CN104156728B - Polarized SAR image classification method based on stacked code and softmax - Google Patents

Polarized SAR image classification method based on stacked code and softmax Download PDF

Info

Publication number
CN104156728B
CN104156728B CN201410334397.8A CN201410334397A CN104156728B CN 104156728 B CN104156728 B CN 104156728B CN 201410334397 A CN201410334397 A CN 201410334397A CN 104156728 B CN104156728 B CN 104156728B
Authority
CN
China
Prior art keywords
data
classification
sar image
matrix
step
Prior art date
Application number
CN201410334397.8A
Other languages
Chinese (zh)
Other versions
CN104156728A (en
Inventor
王爽
马文萍
谢慧明
霍丽娜
马晶晶
雷晓珍
Original Assignee
西安电子科技大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 西安电子科技大学 filed Critical 西安电子科技大学
Priority to CN201410334397.8A priority Critical patent/CN104156728B/en
Publication of CN104156728A publication Critical patent/CN104156728A/en
Application granted granted Critical
Publication of CN104156728B publication Critical patent/CN104156728B/en

Links

Abstract

The invention belongs to the technical field of image processing, particularly discloses a polarized SAR image classification method based on a stacked code and softmax, and mainly aims to solve the problem that conventional characteristic extraction requires much priori knowledge, and is high in manual labor intensity. The method comprises the following steps: (1) inputting an image and filtering; (2) extracting independent elements; (3) selecting a trained sample and a tested sample and conducting whitening process; (4) constructing a stacked code network to learn superior characteristics better representing the SAR image; (5) training a classifier, and forecasting classification result; (6) calculating accuracy; (7) outputting the result. Compared with the classic classification methods, the method is provided with higher accuracy of ground features classification, is more integrative in uniform district, better in district consistency and classification performance, and can be suitable for ground feature classification and target identification of polarized SAR images.

Description

A kind of Classification of Polarimetric SAR Image method based on stack coding and softmax

Technical field

The invention belongs to technical field of image processing, and in particular to polarization synthetic aperture radar image terrain classification technology is led A kind of polarization synthetic aperture thunder for feature learning being carried out based on stack coding and being classified with softmax graders in domain Up to (Synthetic Aperture Radar, SAR) image classification method.Can be used for the terrain classification of Polarimetric SAR Image and Target identification, can effectively improve the precision of Classification of Polarimetric SAR Image.

Background technology

Synthetic aperture radar (SAR) system can obtain the remote sensing images with round-the-clock, round-the-clock, high resolution, pole It is advanced SAR system to be combined to aperture radar (polarization SAR), describes what is observed by launching and receiving polarization radar wave Land cover pattern thing and target.

In the past twenty years, research shows that polarization SAR is carried in target detection, terrain classification, parametric inversion, landform Taking application aspect can provide useful information more more than single polarization SAR.Nowadays, some Space-bornes, such as TerraSAR-X is defended Star, RADARSAT-2 satellites, and ALOS-PALSAR satellites constantly provide data volume huge polarization SAR data.Solve manually It is not reliable to release these a large amount of extremely complex images.Therefore, automatically or semi-automatically system is come to pole urgently to need exploitation Change SAR image to explain and information excavating.

According to whether training sample and manual intervention, Polarimetric SAR Image is divided into supervised classification and unsupervised classification.It is right In both approaches, feature extraction and sorting technique are two fundamentals.The performance of Classification of Polarimetric SAR Image is largely On depend on feature.For this reason, many Classification of Polarimetric SAR Image methods put forth effort on the extraction of Polarimetric SAR Image feature. The scientific paper that many has been delivered proposes various methods based on the properties of polarization SAR data to obtain polarization SAR feature.

Cloude et al. proposes the Unsupervised classification of polarimetric synthetic aperture radar images based on H/ α goal decompositions, sees Cloude S R, Pottier E.An entropy based classification scheme for land applications of Polarimetric SAR [J] .IEEE Trans.Geosci.Remote Sensing.1997,35 (1):549-557. is exactly A kind of method of feature extraction, the method is mainly decomposes two features of sign polarization data of acquisition H and α by Cloude, Then the H/ α planes that are constituted according to H and α it is artificial be divided into 9 regions, remove an area that can not possibly exist in theory Domain, finally divides an image into 8 classes.The defect that H/ alpha taxonomies are present is that the division in region is excessively dogmatic, when data distribution area May be by the division of mistake, in addition, same category of atural object may be divided less than in different regions, together when on the border of domain When, different classes of atural object is also likely to be present in the same area.

