CN105844297A - Local spatial information-based encapsulation type hyperspectral band selection method - Google Patents
Local spatial information-based encapsulation type hyperspectral band selection method Download PDFInfo
- Publication number
- CN105844297A CN105844297A CN201610164988.4A CN201610164988A CN105844297A CN 105844297 A CN105844297 A CN 105844297A CN 201610164988 A CN201610164988 A CN 201610164988A CN 105844297 A CN105844297 A CN 105844297A
- Authority
- CN
- China
- Prior art keywords
- wave band
- band
- combination
- svm
- wave
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
- G06F18/2193—Validation; Performance evaluation; Active pattern learning techniques based on specific statistical tests
-
- 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
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention provides a local spatial information-based encapsulation type hyperspectral band selection method. The method includes the following steps that: 1) a hyperspectral image to be subjected to band selection is inputted, and the hyperspectral image is converted into matrix type hyperspectral data, the matrix type hyperspectral data are normalized; 2) with a support vector machine (SVM) adopted as a classifier, a cross validation method is adopted to select M initial bands; 3) the M bands are combined with remaining bands one by one, the energy function E (f) of combination bands is calculated; 4) an existing graph cutting method is adopted to perform energy minimization on the energy function E (f) of the combination bands; and 5) band selection is carried out according to the value of the energy function of the combination bands. According to the method of the invention, spectral information and local spatial information in the hyperspectral image are fused effectively, and compared with a method according to which only spectral dimension information is utilized, the recognition performance of the method of the invention for selected bands can be greatly improved, and the method can be applied to data dimensionality reduction of hyperspectral images.
Description
Technical field
The invention belongs to technical field of image processing, further relate to the encapsulation in Hyperspectral imaging waveband selection field
Formula EO-1 hyperion band selection method, can be used for the Data Dimensionality Reduction in Hyperspectral imagery processing.
Background technology
Remote sensing technology, through the development of the second half in 20th century, in theory, technically and has major progress in application,
And wherein hyper-spectral image technique is especially prominent.High spectrum image is by the EO-1 hyperion sensor being mounted on space platform, i.e.
Imaging spectrometer, in ultraviolet, visible ray, near-infrared and the mid infrared region of electromagnetic spectrum, with tens of or even hundreds of continuously and
The spectral band of segmentation is to target area imaging simultaneously.Hyper-spectral image technique makes the spectral resolution of image have the biggest proposing
Height, is an important breakthrough of remote sensing development.One hyperspectral image data collection comprises hundreds of and has catch light spectral resolution
Spectral band, use spectral information abundant in high spectrum image can reach accurate target recognition.But it is huge
Data volume, the data mode of higher-dimension, the high redundancy of information process also to follow-up data and bring huge challenge.Therefore, how
Farthest retain the spectral information that high-spectral data is abundant, reduce the most again data dimension and become Hyperspectral imagery processing
One of important technological problems.
The dimension reduction method that high spectrum image is conventional has two kinds, is feature extraction and waveband selection respectively.Feature extraction leads to
Cross and combine original Band Set to generate new Band Set, generally by method linearly or nonlinearly, higher-dimension wave band is empty
Between map to low-dimensional wave band space, newly-generated Band Set dimension is less than original Band Set dimension, thus reaches dimensionality reduction
Purpose.Common method has PCA PCA, FLD linear discriminant analysis etc..The newly-generated wave band of feature extraction is original
Wave band obtains after conversion combination, changes original data, destroys the inherent physical significance comprised in initial data.
Waveband selection method is the Band Set new by picking out subband composition in original Band Set, does not appoints it
What changes, and maintains the integrity of data, also remains the physical meaning corresponding to wave band simultaneously, and reaches to reduce data
The purposes such as amount, elimination noise wave band.
