CN105069479A - Polarized SAR image classification method based on online sequence limit learning machine - Google Patents

Abstract

The invention discloses a polarized SAR image classification method based on an online sequence limit learning machine, and mainly solves the problems that polarized SAR image classification is long in training time and large in memory consumption. The method is realized by the steps that 1) mark information and a coherence matrix of a polarized SAR image to be classified are input to carry out Lee filtering; 2) characteristics of the filtered coherence matrix are extracted, and normalized to obtain a data set, and a mark set is obtained from the data set; 3) the mark set is divided into an initial training set and online training sets; 4) the initial training set is used for learning in the initial stage, and a model learned in the initial stage is used to carry out initial classification o the polarized SAR image; and 5) the training sets are learned online via iteration, and a model learned in the online stage is used to carry out final classification on the polarized SAR image. The method can be used to process incremental training data, is relatively high in the classification precision, reduces training time and memory consumption, and can be used for terrain classification and object identification.

Description

Based on the Classification of Polarimetric SAR Image method of online limit of sequence learning machine

Technical field

The invention belongs to technical field of image processing, further relate to Classification of Polarimetric SAR Image method, can be used for terrain classification and target identification.

Background technology

Polarization SAR with scattering matrix or coherence matrix, covariance matrix record terrestrial object information.Different target is due to physical characteristics difference, and the amplitude under different polarization state, phase place, polarization ratio, scattering entropy all exists difference, therefore can obtain the information of abundanter ground object target.The task of Classification of Polarimetric SAR Image is exactly to have the atural object of similar quality to be divided into a class, determines the classification corresponding to each pixel of Polarimetric SAR Image specifically.

According to the need of artificial guidance, Classification of Polarimetric SAR Image can be divided into Supervised classification, semisupervised classification and unsupervised segmentation.Batch study and online Sequence Learning can be divided into according to the difference of training patterns.Batch study can regard a kind of special circumstances of online Sequence Learning as, and online Sequence Learning mode is more flexible, is more suitable for practical application.Existing Classification of Polarimetric SAR Image method is batch mode of learning.

The paper " Classification of Polarimetric SAR Image based on support vector machine " " modern radar " that Wu Yonghui delivers, article is numbered: disclose the method for a kind of support vector machines to Classification of Polarimetric SAR Image in 1004-7859.2007.06.017.The implementation procedure of the method is: first carry out feature extraction to Polarimetric SAR Image, and be normalized, and trains SVM classifier finally to classify to Polarimetric SAR Image by the SVM classifier trained afterwards, obtains classification results.The method is batch mode of learning, when increase new samples is used for training, then needs training sample to be before reused for training, adds the training time.

Summary of the invention

The object of the invention is to a kind of online classification method proposing Classification of Polarimetric SAR Image based on online limit of sequence learning machine, to solve the problem that the training time is long and memory consumption is large of the batch training study mode of prior art.

For achieving the above object, the present invention includes following steps:

(1) Polarimetric SAR Image label information to be sorted is inputted, input the coherence matrix that a width size is the Polarimetric SAR Image to be sorted of 3 × 3 × M, and with Lee wave filter filtering coherent noise, obtain filtered coherence matrix, wherein, in filtered coherence matrix, each element is 3 × 3 matrixes, and M represents the sum of Polarimetric SAR Image pixel to be sorted;

(2) by 3 × 3 corresponding for element each in coherence matrix after filtering matrixes, pull into the proper vector of one 9 dimension, obtaining size is the data matrix of 9 × M;

(3) data matrix is normalized, obtains data set wherein n=9 is the dimension of feature, x _ifor proper vector;

(4) test set N is obtained respectively from data centralization _p, initial training collection with on-line training collection

(4a) according to label information from data set middle acquisition label sets

wherein m is classification sum, and N is the sum of marker samples, t _ifor categorization vector;

(4b) using label sets random selecting 90% as test set: remain 10% as training set, wherein N _pfor the number of samples of test set;

(4c) training set is divided into 5 parts, namely 1 part is initial training collection with 4 parts of on-line training collection

k is on-line study iteration mark, k=0,1,2,3.

