Hyperspectral target detection method fusing multi-layer information

Abstract

The invention relates to the technical field of hyperspectral remote sensing image target detection, and discloses a hyperspectral target detection method fusing multi-layer information, which comprises the following steps of: firstly, denoising the spatial dimension of a hyperspectral image by adopting Gaussian filtering so as to maintain the edge details of the image and realize segmented smoothing; secondly, discrete Fourier transform is applied to the spectral dimension to extract spectral curve characteristics, so that the spectral contrast between the target and the background is improved; then, extracting an interested target area according to the prior spatial spectrum information of the target; further, the background and the abnormal target are respectively represented as low-rank and sparse components through an abnormal feature decomposition technology, and abnormal pixels are quantized by adopting a non-parameterized Manhattan distance detection model to obtain abnormal information; finally, accurate detection of an interested target is realized by fusing multi-level information, and the method provided by the invention shows relatively good detection performance in detection and identification of scattered household garbage targets in a wetland environment and a river and lake shoreline scene.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing image target detection, and in particular to a hyperspectral target detection method that fuses multi-layer information. Background Art

[0002] Currently, there are often artificial objects of various material types left in wetlands and other complex environments where human activities occur, such as plastic bottles, metal cans, fabrics, etc., and meteorological phenomena such as storms will also generate damaged wooden garbage in environments such as wetlands and river and lake shorelines; garbage of various materials will cause certain harm to the natural ecosystem. Often, the shapes, colors, and textures of these targets may be highly similar to the surrounding environment and are easily masked by the environment. For some small and low-contrast garbage targets, it is difficult to be accurately detected and identified using conventional detection techniques; with the development of remote sensing technology, hyperspectral remote sensing technology can effectively solve this problem. Different from traditional remote sensing technology, it can obtain spectral band information at the nanometer level of target materials by detecting hundreds of narrow bands, so as to detect and identify using the difference between the target and background bands; spectral images with a spectral resolution in the range of 10 -2 λ order of magnitude are called hyperspectral images.

[0003] The invention patent "A Small Target Detection Method for Hyperspectral Images Based on Unsupervised Segmentation" with the patent application number CN202211162154.1 provides a small target detection method for hyperspectral images based on unsupervised segmentation. It performs normalization processing on the hyperspectral target image to be detected and establishes a hyperspectral image segmentation model; according to the hyperspectral image segmentation result map, it counts the number of pixels contained in each connected domain in the segmentation result map and determines whether the number of pixels in each connected domain is less than the set threshold N; divides the image to be detected into multiple connected domains, and determines the connected domains with the number of pixels less than the set threshold N in the connected domain as the regions of interest for the small targets to be detected; uses the pixel spectra in the regions of interest and known spectral templates as the input of an adaptive cosine estimation hyperspectral detector to detect whether the pixels in the regions of interest are target pixels; finally outputs the detection result of the hyperspectral target image; the above technical solution uses the pixel spectra in the regions of interest and known spectral templates as the input of an adaptive cosine estimation hyperspectral detector, which overly relies on the prior information of the target material, resulting in weak model generalization ability and low robustness, prone to some false detections and missed detections, and ignores the spatial dimension information in the three-dimensional structure of hyperspectral images, resulting in low detection accuracy. In addition, relying only on the hyperspectral image segmentation model based on a deep learning network and the subsequent adaptive cosine estimation hyperspectral detector has the limitations of a single model and cannot effectively handle complex hyperspectral image scenarios. Summary of the Invention

[0004] The object of the present invention is to provide a hyperspectral target detection method that fuses multi-layer information, which can be used for target detection of hyperspectral images collected by aerial remote sensing equipment, and is particularly suitable for the detection and identification of scattered domestic waste targets in complex scenes.

