CN112949776A - Polarized SAR image classification method - Google Patents

Abstract

The invention discloses a polarized SAR image classification method, and belongs to the technical field of image processing. The method comprises the following steps: carrying out refined Lee filtering on the polarized SAR image; acquiring polarization coherent matrix T and H/A/alpha decomposition characteristics, and setting the polarization coherent matrix T and H/A/alpha decomposition characteristics as image original characteristics; normalizing each element of the original features of the image; training a half-coupled projection dictionary pair learning model by using the training samples; performing feature extraction on the polarized SAR image by using a half-coupling projection dictionary to the learning model; and inputting the extracted features into an RFC classifier to obtain a classification result. The half-coupling projection dictionary can obtain essential relation among different characteristics for a learning model, can improve the classification result of the polarized SAR image by combining an RFC classifier, and can solve the problem of low classification precision of the existing polarized SAR image classification method.

Description

Polarized SAR image classification method

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a polarized SAR image classification method.

Background

In recent decades, remote sensing technology has been rapidly developed, and polarized SAR is becoming more important in the field of remote sensing. The polarized SAR is not limited by time and weather during working, and can continuously provide images with rich polarized information, so that the polarized SAR is widely applied to the civil and military fields, such as agricultural production, geological exploration, ocean monitoring, military reconnaissance and the like. The polarized SAR classification is the most basic application in polarized SAR application, and a series of polarized SAR classification algorithms are proposed in succession.

It is well known that polarimetric SAR image classification is actually a high-dimensional mapping problem, and therefore, high-level processing and learning methods work well in polarimetric SAR image classification. The key problem of sparse representation is dictionary learning, however, the lp norm is used in most of the previous dictionary learning models, so that sparsity of sparse parameters is guaranteed, and a great calculation problem is brought. Fortunately, a projection dictionary pair learning model is provided, which does not use lp norm, creatively provides a projection dictionary pair framework and provides a very effective solving mode, so that the projection dictionary pair learning model can obtain a good classification result, and the time consumption of the training and testing stage is obviously reduced compared with the prior dictionary learning method.

The polarization SAR image often has more than one extracted feature, such as a polarization coherence matrix, a polarization covariance matrix, various polarization decompositions, color features, and the like, and the conventional method is to pull the features into a feature vector, however, the operation does not take into consideration the essential relation between the features. Fortunately, a half-coupled dictionary learning model is proposed that can be used to obtain essential connections between different features, but that also uses the lp norm and is therefore time consuming in the training and testing phase.

Most of the prior discrimination dictionary learning models often use sparse coding classifiers, the classifiers are particularly dependent on dictionary learning effects, and no classification help is provided for samples with poor dictionary learning effects, so that the extraction of features through dictionary learning and the application of classifiers with excellent performance such as support vector machines and random forests also become a new idea for polarized SAR image classification.

Disclosure of Invention

The invention aims to solve the problems, and provides a polarized SAR image classification method by taking a projection dictionary learning model and a half-coupled dictionary learning model as basic models and taking a Random Forest Classifier (RFC) as a classifier.

In order to achieve the purpose, the invention adopts the technical scheme that:

a polarized SAR image classification method comprises the following steps:

step 1, carrying out refined Lee filtering on a polarized SAR image, and filtering speckle noise;

step 2, acquiring a polarization coherent matrix T and H/A/alpha decomposition characteristics of the polarization SAR image, and setting the polarization coherent matrix T and the H/A/alpha decomposition characteristics as original characteristics of the image;

step 3, normalizing each element of the original features of the image to an interval [0,1 ];

step 4, randomly selecting a marked sample of a part of polarized SAR image, and setting the marked sample as a training sample;

step 5, training a half-coupling projection dictionary pair learning model by using the image original features of the training samples;

step 6, using the half-coupling projection dictionary obtained by training in the step 5 to perform feature extraction on the polarized SAR image to be classified of the learning model;

and 7, inputting the features extracted in the step 6 into an RFC classifier to obtain a classification result.