Wherein N ₀for the number of samples of initial training collection, N _k+1for on-line training collection number of samples;

(5) activation function arranging extreme learning machine hidden layer node is G (a, b, x)=exp (-b||x-a|| ²), hidden layer unit number is L, and N ₀> L, arranges input weights a at random _jpartially be worth b _j, j=1 ..., L;

(6) initial training collection is used carry out the study of starting stage, obtaining and exporting weights is β ⁰with intermediate parameters P ₀, initial on-line study iteration mark k=0 is set;

(7) by data set be input in the starting stage extreme learning machine trained, obtain starting stage data set classification results be Y ₀:

(7a) input data set calculating hidden layer output matrix is Q,

$Q={\left[\begin{array}{ccccc}G(a_1,b_1,x_1)& ...& G(a_j,b_j,x_1)& ...& G(a_L,b_L,x_1)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_i)& ...& G(a_j,b_j,x_i)& ...& G(a_L,b_L,x_i)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_M)& ...& G(a_j,b_j,x_M)& ...& G(a_L,b_L,x_M)\end{array}\right]}_{M\×L};$

(7b) according to output weights β ⁰with output matrix Q, calculate starting stage data set classification results be Y ₀=Q β ⁰;

(8) online training set is used carry out the study of on-line stage, the classification results obtaining on-line stage is Y _k+1:

(8a) Input Online training set the output matrix calculating hidden layer is H _k+1,

$H_{k+1}={\left[\begin{array}{ccccc}G(a_1,b_1,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})& ...& G(a_j,b_j,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})& ...& G(a_L,b_L,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_i)& ...& G(a_j,b_j,x_i)& ...& G(a_L,b_L,x_i)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})& ...& G(a_j,b_j,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})& ...& G(a_L,b_L,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})\end{array}\right]}_{N_{k+1}\×L};$

(8b) the output weights in on-line study stage are calculated ${\mathrm{\β}}^{(k+1)}={\mathrm{\β}}^{\left(k\right)}-P_{k+1}{H}_{k+1}^T(T_{k+1}-H_{k+1}{\mathrm{\β}}^{\left(k\right)}),$

Wherein upper right mark ^tfor transpose operation, upper right mark ^-1for matrix inversion operation;

(8c) according to output weights β ^(k+1)with output matrix Q, calculate data set the classification results of on-line stage be Y _k+1=Q β ^(k+1);

(8d) on-line study iteration mark k=k+1 is set, returns circulation in step (8a), until terminate when meeting termination condition k=4.

The present invention has the following advantages compared with prior art:

1. the present invention utilizes Lee filtering to carry out pre-service to original polarization SAR, effectively reduces coherent speckle noise, improves quality and the classification performance of image;

2. the present invention utilizes online limit of sequence learning machine method to Classification of Polarimetric SAR Image, and this algorithm Generalization Capability is good, and nicety of grading is high.

3. the present invention utilizes online limit of sequence learning machine method to Classification of Polarimetric SAR Image, and this algorithm is on-line training learning algorithm, can process the training data of increment.Compare traditional batch learning algorithm, training time and memory consumption can be reduced.

4. the present invention utilizes online limit of sequence learning machine method to Classification of Polarimetric SAR Image, and because this algorithm has analytic solution, without the need to iterative weight parameter, so training speed is fast, working time is short;

Accompanying drawing explanation

Fig. 1 is realization flow figure of the present invention;

Fig. 2 be the present invention emulate use area, Fu Laifulan farmland polarization SAR data after filtering after pseudocolour picture;

Fig. 3 is the figure of substance markers practically that the present invention emulates area, the Fu Laifulan farmland polarization SAR data of use;

Fig. 4 is the present invention emulates area, the Fu Laifulan farmland polarization SAR data of use classification results figure to the present invention.

Embodiment

With reference to Fig. 1, specific implementation step of the present invention is as follows:

Step 1, input information, obtains filtered coherence matrix.

Input Polarimetric SAR Image label information to be sorted;

Input the coherence matrix that a width size is the Polarimetric SAR Image to be sorted of 3 × 3 × M;

With the Lee wave filter that window size is 11 × 11, filtering is carried out to the coherence matrix of Polarimetric SAR Image, removing coherent noise, obtains filtered coherence matrix, wherein, in filtered coherence matrix, each element is 3 × 3 matrixes, and M represents the sum of Polarimetric SAR Image pixel to be sorted.

