CN110956221A - Small sample polarization synthetic aperture radar image classification method based on deep recursive network - Google Patents

Abstract

The invention discloses a small sample PolSAR image classification method based on a deep recursive network, which firstly performs feature enhancement on small sample data by using the complexity of the deep recursive network and the advantages of spatial analysis, then selects 0.5% of samples as training samples, and simultaneously uses only 5 feature sequences as training samples at each point. In addition, the invention provides a sample synthesis method of RNBP (radio network Back propagation) to improve the richness of training samples, and simultaneously provides MB and MBW (multi-band wavelet) to improve the robustness of the test sample, thereby achieving a better classification effect of small samples. In the construction of the deep recursion network, the invention uses LSTM as a basic processing unit, and each spatial sample sequence corresponds to an LSTM time unit, thereby fully considering the spatial characteristics of PolSAR and realizing classification. And finally, on the basis of a probability map obtained by a deep recursive network, performing spatial texture analysis by using a CRF (cross domain gradient) to realize an optimal small sample decision classification process.

Description

Small sample polarization synthetic aperture radar image classification method based on deep recursive network

Technical Field

The invention relates to a small sample polarimetric synthetic aperture radar image classification method based on a deep recursive learning network, and belongs to the field of computer vision.

Background

Synthetic Aperture Radar (SAR) is a method for acquiring ground data all day long, all weather, with high resolution and high penetrability, and has high civil and commercial value. Interpreting SAR images can obtain much useful information, and therefore interpretation of SAR images is an important part of the practical application of SAR. Traditional machine learning and deep learning are two main methods for SAR image interpretation. SAR imaging is more stable than other sensors and is less susceptible to weather, light and other conditions. While another advantage of SAR is that a large amount of surface information data can be generated. It is difficult to manually process such a large amount of data.

The computer vision image processing technology based on the traditional machine learning and deep learning can well solve the problem of large data volume. The traditional machine learning method has a strict mathematical theory as a support, the requirement on computing resources is lower than that of a neural network, and meanwhile, the precision of classification and identification can meet the requirement to a certain extent. With the improvement of computer computing power, related processing methods based on neural networks are very colorful, and the classification and identification precision of the methods is often far higher than that of a machine learning method. However, the classification and identification method of the neural network depends on a large amount of training data, so that such a large amount of training data cannot be obtained in practical application and practical conditions, which requires a large amount of labor cost for collection and labeling. Too few training samples eventually lead to neural network overfitting, i.e. high classification or recognition accuracy on the training samples, but the results are very poor in testing and practical use. The deep recursive network depends on the excellent capability of capturing context information, and the complex network architecture can enable the network to have smaller requirements on the training sample size, so that the deep recursive network is more suitable for the classification task of small sample polarization SAR.

In addition, because the neural network has the defect of poor model interpretability, it is difficult to find an instructive direction during optimization. The sample characteristics used by the user are fused with characteristic samples obtained by the traditional unsupervised learning, and the deep recursive network is used for obtaining the relation between the context sequences and deep sample characteristics, so that a better classification effect is achieved. The combination not only exerts the coding capability of the neural network for fully utilizing computing resources, but also partially avoids the problem that the modeling result of the conventional neural network is difficult to explain, and leads the subsequent improvement and optimization to be more traceable. Under realistic conditions, not only a great deal of manpower is required to label the samples, but also the samples are likely to be partially lacking in the category. The random neighborhood pixel block combination (RNBP) proposed by the method can construct a plurality of unknown and possibly-occurring training samples on the basis of the existing training samples to enhance the training sample amount. Meanwhile, a multi-block combination (MB) model and a multi-block combination weight (MBW) model are provided to eliminate outliers possibly existing in test data by utilizing robustness analysis of spatial features of the test samples so as to enhance the test accuracy of the test samples.

Finally, for the classification of the remote sensing image, the probability that adjacent pixels belong to the same class is usually much higher than that of different classes, so that the texture analysis of the space spectrum can be performed on the probability graph model obtained by classification, and a better classification effect can be obtained. Conditional Random Field (CRF) processing enables texture analysis based on probability maps to achieve optimal classification decisions.

