A change detection method for dual-temporal hyperspectral remote sensing images

Abstract

The present invention provides a method for detecting changes in dual-phase hyperspectral remote sensing images, comprising the following steps: with respect to spectral information: first, performing image fusion to obtain two-phase image fusion features, then obtaining a preliminary delimiting value and a comprehensive weight, and finally converting the preliminary delimiting value into a reference delimiting value; with respect to spatial information: first, obtaining a first principal component difference image, then obtaining an amplitude spectrum and a phase spectrum in the frequency domain, and finally obtaining a frequency domain significance level through inverse Fourier transform; and information fusion: performing a convolution operation on the frequency domain significance level and the reference delimiting value to obtain a comprehensive delimiting value, and outputting a change detection result of the dual-phase hyperspectral remote sensing image in combination with a preset false alarm threshold. The present invention provides a method for detecting changes in dual-phase hyperspectral remote sensing images, which solves the problem that current dual-phase hyperspectral remote sensing image change detection technologies are prone to high false alarms, resulting in insufficient detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing, and more particularly to a method for detecting changes in dual-temporal hyperspectral remote sensing images. Background Art

[0002] Dual-temporal hyperspectral remote sensing image change detection, a key application of hyperspectral remote sensing technology, can continuously observe and detect changes in imaged scenes. In recent years, it has been widely used in practical fields such as environmental monitoring, urban expansion, and natural disaster detection and assessment. Essentially, the purpose of dual-temporal hyperspectral remote sensing image change detection is to model the spatial-spectral differences between the previous and next image phases, thereby distinguishing between changed and unchanged areas in the two image phases.

[0003] Current change detection techniques for dual-temporal hyperspectral remote sensing images, represented by algebraic, transformation, and classification methods, primarily focus on mining and characterizing the differences between pixel-level spectral features in hyperspectral remote sensing images to achieve change detection. However, these techniques are extremely sensitive to pixel-level changes and are prone to false alarms, resulting in insufficient detection accuracy. Summary of the Invention

[0004] The present invention provides a method for detecting changes in dual-temporal hyperspectral remote sensing images to overcome the technical defect that the current dual-temporal hyperspectral remote sensing image change detection technology is prone to high false alarms, thereby resulting in insufficient detection accuracy.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] A method for detecting changes in dual-temporal hyperspectral remote sensing images comprises the following steps:

[0007] In terms of spectral information: First, a pair of dual-phase hyperspectral remote sensing images are fused to obtain the fusion features of the two phases of images. Then, preliminary delimitation values and comprehensive weights are obtained based on the image fusion features. Finally, the preliminary delimitation values are converted into reference delimitation values based on the comprehensive weights.

[0008] In terms of spatial information: First, the first principal component difference image of the dual-temporal hyperspectral remote sensing image is obtained. Then, the amplitude spectrum and phase spectrum of the first principal component difference image in the frequency domain are obtained. The amplitude spectrum is filtered using Gaussian filtering to suppress non-changing attributes. Finally, the frequency domain significance level is obtained through inverse Fourier transform.

[0009] Information fusion: performing a convolution operation on the frequency domain significance level and the reference threshold value to obtain a comprehensive threshold value, and outputting the change detection result of the dual-phase hyperspectral remote sensing image in combination with a preset false alarm threshold.

[0010] In the above scheme, the reference boundary value is obtained by quantifying the spectral information of the dual-phase hyperspectral remote sensing image, and the frequency domain significance level is obtained according to the amplitude spectrum and phase spectrum of the first principal component difference image of the dual-phase hyperspectral remote sensing image in the frequency domain by using inverse Fourier transform. Finally, the frequency domain significance level is convolved with the reference boundary value to obtain a comprehensive boundary value, and the binary detection result of the dual-phase hyperspectral remote sensing image change detection is output according to the false alarm threshold, so as to achieve the best balance between false alarms and missed detections in the detection results.