Step 2, obtains data matrix according to filtered coherence matrix.

By 3 × 3 corresponding for element each in coherence matrix after filtering matrixes, pull into the proper vector of one 9 dimension, obtaining size is the data matrix of 9 × M.

Step 3, obtains data set according to data matrix.

Data matrix is normalized, obtains data set wherein n=9 is the dimension of feature, x _ifor proper vector.

Normalized common method has characteristic line convergent-divergent, feature normalization and albefaction, adopts feature normalization process in this example.

Step 4, obtains test set N respectively from data centralization _p, initial training collection with on-line training collection

(4.1) according to label information from data set middle acquisition label sets

wherein m is classification sum, and N is the sum of marker samples, t _ifor categorization vector;

(4.2) using label sets random selecting 90% as test set: remain 10% as training set, wherein N _pfor the number of samples of test set;

(4.3) training set is divided into 5 parts, namely 1 part is initial training collection with 4 parts of on-line training collection

k is on-line study iteration mark, k=0,1,2,3, wherein N ₀for the number of samples of initial training collection, N _k+1for on-line training collection number of samples.

Step 5, the activation function arranging extreme learning machine hidden layer node is G (a, b, x)=exp (-b||x-a|| ²), hidden layer unit number is L, and N ₀> L, arranges input weights a at random _jpartially be worth b _j, j=1 ..., L.

Step 6, uses initial training collection carry out the study of starting stage:

(6.1) training set is inputted calculating hidden layer output matrix is H ₀,

$H_0={\left[\begin{array}{ccccc}G(a_1,b_1,x_1)& ...& G(a_j,b_j,x_1)& ...& G(a_L,b_L,x_1)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_i)& ...& G(a_j,b_j,x_i)& ...& G(a_L,b_L,x_i)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_{N_0})& ...& G(a_j,b_j,x_{N_0})& ...& G(a_L,b_L,x_{N_0})\end{array}\right]}_{N_0\×L};$

(6.2) the output weights calculating initial learning period are β ⁰=P ₀h ₀ ^tt ₀, wherein primary iteration variable is classification matrix is upper right mark ^tfor transpose operation, upper right mark ^-1for matrix inversion operation;

(6.3) initial on-line study iteration mark k=0 is set.

Step 7, by data set be input in the starting stage extreme learning machine trained, obtain starting stage data set classification results be Y ₀.

(7.1) input data set calculating hidden layer output matrix is Q,

$Q={\left[\begin{array}{ccccc}G(a_1,b_1,x_1)& ...& G(a_j,b_j,x_1)& ...& G(a_L,b_L,x_1)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_i)& ...& G(a_j,b_j,x_i)& ...& G(a_L,b_L,x_i)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_M)& ...& G(a_j,b_j,x_M)& ...& G(a_L,b_L,x_M)\end{array}\right]}_{M\×L};$

(7.2) according to output weights β ⁰with output matrix Q, calculate starting stage data set classification results be Y ₀=Q β ⁰.

Step 8, uses online training set carry out the study of on-line stage, the classification results obtaining on-line stage is Y _k+1.

(8.1) Input Online training set the output matrix calculating hidden layer is H _k+1,

$H_{k+1}={\left[\begin{array}{ccccc}G(a_1,b_1,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})& ...& G(a_j,b_j,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})& ...& G(a_L,b_L,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_i)& ...& G(a_j,b_j,x_i)& ...& G(a_L,b_L,x_i)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})& ...& G(a_j,b_j,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})& ...& G(a_L,b_L,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})\end{array}\right]}_{N_{k+1}\×L};$

(8.2) the output weights in on-line study stage are calculated ${\mathrm{\β}}^{(k+1)}={\mathrm{\β}}^{\left(k\right)}-P_{k+1}{H}_{k+1}^T(T_{k+1}-H_{k+1}{\mathrm{\β}}^{\left(k\right)}),$

Wherein online iteration variable is $P_{k+1}=P_k-P_k{H}_{k+1}^T{(I+H_{k+1}{P}_k{H}_{k+1}^T)}^{-1}{H}_{k+1}{P}_k;$

(8.3) according to output weights β ^(k+1)with output matrix Q, calculate data set the classification results of on-line stage be Y _k+1=Q β ^(k+1);

(8.4) on-line study iteration mark k=k+1 is set, returns circulation in step (8.1), until terminate when meeting termination condition k=4.