Yoshio Yamaguchi etc. propose a kind of polarization diagram for being based on four component goal decompositions as unsupervised classification is calculated Method, is shown in Yoshio Yamaguchi, Toshifumi Moriyama, Motoi Ishido, and Hiroyoshi Yamada, “Four-Component Scattering Model for Polarimetric SAR Image Decomposition,” IEEE Trans.Geosci.Remote Sens, vol.43, no.8, Aug.2005. are the methods of another feature extraction.Should Polarization SAR data are decomposed into four simple scattering mechanisms and are combined by method.This four scattering mechanisms are respectively:Plane dissipates Penetrate, dihedral angle is scattered, volume scattering and spiral are scattered.

The method of the extraction of these features is designed according to the problem to be solved and the feature engineer of data , at the same time, the method hand labor intensity that these extract feature is especially big.

Stack coding is a kind of unsupervised feature learning framework, and it can extract multilayer feature.Similar to the level of human brain Model, stack coding can be extracted from low level to high-level feature.In general, the feature of higher level can preferably reflect The property of data, is more beneficial for classification.Stack coding does not need the priori of any data can just to extract feature.Therefore, It is more beneficial for being applied to new problem and new data.

The content of the invention

Deficiency it is an object of the invention to be directed to above-mentioned prior art, proposes a kind of based on stack coding and softmax Classification of Polarimetric SAR Image method, to improve classifying quality and precision.

Realizing technical thought of the invention is:First, polarimetric synthetic aperture radar SAR image is filtered;Secondly, it is right Each pixel extraction goes out 9 independent elements, constitutes a row vector, has just obtained each pixel and has been adapted to stack and encodes be used for The original input data of learning characteristic;Then, encoded to these data learning characteristics with stack;Again with the feature that obtains of study to Amount is classified in being input to softmax graders, and stack is encoded and softmax grader groups with back-propagation algorithm afterwards It is finely adjusted into framework, obtains more preferable feature and classification results.

To achieve the above object, a kind of Classification of Polarimetric SAR Image method based on stack coding and softmax of the present invention, Comprise the following steps:

(1) Polarimetric SAR Image to be sorted is read in, using exquisiteness polarization LEE filtering methods to polarization to be sorted SAR image is filtered, and removes speckle noise, obtains the Polarimetric SAR Image after denoising;

(2) 9 independent elements, structure are extracted from the coherence matrix T of Polarimetric SAR Image each pixel after denoising Into a vector, each pixel is just obtained and has been adapted to stack coding network for the original input data of learning characteristic;

(3) according to real ground substance markers, 10% is randomly selected to each atural object classification of Polarimetric SAR Image respectively has Used as training sample, remaining 90% has mark original input data as test sample to mark original input data, and to choosing The training sample for taking does whitening processing, obtains more preferably carrying out the input data of feature learning;

(4) construction stack coding network learning characteristic vector, the characteristic vector that the study of the stack coding network second layer is obtained It is exactly the advanced features vector of Polarimetric SAR Image;The parameter setting for constructing stack coding network is performed as follows:

(4a) trains stack coding network ground floor:The hidden layer nodes for setting stack network code ground floor are 25, Openness penalty factor β is 3, and openness parameter ρ is 0.1, and attenuation parameter λ is 3 × 10-3

(4b) trains the stack coding network second layer:The hidden layer nodes for setting the stack coding network second layer are 50, Openness penalty factor β is 3, and openness parameter ρ is 0.1, and attenuation parameter λ is 3 × 10-3

(5) grader is trained:The characteristic vector and its real ground substance markers that study is obtained are put into softmax graders In be trained, while with back-propagation algorithm to stack encode and softmax graders constitute framework be finely adjusted;

(6) classification results are predicted and accuracy is calculated:Test sample is done into corresponding whitening processing, then by after albefaction Test data is input to the classification results predicted in the framework for training, will predict the classification results that obtain with really Substance markers are contrasted and are calculated accuracy;

(7) output result:

On Polarimetric SAR Image after sorting, identical category is assigned to identical RGB color in classification results, is painted Classification results figure afterwards, classification results figure after output colouring.

The sliding window size of the exquisite polarization LEE filtering methods described in above-mentioned steps (1) is 7 × 7 pixels.

Described in above-mentioned steps (2) extracted from the coherence matrix T of each pixel of Polarimetric SAR Image 9 it is independent Element, is carried out as follows:

Each pixel of filtered image is read in, these pixels are the coherence matrix T of 3 × 3:

Wherein H represents horizontal polarization, and V represents vertical polarization, SHHExpression level is to transmitting and level to the number of echoes for receiving According to SVVRepresent vertical to transmitting and vertical to the echo data for receiving, SHVExpression level is vertical to the number of echoes for receiving to transmitting According to, () * represents the conjugation of this data,<>Represent average by number is regarded, it can be seen that T matrixes are a conjugation complex matrix, and it is only There are 9 independent elements, respectively T11、T22、T33、T12Real part and imaginary part, T13Real part and imaginary part and T23Real part and Imaginary part;By this 9 element combinations into a row vector, the original input data of each pixel has just been obtained.