[0011] Preferably, the image fusion features are Fusion features with images The edges are filled in a symmetrical manner to ensure that the edges and corner pixels have eight-connected regions;

[0012] By calculation and The spectral absolute distance of the center pixel pair of the eight connected regions is used to obtain a preliminary boundary value. The spectral absolute distance is calculated by the following formula:

[0013]

[0014] Among them, N b Indicates the number of bands of the hyperspectral remote sensing image, express The spectral reflectance value of the i-th band of the corresponding b-th pixel, express The spectral reflectance value of the i-th band corresponding to the b-th pixel.

[0015] Preferably, the similarity between each pixel pair is quantified by introducing the spectral cosine distance function:

[0016]

[0017] The weight scaling factor is constructed by using the gradient correlation of the pixels in the eight connected regions of the central pixel:

[0018]

[0019]

[0020] The comprehensive weight is calculated based on the weight scaling factor:

[0021] s b =ω b ·(1-g b )

[0022] The reference limit value is calculated by the following formula:

[0023] D=sb dist b ,b=1,2,...,N

[0024] in, are the universal symbols of the two-phase image pixels, express The transpose of express The transpose of ω b,c It represents the cosine similarity between the pixel pairs of the eight-connected regions in two time phases. Indicates the cosine similarity value from the second pixel to the last pixel. Represents pixels The cth neighboring pixel of Represents pixels The cth neighborhood pixel of , c represents the counting variable of the number of pixels in the eight-connected region, c = 1, 2, ..., 8; ω b Represents ω b,v The average weight of .

[0025] In the above scheme, the absolute spectral distance, cosine similarity and gradient correlation of eight-connected neighborhood pixels of pixel pairs are fully utilized to quantify the differences in spectral attribute patterns of pixels in dual-temporal hyperspectral remote sensing images.

[0026] Preferably, a principal component analysis algorithm is used to extract the first principal component features of each hyperspectral remote sensing image, and an image difference operation is used to obtain the structural and texture difference features between the first principal component features to obtain a first principal component difference image.

[0027] Preferably, the first principal component difference image G is mapped to the frequency domain using Fourier transform, and its amplitude spectrum is extracted from the image spectrum features:

[0028] A(f)=R(F(G))

[0029] And perform a log transform on A(f) to suppress the upper limit of the amplitude of the non-changing attribute and amplify the difference in the amplitude of the changing attribute:

[0030] L(f)=log(A(f))

[0031] Where F(·) represents Fourier transform, and R(·) represents the modulus of the complex features in the frequency domain.

[0032] Preferably, the Gaussian filter is a two-dimensional Gaussian filter:

[0033]

[0034] The filtered output is:

[0035] G(f)=L(f)*K(x,y)

[0036] Among them, x c with y c Represents the two-dimensional coordinate position of the center pixel,

[0037]

[0038] ω g =2×ceil(2σ)+1,

[0039] ω g represents the local window size, x and y represent the eight-connected region pixels (ω g ×ω g -1), σ represents the standard deviation of the two-dimensional Gaussian filter, and ceil(·) represents rounding up to an integer.

[0040] Preferably, the method further includes extracting phase spectrum features from the frequency domain representation features:

[0041] P(f)=S(F(G))

[0042] in, Represents the function of taking the imaginary part of the complex features in the frequency domain.

[0043] Preferably, G(f) and P(f) are used to perform inverse Fourier transform to obtain the frequency domain significance level:

[0044] G1=F -1 (exp(G(f)+j·P(f))) 2

[0045] Among them, F -1 (·) represents the inverse Fourier transform, and j is the imaginary number symbol.

[0046] Preferably, the comprehensive definition value is calculated by the following formula:

[0047] V c =Norm(D)*Norm(G1)

[0048] Where Norm(·) represents the maximum and minimum normalization function, D represents the reference limit value, and G1 represents the frequency domain significance level.

[0049] Preferably, the preset false alarm threshold is β t =2.5×10 -2 .

[0050] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0051] The present invention provides a dual-phase hyperspectral remote sensing image change detection method, which obtains a reference boundary value by quantifying the spectral information of the dual-phase hyperspectral remote sensing image, and obtains a frequency domain significance level according to the amplitude spectrum and phase spectrum of the first principal component difference image of the dual-phase hyperspectral remote sensing image in the frequency domain by using inverse Fourier transform. Finally, the frequency domain significance level is convolved with the reference boundary value to obtain a comprehensive boundary value, and a binary detection result of the dual-phase hyperspectral remote sensing image change detection is output according to the false alarm threshold, so that the false alarm and missed detection in the detection result are optimally balanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a schematic diagram of the implementation process of the technical solution of the present invention;

[0053] Figure 2 Schematic diagram showing the comparison of detection effects of different methods in the present invention on a river hyperspectral remote sensing image dataset;

[0054] Figure 3 Schematic diagram comparing the detection effects of different methods in the present invention on the farm hyperspectral remote sensing image dataset. DETAILED DESCRIPTION

[0055] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0056] In order to better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size;

[0057] It is understandable to those skilled in the art that some well-known structures and descriptions thereof may be omitted in the drawings.

[0058] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0059] Example 1

[0060] like Figure 1 As shown, a dual-temporal hyperspectral remote sensing image change detection method includes the following steps:

[0061] In terms of spectral information: First, a pair of dual-phase hyperspectral remote sensing images (such as Figure 1 The original hyperspectral images T1 and T2 in the image are fused to obtain the fusion features of the two images (such as Figure 1 Then, a preliminary delimiting value and a comprehensive weight are obtained according to the image fusion features, and finally the preliminary delimiting value is converted into a reference delimiting value according to the comprehensive weight;

[0062] In terms of spatial information: First, the first principal component difference image of the dual-temporal hyperspectral remote sensing image is obtained. Then, the amplitude spectrum and phase spectrum of the first principal component difference image in the frequency domain are obtained. The amplitude spectrum is filtered using Gaussian filtering to suppress non-changing attributes. Finally, the frequency domain significance level is obtained through inverse Fourier transform.

[0063] Information fusion: performing a convolution operation on the frequency domain significance level and the reference threshold value to obtain a comprehensive threshold value, and outputting the change detection result of the dual-phase hyperspectral remote sensing image in combination with a preset false alarm threshold.

[0064] In the specific implementation process, the reference boundary value is obtained by quantifying the spectral information of the dual-phase hyperspectral remote sensing image, and the frequency domain significance level is obtained according to the amplitude spectrum and phase spectrum of the first principal component difference image of the dual-phase hyperspectral remote sensing image in the frequency domain by using inverse Fourier transform. Finally, the frequency domain significance level is convolved with the reference boundary value to obtain a comprehensive boundary value, and the binary detection result of the dual-phase hyperspectral remote sensing image change detection is output according to the false alarm threshold, so as to achieve the best balance between false alarms and missed detections in the detection results.

[0065] Example 2

[0066] A method for detecting changes in dual-temporal hyperspectral remote sensing images comprises the following steps:

[0067] In terms of spectral information: First, a pair of dual-phase hyperspectral remote sensing images (such as Figure 1 The original hyperspectral images T1 and T2 in the image are fused to obtain the fusion features of the two images (such as Figure 1 Then, a preliminary delimiting value and a comprehensive weight are obtained according to the image fusion features, and finally the preliminary delimiting value is converted into a reference delimiting value according to the comprehensive weight;

[0068] More specifically, the image fusion features Fusion features with images The edges of the pixel are filled in a symmetrical manner to ensure that the edge and the corner pixel have an eight-connected area; in this embodiment, the number of rows and columns filled is 2;

[0069] By calculation and The spectral absolute distance of the center pixel pair of the eight connected regions is used to obtain a preliminary boundary value. The spectral absolute distance is calculated by the following formula:

[0070]

[0071] Among them, N b Indicates the number of bands of the hyperspectral remote sensing image, express The spectral reflectance value of the i-th band of the corresponding b-th pixel, express The spectral reflectance value of the i-th band corresponding to the b-th pixel.

