CN107220661B - Spectrum waveband selection method based on multi-mode fusion - Google Patents

Spectrum waveband selection method based on multi-mode fusion Download PDF

Info

Publication number
CN107220661B
CN107220661B CN201710342008.XA CN201710342008A CN107220661B CN 107220661 B CN107220661 B CN 107220661B CN 201710342008 A CN201710342008 A CN 201710342008A CN 107220661 B CN107220661 B CN 107220661B
Authority
CN
China
Prior art keywords
band
wave band
wave
sequence
bands
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201710342008.XA
Other languages
Chinese (zh)
Other versions
CN107220661A (en
Inventor
李照奎
黄林
赵亮
王岩
刘翠微
张德园
石祥滨
徐一民
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenyang Aerospace University
Original Assignee
Shenyang Aerospace University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenyang Aerospace University filed Critical Shenyang Aerospace University
Priority to CN201710342008.XA priority Critical patent/CN107220661B/en
Publication of CN107220661A publication Critical patent/CN107220661A/en
Application granted granted Critical
Publication of CN107220661B publication Critical patent/CN107220661B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
    • G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2323—Non-hierarchical techniques based on graph theory, e.g. minimum spanning trees [MST] or graph cuts

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Health & Medical Sciences (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Discrete Mathematics (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • General Health & Medical Sciences (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Image Analysis (AREA)
  • Investigating Or Analysing Materials By Optical Means (AREA)

Abstract

The invention discloses a spectrum band selection method based on multi-mode fusion, which comprises the following steps: receiving a hyperspectral image sample and category label information; fusing spatial features and textural features in a plurality of spatial neighborhoods, and sequencing all wave bands according to correlation by using a correlation measurement criterion of the wave bands to obtain a wave band sequence 1; arranging all wave bands respectively based on the maximum value, the average value and the variance value of each row in the correlation matrix to obtain wave band sequences 2-4; weighting and summing the sequencing serial numbers of each wave band in the four wave band sequences to obtain the final sequencing serial number of each wave band; and taking the n wave bands with the final sequencing serial numbers at the top as the selection wave bands. According to the spectrum band selection method based on multi-mode fusion, rich multi-mode information is considered at the same time, the band correlation is calculated by adopting a supervision strategy, the priority of bands is considered, the redundancy among the selected bands is also considered, and the spectrum band which is more beneficial to the classification of hyperspectral images can be obtained.