Further, the specific steps of step 2 are as follows:

2a) selecting three elements at the upper triangle of a polarized coherent matrix T and three elements at the diagonal of the polarized coherent matrix T of each polarized SAR image as a first part of the original characteristics,the symbol being X₁；

2b) Performing H/A/alpha decomposition on the polarization coherent matrix T of each polarization SAR image, selecting 3 non-negative characteristic values, entropy, average scattering angle and anisotropic parameters, and marking the non-negative characteristic values, the entropy, the average scattering angle and the anisotropic parameters as a second part of the original characteristics as X₂；

2c) Combining the features obtained in 2a) and 2b), and marking the combined features as original features of the image as X ═ X₁,X₂}。

Further, the half-coupled projection dictionary pair learning model in step 5 is described as follows:

in the formula, τ₁，τ₂，λ₁，λ₂，θ₁，θ₂Is a fixed scalar quantity, X ═ X₁,X₂Denotes the raw features of the polarimetric SAR image as input to the model from the half-coupled projection dictionary, where X₁＝[X₁₁,...,X_k1...,X_K1]，X₂＝[X₁₂,...,X_k2...,X_K2]K denotes the number of classes of the polarized SAR image, a ═ a₁,A₂Denotes an encoding matrix, where A₁＝[A₁₁,...,A_k1...,A_K1]，A₂＝[A₁₂,...,A_k2...,A_K2]，D＝{D₁,D₂Denotes a composite dictionary, in which D₁＝[D₁₁,...,D_k1...,D_K1]，D₂＝[D₁₂,...,D_k2...,D_K2]，P＝{P₁,P₂Denotes an analytical dictionary, where P₁＝[P₁₁,...,P_k1...,P_K1]，P₂＝[P₁₂,...,P_k2...,P_K2]Synthesizing the dictionary D₁And an analysis dictionary P₁Representing input data X₁Dictionary pair of, composite dictionary D₂And an analysis dictionary P₂Representing input data X₂The pair of dictionaries of (1),

represents X_k1At X₁The complement of (1) is set as (c),

represents X_k2At X₂W represents a mapping matrix, d_i1And d_i2Each represents D₁And D₂The (i) th atom of (c),

and

a representation of the fidelity item of the data,

and

a discrimination item is represented by a number of items,

is represented by A₁And A₂The mapping equation of (1).

Further, the specific steps of step 6 are as follows:

6a) a is to be₁Arranged to extract a first part of the feature, A₁＝[A₁₁,...,A_k1...,A_K1]Wherein A is_k1＝P_k1X_k1；

6b) A is to be₂Arranged to extract a second part of the feature, A₂＝[A₁₂,...,A_k2...,A_K2]Wherein A is_k2＝P_k2X_k2；

6c) Combining the features obtained from 6a) and 6b), and marking as a ═ a₁,A₂}；

6d) And inputting the characteristic A into the half-coupling projection dictionary pair learning model obtained by training in the step 5, thereby extracting the characteristic of the polarized SAR image.

Further, the specific steps of step 7 are as follows:

7a) using a semi-coupling projection dictionary to extract the features of the training samples from the learning model as input data of an RFC classifier, and training the RFC classifier;

7b) after the RFC classifier is trained, inputting the feature extraction result of the polarized SAR image to be classified into the RFC classifier to obtain the classification result of the whole polarized SAR image to be classified.

Compared with the prior art, the invention has the following advantages:

1. the invention adopts the T matrix and the H/A/alpha decomposition characteristic as the original characteristic, the H/A/alpha decomposition characteristic is obtained by carrying out H/A/alpha polarization decomposition on the T matrix, and the two characteristics are connected with each other, so that the invention is very suitable for taking the two characteristics as the input characteristic of learning of the half-coupling projection dictionary pair.

2. The traditional method is to pull the extracted different features into a feature vector, which can cause some relations between different features to be not fully considered. The half-coupling projection dictionary can acquire essential relation among different characteristics for a learning mode, and is beneficial to improving a classification result.

3. The RFC classifier adopted by the invention is a classifier with excellent classification performance, and is beneficial to obtaining a better classification result.

