CN107358256B - Weakly supervised polarization SAR classification method based on card side-chessboard distance measurement - Google Patents
Weakly supervised polarization SAR classification method based on card side-chessboard distance measurement Download PDFInfo
- Publication number
- CN107358256B CN107358256B CN201710550325.0A CN201710550325A CN107358256B CN 107358256 B CN107358256 B CN 107358256B CN 201710550325 A CN201710550325 A CN 201710550325A CN 107358256 B CN107358256 B CN 107358256B
- Authority
- CN
- China
- Prior art keywords
- sample
- polarization
- training sample
- test sample
- scattering
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (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)
- Radar Systems Or Details Thereof (AREA)
- Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
Abstract
The invention discloses a kind of Weakly supervised polarization SAR classification method based on card side-chessboard distance measurement, mainly solve the problems, such as that the nicety of grading of current Weakly supervised method is low.Its implementation includes: 1) to read in a width polarimetric SAR image, is based on Cloude-Pottier goal decomposition method, constructs the polarization matrix with four dimensional features；2) from polarization matrix in choose 1% marker samples as training sample；3) chi-Square measure and space chessboard distance of decoupled method test sample and training sample；4) combined distance is obtained according to chi-Square measure and chessboard distance, the standard of neighbour is judged as arest neighbors method, realize Classification of Polarimetric SAR Image.The present invention utilizes nearest neighbor classifier only with a small amount of training sample, accurately realizes the terrain classification of polarization SAR.
Description
Technical field
The invention belongs to technical field of image processing, and in particular to the terrain classification method of polarization SAR can be used for atural object point
Class.
Background technique
Synthetic aperture radar SAR is a kind of with high-resolution imaging radar all-time anf all-weather, since SAR has earth's surface
There is certain penetration capacity, the interference of cloud and mist vegetation and leaf can be excluded, in mapping and military affairs, environmental monitoring, disaster prison
Survey, marine monitoring etc. have advantageous advantage, thus also attract the country for the analysis of the data of SAR image and interpretation
The extensive concern of outer scientific research personnel.The appearance of polarization SAR makes each pixel of image contain more polarization information numbers of atural object
According to providing more analysis foundations for terrain classification.
The classification method of polarization SAR mainly has measure of supervision, unsupervised approaches and semi-supervised method at present.There is prison
The method superintended and directed generally depends on a large amount of marker samples, however the acquisition of a large amount of marker samples be it is very expensive and difficult, this
The difficulty of image data interpretation is increased, and the polarization SAR data largely marked are trained and also increase time complexity
Degree and computation complexity, it can be seen that, the method for having supervision is not an effective method.Current unsupervised side
Method also comparative maturity, independent of any marker samples, but unsupervised approaches often separate many additional classifications, this is big
The classification number for needing to classify is had exceeded greatly, and the processing in later period can have too many interference.Semi-supervised method can utilize a small amount of label
The strategy of sample and a large amount of unmarked samples, when appearing to be a kind of very promising method at present, but still having higher
Between complexity and model complexity.
Summary of the invention
The purpose of the present invention is being directed to above-mentioned many deficiencies in the prior art, propose a kind of based on card side-chessboard distance
The Weakly supervised polarization SAR classification method of measurement improves polarization SAR point to reduce the complexity and time complexity of disaggregated model
Class speed and nicety of grading.