Description

Spectrum waveband selection method based on multi-mode fusion
Technical Field
The invention relates to the field of hyperspectral image processing, and particularly provides a spectrum band selection method based on multimodal fusion.
Background
The hyperspectral remote sensing image has various modal information such as textures, correlation and spectra, wherein abundant spectral information can reflect diagnostic spectral characteristics for distinguishing different substances, so that the hyperspectral remote sensing can detect more ground object information, and the cognitive ability of human beings to the objective world is greatly improved.
More in the current research is to select a relevant metric function as a metric criterion of spectral sorting, and then sort the spectral information based on the metric function, wherein the band selection based on the spectral sorting only considers the band priority of a given task, often neglects the possible redundancy between the selected bands, and regarding the problem of redundancy, the current method for removing the redundant band is to remove the maximum correlation between a certain sorted band and all other bands as a redundant band, which often results in some bands with higher priority being replaced by bands with lower priority, and no relevant criterion is given to the trade-off between the band priority and the band redundancy.
In the other partial band selection algorithms, classification accuracy is used as a target function, and after an evaluation function is determined, an effective search strategy is required to ensure band selection accuracy. The royal nation and the like apply the bee colony algorithm to the wave band selection, the Gao and the like apply the particle swarm optimization algorithm to the wave band selection method, and in addition, the introduction of the technologies such as wave band clustering, sparse non-negative matrix factorization clustering and the like is also beneficial to the improvement of the wave band selection precision.
In summary, the conventional known technologies cannot fully consider the trade-off criteria between the correlation and redundancy among the spectral bands, and cannot well select a band with larger useful information, which is more beneficial for classifying the hyperspectral images.
Disclosure of Invention
In view of this, the present invention aims to provide a spectrum band selection method based on multi-modal fusion, so as to solve the problem that the existing spectrum band selection method cannot comprehensively consider the trade-off criteria between the correlation and redundancy among spectrum bands, and cannot well select a band with larger useful information, which is more beneficial for classifying hyperspectral images.
The technical scheme provided by the invention is as follows: a spectrum band selection method based on multi-mode fusion comprises the following steps:
s1, receiving hyperspectral image samples and class label information of the samples;
s2, fusing spatial features and textural features in a plurality of spatial neighborhoods, and sequencing all the wave bands from low to high according to the correlation by using the correlation measurement criterion of the wave bands to obtain a wave band sequence 1;
s3, adjusting all received hyperspectral image samples according to the wave band sequence 1, converting the hyperspectral image samples into a two-dimensional reflectivity matrix, calculating a correlation matrix among wave bands by using the two-dimensional reflectivity matrix, and arranging all the wave bands respectively based on the maximum value, the average value and the variance value of each row in the correlation matrix to obtain a wave band sequence 2, a wave band sequence 3 and a wave band sequence 4, wherein the wave band sequence is ordered based on the maximum value, the smaller the maximum value is, the more the wave band sequence is, the sorted based on the average value is, the smaller the average value is, the more the wave band sequence is, the sorted based on the variance value is, the larger the variance value is, the more the;
s4, weighting and summing the sequence numbers of each wave band in the four wave band sequences to obtain the final sequence number of each wave band;
and S5, selecting n wave bands with the top sequence numbers as selected wave bands, wherein n is the minimum wave band number when the classification accuracy tends to be stable.
Preferably, the S2 specifically includes the following steps:
s201, firstly, extracting local texture features of each sample under each wave band under each neighborhood by using an LBP operator, wherein the local texture features of the jth sample with the category marked as c under the wave band b under the omega neighborhood are calculated according to a formula (1)