[0072] More specifically, the spectral cosine distance function is introduced to quantify the similarity between pixel pairs:

[0073]

[0074] The weight scaling factor is constructed by using the gradient correlation of the pixels in the eight connected regions of the central pixel:

[0075]

[0076]

[0077] The comprehensive weight is calculated based on the weight scaling factor:

[0078] s b =ω b ·(1-g b )

[0079] The reference limit value is calculated by the following formula:

[0080] D=s b dist b ,b=1,2,...,N

[0081] in, are the universal symbols of the two-phase image pixels, Representation The transpose of express The transpose of ω b,c It represents the cosine similarity between the pixel pairs of the eight-connected regions in two time phases. Indicates the cosine similarity value from the second pixel to the last pixel. Represents pixels The cth neighboring pixel of Represents pixels The cth neighborhood pixel of , c represents the counting variable of the number of pixels in the eight-connected region, c = 1, 2, ..., 8; ω b Represents ω b,c The average weight of .

[0082] During the specific implementation process, the absolute spectral distance, cosine similarity, and gradient correlation of eight-connected neighboring pixels of a pixel pair were fully utilized to quantify the differences in the spectral attribute patterns of pixels in dual-phase hyperspectral remote sensing images. If only the absolute spectral distance and cosine similarity were used to define the attribute patterns of pixel pairs, the attribute patterns would tend to be variable. This is because the acquisition of hyperspectral remote sensing data is often affected by natural and unnatural factors such as cloud obscuration and sensor jitter. To this end, this embodiment uses the gradient correlation of pixels in the eight-connected region of the central pixel to construct a weight scaling factor, effectively overcoming the false alarm phenomenon.

[0083] In terms of spatial information: First, the first principal component difference image of the dual-temporal hyperspectral remote sensing image is obtained. Then, the amplitude spectrum and phase spectrum of the first principal component difference image in the frequency domain are obtained. The amplitude spectrum is filtered using Gaussian filtering to suppress non-changing attributes. Finally, the frequency domain significance level is obtained through inverse Fourier transform.

[0084] More specifically, the principal component analysis algorithm is used to extract the first principal component features of each hyperspectral remote sensing image, and the image difference operation is used to obtain the structural and texture difference features between the first principal component features to obtain the first principal component difference image.

[0085] More specifically, the first principal component difference image G is mapped to the frequency domain using Fourier transform, and its amplitude spectrum is extracted from the image spectrum features:

[0086] A(f)=R(F(G))

[0087] And perform a log transform on A(f) to suppress the upper limit of the amplitude of the non-changing attribute and amplify the difference in the amplitude of the changing attribute:

[0088] L(f)=log(A(f))

[0089] Where F(·) represents Fourier transform, and R(·) represents the modulus of the complex features in the frequency domain.

[0090] More specifically, the Gaussian filter is a two-dimensional Gaussian filter:

[0091]

[0092] The filtered output is:

[0093] G(f)=L(f)*K(x,y)

[0094] Among them, x c with y c Represents the two-dimensional coordinate position of the center pixel,

[0095]

[0096] ω g =2×ceil(2σ)+1,

[0097] ω g represents the local window size, x and y represent the eight-connected region pixels (ω g ×ω g -1), σ represents the standard deviation of the two-dimensional Gaussian filter, and ceil(·) represents rounding up to an integer.

[0098] In the specific implementation process, a two-dimensional Gaussian filter is introduced to perform a standard deviation-induced window ω on the log amplitude spectrum of the image. g ×ω g Filtering is performed to further suppress the amplitude of the non-changing attribute, thereby improving the separability of the changing and non-changing attribute pixels.

[0099] More specifically, it also includes phase spectrum feature extraction of frequency domain representation features:

[0100] P(f)=S(F(G))

[0101] in, Represents the function of taking the imaginary part of the complex features in the frequency domain.

[0102] More specifically, G(f) and P(f) are used to perform inverse Fourier transform to obtain the frequency domain significance level:

[0103] G1=F -1 (exp(G(f)+j·P(f))) 2

[0104] Among them, F -1 (·) represents the inverse Fourier transform, and j is the imaginary number symbol.