Disclosure of Invention

The invention mainly aims to provide a method for realizing image classification of a polarimetric synthetic aperture radar under the condition of a small sample based on a deep recursive network.

The invention provides the small sample classification aiming at the real data of the polarized SAR under the practical condition after the relevant direction of the small sample target identification is fully investigated. The method is different from the traditional deep learning method in that the prediction type of the output sample is different after the sample to be classified is directly input during classification, and the method enhances the feature expression of the sample by using the mechanism of the traditional unsupervised scattering model and various features obtained under different decomposition models. Meanwhile, in view of the similarity of adjacent samples, the invention provides a random neighborhood pixel block combination (RNBP) method for artificially synthesizing a new sample to supplement the incompleteness of a training sample, and simultaneously invents a multi-block combination (MB) and a multi-block combination weight (MBW) model to realize the robustness processing of a test sample. The invention creatively provides a method for extracting the spatial information of a small sample by using a deep recursive network, which enhances the capture capability of the network on the spatial information, and the deep recursive network mainly focuses on the extraction of deep features in the aspect of the learning of pixel features in view of the high-efficiency expression capability of unsupervised decomposition features. And the probability map obtained by the deep recursive network is subjected to spatial texture analysis through the CRF, so that the optimal classification effect is obtained.

The technical scheme of the invention specifically comprises the following technical contents:

1. and (5) constructing a decomposition model. For a small sample PolSAR, the polarization data may not be enough to express the intrinsic properties of the sample, so the sample characteristics can be enriched by an unsupervised characteristic scattering model to enhance the characteristic properties of different ground feature classes.

2. Small sample data extraction and sample enhancement. Because of the limitation of a small sample data set, the number and the size of training samples are possibly limited, so in order to enhance the expression capacity of the training sample set, besides rich characteristic expression of an unsupervised theory, the invention provides a random neighborhood pixel block combination (RNBP) method to artificially synthesize new samples so as to improve the expression capacity of the training set, and simultaneously, in order to reduce the data volume, the invention adopts an interval sampling method to select a small number of samples so as to increase the training efficiency of a network. The test sample is different from the training sample, the test sample is less restricted by the size of the sample, but because the test sample has more uncertainty, the robust analysis is carried out on the test sample, the interference of an outlier is reduced, and the identification capability of a network to the test sample is improved more effectively.

3. A deep recursive network. At present, the popular depth recursive network is more important than feature mining of shallow samples, the polarization attribute of one pixel is represented by a feature enhancement method of multiple scattering mechanisms, and the depth recursive network is used for mining the recursive relationship among neighborhood feature sample sequences, so that the neighborhood feature attribute and the deep polarization feature are mined, and more effective classification is achieved.

4. And (5) probability map spatial texture analysis. In a probability graph model obtained by a deep recursive network, the category attribute of each pixel point is expressed by a probability value, and the sum of the category attribute probabilities of a certain pixel is 1. In addition, although the deep recursive network can take into account the spatial relationship of the samples, due to the limitation of small samples, the network model is not perfect, and the texture information of the samples is not integrated, so that the optimal classification effect cannot be realized. On the basis of the probability map, the CRF is utilized to take the class probability of the pixel points as a unitary energy potential cluster and take the relation between the fields as a binary energy potential cluster to jointly construct the energy attribute of the sample, thereby achieving a better classification effect.

An SAR target recognition method based on an incomplete training set of a twin network comprises the following implementation flows:

the implementation flow of the method is as follows:

step 1, data preprocessing: performing feature extraction on four-channel polarization data (HH/HV/VH/VV) by using different scattering models, and combining the four-channel polarization data into a three-dimensional feature image:

1) for the full polarization data, firstly, a polarization coherent matrix T and a polarization covariance matrix C are solved;

2) denoising the T matrix and the C matrix by using an LEE filter;

3) finally, different decomposition characteristics are obtained by using different decomposition mechanisms, and the method comprises the following steps:

① two-component decomposition method based on kennaugh matrix K (e.g., Huynen, Barnes and Yang);

② decomposition covariance matrix C based on scattering model₃Or a coherence matrix T₃The method of (Freeman and Durden, Yamaguchi, Dong);

③ based on covariance matrix C₃Or a coherence matrix T₃Methods of feature vector or eigenvalue analysis (cloud, Holm, van Zyl, cloud and Pottier, the multiple-component learning model (MCSM));

④ is based on the method of coherent decomposition of the scattering matrix S (Krogager, Touzi).