Figure GDA0002589160480000031
Sample vjHaving different reflectivities in different wavelength bands, formula (1) vb,jRepresents the sample vjReflectivity at band b，
Figure GDA0002589160480000032
Omega is a sample vjLocal neighborhood of mΩIs a sample vjThe number of neighbor samples in the outermost circle of the omega neighborhood;
s202, calculating local texture features of all samples under each category under each neighborhood under each waveband by using a formula (2), and forming texture feature vectors, wherein the local texture features of all samples marked as c under the category under the omega neighborhood under the waveband b are calculated according to the formula (2)
Figure GDA0002589160480000033
Wherein N iscRepresenting the number of samples with class mark c;
texture feature vector at band b for all samples with class label c
Figure GDA0002589160480000034
Is represented by, whereinkExpressing the kth neighborhood, wherein k expresses the number of neighborhoods;
s203, calculating a correlation coefficient rho between any two categories by using a cosine measurement method for each waveband, counting the correlation coefficient of the waveband between any two categories, and calculating a new waveband correlation measurement criterion ξ, wherein the waveband b is in the category cxAnd cyThe correlation coefficient p between the two is calculated by the formula (3), and the band correlation metric criterion ξbCalculated by formula (4), wherein C in formula (4) represents the number of categories,
Figure GDA0002589160480000041
Figure GDA0002589160480000042
and S204, sequencing the relevance ξ of all the wave bands obtained in S203, wherein the wave band with the lowest relevance is the wave band with the high selected priority, and the wave band with the lowest relevance is arranged at the forefront of the sequence to finally obtain the wave band sequence 1.
More preferably, the S3 specifically includes the following steps:
s301, rearranging the wave bands of all the marked samples according to the wave band sequence 1 to obtain a two-dimensional reflectivity matrix V ∈ RB*MWhere B is the number of bands, M is the number of samples with class mark information, and each element V in the matrix Vb,mDenotes the reflectivity, R, of the mth sample in the b-th bandB*MRepresenting a matrix of B rows and M columns;
s302, calculating correlation matrixes D ∈ R of all wave bands by using the two-dimensional reflectivity matrix VB*BWherein each element of the correlation matrix D
Figure GDA0002589160480000043
Representing the correlation between the band omega and the band, RB*BRepresenting a matrix of B rows and B columns, Vω,:Represents the value, V, of all samples at the band ω,:Values of all samples at the band;
and S303, arranging all the wave bands respectively based on the maximum value, the average value and the variance value of each row in the correlation matrix D to obtain a wave band sequence 2, a wave band sequence 3 and a wave band sequence 4, wherein the wave band sequences are ordered based on the maximum value, the smaller the maximum value is, the more the wave band sequences are, the smaller the average value is, the more the wave band sequences are, and the larger the variance value is, the more the wave band sequences are.
More preferably, the S4 is specifically as follows:
calculating the final sequence number of each band by formula (5)
Figure GDA0002589160480000051
Wherein the content of the first and second substances,
Figure GDA0002589160480000052
denotes the position number, w, of band b in the r-th sequencerRepresents the weight of the r-th sequence in all sequencesAnd (4) weighing values.
The spectrum band selection method based on multi-mode fusion provided by the invention combines spatial neighborhood information and textural features to calculate the correlation of the same band among different classes, and sorts all bands according to the correlation; calculating the maximum value, the average value and the variance value of each row in the correlation matrix through the correlation matrix, then reordering all the wave bands based on the maximum value, the average value and the variance value respectively, so as to obtain 4 ordering sequences, then weighting and summing the ordering serial numbers of each wave band in the four wave band sequences to obtain the final ordering serial number of each wave band, finally selecting n wave bands with the front serial numbers as the selection wave bands according to the requirement, and removing the following wave bands, thus finishing the effective selection of the wave bands.
