Hyperspectral Image Classification Via Quaternion Convolution Neural Networks

Hyperspectral images（HSIs）contain the rich spatial and spectral information,whose classification has been widely used in geological survey,precision agriculture,marine protection,and other fields.Considering the complexity of the types of land covers and their distribution,it is significantly difficult to meet the demand of actual applications just using a single HSI data.Different from HSI data,light detection and ranging（Li DAR）data consists of the detailed elevation information,which can effectively improve the characterization of HSI scenes.However,simply stacking or cascading data and their features fails to model their correlations,resulting in the loss of information.Therefore,it is a hot issue to design an effective algorithm,a method to effectively use the spatial information of Li DAR data to assist HSI data,improving the HSI classification accuracy.To this end,using the latest research progresses in the theories of quaternion,deep learning,and attention mechanism,two novel deep quaternion convolution neural networks are designed for HSI classification,which mainly include the following two aspects:In order to make full use of the correlations between different features,the idea of the quaternion neural networks is introduced into HSI classification,thus forming a deep classification framework（quaternion convolutional neural network,QCNet）,which includes two models（i.e.,2DQCNet and 3DQCNet）.Firstly,the extended morphological attribute profiles algorithm is applied to process the original data,and the morphological features of different attributes of Li DAR and HSI data are extracted respectively,which are further extended to the quaternion domain.Subsequently,the extracted features are fed into the2 DQCNet and 3DQCNet models to perform the end-to-end classification.The experimental results on two remote sensing data sets verify the excellent classification performance of the proposed QCNet framework.Aiming at the lack of elevation information in HSI data,a dual channel QCNet（DualChannel QCNet）model is proposed to learn the complementary information of the HSI data and Li DAR data for better classification accuracy.The proposed model includes Li DAR channel and HSI channel.Firstly,the 2DQCNet and 3DQCNet models are applied to learn the spatial-spectral representation of HSI data and the elevation information of Li DAR data.Then,to effectively utilize the complementary information provided by Li DAR data,a quaternion spatial attention block（QSAB）and a fusion strategy are further designed,thus improving the classification performance of HSIs.Experiments conducted on two multisource remote sensing data sets show that the proposed model can be superior to other deep networks.