Of the four classes of object (TD) algorithms, the first three classes are non-coherent decompositions, which are commonly used to process the coherence matrix [ T ], the covariance matrix [ C ] and the Kennaugh matrix [ K ], which are expressed as linear combinations. The coherent TD algorithm is used to process the scattering matrix S for representing polarization information of point objects and distributed objects. These decomposition features together construct a three-dimensional feature model of PolSAR (data size is m × n × 107, where m × n is the image size and 107 is the number of features). Wherein, 107 characteristic components are shown in the following table:

step 2, extracting training and testing data sets:

1) normalizing each feature dimension of the constructed three-dimensional feature model to eliminate dimensional interference among different features;

2) dividing data blocks in a pixel space dimension, taking each pixel as a center, selecting a 3 multiplied by 3 neighborhood block, copying a value of an adjacent row or column as a substitute value when the boundary exceeds a part, and then obtaining a corresponding pixel feature block with the size of 3 multiplied by 107;

3) selecting 5 pixel points far away from each other from 3 x 3 dimensions of the pixel block as a spatial feature sequence, combining the pixel points into a 5 x 107 data set, and using the data set as a standard network input data set;

4) randomly selecting no more than 0.5% of samples in the standard data set, wherein the number of each type of samples is no less than 6 samples to serve as training samples, and the others to serve as test samples.

Step 3, sample enhancement, which is used for enhancing the expression capacity of the sample:

1) because the probability of the same class between adjacent ground features is greater than that of the different classes, a random neighborhood pixel block combination (RNBP) method is provided to expand training samples, and the model architecture is shown in FIG. 1.

The sample center of the same type of label is extracted, and then the sequences are randomly and freely combined to obtain a new training sample, and meanwhile, the sample label is also known. The new training samples augment the training data set and are also used together to train the deep recursive network model. Specifically, assume that a certain class of training samples is n, where n is>5, the random free combination method of the synthetic samples can be used for obtaining

And (4) seed preparation.

2) The label of the test specimen is unknown and sample enhancement cannot be performed using the above method. However, the central sample value of the test sample is adjusted through the weight between the test sample and the surrounding samples, the singular value error of the test sample is restrained, the data robustness is enhanced, and the classification effect is improved. The input test sample center can be calculated by the following formula:

where m is 9 for a 3 × 3 pixel block, for the pixel if

Is an average value, then

If the weight is used to adjust the ratio of similar samples and reduce the influence of different samples, then

Wherein

The process flow is shown in fig. 2.

Step 4, constructing a deep recursive learning model, and training a neural network by using the training samples and the extended training samples:

1) constructing a deep recurrent neural network

The network model consists of two parts: a deep recursive network layer and a fully connected layer. In the deep recursive network, the elements of the network are LSTM architecture, and the basic architecture of the network is shown in fig. 3.

The LSTM unit includes an input gate, a forgetting gate, and an output gate. First, feature input x for the current time_tAnd the LSTM output value h input at the previous moment_t-1Respectively integrating and adding information through a full connection layer, inputting the information into an activation function sigma (-) and a tanh (-) to obtain input data i_tAnd I_tThe calculation method is as follows:

InputGate:

i_t＝σ(W_ix·x_t+W_ih·h_t-1+b_i) (1)

I_t＝tanh(W_Ix·x_t+W_Ih·h_t-1+b_I) (2)

in the formula, W_ix,W_ih,W_Ix,W_IhIs a full connection weight, b_i,b_IIs an offset. At the same time, xt and ht-1 are input into the information obtained by full-link integration to determine the state of the current information, so as to preserve the long-term cell state, namely:

ForgetGate:

f_t＝σ(W_fh·h_t-1+W_fx·x_t+b_f) (3)

in the formula, W_fh,W_fxIs a full connection weight, b_fIs an offset. Due to forgetting the output f of the door_tHas a value range of [0,1 ]]Therefore, the choice of long time state information can be determined together with the cell state value at time t-1, namely: f. of_t*c_t-1. The information obtained by the input gate jointly determines the acceptance or rejection of the information under the current time state, namely: i.e. i_t*I_t. Thus the LSTM cell state at the current time may be represented as:

MemoryCell:

c_t＝f_t*c_t-1+i_t*I_t(4)

finally, the value o is output_tIs formed by an input value parameter x_tAnd h_t-1Obtained by full concatenation post-accumulation, namely:

o_t＝σ(W_ox·x_t+W_oh·h_t-1+b_o) (5)

the output of the LSTM unit is determined by the state and output value of the LSTM unit, i.e.:

h_t＝o_t*tanh(c_t) (6)

in the formula, W_ox,W_ohIs a full connection weight, b_oFor bias, "+" denotes matrix addition, "· denotes matrix multiplication (full connection)," + "denotes dot multiplication. h is_tFor passing to the LSTM unit of the next time instant of the same layer and the LSTM unit of the same time instant of the next layer.

The classification model is built by adopting three layers of LSTM recursive networks and one layer of fully-connected network. In the three-layer LSTM network model: the node number of the first layer is 128, the network input data size is (5,107), and the output size is (5, 128); the node number set in the second layer is 64, the network input data size is (5, 128), and the output size is (5, 64); the node number set in the third layer is 32, the network input data size is (5, 64), and the output size is (1, 32). And the last layer of full connection layer is provided with the node number which is the same as the category number, and the activation function is sigmod to construct the classification probability. Specifically, for 11 classified PolSAR images, the network parameter cases are as follows:

network layer name	Network layer structure	Input size	Output size	Amount of ginseng
					Recursive network Layer (LSTM)	5 LSTM units, 128 nodes for each unit	(5，107)	(5，128)	120832
Recursive network Layer (LSTM)	5 LSTM units, each unit having 64 nodes	(5，128)	(5，64)	49408
					Recursive network Layer (LSTM)	5 LSTM units, 32 nodes per unit	(5，64)	(1，32)	12416
Full connection layer	11 full-connection nodes	(1，32)	(1，11)	363

The network input size is (5,107), where 107 corresponds to the feature sequence dimension of the artificial feature construction and 5 corresponds to 5 spatial pixels. The output size is (1, 11), and the belonged probability of the corresponding 11 types of samples is output.

2) Training and testing of deep recursive networks

After the deep recursive network construction is completed, training data samples with the size of (5,107) and the extended training samples are input into the deep recursive network model for training, the number of iterations is set to be 150, and the batch _ size is set to be 8. In the network, pixel feature sequences corresponding to 5 neighborhood pixel points are respectively used as sequence features corresponding to 5 time points of the LSTM, and recursive learning is performed in the network in a circulating mode.

And after the network training is finished, inputting the processed test sample into the deep recursion network, outputting a probability graph consisting of the probabilities of all categories through the network according to the corresponding output probability of the network, and performing the next texture segmentation.

Step 5, performing texture analysis by using the conditional random field to give a final image classification result

In order to further enhance the expression capability of the image on the special region of the image space spectrum feature and further improve the classification precision, after the depth recursive network model, the obtained probability map further realizes texture analysis by using a Conditional Random Field (CRF), thereby improving the classification precision. The basic architecture of a CRF network is shown in fig. 4.

In the figure e_jAnd k denotes a center node y_iAnd the domain node y_jThe green part represents the distance y_iDistant nodes, they are paired with y_iThe influence of (2) is small; red indicates the distance y_iCloser nodes, they are to y_iThe effect is large. Thus, the energy functional relationship between the center node and the domain is represented as:

in the formula, sigma_iξ_u(x_i) Representing a univariate potential function which combines shape, texture, color and position information of the image, Σ_i＜jξ_p(x_i,x_j) Is a potential energy function, and the calculation formula is as follows:

wherein u (x)_i,x_j) Is a class check function if x_i,x_jIf the difference is not the same, the value is 0,

can be expressed as:

in the formula (I), the compound is shown in the specification,

with the addition of the proximity location information,

the color information is added to the color information,

for eliminating small areas of non-smoothness in the vicinity. In the experiment, the probability information of each sample is used as a unigram, and the probability of a bigram is determined by the relationship between the adjacent samples. And finally outputting the finally obtained potential energy serving as final output to judge the final class of the points, so that an optimal classification result is obtained.