[0105] Information fusion: performing a convolution operation on the frequency domain significance level and the reference threshold value to obtain a comprehensive threshold value, and outputting the change detection result of the dual-phase hyperspectral remote sensing image in combination with a preset false alarm threshold.

[0106] More specifically, the comprehensive threshold value is calculated using the following formula:

[0107] V c =Norm(D)*Norm(G1)

[0108] Where Norm(·) represents the maximum and minimum normalization function, D represents the reference limit value, and G1 represents the frequency domain significance level.

[0109] More specifically, the preset false alarm threshold is β t =2.5×10 -2 .

[0110] In the specific implementation process, in order to obtain the binary classification results of the dual-temporal hyperspectral remote sensing image change detection, this embodiment uses V c The value of is divided into 5000 candidate thresholds with equal intervals from the minimum to the maximum value, and the recall rate and false alarm rate of the method described in this embodiment are counted under each candidate threshold. Finally, at the false alarm rate β t =2.5×10 -2 At the horizontal level, the entire image is binarized using a threshold.

[0111] Example 3

[0112] To verify the effectiveness of the dual-temporal hyperspectral remote sensing image change detection method (FDSE method) described in this example, a comparative analysis was conducted using the FDSE method and other transformation detection methods. The comparison methods included the classic CVA, PCA-CVA, and IRMAD methods, as well as the advanced PTCD, SALA, and MMPs methods. Based on the binary confusion matrix, this example introduced five performance evaluation metrics: overall accuracy (OA), average accuracy (AA), Kappa coefficient (Kappa), intersection over union (IoU), and F1 score (F1). These metrics objectively and impartially assess the performance of various dual-temporal hyperspectral remote sensing image change detection algorithms. Furthermore, in this example, the overall accuracy of changed pixels is denoted as OAc, and the overall accuracy of non-changed pixels is denoted as OAu.

[0113] Table 1 shows the OA, AA, Kappa, F1, IoU and time consumption indicators obtained by CVA, PCA-CVA, IRMAD, PTCD, SALA, and MMPs using the FDSE method of this embodiment on the river hyperspectral remote sensing image dataset.

[0114] Table 1

[0115]

[0116]

[0117] In terms of the classic change detection methods CVA, PCA-CVA, and IRMAD analysis, although FDSE's OAc metric is lower than those of CVA and PCA-CVA, FDSE's OAc is 0.97504, achieving the highest accuracy. Furthermore, the FDSE method's F1 (0.98076), IoU (0.96225), and OA (0.96508) metrics are also significantly higher than those of the three classic methods. The results in Table 1 demonstrate that, compared with classic methods, the FDSE method of this embodiment, which utilizes spectral and frequency domain feature fusion, is effective in balancing false alarms and missed detections in change detection. Furthermore, compared with advanced change detection methods, namely PTCD, SALA, and MMPs, the FDSE method of this embodiment achieves OAc, Kappa, F1, IoU, and OAc of 0.97504, 0.79152, 0.98076, 0.96225, and 0.96508, respectively, surpassing the corresponding accuracy metrics of the advanced methods. Furthermore, in terms of algorithm execution time, the FDSE method of this embodiment requires more time than traditional methods, but significantly less time than advanced detection methods. The experimental results fully demonstrate the effectiveness and feasibility of the FDSE method of this embodiment in terms of detection accuracy and time consumption.

[0118] Table 2 shows the OA, AA, Kappa, F1, IoU and time consumption indicators obtained by CVA, PCA-CVA, IRMAD, PTCD, SALA, MMPs and FDSE methods on the farm hyperspectral remote sensing image dataset.

