Research On Hyperspectral Image Classification Based On Space Spectral Union And Novel Activation Function

In recent years,with the continuous development of remote sensing image processing technology,the application of hyperspectral image classification in environmental monitoring,national defense and military,urban planning and agriculture,forestry and agriculture has received extensive attention.Domestic and foreign researchers have also successively put forward some excellent classification methods.However,due to the special high-dimensionality of hyperspectral images,strong correlation between adjacent bands,highly nonlinear data structure,and few training samples,there are still many difficulties in the classification of hyperspectral images.This paper proposes a combined hyperspectral image classification method based on space-spectral features to solve the problem that only using spectral features for classification of hyperspectral images is easy to cause Hughes phenomenon;a new type of activation function for hyperspectral image classification is proposed to solve the problem that activation functions in specific tasks are often limited by models,datasets and coefficient factors.Experiments are conducted on three of the most widely used hyperspectral image datasets,and the experimental results demonstrate the superiority of the proposed method and activation function.The innovations of this paper are as follows:(1)A hyperspectral image classification method combining spatial feature information and spectral feature information is proposed.Aiming at the high-dimensional data characteristics of hyperspectral images,this method uses principal component analysis and linear discriminant analysis to preprocess the input image data for dimensionality reduction,and projects the high-dimensional data into the low-dimensional feature space,so that the data after dimensionality reduction is not only The main information is retained,and the distance ratio between classes and within classes is increased;Aiming at the extraction of spatial features and spectral features and the combination of features,this method combines two-dimensional Gabor filter with random block convolution feature extraction,not only using the feature extraction ability of random block convolution and the advantages of Gabor filter,Multi-layer fusion of feature maps is also implemented,giving the classification network the advantage of multi-scale.Through comparative experiments on three datasets,it is proved that this method can not only improve the classification accuracy of multiple types of ground objects,but also has robustness in experiments with limited boundaries of different types of ground objects and limited training samples.(2)A piecewise combinatorial modified linear unit activation function for hyperspectral image classification is proposed.The activation function has the non-monotonicity of the negative semi-axis,the nonlinearity of the positive semi-axis and the sparsity at the same time through the combination of pieces.The negative semi-axis approaches zero as the negative input value decreases,introducing negative activation values and non-zero derivative values for the negative semi-axis while maintaining the sparsity of the negative input values;the positive semi-axis does not use an identity map,but with An increase in the input value,closer to the identity map,brings nonlinearity to positive input values.Finally,a comparative experiment of ground object classification is carried out on three datasets using convolutional neural network models with different structures,and the experimental results verify the superiority of the activation function.(3)Based on the hyperspectral image classification algorithm based on the above-mentioned combination of spatial spectral features,a classification system applied to hyperspectral images is implemented.The system is mainly composed of modules such as image reading,visualization and classification of hyperspectral images,performance evaluation,preservation of classification results,and historical viewing.The selected hyperspectral images can be directly classified,and the classification result graph and classification accuracy can be displayed intuitively after classification,which promotes the development of the hyperspectral image classification system.