A Hyperspectral Image Classification Algorithm And System Implementation Combining Artificial And Deep Spatial Spectral Features

As a basic research problem in the field of hyperspectral remote sensing image analysis,hyperspectral image classification has received a lot of attention in academia and even industry.Researchers at home and abroad have proposed numerous excellent classification algorithms.However,due to the complexity of the spectral and spatial feature of hyperspectral images,there are still many problems to be solved in hyperspectral image classification,which is still an extremely difficult research topic.In this paper,we carry on an investigation into hyperspectral image classification algorithms based on deep learning and propose two classification models based on deep feature extraction.At the same time,we use appropriate artificial methods to optimize the extracted features,Finally,the effectiveness of the algorithms is verified by experiments.The innovations of this article can be listed as follows:(1)This dissertation proposes a hybrid convolutional neural network with anisotropic diffusion for hyperspectral image classification.Due to the characteristics of hyperspectral image cube data,we use two-dimensional and three-dimensional mixed convolutional layers to extract features and take advantage of 2D and 3D convolution in extracting features of hyperspectral images.To solve the problem that convolution operations with different dimensions in deep neural networks usually adopt an isotropic template structure,we put forward the use of anisotropic diffusion for feature propagation to extract deep features inanisotropic neighborhoods and propagate the intrinsical discriminative features in each class.Experiments on three datasets show that this method can propagate the deep features in anisotropic neighborhoods,eliminate the classification mistakes of traditional convolutional neural networks with a small amount of training data and maintain the boundary of different types of features better.(2)This dissertation proposes a multi-scale dense spectral–spatial convolution network for hyperspectral images classification.The algorithm introduces a densely connected structure to deepen the network,and uses spectral dense blocks to learn the spectral characteristics of hyperspectral images.And considering the multi-scale characteristics of spatial features of hyperspectral images,we use a two-dimensional dense structure with horizontal and vertical connections to realize the learning of spatial features.The comprehensive evaluation results on three datasets verify the effectiveness of the method and the ability of small sample classification.(3)This dissertation designs a hyperspectral image classification method based on artificial features and deep features.In order to adress the problems of low classification accuracy after the fusion of multiple artificial features and the proneness to overfitting during the training of multi-channel deep models.We introduce appropriate artificial features into the deep model.After that,the number of multi-channel deep model parameters is reduced and overfitting is suppressed.Experiments on three datasets show that by introducing appropriate artificial features in the deep model,the scale of parameters in the deep model can be reduced,overfitting during the training process can be suppressed,and classification accuracy can be improved.(4)This dissertation designes a hyperspectral image classification system.The system inherits the eight algorithms and three artificial feature extraction algorithms such as the above algorithms and comparison methods;and has designed three core modules: the hyperspectral image visualization module,the hyperspectral image classification module,and the classification result evaluation module.A number of functions including the visualization of original images,the setting of algorithm parameters,and the display of classification results have been implemented.