CN105303198B - A kind of remote sensing image semisupervised classification method learnt from fixed step size - Google Patents
A kind of remote sensing image semisupervised classification method learnt from fixed step size Download PDFInfo
- Publication number
- CN105303198B CN105303198B CN201510788769.9A CN201510788769A CN105303198B CN 105303198 B CN105303198 B CN 105303198B CN 201510788769 A CN201510788769 A CN 201510788769A CN 105303198 B CN105303198 B CN 105303198B
- Authority
- CN
- China
- Prior art keywords
- sample
- remote sensing
- classification
- sensing image
- marked
- 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.)
- Expired - Fee Related
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
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
Abstract
The present invention relates to a kind of remote sensing image semisupervised classification methods learnt from fixed step size, include the following steps:Remote sensing image is pre-processed, the marked sample of all kinds of atural objects is obtained;Selected section or all unmarked sample combine the marked sample structure sparse graph, and carry out initial category calibration to unmarked sample, to expand the quantity of marked sample;Based on from fixed step size learning algorithm, the classification information with initial markers sample after expansion is selected or rejected;Supervised classifier is selected, remote sensing image classify by pixel.The present invention is expanding number of training purpose simultaneously, reduces the purpose that sample is accidentally demarcated in training sample.
Description
Technical field
The present invention relates to a kind of remote sensing image semisupervised classification methods learnt from fixed step size.
Background technology
Remotely-sensed data feature efficient with its, economic, synchronizing on a large scale, has become earth resource environmental monitoring, planning
One of with the important technical of management, and it is widely used in region or even global range.Classification of remote-sensing images
Cardinal principle be that show different spectrum, texture and space geometry characteristic etc. in remote sensing image according to type of ground objects special
Sign, to carry out the identification of terrain object attribute classification.There are mainly two types of mode, visual interpretation and computers point for classification of remote-sensing images at present
Class.Wherein computer automatic sorting has the advantages that efficient, objective, is extensive, repeated Remote Sensing Data Processing and classification
Effective ways.
It is well known that during the computer classes of remote sensing image, how to obtain sufficient amount of training sample is meter
The critical issue of calculation machine image classification.It is to interpret manually to select in image by visual observation to obtain the major way of training sample at present
It takes or live on-site inspection samples.The specialized capability for interpreting expert is wanted however, obtaining training sample in a manner of visual interpretation
Ask higher, and subjective degree is larger, largely there may be certain errors when selection sample;On the other hand, with live on-site inspection
Though mode obtain training sample and can obtain the training sample of degree of precision, working efficiency is low, and field investigation
It is costly, the general selection for being only applicable in a small amount of training sample.It is indicated above in the computer classes of remote sensing image, training sample
Acquisition will face two aspect the problem of:(1) or training sample data have certain error or noise;It (2) or can only
Obtain a small amount of high-precision sample.Both of which is unfavorable for the classification processing of remote sensing image, this is because if sample number
According to containing error either quantity is very few will all cause used in sample is under-represented to certain ground class or representative presence
The problem of deviation.For the processing of the above high dimensional data only small sample the case where, remote sensing circle propose using in image it is a large amount of not
The sample information of label carries out certain shape using the semi-supervised treatment technology in machine learning to insufficient amount of marked sample
The supplement of formula, to achieve the purpose that carry out deviation correction to representative bad ground class.
The common semi-supervised technology of remote sensing image processing mainly has following four mode:Based on generate model method, from
Study mutually learns, the method for direct-push and figure.The process of wherein semi-supervised self study is a kind of mistake of grader recurrence fitting
Journey, the nonstandard random sample that setting confidence threshold value will be only met in each recursive procedure are originally brought into marked sample set, and
Participate in next recurrence fitting.Method based on figure does not need parameter fitting process, but utilizes certain similarity criterion to participating in transporting
The unmarked sample calculated directly is differentiated, thus has the characteristics that high-efficient simple.However, self study or drawing method is semi-supervised
Learning method has respective limitation, for example the accumulation being easy to cause error based on self study grader recurrence fit procedure is passed
It broadcasts;And the graph model based on marked sample similarity is easy to be limited by initial sample information.