[0005] The present invention adopts the following technical solutions:

[0006] A hyperspectral target detection method that fuses multi-layer information, comprising the following steps:

[0007] S1: Collect hyperspectral remote sensing images, and perform standardization processing on the original hyperspectral data in the hyperspectral remote sensing images;

[0008] S2: Extract multiple points of interest as the target of interest from the standardized hyperspectral data, perform arithmetic averaging on the spectral feature curves corresponding to the multiple extracted points of interest to obtain the average prior spectrum of the target of interest, and construct a spectral database according to the average prior spectrum;

[0009] S3: Perform filtering and noise reduction processing on the spatial domain data of the images in the spectral database, and then perform discrete Fourier transform processing on the spectral domain data of the processed images to obtain the fused prior spatial-spectral domain information that combines the spatial domain data and the spectral domain data. According to the fused prior spatial-spectral domain information, construct an adaptive cosine estimation model and obtain the information D1 of the normal target of interest;

[0010] S4: Decompose the hyperspectral data in the spectral database by means of low-rank sparse representation and matrix decomposition, optimize the obtained low-rank matrix and sparse matrix, design a non-Gaussian detector according to the optimized low-rank matrix and sparse matrix, and use the constructed non-parametric controlled Manhattan distance detection model to evaluate abnormal pixels to obtain the information D2 of the abnormal target of interest;

[0011] S5: Fuse the information D1 of the normal target of interest and the information D2 of the abnormal target of interest to finally obtain the target detection result.

[0012] Preferably, in the step S1, the original hyperspectral data is standardized by using the maximum value and minimum value method.

[0013] Preferably, in the step S2, the random sampling technique is used to extract multiple points of interest within the established region of interest from the standardized hyperspectral data.

[0014] Preferably, in the step S3, two-dimensional Gaussian filtering and noise reduction processing are performed on the spatial domain data of the images in the spectral database.

[0015] Preferably, in the step S3, the denoised image spectral domain data is further subjected to two-dimensional discrete Fourier transform processing to obtain the fused prior spatial-spectral domain information that combines the spatial domain data and the spectral domain data.

[0016] Preferably, in the step S3, finally, according to the fused prior spatial-spectral domain information, an adaptive cosine estimation model that combines the spatial domain and the spectral domain information is constructed, and the normal target information D1 of interest is calculated.

[0017] Preferably, constructing the adaptive cosine estimation model includes the following steps:

[0018] S3301: The hyperspectral data follows a multivariate normal distribution, and a binary hypothesis model that follows the multivariate normal distribution is established;

[0019] S3302: The binary hypothesis model is derived and solved to obtain the maximum likelihood estimates of the background noise variance and the joint variance of the target signal and the noise in the binary hypothesis model;

[0020] S3303: According to the obtained maximum likelihood estimates, an adaptive cosine estimation model is constructed, and the normal target information D1 of interest is calculated.

[0021] Preferably, in the step S4, it specifically includes the following steps:

[0022] S410: Using the method of low-rank sparse representation and matrix decomposition, the hyperspectral data is decomposed into a low-rank matrix, a sparse matrix, and a noise matrix; the decomposed sparse matrix and low-rank matrix are output to step S420;

[0023] S420: Using the GoDec algorithm, under the rank and sparsity constraints, the low-rank matrix and the sparse matrix obtained in step S410 are optimized by minimizing the decomposition error, and the optimized low-rank matrix and sparse components are output to step S430;

[0024] S430: According to the optimized low-rank matrix and sparse matrix, a non-Gaussian detector is designed, and the non-parametric controlled Manhattan distance detection model is used to evaluate the abnormal pixels and obtain the abnormal target information D2 of interest.

[0025] Preferably, in the S430, according to the low-rank matrix and the sparse matrix obtained from step S420, a non-Gaussian detector is designed; the spectral vector set extracted from the sparse matrix is used to calculate the standard vector representing the highest probability background by minimizing the l1 norm, and then a non-parametric controlled Manhattan distance detection model is constructed; finally, the Manhattan distance obtained from the Manhattan distance detection model is used to evaluate the abnormal pixels, and the abnormal target information D2 of interest is obtained.

[0026] Preferably, in step S5, the normal target information D1 of interest obtained in step S3 and the abnormal target information D2 of interest obtained in step S4 are fused by weighted signal-to-noise ratio index to finally obtain the target detection result D.