In conclusion, the method is based on the learning and RFC of the half-coupling projection dictionary pair, and can be used for the terrain classification and the target recognition of the polarized SAR image.

Drawings

FIG. 1 is a general flow diagram of the process of the present invention;

FIG. 2 is a schematic diagram of the method of the present invention;

FIG. 3 is an image of the region of the Chinese Weinweih simulated by the present invention, wherein (a) - (c) respectively show Pauli pseudo-color drawing, real quasi-color drawing and color scheme;

FIG. 4 is a graph of the classification results for the image of FIG. 3 using the method of the present invention;

FIG. 5 is an image of the golden Gate bridge around the U.S. san Francisco bay, used in the simulation of the present invention, where (a) - (c) represent Pauli pseudo-color drawing, real pseudo-color drawing, and color scheme, respectively;

fig. 6 is a graph of the classification results for the image of fig. 5 using the method of the present invention.

Detailed Description

The technical solution of the present invention will be described in detail with reference to the accompanying drawings and the detailed description.

Referring to fig. 1 and 2, a polarized SAR image classification method includes the following steps:

step 1, inputting a polarized SAR image to be classified, and carrying out refined Lee filtering on the image to filter speckle noise.

Step 2, acquiring polarization coherent matrix T and H/A/alpha decomposition characteristics, and setting the characteristics as image original characteristics, specifically comprising the following steps:

2a) selecting the real part and imaginary part of three elements at the upper triangle of the T matrix of each sample and three elements on the diagonal of the T matrix as the first part of the original characteristic, marked as X₁；

2b) Performing H/A/alpha decomposition on the T matrix of each sample, selecting 3 non-negative characteristic values, entropy, average scattering angle and anisotropic parameters, and marking the values as a second part of the original characteristics as X₂；

2c) Combining the features obtained in 2a) and 2b), and marking the combined features as original features of the image as X ═ X₁,X₂}。

And 3, normalizing each element of the original features of the image to [0,1 ].

And 4, randomly selecting marked samples of the 5% polarized SAR images as training samples, and setting the rest 95% marked samples as test samples.

And 5, training a learning model of the half-coupling projection dictionary pair by using the training samples obtained in the step 4, wherein the learning model of the half-coupling projection dictionary pair is described as follows:

in the formula, τ₁，τ₂，λ₁，λ₂，θ₁，θ₂Is a fixed scalar quantity, X ═ X₁,X₂Representing original features of the polarized SAR image as input of a half-coupled projection dictionary to the model, wherein X₁＝[X₁₁,...,X_k1...,X_K1]，X₂＝[X₁₂,...,X_k2...,X_K2]Where K denotes the number of polarimetric SAR image classes, a ═ a₁,A₂Denotes an encoding matrix, where A₁＝[A₁₁,...,A_k1...,A_K1]，A₂＝[A₁₂,...,A_k2...,A_K2]，D＝{D₁,D₂Denotes a composite dictionary, in which D₁＝[D₁₁,...,D_k1...,D_K1]，D₂＝[D₁₂,...,D_k2...,D_K2]，P＝{P₁,P₂Denotes an analytical dictionary, where P₁＝[P₁₁,...,P_k1...,P_K1]，P₂＝[P₁₂,...,P_k2...,P_K2]Synthesizing the dictionary D₁And an analysis dictionary P₁Representing input data X₁Dictionary pair of, composite dictionary D₂And an analysis dictionary P₂Representing input data X₂The pair of dictionaries of (1),

represents X_k1At X₁The complement of (1) is set as (c),

represents X_k2At X₂W represents a mapping matrix, d_i1And d_i2Each represents D₁And D₂The (i) th atom of (c),

and

representing data guaranteesThe true term is the term that is used,

and

a discrimination item is represented by a number of items,

is represented by A₁And A₂The mapping equation of (1).