The training sample to selection described in above-mentioned steps (3) does whitening processing to be carried out as follows:

The first step, calculates the average M and covariance matrix C of training data S;

Second step, feature decomposition is carried out to covariance matrix C, obtains feature value vector matrix V and eigenvalue matrix D, its The diagonal of middle D is characterized value, and remaining element is all 0;

3rd step, whitening matrix P is calculated with following formula:

Wherein, diag () represents the value for taking diagonal of a matrix, and () ' expression takes conjugate transposition to matrix;

4th step, the data X after albefaction is calculated with following formula, and preserves M and P, is made for test data albefaction below With:

X=(S-M) P

Wherein X is the data that training data obtained after whitening processing.

The parameter setting of the training grader described in above-mentioned steps (5) is performed as follows:

Because truly substance markers are 15 classes to the data for using, the nodes for setting output layer are 15, and grader Input data dimension for the stack coding network second layer hidden layer nodes 50, so input layer number be 50, while It is 10 to set attenuation parameter λ-4

Color method refers in three primary colours described in above-mentioned steps (7), on Polarimetric SAR Image after sorting, will be red, green Color, blue three colors mix three primary colours according to different ratios respectively as three primary colours, take following 15 to this three primary colours respectively Class value:(255,0,0)、(255,128,0)、(171,138,80)、(255,255,0)、(183,0,255)、(191,191, 255)、(90,11,255)、(191,255,191)、(0,252,255)、(128,0,0)、(255,182,229)、(0,255,0)、 (0,131,74)、(0,0,255)、(255,217,157);15 kinds of different colors are obtained to paint classification results, wherein dividing Identical color in identical category in class result.

Beneficial effects of the present invention:Compared with prior art, the present invention has advantages below:

1. any prior information of data is required no knowledge about, any new problem and new number can be conveniently used in According to;

2. the scattering properties and statistical property of exploration polarization SAR need not be gone, hand labor can be largely reduced strong Degree;

3. the invention is characterized in that by learning what is obtained, the advanced features for obtaining can preferably reflect the property of data, Classification is more beneficial for, more preferable classification results and nicety of grading can be obtained, classification results region consistency of the present invention is also preferable.

The present invention is described in further details below with reference to accompanying drawing.

Brief description of the drawings

Fig. 1 is general flow chart of the invention;

Fig. 2 is the PauliRGB composite diagrams that the Fu Laifulan that present invention emulation is used saves Flevoland polarization SAR data;

Fig. 3 is polarimetric synthetic aperture radar SAR image real ground substance markers used by the present invention;

Fig. 4 is the classification results with existing supervision wishart sorting techniques;

Fig. 5 is the classification results with existing support vector machines sorting technique;

Fig. 6 is the classification results figure of polarimetric synthetic aperture radar SAR image used by the present invention.

Specific embodiment

Reference picture 1, of the invention to implement step as follows:

Step 1, reads in a Polarimetric SAR Image to be sorted, using exquisiteness polarization LEE filtering methods to pole to be sorted Change SAR image to be filtered, remove speckle noise, obtain filtered Polarimetric SAR Image;

One Polarimetric SAR Image to be sorted of (1a) input option.

(1b) is filtered using exquisite polarization LEE filtering methods to Polarimetric SAR Image to be sorted:

The first step, the sliding window size for setting exquisite polarization LEE filtering is 7 × 7 pixels;

Second step, by sliding window in the pixel of the Polarimetric SAR Image of input, slides, often from left to right, from top to bottom When cunning moves a step, by sliding window according to pixel space position, it is divided into 9 subwindows, sub- window successively from left to right, from top to bottom There is overlap between mouthful, the size of each window is 3 × 3 pixels.Calculate the power (i.e. the sum of T diagonal of a matrixs) of each sub- window Value, resulting average is constituted the average window of 3 × 3 pixels;

3rd step, selection level, the gradient template of the four direction of vertical, 45 degree and 135 degree, by average window respectively with Four templates are weighted, and the gradient template that selection calculates weighted results maximum absolute value is edge direction;

4th step, the pixel to comparing both sides of edges seeks power average, and two for obtaining average is individually subtracted central window The average of mouthful all pixels, using the less side of average difference as direction window;

5th step, according to the following formula, obtains the weights of exquisite polarization LEE filtering;

Wherein, k represents the weights of exquisite polarization LEE filtering, y0Represent the performance number of single pixel point, var (y0) expression side To the variance yields of window internal power,Represent the intraoral power average of Directional Windows, δvRepresent the Polarimetric SAR Image coherent spot of input The standard deviation of noise;

6th step, according to the following formula, obtains filtering the covariance matrix of after-polarization SAR image center pixel:

Wherein,The covariance matrix of filtering after-polarization SAR image center pixel is represented,Represent the intraoral polarization of Directional Windows The average of the covariance matrix of SAR image pixel, k represents the weights of exquisite polarization LEE filtering, and Z represents Polarimetric SAR Image center The covariance matrix of pixel.

