Hyperspectral image band selection method for regional multi-image fusion and community detection

Abstract

The invention discloses a hyperspectral image band selection method for regional multi-image fusion and community detection, relates to the technical field of hyperspectral image processing, and solves the problem that the existing band selection method only considers the related features of a single region to select the band of the single region. And the accuracy of the finally obtained selected wave band is not high in wave band selection. Comprising the following steps: 1, obtaining an original hyperspectral image cube, and generating a plurality of uniform regions based on the original hyperspectral image cube; 2, constructing a plurality of wave band similar graphs of each uniform region; 3, updating the wave band similar graph corresponding to each uniform region according to the influence relationship among the uniform regions to obtain an updated similar graph; 4, generating an optimal similar graph according to the updated similar graph; 5, performing community detection on the optimal similar graph to obtain a plurality of final selected wavebands; and the final waveband selection accuracy is improved.

Description

Technical Field

[0001] The present application belongs to the field of hyperspectral image processing technology, and relates to hyperspectral image processing technology, and specifically to a hyperspectral image band selection method for regional multi-image fusion and community detection. Background Art

[0002] Hyperspectral technology is widely used in many fields, including agricultural monitoring, environmental monitoring, mineral exploration, remote sensing image analysis, medical diagnosis, pharmaceutical quality control, and forensic medicine. By acquiring the spectral information of the target object, it can achieve accurate identification and classification of different substances.

[0003] At present, there are two main methods for band dimensionality reduction in traditional methods for hyperspectral image band selection, namely feature extraction and band selection. The former usually projects the original high-dimensional spatial data into a new low-dimensional feature space according to certain standards, with the aim of learning more discriminative features in the new feature space. For the latter, a band subset is selected from the original data. The selected bands are required to be unique and representative, and can retain the inherent band properties and physical meanings; band selection can provide rich spectral information and better characterize the features; traditional band selection methods mainly consider the hyperspectral band selection problem at the data level. Since the corresponding regions have mutual influences, simply considering the relevant features of different regions from the data level makes the band selection ultimately obtain the selected bands with low accuracy. Therefore, a hyperspectral image band selection method with regional multi-image fusion and community detection is needed. Summary of the invention

[0004] The present application aims to solve at least one of the technical problems existing in the prior art; to this end, the present application proposes a hyperspectral image band selection method for regional multi-image fusion and community detection, which is used to solve the technical problem that the existing band selection method only considers the relevant characteristics of a single region to select the band of a single region, resulting in the band selection ultimately obtaining a low accuracy of the selected band.

[0005] To achieve the above-mentioned purpose, the first aspect of the present application provides a hyperspectral image band selection method for regional multi-image fusion and community detection, comprising the following steps:

[0006] Step 1: Obtain the original hyperspectral image cube and generate several uniform areas based on the original hyperspectral image cube;

[0007] Step 2: Construct several band similarity graphs of several uniform areas;

[0008] Step 3: According to the influence relationship between each uniform area, the band similarity map corresponding to each uniform area is updated to obtain an updated similarity map;

[0009] Step 4: Generate the optimal similarity graph based on the updated similarity graph;

[0010] Step 5: Perform community detection on the optimal similarity graph to obtain several final selected bands.

[0011] The present application obtains a number of uniform regions by dividing the original hyperspectral image cube, and generates a band similarity map of the corresponding uniform region according to the characteristics of the spectral values ​​of each band of each uniform region; then updates the band similarity map according to the relationship between different regions to obtain an updated similarity map, fuses the updated similarity maps of each uniform region to obtain a final optimal similarity map, performs community detection on the optimal similarity map to finally obtain a number of final selected bands, thereby achieving dimensionality reduction of the original hyperspectral image cube; in the process of generating the optimal similarity map, the mutual influence between different uniform regions is taken into account to increase the accuracy of the final band selection.

[0012] Preferably, step one comprises the following steps:

[0013] S11: Get the original hyperspectral image cube;

[0014] S12: using principal component analysis to process the original hyperspectral image cube to obtain the first principal component;

[0015] S13: using superpixel segmentation to segment the first principal component into a plurality of superpixels;

[0016] S14: Treat each superpixel as a uniform region.