And 6, performing feature extraction on the polarized SAR image by using the half-coupling projection dictionary obtained by training in the step 5, wherein the specific operation mode is as follows:

6a)A₁＝[A₁₁,...,A_k1...,A_K1]wherein A is_k1＝P_k1X_k1A is₁A first portion configured to extract features;

6b)A₂＝[A₁₂,...,A_k2...,A_K2]wherein A is_k2＝P_k2X_k2A is₂A second portion configured to extract features;

6c) combining the features obtained from 6a) and 6b), and marking as a ═ a₁,A₂}。

And 7, extracting features and inputting the features into the RFC classifier by using the step 6, and counting the classification result of the test sample, wherein the specific operation steps are as follows:

7a) using the part of the extracted features corresponding to the training samples in the step 6 as input data of RFC, and training an RFC classifier, wherein the number of decision trees of the RFC is set to be 100;

7b) after the RFC classifier is trained, inputting the extracted features in the step 6 into the RFC classifier to obtain a classification result of the whole image;

7c) the classification accuracy of the test samples was counted, and the Overall Accuracy (OA) and Kappa coefficient were used as evaluation indexes.

The effect of the method can be further illustrated by the following simulation experiment:

1. experimental conditions and methods

The hardware platform is as follows: intel (R) core (TM) i5-9400F CPU @2.90GHZ, 32GB RAM;

the software platform is as follows: MATLAB 2018;

the experimental method comprises the following steps: RFC, inventive method (SDPL-RFC).

2. Simulation content and results

Experiment I, the image of the region of the Weinwei of China shown in FIG. 3 is used as a test image, the RFC and the method of the present invention are used for carrying out classification simulation on the image of FIG. 3, the classification result based on the RFC is shown in FIG. 4(a), the classification result based on the method of the present invention is shown in FIG. 4(b), and it can be seen from FIG. 4 that the classification result based on the method of the present invention is obviously improved compared with the classification result based on the RFC. Table 1 shows the overall classification accuracy (OA) and Kappa coefficient of the image of the west-wei-river region. It can be seen that the method of the present invention indeed further improves the classification result of the full convolution network.

TABLE 1 accuracy of image classification in the Weianwei river region

Method	Water area	Grass land	Building area	OA	Kappa coefficient
						RFC	0.8950	0.9198	0.9016	0.9097	0.8505
SDPL-RFC	0.9095	0.9280	0.9346	0.9275	0.8805

Experiment two, the image of the golden gate bridge around the U.S. san francisco bay shown in fig. 5 is used as a test image, the RFC and the method of the present invention are used to perform classification simulation on fig. 5, the classification result based on the RFC is shown in fig. 6(a), the classification result based on the method of the present invention is shown in fig. 6(b), and it can be seen from fig. 6 that the classification result based on the method of the present invention is significantly improved compared with the classification result based on the RFC. Table 2 shows OA and Kappa coefficients of images of the golden gate bridge around the gulf of san francisco, usa. It can be seen that the method of the present invention indeed further improves the classification result of the full convolution network.

TABLE 2 Classification accuracy of Okinawa bridge images around the U.S. san Francisco bay

Method	Water area	Vegetation	Low densityRegion(s)	High dense region	Developed area	OA	Kappa coefficient
								RFC	0.9999	0.9358	0.8889	0.8480	0.8748	0.9404	0.9143
SDPL-RFC	0.9999	0.9412	0.9281	0.8901	0.9351	0.9581	0.9397

In a word, the half-coupling projection dictionary can obtain essential relation among different characteristics for a learning model, can improve the classification result of the polarized SAR image by combining an RFC classifier, and can solve the problem of low classification precision of the existing polarized SAR image classification method.

Claims (5)

1. A polarized SAR image classification method is characterized by comprising the following steps:

step 1, carrying out refined Lee filtering on a polarized SAR image, and filtering speckle noise;

step 2, acquiring a polarization coherent matrix T and H/A/alpha decomposition characteristics of the polarization SAR image, and setting the polarization coherent matrix T and the H/A/alpha decomposition characteristics as original characteristics of the image;

step 3, normalizing each element of the original features of the image to an interval [0,1 ];

step 4, randomly selecting a marked sample of a part of polarized SAR image, and setting the marked sample as a training sample;

step 5, training a half-coupling projection dictionary pair learning model by using the image original features of the training samples;

step 6, using the half-coupling projection dictionary obtained by training in the step 5 to perform feature extraction on the polarized SAR image to be classified of the learning model;

and 7, inputting the features extracted in the step 6 into an RFC classifier to obtain a classification result.