[0027] The beneficial effect of the present invention is that for various types of garbage that are easily covered by the environment, including plastic bottles, metal cans, fabrics left by humans, and damaged wooden garbage produced by meteorological disasters, by extracting spectral characteristic curves from pre-processed hyperspectral images and constructing a spectral database, the unique spectral characteristics of garbage targets that are highly similar to the environment, small in size, and low in contrast are accurately captured, breaking through the limitations of traditional detection methods that are affected by environmental interference and greatly improving detection accuracy;

[0028] The adaptive cosine estimation model built based on the spatial and spectral domain information of the prior target can adaptively adjust the detection strategy according to the diversity of the wetland environment and garbage targets, obtain the information of normal targets of interest, and significantly enhance the adaptability and reliability of the detection system to complex and changing scenes, overcoming the dilemma that traditional methods are difficult to cope with complex environments.

[0029] The Manhattan distance detection model constructed by using the abnormal feature decomposition method can effectively mine the abnormal features of tiny garbage targets and obtain the information of interesting abnormal targets, solving the problem that tiny garbage is easily ignored, and providing strong support for the comprehensive maintenance of the ecological environment.

[0030] Finally, the target detection results are obtained by fusing the above-obtained multi-layer information, integrating multi-dimensional data, further reducing the false detection and missed detection rates, and comprehensively improving the detection efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flowchart of a hyperspectral target detection method integrating multi-layer information according to the present invention;

[0032] Figure 2 This is a flow chart of obtaining the target of interest D1(s) by using the target prior space spectrum domain information in the present invention;

[0033] Figure 3 This is a flow chart of obtaining the abnormal target D2(s) of interest by using abnormal feature decomposition in the present invention;

[0034] Figure 4 This is a flow chart of finally obtaining the target detection result by fusing the normal target information D1 of interest and the abnormal target information D2 of interest in the present invention;

[0035] Figure 5These are the false color images, ground truth images, and detection result images of different methods in the present invention for the withered grass background wooden block dataset; where (a) is the false color image; (b) is the ground truth image; (c) is CEM; (d) is ACE; (e) is RX; (f) is LRASR; (g) is LRSNCR; (h) is Proposed;

[0036] Figure 6 These are the false color images, ground truth images, and detection result images of different methods in the present invention for the water surface background life jacket dataset; where (a) is the false color image; (b) is the ground truth image; (c) is CEM; (d) is ACE; (e) is RX; (f) is LRASR; (g) is LRSNCR; (h) is Proposed;

[0037] Figure 7 These are the ROC curve graphs of different methods in the present invention for the withered grass background wooden block dataset;

[0038] Figure 8 These are the ROC curve graphs of different methods in the present invention for the water surface background life jacket dataset. Detailed implementation manners

[0039] The present invention will be described in detail below with reference to the accompanying drawings and embodiments:

[0040] As shown in the Figures 1 to 8 accompanying drawings, the flowchart of a hyperspectral target detection method for fusing multi - layer information according to the present invention specifically includes the following steps:

[0041] S1: Collect hyperspectral remote sensing images, and perform standardization processing on the original hyperspectral data in the hyperspectral remote sensing images;

[0042] S2: Extract multiple points of interest as the target of interest from the standardized hyperspectral data, perform arithmetic averaging on the spectral feature curves corresponding to the multiple extracted points of interest to obtain the average prior spectrum of the target of interest, and construct a spectral database based on the average prior spectrum;

[0043] S3: Perform filtering and noise reduction processing on the spatial domain data of the images in the spectral database, then perform discrete Fourier transform processing on the spectral domain data of the processed images to obtain the fused prior spatial - spectral domain information that combines the spatial domain data and the spectral domain data. Based on the fused prior spatial - spectral domain information, construct an adaptive cosine estimation model and obtain the information D1 of the normal target of interest;

[0044] S4: Decompose the hyperspectral data in the spectral database by means of low-rank sparse representation and matrix decomposition, optimize the obtained low-rank matrix B and sparse matrix S, design a non-Gaussian detector based on the optimized low-rank matrix B and sparse matrix S, and use the constructed non-parametric controlled Manhattan distance detection model to evaluate abnormal pixels to obtain abnormal target information D2 of interest;

[0045] S5: Fuse the normal target information D1 of interest and the abnormal target information D2 of interest to finally obtain the target detection result.