According to the different combinations of waveband selection step from sorter model, waveband selection can be divided into three types,
It is filtering type, packaged type and embedded, wherein respectively:
Filtering type waveband selection, its step is totally independent of sorter model, obtains ripple by the character of data self
The dependency of section.The method can calculate a relevance values to each wave band, selects wave band according to relevance values, generally selects
Select the wave band that relevance values is big, weed out the wave band that relevance values is little, after waveband selection completes, by waveband selection subset pair
The data answered are input in grader.Filtering type waveband selection have ignored contacting between waveband selection and grader, selection
Band subset can not be well matched with learning algorithm, therefore performance bad in subsequent classification application.
Packaged type waveband selection, its step is fully integrated together with sorter model.Concrete implementation mode is: often
During the waveband selection of one step, by wave band to be selected with select band combination to get up to be input to grader, use grader point
Class precision, as index, selects the wave band that classifier performance of sening as an envoy to reaches optimum, until reaching termination condition, so encapsulation
The classification performance of the band subset selected by formula method compared with filtering type and embedded more preferably.Packaged type is instructed by grader completely
The process of waveband selection, in close relations with selected grader, but existing grader, only make use of the spectrum of high-spectral data to believe
Breath, fails to organically combine spatial information with spectrum information.
Embedded waveband selection, it combines learning algorithm and waveband selection mechanism is gone to evaluate and is considered in learning process
Wave band, learning training and waveband selection are carried out simultaneously, are combined with each other, and the process of structural classification model is exactly the mistake selecting wave band
Journey, repetitive cycling iteration, at the end of disaggregated model constructs, the end product of waveband selection is the ripple that disaggregated model includes
Section.This embedded band selection method is owing in combination with filtering type and packaged type, therefore processing procedure is increasingly complex.
Summary of the invention
Present invention aims to above the deficiencies in the prior art, propose a kind of encapsulation based on local spatial information
Formula EO-1 hyperion band selection method, to improve nicety of grading.
The technical thought of the present invention is: utilizes the spectral information of high-spectral data and the spatial information of image simultaneously, will build
Stand in the standard as evaluation wave band performance of the graph-cut on the basis of energy function minimizes, thus obtain better performances
Wave band, reaches the purpose of high-spectral data dimensionality reduction.Its implementation includes the following:
(1) high spectrum image of waveband selection is treated in input, it is assumed that the original wave band number of image is P, and by this EO-1 hyperion
Image is converted to the high-spectral data of matrix form;
(2) high-spectral data is normalized, its spectral value is normalized between 0-1;
(3) use SVM SVM as grader, the method for employing cross validation high-spectrum after normalization
M initial wave band is selected in Xiang;
(4) by M the wave band selected, remaining unselected wave band is combined into composite wave one by one with original Band Set respectively
Section, the initial labels collection f of pixel samples in each combination wave band and pixel samples are belonging respectively to utilize SVM to draw
The estimated probability p of certain classification of high spectrum image;
(5) the classification estimated probability p according to each pixel calculates the spectral energy item of pixel itself, by all pixels
Spectral energy item be added obtain spectral cterm Ed;
(6) each pixel and its spatially dimensional energy item between neighbor are calculated according to potts model, by institute
The dimensional energy item having pixel is added and obtains space item Es;
(7) by spectral cterm EdWith space item EsIt is added energy function E (f) obtaining each combination wave band;
(8) utilize graph-cut figure blanking method that the energy function of each combination wave band is carried out energy minimization, obtain each group
Close the final tally set of wave bandAnd energy
(9) energy of all combination wave bands is contrastedSelect whereinMinimum combination wave band, this combination wave band
In the M+1 wave band be required wave band.
The present invention compared with prior art has the advantage that
The most existing band selection method the most only make use of the spectrum dimension information of high-spectral data, and have ignored its space
Information, the spatial information of EO-1 hyperion is joined in the standard evaluating wave band performance by the present invention, selected by the wave band that goes out can be simultaneously
Reflection spectral information and spatial information, ability to express is higher.