Effect of the present invention can be further illustrated by experiment simulation below:

1, emulation platform

Hardware platform is: Intel third generation Duo i5-3230M2.60GHz double-core, 8GBRAM;

Software platform is: MATLABR2013a;

The Polarimetric SAR Image that emulation uses as shown in Figure 2, image size is 750 × 1024, and resolution is 12 × 6 meters, and this Fig. 2 is area, the Fu Laifulan farmland polarization SAR data that AIRSAR system obtained in 1989, wherein have 15 class atural objects, substance markers figure as shown in Figure 3 practically.

The online limit of sequence learning machine OS-ELM of the present invention classifies to area, Fu Laifulan farmland Polarimetric SAR Image, and control methods has: based on the reverse transmittance nerve network method BPNN of batch study, based on the support vector machine method SVM of batch study.Five learning processes are carried out in emulation, and in each learning process, training sample amount increases with 2% of marker samples quantity, is contrasted by the operation result of above distinct methods.

2, content and result is emulated

Emulation 1, classify to area, Fu Laifulan farmland Polarimetric SAR Image by the present invention's online limit of sequence learning machine method, the final classification results of on-line study as shown in Figure 4.As seen from Figure 4, except a little wrong point spot, region, every class farmland all obtains more accurate classification results, and the boundary edges between different croplands region is comparatively level and smooth, and the junction edge clear in waters and farmland can be distinguished.

Emulation 2, classifies to area, Fu Laifulan farmland Polarimetric SAR Image by existing batch BPNN method and batch SVM method respectively.

Five learning processes are carried out in above-mentioned each emulation, and in each learning process, training sample amount increases with 2% of marker samples quantity, and accuracy rate and the working time of its classification are as shown in table 1.

Table 1: the classification accuracy of the present invention and prior art and working time Performance comparision

As seen from Table 1, along with the increase of training sample, the accuracy rate of three kinds of methods increases all simultaneously, and the increase trend of the inventive method is more obvious.When training sample amount is 10% of marker samples quantity time, the accuracy rate of three kinds of methods is suitable, is all 0.966.Along with the increase of training sample, learning time of the present invention does not almost increase, and the initial learn time is roughly 5 seconds, and the on-line study time is roughly 9 seconds.And the existing learning time based on learning method BPNN in batches and SVM constantly increases along with the increase of training sample, the learning time of BPNN was increased to 251 seconds from 26 seconds, and the learning time of SVM was increased to 142 seconds from 13 seconds, all consuming time than the present invention.

To sum up, by process Classification of Polarimetric SAR Image problem of the present invention, the training data of increment can be processed, higher nicety of grading can be obtained, training time and the memory consumption of Classification of Polarimetric SAR Image issue handling can also be reduced.

Claims (4)

1., based on a Classification of Polarimetric SAR Image method for online limit of sequence learning machine, comprise the steps:

(1) Polarimetric SAR Image label information to be sorted is inputted, input the coherence matrix that a width size is the Polarimetric SAR Image to be sorted of 3 × 3 × M, and with Lee wave filter filtering coherent noise, obtain filtered coherence matrix, wherein, in filtered coherence matrix, each element is 3 × 3 matrixes, and M represents the sum of Polarimetric SAR Image pixel to be sorted;

(2) by 3 × 3 corresponding for element each in coherence matrix after filtering matrixes, pull into the proper vector of one 9 dimension, obtaining size is the data matrix of 9 × M;

(3) data matrix is normalized, obtains data set wherein n=9 is the dimension of feature, x _ifor proper vector;

(4) test set N is obtained respectively from data centralization _p, initial training collection with on-line training collection

(4a) according to label information from data set middle acquisition label sets

wherein m is classification sum, and N is the sum of marker samples, t _ifor categorization vector;

(4b) using label sets random selecting 90% as test set: remain 10% as training set, wherein N _pfor the number of samples of test set;

(4c) training set is divided into 5 parts, namely 1 part is initial training collection with 4 parts of on-line training collection

k is on-line study iteration mark, k=0,1,2,3.