Step 2,9 independent units are extracted from the coherence matrix T of Polarimetric SAR Image each pixel after denoising Element, T matrixes are a conjugation complex matrix, and it only has 9 independent elements, used as the data that next step is processed.

Each pixel of filtered image is read in, these pixels are the coherence matrix T of 3 × 3:

Wherein H represents horizontal polarization, and V represents vertical polarization, SHHExpression level is to transmitting and level to the number of echoes for receiving According to SVVRepresent vertical to transmitting and vertical to the echo data for receiving, SHVExpression level is vertical to the number of echoes for receiving to transmitting According to, () * represents the conjugation of this data,<>Represent average by number is regarded.It can be seen that T matrixes are a conjugation complex matrix, it is only There are 9 independent elements, respectively T11、T22、T33、T12Real part and imaginary part, T13Real part and imaginary part and T23Real part and Imaginary part.By this 9 element combinations into a row vector, the original input data of each pixel has just been obtained.

Step 3, chooses training sample and carries out whitening processing.

(3a) randomly selects 10% respectively according to real ground substance markers, each the atural object classification to polarization SAR data There is mark original input data as training sample, remaining 90% has mark original input data as training sample.

(3b) does whitening processing to the training sample chosen.

The first step, calculates the average M and covariance matrix C of training data S;

Second step, feature decomposition is carried out to covariance matrix C, obtains feature value vector matrix V and eigenvalue matrix D, its The diagonal of middle D is characterized value, and remaining element is all 0;

3rd step, whitening matrix P is calculated with following formula:

Wherein, diag () represents the value for taking diagonal of a matrix, and () ' expression takes conjugate transposition to matrix.

4th step, the data X after albefaction is calculated with following formula, and preserves M and P, is made for test data albefaction below With:

X=(S-M) P

Step 4, construction stack coding network learning characteristic vector.

The stack coding network for using is a neutral net being made up of the sparse self-encoding encoder of two-layer, and its preceding layer is sparse The characteristic vector for learning of self-encoding encoder as the sparse self-encoding encoder of its later layer input.Sparse self-encoding encoder is by being input into Layer, hidden layer, the part of output layer 3 are constituted.Own coding neutral net makes input layer try one's best equal to output by back-propagation algorithm Layer, when input layer is less than the threshold value of setting or meets certain iterations with the loss function of output interlayer, hidden layer The characteristic vector output that just obtains as own coding neural network learning of activation value vector.

Be input to training sample after albefaction in the sparse own coding of ground floor and be trained by (4a), obtains that input can be characterized The characteristic vector of data.

The first step, arrange parameter:It is 25 to set hidden layer nodes, and openness penalty factor β is 3, and openness parameter ρ is 0.1, attenuation parameter λ are 3 × 10-3

Second step, is here 9 dimensions according to the dimension of input data, so input layer number is 9, because to allow output Data are tried one's best equal to input data, so output node is also 9.An encoder matrix W is generated at random11∈R25×9With a decoding Matrix W12∈R9×25And coding biasing b11∈R25×1With decoding biasing b12∈R9×1;Wherein encoder matrix is input layer to hiding The weight matrix of layer, coding is biased to input layer to the biasing of hidden layer;Decoding matrix is weights square of the hidden layer to output layer Battle array, decoding is biased to hidden layer to the biasing of output layer.

3rd step, the activation value a of hidden layer is calculated according to following formula1:

a1=f (W11·X+b11)

Wherein X is input data, and f is a nonlinear function:Sigmoid functions, its form is as follows:

F (x)=1/ (1+exp (- x))

4th step, according to the activation value a for obtaining1And following formula calculates average activity

Wherein m is the number of the training sample of input,Represent that in given input be x(i)In the case of, it is sparse self-editing Code neutral net hides the activation value of node layer j;

5th step, according to the activation value a for obtaining1And following formula calculates output valve Y1

Y1=f (W12·a1+b12)

6th step, according to following formula counting loss value:

Wherein x(i)I-th input value of X is represented,Represent Y1I-th output valve.

7th step, W is updated according to following formula11、W12、b11And b12

WhereinIt is a1The value of the row of jth row i-th.

The 3rd step to the 7th step is repeated, until the difference of front and rear cost twice is less than 10-6Or iterations is more than 400 It is secondary, then stop iteration and preserve a1、W11And b11Value.