2. the spatial information that existing band selection method is ignored is attached in spectral information by the present invention, make use of Gao Guang
The more information of modal data, so its nicety of grading is higher.
Accompanying drawing explanation
Fig. 1 is the flowchart of the present invention;
Fig. 2 is the present invention and the Selection effect correlation curve figure of prior art on the high spectrum image of the state of Indiana;
Fig. 3 is the present invention and the Selection effect correlation curve figure of prior art on Botswana's high spectrum image;
Fig. 4 is the present invention and the Selection effect correlation curve of prior art on Kennedy Space Center's high spectrum image
Figure.
Detailed description of the invention
Referring to the drawings 1, the present invention to realize step as follows:
Step 1, input data.
The high spectrum image of waveband selection is treated in input, it is assumed that the original wave band number of image is P, converts the image into matrix
The high-spectral data of form.
Step 2, high-spectral data is made normalized.
Process for convenience of follow-up data, high-spectral data is made normalized, by the spectral value normalizing of high spectrum image
Change between 0-1.
Step 3, choose marker samples.
From high spectrum image, choosing has the pixel of classification information as sample point, then randomly selects hundred from sample point
/ ten as marker samples, participates in subsequent arithmetic.
Step 4, based on cross validation method use SVM SVM in original wave band, carry out waveband selection.
(4a) setting and have been chosen by Band Set as S, time initial, S be sky, i.e. S=[];
(4b) first wave band is selected
(4b.1) method using 5 folding cross validations, is divided into five equal portions by known mark sample mean, is used alternatingly wherein
Marker samples, as training sample, as test sample, is classified for remaining four parts by portion with SVM SVM, this
In example, SVM classifier uses libsvm-3.20 to realize, and parameter is c=1024, g=2-7, remaining parameter is default value, obtains P
The classification accuracy of original wave band;
(4b.2) wave band that selection sort accuracy is the highest from the classification accuracy of P original wave band, remembers this wave band
For s1;
(4b.3) in original Band Set, wave band s is removed1, by wave band s1Add and selected Band Set S, the most selected wave band
Set S=[s1];
(4c) second wave band is selected:
(4c.1) method using 5 folding cross validations, carries out the selection of second wave band, will the most selected wave band
s1Respectively with the unselected band combination of residue, obtain P-1 combination wave band, with SVM SVM, marker samples classified,
Obtain the classification accuracy of P-1 combination wave band;
(4c.2) the combination wave band that accuracy of selection is the highest from the classification accuracy of P-1 combination wave band, by this combination
Second wave band in wave band is denoted as s2;
(4c.3) in original Band Set, wave band s is removed2, by wave band s2Add and selected Band Set S, the most selected wave band
Set S=[s1,s2];
(4d) the m+1 wave band is selected:
(4d.1) method using 5 folding cross validations, will be denoted as by P-m any one wave band remained in original wave band
B, has selected wave band S=[s the most1,s2...sm], will select wave band S Yu b combination obtain P-m combine wave band, be denoted as Q=[...
sm, b], with SVM SVM, marker samples is classified, obtain the classification accuracy of P-m combination wave band;
(4d.2) the combination wave band that accuracy of selection is the highest from the classification accuracy of P-m combination wave band, by this combination
Wave band b in wave band is denoted as sm+1;
(4d.3) in remaining original wave band, wave band s is removedm+1, by sm+1Add and selected Band Set S, the most selected ripple
Duan Jihe S=[s1,s2...sm,sm+1];
(4e) repeated execution of steps (4d), until selection wave band reaches required wave band number, the most selected works conjunction S and is
Required wave band, have selected M wave band altogether in this example.