According to the spectrum band selection method based on multi-mode fusion, rich multi-mode information is considered, band correlation is calculated by adopting a supervision strategy, the priority of bands is considered, the redundancy among the selected bands is also considered, the redundant bands are removed by utilizing the correlation distribution among the bands, and the spectrum bands which are more beneficial to the classification of hyperspectral images are obtained.
Drawings
The invention is described in further detail below with reference to the following figures and embodiments:
FIG. 1 is a flow chart of a method of spectral band selection based on multimodal fusion;
FIG. 2 is a detailed flowchart of step 2;
FIG. 3 is a flowchart illustrating the step 3.
Detailed Description
The invention will be further explained with reference to specific embodiments, without limiting the invention.
As shown in FIG. 1, the present invention provides a spectral band selection method based on multi-modal fusion, comprising the following steps:
s1, receiving hyperspectral image samples and class label information of the samples;
s2, fusing spatial features and textural features in a plurality of spatial neighborhoods, and sequencing all the wave bands from low to high according to the correlation by using the correlation measurement criterion of the wave bands to obtain a wave band sequence 1;
s3, adjusting all received hyperspectral image samples according to the wave band sequence 1, converting the hyperspectral image samples into a two-dimensional reflectivity matrix, calculating a correlation matrix among wave bands by using the two-dimensional reflectivity matrix, and arranging all the wave bands respectively based on the maximum value, the average value and the variance value of each row in the correlation matrix to obtain a wave band sequence 2, a wave band sequence 3 and a wave band sequence 4, wherein the wave band sequence is ordered based on the maximum value, the smaller the maximum value is, the more the wave band sequence is, the sorted based on the average value is, the smaller the average value is, the more the wave band sequence is, the sorted based on the variance value is, the larger the variance value is, the more the;
s4, weighting and summing the sequence numbers of each wave band in the four wave band sequences to obtain the final sequence number of each wave band;
and S5, selecting n wave bands with the top sequence numbers as selected wave bands, wherein n is the minimum wave band number when the classification accuracy tends to be stable.
According to the spectrum band selection method based on multi-mode fusion, rich multi-mode information is considered at the same time, band correlation is calculated by adopting a supervision strategy, the priority of bands is considered, the redundancy among the selected bands is also considered, the redundant bands are removed by utilizing the correlation distribution among the bands, and the spectrum band which is more beneficial to the classification of hyperspectral images is obtained.
As an improvement of the technical solution, the S2 specifically includes the following steps:
s201, firstly, extracting local texture features of each sample under each wave band under each neighborhood by using an LBP operator, wherein the local texture features of the jth sample with the category marked as c under the wave band b under the omega neighborhood are calculated according to a formula (1)
Figure GDA0002589160480000071
Sample vjIn different wave bandsThe same reflectance, in equation (1), vb,jRepresents the sample vjThe reflectivity at the band b of the light,
Figure GDA0002589160480000072
omega is a sample vjLocal neighborhood of mΩIs a sample vjThe number of neighbor samples in the outermost circle of the omega neighborhood;
s202, calculating local texture features of all samples under each category under each neighborhood under each waveband by using a formula (2), and forming texture feature vectors, wherein the local texture features of all samples marked as c under the category under the omega neighborhood under the waveband b are calculated according to the formula (2)
Figure GDA0002589160480000073
Wherein N iscRepresenting the number of samples with class mark c;
texture feature vector at band b for all samples with class label c
Figure GDA0002589160480000074
Is represented by, whereinkExpressing the kth neighborhood, wherein k expresses the number of neighborhoods;
s203, calculating a correlation coefficient rho between any two categories by using a cosine measurement method for each waveband, counting the correlation coefficient of the waveband between any two categories, and calculating a new waveband correlation measurement criterion ξ, wherein the waveband b is in the category cxAnd cyThe correlation coefficient p between the two is calculated by the formula (3), and the band correlation metric criterion ξbCalculated by formula (4), wherein C in formula (4) represents the number of categories,