The technical scheme is that the method based on existing goal decomposition, it is only necessary to a small amount of marker samples, using card side-chess
Neighbour judging basis of the disk combination metric mode as nearest neighbor classifier is realized to realize Weakly supervised polarization SAR classification
Step includes the following:
(1) a width polarization SAR image to be classified data are read in, the polarization coherence matrix T of 3 × 3 target is obtained, with relevant
Matrix T carries out the goal decomposition of Cloude-Pottier as input, obtains the polarization matrix X with four dimensional features:
Wherein xiFor i-th of sample,N is total sample number, and P is echo strength parameter, H
For scattering entropy, α is angle of scattering, and A is each Xiang Yidu；
(2) sample for choosing 1% from sample set X at random is training sample set S, remaining is test sample collection T:
Wherein siIt is i-th of sample of training sample set, i=1,2 ..., n, tjIt is j-th of sample of test sample collection, j
=1,2 ..., k, n are training sample sum, n=N × 0.01, every class selection ncA training sample: nc=n/NC, NCFor class of classifying
Not total, k is test sample sum, k=N-n；
(3) decoupled method test sample tjWith training sample siChi-Square measureWith space chessboard distance dchess
(L, L '):
dchess(L, L ')=max (| l1-l1′|,|l2-l2' |),
Wherein simIt is siM dimension value, tjmIt is tjM dimension value, m=1,2,3,4；
For training sample siSpace coordinate,For test sample tjSpace coordinate, l1It is siIn sky
Between in abscissa, l2It is siOrdinate in space, l1' it is tjAbscissa in space, l2' it is tjVertical seat in space
Mark；
(4) according to step (3) as a result, calculating test sample tjWith training sample siCombined distance d (si,tj):
Wherein e(·)Expression takes index operation, and ln () expression takes log operations；
(5) nearest neighbor classifier NN is utilized, test sample t is calculatedjCombination with each sample in training sample set S away from
From d (si,tj), obtain distance set { dj1,dj2,…,djξ,…,djn, it will be apart from the smallest djξCorresponding training sample sξ's
Label is set as test sample tjLabel；
(6) above-mentioned (3)-(5) are repeated and Classification of Polarimetric SAR Image is finally realized to remaining test sample prediction label.
The beneficial effects of the present invention are:
1) method based on goal decomposition is used, the polarization matrix with four dimensional features is constructed, is effectively simplified
The data structure of polarimetric SAR image.
2) present invention is only to need the Weakly supervised polarization SAR method of a small amount of marker samples.
3) present invention uses neighbour judging basis of the specific combination metric form as nearest neighbor classifier NN,
To reduce the complexity and time complexity of model, improve polarization SAR nicety of grading and
Classification speed.
The present invention is described in further details below with reference to attached drawing.
Detailed description of the invention
Implementation flow chart Fig. 1 of the invention；
Fig. 2 is true, and class is marked on a map；
Fig. 3 is using the present invention with existing two methods to the terrain classification figure of Fig. 2.
Specific embodiment
In order to make those skilled in the art better understand the present invention, the present invention is carried out specifically with reference to the accompanying drawings
It is bright.
Referring to Fig.1, realization step of the invention includes the following:
Step 1, a width polarization SAR image to be classified data are read in, polarization scattering matrix S is obtainedc, by orthogonal multiple
Pauli matrix base is by ScVector quantization, by Scattering of VectorObtain the polarization coherence matrix T of 3 × 3 target.
(1.1) a width polarimetric SAR image data is read in, polarization scattering matrix S is obtainedc:
Wherein shhExpression is emitted with horizontal direction, and horizontal direction receives, shvExpression is emitted with horizontal direction, vertical direction
It receives, svhExpression is emitted with vertical direction, and horizontal direction receives, svvExpression is emitted with vertical direction, and vertical direction receives, by
S known to reciprocal theoremhv=Svh；
(1.2) Pauli matrix base P is answered by rotating orthogonall, by polarization scattering matrix ScVector quantization obtains Pauli scattering
Vector
WhereinTrace () is to seek square
The mark of battle array, subscript " T " indicate transposition,
(1.3) by Scattering of VectorObtain 3 × 3 coherence matrix T:
Wherein subscript " * " indicates conjugate transposition.
Step 2, using coherence matrix T as input, the goal decomposition of Cloude-Pottier is carried out, obtains having four-dimensional special
The polarization matrix X of sign.