Step 5, calculating spectral cterm Ed。
(5a) select a wave band in P-M wave band to be designated as b, by wave band b and select Band Set S to be combined into a group
Close wave band, be designated as G=[S b];
(5b) SVM SVM is used to obtain combining the classification estimated probability p (L of sample point in wave band Gi|xi), wherein
xiRepresent the sample point in high spectrum image, LiRepresent the class mark of high spectrum image, calculate and weigh to sample xiIt is assigned to class mark Li's
Error:
Vi(Li)=-ln (p (Li|xi))
Wherein, 1≤i≤n, n represent the number of sample point in high spectrum image;
(5c) spectral cterm E of combination wave band G is calculatedd:
Step 6, calculating space item Es。
(6a) select a wave band in P-M wave band to be designated as b, by wave band b and select Band Set S to be combined into a group
Close wave band, be designated as G=[S b];
(6b) use SVM SVM to obtain the initial labels collection f of sample point of G, f is expressed as [L1,
L2...Li...Lj...];
(6b) spatially two sample x are calculatediAnd xjDiversity:
Wi.j(Li,Lj)=β (1-δ (Li,Lj))
Wherein, β is the parameter controlling spatial smoothness significance level, and value is β=1;
(6c) the space item E of combination wave band G is calculateds:
Wherein, adjacent pair sample on i~j representation space.
Step 7, energy function E (f) of calculating combination wave band G.
E (f)=Ed+Es
Wherein, EdRepresent the spectral cterm of Band Set, EsRepresent the space item of Band Set.
Step 8, use graph-cut figure blanking method carry out energy minimization to combination wave band.
(8a) random by the part labels L in initial labels collection fiChanging label α into, wherein α is the class mark of high spectrum image
One of, the tally set f' after being adjusted;
(8b) according to the tally set f' after adjusting, utilize argmin to seek ginseng function and calculate the energy function making combination wave band G
Minimum candidate's tally set:Wherein, E (f') represents the energy function of the tally set f' after adjusting;
(8c) final tally set is calculated
Wherein,Represent the tally set after adjustingEnergy function;
(8d) step that iterates (8a)-(8c), until tally setNo longer till change, then this tally setCorresponding
Energy functionIt is the energy value of combination wave band G.
Step 9, carry out waveband selection.
The relatively energy function of all combination wave bandsSelect whereinMinimum combination wave band G, by the ripple in G
Section b, as selecting wave band to add Band Set S, concentrates from unselected wave band simultaneously and removes this wave band;
Repeated execution of steps 5 to step 8, until selection wave band reaches required wave band number, the most selected works conjunction S and is
Required wave band.
Below in conjunction with analogous diagram, the effect of the present invention is described further.
1. simulated conditions:
Hardware platform is: processor is Inter Core i7-5557U, and dominant frequency is 3.10GHz, inside saves as 4GB;
Software platform is: Windows 10 home edition 64 bit manipulation system, MatlabR2015b.
2. emulation data:
This example is used the state of Indiana obtained by airborne visible ray and Infrared Imaging Spectrometer AVIRIS high
The Kennedy Space that spectrum picture, NASA E0-1 Botswana's high spectrum image obtained, and NASA AVIRIS obtain
Center high spectrum image carries out sorting algorithm emulation;State of Indiana high spectrum image is designated as Indian pines, by rich thatch
High spectrum image of watt receiving is designated as Botswana, and Kennedy Space Center's high spectrum image is designated as KSC.
The high-spectral data of three width images and the ground truth figure of correspondence come from:
http://www.ehu.es/ccwintco/index.php?Title=Hyperspectral_Remote_
Sensing_Sce nes.
3. emulation content:
In order to verify the effectiveness of waveband selection, typically after carrying out waveband selection, can be with in ground truth figure
Classification as label, carry out classification hyperspectral imagery experiment, using the accuracy criterion as waveband selection performance of classifying.
(3.1) emulation one:
Existing four the representative band selection methods of experimental selection compare with the present invention, and these four method is divided
It not based on similarity without supervision EO-1 hyperion band selection method SBBS, SVM wave band based on cross validation precision
System of selection SVMCV, represent component analysis method ECA, maximal correlation minimal redundancy method MRMR;
With existing four kinds of band selection methods, Indian pines image is made waveband selection by the present invention, choose 100 altogether
Individual wave band, the sample set X new to wave band composition selected by all sample extraction of high spectrum image, according to new sample set pair
High-spectral data is classified, and randomly chooses the sample of 10% as training sample from X, residue sample as test sample,
Use SVM SVM classifier to do classification experiments, respectively obtain the classification accuracy of five kinds of methods;In this experiment, SVM divides
Class device uses libsvm-3.20 to realize, and SVM uses RBF core, and parameter is c=1024, g=2-7, remaining parameter is silent
Recognize value;Result such as accompanying drawing 2;
Accompanying drawing 2 presents the accuracy obtained for classification after five kinds of methods pick out 10 to 100 wave bands respectively.From
It can be seen that the effect of the present invention is better than existing four kinds of methods, the especially present invention picks out acquired by 20 wave bands in Fig. 2
Effect has been better than performance when existing four kinds of methods, 50 wave bands of acquisition, illustrates that the waveband selection performance of the present invention is more excellent
Elegant.
(3.2) emulation two:
With above-mentioned existing four kinds of band selection methods, Botswana image is made waveband selection by the present invention, choose 50 altogether
Wave band, the sample set Y new to wave band composition selected by all sample extraction of high spectrum image, according to new sample set to height
Spectroscopic data is classified, and randomly chooses the sample of 10% and as test sample, make as training sample, residue sample from Y
Do classification experiments by SVM SVM classifier, respectively obtain the classification accuracy of five kinds of methods;Svm classifier in this experiment
Device uses libsvm-3.20 to realize, and SVM uses RBF core, and parameter is c=1024, g=2-7, remaining parameter is acquiescence
Value;Result such as accompanying drawing 3;
Accompanying drawing 3 presents the accuracy obtained for classification after five kinds of methods pick out 5 to 50 wave bands respectively.From figure
It can be seen that the effect of the present invention is better than existing four kinds of methods in 3, the especially present invention picks out the effect acquired by 15 wave bands
Fruit has been better than performance when existing four kinds of methods, 50 wave bands of acquisition, illustrates that the waveband selection performance of the present invention is more excellent
Elegant.
(3.3) emulation three:
With above-mentioned existing four kinds of band selection methods, KSC image is made waveband selection by the present invention, choose 100 ripples altogether
Section, the sample set Z new to wave band composition selected by all sample extraction of high spectrum image, according to new sample set to Gao Guang
Modal data is classified, and randomly chooses the sample of 10% and as test sample, use as training sample, residue sample from Z
SVM SVM classifier does classification experiments, respectively obtains the classification accuracy of five kinds of methods;SVM classifier in this experiment
Using libsvm-3.20 to realize, SVM uses RBF core, and parameter is c=1024, g=2-7, remaining parameter is default value;
Result such as accompanying drawing 4;
Accompanying drawing 4 presents the accuracy obtained for classification after five kinds of methods pick out 10 to 100 wave bands respectively.From
It can be seen that the effect of the present invention is better than existing four kinds of methods, the especially present invention picks out acquired by 20 wave bands in Fig. 4
Effect has been better than performance when existing four kinds of methods, 100 wave bands of acquisition, the waveband selection performance of the present invention is described more
Outstanding.
To sum up, the effect of the present invention is substantially better than other several method, in the case of different-waveband number, relatively other
Method performance has had and has been obviously improved, and shows that the present invention is highly effective.
Claims (5)
1. a packaged type EO-1 hyperion band selection method based on local spatial information, including:
(1) high spectrum image of waveband selection is treated in input, it is assumed that the original wave band number of image is P, and by this high spectrum image
Be converted to the high-spectral data of matrix form;
(2) high-spectral data is normalized, its spectral value is normalized between 0-1;
(3) use SVM SVM is as grader, the method for employing cross validation high spectrum image after normalization
Select M initial wave band;
(4) by M the wave band selected, remaining unselected wave band is combined into combination wave band one by one with original Band Set respectively, profit
Show that the initial labels collection f of pixel samples in each combination wave band and pixel samples are belonging respectively to EO-1 hyperion with SVM
The estimated probability p of certain classification of image;
(5) the classification estimated probability p according to each pixel calculates the spectral energy item of pixel itself, by the light of all pixels
Spectrum energy item is added and obtains spectral cterm Ed;
(6) each pixel and its spatially dimensional energy item between neighbor are calculated according to potts model, by all pictures
The dimensional energy item of element is added and obtains space item Es;
(7) by spectral cterm EdWith space item EsIt is added energy function E (f) obtaining each combination wave band;
(8) utilize graph-cut figure blanking method that the energy function of each combination wave band is carried out energy minimization, obtain each composite wave
The final tally set of sectionAnd energy
(9) energy of all combination wave bands is contrastedSelect whereinMinimum combination wave band, the in this combination wave band
M+1 wave band is required wave band.
2. according to the packaged type EO-1 hyperion band selection method based on local spatial information described in claim 1, its feature
It is, step (3) uses SVM SVM carry out waveband selection, carry out as follows:
(3a) select that high spectrum image has the pixel of classification information as sample point, and randomly select in sample point
10% as marker samples;
(3b) setting and have been chosen by Band Set as S, time initial, S be sky, i.e. S=[];
(3c) method using 5 folding cross validations, carries out the selection of first wave band, will be divided into five by known mark sample mean
Equal portions, are used alternatingly a copy of it as training sample, remaining four parts as test sample, with SVM SVM to mark
Note sample is classified, and obtains the classification accuracy of P original wave band, selects the wave band that wherein accuracy is the highest, by this wave band
It is designated as s1, and in original Band Set, remove wave band s1, by wave band s1Add and selected Band Set S, the most selected Band Set S
=[s1];
(3d) method using 5 folding cross validations, carries out the selection of second wave band, will the most selected wave band s1Respectively
With the unselected band combination of residue, obtain P-1 combination wave band;With SVM SVM, marker samples is classified, obtain P-
The classification accuracy of 1 combination wave band, selects the unselected wave band in the combination wave band that wherein accuracy is the highest, is denoted as by this wave band
s2, and in original Band Set, remove wave band s2, by s2Add and selected Band Set S, the most selected Band Set S=[s1, s2];
(3e) method using 5 folding cross validations, carries out the selection of m+1 wave band, will P-m remain in original wave band
Any one wave band is denoted as b, has the most selected wave band S=[s1, s2...sm], obtain P-m combination by selecting wave band S Yu b combination
Wave band, is denoted as Q=[s1...sm, b], with SVM SVM, marker samples is classified, obtain P-m combination wave band
Classification accuracy, selects the unselected wave band in the combination wave band that wherein accuracy is the highest, this unselected wave band is denoted as sm+1, and
Remain and original wave band is got rid of this wave band, by sm+1Add and selected Band Set S, the most selected Band Set S=[s1,
s2...sm,sm+1];
(3f) repeated execution of steps (3e), selects wave band until reaching required wave band number, and selected works conjunction S the most is required
Wave band.
3. according to the packaged type EO-1 hyperion band selection method based on local message described in claim 1, it is characterised in that
Step (5) calculates spectral cterm Ed, carry out as follows:
(5a) select a wave band in unselected wave band to be designated as b, by wave band b and select Band Set S to be combined into a composite wave
Section, is designated as G=[S b];
(5b) SVM SVM is used to obtain the classification estimated probability p (L of sample point in Gi|xi), wherein xiRepresent EO-1 hyperion
Sample point in image, LiRepresent the class mark of high spectrum image, calculate and weigh to sample xiIt is assigned to class mark LiError:
Vi(Li)=-ln (p (Li|xi))
Wherein, 1≤i≤n, n represent the number of sample point in high spectrum image;
(5c) spectral cterm E of combination wave band G is calculatedd:
4. according to the packaged type EO-1 hyperion band selection method based on local message described in claim 1, it is characterised in that
Step (6) calculates space item Es, carry out as follows:
(6a) select a wave band in unselected wave band to be designated as b, by wave band b and select Band Set S to be combined into a composite wave
Section, is designated as G=[S b];
(6b) use SVM SVM to obtain the initial labels collection f of sample point of G, f is expressed as [L1,
L2...Li...Lj...];
(6b) spatially two sample x are calculatediAnd xjDiversity:
Wi.j(Li,Lj)=β (1-δ (Li,Lj))
Wherein, β is the parameter controlling spatial smoothness significance level, and value is β=1;
(6c) the space item E of combination wave band G is calculateds:
Wherein, adjacent pair sample on i~j representation space.