[0046] In the present invention, in step S1, the original hyperspectral data is normalized by using the maximum value and minimum value method, and the formula is as follows:

[0047]

[0048] Among them, X0 represents the collected original hyperspectral data; X represents the hyperspectral data after normalization.

[0049] In the present invention, in step S2, by using the random sampling technique, for the hyperspectral data after normalization, multiple points of interest are extracted as targets of interest within the established region of interest, and the spectral feature curves corresponding to the multiple extracted points of interest are arithmetically averaged to obtain the average prior spectrum of the target of interest and construct a spectral database. The formula is as follows:

[0050]

[0051] Among them, S mean (w) represents the average prior spectrum of the target of interest, N1 represents the total number of selected points of interest; f() is the radiation response function of the spectrum; w represents the wavelength of the hyperspectral image; i r represents the index of the random point of interest; f(w,i r ) represents the prior spectrum of the i r th random point of interest at wavelength w.

[0052] In the present invention, in step S3, it specifically includes the following steps:

[0053] S310: Perform two-dimensional Gaussian filtering denoising processing on the spatial domain data of the image in the spectral database to improve the estimation of the background part;

[0054] When the hyperspectral image is converted from a three-dimensional stereoscopic image to a two-dimensional matrix, it will cause uneven local spatial information and some noise. By using a Gaussian filter to process the spatial domain of the hyperspectral image, the edge information of the image can be retained, the piecewise smoothness of the image can be promoted, so as to improve the estimation of the background part; the Gaussian function used for two-dimensional Gaussian filtering denoising processing is:

[0055]

[0056] Among them, x and y represent the coordinates of the image pixels; σ h represents the standard deviation for Gaussian filtering.

[0057] S320: Perform two-dimensional discrete Fourier transform processing on the spectral domain data of the image after noise reduction processing to obtain the representation of the pixel in the frequency domain;

[0058] In the present invention, in step S320, for a one-dimensional sequence β(n), where n = 0, 1,..., N2 - 1, the discrete Fourier transform can be defined as:

[0059]

[0060] Among them, X(α) represents the complex representation in the frequency domain; j represents the imaginary unit; N2 represents the sequence length; β(n) represents a one-dimensional sequence; n represents the sampling point serial number in the time domain; α represents the frequency index in the frequency domain, α = 0, 1,..., N2 - 1.

[0061] In the field of hyperspectral image analysis, each pixel unit does not carry only single intensity information, but contains a set of spectral information distributed with wavelengths. For these spectral information, discrete Fourier transform can be performed on each pixel to obtain the representation of the pixel in the frequency domain. When performing two-dimensional discrete Fourier transform on the entire image, the two-dimensional Fourier transform is defined as:

[0062]

[0063] Among them, F(u, v) is the complex representation in the frequency domain; M represents the size of the image in the x direction; N3 represents the size of the image in the y direction; u represents the index in the x direction; v represents the index in the y direction; g(x, y) represents the pixel intensity in the spatial domain; (x, y) represents the coordinates of the image pixels.

[0064] S330: Construct an adaptive cosine estimation model that fuses spatial domain and spectral domain information, and calculate the normal target information D1 of interest in the detection result.

[0065] In the present invention, in step S330, it specifically includes the following process,

[0066] S3301: The hyperspectral data follows a multivariate normal distribution, and establish a binary hypothesis model that follows a multivariate normal distribution;

[0067] Observation data set where pixel s m is a data set pixel vector, s m =(sm1 , s m2 , …, s mL ) T , where \(1\leq m\leq N_4\), \(N_4\) represents the number of pixels in the hyperspectral image; \(L\) is the number of bands in the hyperspectral image; when the background is uniform and follows a multivariate normal distribution, and the target and background spectra act in a linearly additive manner, the following binary hypothesis model is established:

[0068]

[0069] Among them, \(H_0\) represents the case without a target; \(H_1\) represents the case with a target; \(s m represents the spectral vector of the \(m\)th pixel in the observed data set; \(\xi\) represents the background noise vector, an \(L\)-dimensional normal random variable with zero mean; \(D\) represents the target signal matrix; \(\alpha d represents the target abundance.