Figure GDA0002589160480000075
Figure GDA0002589160480000081
and S204, sequencing the relevance ξ of all the wave bands obtained in S203, wherein the wave band with the lowest relevance is the wave band with the high selected priority, and the wave band with the lowest relevance is arranged at the forefront of the sequence to finally obtain the wave band sequence 1.
As an improvement of the technical solution, the S3 specifically includes the following steps:
s301, rearranging the wave bands of all the marked samples according to the wave band sequence 1 to obtain a two-dimensional reflectivity matrix V ∈ RB*MWhere B is the number of bands, M is the number of samples with class mark information, and each element V in the matrix Vb,mDenotes the reflectivity, R, of the mth sample in the b-th bandB*MRepresenting a matrix of B rows and M columns;
s302, calculating correlation matrixes D ∈ R of all wave bands by using the two-dimensional reflectivity matrix VB*BWherein each element of the correlation matrix D
Figure GDA0002589160480000082
Representing the correlation between the band omega and the band, RB*BRepresenting a matrix of B rows and B columns, Vω,:Represents the value, V, of all samples at the band ω,:Values of all samples at the band;
and S303, arranging all the wave bands respectively based on the maximum value, the average value and the variance value of each row in the correlation matrix D to obtain a wave band sequence 2, a wave band sequence 3 and a wave band sequence 4, wherein the wave band sequences are ordered based on the maximum value, the smaller the maximum value is, the more the wave band sequences are, the smaller the average value is, the more the wave band sequences are, and the larger the variance value is, the more the wave band sequences are.
As an improvement of the technical solution, the S4 is specifically as follows:
calculating the final sequence number of each band by formula (5)
Figure GDA0002589160480000083
Wherein the content of the first and second substances,
Figure GDA0002589160480000084
denotes the position number, w, of band b in the r-th sequencerRepresenting the weight value occupied by the r-th sequence in all sequences.
Example 1
S1, receiving hyperspectral image samples and class label information of the samples;
taking indian dataset as an example, the input 145 x 200 hyperspectral image samples (where the samples include 145 x 145 samples and 200 wavebands), and the category label information corresponding to the dataset are 1-16 in total, representing 16 sample categories.
S2, fusing spatial features and textural features in a plurality of spatial neighborhoods, and sequencing all the wave bands from low to high according to the correlation by using the correlation measurement rule of the wave bands to obtain a wave band sequence 1;
the method comprises the following specific steps:
s201, taking 5 neighborhoods of 3 x 3, 5 x 5, 7 x 7, 9 x 9 and 11 x 11 of each sample, firstly, inputting all samples under each wave band of the 3 x 3 neighborhoods into an LBP filter function for filtering, outputting filtered values of 145 x 145 samples under the 3 x 3 neighborhoods of each wave band, taking the wave band 1 as an example, inputting 145 x 145 pixels in the wave band 1 into the LBP filter function, and utilizing a formula to input 145 pixels in the wave band 1 into the LBP filter function
Figure GDA0002589160480000091
And formula
Figure GDA0002589160480000092
And obtaining filtered values of 145 × 145 samples of the waveband 1 in the neighborhood of 3 × 3, calculating the values of all pixels of the waveband in the neighborhood of 5 after LBP filtering in the same way, and finally outputting a 4-dimensional matrix of 145 × 5 × 200, wherein the first two dimensions represent the number of samples, the third dimension represents the number of the neighborhood, and the fourth dimension represents the total number of the waveband.
S202, local texture features of all samples in each category in each wave band in each neighborhood are calculated, wherein the texture features of all samples marked as c in the category in the wave band b in the field omega are calculated according to a formula (2)
Figure GDA0002589160480000101
Wherein N iscIndicating the number of samples of the class labeled c. Take the example of calculating the local texture features of band 1 in 3 × 3 neighborhood for all samples in category 1: firstly, calculating the sum of LBP filtered values of all samples with the category of 1 under the wave band 1 in the 3 x 3 neighborhood, and dividing the sum by the number of all samples with the category of 1 under the wave band 1 to obtain the total LBP filtered value