Invention content
In view of this, the purpose of the present invention is to provide a kind of from the remote sensing image semisupervised classification side that fixed step size learns
Method is expanding number of training purpose simultaneously, reduces the purpose that sample is accidentally demarcated in training sample.
To achieve the above object, the present invention adopts the following technical scheme that:A kind of half prison of the remote sensing image learnt from fixed step size
Superintend and direct sorting technique, it is characterised in that include the following steps:
Step S1:Remote sensing image is pre-processed, the marked sample of all kinds of atural objects is obtained;
Step S2:Selected section or all unmarked sample combine the marked sample and carry out sparse composition, and to not
Marker samples carry out initial category calibration, to expand the quantity of marked sample;
Step S3:Based on from fixed step size learning algorithm, the classification information with initial markers sample after expansion is selected
It selects or rejects;
Step S4:Supervised classifier is selected, remote sensing image classify by pixel.
Further, the particular content of the step S1 includes:
Step S11:The remote sensing image of acquisition is pre-processed, the pretreatment includes geometry and radiant correction, image
Splicing is extracted with cutting, visual fusion enhancing and image characteristics;
Step S12:It obtains per the marked sample of class atural object in remote sensing image, and records the classification of the marked sample
Mark, the remote sensing image include c atural object classification, ci∈ (1,2 ..., c) be the i-th class sample classification logotype, Xl=[X1
X2 ... Xc] be marked sample set, whereinIndicate the sample of i classifications, wherein c, i are
Natural number, niFor the number of samples of the i-th classification, xijFor a column vector, the spectral vector of i-th j-th of sample of class atural object is indicated
Or feature vector;
Step S13:To marked sample set XlWith the spectral vector or feature vector x of unmarked sampleijIt is normalized,
So that two norm value of pixel spectral vector is 1.
Further, the particular content of the step S2 includes:
Step S21:Selected section or whole unmarked samples are denoted as X as unmarked sample setu, wherein in subset not
The number of marker samples is denoted as Nu;
Step S22:Combine marked sample set XlWith unmarked sample Xu, construct new matrix D ic=[X1 Xu]B×N,
Middle B is the spectral band number of remote sensing image, and N is marked sample collection XlWith unmarked sample set XuThe sum of number of samples;
Step S23:To matrix D ic={ yi| i=1,2 ..., N in each column element yi, resolve sparse model
WhereinOperation be that s.t indicates to solve to be subject to when the equation to square of respective vectors or matrix F-norm
Restrictive condition, | | αi||0≤K0Indicate αiThe number of middle nonzero element is not more than K0,αiiIndicate αiIn i-th of element, wait owning
YiAfter (i=1,2 ..., N) is calculated, corresponding sparse coding αiSequential storage is to matrix A=[α1 α2 ... αN];
Step S24:Construction similarity matrix W=(| A |+| AT|)/2, wherein ATIndicate the transposition operation to A, | A | it indicates
It takes absolute value operation to matrix A;
Step S25:The initial labels that unmarked sample is predicted using label pass-algorithm, using marked sample come pre-
Survey the classification information of unmarked sample
Step S26:LuIn each row vector liEach unmarked sample z is corresponded toiCategory attribute, and by each
Unmarked sample is added to corresponding marked sample set X according to the classification information of predictionlIn subset of all categories, i.e. X'i=Xi∪
{zi, to tentatively obtain markd new samples set X'l=[X'1 X'2 ... X'c], wherein Xi∪{ziIndicate sample zi
It is added to subset X by " simultaneously " operation of setiIn.
Further, the particular content of the step S3 includes:
Step S31:Count the new samples set X'l=[X'1 X'2 ... X'c] in number of all categories, be denoted as n'i,
I=1,2 ... c;
Step S32:Initialize Mark-matrix S=[S of Different categories of samples1 S2 ... Sc], wherein the label of the i-th classification beWherein n'i> ni,sij={ 0,1 }, i=1,2 ..., c, j=1,2 ..., n'i, indicate such as
Fruit sample point sijIt is selected then be 1, it is unselected then be 0, initialize the Mark-matrix S method be X'lIn come from Xl
Sample labeling be 1, remaining it is extended produced by sample labeling be 0;
Step S33:According to the Mark-matrix S, selects to mark the sample for corresponding to 1 in every group, utilize selected prison
Classifier training sorter model is superintended and directed, and goes out every group of X' using the model predictioniThe category attribute of sample in subset, calculates every group
In penalty values after all sample classifications, U=[U1 U2 ... Uc], whereinLoss function
Computational methods are:uij=1-Pij, wherein PijThe posterior probability classified under train classification models for i-th group of j-th of sample.
Step S34:Samples selection Mark-matrix S is updated, according to the threshold value λ of initial setting up, to every group of loss function value UiIn
It is 1 less than the sample labeling at position corresponding to the threshold value λ, remaining is 0;
Step S35:Increase the step-length of threshold value λ:λ=μ * λ, the wherein value of μ are the constant more than 1;
Step S36:Iteration executes step S32 to step S35, when iterations reach setting number or λ more than 0.5,
Then iterative process terminates;
Step S37:According to the S of last every group of sample labelingiSituation marks all as corresponding sample as last
The training sample set of classification, and it is recorded as newX=[newX1 newX2 ... newXc], wherein newXiMost for the i-th classification
Whole marked sample.
Further, the supervised classifier in the step S33 is maximum likelihood classifier or support vector machine classifier.
Further, the value range of λ is (0.01,0.1) in the step S34.
Further, μ=1.1 in the step S35.
Further, the particular content of the step S4 is:
Step S41:Common supervised classifier is selected, supervised classifier herein is SVM or k- nearest neighbour classification devices, profit
Sorter model is trained with updated new samples newX;
Step S42:Using trained sorter model, the supervised classification by pixel is carried out to raw video data.
The present invention has the advantages that compared with prior art:
1, the present invention has universality, is suitable for various types of classification of remote-sensing images and handles, is also applied for various classification
The semisupervised classification algorithm of device.For example, for target in hyperspectral remotely sensed image, can be selected in self study grader recurrence fit procedure
Select support vector machines (SVM) grader etc.;Maximum likelihood classifier can then be used for the multispectral image of middle low resolution
Etc.;
2, the present invention is small for initial markers sample requirement amount, and the present invention has preferable samples selection and rejecting mechanism, adopts
With the quantity of mode Dynamic Recurrent exptended sample from the easier to the more advanced;
3, anti-noise ability of the present invention is strong, and during taxonomy model of the invention, the same of nonstandard random sample sheet is selected expanding
When, sample of poor quality, that noise is big also can be dynamically rejected, therefore, method of the invention has stronger anti-noise ability;
4, the expansible application of the present invention general does not demarcate Method of Sample Selection and rejecting machine process provides a kind of
System, in addition to it can apply semisupervised classification, applies also for the image processing of image feature feature extraction and dimensionality reduction, target acquisition etc..
Description of the drawings
Fig. 1 is the method for the present invention flow chart.
Fig. 2 is that the present invention is based on the samples of sparse patterning process to expand flow chart.
Fig. 3 is that the present invention is based on the samples selections learnt from fixed step size and elimination method flow chart.
Specific implementation mode
The present invention will be further described with reference to the accompanying drawings and embodiments.
As shown in Figure 1, the present invention provides a kind of remote sensing image semisupervised classification method learnt from fixed step size, feature exists
In including the following steps:
Step S1:Remote sensing image is pre-processed, obtain a small amount of qualified all kinds of atural objects has label (classification information)
Sample, particular content include:
Step S11:According to the quality of image in the remote sensing image data source of acquisition, the remote sensing image of acquisition is located in advance
Reason, the pretreatment include geometry and radiant correction, image joint and cutting, visual fusion enhancing and feature extraction;
Step S12:With being obtained according to the data of field factual survey or in a manner of visual interpretation in remote sensing image per class
The fraction of marked sample of object, and record the classification logotype of the marked sample, comprising c a the species of the remote sensing image
Not, ci∈ (1,2 ..., c) be the i-th class sample classification logotype, Xl=[X1 X2 ... Xc] be marked sample set,
InIndicate that the sample of i classifications, corresponding category label are
Wherein, c, i are natural number, niFor the number of samples of the i-th classification, xijFor a column vector, the i-th class atural object the is indicated
The spectral vector or feature vector of j sample;
Step S13:To all marked sample set XlWith the spectral vector or feature vector x of unmarked sampleijReturned
One changes so that each pixel vector meetsI.e. two norm value of pixel spectral vector is 1.