5. according to the packaged type EO-1 hyperion band selection method based on local message described in claim 1, it is characterised in that
Step (8) uses graph-cut figure blanking method combination wave band is carried out energy minimization, carries out as follows:
(8a) random by the part labels L in initial labels collection fiChanging label α into, wherein α is one of class mark of high spectrum image,
Tally set f' after being adjusted;
(8b) according to the tally set f' after adjusting, utilize argmin to seek ginseng function and calculate the energy function minimum making combination wave band G
Candidate's tally set Wherein, E (f') represents the energy function of the tally set f' after adjusting;
(8c) final tally set is calculated
Wherein,Represent the tally set after adjustingEnergy function;
(8d) step that iterates (8a)-(8c), until tally setNo longer till change, then this tally setCorresponding energy
FunctionIt is the energy value of combination wave band G.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610164988.4A CN105844297A (en) | 2016-03-22 | 2016-03-22 | Local spatial information-based encapsulation type hyperspectral band selection method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610164988.4A CN105844297A (en) | 2016-03-22 | 2016-03-22 | Local spatial information-based encapsulation type hyperspectral band selection method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN105844297A true CN105844297A (en) | 2016-08-10 |
Family
ID=56587881
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610164988.4A Pending CN105844297A (en) | 2016-03-22 | 2016-03-22 | Local spatial information-based encapsulation type hyperspectral band selection method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105844297A (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106529563A (en) * | 2016-09-19 | 2017-03-22 | 西安电子科技大学 | High-spectral band selection method based on double-graph sparse non-negative matrix factorization |
CN107832793A (en) * | 2017-11-08 | 2018-03-23 | 深圳大学 | The sorting technique and system of a kind of high spectrum image |
CN117274236A (en) * | 2023-11-10 | 2023-12-22 | 山东第一医科大学第一附属医院(山东省千佛山医院) | Urine component abnormality detection method and system based on hyperspectral image |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104820841A (en) * | 2015-05-08 | 2015-08-05 | 西安电子科技大学 | Hyper-spectral classification method based on low-order mutual information and spectral context band selection |
CN105046276A (en) * | 2015-07-14 | 2015-11-11 | 西安电子科技大学 | Hyperspectral image band selection method based on low-rank expression |
CN105184314A (en) * | 2015-08-25 | 2015-12-23 | 西安电子科技大学 | wrapper-type hyperspectral waveband selection method based on pixel clustering |
-
2016
- 2016-03-22 CN CN201610164988.4A patent/CN105844297A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104820841A (en) * | 2015-05-08 | 2015-08-05 | 西安电子科技大学 | Hyper-spectral classification method based on low-order mutual information and spectral context band selection |
CN105046276A (en) * | 2015-07-14 | 2015-11-11 | 西安电子科技大学 | Hyperspectral image band selection method based on low-rank expression |
CN105184314A (en) * | 2015-08-25 | 2015-12-23 | 西安电子科技大学 | wrapper-type hyperspectral waveband selection method based on pixel clustering |
Non-Patent Citations (4)
Title |
---|
XIANGHAI CAO等: "Supervised Band Selection Using Local Spatial Information for Hyperspectral Image", 《IEEE》 * |
YULIYA TARABALKA等: "Graph-cut-based model for spectral-spatial classification of hyperspectral images", 《IEEE》 * |
YURI BOYKOV等: "Fast approximate energy minimization via graph cuts", 《IEEE》 * |
于绍慧等: "基于矩阵模式的高光谱波段选择方法", 《光电工程》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106529563A (en) * | 2016-09-19 | 2017-03-22 | 西安电子科技大学 | High-spectral band selection method based on double-graph sparse non-negative matrix factorization |