(2.1) following feature decomposition is carried out to coherence matrix T:
Wherein λtIt is t-th of characteristic value of coherence matrix T, t=1,2,3, subscript " * " indicates conjugate transposition, etIt is relevant square
T-th of unit character vector after battle array T orthogonalization；
(2.2) the polarization entropy H of coherence matrix T is calculated:
WhereinFor t kind scattering mechanism occur probability, t=1,2,3；
(2.3) the angle of scattering α of coherence matrix T is calculated:
Wherein αtIt is the type of t kind scattering mechanism, t=1,2,3；
(2.4) each to different degree A of coherence matrix T is calculated:
Wherein p2It is the probability that the 2nd kind of scattering mechanism occurs, p3It is the probability that the 3rd kind of scattering mechanism occurs；
(2.5) the echo strength parameter P of coherence matrix T is calculated:
(2.6) i-th of sample x is constructed according to (2.2)-(2.5)iFour dimensional features:
Wherein P is echo strength parameter, and H is scattering entropy, and α is angle of scattering, and A is each Xiang Yidu；
(2.7) remaining sample repeats (2.6) and obtains the polarization matrix X with four dimensional features:
Wherein N is total sample number.
Step 3, training, test sample are chosen, both samples combined distance d (s is calculatedi,tj)。
(3.1) sample for choosing 1% from sample set X at random is training sample set S, remaining is test sample collection T:
Wherein siIt is i-th of sample of training sample set, i=1,2 ..., n, tjIt is j-th of sample of test sample collection, j
=1,2 ..., k, n are training sample sum, n=N × 0.01, every class selection ncA training sample: nc=n/NC, NCFor class of classifying
Not total, k is test sample sum, k=N-n.
(3.2) test sample t is calculatedjWith training sample siChi-Square measure
(3.3) test sample t is calculatedjWith training sample siSpace chessboard distance dchess(L, L '):
dchess(L, L ')=max (| l1-l1′|,|l2-l2' |),
Wherein simIt is siM dimension value, tjmIt is tjM dimension value, m=1,2,3,4；
For training sample siSpace coordinate,For test sample tjSpace coordinate, l1It is siIn sky
Between in abscissa, l2It is siOrdinate in space, l1' it is tjAbscissa in space, l2' it is tjVertical seat in space
Mark.
(3.4) according to step (3.2)-(3.3) as a result, calculating test sample tjWith training sample siCombined distance d
(si,tj):
Wherein e(·)Expression takes index operation, and ln () expression takes log operations.
Step 4, according to combined distance d (si,tj), realize the terrain classification of polarization SAR.
(4.1) nearest neighbor classifier NN is utilized, test sample t is calculatedjWith the combination of each sample in training sample set S
Distance obtains distance set { dj1,dj2,…,djξ,…,djn},djξIt is training sample sξWith test sample tjCombined distance d
(sξ,tj), ξ=1,2 ..., n, n are training sample number；
(4.2) by { d in distance setj1,dj2,…,djξ,…,djnDistance arrange from small to large, with the distance minimum
It is worth corresponding training sample label as j-th of test sample tjLabel；
(4.3) repeat the above steps (4.1)-(4.2), to remaining test sample prediction label, final realization polarization SAR
Object classification.
Effect of the invention can be further illustrated by following emulation experiment.
1. experiment condition: this experiment is Intel (R) Xeon (R) in CPU, and dominant frequency 2.40GHz inside saves as 16G's
It is emulated in 7 system of WINDOWS using software MATLAB R2013b.
The L-band data in the area Dutch Flevoland that this experimental selection NASA/JPLARISAR is obtained, utilize its one
A subgraph Flevoland1, as shown in Fig. 2, size is 300 × 270.The region shares 6 class atural objects, is potato, sweet tea respectively
Dish, bare area, barley, wheat, pea.Total sample number is 81000, according to selecting training sample, marked sample described in (3.1)
Number is 46659, chooses and wherein 1% is used as training sample, training sample sum is 468, and it is 78 that every class, which chooses sample number,.
2. experiment content
Experiment classifies to Fig. 2 there are two types of control methods using the present invention and now, and wherein control methods 1 is based on European
The arest neighbors classification method of distance, control methods 2 are the svm classifier method based on core measurement, classification results such as Fig. 3, in which:
Fig. 3 (a) is the classification results figure of the method for the present invention,
Fig. 3 (b) is the classification results figure of control methods 1,
Fig. 3 (c) is the classification results figure of control methods 2.