Be input to the characteristic vector that the sparse own coding training of ground floor is obtained in the sparse own coding of the second layer and carry out by (4b) Training, obtains characterizing the second layer characteristic vector of the characteristic vector obtained by the sparse own coding of ground floor, and this characteristic vector is just It is the stack coding network characteristic vector that obtains of study;

The first step, arrange parameter:It is 50 to set hidden layer nodes, and openness penalty factor β is 3, and openness parameter ρ is 0.1, attenuation parameter λ are 3 × 10-3

Second step, is here 25 dimensions according to the dimension of input data, so input layer number is 25, because to allow defeated Go out data to try one's best equal to input data, so output node is also 25.An encoder matrix W is generated at random21∈R50×25With one Decoding matrix W21∈R25×50And coding biasing b21∈R50×1With decoding biasing b22∈R25×1

3rd step, the activation value a of hidden layer is calculated according to following formula2:

a2=f (W21·X+b21)

Wherein a1For the characteristic vector that the sparse own coding study of ground floor is obtained, f is a nonlinear function:sigmoid Function, its form is as follows:

F (x)=1/ (1+exp (- x))

4th step, according to the activation value a for obtaining2And following formula calculates average activity

Wherein m is the number of the training sample of input,Represent and be in given inputIn the case of, it is sparse self-editing Code neutral net hides the activation value of node layer j, whereinIt is a1The data of the i-th row.

5th step, according to the activation value a for obtaining2And following formula calculates output valve Y2

Y1=f (W22·a21+b22)

6th step, according to following formula counting loss value:

WhereinIt is a1The data of the i-th row,Represent Y2I-th output valve.

7th step, W is updated according to following formula21、W22、b21And b22

WhereinIt is a2The value of the row of jth row i-th.

The 3rd step to the 7th step is repeated, until the difference of front and rear cost twice is less than 10-6Or iterations is more than 400 It is secondary, then stop iteration and preserve a2、W21And b21Value.

Step 5, trains grader.

Be put into the characteristic vector that the study of stack coding network is obtained in softmax graders and be trained by (5a), obtains The grader for training.Softmax graders are a multi-categorizers, are made up of input layer and output layer.

The first step, arrange parameter:Because truly substance markers are 15 classes to the data for using, the node of output layer is set Number is 15, and the input data of grader is a2, so input layer number is 50, while it is 10 to set attenuation parameter λ-4

Second step, generates a weight matrix W at random31∈R50×15

3rd step, input is according to the category Y that truly substance markers are obtained3

4th step, according to following formula counting loss value:

1 { } therein is an indicative function, i.e., when the value in braces is true, the result of the function is just 1, Otherwise its result is just the number that 0, m is the training sample being input into,It is a1The data of the i-th row,It is W31The number of jth row According to,It is i-th mark of data.

5th step, W is updated according to following formula31

Fourth, fifth step is repeated, until the difference of front and rear cost twice is less than 10-6Or iterations is more than 400 times, then stop Only iteration and preserve W31Value.

(5b) is finely adjusted with back-propagation algorithm to the framework that stack coding network and softmax graders are constituted, and is obtained To the taxonomy model of the feature learning for training.

The first step, arrange parameter:It is 3 × 10 to set attenuation parameter λ-3

Second step, a is calculated according to following formula1、a2Value:

a1=f (W11·X+b11)

a2=f (W21·X+b21)

Wherein X is to obtain being data after training data carries out whitening processing

3rd step, according to following formula counting loss value:

Wherein W11、W21And W31All it is the value of preservation in step (6).

4th step, W is updated according to following formula11、b11、W21、b21、W31And a1

WhereinThat represent is d2The value of the i-th row.

d1=-(W '21·d2)·X(1-X)

Wherein b11、b21It is also the value of preservation in step (6),That represent is d2The value of the i-th row

Second step to the 4th step is repeated, until the difference of front and rear cost twice is less than 10-6Or iterations is more than 400 It is secondary, then stop iteration and preserve W11、b11、W21、b21And W31Value.

Step 6, predicts classification results.

Test sample is done corresponding whitening processing by (6a).

Input test data S1, the data X after albefaction is obtained according to following formula1

X1=(S1-M)·P

Wherein M and P are the M and P of preservation in step (4).

(6b) test data after albefaction is input to the classification results predicted in the framework for training.

Output Y can be tried to achieve with following formula:

Y=W31·f(W21·f(W11·X1+b11)+b21)

Wherein f is sigmoid functions, X1Test data carries out the data after whitening processing.Each column correspondence is every number in Y According to predict the outcome, the position of the wherein maximum corresponding to each column is exactly the prediction category of the data.

(6b) calculates accuracy.

To predict that the classification results for obtaining are contrasted with real ground substance markers.For each classification, classification is correct In number of pixels and test sample the ratio of the total number of pixels of respective classes as this classification accuracy;Total classification is just True pixel pixel and the ratio of the total number of pixels of test sample are used as the overall classification accuracy rate of Polarimetric SAR Image.

Step 7, output result.