Figure GDA0002589160480000102
In the same way, can obtain
Figure GDA0002589160480000103
The values under these 5 neighborhoods, i.e.
Figure GDA0002589160480000104
Similarly, the feature vectors of 16 classes in band 1 in 5 neighborhoods can be calculated
Figure GDA0002589160480000105
Finally, a 16 × 5 matrix can be obtained under each wave band, the first dimension represents the number of categories, and the second dimension represents the number of neighborhoods, that is, the feature vector of each category is understood to be composed of 5 neighborhood values.
S203, for each wave band, calculating a correlation coefficient between any two categories by using a cosine measurement method, and calculating the correlation ξ between different categories of each wave band according to a correlation measurement criterion, taking the wave band 1 as an example, any two categories c under the wave band 1xAnd cyPass through
Figure GDA0002589160480000106
Calculating the correlation coefficient between them, e.g. calculating the correlation coefficient between class 1 and class 2 in band 1, i.e. calculating the correlation coefficient between them
Figure GDA0002589160480000107
By passing
Figure GDA0002589160480000108
The correlation metric criteria to calculate ξ for each band.
And S204, sequencing the relevance ξ of all the wave bands obtained in S203, wherein the wave band with the lowest relevance is the wave band with the high selected priority, and the wave band with the lowest relevance is arranged at the forefront of the sequence to finally obtain the wave band sequence 1.
S3, adjusting all received hyperspectral image samples according to the wave band sequence 1 and converting the hyperspectral image samples into a two-dimensional reflectivity matrix V ∈ RB*MWhere B is the number of bands, M is the number of samples with class mark information, and each element V in the matrix Vb,mDenotes the reflectivity, R, of the mth sample in the b-th bandB*MRepresenting a matrix of B rows and M columns; the method comprises the following specific steps:
s301, in indian data set, adjusting original image samples according to the wave band sequence 1, and converting a three-dimensional matrix into a two-dimensional matrix, namely V ∈ RB*MB is the number of bands 200, and M is the number of samples with category label information 145 × 145 — 21025.
S302, utilizing the two-dimensional reflectivity matrix V ∈ R200*21025Calculating the correlation matrix D ∈ R of all the wave bands200 *200Wherein each element of the correlation matrix D
Figure GDA0002589160480000111
Representing the correlation between the band omega and the band.
And S303, arranging all the wave bands respectively based on the maximum value, the average value and the variance value of each row in the correlation matrix D to obtain a wave band sequence 2, a wave band sequence 3 and a wave band sequence 4, wherein the wave band sequences are ordered based on the maximum value, the smaller the maximum value is, the more the wave band sequences are, the smaller the average value is, the more the wave band sequences are, and the larger the variance value is, the more the wave band sequences are.
S4, using formula
Figure GDA0002589160480000112
And calculating the final sequencing serial number of each waveband.
After the above operations, the final position of band 1 at 187 in band sequence 1, 65 in band sequence 2, 89 in band sequence 3, 112 bits in band sequence 4, and band 1 is calculated by cross-validation a plurality of times, and w is taken as the one with higher accuracyrThe weight of (2). After the final sequence is obtained, through the curve trend of the accuracy, it is found that the accuracy basically tends to be flat when the first 120 wave bands are taken, so that the value of the wave band n is 120 for the data set.
Example 2
The method comprises the steps of selecting 80% of samples of each category on an indian data set for training, processing image samples by the method, finally classifying by an SVM classifier, wherein the accuracy of classification is 5% higher than that of the method of selecting all bands and then using the SVM classifier, so that the effectiveness of the method in the embodiment 1 can be proved, and 120 bands selected can be bands with larger useful information and better beneficial to classification of hyperspectral images.
The embodiments of the present invention have been written in a progressive manner with emphasis placed on the differences between the various embodiments, and similar elements may be found in relation to each other.
The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.

