,this paper proposes a network based on multiscale deep middle-level feature fusion.The network fully fuses the strong complementary and related information between different scale features,and extracts the multiscale middle-level features.The multiscale middle-level features are fused in the convolutional layer,and the subsequent residual blocks can extract more discriminative features to classify.On four different data sets,especially in the case of limited training samples,the superiority of this network is proved.