[0070] Since the hyperspectral data follows a multivariate normal distribution, the binary hypothesis model can be expressed as

[0071]

[0072] Among them, \(N_4\) represents the number of pixels in the hyperspectral image; \(\Gamma\) represents the covariance matrix of the background clutter; \(\sigma_0 2 represents the variance of the background noise; \(D\) represents the target signal matrix; \(\sigma_1 2 represents the joint variance of the target signal and the noise; \(\alpha d represents the target abundance; \(D\alpha d represents the linear combination of the target signal and the abundance.

[0073] S3302: Derive and solve the binary hypothesis model to obtain the maximum likelihood estimates of the background noise variance and the joint variance of the target signal and the noise in the binary hypothesis model;

[0074] Deriving and solving Equation (7) gives the maximum likelihood estimates of \(\sigma_0 2 and \(\sigma_1 2 as follows:

[0075]

[0076] Among them, represents the maximum likelihood estimate of \(\sigma_0 2 ; represents the maximum likelihood estimate of \(\sigma_1 2 ; \(s m T represents the transpose of the spectral vector \(s\) of the \(m\)th pixel; \(\Gamma m represents the inverse matrix of the background clutter covariance matrix \(\Gamma\); \(L\) represents the number of bands in the hyperspectral image; \(D\alpha -1 ; dRepresents a linear combination of the target signal and the abundance.

[0077] S3303: Construct an adaptive cosine estimation model based on the obtained maximum likelihood estimate, and calculate the normal target information D1 of interest;

[0078] The obtained maximum likelihood estimate and Used to construct a generalized likelihood ratio test. By substituting σ0 2 and σ1 2 into the likelihood ratio formula and simplifying, finally construct an adaptive cosine estimation model as follows:

[0079]

[0080] where d represents the target signal vector; s represents a specific pixel spectral vector; s T represents the transpose of the specific pixel spectral vector s; Γ -1 represents the inverse matrix of the background clutter covariance matrix Γ;

[0081] When d = D, Equation (9) can be rewritten as:

[0082]

[0083] Calculate the target D1 information of interest in the detection result, that is, D1 = ACE(s).

[0084] In the present invention, the adaptive cosine estimation model is based on an unstructured background model. In the model identification stage, it does not distinguish between the background and noise, but assumes that the background has consistency and follows the multivariate normal distribution law, and the target spectrum and the background spectrum are in a linear mixing relationship; therefore, the abnormal target information D2 of interest is obtained by using eigen - decomposition. Since in hyperspectral abnormal target detection, the abnormal target contains fewer pixels and the background contains more pixels, in matrix decomposition, the target can be represented by sparse components and the background can be represented by low - rank components. In step S4, the specific process includes the following:

[0085] S410: Use the method of low - rank sparse representation and matrix decomposition to decompose the hyperspectral image into a low - rank matrix B, a sparse matrix S, and a noise matrix N5. The decomposed low - rank matrix B and sparse matrix S are output to step S420;

[0086] In this embodiment, in step S410, the extraction model can be expressed as:

[0087] X = B + S + N5 (11)

[0088] Among them, X represents the hyperspectral data after normalization; B is a low-rank matrix representing the background component; S is a sparse matrix representing the target component; and N5 is a noise matrix representing data perturbation.

[0089] S420: Using the GoDec algorithm under rank and sparse constraints, optimize the low-rank matrix B and the sparse matrix S obtained in step S410 by minimizing the decomposition error, and output the optimized sparse matrix S and low-rank matrix B to step S430;

[0090] In this embodiment, in step S420, the GoDec algorithm is used to solve the low-rank background component and the sparse component by minimizing the decomposition error under low-rank and sparse constraints, as follows:

[0091]

[0092] Among them, represents jointly optimizing B and S to minimize the objective function under given constraints; ‖·‖ F represents the Frobenius norm of the matrix, making the reconstruction error after low-rank sparse decomposition the smallest; rank() represents the function of finding the rank of the matrix; r represents the upper limit of the rank of the low-rank matrix B, which can be set according to the main background endmember; card( ) represents the sparsity function of the matrix; γ represents the upper limit of the number of non-zero elements in the sparse matrix S, called the l0 norm of the sparse matrix S. The smaller the value of γ, the sparser the sparse matrix S.