Claims (2)

1. The method for selecting the spectral band based on the multi-mode fusion is characterized by comprising the following steps of:
s1, receiving hyperspectral image samples and class label information of the samples;
s2, fusing spatial features and textural features in a plurality of spatial neighborhoods, and sequencing all the wave bands from low to high according to the correlation by using the correlation measurement criterion of the wave bands to obtain a wave band sequence 1;
s3, adjusting all received hyperspectral image samples according to the wave band sequence 1, converting the hyperspectral image samples into a two-dimensional reflectivity matrix, calculating a correlation matrix among wave bands by using the two-dimensional reflectivity matrix, and arranging all the wave bands respectively based on the maximum value, the average value and the variance value of each row in the correlation matrix to obtain a wave band sequence 2, a wave band sequence 3 and a wave band sequence 4, wherein the wave band sequence is ordered based on the maximum value, the smaller the maximum value is, the more the wave band sequence is, the sorted based on the average value is, the smaller the average value is, the more the wave band sequence is, the sorted based on the variance value is, the larger the variance value is, the more the;
s4, weighting and summing the sequence numbers of each wave band in the four wave band sequences to obtain the final sequence number of each wave band;
s5, selecting n wave bands with the top sequence numbers as selected wave bands, wherein n is the minimum wave band number when the classification accuracy tends to be stable;
wherein, the S2 specifically includes the following steps:
s201, firstly, extracting local texture features of each sample under each wave band under each neighborhood by using an LBP operator, wherein the local texture features of the jth sample with the category marked as c under the wave band b under the omega neighborhood are calculated according to a formula (1)
Figure FDA0002589160470000011
Sample vjHaving different reflectivities in different wavelength bands, formula (1) vb,jRepresents the sample vjThe reflectivity at the band b of the light,
Figure FDA0002589160470000012
omega is a sample vjLocal neighborhood of mΩIs a sample vjThe number of neighbor samples in the outermost circle of the omega neighborhood;
s202, calculating local texture features of all samples under each category under each neighborhood under each waveband by using a formula (2), and forming texture feature vectors, wherein the local texture features of all samples marked as c under the category under the omega neighborhood under the waveband b are calculated according to the formula (2)
Figure FDA0002589160470000021
Wherein N iscRepresenting the number of samples with class mark c;
texture feature vector at band b for all samples with class label c
Figure FDA0002589160470000022
Is represented by, whereinkExpressing the kth neighborhood, wherein k expresses the number of neighborhoods;
s203, calculating a correlation coefficient rho between any two categories by using a cosine measurement method for each waveband, counting the correlation coefficient of the waveband between any two categories, and calculating a new waveband correlation measurement criterion ξ, wherein the waveband b is in the category cxAnd cyThe correlation coefficient p between the two is calculated by the formula (3), and the band correlation metric criterion ξbCalculated by formula (4), wherein C in formula (4) represents the number of categories,
Figure FDA0002589160470000023
Figure FDA0002589160470000024
s204, sequencing the relevance ξ of all the wave bands obtained in S203, wherein the wave band with the lowest relevance is the wave band with the high priority, and the wave band with the lowest relevance is arranged at the forefront of the sequence to finally obtain a wave band sequence 1;
wherein, the S3 specifically includes the following steps:
s301, rearranging the wave bands of all the marked samples according to the wave band sequence 1 to obtain a two-dimensional reflectivity matrix V ∈ RB *MWhere B is the number of bands, M is the number of samples with class mark information, and each element V in the matrix Vb,mDenotes the reflectivity, R, of the mth sample in the b-th bandB*MRepresenting a matrix of B rows and M columns;
s302, calculating correlation matrixes D ∈ R of all wave bands by using the two-dimensional reflectivity matrix VB*BWherein each element of the correlation matrix D
Figure FDA0002589160470000031
Representing the correlation between the band omega and the band, RB*BRepresenting a matrix of B rows and B columns, Vω,:Represents the value, V, of all samples at the band ω,:Values of all samples at the band;
and S303, arranging all the wave bands respectively based on the maximum value, the average value and the variance value of each row in the correlation matrix D to obtain a wave band sequence 2, a wave band sequence 3 and a wave band sequence 4, wherein the wave band sequences are ordered based on the maximum value, the smaller the maximum value is, the more the wave band sequences are, the smaller the average value is, the more the wave band sequences are, and the larger the variance value is, the more the wave band sequences are.
2. The method for selecting spectral bands based on multimodal fusion according to claim 1, wherein: the S4 is specifically as follows:
calculating the final sequence number of each band by formula (5)
Figure FDA0002589160470000032
Wherein the content of the first and second substances,
Figure FDA0002589160470000033
denotes the position number, w, of band b in the r-th sequencerRepresenting the weight value occupied by the r-th sequence in all sequences.
CN201710342008.XA 2017-05-16 2017-05-16 Spectrum waveband selection method based on multi-mode fusion Active CN107220661B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710342008.XA CN107220661B (en) 2017-05-16 2017-05-16 Spectrum waveband selection method based on multi-mode fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710342008.XA CN107220661B (en) 2017-05-16 2017-05-16 Spectrum waveband selection method based on multi-mode fusion