CN106529563B (en) * | 2016-09-19 | 2019-03-26 | 西安电子科技大学 | EO-1 hyperion band selection method based on the sparse Non-negative Matrix Factorization of digraph |
CN107832793A (en) * | 2017-11-08 | 2018-03-23 | 深圳大学 | The sorting technique and system of a kind of high spectrum image |
CN107832793B (en) * | 2017-11-08 | 2021-08-06 | 深圳大学 | Hyperspectral image classification method and system |
CN117274236A (en) * | 2023-11-10 | 2023-12-22 | 山东第一医科大学第一附属医院(山东省千佛山医院) | Urine component abnormality detection method and system based on hyperspectral image |
CN117274236B (en) * | 2023-11-10 | 2024-03-08 | 山东第一医科大学第一附属医院(山东省千佛山医院) | Urine component abnormality detection method and system based on hyperspectral image |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Kang et al. | Classification of hyperspectral images by Gabor filtering based deep network | |
CN109948693B (en) | Hyperspectral image classification method based on superpixel sample expansion and generation countermeasure network | |
Song et al. | Hyperspectral image classification with deep feature fusion network | |
CN107451614B (en) | Hyperspectral classification method based on fusion of space coordinates and space spectrum features | |
CN111274869B (en) | Method for classifying hyperspectral images based on parallel attention mechanism residual error network | |
Li et al. | Gaussian pyramid based multiscale feature fusion for hyperspectral image classification | |
CN102324047B (en) | Hyper-spectral image ground object recognition method based on sparse kernel representation (SKR) | |
Chen et al. | Hyperspectral classification based on spectral–spatial convolutional neural networks | |
Zare et al. | Hyperspectral band selection and endmember detection using sparsity promoting priors | |
CN107145836B (en) | Hyperspectral image classification method based on stacked boundary identification self-encoder | |
CN107239759B (en) | High-spatial-resolution remote sensing image transfer learning method based on depth features | |
CN109766858A (en) | Three-dimensional convolution neural network hyperspectral image classification method combined with bilateral filtering | |
CN107145830A (en) | Hyperspectral image classification method with depth belief network is strengthened based on spatial information | |
Imani et al. | Principal component discriminant analysis for feature extraction and classification of hyperspectral images | |
CN111783884B (en) | Unsupervised hyperspectral image classification method based on deep learning | |
CN105184314B (en) | Wrapper formula EO-1 hyperion band selection methods based on pixel cluster | |
CN104866871B (en) | Hyperspectral image classification method based on projection structure sparse coding | |
CN107844751A (en) | The sorting technique of guiding filtering length Memory Neural Networks high-spectrum remote sensing | |
Li et al. | The effect of ground truth on performance evaluation of hyperspectral image classification | |
CN105913092A (en) | Figure regular hyperspectral image band selection method based on subspace learning | |
CN109472287A (en) | Three-dimensional fluorescence spectrum feature extracting method based on Two-Dimensional Gabor Wavelets | |
CN115564996A (en) | Hyperspectral remote sensing image classification method based on attention union network | |
Jahan et al. | Fusion of hyperspectral and LiDAR data using discriminant correlation analysis for land cover classification | |
CN113673556B (en) | Hyperspectral image classification method based on multiscale dense convolution network | |
CN104008394A (en) | Semi-supervision hyperspectral data dimension descending method based on largest neighbor boundary principle |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20160810 |