Wherein N ₀for the number of samples of initial training collection, N _k+1for on-line training collection number of samples;

(5) activation function arranging extreme learning machine hidden layer node is G (a, b, x)=exp (-b||x-a|| ²), hidden layer unit number is L, and N ₀> L, arranges input weights a at random _jpartially be worth b _j, j=1 ..., L;

(6) initial training collection is used carry out the study of starting stage, obtaining and exporting weights is β ⁰with intermediate parameters P ₀, initial on-line study iteration mark k=0 is set;

(7) by data set be input in the starting stage extreme learning machine trained, obtain starting stage data set classification results be Y ₀:

(7a) input data set calculating hidden layer output matrix is Q,

$Q={\left[\begin{array}{ccccc}G(a_1,b_1,x_1)& ...& G(a_j,b_j,x_1)& ...& G(a_L,b_L,x_1)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_i)& ...& G(a_j,b_j,x_i)& ...& G(a_L,b_L,x_i)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_M)& ...& G(a_j,b_j,x_M)& ...& G(a_L,b_L,x_M)\end{array}\right]}_{M\×L};$

(7b) according to output weights β ⁰with output matrix Q, calculate starting stage data set classification results be Y ₀=Q β ⁰;

(8) online training set is used carry out the study of on-line stage, the classification results obtaining on-line stage is Y _k+1:

(8a) Input Online training set the output matrix calculating hidden layer is H _k+1,

$H_{k+1}={\left[\begin{array}{ccccc}G(a_1,b_1,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})& ...& G(a_j,b_j,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})& ...& G(a_L,b_L,x_{\left({\mathrm{\Σ}}_{j=0}^k{N}_j\right)+1})\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_i)& ...& G(a_j,b_j,x_i)& ...& G(a_L,b_L,x_i)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})& ...& G(a_j,b_j,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})& ...& G(a_L,b_L,x_{{\mathrm{\Σ}}_{j=0}^{k+1}{N}_j})\end{array}\right]}_{N_{k+1}\×L};$

(8b) the output weights in on-line study stage are calculated ${\mathrm{\β}}^{(k+1)}={\mathrm{\β}}^{\left(k\right)}-P_{k+1}{H}_{k+1}^T(T_{k+1}-H_{k+1}{\mathrm{\β}}^{\left(k\right)}),$

Wherein $P_{k+1}=P_k-P_k{H}_{k+1}^T{(I+H_{k+1}{P}_k{H}_{k+1}^T)}^{-1}{H}_{k+1}{P}_k,$ Upper right mark ^tfor transpose operation, upper right mark ^-1for matrix inversion operation;

(8c) according to output weights β ^(k+1)with output matrix Q, calculate data set the classification results of on-line stage be Y _k+1=Q β ^(k+1);

(8d) on-line study iteration mark k=k+1 is set, returns circulation in step (8a), until terminate when meeting termination condition k=4.

2. the Polarimetric SAR Image terrain classification method based on extreme learning machine according to claim 1, in wherein said step (1), the window size of Lee wave filter is 11 × 11.

3. the Polarimetric SAR Image terrain classification method based on extreme learning machine according to claim 1, in wherein said step (3), normalized takes feature normalization, makes every one-dimensional characteristic of data set have zero-mean and unit variance.

4. the Polarimetric SAR Image terrain classification method based on extreme learning machine according to claim 1, calculates in wherein said step (6) and exports weights β ⁰with intermediate parameters P ₀, carry out as follows:

(6a) training set is inputted calculating hidden layer output matrix is H ₀,

$H_0={\left[\begin{array}{ccccc}G(a_1,b_1,x_1)& ...& G(a_j,b_j,x_1)& ...& G(a_L,b_L,x_1)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_i)& ...& G(a_j,b_j,x_i)& ...& G(a_L,b_L,x_i)\\ \·& & & & \·\\ \·& ...& & ...& \·\\ \·& & & & \·\\ G(a_1,b_1,x_{N_0})& ...& G(a_j,b_j,x_{N_0})& ...& G(a_L,b_L,x_{N_0})\end{array}\right]}_{N_0\×L};$

(6b) the output weights calculating initial learning period are β ⁰=P ₀h ₀ ^tt ₀, wherein $T_0={[t_1,....t_{N_0}]}^T.$