[0093] Transform the optimization problem in Equation (12) into alternately solving the following two sub-formulas until convergence:

[0094]

[0095] Among them, B t represents the low-rank matrix solved in the t-th iteration; S t represents the sparse matrix solved in the t-th iteration; t represents the number of iterations; N4 represents the number of pixels of the hyperspectral image;

[0096] Use the low-order approximation theory to solve the sub-problem in (13). When

[0097]

[0098] Among them, Y1 represents the left projection matrix generated by bilateral random projection; A1 ∈ R L×r , R L×r , They are all random matrices, but with different dimensions. L represents the number of bands of the hyperspectral image; r is the upper limit of the rank of the low-rank matrix B; A1 can be obtained through the randn function, which can generate a random matrix conforming to the standard normal distribution; Y2 represents the right projection matrix generated by bilateral random projection; A2 can be obtained through A2 = Y1 = XA1.

[0099] Based on bilateral random projection The approximate matrix of rank r is:

[0100]

[0101] For the sub-problem in (14), S t is updated by the element-wise hard thresholding method of X–B t-1 That is:

[0102]

[0103] where P Ω () is the projection of the matrix on the set Ω, and Ω is the non-zero subset composed of the first γN4 elements arranged in descending order of the absolute values in |X–B t-1 |.

[0104] S430: According to the optimized low-rank matrix B and sparse matrix S, design a non-Gaussian detector, and use the constructed non-parametric controlled Manhattan distance detection model to evaluate abnormal pixels and obtain the information D2 of the abnormal target of interest.

[0105] Since the bands are segmented at specific spectra, the Manhattan distance can be effectively used for anomaly detection.

[0106] In step S430, constructing the Manhattan distance detection model mainly includes the following process:

[0107] Before calculating the Manhattan distance, it is necessary to calculate the standard vector e B ,

[0108]

[0109] where ‖·‖0 represents the l0 norm of the matrix; E a represents the set of spectral vectors extracted from the optimized sparse matrix S, where each vector corresponds to the spectral characteristics of a pixel in the hyperspectral image; e ij is the vector at the i-th row and j-th column of the matrix E a ;

[0110] The background in the abnormal component is sparse and close to zero, and the l0 norm is not suitable for finding the standard vector. The l0 norm is relaxed to the l1 norm to obtain the standard vector:

[0111]

[0112] Among them, ‖·‖1 represents the l1 norm of the matrix.

[0113] In hyperspectral anomaly target detection, since the anomaly target contains fewer pixels and the background contains more pixels, in matrix decomposition, the target component can be represented by a sparse matrix and the background component can be represented by a low-rank matrix; applying the Manhattan distance can effectively avoid the problem that in extreme cases, due to the excessive sparsity of the target components, the Gaussian covariance matrix (i.e., Euclidean distance or Mahalanobis distance) cannot be inverted. In addition, considering the offset caused by the anomaly, the vector that minimizes the l1 norm is used as the standard vector e representing the background with the highest probability. B It is more reasonable to replace the average value; the detection output is achieved by calculating the Manhattan distance, and it is expected to be more suitable for obtaining the outlier residuals in hyperspectral anomaly detection; therefore, the constructed Manhattan distance detection model is more conducive to obtaining the outlier residuals in hyperspectral anomaly detection.

[0114] According to the obtained standard vector e B , a non-parametric controlled Manhattan distance detection model is constructed, and the detection model can be expressed as:

[0115]

[0116] Among them, r p is the Manhattan distance of the pixel to be measured; e pk represents the k-th element to be measured; T is the threshold for anomaly detection; L represents the number of bands of the hyperspectral image; e Bk represents the k-th element of the background standard vector e B .

[0117] The Manhattan distance obtained according to the Manhattan distance detection model is used to evaluate the anomaly pixels, and the information D2 of the anomaly target of interest is obtained. D2(s) = r p (s), where s represents a specific pixel spectral vector.

[0118] In the present invention, in step S5, the information D1 of the normal target of interest obtained in step S3 and the information D2 of the anomaly target of interest obtained in step S4 are fused by a weighted signal-to-noise ratio index to obtain the target detection result D, and the formula is as follows:

[0119]

[0120] Among them, D is the target information; D1 represents the normal target information of interest obtained through the adaptive cosine estimation model; D2 represents the abnormal target information of interest obtained through the method of abnormal feature decomposition; SNR1 represents the signal-to-noise ratio of the normal target information D1 of interest, and SNR2 is the signal-to-noise ratio of the abnormal target information D2 of interest. Specific Embodiment 1:

[0122] It is obtained by using the aerial hyperspectral remote sensing method of UAV + hyperspectral imager. The UAV is the DJI M300 RTK multi-rotor UAV, and the imager is the Anzhou Technology Cubert X20P airborne hyperspectral imager; their spectral range is from 350 nanometers to 1000 nanometers, covering the visible light and part of the near-infrared spectral bands. The spatial resolution at a distance of 110 meters is 3.73 centimeters, and the spectral resolution is 4 nanometers.

[0123] The background of the dataset contains withered grass and water, and the targets contain wooden blocks and life jackets. The prior spectra of the targets are obtained through the collected data. The false color map and the ground truth map of the dataset are shown in Appendix Figure 5 (a)-(b) and Appendix Figure 6 (a)-(b) respectively.

[0124] Figure 5 (c) to (h) and Figure 6 (c) to (h) respectively show the detection result maps of different methods on the wooden block dataset with withered grass background and the life jacket dataset with water surface background. Among them, "Proposed" is the detection result corresponding to the proposed method; it can be seen from the detection result maps that the proposed method of the present invention can detect all targets while well suppressing the background; however, some of the comparison methods detect fewer target pixels, and some methods have poor background suppression and are almost confused with the targets.

[0125] Figure 7 and Figure 8 respectively show the Receiver Operating Characteristic (ROC) curves of different methods on the two collected datasets. In the figure, "Proposed" is the ROC curve corresponding to the proposed method of the present invention; in Figure 7 when the false alarm probability is 0.1, the detection probability of the proposed method of the present invention is 1.0, while the detection probabilities of other methods are all lower than 1.0, and the detection performance is significantly better than that of other algorithms. When the detection probability is 0.9, the false alarm probability of the proposed method of the present invention is 0.0051, while the false alarm probabilities of other methods are all higher than 0.0051, and the false alarm probability is significantly better than that of other algorithms; in Figure 8In [the case where] the false alarm probability is 0.1, the detection probability of the method proposed by the present invention is 1.0, while the detection probabilities of other methods are all lower than 1.0. The detection probability is significantly better than the detection performance of other algorithms. When the detection probability is 0.9, the false alarm probability of the method proposed by the present invention is 0.0038, while the false alarm probabilities of other methods are all higher than 0.0048. The false alarm probability is significantly better than the detection performance of other algorithms.

[0126] It can be seen from Figure 7 and Figure 8 that in the two ROC curve graphs, the ROC curves of the method proposed by the present invention are both located in the upper left of the comparison methods. This indicates that the performance of the method proposed by the present invention on the ROC curve is significantly better than that of other algorithms.

[0127] The AUC values of different methods in two experimental data sets are as follows in the table,

[0128] Table 1. AUC values of different methods

[0129]

[0130]

[0131] Table 1 shows the AUC values of different methods in two experimental data sets. The AUC values of the proposed method in the two data sets are 0.9979 and 0.9990 respectively, both of which are higher than the used comparison methods. It shows that the method proposed by the present invention has excellent detection performance.

