Hyperspectral image feature extraction selection method based on adaptive multi-modal spectral embedding

Abstract

The application discloses a hyperspectral image feature extraction selection method based on adaptive multi-mode spectrum embedding and relates to the technical field of hyperspectral image feature mining, and comprises the following steps: constructing an adaptive multi-mode spectrum embedding algorithm; constructing an adaptive multi-mode feature extraction selection algorithm based on the adaptive multi-mode spectrum embedding algorithm; training the adaptive multi-mode feature extraction selection algorithm by using a hyperspectral training data set; and performing actual hyperspectral image feature extraction selection by using the trained adaptive multi-mode feature extraction selection algorithm.The adaptive multi-mode feature extraction selection algorithm can simultaneously perform feature extraction and feature selection, adaptively assign weights, only use a part of features in the feature extraction process, and the projection subspace has clear physical interpretability, so that the effectiveness of feature extraction selection is improved, and the subsequent classification precision is improved.

Claims

1. A hyperspectral image feature extraction and selection method based on adaptive multi-mode spectral embedding, characterized in that, Includes the following steps: Step 1: Construct an adaptive multi-mode spectral embedding algorithm; Step 2: Construct an adaptive multi-mode feature extraction and selection algorithm based on the adaptive multi-mode spectral embedding algorithm; Step 3: Train the adaptive multi-mode feature extraction and selection algorithm using a hyperspectral training dataset; Step four: Use the trained adaptive multi-mode feature extraction and selection algorithm to perform actual hyperspectral image feature extraction and selection; In step two, the objective function for constructing the adaptive multi-modal feature extraction and selection algorithm is: ， in, Indicates inclusion Multimode data matrix of a hyperspectral image dataset of samples , , Represents dimension. , This indicates the number of patterns in multimodal data. Represents the characteristic transformation matrix. Indicates the low-dimensional embedding dimension. Indicates the first The Laplace matrix of the pattern data, It is a regular expression parameter. Represents a matrix beg Norm, It is a matrix row vectors Norm; The adaptive multimodal feature extraction and selection algorithm is solved by alternately optimizing the weight coefficients and the feature transformation matrix; The solution steps of the adaptive multimodal feature extraction and selection algorithm include: Step S21, input the multi-mode data matrix , , , The spatial embedding dimension of the algorithm is the low-dimensional embedding dimension. The nearest neighbor number of the weight matrix is Initialize the weight coefficients for each mode. ,initialization It is a random matrix; Step S22: Calculate the Laplacian matrix of the multi-model data. and diagonal matrix ; Step S23: Solve the generalized eigenvalue problem ,in, Represents the eigenvector. Representing eigenvalues, derived from the previous... Minimum eigenvalues corresponding feature vector Constructing the characteristic transformation matrix ; Step S24: Determine whether the objective function of the adaptive multi-mode feature extraction and selection algorithm has converged. If it has not converged, update the weight coefficients of each mode. Then return to step S22. If convergence is achieved, output the low-dimensional feature data. .

2. The hyperspectral image feature extraction and selection method based on adaptive multi-mode spectral embedding according to claim 1, characterized in that, In step one, the objective function for constructing the adaptive multi-mode spectral embedding algorithm is: ， in, Indicates inclusion Multimode data matrix of a hyperspectral image dataset of samples , , Represents dimension. , This indicates the number of patterns in multimodal data. Represents the characteristic transformation matrix. Indicates the low-dimensional embedding dimension. Indicates the first The Laplace matrix of the pattern data.

3. The hyperspectral image feature extraction and selection method based on adaptive multi-mode spectral embedding according to claim 2, characterized in that, The adaptive multi-mode spectral embedding algorithm is solved by alternately optimizing the weight coefficients and the feature transformation matrix.

4. The hyperspectral image feature extraction and selection method based on adaptive multimodal spectral embedding according to claim 2, characterized in that, The solution steps of the adaptive multi-mode spectral embedding algorithm include: Step S11, Input the multi-mode data matrix , , , The spatial embedding dimension of the algorithm is the low-dimensional embedding dimension. The nearest neighbor number of the weight matrix is Initialize the weight coefficients for each mode. ; Step S12: Calculate the Laplacian matrix of the multi-model data. ; Step S13: Solve the generalized eigenvalue problem ,in, Represents the eigenvector. Representing eigenvalues, derived from the previous... Minimum eigenvalues corresponding feature vector Constructing the characteristic transformation matrix ; Step S14: Determine whether the objective function of the adaptive multi-mode spectral embedding algorithm has converged. If it has not converged, update the weight coefficients of each mode. Then return to step S12. If convergence is achieved, output the low-dimensional feature data. .

5. The hyperspectral image feature extraction and selection method based on adaptive multimodal spectral embedding according to claim 1, characterized in that, In step three, the low-dimensional feature data extracted and selected by the adaptive multi-modal feature extraction and selection algorithm are classified, and the overall accuracy, average accuracy and Kappa coefficient are used as evaluation indicators to judge the performance of the adaptive multi-modal feature extraction and selection algorithm based on the classification accuracy.

Images