[0017] Preferably, the step of constructing a plurality of band similarity graphs of a plurality of uniform regions comprises:

[0018] Get the spectral values ​​of each pixel point in the uniform area in different bands, and integrate the spectral values ​​of each pixel point in the same band into the band feature vector of the corresponding uniform area, and mark the band feature vectors of the i-th band and the j-th band of the same uniform area as and p is the number of the uniform region;

[0019] By formula Calculate the similarity S between the band feature vectors corresponding to the i-th band and the j-th band in the uniform area numbered p p (i, j); where express The kth nearest neighbor set of express The kth nearest neighbor set of ; σ is the width of the Gaussian kernel function;

[0020] The similarities between the corresponding band feature vectors of several bands in the uniform area numbered p are obtained, and each similarity is integrated into a band similarity graph.

[0021] Preferably, step three comprises the following steps:

[0022] S31: Obtaining the band similarity graph of each uniform area and the band feature vector corresponding to each band, as well as the central pixel coordinates of each uniform area;

[0023] S32: integrating the band feature vectors corresponding to each band in the same uniform region into a feature vector corresponding to the uniform region;

[0024] S33: Select any uniform area as the update area; the remaining uniform areas are recorded as the affected areas;

[0025] S34: generating an influence weight of the influence area on the update area according to the feature vector and the center pixel coordinates of the influence area and the feature vector and the center pixel coordinates of the update area;

[0026] S35: Obtaining the influence weights of each influence region on the update region, and updating the update region according to the influence weights to obtain a corresponding update similarity graph;

[0027] S36: taking each uniform region as an update region in turn and calculating its corresponding update similarity graph.

[0028] Preferably, generating the influence weight of the influence area on the update area according to the feature vector and the central pixel coordinates of the influence area and the feature vector and the central pixel coordinates of the update area includes:

[0029] Get the feature vector x of the update region p p and the center pixel coordinate I P , and the eigenvector x of any influence region q q and the center pixel coordinate I q ; Through the formula:

[0030]

[0031] Calculate the influence weight ω of the influence area q on the update area p pq , where σ sv and σ sl are the spectral sensitivity and spatial sensitivity of the control elements respectively; N is the number of uniform regions; It is a double summation, that is, when p=1, q is summed from 1 to N; when p=2, q is summed from 1 to N; until p=N;

[0032] Preferably, updating the update region according to a number of influence weights to obtain a corresponding update similarity graph includes:

[0033] Get the influence weight of each influence area on the update area through the formula:

[0034]

[0035] Calculate the similarity graph of the band corresponding to the updated area p after t+1 diffusions in, is the similarity graph after t times of diffusion of the band similarity graph corresponding to the affected area q; S p is to update the band similarity of region p; α is an adjustment parameter, the specific value is set according to experience, and α in this embodiment is 0.7; is the matrix parameter corresponding to the update area p; Transpose the matrix parameters corresponding to the update area p;

[0036] When the set diffusion times are reached, the corresponding similarity graph is used as the updated similarity graph of the corresponding area.

[0037] Preferably, generating an optimal similarity graph according to the updated similarity graph comprises:

[0038] Get the updated similarity graph corresponding to each uniform area

[0039] By formula Calculate the optimal similarity graph S * ;in, It is the updated similarity graph obtained after T times diffusion of the band similarity graph corresponding to the uniform region numbered p.

[0040] After the original band similarity graph is diffused for a set number of times T, a stable updated similarity graph can be obtained.

[0041] Preferably, the step five comprises the following steps:

[0042] S51: Obtaining the optimal similarity graph and the required number of bands;

[0043] S52: performing community detection on the optimal similarity graph to obtain several initial band subsets;

[0044] S53: Obtain the number of band subsets of the initial band subset, and generate the final selected band according to the number of band subsets and the required number of bands.

[0045] Preferably, generating the final selected bands according to the number of band subsets and the required number of bands includes:

[0046] Mark the number of subsets of the initial wavelength band subset as M; mark the required number of wavelength bands as K;

[0047] When M < K, perform community detection and division on the initial wavelength band subset again to obtain several secondary wavelength band subsets, obtain the number of subsets of the secondary wavelength band subset, and generate the final selected wavelength bands according to the number of subsets of the wavelength band and the required number of wavelength bands;

[0048] When M > K, merge the appropriate wavelength band subsets, and obtain the number of subsets of the merged wavelength band subsets; generate the final selected wavelength bands according to the number of subsets of the wavelength band and the required number of wavelength bands;

[0049] When M = K, select the most suitable wavelength band from each wavelength band subset, and use this wavelength band as the final selected wavelength band.

[0050] Compared with the prior art, the beneficial effects of this application are:

[0051] 1. In this application, several uniform regions are obtained by dividing the original hyperspectral image cube, and a band similarity map corresponding to each uniform region is generated according to the characteristics of the spectral values of each band in each uniform region; then, the band similarity map is updated according to the mutual relationship between different regions to obtain an updated similarity map, and the updated similarity maps of each uniform region are fused to obtain the final optimal similarity map. Community detection is performed on the optimal similarity map to finally obtain several final selected wavelength bands, realizing the dimensionality reduction of the original hyperspectral image cube; during the generation process of the optimal similarity map, the mutual influence between different uniform regions is considered to increase the accuracy of the final wavelength band selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of this application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0053] Figure 1 It is a schematic diagram of the steps of the wavelength band selection method of this application;

[0054] Figure 2 It is a framework structure diagram of the hyperspectral image wavelength band selection method of this application;

[0055] Figure 3 It is a schematic diagram of the overall process of the wavelength band selection method in this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The technical solution of the present application will be described clearly and completely in conjunction with the embodiments below. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0057] See also Figure 1-Figure 3 The first aspect of the present application provides a method for selecting a hyperspectral image band for regional multi-image fusion and community detection, comprising the following steps:

[0058] Step 1: Obtain the original hyperspectral image cube and generate several uniform areas based on the original hyperspectral image cube;

[0059] Step 2: construct several band similarity graphs of several uniform regions; the band similarity graph is a similarity graph composed of the similarity between the spectral values ​​of different bands in the uniform region;

[0060] Step 3: updating the band similarity graph corresponding to each uniform area according to the influence relationship between each uniform area to obtain an updated similarity graph; the updated similarity graph is a new similarity graph obtained by updating the original band similarity graph according to the mutual influence between different uniform areas;

[0061] Step 4: Generate the optimal similarity graph based on the updated similarity graph;

[0062] Step 5: Perform community detection on the optimal similarity graph to obtain several final selected bands.

[0063] In this embodiment, the original hyperspectral image cube is divided into several uniform areas, and a band similarity map of the corresponding uniform area is generated according to the characteristics of the spectral values ​​of each band of each uniform area; the band similarity map is then updated according to the relationship between different areas to obtain an updated similarity map, the updated similarity maps of each uniform area are fused to obtain a final optimal similarity map, community detection is performed on the optimal similarity map to finally obtain several final selected bands, thereby achieving dimensionality reduction of the original hyperspectral image cube; in the process of generating the optimal similarity map, the mutual influence between different uniform areas is taken into account to increase the accuracy of the final band selection.

[0064] Step one includes the following steps:

[0065] S11: Get the original hyperspectral image cube;

[0066] S12: using principal component analysis to process the original hyperspectral image cube to obtain the first principal component;

[0067] S13: using superpixel segmentation to segment the first principal component into a plurality of superpixels;

[0068] S14: Treat each superpixel as a uniform region.

[0069] Since different regions often have different reflectance characteristics, different regions should be treated differently in the original hyperspectral image cube; obtain an original hyperspectral image cube of a given size of H×W×B, where H is the image height of the original hyperspectral image cube, and W is the image width of the original hyperspectral image cube; B is the band composition of the original hyperspectral image cube; in order to reduce the subsequent computational cost of obtaining the main information of the original hyperspectral image cube, this embodiment first uses principal component analysis to obtain the first principal component of the original hyperspectral image; the principal component analysis step of the hyperspectral image is a prior art and will not be repeated here; the first principal component is segmented into superpixels to obtain a plurality of superpixels; specifically, the first principal component is marked as H f ; Through the formula Divide the first principal component into N superpixels;

[0070] Among them, R p is the superpixel numbered p; R q is the superpixel numbered q; st means that the previous formula x needs to satisfy the following conditions.

[0071] Constructing several band similarity maps of several uniform regions, including: obtaining the spectral values ​​of each pixel point in the uniform region in different bands, integrating the spectral values ​​of each pixel point in the same band into the band feature vector of the corresponding uniform region, and marking the band feature vectors of the i-th band and the j-th band of the same uniform region as and p is the number of the uniform region;

[0072] By formula Calculate the similarity S between the band feature vectors corresponding to the i-th band and the j-th band in the uniform area numbered p p (i, j); where express The kth nearest neighbor set of express The kth nearest neighbor set of ; σ is the width of the Gaussian kernel function;

[0073] The similarities between the corresponding band feature vectors of several bands in the uniform area numbered p are obtained, and each similarity is integrated into a band similarity graph.

[0074] This embodiment generates a band similarity map of the uniform region through the spectral value features of different bands in the same uniform region, thereby ensuring that the band similarity map contains the features of the uniform region.

[0075] Step three includes the following steps:

[0076] S31: Obtaining the band similarity graph of each uniform area and the band feature vector corresponding to each band, as well as the central pixel coordinates of each uniform area;

[0077] S32: integrating the band feature vectors corresponding to each band in the same uniform region into a feature vector corresponding to the uniform region;

[0078] S33: Select any uniform area as the update area; the remaining uniform areas are recorded as the affected areas;

[0079] S34: Generate an influence weight of the influence area on the update area according to the feature vector and the center pixel coordinates of the influence area and the feature vector and the center pixel coordinates of the update area; the weight is calculated by the following formula:

[0080]

[0081] Among them, ω pq I is the influence weight of the influence area q on the update area p; P is the center pixel coordinate of the update area p; q is the center pixel coordinate of the affected area q; σ sv and σ sl are the spectral sensitivity and spatial sensitivity of the control elements respectively; N is the number of uniform regions; is a double summation, that is, when p = 1, q is summed from 1 to N; when p = 2, q is summed from 1 to N; until p = N; the first term in the above formula measures the spectral similarity between two superpixels, that is, the spectral similarity between two uniform regions; the second term measures the spatial distance between two superpixels, that is, the spatial distance between two uniform regions;

[0082] S35: Obtaining the influence weights of each influence region on the update region, and updating the update region according to the influence weights to obtain a corresponding update similarity graph;

[0083] The conventional diffusion method is as follows: Get an M×M similarity graph S, and according to the popular sorting model, use the formula Realize the optimization of the similarity graph S to obtain a new similarity graph Where μ is the regularization parameter, and μ>0; D is a diagonal matrix, and the value of the diagonal element in the i-th row and i-th column is given by the formula Calculated; S ij is the similarity between data point i and data point j, that is, the element value of the i-th row and j-th column in the similarity graph; S pq is the similarity between data point q and data point p, that is, the element value of the pth row and qth column in the similarity graph; is the similarity between data point i and data point p in the new similarity graph, that is, the element value of the i-th row and p-th column in the new similarity graph; is the similarity between data point j and data point q in the new similarity graph, that is, the element value of the jth row and qth column in the new similarity graph; is the similarity between data point i and data point j in the new similarity graph, that is, the element value of the i-th row and j-th column in the new similarity graph;

[0084] The above formula consists of two items. The first item is to compare the local and global. If the data point i and data point j on the original similarity graph S are similar, and the data point q and data point p are also similar, then the similarity in the new similarity graph is and should also be similar; the second term can be expressed as a fitting term, which regulates that the final similarity graph will not change much from the original similarity graph;