2. The polarized SAR image classification method according to claim 1, characterized in that the specific steps of step 2 are as follows:

2a) selecting three elements at the upper triangle of a polarized coherent matrix T and three elements at the diagonal of the polarized coherent matrix T of each polarized SAR image, and taking the three elements as a first part of the original characteristics and marking the three elements as X₁；

2b) Performing H/A/alpha decomposition on the polarization coherent matrix T of each polarization SAR image, selecting 3 non-negative characteristic values, entropy, average scattering angle and anisotropic parameters, and marking the non-negative characteristic values, the entropy, the average scattering angle and the anisotropic parameters as a second part of the original characteristics as X₂；

2c) Combining the features obtained in 2a) and 2b), and marking the combined features as original features of the image as X ═ X₁,X₂}。

3. The polarized SAR image classification method according to claim 1, characterized in that the half-coupled projection dictionary-to-learning model of step 5 is described as follows:

in the formula, τ₁，τ₂，λ₁，λ₂，θ₁，θ₂Is a fixed scalar quantity, X ═ X₁,X₂Denotes the raw features of the polarimetric SAR image as input to the model from the half-coupled projection dictionary, where X₁＝[X₁₁,...,X_k1...,X_K1]，X₂＝[X₁₂,...,X_k2...,X_K2]K denotes the number of classes of the polarized SAR image, a ═ a₁,A₂Denotes an encoding matrix, where A₁＝[A₁₁,...,A_k1...,A_K1]，A₂＝[A₁₂,...,A_k2...,A_K2]，D＝{D₁,D₂Denotes a composite dictionary, in which D₁＝[D₁₁,...,D_k1...,D_K1]，D₂＝[D₁₂,...,D_k2...,D_K2]，P＝{P₁,P₂Denotes an analytical dictionary, where P₁＝[P₁₁,...,P_k1...,P_K1]，P₂＝[P₁₂,...,P_k2...,P_K2]Synthesizing the dictionary D₁And an analysis dictionary P₁Representing input data X₁Dictionary pair of, composite dictionary D₂And an analysis dictionary P₂Representing input data X₂The pair of dictionaries of (1),

represents X_k1At X₁The complement of (1) is set as (c),

represents X_k2At X₂W represents a mapping matrix, d_i1And d_i2Each represents D₁And D₂The (i) th atom of (c),

and

representing dataThe fidelity items are the items that are presented,

and

a discrimination item is represented by a number of items,

is represented by A₁And A₂The mapping equation of (1).

4. The polarized SAR image classification method according to claim 3, characterized in that, the specific steps of step 6 are as follows:

6a) a is to be₁Arranged to extract a first part of the feature, A₁＝[A₁₁,...,A_k1...,A_K1]Wherein A is_k1＝P_k1X_k1；

6b) A is to be₂Arranged to extract a second part of the feature, A₂＝[A₁₂,...,A_k2...,A_K2]Wherein A is_k2＝P_k2X_k2；

6c) Combining the features obtained from 6a) and 6b), and marking as a ═ a₁,A₂}；

6d) And inputting the characteristic A into the half-coupling projection dictionary pair learning model obtained by training in the step 5, thereby extracting the characteristic of the polarized SAR image.

5. The polarized SAR image classification method according to claim 1, characterized in that, the specific steps of step 7 are as follows:

7a) using a semi-coupling projection dictionary to extract the features of the training samples from the learning model as input data of an RFC classifier, and training the RFC classifier;

7b) after the RFC classifier is trained, inputting the feature extraction result of the polarized SAR image to be classified into the RFC classifier to obtain the classification result of the whole polarized SAR image to be classified.

Images