Fig. 3 illustrates the method for the present invention and two kinds of rough classifying qualities of control methods, it can be seen that the present invention is obviously excellent
In control methods.
The overall classification accuracy and Kappa coefficient of the method for the present invention and control methods are calculated, as a result such as table 1.
The classification results of table 1 the method for the present invention and two kinds of control methods
| The method of the present invention | Control methods 1 | Control methods 2 | |
| Overall classification accuracy (%) | 97.23 | 52.23 | 53.14 |
| Kappa coefficient | 0.96 | 0.4059 | 0.4206 |
| Time (s) | 14.62 | 10.7 | 8.42 |
As it can be seen from table 1 45% and 44.09% is respectively increased than the nicety of grading of two kinds of control methods in the present invention, together
When Kappa coefficient be much higher than both control methods, illustrate that the consistency classified of the present invention is best.
The above results show that the present invention can accurately realize that polarization SAR is classified.
Claims (2)
1. the Weakly supervised polarization SAR classification method based on card side-chessboard distance measurement, comprising:
(1) a width polarization SAR image to be classified data are read in, the polarization coherence matrix T of 3 × 3 target are obtained, with coherence matrix
T carries out the goal decomposition of Cloude-Pottier as input, obtains the polarization matrix X with four dimensional features:
Wherein xiFor i-th of sample,N is total sample number, and P is echo strength parameter, and H is scattered
Entropy is penetrated, α is angle of scattering, and A is each Xiang Yidu；
(2) sample for choosing 1% from sample set X at random is training sample set S, remaining is test sample collection T:
Wherein soIt is o-th of sample of training sample set, o=1,2, n, tjIt is j-th of sample of test sample collection, j=1,
2 ..., k, n are training sample sum, n=N × 0.01, every class selection ncA training sample: nc=n/NC, NCIt is total for class categories
Number, k are test sample sum, k=N-n；
(3) decoupled method test sample tjWith training sample soChi-Square measureWith space chessboard distance dchess(L,
L '):
dchess(L, L ')=max (| l1-l′1|,|l2-l′2|),
Wherein somIt is soM dimension value, tjmIt is tjM dimension value, m=1,2,3,4；
For training sample soSpace coordinate,For test sample tjSpace coordinate, l1It is soIn space
Abscissa, l2It is soOrdinate in space, l '1It is tjAbscissa in space, l2' it is tjOrdinate in space；
(4) according to (3) as a result, calculating test sample tjWith training sample soCombined distance d (so,tj):
Wherein e(·)Expression takes index operation, and ln () expression takes log operations；
(5) nearest neighbor classifier NN is utilized, test sample t is calculatedjWith the combined distance d of each sample in training sample set S
(so,tj), obtain distance set { dj1,dj2,…,djξ,…,djn, it will be apart from the smallest djξCorresponding training sample sξLabel
It is set as test sample tjLabel；
(6) above-mentioned (3)-(5) are repeated and Classification of Polarimetric SAR Image is finally realized to remaining test sample prediction label.