On Polarimetric SAR Image after sorting, using red, green, blue three colors as three primary colours, will be red, green Color, blue three colors mix three primary colours according to different ratios respectively as three primary colours, take following 15 to this three primary colours respectively Class value:(255,0,0)、(255,128,0)、(171,138,80)、(255,255,0)、(183,0,255)、(191,191, 255)、(90,11,255)、(191,255,191)、(0,252,255)、(128,0,0)、(255,182,229)、(0,255,0)、 (0,131,74)、(0,0,255)、(255,217,157).15 kinds of different colors are obtained to paint classification results, wherein dividing Identical color in identical category in class result.Identical color in identical category in classification results, the classification after being painted Result figure, classification results figure after output colouring.

The present invention has advantages below compared with prior art:

1st, any prior information of data is required no knowledge about, any new problem and new number can be conveniently used in According to;

2nd, the scattering properties and statistical property of exploration polarization SAR need not be gone, hand labor can be largely reduced strong Degree;

3rd, the invention is characterized in that by learning what is obtained, the advanced features for obtaining can preferably reflect the property of data, Classification is more beneficial for, more preferable classification results and nicety of grading can be obtained, classification results region consistency of the present invention is also preferable.

Emulation content of the invention and result

The Fu Laifulan shown in Fig. 2 is saved Flevoland polarization SARs data as test image for the present invention, and size is 750 ×1024.Truly substance markers according to Fig. 3 randomly select 10% as training sample to each classification.There is mark with remaining 90% data as test data.

Emulation one, is classified with existing supervision wishart sorting techniques to Fig. 2, and classification results are shown in Fig. 4.It is wherein every The accuracy of class and total accuracy see the table below.

Emulation two, is classified with existing support vector machines sorting technique to Fig. 2, and classification results are shown in Fig. 5.Wherein As input data after training sample whitening processing.Test data is also required to be tested again after doing corresponding albefaction.Per class just True rate and total accuracy see the table below.

Emulation three, is classified with the present invention to Fig. 2, and classification results are shown in Fig. 6.Accuracy and total accuracy per class are shown in Following table:

Be can be seen that by upper table and Fig. 4, Fig. 5 and Fig. 6:The present invention obtains highest correct among three kinds of methods Rate (93.58%).The present invention either in precision, or in visual effect all than supervision wishart methods and support to Amount machine SVM methods are more preferable.Supervision wishart by the region of many water for mistake has been divided into exposed soil, and the present invention and SVMs SVM methods are not the case.It is also worth noting that support vector machines method is in building, and this atural object is obtained Very low accuracy (10.44%).The method classifies for forest most of building mistake.The all atural objects of the present invention Classification accuracy rate is both greater than 84%, and the homogeneous region of the method is more complete than other two methods, and region consistency is more preferable.Explanation The present invention is in polarization SAR data classification performance more preferable.

It is exemplified as above be only to of the invention for example, do not constitute the limitation to protection scope of the present invention, it is all It is that design same or analogous with the present invention is belonged within protection scope of the present invention.

Claims (5)