Step S2:Selected section or all unmarked sample are combined the sparse composition of the marked sample progress and (are utilized dilute
Dredge expression technology construction sample similarity figure), and initial category calibration is carried out to unmarked sample, to expand marked sample
Quantity, as shown in Fig. 2, particular content includes:
Step S21:According to remote sensing image actual size, depending on computational efficiency situation from remote sensing image selected section or whole
Unmarked sample is denoted as X as unmarked sample setu, wherein the number of unmarked sample is denoted as N in subsetu;
Step S22:Combine marked sample set XlWith unmarked sample Xu, construct new matrix D ic=[X1 Xu]B×N,
Middle B is the spectral band number or number of features of remote sensing image, and N is marked sample collection XlWith unmarked sample set XuSample
The sum of number;
Step S23:To Dic={ yi| i=1,2 ..., N in each column element yi, resolve sparse model
WhereinOperation be that s.t indicates to solve to be subject to when the equation to square of respective vectors or matrix F-norm
Restrictive condition, | | αi||0≤K0Indicate αiThe number of middle nonzero element is not more than K0, αiiIndicate αiI-th of element in vector, ginseng
Number K0=20, then use Open-Source Tools SPAMS to can get sparse expression coefficient αi, wait for all yi(i=1,2 ..., N) it calculates
After, corresponding sparse coding αiSequential storage is to matrix A=[α1 α2 ... αN];
Step S24:Construction similarity matrix W=(| A |+| AT|)/2, wherein ATIndicate the transposition operation to A, | A | it indicates
It takes absolute value operation to matrix A;Since W can be further represented asThus each column in W matrixes (or
Often go) numerical response of vector figure connection (similar) situation of corresponding sample in entire Dic, wherein 0 indicates not connected, it is non-
0 indicates connection;
Step S25:The initial labels of unmarked sample are predicted using label (classification information) pass-algorithm, using having marked
Sample is remembered to predict the classification information of unmarked sampleMethod is as follows:Label information is first converted into matrix form
Ll, then Lu=(Duu-Wuu)-1*Wul*Ll, wherein D is diagonal matrix, and make is to be carried out by row accumulation operations, i.e., to W matrixesLlFor the label information (L of marked samplei=[0 ..., 1 ..., 0] indicate i-th
It is set to 1, remaining is 0) result of calculationThe initial markers information assigned after being computed for unmarked sample.
Step S26:LuIn each row vector liEach unmarked sample z is corresponded toiCategory attribute, determine rule
It is then i=arg mini li, each unmarked sample is added to correspondence by i=1,2 ..., c according to the classification information of prediction
Marked sample set XlIn subset of all categories, i.e. X'i=Xi∪{zi(it is assumed that ziBelong to the i-th class), to tentatively be had
The new samples set X' of labell=[X'1 X'2 ... X'c], wherein Xi∪{ziIndicate sample ziPass through " simultaneously " operation of set
It is added to subset XiIn.
Step S3:Based on from fixed step size learning algorithm, the classification information with initial markers sample after expansion is selected
It selects or rejects, as shown in figure 3, particular content includes:
Step S31:Count the new samples set X'l=[X'1 X'2 ... X'c] in number of all categories, be denoted as n'i,
I=1,2 ... iterations parameter t=0, initial abstraction threshold value λ=0.01 and threshold increase steps μ=1.1 are arranged in c;
Step S32:Initialize Mark-matrix S=[S of Different categories of samples1 S2 ... Sc], wherein the label of the i-th classification beWherein n'i> ni,sij={ 0,1 }, i=1,2 ..., c, j=1,2 ..., n'i, indicate such as
Fruit sample point sijIt is selected then be 1, it is unselected then be 0, initialize the Mark-matrix S method be X'lIn come from Xl
Sample labeling be 1, remaining it is extended produced by sample labeling be 0;
Step S33:According to the Mark-matrix S, selects to mark the sample for corresponding to 1 in every group, utilize selected prison
Classifier training sorter model is superintended and directed, the supervised classifier is maximum likelihood classifier or SVM classifier (if selecting SVM
Grader then needs that SVM output valves are converted into posterior probability using logistic regression method), and gone out often using the model prediction
Group X'iThe category attribute of sample in subset calculates the penalty values after all sample classifications in every group, U=[U1 U2 ... Uc],
WhereinLoss function computational methods are:uij=1-Pij, wherein PijFor i-th group of j-th of sample
The posterior probability classified under train classification models.
