A subspace sparse representation hyperspectral target detection method based on spectral correlation

Abstract

A spectral correlation-based subspace sparse representation hyperspectral target detection method, based on the hyperspectral image data to be measured, quantitatively estimates the correlation between spectral information to obtain a spectral correlation coefficient matrix; based on the spectral correlation coefficient matrix and the spectral data in the target spectral dictionary, the minimum variance principle and eigenvalue decomposition method are used to calculate the characteristic subspace of the target spectral dictionary, sort the obtained eigenvalues ​​from small to large, take the first few eigenvalues ​​with a proportion of p, and their corresponding eigenvectors form the projection vector of the characteristic subspace; the hyperspectral image data to be measured and the target spectral dictionary are all projected into the characteristic subspace, and a sparse representation method is used to achieve target detection. The present invention removes information redundancy between spectra while ensuring sufficient dimensions for differentiation, and can effectively extract the common features of the prior spectral information in the target spectral dictionary, more accurately and efficiently detecting and identifying targets, with superior performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of hyperspectral remote sensing target detection, and in particular relates to a subspace sparse representation hyperspectral target detection method based on spectral correlation. Background Art

[0002] Hyperspectral remote sensing is a cutting-edge remote sensing technology that simultaneously acquires hyperspectral information from two-dimensional space and one-dimensional targets. It represents a significant advancement in remote sensing. Its most significant feature is the fusion of imaging and spectral detection technologies, enabling the simultaneous acquisition of regional image information and spectral signatures. Hyperspectral imagery (HSI) contains dozens to hundreds of adjacent narrow spectral bands. Compared to traditional multispectral imagery, it can more accurately identify different materials, enabling a variety of potential applications and opening up new avenues for remote sensing.

[0003] In recent years, significant progress has been made in identifying various substances of interest using hyperspectral remote sensing technology, a technique known as hyperspectral target detection (HTD). Target detection in remote sensing seeks to identify man-made objects or low-probability materials, which often have spectral signatures contaminated by background noise. Hyperspectral imagery, with its high spectral resolution, offers powerful recognition capabilities and an advantage in detecting these objects. Numerous target detection algorithms have been proposed and applied to hyperspectral remote sensing. These algorithms extract low-probability targets with distinct spectral signatures based on the information distribution within the image. These algorithms hold significant application value and promise in numerous fields, including military reconnaissance, Earth resource surveys, environmental health monitoring, natural disaster forecasting, and atmospheric exploration.

[0004] There are many methods for target detection in hyperspectral remote sensing images, which can be mainly divided into two categories: supervised detection and unsupervised detection. Supervised detection methods are generally based on prior information and have high detection accuracy. However, the acquisition of prior information is limited by the spectral library, reflectivity, and spectral mixing phenomena, which to a certain extent restricts the application of this method in practice. Classic supervised detection methods include spectral angle matching (SAM) and matched filter (MF). Hyperspectral unsupervised detection methods do not require prior information of the target. They mainly select points with a low probability of occurrence from global or local blocks based on the statistical characteristics of the data as targets. The disadvantage is that the real scene is complex and changeable, and it is impossible to model the target background through a pure mathematical model. In practical applications, the robustness is poor. Classic unsupervised detection algorithms include RX anomaly detection algorithm and SVDD algorithm.

[0005] In recent years, with the continuous expansion and improvement of hyperspectral databases, supervised target detection algorithms based on prior spectral information of targets have received increasing attention, especially methods based on machine learning and multiple prior information. However, in hyperspectral remote sensing, the spectral curves of ground objects have strong variability and strong correlation between bands. How to effectively reduce spectral redundancy, better extract the common features of multiple prior spectral information, and detect and identify specific objects is a major challenge.

[0006] Therefore, this field needs a hyperspectral remote sensing image target detection method based on multiple prior information, which can reduce the data information redundancy caused by the correlation between spectra and fully extract the features therein to meet the needs of ground object detection and identification. Summary of the Invention

[0007] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a subspace sparse representation hyperspectral target detection method based on spectral correlation, so as to solve the problem of common feature extraction and target detection effect improvement of machine learning methods in the prior art for processing multiple prior spectral information in hyperspectral images.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is:

[0009] A subspace sparse representation hyperspectral target detection method based on spectral correlation is proposed. Based on the hyperspectral image data to be measured, the correlation between spectral information is quantitatively estimated to obtain a spectral correlation coefficient matrix. Based on the spectral correlation coefficient matrix and the spectral data in the target spectral dictionary, the minimum variance principle and the eigenvalue decomposition method are used to calculate the characteristic subspace of the target spectral dictionary. The obtained eigenvalues ​​are sorted from small to large, and the first few eigenvalues ​​with a proportion of p are taken. The corresponding eigenvectors constitute the projection vector of the characteristic subspace. The hyperspectral image data to be measured and the target spectral dictionary are all projected into the characteristic subspace, and the sparse representation method is used to achieve target detection.