Claims

1. A hyperspectral target detection method integrating multi-layer information, characterized in that: The steps include: S1: Collect hyperspectral remote sensing images and standardize the original hyperspectral data in the hyperspectral remote sensing images; S2: extract multiple points of interest from the standardized hyperspectral data as targets of interest, perform arithmetic averaging on the spectral characteristic curves corresponding to the extracted multiple points of interest to obtain the average prior spectrum of the target of interest, and construct a spectral database based on the average prior spectrum; S3: Filter and de-noise the spatial domain data of the image in the spectral database, and then perform discrete Fourier transform on the spectral domain data of the processed image to obtain the fused prior space-spectral domain information that combines the spatial domain data and the spectral domain data. Based on the fused prior space-spectral domain information, an adaptive cosine estimation model is constructed to obtain the normal target information of interest. ; S4: Decompose the hyperspectral data in the spectral database by low-rank sparse representation and matrix decomposition, optimize the obtained low-rank matrix and sparse matrix, design a non-Gaussian detector based on the optimized low-rank matrix and sparse matrix, and use the constructed non-parametric controlled Manhattan distance detection model to evaluate abnormal pixels and obtain the abnormal target information of interest. ; S5: Fusion of normal target information of interest and interesting abnormal target information , and finally obtain the target detection result.

2. The hyperspectral target detection method integrating multi-layer information according to claim 1, characterized in that: In the step S1, the original hyperspectral data is standardized using the maximum and minimum method.

3. The hyperspectral target detection method integrating multi-layer information according to claim 1, characterized in that: In the step S2, a random sampling technique is used to extract a plurality of points of interest from the standardized hyperspectral data within a predetermined region of interest.

4. The hyperspectral target detection method integrating multi-layer information according to claim 1, characterized in that: In the step S3, two-dimensional Gaussian filtering and noise reduction processing is performed on the image space domain data in the spectral database.

5. The hyperspectral target detection method integrating multi-layer information according to claim 4 is characterized in that: In the step S3, two-dimensional discrete Fourier transform is further performed on the spectral domain data of the image after the noise reduction process to obtain fused prior space-spectral domain information that fuses the spatial domain data and the spectral domain data.

6. The hyperspectral target detection method integrating multi-layer information according to claim 5, characterized in that: In the step S3, finally, based on the fusion of prior spatial-spectral domain information, an adaptive cosine estimation model integrating spatial domain and spectral domain information is constructed, and the normal target information of interest is calculated. .

7. The hyperspectral target detection method integrating multi-layer information according to claim 6 is characterized in that: Building an adaptive cosine estimation model includes the following steps: S3301: Hyperspectral data obeys multivariate normal distribution, and a binary hypothesis model obeying multivariate normal distribution is established; S3302: deriving and solving the binary hypothesis model to obtain the maximum likelihood estimation of the background noise variance and the joint variance of the target signal and the noise in the binary hypothesis model; S3303: Construct an adaptive cosine estimation model based on the obtained maximum likelihood estimation, and calculate the normal target information of interest .

8. The hyperspectral target detection method integrating multi-layer information according to claim 1, characterized in that: The step S4 specifically includes the following steps: S410: Decomposing the hyperspectral data into a low-rank matrix, a sparse matrix and a noise matrix by using a low-rank sparse representation and matrix decomposition method; outputting the decomposed sparse matrix and low-rank matrix to step S420; S420: Using the GoDec algorithm to optimize the low-rank matrix and sparse matrix obtained in step S410 by minimizing the decomposition error under rank and sparsity constraints, and outputting the optimized low-rank matrix and sparse components to step S430; S430: Based on the optimized low-rank matrix and sparse matrix, a non-Gaussian detector is designed, and the non-parametric controlled Manhattan distance detection model is used to evaluate abnormal pixels and obtain the information of the abnormal target of interest. .

9. The hyperspectral target detection method integrating multi-layer information according to claim 8, characterized in that: In the above S430, a non-Gaussian detector is designed according to the low-rank matrix and the sparse matrix obtained from step S420; the set of spectral vectors extracted from the sparse matrix is ​​minimized. The norm calculation obtains the standard vector representing the highest probability background, and then constructs a non-parametric controlled Manhattan distance detection model; finally, the Manhattan distance obtained by the Manhattan distance detection model is used to evaluate the abnormal pixels and obtain the abnormal target information of interest. .

10. The hyperspectral target detection method integrating multi-layer information according to claim 1, characterized in that: In step S5, the normal target information of interest obtained in step S3 is and the abnormal target information of interest obtained in step S4 Perform weighted signal-to-noise ratio index fusion to finally obtain the target detection result .

Images