Step S34:Update samples selection Mark-matrix S, update selected marker SiThe method of (i=1,2 ..., c) is evidence
The value range of the threshold value λ, λ of initial setting up are 0.01~0.1, to every group of loss function value UiIn be less than the threshold value λ institute it is right
It is 1 to answer the sample labeling at position, remaining is 0;
Step S35:Iterations t=t+1 is enabled, the step-length of threshold value λ is increased:λ=μ * λ;
Step S36:Iteration executes step S32 to step S35, when iterations reach setting number or λ more than 0.5,
Then iterative process terminates;
Step S37:According to the S of last every group of sample labelingiSituation marks all as corresponding sample as last
The training sample set of classification, and it is recorded as newX=[newX1 newX2 ... newXc], wherein newXiIt is final for i-th group
Marked sample.
Step S4:Supervised classifier is selected, remote sensing image by pixel classify, particular content is:
Step S41:Common supervised classifier is selected, supervised classifier herein is SVM or k-NN, and utilization is updated
New samples newX trains sorter model;
Step S42:Using trained sorter model, the supervised classification by pixel is carried out to raw video data.
The purposes of the present invention essentially consists in:In the case where marked sample is less for remote sensing image, quantity is automatically provided
Enough, the higher sample data of precision.In the supervised classification of remote sensing image, feature extraction and dimensionality reduction, target acquisition etc. tool
It plays an important role, and has use value with production practices in teaching and scientific research.
The classification of remote-sensing images learnt from fixed step size is used together general with ENVI softwares.This is realized using IDL language
The illustrated method of invention, makes card format, as ENVI by the remote sensing image semisupervised classification method learnt from fixed step size
The Function Extension of software can also make separate functional blocks, provide that quantity is enough, the higher sample data file of precision.
The foregoing is merely presently preferred embodiments of the present invention, all equivalent changes done according to scope of the present invention patent with
Modification should all belong to the covering scope of the present invention.
Claims (6)
1. a kind of remote sensing image semisupervised classification method learnt from fixed step size, it is characterised in that include the following steps:
Step S1:Remote sensing image is pre-processed, the marked sample of all kinds of atural objects is obtained;
Step S2:Selected section or all unmarked sample, the joint marked sample carry out sparse composition, and to unmarked
Sample carries out initial category calibration, to expand the quantity of marked sample;
Step S3:Based on from fixed step size learning algorithm, to the classification information with initial markers sample after expansion carry out selection or
It rejects;
Step S4:Supervised classifier is selected, remote sensing image classify by pixel;
The particular content of the step S1 includes:
Step S11:The remote sensing image of acquisition is pre-processed, the pretreatment includes geometry and radiant correction, image joint
It is extracted with cutting, visual fusion enhancing and image characteristics;
Step S12:It obtains per the marked sample of class atural object in remote sensing image, and records the classification mark of the marked sample
Know, the remote sensing image includes c atural object classification, ci∈ (1,2 ..., c) be the i-th class sample classification logotype, Xl=[X1 X2
... Xc] be marked sample set, whereinIndicate the sample of i classifications, wherein c, i are certainly
So number, niFor the number of samples of the i-th classification, xijFor a column vector, indicate i-th j-th of sample of class atural object spectral vector or
Feature vector;
Step S13:To marked sample set XlWith the spectral vector or feature vector x of unmarked sampleijIt is normalized so that
Two norm value of pixel spectral vector is 1;
The particular content of the step S2 includes:
Step S21:Selected section or whole unmarked samples are denoted as X as unmarked sample setu, unmarked sample wherein in subset
This number is denoted as Nu;
Step S22:Combine marked sample set XlWith unmarked sample Xu, construct new matrix D ic=[X1 Xu]B×N, wherein B
For the spectral band number of remote sensing image, N is marked sample set XlWith unmarked sample set XuThe sum of number of samples;
Step S23:To matrix D ic={ yi| i=1,2 ..., N in each column element yi, resolve sparse model
WhereinOperation be to square of respective vectors or matrix F-norm, by being limited when s.t indicates to solve the equation
Condition, | | αi||0≤K0Indicate αiThe number of middle nonzero element is not more than K0,αiiIndicate αiIn i-th of element, wait for all yi
After (i=1,2 ..., N) is calculated, corresponding sparse coding αiSequential storage is to matrix A=[α1 α2 ... αN];
Step S24:Construction similarity matrix W=(| A |+| AT|)/2, wherein ATIndicate the transposition operation to A, | A | it indicates to square
Battle array A takes absolute value operation;
Step S25:The initial labels that unmarked sample is predicted using label pass-algorithm are predicted not using marked sample
The classification information of marker samples
Step S26:LuIn each row vector liEach unmarked sample z is corresponded toiCategory attribute, and each is not marked
Note sample is added to corresponding marked sample set X according to the classification information of predictionlIn subset of all categories, i.e. X'i=Xi∪{zi,
To tentatively obtain marked new samples set X'l=[X'1 X'2 ... X'c], wherein Xi∪{ziIndicate sample ziPass through
" simultaneously " operation of set is added to subset XiIn.
2. the remote sensing image semisupervised classification method according to claim 1 learnt from fixed step size, it is characterised in that:It is described
The particular content of step S3 includes:
Step S31:Count the new samples set X'l=[X'1 X'2 ... X'c] in number of all categories, be denoted as n'i, i=
1,2,...c;
Step S32:Initialize Mark-matrix S=[S of Different categories of samples1 S2 ... Sc], wherein the label of the i-th classification beWherein n'i> ni,sij={ 0,1 }, i=1,2 ..., c, j=1,2 ..., n'i, indicate such as
Fruit sample point sijIt is selected then be 1, it is unselected then be 0, initialize the Mark-matrix S method be X'lIn come from Xl
Sample labeling be 1, remaining it is extended produced by sample labeling be 0;
Step S33:According to the Mark-matrix S, selects to mark the sample for corresponding to 1 in every group, utilize selected supervision point
Class device trains sorter model, and goes out every group of X' using the model predictioniThe category attribute of sample in subset calculates institute in every group
There are the penalty values after sample classification, U=[U1 U2 ... Uc], whereinLoss function calculates
Method is:uij=1-Pij, wherein PijThe posterior probability classified under train classification models for i-th group of j-th of sample;
Step S34:Samples selection Mark-matrix S is updated, according to the threshold value λ of initial setting up, to every group of loss function value UiIn be less than institute
The sample labeling stated at position corresponding to threshold value λ is 1, remaining is 0;
Step S35:Increase the step-length of threshold value λ:λ=μ * λ, the wherein value of μ are the constant more than 1;Step 3.6:Iteration executes step
Rapid S32 to step S35, when iterations reach setting number or λ more than 0.5, then iterative process terminates;
Step S37:According to the S of last every group of sample labelingiSituation will have been labeled as 1 corresponding sample as finally classifying
Training sample set, and it is recorded as newX=[newX1 newX2 ... newXc], wherein newXiFinal for the i-th classification has been marked
Remember sample.
3. the remote sensing image semisupervised classification method according to claim 2 learnt from fixed step size, it is characterised in that:It is described
Supervised classifier in step S33 is maximum likelihood classifier or support vector machine classifier.
4. the remote sensing image semisupervised classification method according to claim 2 learnt from fixed step size, it is characterised in that:It is described
The value range of λ is (0.01,0.1) in step S34.
5. the remote sensing image semisupervised classification method according to claim 2 learnt from fixed step size, it is characterised in that:It is described
μ=1.1 in step S35.