2. according to the method described in claim 1, wherein carrying out Cloude- using coherence matrix T as input in step (1)
The goal decomposition of Pottier carries out as follows:
Following feature decomposition 1a) is carried out to coherence matrix T:
Wherein λtIt is t-th of characteristic value of coherence matrix T, t=1,2,3, subscript " * " indicates conjugate transposition, etIt is coherence matrix T
T-th of unit character vector after orthogonalization, v=1,2,3；
1b) calculate polarization entropy H:
WhereinFor w kind scattering mechanism occur probability, w=1,2,3；
1c) calculate angle of scattering α:
Wherein αwIt is the type of w kind scattering mechanism, w=1,2,3；
It 1d) calculates each to different degree A:
Wherein p2It is the probability that the 2nd kind of scattering mechanism occurs, p3It is the probability that the 3rd kind of scattering mechanism occurs；
1e) calculate echo strength parameter P:
1f) according to 1b) -1e) i-th of sample x of buildingiFour dimensional features:
Wherein P is echo strength parameter, and H is scattering entropy, and α is angle of scattering, and A is each Xiang Yidu；
1g) remaining sample repeats 1f) obtain the polarization matrix X with four dimensional features:
Wherein N is total sample number.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710550325.0A CN107358256B (en) | 2017-07-07 | 2017-07-07 | Weakly supervised polarization SAR classification method based on card side-chessboard distance measurement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710550325.0A CN107358256B (en) | 2017-07-07 | 2017-07-07 | Weakly supervised polarization SAR classification method based on card side-chessboard distance measurement |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107358256A CN107358256A (en) | 2017-11-17 |
| CN107358256B true CN107358256B (en) | 2019-07-23 |
Family
ID=60292777
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710550325.0A Active CN107358256B (en) | 2017-07-07 | 2017-07-07 | Weakly supervised polarization SAR classification method based on card side-chessboard distance measurement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107358256B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109615025A (en) * | 2018-12-28 | 2019-04-12 | 内蒙古工业大学 | Quick atural object classification method |
Family Cites Families (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8125370B1 (en) * | 2007-04-16 | 2012-02-28 | The United States Of America As Represented By The Secretary Of The Navy | Polarimetric synthetic aperture radar signature detector |
| CN101894275B (en) * | 2010-06-29 | 2013-07-24 | 武汉大学 | Weakly supervised method for classifying SAR images |
| CN103246883B (en) * | 2013-05-20 | 2016-02-17 | 中国矿业大学(北京) | A kind of underground coal mine thermal infrared images face identification method |
-
2017
- 2017-07-07 CN CN201710550325.0A patent/CN107358256B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN107358256A (en) | 2017-11-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104077599B (en) | Polarization SAR image classification method based on deep neural network | |
| CN107944370B (en) | Classification of Polarimetric SAR Image method based on DCCGAN model | |
| CN103824084B (en) | Classification of Polarimetric SAR Image method based on SDIT and SVM | |
| CN102982338B (en) | Classification of Polarimetric SAR Image method based on spectral clustering | |
| CN107330457B (en) | A kind of Classification of Polarimetric SAR Image method based on multi-feature fusion | |
| CN102968640B (en) | Decompose and the Classification of Polarimetric SAR Image method of data distribution characteristics based on Freeman | |
| CN102999762B (en) | Decompose and the Classification of Polarimetric SAR Image method of spectral clustering based on Freeman | |
| CN105718942B (en) | High spectrum image imbalance classification method based on average drifting and over-sampling | |
| CN103839073B (en) | Polarization SAR image classification method based on polarization features and affinity propagation clustering | |
| CN104751185B (en) | SAR image change detection based on average drifting genetic cluster | |
| CN105868793A (en) | Polarization SAR image classification method based on multi-scale depth filter | |
| JP7438655B2 (en) | Weather analysis equipment, weather analysis methods, and programs | |
| CN103996047A (en) | Hyperspectral image classification method based on compression spectrum clustering integration | |
| CN105160353B (en) | Polarization SAR data terrain classification method based on multiple features collection | |
| CN106611422B (en) | Stochastic gradient Bayes's SAR image segmentation method based on sketch structure | |
| CN106651884B (en) | Mean field variation Bayes's SAR image segmentation method based on sketch structure | |
| CN109840542B (en) | Self-adaptive dimension decision tree classification method based on polarization characteristics | |
| CN104700116B (en) | The sorting technique of the Polarimetric SAR Image atural object represented based on multi-layer quantum ridge ripple | |
| CN106683102A (en) | SAR image segmentation method based on ridgelet filters and convolution structure model | |
| CN105138966B (en) | Classification of Polarimetric SAR Image method based on fast density peak value cluster | |
| CN104408472B (en) | Classification of Polarimetric SAR Image method based on Wishart and SVM | |
| CN103714353B (en) | The Classification of Polarimetric SAR Image method of view-based access control model prior model | |
| CN103870842B (en) | Polarized SAR image classification method combining polarization feature and watershed | |
| Gou et al. | Eigenvalue analysis-based approach for POL-SAR image classification | |
| CN104463227B (en) | Classification of Polarimetric SAR Image method based on FQPSO and goal decomposition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |