Hyperspectral Images Classification With Attention Based CNN And Sparse ELM

With the development of hyperspectral technology,the dimension of hyperspectral remote sensing image is increasing.How to use the rich spectral information and spatial information of hyperspectral image to classify objects is one of the current research hotspots.Hyperspectral image data is characterized by multi-dimensionality,correlation,nonlinearity,and large amount of data.Therefore,it is particularly important to extract deeper features in hyperspectral images.The spatial resolution of hyperspectral images is continuously increasing,which makes feature extraction and classification of hyperspectral images using spatial information has become an important part of hyperspectral image analysis and processing.As an important model in deep learning,the current deep convolutional neural network is widely used in the field of visual image with its natural advantages in image spatial information learning.On the existing hyperspectral image processing methods based on convolution neural network,this paper deeply studies the learning methods of local key features in hyperspectral images,which is used to enhance the learning of local features in hyperspectral images,and how to improve the performance of pixel classification of hyperspectral images is also studied.Therefore,the research work on local feature extraction and pixel classification of hyperspectral images is summarized as follows:(1)Aiming at the problem of information redundancy in original hyperspectral image,a feature fusion method based on PCA principal component analysis and guided filter is proposed to extract the spatial-spectral features from original hyperspectral images and remove noise.The experimental results show that the proposed method can effectively reduce the redundancy of the original hyperspectral images and extract spatialspectral features.(2)Aiming at the problem that it is difficult to extract the local features of hyperspectral images by traditional convolutional neural network,a feature extraction method based on spatial-spectral attention is proposed,which enhances the learning of local key features in spatial and spectral domains of hyperspectral images.Among them,the input of the attention model is the low-level features of convolutional neural network,and the output of the attention model is used to extract the high-level features.The experimental results show that the proposed method can effectively learn the local key features of hyperspectral images and achieve better results than traditional convolutional neural networks.(3)Aiming at the difficulty of hyperspectral image pixel classification,a multiobjective optimization based sparse extreme learning machine method for pixel classification is proposed.In order to solve the problems of information redundancy and overfitting of the algorithm,this method defines the sparse connecting structure of extreme learning machine,and builds a multiobjective model to simultaneously optimize the network parameters and the sparse connecting structure of sparse extreme learning machine.Finally,it makes decision on the multiobjective optimal model by using ensemble learning.Experimental results show that the proposed sparse extreme learning machine is feasible and effectively improves the pixel classification performance of hyperspectral images.In summary,this paper analyses the shortcomings of hyperspectral image feature extraction method and extreme learning machine classification method,and then studies the feature extraction algorithm based on attention mechanism and sparse extreme learning machine classification algorithm.Several experimental results demonstrate the effectiveness of the proposed algorithm.