[0085] The above formula has a closed solution Where α = 1 / (1+μ); vec(·) is an operator that vectorizes the input matrix by stacking its columns one by one; vec(·) -1 is the inverse operator of vec(·); A is the matrix parameter, through the formula A T Represents the transpose of the matrix parameter A; iterate the closed solution to obtain the final expression However, this expression can only update a single similarity graph, and its applicability is low. This embodiment improves the above formula to obtain a formula for calculating the updated similarity graph:

[0086]

[0087] in, is the similarity graph after t+1 diffusion of the band similarity graph corresponding to the update area p; is the similarity graph after t times of diffusion of the band similarity graph corresponding to the affected area q; S p is to update the band similarity of region p; α is an adjustment parameter, the specific value is set according to experience, and α in this embodiment is 0.7; To update the matrix parameters corresponding to region p, use the formula in the above inference and formula It is easy to deduce the matrix parameters Specifically, the feature vector of the update region is recorded as S and substituted into the above formula Obtain the matrix D, and then substitute the matrix D and S into Find the matrix parameters Transpose the matrix parameters corresponding to the update region p; when the set diffusion times are reached, the corresponding similarity graph is used as the updated similarity graph of the corresponding region;

[0088] This embodiment updates the overly high similarity graphs corresponding to the multiple uniform regions by making full use of the complementarity between the different similarity graphs in the different uniform regions;

[0089] S36: taking each uniform region as an update region in turn and calculating its corresponding update similarity graph.

[0090] Since the corresponding band similarity map obtained for the features in a single uniform area only considers the relevant features of the corresponding uniform area, this will cause the obtained band similarity map to be inaccurate; therefore, this embodiment diffuses the band similarity maps of each uniform area multiple times according to the implicit information between different band similarity maps to achieve the update of the band similarity map and obtain an updated similarity map, so as to ensure that the updated similarity map can better reflect the features of the corresponding uniform area, thereby ensuring that the subsequently obtained selected bands are more accurate.

[0091] At the same time, this embodiment improves the conventional diffusion method by making full use of the complementarity between different similarity graphs in different uniform regions to update the similarity graphs corresponding to multiple uniform regions; multiple updated graphs are used to generate the optimal similarity graph, and multiple graphs are fused to perform clustering-based hyperspectral band selection; instead of taking a certain hyperspectral band as a whole to calculate the similarity between band pairs, the hyperspectral bands are divided into uniform regions taking into account that different objects often have different reflective characteristics; we use a graph diffusion strategy to fuse multiple graphs to learn a consensus similarity graph that reflects the global relationship of all hyperspectral bands, which further improves the accuracy of updating the band similarity graph.

[0092] Generating an optimal similarity graph according to the updated similarity graph includes: obtaining an updated similarity graph corresponding to each uniform region;

[0093] By formula Calculate the optimal similarity graph S * ;in, It is the updated similarity graph obtained after T times diffusion of the band similarity graph corresponding to the uniform region numbered p.

[0094] The original band similarity map can be diffused for a set number of times T to obtain a stable updated similarity map; the most stable similarity maps in each region are fused to obtain the optimal similarity map, so that the optimal similarity map can better represent the characteristics of the original hyperspectral image cube.

[0095] Step five includes the following steps:

[0096] S51: Obtaining the optimal similarity graph and the required number of bands;

[0097] S52: Perform community detection on the optimal similarity graph to obtain several initial wavelength subsets;

[0098] S53: Obtain the number of wavelength subsets in the initial wavelength subsets, and generate the final selected wavelengths according to the number of wavelength subsets and the required number of wavelengths.

[0099] Specifically, in this embodiment, the Louvain community detection method is used to preliminarily partition the optimal similarity graph through module optimization to obtain several initial wavelength subsets; considering the limitations of traditional clustering methods in this embodiment, the community detection algorithm is used to replace clustering, and community detection is performed on the generated diffusion similarity graph results. When the Louvain community detection algorithm is applied to wavelength selection, it can better group wavelengths with high similarity together, effectively reducing the redundancy of wavelength selection.

[0100] Generating the final selected wavelengths according to the number of wavelength subsets and the required number of wavelengths includes: marking the number of wavelength subsets in the initial wavelength subsets as M; marking the required number of wavelengths as K;

[0101] When M < K, perform community detection and partitioning on the initial wavelength subsets again to obtain several secondary wavelength subsets, obtain the number of wavelength subsets in the secondary wavelength subsets, and generate the final selected wavelengths according to the number of wavelength subsets and the required number of wavelengths;

[0102] When M > K, merge appropriate wavelength subsets. Specifically, wavelength subsets with high similarity are merged based on wavelength subsets with the same modularity in community detection; obtain the number of wavelength subsets after merging; generate the final selected wavelengths according to the number of wavelength subsets and the required number of wavelengths;

[0103] When M = K, select the most appropriate wavelength from each wavelength subset, and the most appropriate wavelength is set according to expert experience; and use this wavelength as the final selected wavelength.

[0104] Since the number of communities after community detection is uncontrollable, the main problem to be solved is to correct the matching between the number of communities and the specified number of wavelengths; in this embodiment, the wavelength subsets are processed through the above steps to output the same number of final selected wavelengths as the required number of wavelength subsets.

[0105] Some of the data in the above formula are calculated by removing the dimension and taking their numerical values. The formula is the one closest to the actual situation obtained through software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

[0106] The working principle of this application:

[0107] The present application obtains a number of uniform regions by dividing the original hyperspectral image cube, and generates a band similarity map of the corresponding uniform region according to the characteristics of the spectral values ​​of each band of each uniform region; then updates the band similarity map according to the relationship between different regions to obtain an updated similarity map, fuses the updated similarity maps of each uniform region to obtain a final optimal similarity map, performs community detection on the optimal similarity map to finally obtain a number of final selected bands, thereby achieving dimensionality reduction of the original hyperspectral image cube; in the process of generating the optimal similarity map, the mutual influence between different uniform regions is taken into account to increase the accuracy of the final band selection.

[0108] The above embodiments are only used to illustrate the technical method of the present application and are not intended to limit it. Although the present application has been described in detail with reference to the preferred embodiments, a person of ordinary skill in the art should understand that the technical method of the present application may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present application.

Claims

1. A hyperspectral image band selection method for regional multi-image fusion and community detection, characterized in that: The following steps are involved: Step 1: Obtain the original hyperspectral image cube and generate several uniform areas based on the original hyperspectral image cube; Step 2: Construct several band similarity graphs for each uniform area; Step 3: updating the band similarity graph corresponding to each uniform area according to the influence relationship between each uniform area to obtain an updated similarity graph; Step 4: Generate the optimal similarity graph based on the updated similarity graph; Step 5: Perform community detection on the optimal similarity graph to obtain several final selected bands.

2. The method for selecting hyperspectral image bands for regional multi-image fusion and community detection according to claim 1, characterized in that: Step one includes the following steps: S11: Get the original hyperspectral image cube; S12: using principal component analysis to process the original hyperspectral image cube to obtain the first principal component; S13: using superpixel segmentation to segment the first principal component into a plurality of superpixels; S14: Treat each superpixel as a uniform region.

3. The method for selecting hyperspectral image bands for regional multi-image fusion and community detection according to claim 1, characterized in that: The step of constructing a plurality of band similarity graphs of a plurality of uniform regions comprises: Get the spectral values ​​of each pixel point in the uniform area in different bands, and integrate the spectral values ​​of each pixel point in the same band into the band feature vector of the corresponding uniform area, and mark the band feature vectors of the i-th band and the j-th band of the same uniform area as and p is the number of the uniform region; By formula Calculate the similarity S between the band feature vectors corresponding to the i-th band and the j-th band in the uniform area numbered p p (i, j); where express The kth nearest neighbor set of express The kth nearest neighbor set of ; σ is the width of the Gaussian kernel function; The similarities between the corresponding band feature vectors of several bands in the uniform area numbered p are obtained, and each similarity is integrated into a band similarity graph.

4. The method for selecting hyperspectral image bands for regional multi-image fusion and community detection according to claim 1, characterized in that: Step three includes the following steps: S31: Obtaining the band similarity graph of each uniform area and the band feature vector corresponding to each band, as well as the central pixel coordinates of each uniform area; S32: integrating the band feature vectors corresponding to each band in the same uniform region into a feature vector corresponding to the uniform region; S33: Select any uniform area as the update area; the remaining uniform areas are recorded as the affected areas; S34: generating an influence weight of the influence area on the update area according to the feature vector and the center pixel coordinates of the influence area and the feature vector and the center pixel coordinates of the update area; S35: Obtaining the influence weights of each influence region on the update region, and updating the update region according to the influence weights to obtain a corresponding update similarity graph; S36: taking each uniform region as an update region in turn and calculating its corresponding update similarity graph.

5. The method for selecting hyperspectral image bands for regional multi-image fusion and community detection according to claim 4, characterized in that: The step of generating the influence weight of the influence area on the update area according to the feature vector and the central pixel coordinates of the influence area and the feature vector and the central pixel coordinates of the update area includes: Get the feature vector x of the update region p p and the center pixel coordinate I P , and the eigenvector x of any influence region q q and the center pixel coordinate I q ; Through the formula: Calculate the influence weight ω of the influence area q on the update area p pq , where σ sv and σ sl are the spectral sensitivity and spatial sensitivity of the control elements respectively; N is the number of uniform regions.

6. The method for selecting hyperspectral image bands for regional multi-image fusion and community detection according to claim 4, characterized in that: The updating region is updated according to a number of influence weights to obtain a corresponding updated similarity graph, including: Get the influence weight of each influence area on the update area through the formula: Calculate the similarity graph of the band corresponding to the updated area p after t+1 diffusions in, is the similarity graph after t times of diffusion of the band similarity graph corresponding to the affected area q; S p is the band similarity of the updated region p; α is the adjustment parameter; is the matrix parameter corresponding to the update area p; Transpose the matrix parameters corresponding to the update area p; When the set diffusion times are reached, the corresponding similarity graph is used as the updated similarity graph of the corresponding area.

7. The method for selecting hyperspectral image bands for regional multi-image fusion and community detection according to claim 1, characterized in that: The step of generating an optimal similarity graph according to the updated similarity graph comprises: Get the updated similarity graph corresponding to each uniform area By formula Calculate the optimal similarity graph S * ;in, It is the updated similarity graph obtained after T times diffusion of the band similarity graph corresponding to the uniform region numbered p.

8. The method for selecting hyperspectral image bands for regional multi-image fusion and community detection according to claim 1, characterized in that: The step five comprises the following steps: S51: Obtaining the optimal similarity graph and the required number of bands; S52: performing community detection on the optimal similarity graph to obtain several initial band subsets; S53: Obtain the number of band subsets of the initial band subset, and generate the final selected band according to the number of band subsets and the required number of bands.

9. The method for selecting hyperspectral image bands for regional multi-image fusion and community detection according to claim 8, characterized in that: The generating the final selected bands according to the number of band subsets and the required number of bands comprises: The number of band subsets of the initial band subset is marked as M; the number of required bands is marked as K; When M < K, the initial wavelength subset is subjected to community detection and division again to obtain a number of secondary wavelength subsets. The number of wavelength subsets in the secondary wavelength subsets is obtained, and the final selected wavelengths are generated according to the number of wavelength subsets and the required number of wavelengths; When M > K, the appropriate wavelength subsets are merged to obtain the number of wavelength subsets after merging; the final selected wavelengths are generated according to the number of wavelength subsets and the required number of wavelengths; When M = K, the most suitable wavelength is selected from each wavelength subset, and this wavelength is used as the final selected wavelength.

Images