CN106971204A - A kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set - Google Patents
A kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set Download PDFInfo
- Publication number
- CN106971204A CN106971204A CN201710209001.0A CN201710209001A CN106971204A CN 106971204 A CN106971204 A CN 106971204A CN 201710209001 A CN201710209001 A CN 201710209001A CN 106971204 A CN106971204 A CN 106971204A
- Authority
- CN
- China
- Prior art keywords
- attribute
- rough set
- remote sensing
- sensing data
- wave band
- 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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to a kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set, it the described method comprises the following steps:Original target in hyperspectral remotely sensed image is pre-processed, cancelling noise jammr band, preselect type of ground objects, determine its topological structure;Rough Set Attribute Reduction, redundancy wave band is removed using Rough Set, retains important wave band;Comentropy sorts, and carries out importance sorting to important wave band according to comentropy, filtering out influences big band combination on classification results;Nicety of grading is carried out with PCA methods to be compared, verify dimensionality reduction effect by the band combination filtered out.Its advantage is shown:The effective dimensionality reduction of EO-1 hyperion can be realized, the amount of storage and transmission quantity of high-spectral data, subsequent treatment and analysis beneficial to high spectrum image is reduced.
Description
Technical field
It is a kind of new based on rough set specifically the present invention relates to the dimension reduction method technical field of high-spectral data
The dimension reduction method of high-spectrum remote sensing data.
Background technology
With developing rapidly for remote sensing technology, the resolution ratio of new remote sensing images, such as:Spectral resolution, spatial resolution
There is provided on atural object more careful spectral information improving constantly with temporal resolution.But meanwhile, also result in EO-1 hyperion
The data volume of storage and the transmission of view data is greatly increased so that artificial interpretation becomes to waste time and energy, but also can be influenceed point
Class effect is deteriorated.How on the premise of the important information of Objects recognition and classification is preserved for, realize to EO-1 hyperion redundant digit
According to effectively processing it is challenging.
Principal component analysis (PCA) is classical high-spectral data dimension-reduction treatment method.This method can not only remove bloom
The second order correlation between image band is composed, and main energy can be concentrated on to preceding several principal components, therefore can be passed through
The big principal component of preceding several characteristic values for retaining PCA realizes the dimensionality reduction of high-spectral data.The open source literatures such as author Yu Ting " are improved
PCA dimension-reduction algorithms and its application in multivariate process quality control ", designs a kind of improved PCA algorithms, there is phase for handling
The Data Dimensionality Reduction problem of closing property.The pedestrian detection of the multiple features cascade based on PCA dimensionality reductions such as sweet tinkling of pieces of jade of literature author, utilizes PCA pairs
HOG features carry out dimensionality reduction, HOG features and Gabor characteristic, color characteristic secondly are cascaded into the feature as pedestrian detection, finally
Classified using SVM radial direction base (RBF) kernel function.Document disclosed in author Zang Zhuo etc. " carries out arbor tree using PCA algorithms
Plant high-spectral data dimensionality reduction and classification ", using PCA respectively to arbor species hyper spectral reflectance initial data and 3
Plant preprocessed data and carry out dimensionality reduction computing, reuse 4 kinds of sorting algorithms of SVM-RBF, SVM-Linear, BP, Fisher, to drop
Data after dimension carry out class test.
Independent component analysis (ICA) is conceived to the higher order statistical characteristic between data so that between each component after conversion not
It is only orthogonal, and also statistical iteration as far as possible.Therefore the ICA methods based on higher order statistical specificity analysis can be disclosed more comprehensively
Essential structure between data, has unique advantage for processing non-Gaussian signal particularly view data.Author Wang Yi
Document disclosed in sail etc. " the multispectral data dimension reduction method based on PCA and ICA ", it is proposed that one kind combination principal component analysis
(PCA) and independent component analysis (ICA) multispectral data dimension reduction method, realize the linear combination with low-dimensional base vector come table
Show the spectroscopic data of higher-dimension." the complex outline Identifying Outliers side based on ISOMAP dimensionality reductions of document disclosed in author Nie Bin etc.
Method ", using higher-dimension complex outline as research object, sets up nonparametric profile matrix model, and the ISOMAP based on geodesic distance is non-
Linear dimensionality reduction technology is combined with the control figure of χ~2, the new profile Identifying Outliers method of proposition, to realize higher-dimension complex outline
Abnormity point is accurately identified.Document disclosed in author Luo Lianling etc. " ICA and svm classifier of remote sensing images crop type textural characteristics ",
Progress independent component analysis ICA dimensionality reductions are proposed, textural characteristics is obtained by calculating gray level co-occurrence matrixes, using svm classifier, grinds
Study carefully the rapid classification method of Type of Forest Land.
Minimal noise separation conversion (MNF) instrument be used to judge in view data dimension (i.e. wave band number), separation number
Noise in, is reduced with the calculating demand in post processing, MNF is substantially the principal component transform being laminated twice.Tang Yuanyuan
Deng document disclosed in author " the target in hyperspectral remotely sensed image Study on Data Processing based on CPU/GPU heterogeneous schemas is with realizing ", it is based on
CPU/GPU heterogeneous schemas realizes the parallelization of target in hyperspectral remotely sensed image MNF dimensionality reductions, by with serial program and shared storage
OpenMP isomorphisms pattern is contrasted, and demonstrates development potentiality of the heterogeneous schemas in high-spectrum remote-sensing process field.Author Liu Pingping etc.
Disclosed document " the dimension-reduction treatment technique study of high-spectral data ", carries out Subspace partition before waveband selection, reject related
Property big wave band, then effective end member number of high-spectral data can carry for the further analysis and application of image after being converted by MNF
For reference.Document disclosed in author Zhang Jingjing etc. " a kind of improved large scale EO-1 hyperion manifold dimension-reducing algorithm ", it is proposed that a kind of
The EO-1 hyperion manifold dimension-reducing algorithm IISOMAP-LLE combined based on increment Isometric Maps (IISOMAP) and LLE, and for manifold
The non-linear knot that the more linear dimension-reduction algorithm minimal noise separation (MNF) of dimension-reduction algorithm can be preferably excavated out in high-spectral data
The advantage of structure, the feasibility and superiority of algorithm are demonstrated by AVIRIS and OMIS-II data experiments.
Above-mentioned document carries out dimension-reduction treatment from different square high-spectral datas respectively, but how to be preserved for atural object knowledge
Not with the important information of classification on the premise of, realize challenging to the effectively processing of EO-1 hyperion redundant data.
Need a kind of high-spectrum image dimensionality reduction towards analysis based on rough set (Rough Set) and comentropy badly in summary
And compression method, it is possible to achieve the effective dimensionality reduction of EO-1 hyperion, reduces the amount of storage and transmission quantity of high-spectral data, beneficial to EO-1 hyperion
The subsequent treatment of image and analysis.And this method yet there are no report.
The content of the invention
The purpose of the present invention is for of the prior art not enough a kind of based on rough set (Rough Set) and letter there is provided one
Cease the high-spectrum image dimensionality reduction and compression method towards analysis of entropy, it is possible to achieve the effective dimensionality reduction of EO-1 hyperion, reduce EO-1 hyperion
The amount of storage and transmission quantity of data, subsequent treatment and analysis beneficial to high spectrum image.
To achieve the above object, the present invention is adopted the technical scheme that:
A kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set, the described method comprises the following steps:
Step S1:Original target in hyperspectral remotely sensed image is pre-processed, cancelling noise jammr band, preselect type of ground objects, really
Its fixed topological structure;
Step S2:Rough Set Attribute Reduction, redundancy wave band is removed using Rough Set, retains important wave band;
Step S3:Comentropy sorts, and carries out importance sorting to the important wave band in step S2 according to comentropy, filters out
Classification results are influenceed with big band combination;
Step S4:Nicety of grading is carried out with PCA methods to be compared, verify dimensionality reduction effect by the band combination filtered out.
As a kind of perferred technical scheme, the pre-selection type of ground objects in step S1 includes water body, naked rock, vegetation and building
Thing.
As a kind of perferred technical scheme, important wave band and redundancy wave band are obtained by differential matrix in step S2.
As a kind of perferred technical scheme, during the wave band described in PCA methods is used in step S4 for step S1 after denoising
The wave band obtained.
As a kind of perferred technical scheme, nicety of grading compares the comparison for needing to carry out time efficiency in step S4, its
The comparison of time efficiency includes maximum likelihood, minimum range, parallelepiped, mahalanobis distance, binary coding, supporting vector
Machine.
As a kind of perferred technical scheme, it is that sorted precision is obtained by confusion matrix in step S4.
As a kind of perferred technical scheme, the definition method of the Rough Set Attribute Reduction in step S2 is as follows:
If S=(U, A, V, are f) decision table, wherein:
U is domain, is a nonempty finite object set, and A is the attribute set of nonempty finite object, A=C ∪ D,What C, D were each represented is conditional attribute and decision attribute;V=∪a∈AVaIt is property value, VaIt is attribute a value
Domain;f:U × A → V is referred to as information function, is that each attribute of each object gives a value of information;What f was represented is according to determination
Domain and attribute, obtain corresponding property value.
For decision table S=, (U, A, V, f), C and D are defined in two equivalence relations on U, and the D positive domains of C are denoted as POSC
(D)=∪X∈U/DC_ (X), if POSC(D)=POS(C-{b})(D) it in C is redundancy that, then attribute b, which is called, conversely, b is called to be in C
It is necessary;If all attributes are necessary in C, it is independent to claim C;If C D independent subsetsThere is POSC′(D)=
POSC(D), then C ' is called a C yojan, and the relation race that C all yojan are constituted is designated as redC(D);The friendship of all yojan in C
Collection is referred to as C D cores, i.e.,:coreD(C)=∩ redD(C)。
As a kind of perferred technical scheme, the definition method of the comentropy in step S3 is as follows:
If what U was represented is the domain of an information system, if domain is divided into m different type by decision attribute D;For
Some Di, i ∈ [1,2 ..., m], it includes element number and is designated as di, any one object belongs to DiProbability be di/|U|;Examine
Conditional attribute C is considered, for any a ∈ C, if attribute a value is a1, a2..., an, it divides domain U for v parts of { u1, u2...,
uv};Wherein, ujSame value a is taken comprising attribute ajData row;ujInclude uijIndividual DiData object;According to attribute a value pair
Current data is divided the entropy that obtained information is known as attribute a;
Its calculation formula is:
Learnt from formula, H (a) essence is conditional entropies of the attribute a on decision attribute D, i.e. H (a)=H (D | a);
Attribute a information gain definition is Gain (a)=H (D)-H (a);What it was represented is according to decision attribute D probability point
The entropy of cloth formation with conditional attribute a is divided after, the difference between the entropy of obtained probability distribution formation.
As a kind of perferred technical scheme, the definition method of differential matrix is as follows:
The differential matrix of decision table is a symmetrical n rank square formation, wherein, matrix element mijIt is defined as:
Wherein Xi, Xj∈U
When building the differential matrix of decision table, only in Xi, XjOn the premise of being not belonging to same Decision Classes, mijIt is area
Divide Xi, XjAll properties set;If Xi, XjBelong to same Decision Classes, then element m in differential matrixijFor empty set.
C D cores are the m of all individual elements in differential matrixijAnd, i.e.,:
CORED(C)=a ∈ C | mij={ a } 1≤i, j≤n }
The differential matrix function of decision table is defined as follows:
ρ*=∧ { ∨ mij}
Each conjunction expression corresponds to C D yojan respectively in function ρ * minimal disjunctive normal form.
The invention has the advantages that:
1st, the dimension reduction method of a kind of new high-spectrum remote sensing data based on rough set of the invention, this method is first by thick
The spectral signature that collection extracts EO-1 hyperion realizes high-spectrum image dimensionality reduction, that is, the redundant attributes between spectrum is removed, secondly by comentropy
Remaining spectral information is ranked up by method by importance, the carry out Optimal Bands Selection based on information content principle, Neng Gougeng
The feature in high spectrum image is extracted well, with stronger classification capacity, facilitates follow-up analyzing and processing.
Brief description of the drawings
A kind of FB(flow block) of the dimension reduction method of new high-spectrum remote sensing data based on rough set of the present invention of accompanying drawing 1.
Accompanying drawing 2 is experimental study area schematic.
Accompanying drawing 3 is 7 wave bands and 204 wave band time efficiency comparison schematic diagrams.
Embodiment
The embodiment that the present invention is provided is elaborated below in conjunction with the accompanying drawings.
Fig. 1 is refer to, Fig. 1 is a kind of stream of the dimension reduction method of new high-spectrum remote sensing data based on rough set of the present invention
Journey block diagram.A kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set, the described method comprises the following steps:
Step S1:Original target in hyperspectral remotely sensed image is pre-processed, cancelling noise jammr band, preselect type of ground objects, really
Its fixed topological structure;Rough Set Attribute Reduction, redundancy wave band is removed using Rough Set, retains important wave band;Comentropy sorts, according to letter
Cease entropy and importance sorting is carried out to wave band, filtering out influences big band combination on classification results;Pass through the band group filtered out
Close and compared with PCA methods progress nicety of grading, verify dimensionality reduction effect.
It should be noted that:
Rough set
For information system, people often choose multiple indexs as judgment criteria and carry out problem analysis.But it is traditional
Decision-making technique think that all indexs are all important, simply shared weighted, the problem of typically not considering redundancy.And
The weight of index is typically provided by the experience of expert, because value and the preference of expert are had nothing in common with each other, the result for causing to draw
Objectivity and fairness are affected.Rough Set has based on the fact that the characteristic of reasoning and data-driven, it is possible to use rough set
The problem of method of attribute reduction is to solve subjective arbitrariness in the redundancy of index, and index weights determination.
The essence of attribute reduction is to find to keep information system to divide constant a series of minimal attribute set, passes through yojan
Afterwards, core attributes will be found, it is not necessary to which attribute will be deleted, and this can judge the important pass between attribute to a certain extent naturally
System.The angle established from weight, certain attribute, can be by judging to delete after the attribute, entirely in the percentage contribution of information system
Whether system architecture is changed, and the size of the degree that changes is described.Specifically, can be with the contribution of defined attribute
In the case of spending to keep other attributes constant, delete after the attribute, the size of system architecture change degree.This change degree is got over
Greatly, the attribute is bigger to the significance level for maintaining whole system stability, and its assignment also should be bigger.
Rough set is a kind of fuzzy and uncertain knowledge the mathematical tool of new solution.Its central idea is exactly to keep classification
On the premise of ability is constant, by Reduction of Knowledge, decision-making and the classifying rules of problem are obtained.
If S=(U, A, V, are f) decision table, wherein:
U is domain, is a nonempty finite object set.A is the attribute set of nonempty finite object, A=C ∪ D,What C, D were each represented is conditional attribute and decision attribute.V=∪a∈AVaIt is property value, VaIt is attribute a value
Domain.f:U × A → V is referred to as information function, is that each attribute of each object gives a value of information.What f was represented is according to determination
Domain and attribute, obtain corresponding property value.
For decision table S=, (U, A, V, f), C and D are defined in two equivalence relations on U, and the D positive domains of C are denoted as POSC
(D)=∪X∈U/DC_ (X), if POSC(D)=POS(C-{b})(D) it in C is redundancy that, then attribute b, which is called, conversely, b is called to be in C
It is necessary.If all attributes are necessary in C, it is independent to claim C.If C D independent subsetsThere is POSC′(D)=
POSC(D), then C ' is called a C yojan, and the relation race that C all yojan are constituted is designated as redC(D).The friendship of all yojan in C
Collection is referred to as C D cores, i.e.,:coreD(C)=∩ redD(C)。
Comentropy
Comentropy is a viewpoint for being used for metric amount in information theory, is that a kind of description things is probabilistic important
Measure.In information system, the acquisition of entropy means the loss of information.The higher system of one order degree, then entropy get over
Small, contained information content is bigger;Otherwise, unordered degree is higher, then entropy is bigger, and information content is smaller.So, information and entropy
It is complementary, information is exactly negentropy.
The size of information gain is the basis for estimation that people construct decision tree nodes attribute.Information gain is bigger, and it is to opinion
The discrimination of element is bigger in domain, and the importance to information system is also bigger, and the important of attribute is also reflected to a certain degree
Degree, can as attribute weight a kind of building method.Due to information gain related algorithm can retain information content compared with
Big attribute, rejects the less attribute of information content, the superperformance with true reflection Attribute Significance.
If what U was represented is the domain of an information system, if domain is divided into m different type by decision attribute D;For
Some Di, i ∈ [1,2 ..., m], it includes element number and is designated as di, any one object belongs to DiProbability be di/|U|.Examine
Conditional attribute C is considered, for any a ∈ C, if attribute a value is a1, a2..., an, it divides domain U for v parts of { u1, u2...,
uv}.Wherein, ujSame value a is taken comprising attribute ajData row;ujInclude uijIndividual DiData object.According to attribute a value pair
Current data is divided the entropy that obtained information is known as attribute a.
Its calculation formula is:
Learnt from formula, H (a) essence is conditional entropies of the attribute a on decision attribute D, i.e. H (a)=H (D | a).
Attribute a information gain definition is Gain (a)=H (D)-H (a).What it was represented is according to decision attribute D probability point
The entropy of cloth formation with conditional attribute a is divided after, the difference between the entropy of obtained probability distribution formation.
