Research On Hyperspectral Image Classification Based On Deep Feature Representation Learning

Hyperspectral image(HSI)classification is one of the most concerned research topics in geoscience and remote sensing imagery processing tasks.HSI involves hundreds of different spectral bands,which has more advantages in spectral information richness than traditional panchromatic and multi-spectral remote sensing images.However,hyperspectral remote sensing images have the characteristics of huge data,nonlinearity,and high complexity,resulting in great challenges to HSI classification.Recently,deep learning-based methods,especially neural networks,have performed well on HSI classification tasks,which not only possess powerful feature representation ability,but also have favorable universality.Therefore,the thesis conducts in-depth research based on neural networks from the following three aspects: how to make full use of spectral and spatial features in HSI,how to effectively avoid the problem of grid effect while increasing the receptive field of the model,and how to deeply integrate local and global features in HSI.The main contributions of the research are as follows:(1)A HSI classification method based on residual dense asymmetric convolutional neural network is proposed in this thesis.The algorithm firstly combines residual learning and dense connections to effectively reuse the spectral-spatial information of the previous layers,thus achieving discriminative feature extraction.Secondly,standard convolutional layers are replaced with asymmetric convolutional layers to reduce network parameters and enhance the diversity of convolution kernels in the model.In addition,the Mish activation function is used to replace the ReLU activation function,which makes the hyperspectral information flow more smoothly in the network,thus further enhancing the classification accuracy and generalization ability of the algorithm.Finally,a residual dense asymmetric convolutional neural network is designed by stacking multiple residual dense asymmetric convolution blocks to achieve the purpose of spectral and spatial feature extraction.Experimental results on several HSI datasets show that the algorithm can achieve competitive classification performance.(2)A HSI classification method based on densely connected pyramidal dilated convolutional network is proposed in this thesis.The algorithm can effectively alleviate the gridding effect caused by the traditional dilated convolution operation and reduce the loss of hyperspectral information that is effective for classification.Firstly,the problems of blind spots and unrecognized regions can be avoided while the receptive field is increased by properly setting the dilation factors in dilated convolutional layers and gradually increasing the width of the network.Different from the traditional method to increase the network width by increasing the channel dimension,each dilated convolutional layer contains different numbers of sub-dilated convolutional layers as the network deepens in the algorithm,thereby achieving the increase of network width.Finally,a densely connected pyramidal dilated convolutional network is designed by stacking multiple densely connected pyramidal dilated convolutional blocks to increase the receptive field of the model.Experimental results on several HSI datasets indicate that the algorithm can effectively improve the classification accuracy.(3)A HSI classification method based on convolution transformer mixer is proposed in this thesis.Recently,vision transformer with powerful global feature modeling ability has performed well on computer vision tasks.Inspired by the vision transformer,firstly,a dualbranch network structure with convolutional neural network and vision transformer is designed to extract local and global features of HSIs.Secondly,a local-global multi-head attention mechanism is constructed to achieve a deep combination of convolution operations and attention mechanism.Finally,the original multi-layer perceptron in vision transformer is replaced with a residual bottleneck block to introduce more local information.Experimental results on several HSI datasets verify that the algorithm can achieve effective performance improvement.(4)The similarity relationship between pixels is also important for hyperspectral image classification.On the basis of capturing local and global features,a hyperspectral image classification method based on deep aggregation vision transformer is designed by introducing graph convolution network,which can mine the similarity relationship between pixels in HSIs,and realize the elegant integration of three powerful paradigms,namely convolutional neural network,vision transformer,and graph convolution network.Firstly,a larger number of convolution operations are introduced into vision transformer to achieve efficient extraction of local and global features.Secondly,since the graph convolution network can effectively represent and analyze irregular data in HSIs,the algorithm adds a new branch on the transformer encoder,namely the graph convolution module,to capture the similarity relationship between pixels.Therefore,the deep aggregation vision transformer proposed in this thesis can capture local features,global features and the relationship between different pixels,which is helpful to improve the feature representation ability of the network.Experimental results on several HSI datasets confirm that the algorithm can achieve high classification accuracy.