The overall architecture of the network is shown in fig. 5. The whole process comprises the following steps:

1) feature enhancement is performed using a scatter decomposition model.

2) Expanding a training sample and testing the sample after robustness analysis;

3) training a deep recursion network model;

4) inputting the test data into a network to obtain a probability chart;

5) and performing space spectrum analysis by using the CRF to obtain a final classification result.

Drawings

FIG. 1 shows a random neighborhood pixel block combination model.

FIG. 2. test sample pretreatment model.

FIG. 3 is an LSTM cell structure.

FIG. 4.CRF architecture.

FIG. 5 is a diagram of the overall architecture of the network

Detailed Description

The basic flow of the small sample polarized synthetic aperture radar image classification method under the deep recursive network is shown in fig. 5, and the method specifically comprises the following steps:

1) the PolSAR polarization data comprises polarization data of four channels of HH, HV, VH and VV, 107 artificial features are constructed on the basis of various target decomposition and decomposition on the basis of the polarization data, the artificial features can express basic attributes of the scattering characteristics of the ground objects, and the 107 artificial features are combined into three-dimensional data to form basic polarization feature data.

2) The polarization data can be divided into labeled samples and non-labeled samples, and the invention is a model constructed under small samples, so that the training sample is 0.5 percent of the total sample, and the rest is a test sample. In order to embody the constraint of small sample size, only 3 × 3 neighborhood samples are used as training sample samples in the experiment, and in order to reduce the data volume, only 5 pixels are sampled at intervals to realize the construction of the training sample. On the other hand, in order to enhance the characterization capability of the training samples, the training samples are synthesized by an RNBP method, and the richness of the samples can be increased by the synthesized new training samples, so that the robustness of the network is improved. The test samples are subjected to robustness analysis to eliminate outliers of the samples, so that the recognition capability of the network on the test samples is improved.

3) Inputting a spatial feature sequence with a sample size of (5,107) into a three-layer recursive network model to realize a training process, wherein spatial pixels correspond to the time dimension of the recursive network, the recursive network uses a standard LSTM model, and all connection layers in each layer of LSTM unit are set to be 128, 64 and 32; the number of nodes of the last layer of full connection layer is set as the number of categories. In the network training process, the network step size is set to 8, and the number of cyclic training is set to 150. The last layer uses Sigmoid function to realize probability output. And finally, converting the probability into a probability map and transmitting the probability map to the CRF network.

4) And after obtaining the probability map, performing texture analysis by using a CRF (gradient frequency model), wherein the iteration number is set to be 5, the space size is set to be 5, and the proportion of the binary potential clusters is set to be 20 in the CRF model. The original probability of a single point of the probability map is set as the energy of a univariate potential group, the texture relation between pixels is set as a binary potential group, the sum of the potential groups is the energy of a certain point, and the category attribute of the sample is judged according to the energy.

Examples

The small sample PolSAR image classification method based on the deep recursive learning network is characterized by comprising the following steps: the implementation flow of the method is as follows,

the invention mainly aims at the classification research of polarized Synthetic Aperture Radar (PolSAR) imaging, and is a new research hotspot of SAR (Synthetic Aperture Radar) remote sensing neighborhood. The radar polarization is studied on the complete vector characteristic of the polarized electromagnetic wave in the propagation process. Compared with a single-polarization SAR system (SAR for short), the multi-polarization SAR has all the traditional advantages of single polarization, records the backscattering echo signals (amplitude and phase information) of the polarization state (HH/HV/VH/VV) in the ground object 4 by utilizing the vector characteristics of electromagnetic waves, completely contains the state information of the electromagnetic waves and radar targets, can provide more target information and distinguishing characteristics, records the information of radar target geometry, physical characteristics and the like more comprehensively, and becomes a research frontier and a hotspot of radar remote sensing. For example, homopolar (HH/VV) SAR images reflect wavelength dependence on surface roughness, whereas cross-polar (HV/VH) SAR images characterize volume scattering. In addition, the L-band VV polarized SAR image shows scattering characteristics of the dense vegetation of the forest, the C-band HV polarized SAR image reflects the volume scattering of crops, and the L, C-band HH polarized SAR image mainly reflects surface scattering characteristics and the like. Therefore, PolSAR has greater application potential in terrain classification, target detection and information extraction. This patent is directed to a classification algorithm designed for fully polarimetric SAR images. The method comprises the following basic steps:

step 1, data preprocessing: carrying out feature extraction on the four-channel polarization data by using different scattering models, and combining the four-channel polarization data into a three-dimensional feature image:

5) for the full polarization data, firstly, a polarization coherent matrix T and a polarization covariance matrix C are solved;

6) denoising the T matrix and the C matrix by using an LEE filter;

7) finally, different decomposition characteristics are obtained by using different decomposition mechanisms, and the method comprises the following steps:

① two-component decomposition method based on kennaugh matrix K;

② decomposition covariance matrix C based on scattering model₃Or a coherence matrix T₃The method of (1);

③ based on covariance matrix C₃Or a coherence matrix T₃A method of feature vector or feature value analysis;

④ is based on a method of coherent decomposition of the scattering matrix S.

Of the four classes of object (TD) algorithms, the first three classes are non-coherent decompositions, which are commonly used to process the coherence matrix [ T ], the covariance matrix [ C ] and the Kennaugh matrix [ K ], which are expressed as linear combinations. The coherent TD algorithm is used to process the scattering matrix S for representing polarization information of point objects and distributed objects. The decomposition features together construct a three-dimensional feature model of the PolSAR;

this step comprises

Inputting: polarization data of four channels for constructing three-dimensional features of polarized image

And (3) outputting: inputting the constructed three-dimensional features into the second step for data extraction

Step 2, extracting training and testing data sets:

4) normalizing each feature dimension of the constructed three-dimensional feature model to eliminate dimensional interference among different features;

5) dividing data blocks in a pixel space dimension, taking each pixel as a center, selecting a neighborhood block, copying a value of an adjacent row or column as a substitute value when the boundary exceeds a part, and then obtaining a corresponding pixel characteristic block;

6) selecting pixel points from the dimension of a pixel block as a spatial characteristic sequence, and combining a data set as a standard network input data set;

8) randomly selecting no more than 0.5% of samples in the standard data set, wherein the number of each type of samples is no less than 6 samples to serve as training samples, and the others to serve as test samples.

This step comprises

Inputting: three-dimensional polarization characteristic image constructed in first step

And (3) outputting: a normalized set of training samples and test samples, wherein: training samples are used for constructing an enhanced sample in the third step and network training in the fourth step; and the test sample is used for constructing a robustness test sample set in the third step.

Step 3, sample enhancement, which is used for enhancing the expression capacity of the sample:

3) because the probability of the same type between adjacent ground objects is greater than that of the different types, a random neighborhood pixel block combination method is provided to expand the training sample. The sample center of the same type of label is extracted, and then the sequences are randomly and freely combined to obtain a new training sample, and meanwhile, the sample label is also known.

4) The input test sample center is calculated by the following formula:

for if

To test the mean value of the sample, then

If the weight is used to adjust the similarityThe ratio of the samples reduces the influence of different samples

Wherein

i denotes the sequence label of the neighborhood samples, j denotes the sequence label of the feature, and m denotes the total number of the neighborhood pixels.

This step comprises

Inputting: three-dimensional polarization characteristic image constructed in first step

And (3) outputting: a normalized set of training samples and test samples, wherein: training samples are used for constructing an enhanced sample in the third step and network training in the fourth step; the test samples are used to construct a robustness test sample set.

Step 4, constructing a deep recursive learning model, and training a neural network by using the training samples and the extended training samples:

3) constructing a deep recurrent neural network

The network model consists of two parts: a deep recursive network layer and a fully connected layer. Wherein, in the deep recursive network, the network unit is the LSTM architecture.

First, feature input x for the current time_tAnd the LSTM output value h input at the previous moment_t-1Respectively integrating and adding information through a full connection layer, inputting the information into an activation function sigma (-) and a tanh (-) to obtain input data i_tAnd I_tThe calculation method is as follows:

input gate

i_t＝σ(W_ix·x_t+W_ih·h_t-1+b_i) (1)

I_t＝tanh(W_Ix·x_t+W_Ih·h_t-1+b_I) (2)

In the formula, W_ix,W_ih,W_Ix,W_IhIs totally connected withIs weighted, wherein W_ixIs the full connection weight, W, between the input sample x and the output i_ihIs the full connection weight between the output characteristic h and the output i of the unit at the previous time point, W_IxIs the full connection weight, W, between the input sample x and the output I_IhThe full connection weight between the unit output characteristic h and the output i of the last time point; b_i,b_IIs an offset, wherein b_iFor full connection bias between input sample, last time unit output signature and output i, b_IThe full connection bias between the input sample, the last time unit output signature and output I. At the same time, x_tAnd h_t-1Inputting information obtained by full-link integration to determine the state of the current information, so as to preserve the long-term cell state, namely:

forgetting door

f_t＝σ(W_fh·h_t-1+W_fx·x_t+b_f) (3)

In the formula, W_fh,W_fxIs a full connection weight, wherein W_fxFor input sample x and output f_tAll connection weights between, W_fhOutputting the characteristic h and the output f for the unit at the previous time point_tAll connection weights between, b_fBeing their common bias. Due to forgetting the output f of the door_tHas a value range of [0,1 ]]Therefore, the choice of long time status information is determined together with the cell status value at time t-1, namely: f. of_t*c_t-1. The information obtained by the input gate jointly determines the acceptance or rejection of the state information at the current time, namely: i.e. i_t*I_t. Thus the LSTM architecture state at the current time is represented as:

MemoryCell:

c_t＝f_t*c_t-1+i_t*I_t(4)

finally, the value o is output_tIs formed by an input value parameter x_tAnd h_t-1Obtained by full concatenation post-accumulation, namely:

o_t＝σ(W_ox·x_t+W_oh·h_t-1+b_o) (5)

the output of the LSTM unit is determined by the state and output value of the LSTM unit, i.e.:

h_t＝o_t*tanh(c_t) (6)

in the formula, W_ox,W_ohIs a full connection weight, wherein W_oxFor input sample x and output o_tAll connection weights between, W_ohOutputting the characteristic h and the output o for the unit at the previous time point_tAll connection weights between, b_oFor their common bias, "+" denotes matrix addition, "· denotes matrix multiplication and" · "denotes dot multiplication. h is_tFor passing to the LSTM unit of the next time instant of the same layer and the LSTM unit of the same time instant of the next layer.

This step comprises

Inputting: training sample set, extended training sample set, and smoothed test sample set

And (3) outputting: probability map output after passing through LSTM network

4) Training and testing of deep recursive networks

After the construction of the deep recursive network is completed, the training data sample and the extended training sample are input into a deep recursive network model for training; and (4) respectively taking the pixel characteristic sequences corresponding to the neighborhood pixel points as the sequence characteristics corresponding to the time points of the LSTM, and circularly and recursively learning in the network.

And after the network training is finished, inputting the processed test sample into the deep recursion network, outputting a probability graph consisting of the probabilities of all categories through the network according to the corresponding output probability of the network, and performing the next texture segmentation.

Step 5, performing texture analysis by using the conditional random field to give a final image classification result

After the deep recursion network model, the obtained probability graph further realizes texture analysis by using a conditional random field, so that the classification precision is improved.

e_jAnd k denotes a center node y_iAnd the domain node y_jEnergy relationship between, centerNode y_iAnd the domain node y_jThe energy functional relationship between e (x) is expressed as:

in the formula (E)_iξ_u(x_i) Representing a univariate potential function, which combines shape, texture, color and position information of the image, Σ_i＜jξ_p(x_i,x_j) Is a potential energy function, and the calculation formula is as follows:

wherein K^(m)(f_i,f_j) A pixel vector f representing pixels i, j for a Gaussian kernel_iAnd f_jPotential feature space relationships between; u (x)_i,x_j) Is a class check function if x_i,x_jIf the difference is not the same, the value is 0,

expressed as:

in the formula (I), the compound is shown in the specification,

with the addition of the proximity location information,

the color information is added to the color information,

for eliminating small areas of non-smoothness in the vicinity. The probability information of each sample is used as a univariate potential group, and the probability of the binary potential group is determined by the relation between the neighborhood samples. Finally, the energy of the potential group is output as the final output to judge the final pointAnd the classification is carried out, so that the optimal classification result is obtained.

This step comprises

Inputting: probability map output after passing through LSTM network

And (3) outputting: classification label graph of test samples

The overall architecture of the network is therefore: data preprocessing, training sample data expansion, test enhancement, LSTM training and testing, CRF probability map segmentation and classification final result label map

The whole network architecture process comprises the following steps:

6) feature enhancement is performed using a scatter decomposition model.

7) Expanding a training sample and testing the sample after robustness analysis;

8) training a deep recursion network model;

9) inputting the test data into a network to obtain a probability chart;

10) and performing space spectrum analysis by using the CRF to obtain a final classification result.

Claims (2)

1. The small sample PolSAR image classification method based on the deep recursive learning network is characterized by comprising the following steps: the implementation flow of the method is as follows,

step 1, data preprocessing: carrying out feature extraction on the four-channel polarization data by using different scattering models, and combining the four-channel polarization data into a three-dimensional feature image:

1) for the full polarization data, firstly, a polarization coherent matrix T and a polarization covariance matrix C are solved;

2) denoising the T matrix and the C matrix by using an LEE filter;

3) finally, different decomposition characteristics are obtained by using different decomposition mechanisms;

step 2, extracting training and testing data sets:

1) normalizing each feature dimension of the constructed three-dimensional feature model to eliminate dimensional interference among different features;

2) dividing data blocks in a pixel space dimension, taking each pixel as a center, selecting a neighborhood block, copying a value of an adjacent row or column as a substitute value when the boundary exceeds a part, and then obtaining a corresponding pixel characteristic block;

3) selecting pixel points from the dimension of a pixel block as a spatial characteristic sequence, and combining a data set as a standard network input data set;

4) randomly selecting not more than 0.5 percent and not less than 6 samples in each class in the standard data set as training samples, and taking the others as test samples;

step 3, sample enhancement, which is used for enhancing the expression capacity of the sample:

because the probability of the same type between adjacent ground objects is greater than that of the different types, a random neighborhood pixel block combination method is provided to expand the training sample; extracting the sample centers of the same type of labels, then randomly and freely combining the sequences to obtain a new training sample, wherein the sample labels are also known;

step 4, constructing a deep recursive learning model, and training a neural network by using the training samples and the extended training samples:

1) constructing a deep recurrent neural network

The network model consists of two parts: a deep recursion network layer and a full connection layer; in the deep recursive network, the network unit is an LSTM architecture;

2) training and testing of deep recursive networks

After the construction of the deep recursive network is completed, the training data sample and the extended training sample are input into a deep recursive network model for training; pixel feature sequences corresponding to the neighborhood pixel points are respectively used as sequence features corresponding to the time points of the LSTM, and the cyclic recursive learning is carried out in the network;

after the network training is finished, inputting the processed test sample into a deep recursion network, outputting a probability graph consisting of the probabilities of all categories through the network according to the corresponding output probability of the network, and performing the next texture segmentation;

step 5, performing texture analysis by using the conditional random field to give a final image classification result

After the deep recursive network model, the obtained probability graph further realizes texture analysis by using a conditional random field, so that the classification precision is improved; and finally outputting the finally obtained potential energy serving as final output to judge the final class of the points, thereby obtaining an optimal classification result.

2. The small-sample PolSAR image classification method based on deep recursive learning network according to claim 1, characterized in that: the whole network architecture process comprises the following steps:

1) performing feature enhancement by using a scattering decomposition model;

2) expanding a training sample and testing the sample after robustness analysis;

3) training a deep recursion network model;

4) inputting the test data into a network to obtain a probability chart;

5) and performing space spectrum analysis by using the CRF to obtain a final classification result.

Images