6. the remote sensing image semisupervised classification method according to claim 2 learnt from fixed step size, it is characterised in that:It is described
The particular content of step S4 is:
Step S41:Common supervised classifier is selected, supervised classifier herein is SVM or k- nearest neighbour classification devices, using more
New samples newX after new trains sorter model;
Step S42:Using trained sorter model, the supervised classification by pixel is carried out to raw video data.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510788769.9A CN105303198B (en) | 2015-11-17 | 2015-11-17 | A kind of remote sensing image semisupervised classification method learnt from fixed step size |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510788769.9A CN105303198B (en) | 2015-11-17 | 2015-11-17 | A kind of remote sensing image semisupervised classification method learnt from fixed step size |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105303198A CN105303198A (en) | 2016-02-03 |
CN105303198B true CN105303198B (en) | 2018-08-17 |
Family
ID=55200440
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510788769.9A Expired - Fee Related CN105303198B (en) | 2015-11-17 | 2015-11-17 | A kind of remote sensing image semisupervised classification method learnt from fixed step size |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105303198B (en) |
Families Citing this family (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106250836B (en) * | 2016-07-26 | 2018-09-14 | 山东师范大学 | Two benches facial image sorting technique and system under a kind of condition of small sample |
CN106650780B (en) | 2016-10-18 | 2021-02-12 | 腾讯科技(深圳)有限公司 | Data processing method and device, classifier training method and system |
IT201700021585A1 (en) * | 2017-02-27 | 2018-08-27 | St Microelectronics Srl | CORRESPONDENT LEARNING PROCEDURE, SYSTEM, DEVICE AND COMPUTER PRODUCT |
CN107239741B (en) * | 2017-05-10 | 2020-09-08 | 厦门瞳景智能科技有限公司 | Single-sample face recognition method based on sparse reconstruction |
CN107203755B (en) * | 2017-05-31 | 2021-08-03 | 中国科学院遥感与数字地球研究所 | Method, device and system for automatically adding new time sequence mark samples of remote sensing images |
CN107194373A (en) * | 2017-06-16 | 2017-09-22 | 河海大学 | A kind of target in hyperspectral remotely sensed image feature extraction and classifying method |
CN107657282A (en) * | 2017-09-29 | 2018-02-02 | 中国石油大学(华东) | Peptide identification from step-length learning method |
CN107944479B (en) * | 2017-11-16 | 2020-10-30 | 哈尔滨工业大学 | Disease prediction model establishing method and device based on semi-supervised learning |
CN107958216A (en) * | 2017-11-27 | 2018-04-24 | 沈阳航空航天大学 | Based on semi-supervised multi-modal deep learning sorting technique |
CN107909120A (en) * | 2017-12-28 | 2018-04-13 | 南京理工大学 | Based on alternative label K SVD and multiple dimensioned sparse hyperspectral image classification method |
CN108388907B (en) * | 2017-12-29 | 2021-03-02 | 中国科学院自动化研究所 | Real-time updating method of polarized SAR data classifier based on multi-view learning |
CN108333140B (en) * | 2018-02-01 | 2020-10-09 | 福州大学 | Characteristic spectral index for detecting damage of bamboo moths and detection method |
CN110298211B (en) * | 2018-03-21 | 2021-03-23 | 北京大学 | River network extraction method based on deep learning and high-resolution remote sensing image |
CN109146847B (en) * | 2018-07-18 | 2022-04-05 | 浙江大学 | Wafer map batch analysis method based on semi-supervised learning |
CN110874465B (en) * | 2018-08-31 | 2022-01-28 | 浙江大学 | Mobile equipment entity identification method and device based on semi-supervised learning algorithm |
CN109447111B (en) * | 2018-09-20 | 2021-12-03 | 杭州师范大学 | Remote sensing supervision classification method based on subclass training samples |
CN109902390B (en) * | 2018-12-13 | 2023-10-24 | 中国石油大学(华东) | Method for predicting development area of favorable reservoir based on small sample expansion |
CN109726641B (en) * | 2019-01-24 | 2023-02-28 | 常州大学 | Remote sensing image cyclic classification method based on automatic optimization of training samples |
CN110096997B (en) * | 2019-04-29 | 2020-04-14 | 宁波大学 | Coastal zone fine classification method integrating multi-source remote sensing data |
CN110232317B (en) * | 2019-05-05 | 2023-01-03 | 五邑大学 | Hyper-spectral image classification method, device and medium based on super-pixel segmentation and two-stage classification strategy |
CN110796262B (en) * | 2019-09-26 | 2023-09-29 | 北京淇瑀信息科技有限公司 | Test data optimization method and device of machine learning model and electronic equipment |