[0119] Table 2

[0120]

[0121] As shown in Table 2, the FDSE method of this embodiment achieves the highest accuracy for AA, Kappa, F1, IoU, and OA, respectively, exceeding the next highest accuracy by 0.01985, 0.05653, 0.01474, 0.02798, and 0.02187. The results in Table 2 demonstrate that, from a comprehensive perspective of detection accuracy and time consumption, the FDSE method of this embodiment balances false alarms and missed detections in the hyperspectral remote sensing image change detection task by fusing the gradient correlation spectral absolute distance with the global significance enhancement results, significantly improving the overall recognition and detection accuracy of change detection.

[0122] in addition, Figure 2-3The detection results of different change detection methods on river and farm hyperspectral remote sensing image datasets are given respectively. Among them, (2a) is the river hyperspectral remote sensing image of the previous phase, (2b) is the river hyperspectral remote sensing image of the next phase, (2c) is the ground truth map of the river hyperspectral remote sensing image, (2d) is the detection result of the CVA method on the river hyperspectral remote sensing image dataset, (2e) is the detection result of the PCA-CVA method on the river hyperspectral remote sensing image dataset, (2f) is the detection result of the IRMAD method on the river hyperspectral remote sensing image dataset, (2g) is the detection result of the PTCD method on the river hyperspectral remote sensing image dataset, (2h) is the detection result of the SALA method on the river hyperspectral remote sensing image dataset, (2i) is the detection result of the MMPs method on the river hyperspectral remote sensing image dataset, and (2j) is the detection result of the FDSE method on the river hyperspectral remote sensing image dataset. Results: (3a) is the hyperspectral remote sensing image of farmland in the previous phase, (3b) is the hyperspectral remote sensing image of farmland in the next phase, (3c) is the ground truth map of the hyperspectral remote sensing image of farmland, (3d) is the detection result of the CVA method in the hyperspectral remote sensing image dataset of farmland, (3e) is the detection result of the PCA-CVA method in the hyperspectral remote sensing image dataset of farmland, (3f) is the detection result of the IRMAD method in the hyperspectral remote sensing image dataset of farmland, (3g) is the detection result of the PTCD method in the hyperspectral remote sensing image dataset of farmland, (3h) is the detection result of the SALA method in the hyperspectral remote sensing image dataset of farmland, (3i) is the detection result of the MMPs method in the hyperspectral remote sensing image dataset of farmland, and (3j) is the detection result of the FDSE method in the hyperspectral remote sensing image dataset of farmland.

[0123] Depend on Figure 2 It can be seen that the five comparisons of CVA, PCA-CVA, PTCD, SALA, and MMPs all showed false alarms in different areas to varying degrees, while the IRMAD method had serious missed detections. In contrast, the FDSE method of this embodiment achieved a better balance between false alarms and missed detections. Therefore, from the perspective of the visual binarization results, it is further verified that the FDSE method of this embodiment has a better performance than traditional and latest detection methods. In addition, Figure 3 It can be seen that CVA, PTCD, SALA, and MMPs methods all have varying degrees of false alarms, while PCA-CVA and IRMAD have varying degrees of missed detections. In contrast, the FDSE method of this embodiment also achieves a balance between false alarms and missed detections, achieving better overall detection accuracy. Figure 2-3 This further fully demonstrates the robustness and generalization of the FDSE method in this embodiment in the task of change detection in hyperspectral remote sensing images.