Nicety of grading is evaluated
Classification in Remote Sensing Image precision evaluation is usually to be carried out with quantitative method in terms of the reliability and variability two of result
's.Fail-safe analysis process refers to the classification degree of accuracy calculating process in the case where certain probability ensures;Variability Analysis is statistics
On classification degree of accuracy interval estimation.Fail-safe analysis is it is emphasised that the degree of closeness of the classification degree of accuracy and truth value, variation
Property analysis be conceived to classification the degree of accuracy amplitude of variation.Therefore, nicety of grading in the narrow sense is exactly to classification accuracy variations
Quantitative description.Broad category precision includes two aspects of classification results reliability and variability quantitative expression.Classification in Remote Sensing Image knot
Fruit objectively needs quantitative reliability and Variability Analysis.Because how classification results are not only interested in user, and
It is the problem of data producer should be answered.This quantitative reliability and Variability Analysis process are exactly that nicety of grading is commented
Valency process.
Evaluation to Classification Precision of RS Images can be summarized as two classes, and one is based on laboratory level or bench-scale testing
And it is mainly used in the evaluation of sorting technique, technique study, two be the main clothes of the businessization operation based on region or country scale
It is engaged in the evaluation of production.Because the application field of evaluation object is different, evaluation method respectively has feature, and its applicability is also different, pin
Research to two class methods also shows different characteristics.
Differential matrix
The definition of differential matrix is defined as follows:
1) differential matrix of decision table is a symmetrical n rank square formation, wherein, matrix element mijIt is defined as:
Wherein Xi, Xj∈U
When building the differential matrix of decision table, only in Xi, XjOn the premise of being not belonging to same Decision Classes, mijIt is area
Divide Xi, XjAll properties set;If Xi, XjBelong to same Decision Classes, then element m in differential matrixijFor empty set.
1) C D cores are the m of all individual elements in differential matrixijAnd, i.e.,:
CORED(C)=a ∈ C | mij={ a } 1≤i, j≤n }
2) the differential matrix function of decision table is defined as follows:
ρ*=∧ { ∨ mij}
Each conjunction expression corresponds to C D yojan respectively in function ρ * minimal disjunctive normal form.
Experimental data
This experiment chooses 1997 in California, USA a part of AVIRIS EO-1 hyperions that Fitow remote sensing test block is not shot
Remotely-sensed data (as shown in Figure 2), wave band keeps count of as 224.Got rid of from original wave band is influenceed serious by steam noise etc.
Wave band (wave band 1~3,108~113,152~161,223~224), retain remaining 204 wave bands to be tested, preselect
4 kinds of types of ground objects, respectively water body, naked rock, vegetation and building, each type are chosen 10 points and classified, and domain is
40。
The some experimental data of table 1
Experimentation
1. important wave band and redundancy wave band are obtained by differential matrix.
The differential matrix of the decision table of table 2
Differential matrix is as follows:
ρ*=(B4 ∨ B5 ... ∨ B222) ∧ (B5 ∨ B6 ∨ ... ∨ B222) ∧ ... ∧ (B4 ∨ B5 ... ∨ B222)
=B3 ∧ B16 ∧ B19 ∧ B20 ∧ B21 ∧ B23 ∧ B26 ∧ B27 ∧ B29 ∧ B31 ∧ B33 ∧ B34 ∧ B35 ∧ B36
∧B38∧B40∧B42∧B46∧B47∧B49∧B50∧B53∧B54∧B55∧B56∧B57∧B58∧B60∧B61∧
B62∧B63∧B64∧B67∧B70∧B72∧B74∧B76∧B77∧B80∧B81∧B83∧B86∧B87∧B90∧
B91∧B92∧B94∧B95∧B102∧B103∧B105∧B115∧B116∧B117∧B119∧B121∧B123∧
B124∧B125∧B131∧B139∧B144∧B145∧B151∧B152∧B169∧B170∧B172∧B176∧B183
∧B184∧B186∧B187∧B188∧B209∧B211∧B212∧B213∧B214∧B215∧B217∧B218∧
B219∧B220
In differential matrix, 204 wave bands, there are 84 important attributes, remaining is conditional attribute C relative to decision-making
The omissible attributes of attribute D, that is, excluding redundancy rate is:(204-84)/204=58.82%.
2, by remaining 84 important wave bands, importance sorting are carried out according to comentropy formula
The wave band importance sorting of table 3
As shown in table 3, the sequence of importance is carried out to wave band,
3 select preceding 7 wave bands to carry out 204 wave bands after band overlapping, with denoising according to importance carries out time efficiency
Comparison.
The wave band time efficiency of table 4 compares
Fig. 3 is refer to, Fig. 3 is 7 wave bands and 204 wave band time efficiency comparison schematic diagrams.It can be apparent from from Fig. 2:Ripple
The time efficiency of 7 wave bands after section superposition, hence it is evident that the time efficiency than 204 wave bands after denoising is much higher.Particularly most
Time efficiency of the time efficiency of maximum-likelihood method, mahalanobis distance and SVMs well below 204 wave bands.
4 select preceding 7 wave bands to carry out band overlapping according to importance, 7 wave bands obtained with principal component analysis (PCA)
Nicety of grading is compared.
The confusion matrix of preceding 7 wave bands of the band overlapping of table 5
Precision is 295406/314368=93.9682% after obtained classification
Kappa coefficient=0.9043
The confusion matrix of the principal component analysis of table 6 (PCA)
Precision is 242922/314368=77.2731% after obtained classification
Kappa coefficient=0.6356
From table 5 and table 6, we can draw:Under the conditions of identical Spectral dimension, preceding 7 wave bands after band overlapping
Nicety of grading is better than PCA nicety of grading, and nicety of grading improves 16.70%, the dimension reduction method energy based on rough set and comentropy
Enough features being preferably extracted in high spectrum image, with stronger classification capacity, facilitate follow-up analyzing and processing.
A kind of dimension reduction method of new high-spectrum remote sensing data based on rough set of the present invention, this method passes through rough set first
The spectral signature for extracting EO-1 hyperion realizes high-spectrum image dimensionality reduction, that is, the redundant attributes between spectrum is removed, secondly by comentropy side
Remaining spectral information is ranked up by method by importance, the carry out Optimal Bands Selection based on information content principle, can be more preferable
Ground is extracted the feature in high spectrum image, with stronger classification capacity, facilitates follow-up analyzing and processing.
Described above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
Member, on the premise of the inventive method is not departed from, can also make some improvement and supplement, and these are improved and supplement also should be regarded as
Protection scope of the present invention.
Claims (9)
1. a kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set, it is characterised in that methods described includes following
Step:
Step S1:Original target in hyperspectral remotely sensed image is pre-processed, cancelling noise jammr band, preselect type of ground objects, determine it
Topological structure;
Step S2:Rough Set Attribute Reduction, redundancy wave band is removed using Rough Set, retains important wave band;
Step S3:Comentropy sorts, and carries out importance sorting to the important wave band in step S2 according to comentropy, filters out to dividing
Class result influences big band combination;
Step S4:Nicety of grading is carried out with PCA methods to be compared, verify dimensionality reduction effect by the band combination filtered out.
2. the dimension reduction method of the high-spectrum remote sensing data according to claim 1 based on rough set, it is characterised in that step S1
In pre-selection type of ground objects include water body, naked rock, vegetation and building.
3. the dimension reduction method of the high-spectrum remote sensing data according to claim 1 based on rough set, it is characterised in that step S2
In important wave band and redundancy wave band are obtained by differential matrix.
4. the dimension reduction method of the high-spectrum remote sensing data according to claim 1 based on rough set, it is characterised in that step S4
The middle wave band used described in PCA methods is by the wave band that is obtained after denoising in step S1.
5. the dimension reduction method of the high-spectrum remote sensing data according to claim 1 based on rough set, it is characterised in that step S4
Middle nicety of grading compares the comparison for needing to carry out time efficiency, and the comparison of its time efficiency includes maximum likelihood, minimum range, put down
Row hexahedron, mahalanobis distance, binary coding, SVMs.
6. the dimension reduction method of the high-spectrum remote sensing data according to claim 1 based on rough set, it is characterised in that step S4
In be that sorted precision is obtained by confusion matrix.
7. the dimension reduction method of the high-spectrum remote sensing data according to claim 1 based on rough set, it is characterised in that step S2
In Rough Set Attribute Reduction definition method it is as follows:
If S=(U, A, V, are f) decision table, wherein:
U is domain, is a nonempty finite object set, and A is the attribute set of nonempty finite object, A=C ∪ D,What C, D were each represented is conditional attribute and decision attribute;V=∪a∈AVaIt is property value, VaIt is attribute a value
Domain;f:U × A → V is referred to as information function, is that each attribute of each object gives a value of information;What f was represented is according to determination
Domain and attribute, obtain corresponding property value.
For decision table S=, (U, A, V, f), C and D are defined in two equivalence relations on U, and the D positive domains of C are denoted as POSC(D)=
∪X∈U/DC_ (X), if POSC(D)=POS(C-{b})(D) it in C is redundancy that, then attribute b, which is called, conversely, it in C is necessary that b, which is called,
's;If all attributes are necessary in C, it is independent to claim C;If C D independent subsetsThere is POSC′(D)=POSC
(D), then C ' is called a C yojan, and the relation race that C all yojan are constituted is designated as redC(D);The common factor of all yojan in C
Referred to as C D cores, i.e.,:coreD(C)=∩ redD(C)。
8. the dimension reduction method of the high-spectrum remote sensing data according to claim 1 based on rough set, it is characterised in that step S3
In comentropy definition method it is as follows:
If what U was represented is the domain of an information system, if domain is divided into m different type by decision attribute D;For some
Di, i ∈ [1,2 ..., m], it includes element number and is designated as di, any one object belongs to DiProbability be di/|U|;Consider bar
Part attribute C, for any a ∈ C, if attribute a value is a1, a2..., an, it divides domain U for v parts of { u1, u2..., uv};
Wherein, ujSame value a is taken comprising attribute ajData row;ujInclude uijIndividual Di data objects;According to attribute a value to current
Data are divided the entropy that obtained information is known as attribute a;Its calculation formula is:
Learnt from formula, H (a) essence is conditional entropies of the attribute a on decision attribute D, i.e. H (a)=H (D | a);
Attribute a information gain definition is Gain (a)=H (D)-H (a);What it was represented is according to decision attribute D probability distribution shapes
Into entropy with conditional attribute a is divided after, the difference between the entropy of obtained probability distribution formation.
9. the dimension reduction method of the high-spectrum remote sensing data according to claim 3 based on rough set, it is characterised in that difference square
The definition method of battle array is as follows:
The differential matrix of decision table is a symmetrical n rank square formation, wherein, matrix element mijIt is defined as:
Wherein Xi, Xj∈U
When building the differential matrix of decision table, only in Xi, XjOn the premise of being not belonging to same Decision Classes, mijIt is to discriminate between Xi,
XjAll properties set;If Xi, XjBelong to same Decision Classes, then element m in differential matrixijFor empty set.
C D cores are the m of all individual elements in differential matrixijAnd, i.e.,:
CORED(C)=a ∈ C | mij={ a } 1≤i, j≤n }
The differential matrix function of decision table is defined as follows:
ρ*=∧ { ∨ mij}
Each conjunction expression corresponds to C D yojan respectively in function ρ * minimal disjunctive normal form.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710209001.0A CN106971204A (en) | 2017-03-31 | 2017-03-31 | A kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710209001.0A CN106971204A (en) | 2017-03-31 | 2017-03-31 | A kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106971204A true CN106971204A (en) | 2017-07-21 |
Family
ID=59336939
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710209001.0A Pending CN106971204A (en) | 2017-03-31 | 2017-03-31 | A kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106971204A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107909062A (en) * | 2017-12-11 | 2018-04-13 | 海南大学 | A kind of Algorithm for Discretization of Remote Sensing Image Attribute and system based on comentropy |
CN108197656A (en) * | 2018-01-04 | 2018-06-22 | 西南交通大学 | A kind of attribute reduction method based on CUDA |
CN108763377A (en) * | 2018-05-18 | 2018-11-06 | 郑州轻工业学院 | Multi-source telemetering big data feature extraction preprocess method is diagnosed based on satellite failure |
CN109241819A (en) * | 2018-07-07 | 2019-01-18 | 西安电子科技大学 | Based on quickly multiple dimensioned and joint template matching multiple target pedestrian detection method |
CN111651720A (en) * | 2020-06-11 | 2020-09-11 | 北京理工大学 | Multispectral reconstruction method and device based on L4 norm optimization |
CN112446256A (en) * | 2019-09-02 | 2021-03-05 | 中国林业科学研究院资源信息研究所 | Vegetation type identification method based on deep ISA data fusion |
WO2022127075A1 (en) * | 2020-12-15 | 2022-06-23 | 海南大学 | Feature discretization method for remote sensing image on the basis of rough fuzzy model |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120183225A1 (en) * | 2010-11-24 | 2012-07-19 | Indian Statistical Institute | Rough wavelet granular space and classification of multispectral remote sensing image |
CN102879677A (en) * | 2012-09-24 | 2013-01-16 | 西北工业大学 | Intelligent fault diagnosis method based on rough Bayesian network classifier |
CN105893748A (en) * | 2016-03-29 | 2016-08-24 | 安徽大学 | Happy index evaluation method and system based on rough set theory |
-
2017
- 2017-03-31 CN CN201710209001.0A patent/CN106971204A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20120183225A1 (en) * | 2010-11-24 | 2012-07-19 | Indian Statistical Institute | Rough wavelet granular space and classification of multispectral remote sensing image |
CN102879677A (en) * | 2012-09-24 | 2013-01-16 | 西北工业大学 | Intelligent fault diagnosis method based on rough Bayesian network classifier |
CN105893748A (en) * | 2016-03-29 | 2016-08-24 | 安徽大学 | Happy index evaluation method and system based on rough set theory |
Non-Patent Citations (5)
Title |
---|
丁华: "《基于知识工程的电牵引采煤机现代设计》", 31 December 2015, 国防工业出版社 * |
刘盾: "基于粗糙集理论的多属性决策方法", 《万方在线学位论文》 * |
孟科等: "基于粗糙集的RBF神经网络在装备研制费用预测中的应用", 《装备指挥技术学院学报》 * |
景运革: "《基于粗糙集属性约简算法的研究》", 31 July 2013, 西南交通大学出版社 * |
毕景芝: "基于粗糙集的高光谱遥感数据降维算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107909062A (en) * | 2017-12-11 | 2018-04-13 | 海南大学 | A kind of Algorithm for Discretization of Remote Sensing Image Attribute and system based on comentropy |
CN107909062B (en) * | 2017-12-11 | 2021-05-07 | 海南大学 | Information entropy-based remote sensing image feature discretization method and system |
CN108197656A (en) * | 2018-01-04 | 2018-06-22 | 西南交通大学 | A kind of attribute reduction method based on CUDA |
CN108197656B (en) * | 2018-01-04 | 2020-07-28 | 西南交通大学 | CUDA-based attribute reduction method |
CN108763377A (en) * | 2018-05-18 | 2018-11-06 | 郑州轻工业学院 | Multi-source telemetering big data feature extraction preprocess method is diagnosed based on satellite failure |
CN109241819A (en) * | 2018-07-07 | 2019-01-18 | 西安电子科技大学 | Based on quickly multiple dimensioned and joint template matching multiple target pedestrian detection method |
CN112446256A (en) * | 2019-09-02 | 2021-03-05 | 中国林业科学研究院资源信息研究所 | Vegetation type identification method based on deep ISA data fusion |
CN111651720A (en) * | 2020-06-11 | 2020-09-11 | 北京理工大学 | Multispectral reconstruction method and device based on L4 norm optimization |
CN111651720B (en) * | 2020-06-11 | 2023-03-31 | 北京理工大学 | Multispectral reconstruction method and device based on L4 norm optimization |
WO2022127075A1 (en) * | 2020-12-15 | 2022-06-23 | 海南大学 | Feature discretization method for remote sensing image on the basis of rough fuzzy model |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106971204A (en) | A kind of dimension reduction method of the new high-spectrum remote sensing data based on rough set | |
Dong et al. | A novel hybrid genetic algorithm with granular information for feature selection and optimization | |
Ala'raj et al. | A new hybrid ensemble credit scoring model based on classifiers consensus system approach | |
Khaidem et al. | Predicting the direction of stock market prices using random forest | |
Gupta et al. | Comparing the performance of machine learning algorithms using estimated accuracy | |
Sha et al. | Semisupervised classification for hyperspectral images using graph attention networks | |
Alshammari et al. | Refining a k-nearest neighbor graph for a computationally efficient spectral clustering | |
Fu et al. | Low-level feature extraction for edge detection using genetic programming | |
Cardoso et al. | Financial credit analysis via a clustering weightless neural classifier | |
Zhuang et al. | Representation learning via semi-supervised autoencoder for multi-task learning | |
Lamba et al. | Classification of plant diseases using machine and deep learning | |
CN112668688B (en) | Intrusion detection method, system, equipment and readable storage medium | |
Salhi et al. | Fast and efficient face recognition system using random forest and histograms of oriented gradients | |
Ali et al. | Detection of deficiency of nutrients in grape leaves using deep network | |
Klikowski et al. | Deterministic sampling classifier with weighted bagging for drifted imbalanced data stream classification | |
Tschannerl et al. | Segmented autoencoders for unsupervised embedded hyperspectral band selection | |
Jivani et al. | The adept K-nearest neighbour algorithm-an optimization to the conventional K-nearest neighbour algorithm | |
Kappaun et al. | Evaluating binary encoding techniques for wisard | |
He et al. | Self-Adaptive bagging approach to credit rating | |
Başaran et al. | Data clustering | |
De Araujo et al. | Automatic cluster labeling based on phylogram analysis | |
Dutta et al. | Improved outlier detection using sparse coding-based methods | |
Georgieva et al. | Classification of healthy and diseased vine leaves using color features | |
Swain et al. | Dimensionality reduction and classification in hyperspectral images using deep learning | |
Chava et al. | Hyper-parameters tuning for hyperspectral image classification |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170721 |