[0010] In one embodiment, the method for quantitatively estimating the correlation between spectral information is:

[0011] The radiation value of the hyperspectral image to be measured is standardized, and the spectral correlation coefficient matrix of the standardized data is calculated to achieve quantitative estimation of the correlation between spectral information.

[0012] In one embodiment, the spectral correlation coefficient matrix may be calculated using a Gram matrix, or by using a Pearson correlation coefficient model, a Spearman correlation coefficient model, a mutual information model, or the like.

[0013] In one embodiment, the normalization is to divide the radiation value of each band in the spectral data by the maximum radiation value; and the same normalization operation is performed on the spectral data in the target spectral dictionary and the spectral data to be measured in the hyperspectral image data.

[0014] In one embodiment, according to the minimum variance principle and incorporating the spectral correlation coefficient matrix, the variance of the spectral information contained in the target spectral dictionary is subjected to eigenvalue decomposition to obtain eigenvalues, and the corresponding eigenvectors are taken to form the projection vector of the characteristic subspace, as follows:

[0015] Step 1: Let the target spectrum dictionary be D t , calculate the correlation coefficient matrix C based on the input hyperspectral image data;

[0016] Step 2, remember After integrating into the spectral correlation coefficient matrix, the variance of the target spectral dictionary is obtained; the eigenvalue decomposition is performed, the obtained eigenvalues ​​are sorted from small to large, and several eigenvalues ​​with a proportion of p are taken, and their corresponding eigenvectors constitute the projection vector of the feature subspace.

[0017] In one embodiment, a sparse representation method is used in the feature subspace to achieve target detection. The process is as follows:

[0018] Step 1: The projected spectral image data to be measured is sparsely represented using the projected target spectrum dictionary, and the sparse representation coefficients are calculated using the orthogonal matching pursuit (OMP) algorithm.

[0019] Step 2: Calculate the representation residual As detection output;

[0020] in, are the hyperspectral image data to be measured and the target spectrum dictionary projected into the feature subspace respectively;

[0021] Step 3: Separate the background and target through threshold segmentation to achieve target detection.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] (1) The present invention uses multiple prior information to construct a target spectrum dictionary, which can meet the demand for gradually improving algorithm effects as the spectrum database continues to expand.

[0024] (2) When calculating the feature subspace, the present invention takes into account the strong correlation between the hyperspectral data bands, and proposes to calculate the spectral correlation coefficient matrix based on the hyperspectral image data to be measured, and uses the size of the variance as the basis for judging the degree of data feature chaos in the feature space. The feature subspace of the target spectral dictionary is calculated based on the principle of minimum variance, and is integrated into the spectral correlation coefficient matrix to better consider the correlation between spectra and remove redundant information.

[0025] (3) When the present invention obtains the projection vector of the feature subspace based on eigenvalue decomposition, it is proposed to remove the first few eigenvalues ​​with the largest proportion, that is, to sort the eigenvalues ​​from small to large, take the eigenvalue with a proportion of p, and select the corresponding eigenvector as the projection vector of the feature subspace, so that the target spectrum dictionary in this feature subspace has stable target spectrum information characteristics.

[0026] (4) When the present invention adopts sparse representation in the feature subspace to obtain the representation residual, only the residual of the sparse representation based on the target spectrum dictionary in the projected feature subspace is obtained, and the representation residual of the background dictionary is discarded, which can improve the detection speed of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of the overall framework of the present invention. DETAILED DESCRIPTION

[0028] In order to facilitate ordinary technicians in this field to understand and implement the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific implementation examples. It should be understood that the implementation examples described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0029] The present invention is a subspace sparse representation hyperspectral target detection method based on spectral correlation. Taking into account the data redundancy caused by the strong correlation between the actual hyperspectral data bands and the insufficient extraction of target spectral feature information in the target spectral dictionary, the spectral correlation coefficient matrix of the hyperspectral data is analyzed, and the spectral data is projected into the feature subspace for sparse representation. When calculating the feature subspace, the spectral correlation coefficient matrix is ​​incorporated to obtain the feature subspace more accurately, and the variance is used as the measure of information difference. With the minimization of variance as the criterion, the eigenvalue decomposition method is adopted to eliminate several eigenvalues ​​with a large proportion, and the eigenvectors corresponding to the remaining eigenvalues ​​are selected as the feature subspace projection vectors. This ensures that there are sufficient dimensions for distinction and better extracts the common features of the prior spectral information, so as to detect and identify targets more accurately and efficiently, and has superior performance.

[0030] The method of the present invention includes three parts: estimating the spectral correlation coefficient matrix based on the hyperspectral data to be measured, calculating the characteristic subspace based on the prior spectral information and the spectral correlation coefficient matrix, and performing sparse representation in the characteristic subspace. Figure 1 As shown, the calculation of the spectral correlation coefficient matrix, the characteristic subspace selection and the subspace sparse representation method in the method of the present invention are introduced. The specific steps are as follows:

[0031] Step 1: Based on the hyperspectral image data to be measured, quantitative estimation of the correlation between spectral information is achieved to obtain the spectral correlation coefficient matrix.

[0032] In this step, the radiation value of the hyperspectral image to be measured is first standardized, and then the Gram matrix of the standardized data is calculated as the spectral correlation coefficient matrix, thereby achieving a quantitative estimation of the correlation between spectral information.

[0033] The Gram matrix is ​​calculated as follows:

[0034]

[0035] Where H represents the hyperspectral image data, and the value of n is equal to the total number of pixels in the hyperspectral image data.

[0036] In more embodiments, the present invention may also adopt Pearson correlation coefficient model, Spearman correlation coefficient model, mutual information model and other technical means to obtain spectral correlation coefficient.

[0037] For example, the present invention provides a method for normalization, which is achieved by dividing the radiance value of each band in the spectral data by the maximum radiance value.

[0038] Step 2: Based on the spectral correlation coefficient matrix and the spectral data in the target spectral dictionary, the minimum variance principle and eigenvalue decomposition method are used to calculate the characteristic subspace of the target spectral dictionary. The obtained eigenvalues ​​are sorted from small to large, and the first few eigenvalues ​​with a proportion of p are selected. The corresponding eigenvectors form the projection vector of the characteristic subspace. In this characteristic subspace, the spectral information in the target spectral dictionary has stable characteristics, which allows for more efficient and accurate detection and identification of targets.

[0039] In this step, the target spectrum dictionary consists of target spectral information extracted from the spectral information database. Based on the actual requirements of the target detection and identification task, the target spectrum information to be detected is selected to form the target spectrum dictionary, such as a certain aircraft or ship. The spectral data in the target spectrum dictionary needs to be standardized in the same way as the hyperspectral image data to be measured.

[0040] The minimum variance principle in this aspect refers to using the variance of the spectral information contained in the target spectral dictionary as the basis for judging the degree of information chaos in the feature subspace. The target spectral dictionary is denoted as D t , the correlation coefficient matrix obtained according to the hyperspectral image data is C. Based on this principle, the spectral correlation coefficient matrix C is integrated, and the variance of the spectral information contained in the target spectral dictionary is decomposed using the eigenvalue method to obtain the eigenvalue, and the corresponding eigenvector is taken to form the projection vector of the characteristic subspace.

[0041] The variance of the target spectrum dictionary after being incorporated into the spectrum correlation coefficient matrix is ​​recorded as Σ′, and the eigenvalue decomposition of the present invention is based on the following formula:

[0042]

[0043] After performing eigenvalue decomposition on Σ′, the obtained eigenvalues ​​are sorted from small to large, and several eigenvalues ​​with a proportion of p are selected (i.e., the sum of these eigenvalues ​​is p), which can be set to p = 0.01% to 0.1%. The corresponding eigenvectors constitute the projection vector of the feature subspace. The hyperspectral image data to be measured and the target spectrum dictionary are multiplied by it to complete the projection operation to the feature subspace, that is:

[0044]

[0045]

[0046] Where W is the obtained projection vector, They are the hyperspectral image data and target spectrum dictionary projected into the feature subspace respectively.

[0047] In this step, since the number of discarded feature dimensions is very small, the remaining feature dimensions are not much different from the original feature dimensions.

[0048] Step 3: Project all the hyperspectral image data to be tested and the target spectrum dictionary into the feature subspace, and use the sparse representation method to achieve target detection.

[0049] In this step, sparse representation is used in the feature subspace to detect the target. The process is as follows:

[0050] Step 31: sparsely represent the projected spectral image data using the projected target spectrum dictionary, and use the orthogonal matching pursuit (OMP) algorithm to solve the following optimization problem:

[0051]

[0052] K is the sparsity, is the sparse coefficient. Sparsity is the number of non-zero elements in the sparse coefficient, which can be used as a preset value. It is generally recommended to select about 20% of the number of prior spectral information. According to this formula, the sparse representation coefficient can be obtained

[0053] Step 32, calculate the representation residual As the detection output, the difference between the sparse representation of the target dictionary and the original spectral data is obtained by calculating the following formula:

[0054]

[0055] in, are the hyperspectral image data to be measured and the target spectrum dictionary projected into the feature subspace respectively.

[0056] In step 33, the background and the target are separated by threshold segmentation to achieve target detection.

[0057] The effectiveness of the present invention is verified by simulation experiments. The simulation experiment hardware equipment is as follows: processor Intel Xeon Gold 6226R CPU, memory 64-GB RAM, NVIDIA RTX 2080Ti 11-GB GPU.

[0058] Three publicly available hyperspectral remote sensing datasets are tested: HYDICE, University of Pavia, and KSC. 30% of the ground truth is selected as the prior spectral information to form the target dictionary, with the feature parameter p set to 99.99% and K set to 5.

[0059] Table 1

[0060]

[0061] The AUC value corresponding to the ROC curve of the output result is obtained. As can be seen from Table 1, the proposed method has higher accuracy and faster calculation speed than the existing supervised hyperspectral image target detection method.

[0062] In summary, this invention combines spectral correlation coefficient estimation, feature subspace, and sparse representation techniques to address the issues of information redundancy and insufficient feature extraction caused by strong inter-spectral correlation when multiple prior information is applied to hyperspectral target detection. By combining the spectral correlation coefficient matrix to calculate the minimum variance feature subspace, the method extracts common features of the prior spectral information in the target spectral dictionary and uses sparse representation in the feature subspace for target detection. Compared with existing supervised hyperspectral target detection algorithms, this method has better detection and recognition capabilities.

Claims

1. A hyperspectral target detection method based on spectral correlation subspace sparse representation, characterized in that: Based on the hyperspectral image data to be measured, the correlation between spectral information is quantitatively estimated to obtain the spectral correlation coefficient matrix; Based on the spectral correlation coefficient matrix and the spectral data in the target spectral dictionary, the minimum variance principle and eigenvalue decomposition method are used to calculate the characteristic subspace of the target spectral dictionary. The obtained eigenvalues ​​are sorted from small to large, and the first few eigenvalues ​​with a proportion of p are taken. The corresponding eigenvectors constitute the projection vector of the characteristic subspace; the hyperspectral image data to be tested and the target spectral dictionary are all projected into the characteristic subspace, and the sparse representation method is used to achieve target detection; Among them, according to the minimum variance principle and incorporating the spectral correlation coefficient matrix, the variance of the spectral information contained in the target spectral dictionary is decomposed using the eigenvalue method to obtain the eigenvalue, and the corresponding eigenvector is taken to form the projection vector of the characteristic subspace. The method is as follows: Step 1: Let the target spectrum dictionary be D t , calculate the correlation coefficient matrix C based on the input hyperspectral image data; Step 2, remember After integrating into the spectral correlation coefficient matrix, the variance of the target spectral dictionary is obtained; the eigenvalue decomposition is performed on it, the obtained eigenvalues ​​are sorted from small to large, and several eigenvalues ​​with a proportion of p are taken, and their corresponding eigenvectors constitute the projection vector of the characteristic subspace; The sparse representation method is used in the feature subspace to achieve target detection. The process is as follows: Step 1: The projected spectral image data to be measured is sparsely represented using the projected target spectrum dictionary, and the sparse representation coefficients are calculated using the orthogonal matching pursuit (OMP) algorithm. Step 2: Calculate the representation residual As detection output; in, are the hyperspectral image data to be measured and the target spectrum dictionary projected into the feature subspace respectively; Step 3: Separate the background and target through threshold segmentation to achieve target detection.

2. The method for hyperspectral target detection based on spectral correlation subspace sparse representation according to claim 1 is characterized in that: The method for quantitatively estimating the correlation between spectral information is: The radiation value of the hyperspectral image to be measured is standardized, and the spectral correlation coefficient matrix of the standardized data is calculated to achieve quantitative estimation of the correlation between spectral information.

3. The method for hyperspectral target detection based on spectral correlation subspace sparse representation according to claim 2, characterized in that: The standardization is to divide the radiation value of each band in the spectral data by the maximum radiation value; the same standardization operation is performed on the spectral data in the target spectral dictionary and the spectral data in the hyperspectral image data to be measured.

4. The method for hyperspectral target detection based on spectral correlation subspace sparse representation according to claim 2, characterized in that: The spectral correlation coefficient matrix is ​​obtained by calculation using a Gram matrix, or using a Pearson correlation coefficient model, a Spearman correlation coefficient model, or a mutual information model.

Images