1. a kind of Classification of Polarimetric SAR Image method based on stack coding and softmax, comprises the following steps:
(1) Polarimetric SAR Image to be sorted is read in, using exquisiteness polarization LEE filtering methods to polarization SAR figure to be sorted As being filtered, speckle noise is removed, obtain the Polarimetric SAR Image after denoising;
(2) 9 independent elements are extracted from the coherence matrix T of Polarimetric SAR Image each pixel after denoising, one is constituted Individual vector, has just obtained each pixel and has been adapted to stack coding network for the original input data of learning characteristic;
(3) according to real ground substance markers, 10% is randomly selected to each atural object classification of Polarimetric SAR Image respectively mark Used as training sample, remaining 90% has mark original input data as test sample to original input data, and to selection Training sample does whitening processing, obtains more preferably carrying out the input data of feature learning;
(4) construction stack coding network learning characteristic is vectorial, and the characteristic vector that the study of the stack coding network second layer is obtained is exactly The advanced features vector of Polarimetric SAR Image;The parameter setting for constructing stack coding network is performed as follows:
(4a) trains stack coding network ground floor:The hidden layer nodes for setting stack network code ground floor are 25, sparse Property penalty factor β be 3, openness parameter ρ is 0.1, and attenuation parameter λ is 3 × 10-3
(4b) trains the stack coding network second layer:The hidden layer nodes for setting the stack coding network second layer are 50, sparse Property penalty factor β be 3, openness parameter ρ is 0.1, and attenuation parameter λ is 3 × 10-3
(5) grader is trained:The characteristic vector and its real ground substance markers that study is obtained are put into softmax graders Row training, while being finely adjusted to the framework that stack coding and softmax graders are constituted with back-propagation algorithm;
(6) classification results are predicted and accuracy is calculated:Test sample is done into corresponding whitening processing, then by the test after albefaction Data input will predict the classification results and real atural object mark for obtaining to the classification results predicted in the framework for training Remember row contrast into and calculate accuracy;
(7) output result:
On Polarimetric SAR Image after sorting, identical category is assigned to identical RGB color in classification results, after being painted Classification results figure, classification results figure after output colouring.
2. a kind of Classification of Polarimetric SAR Image method based on stack coding and softmax according to claim 1, it is special Levy and be:The sliding window size of the exquisite polarization LEE filtering methods described in step (1) is 7 × 7 pixels.
3. a kind of Classification of Polarimetric SAR Image method based on stack coding and softmax according to claim 1, it is special Levy and be:9 independent elements are extracted from the coherence matrix T of each pixel of Polarimetric SAR Image described in step (2), Carry out as follows:
Each pixel of filtered image is read in, these pixels are the coherence matrix T of 3 × 3:
T = < | S H H + S V V | 2 > < ( S H H + S V V ) ( S H H - S V V ) * > 2 < ( S H H + S V V ) S H V * > < ( S H H - S V V ) ( S H H + S V V ) * > < | S H H - S V V | 2 > 2 < ( S H H - S V V ) S H V * > 2 < S H V ( S H H + S V V ) * > 2 < S H V ( S H H - S V V ) * > 4 < | S H V | 2 >
Wherein H represents horizontal polarization, and V represents vertical polarization, SHHExpression level is to transmitting and level to the echo data for receiving, SVV Represent vertical to transmitting and vertical to the echo data for receiving, SHVExpression level is vertical to the echo data for receiving to transmitting, () * represents the conjugation of this data,<>Represent average by number is regarded, it can be seen that T matrixes are a conjugation complex matrix, and it only has 9 Individual independent element, respectively T11、T22、T33、T12Real part and imaginary part, T13Real part and imaginary part and T23Real part and void Portion;By this 9 element combinations into a row vector, the original input data of each pixel has just been obtained.
4. a kind of Classification of Polarimetric SAR Image method based on stack coding and softmax according to claim 1, it is special Levy and be:The training sample to selection described in step (3) does whitening processing to be carried out as follows:
The first step, calculates the average M and covariance matrix C of training data S;
Second step, feature decomposition is carried out to covariance matrix C, obtains feature value vector matrix V and eigenvalue matrix D, wherein D's Diagonal is characterized value, and remaining element is all 0;
3rd step, whitening matrix P is calculated with following formula:
P = V &CenterDot; d i a g ( 1 / d i a g ( D ) ) &CenterDot; V &prime;
Wherein, diag () represents the value for taking diagonal of a matrix, and () ' expression takes conjugate transposition to matrix;
4th step, the data X after albefaction is calculated with following formula, and preserves M and P, is used for test data albefaction below:
X=(S-M) P
Wherein X is the data that training data obtained after whitening processing.
5. a kind of Classification of Polarimetric SAR Image method based on stack coding and softmax according to claim 1, it is special Levy and be:The parameter setting of the training grader described in step (5) is performed as follows:
Because truly substance markers are 15 classes to the data for using, the nodes for setting output layer are 15, and grader is defeated Enter the hidden layer nodes 50 that data dimension is the stack coding network second layer, so input layer number is 50, while setting Attenuation parameter λ is 10-4
CN201410334397.8A 2014-07-14 2014-07-14 Polarized SAR image classification method based on stacked code and softmax CN104156728B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201410334397.8A CN104156728B (en) 2014-07-14 2014-07-14 Polarized SAR image classification method based on stacked code and softmax

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201410334397.8A CN104156728B (en) 2014-07-14 2014-07-14 Polarized SAR image classification method based on stacked code and softmax

Publications (2)

Publication Number Publication Date
CN104156728A CN104156728A (en) 2014-11-19
CN104156728B true CN104156728B (en) 2017-05-24

Family

ID=51882225

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201410334397.8A CN104156728B (en) 2014-07-14 2014-07-14 Polarized SAR image classification method based on stacked code and softmax

Country Status (1)

Country Link
CN (1) CN104156728B (en)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104680180B (en) * 2015-03-09 2018-06-26 西安电子科技大学 Classification of Polarimetric SAR Image method based on K mean values and sparse own coding
CN105046322A (en) * 2015-07-03 2015-11-11 西南交通大学 Method for diagnosing leading screw faults
CN105825223A (en) * 2016-03-09 2016-08-03 西安电子科技大学 Polarization SAR terrain classification method based on deep learning and distance metric learning
CN106126896B (en) * 2016-06-20 2019-03-22 中国地质大学(武汉) Mixed model wind speed forecasting method and system based on empirical mode decomposition and deep learning
CN106203489B (en) * 2016-07-01 2019-02-15 西安电子科技大学 Classification of Polarimetric SAR Image method based on multiple dimensioned depth direction wave network
CN106503654A (en) * 2016-10-24 2017-03-15 中国地质大学(武汉) A kind of face emotion identification method based on the sparse autoencoder network of depth
CN106529428A (en) * 2016-10-31 2017-03-22 西北工业大学 Underwater target recognition method based on deep learning