CN110852175A (en) * | 2019-10-16 | 2020-02-28 | 华东师范大学 | Depth network hyperspectral remote sensing classification method based on mixed limited Boltzmann machine |
CN111222539B (en) * | 2019-11-22 | 2021-07-30 | 国际竹藤中心 | Method for optimizing and expanding supervision classification samples based on multi-source multi-temporal remote sensing image |
CN112906666A (en) * | 2021-04-07 | 2021-06-04 | 中国农业大学 | Remote sensing identification method for agricultural planting structure |
CN115205689B (en) * | 2022-09-14 | 2022-11-18 | 北京数慧时空信息技术有限公司 | Improved unsupervised remote sensing image anomaly detection method |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102096825A (en) * | 2011-03-23 | 2011-06-15 | 西安电子科技大学 | Graph-based semi-supervised high-spectral remote sensing image classification method |
CN103593676A (en) * | 2013-11-29 | 2014-02-19 | 重庆大学 | High-spectral remote-sensing image classification method based on semi-supervision sparse discriminant embedding |
-
2015
- 2015-11-17 CN CN201510788769.9A patent/CN105303198B/en not_active Expired - Fee Related
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102096825A (en) * | 2011-03-23 | 2011-06-15 | 西安电子科技大学 | Graph-based semi-supervised high-spectral remote sensing image classification method |
CN103593676A (en) * | 2013-11-29 | 2014-02-19 | 重庆大学 | High-spectral remote-sensing image classification method based on semi-supervision sparse discriminant embedding |
Non-Patent Citations (2)
Title |
---|
"基于半监督KFCM及邻域信息的遥感图像分类算法";宋文等;《计算机工程与应用》;20140930;第50卷(第9期);全文 * |
"基于多类支持向量机的遥感图像分类及其半监督式改进策略";祁亨年等;《复旦学报(自然科学版)》;20041031;第43卷(第5期);全文 * |
Also Published As
Publication number | Publication date |
---|---|
CN105303198A (en) | 2016-02-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105303198B (en) | A kind of remote sensing image semisupervised classification method learnt from fixed step size | |
CN108830312B (en) | Integrated learning method based on sample adaptive expansion | |
CN105825511A (en) | Image background definition detection method based on deep learning | |
Mishra et al. | A Deep Learning-Based Novel Approach for Weed Growth Estimation. | |
CN107358142A (en) | Polarimetric SAR Image semisupervised classification method based on random forest composition | |
CN111401426B (en) | Small sample hyperspectral image classification method based on pseudo label learning | |
Zhang et al. | A GANs-based deep learning framework for automatic subsurface object recognition from ground penetrating radar data | |
Mishra et al. | Weed density estimation in soya bean crop using deep convolutional neural networks in smart agriculture | |
CN108985360A (en) | Hyperspectral classification method based on expanding morphology and Active Learning | |
Jia et al. | Bringing automated, remote‐sensed, machine learning methods to monitoring crop landscapes at scale | |
Alguliyev et al. | Plant disease detection based on a deep model | |
Hussain et al. | A simple and efficient deep learning-based framework for automatic fruit recognition | |
Hu et al. | Deep learning techniques for in-crop weed identification: A review | |
Yang et al. | Unsupervised images segmentation via incremental dictionary learning based sparse representation | |
CN115359353A (en) | Flower identification and classification method and device | |
Devi et al. | Change detection techniques–A survey | |
Xu et al. | Multi-modal deep learning for weeds detection in wheat field based on RGB-D images | |
Malik et al. | Computer vision models for comparing spatial patterns: understanding spatial scale | |
Tombe et al. | Adaptive deep co-occurrence feature learning based on classifier-fusion for remote sensing scene classification | |
Sehree et al. | Olive trees cases classification based on deep convolutional neural network from unmanned aerial vehicle imagery | |
Liu et al. | A novel deep framework for change detection of multi-source heterogeneous images | |
Kalpana et al. | An image based classification and prediction of diseases on cotton leaves using deep learning techniques | |
Zhou et al. | Wavelet domain multi-view active learning for hyperspectral image analysis | |
Heidary et al. | Urban change detection by fully convolutional siamese concatenate network with attention | |
Gimenez et al. | Rejection methods for vegetation mapping using hyperspectral airborne data |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20180817 Termination date: 20211117 |
|
CF01 | Termination of patent right due to non-payment of annual fee |