[0124] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting changes in dual-temporal hyperspectral remote sensing images, characterized in that: The following steps are involved: spectrum Information: First, a pair of dual-phase hyperspectral remote sensing images are fused to obtain the fusion features of the two images. Then, preliminary delimitation values and comprehensive weights are obtained based on the image fusion features. Finally, the preliminary delimitation values are converted into reference delimitation values based on the comprehensive weights. Image fusion features Fusion features with images The edges are filled in a symmetrical manner to ensure that the edges and corner pixels have eight-connected regions; By calculation and The spectral absolute distance of the center pixel pair of the eight connected regions is used to obtain a preliminary boundary value. The spectral absolute distance is calculated by the following formula: Among them, N b Indicates the number of bands of the hyperspectral remote sensing image, express The spectral reflectance value of the i-th band of the corresponding b-th pixel, express The corresponding spectral reflectance value of the i-th band of the b-th pixel; By introducing the spectral cosine distance function, the similarity between each pixel pair is quantified: The weight scaling factor is constructed by using the gradient correlation of the pixels in the eight connected regions of the central pixel: The comprehensive weight is calculated based on the weight scaling factor: s b =ω b ·(1-g b ) The reference limit value is calculated by the following formula: D=s b ·dist b ,b=1,2,...,N in, are the universal symbols of the two-phase image pixels, express The transpose of express The transpose of ω b,c It represents the cosine similarity between the pixel pairs of the eight-connected regions in two time phases. Indicates the cosine similarity value from the second pixel to the last pixel. Represents pixels The cth neighboring pixel of Represents pixels The cth neighborhood pixel of , c represents the counting variable of the number of pixels in the eight-connected region, c = 1, 2, ..., 8; ω b Represents ω b,c The average weight of In terms of spatial information: First, the first principal component difference image of the dual-temporal hyperspectral remote sensing image is obtained. Then, the amplitude spectrum and phase spectrum of the first principal component difference image in the frequency domain are obtained. The amplitude spectrum is filtered using Gaussian filtering to suppress non-changing attributes. Finally, the frequency domain significance level is obtained through inverse Fourier transform. Information fusion: performing a convolution operation on the frequency domain significance level and the reference threshold value to obtain a comprehensive threshold value, and outputting the change detection result of the dual-phase hyperspectral remote sensing image in combination with a preset false alarm threshold.

2. The method for detecting changes in dual-temporal hyperspectral remote sensing images according to claim 1, wherein: The principal component analysis algorithm is used to extract the first principal component features of each hyperspectral remote sensing image, and the image difference operation is used to obtain the structural and texture difference features between the first principal component features to obtain the first principal component difference image.

3. The method for detecting changes in dual-temporal hyperspectral remote sensing images according to claim 2, wherein: The first principal component difference image G is mapped to the frequency domain using Fourier transform, and its amplitude spectrum is extracted from the image spectrum features: A(f)=R(F(G)) And perform a log transform on A(f) to suppress the upper limit of the amplitude of the non-changing attribute and amplify the difference in the amplitude of the changing attribute: L(f)=log(A(f)) Where F(·) represents Fourier transform, and R(·) represents the modulus of the complex features in the frequency domain.

4. The method for detecting changes in dual-temporal hyperspectral remote sensing images according to claim 3, wherein: The Gaussian filter is a two-dimensional Gaussian filter: The filtered output is: G(f)=L(f)*K(x,y) Among them, x c with y c Represents the two-dimensional coordinate position of the center pixel, oh g =2×ceil(2σ)+1, ω g represents the local window size, x and y represent the eight-connected region pixels (ω g ×ω g -1), σ represents the standard deviation of the two-dimensional Gaussian filter, and ceil(·) represents rounding up to an integer.

5. The method for detecting changes in dual-temporal hyperspectral remote sensing images according to claim 3, wherein: It also includes phase spectrum feature extraction of frequency domain representation features: P(f)=S(F(G)) in, Represents the function of taking the imaginary part of the complex features in the frequency domain.

6. The method for detecting changes in dual-temporal hyperspectral remote sensing images according to claim 5, characterized in that: Perform inverse Fourier transform using G(f) and P(f) to obtain the frequency domain significance level: G1=F -1 (exp(G(f)+j·P(f))) 2 Among them, F -1 (·) represents the inverse Fourier transform, and j is the imaginary number symbol.

7. The method for detecting changes in dual-temporal hyperspectral remote sensing images according to claim 1, wherein: The comprehensive threshold value is calculated using the following formula: V c =Norm(D)*Norm(G1) Where Norm(·) represents the maximum and minimum normalization function, D represents the reference limit value, and G1 represents the frequency domain significance level.

8. The method for detecting changes in dual-temporal hyperspectral remote sensing images according to claim 1, wherein: The preset false alarm threshold is β t =2.5×10 -2 .

Images