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102156883A (en) * 2011-04-22 2011-08-17 北京航空航天大学 Classifying method utilizing polarizable SAR images under two different frequencies
CN102208031A (en) * 2011-06-17 2011-10-05 西安电子科技大学 Freeman decomposition and homo-polarization rate-based polarized synthetic aperture radar (SAR) image classification method
CN103093432A (en) * 2013-01-25 2013-05-08 西安电子科技大学 Polarized synthetic aperture radar (SAR) image speckle reduction method based on polarization decomposition and image block similarity
CN103839073A (en) * 2014-02-18 2014-06-04 西安电子科技大学 Polarization SAR image classification method based on polarization features and affinity propagation clustering

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7006656B2 (en) * 2001-10-15 2006-02-28 The Research Foundation Of Suny Lossless embedding of data in digital objects
US7653248B1 (en) * 2005-11-07 2010-01-26 Science Applications International Corporation Compression for holographic data and imagery

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102156883A (en) * 2011-04-22 2011-08-17 北京航空航天大学 Classifying method utilizing polarizable SAR images under two different frequencies
CN102208031A (en) * 2011-06-17 2011-10-05 西安电子科技大学 Freeman decomposition and homo-polarization rate-based polarized synthetic aperture radar (SAR) image classification method
CN103093432A (en) * 2013-01-25 2013-05-08 西安电子科技大学 Polarized synthetic aperture radar (SAR) image speckle reduction method based on polarization decomposition and image block similarity
CN103839073A (en) * 2014-02-18 2014-06-04 西安电子科技大学 Polarization SAR image classification method based on polarization features and affinity propagation clustering

Also Published As

Publication number Publication date
CN104156728A (en) 2014-11-19

Similar Documents

Publication Publication Date Title
Mertens et al. Using genetic algorithms in sub-pixel mapping
Dalla Mura et al. Morphological attribute profiles for the analysis of very high resolution images
Maggiori et al. Can semantic labeling methods generalize to any city? the inria aerial image labeling benchmark
Ghamisi et al. Multilevel image segmentation based on fractional-order Darwinian particle swarm optimization
Basu et al. Deepsat: a learning framework for satellite imagery
Corcoran et al. Influence of multi-source and multi-temporal remotely sensed and ancillary data on the accuracy of random forest classification of wetlands in Northern Minnesota
Myint et al. Per-pixel vs. object-based classification of urban land cover extraction using high spatial resolution imagery
Zhu et al. Assessment of spectral, polarimetric, temporal, and spatial dimensions for urban and peri-urban land cover classification using Landsat and SAR data
Kuenzer et al. Remote sensing of mangrove ecosystems: A review
Zhang et al. Fully-fuzzy supervised classification of sub-urban land cover from remotely sensed imagery: statistical and artificial neural network approaches
Zhou et al. Polarimetric SAR image classification using deep convolutional neural networks
Mas et al. Modelling deforestation using GIS and artificial neural networks
Maulik et al. Automatic fuzzy clustering using modified differential evolution for image classification
Mertens et al. A sub‐pixel mapping algorithm based on sub‐pixel/pixel spatial attraction models
Xu et al. Decision tree regression for soft classification of remote sensing data
Pedergnana et al. Classification of remote sensing optical and LiDAR data using extended attribute profiles
Zhong et al. Remote sensing image subpixel mapping based on adaptive differential evolution
Pacifici et al. Urban mapping using coarse SAR and optical data: Outcome of the 2007 GRSS data fusion contest
Tatem et al. Super-resolution land cover pattern prediction using a Hopfield neural network
Pacifici et al. An innovative neural-net method to detect temporal changes in high-resolution optical satellite imagery
Soltani-Farani et al. Spatial-aware dictionary learning for hyperspectral image classification
Puissant et al. Object-oriented mapping of urban trees using Random Forest classifiers
Dai et al. Data fusion using artificial neural networks: a case study on multitemporal change analysis
Rodriguez-Galiano et al. Land cover change analysis of a Mediterranean area in Spain using different sources of data: Multi-seasonal Landsat images, land surface temperature, digital terrain models and texture
Pradhan et al. An easy to use ArcMap based texture analysis program for extraction of flooded areas from TerraSAR-X satellite image

Legal Events

Date Code Title Description
PB01 Publication
C06 Publication
SE01 Entry into force of request for substantive examination
C10 Entry into substantive examination
GR01 Patent grant
GR01 Patent grant