Publications (2)

Publication Number Publication Date
CN107220661A CN107220661A (en) 2017-09-29
CN107220661B true CN107220661B (en) 2020-09-15

Family

ID=59944014

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710342008.XA Active CN107220661B (en) 2017-05-16 2017-05-16 Spectrum waveband selection method based on multi-mode fusion

Country Status (1)

Country Link
CN (1) CN107220661B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108596077A (en) * 2018-04-20 2018-09-28 南阳理工学院 The adaptive spectrum focus wave section selection method of high spectrum image
CN109191443B (en) * 2018-08-28 2022-03-04 广东工业大学 Hyperspectral image waveband selection method based on sequence information and waveband quality
CN111062403B (en) * 2019-12-26 2022-11-22 哈尔滨工业大学 Hyperspectral remote sensing data depth spectral feature extraction method based on one-dimensional group convolution neural network

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101859383A (en) * 2010-06-08 2010-10-13 河海大学 Hyperspectral remote sensing image band selection method based on time sequence important point analysis
CN105869161A (en) * 2016-03-28 2016-08-17 西安电子科技大学 Method for selecting wave bands of hyperspectral image based on image quality assessment

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR100808145B1 (en) * 2006-12-01 2008-02-29 한국전자통신연구원 Detection time-frequency code in mb-ofdm uwb system

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101859383A (en) * 2010-06-08 2010-10-13 河海大学 Hyperspectral remote sensing image band selection method based on time sequence important point analysis
CN105869161A (en) * 2016-03-28 2016-08-17 西安电子科技大学 Method for selecting wave bands of hyperspectral image based on image quality assessment

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Constrained Band Selection for Hyperspectral Imagery;Chein-I Chang等;《IEEE》;20060605;第1575-1584页 *
利用独立成分分析的高光谱图像波段选择方法;何元磊等;《红外与激光工程》;20120331;第41卷(第3期);第818-823页 *

Also Published As

Publication number Publication date
CN107220661A (en) 2017-09-29

Similar Documents

Publication Publication Date Title
Armi et al. Texture image analysis and texture classification methods-A review
Moghimi et al. Ensemble feature selection for plant phenotyping: a journey from hyperspectral to multispectral imaging
WO2018120740A1 (en) Picture classification method, device and robot
CN108596154B (en) Remote sensing image classification method based on high-dimensional feature selection and multilevel fusion
EP3065084A1 (en) Image recognition method, image recognition device, and recording medium
Zare et al. Hyperspectral band selection and endmember detection using sparsity promoting priors
CN104484681B (en) Hyperspectral Remote Sensing Imagery Classification method based on spatial information and integrated study
Eng et al. The use of VARI, GLI, and VIgreen formulas in detecting vegetation in aerial images
CN107220661B (en) Spectrum waveband selection method based on multi-mode fusion
EP2450850B1 (en) System and method for extracting representative feature
Ghodrati et al. How can selection of biologically inspired features improve the performance of a robust object recognition model?
CN108491864B (en) Hyperspectral image classification based on automatic determination of convolution kernel size convolutional neural network
EP1636752A2 (en) Color unmixing and region of interest detection in tissue samples
CN111814685A (en) Hyperspectral image classification method based on double-branch convolution self-encoder
US20150356350A1 (en) unsupervised non-parametric multi-component image segmentation method
Güneş et al. Determination of the varieties and characteristics of wheat seeds grown in Turkey using image processing techniques
CN113052216A (en) Oil spill hyperspectral image detection method based on two-way graph U-NET convolutional network
Li et al. DeepLearnMOR: a deep-learning framework for fluorescence image-based classification of organelle morphology
Faria et al. Time series-based classifier fusion for fine-grained plant species recognition
AU2019101149A4 (en) An Image retrieval System for Brand Logos Based on Deep Learning
Li et al. Improving model robustness for soybean iron deficiency chlorosis rating by unsupervised pre-training on unmanned aircraft system derived images
Uddin et al. Paddy seed variety identification using t20-hog and haralick textural features
Pal Margin-based feature selection for hyperspectral data
Barrera et al. Classification of hyperspectral images of the interior of fruits and vegetables using a 2D convolutional neuronal network
CN109117783A (en) A kind of high spectrum image interactive class method indicated based on nuclear coordination

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant