CN114037896B - PolSAR ground object fine classification method based on multi-index convolution self-encoder - Google Patents
PolSAR ground object fine classification method based on multi-index convolution self-encoder Download PDFInfo
- Publication number
- CN114037896B CN114037896B CN202111319995.4A CN202111319995A CN114037896B CN 114037896 B CN114037896 B CN 114037896B CN 202111319995 A CN202111319995 A CN 202111319995A CN 114037896 B CN114037896 B CN 114037896B
- Authority
- CN
- China
- Prior art keywords
- encoder
- polsar
- index
- matrix
- polarization
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 17
- 238000011176 pooling Methods 0.000 claims abstract description 12
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 230000010287 polarization Effects 0.000 claims description 46
- 239000011159 matrix material Substances 0.000 claims description 32
- 239000013598 vector Substances 0.000 claims description 27
- 238000010586 diagram Methods 0.000 claims description 17
- 238000012545 processing Methods 0.000 claims description 5
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 238000005070 sampling Methods 0.000 claims description 3
- 230000004931 aggregating effect Effects 0.000 claims description 2
- 238000013527 convolutional neural network Methods 0.000 description 15
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000002474 experimental method Methods 0.000 description 5
- 230000006870 function Effects 0.000 description 5
- 238000007635 classification algorithm Methods 0.000 description 3
- 101000822805 Naja atra Cytotoxin A5 Proteins 0.000 description 2
- 101000822803 Naja kaouthia Cytotoxin homolog Proteins 0.000 description 2
- 101000783567 Naja naja Cytotoxin 1 Proteins 0.000 description 2
- 101000822819 Naja naja Cytotoxin-like basic protein Proteins 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000012512 characterization method Methods 0.000 description 2
- 230000000295 complement effect Effects 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 239000006185 dispersion Substances 0.000 description 2
- 230000005284 excitation Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 208000002079 spinocerebellar ataxia with epilepsy Diseases 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a PolSAR ground object fine classification method based on a multi-index convolution self-encoder, which comprises the following steps: 1. preprocessing and slicing the PolSAR data; 2. extracting a feature map of the PolSAR image through convolution and pooling operation, inputting the obtained feature map into an MI-SE module to calculate multiple indexes, and obtaining the weight of each dimension of the input feature map; 3. weighting the importance of the feature map according to the weight distributed to the input feature map to obtain a final feature; 4. inputting the final characteristics into a classifier to obtain a predicted result, and comparing the predicted result with a real result to complete the training process of the model; 5. and inputting the PolSAR image to be classified into a trained multi-index convolution self-encoder model to obtain a classification result. The method can improve the integrality and the fineness of the feature representation of the PolSAR, obtain higher classification precision and classification efficiency, and have better engineering application value.
Description
Technical Field
The invention relates to the technical field of PolSAR ground object fine classification, in particular to a PolSAR ground object fine classification method based on a multi-index convolution self-encoder.
Background
Polarized synthetic aperture radar (Polarimetric synthetic aperture radar, polSAR) as a high resolution imaging sensor can provide detailed ground object information. Compared with an optical image, the PolSAR can obtain more effective and comprehensive ground feature information under all-weather and full-time conditions. Therefore, classification of ground features by using PolSAR data can effectively improve classification accuracy.
The conventional PolSAR terrain classification method can be summarized into an unsupervised classification without training sample tags and a supervised classification with training sample tags. Although a lot of human resources are not consumed in the non-supervision classification process, the operation time is long, and the classification precision is low. The supervised classification based on machine learning mainly comprises two processes of feature extraction and classification, the performance of which is mainly limited by the representation capability of the features, and the classification performance is extremely reduced when the extracted features are difficult to highlight the differences among different feature classes. Supervised classification algorithms based on deep learning have been proposed in recent years, but as the complexity of deep learning networks increases, the training process generally requires a large number of samples and a sufficient amount of time. However the problem of sample lack of labels is just a short plate in the PolSAR clutter classification problem. To address this problem, convolutional self-encoders (convolutional auto-encoders, CAE) are of greater interest because CAE can train network parameters through the decoder under unsupervised conditions, thereby reducing the amount of samples required for final classification. However, because the network depth is low, the extracted depth features cannot fully represent the information of the target and are also easily affected by the interference information, so that the classification result is poor.
There are many existing methods for improving the problems of the CAE network, which improve the conventional CAE network in terms of sample data size, network depth, loss function, etc. For example: stacked convolutional self-encoders (stacked convolutional Autoencoder, SCAE) deepen the entire network by stacking multiple CAE networks and learn network parameters using hierarchical training to obtain deeper ground feature information; there is also an algorithm that uses the wishart distance as the reconstruction loss of the CAE network. However, these improved algorithms merely stack adjacent features into different feature map channels, which is not complete and fine enough for different classes of feature representation.
Disclosure of Invention
The invention aims to solve the defects in the prior art, and provides a PolSAR ground object fine classification method based on a multi-index convolution self-encoder, so that the completeness and the fineness of the feature characterization of the PolSAR ground object can be improved, and higher classification precision and classification efficiency are obtained.
In order to achieve the aim of the invention, the invention adopts the following technical scheme:
the invention relates to a PolSAR ground object fine classification method based on a multi-index convolution self-encoder, which is characterized by comprising the following steps:
step 1: preprocessing and clipping operation of PolSAR data:
step 1.1: the scattering characteristic of any minimum resolution unit in the PolSAR data is represented by using a polarized scattering matrix S; according to the reciprocity principle and Pauli decomposition principle, converting any one polarized scattering matrix S in PolSAR data into a vectorWherein S is HH Representing the complex scattering coefficient when the transmitting polarization mode is horizontal polarization and the receiving polarization mode is horizontal polarization, S VV Representing the complex scattering coefficient when the transmitting polarization mode is vertical polarization and the receiving polarization mode is vertical polarization, S HV Representing the complex scattering coefficient when the transmitting polarization mode is vertical polarization and the receiving polarization mode is horizontal polarization;
step 1.2: based on the loss K, a polarization coherence matrix T is obtained using equation (1):
in the formula (1) (. Cndot. H Represents conjugate transpose, T ij Elements representing the ith row and jth column of the polarization coherence matrix T; and i, j=1, 2,3;
step 1.3: extracting a six-dimensional eigenvector f= [ a, B, C, D, E, F ] from the polarization coherence matrix T using formula (2):
in equation (2), span represents the total scattered power of all polarized channels, and span=t 11 +T 22 +T 33 A represents the decibel form of the total scattered power Span, B and C represent the 2 nd row and 2 nd column elements T, respectively 22 And line 3, column 3 element T 33 D, E and F represent three relative correlation coefficients;
step 1.4: normalizing the six-dimensional feature vector F to obtain a normalized six-dimensional feature vector F', so that preprocessed PolSAR data are formed by the normalized six-dimensional feature vectors of all minimum resolution units;
step 1.5: cutting the preprocessed PolSAR data into slices with the size of L multiplied by 6, thereby obtaining a slice set { s } of the PolSAR data 1 ,...,s n ,...,s N -wherein s n Represents an nth slice, N represents the total number of slices; n is E [1, N]；
Step 2: constructing a multi-index convolutional self-encoder based classification network, comprising: an encoder, a decoder, and a classifier; and aggregating the slices { s } 1 ,...,s n ,...,s N -input into the classification network;
step 2.1: the encoder consists of m convolution layers, m pooling layers, m MI-SE modules, m weighting modules and a full connection layer;
the nth slice s n Inputting the characteristic diagram U into the encoder, and obtaining a characteristic diagram U with a scale of H multiplied by W multiplied by D after the treatment of a first convolution layer and a pooling layer, wherein H and W respectively represent the height and the width of the characteristic diagram U, and D represents the channel number of the characteristic diagram U;
the feature map U is input into a first MI-SE module, and a multi-index is calculated by using a formula (3), and the method comprises the following steps: first moment S 1 Second order central moment S 2 And coefficient of variation S 3 ：
In the formula (3), U (a, b, i) represents all elements of the a-th row and the b-th column of the characteristic diagram U;
the first moment S 1 Second order central moment S 2 And coefficient of variation S 3 Splicing the matrix into a multi-index matrix V with the size of 3 xD, and processing the multi-index matrix V by a convolution layer and a full connection layer in the first MI-SE module to obtain a weight matrix Y with the size of 1 xD;
the first weighting module multiplies the weight matrix Y with the feature map U to obtain a weighted feature map U' with the size of H multiplied by W multiplied by D;
the weighted feature map U' sequentially passes through the rest m-1 convolution layers, the pooling layer, the MI-SE module and the weighting module of the encoder, and finally the output features pass through a full connection layer to obtain the feature vector alpha output by the encoder;
step 2.2: the decoder is formed by sequentially connecting k convolution layers and k up-sampling layers in a staggered manner;
the feature vector alpha is input into a decoder to obtain an nth slice s n Is a reconstructed slice of (a)Re-use of (4) to establish a counter-propagating loss function L MSE ：
In equation (4), MSE represents the minimum mean square error; p is the number of elements per slice,for the nth slice s n The c-th element in the slice reconstructed after decoder,>for the nth slice s n C element of (a);
the encoder and decoder are trained using back propagation when the loss function L MSE When the minimum value is taken, training is stopped, and the optimal feature vector alpha output by the trained encoder is obtained * ；
Step 2.3: the classifier consists of t full-connection layers and one softmax layer;
the optimal feature vector alpha * In the input classifier, firstly, reducing the dimension through t full-connection layers, then inputting the feature vector after the dimension reduction into a softmax layer, and calculating an nth slice s n The center element of (a) corresponds to the posterior probability of each category, and the category corresponding to the maximum posterior probability is selected as the nth slice s n Is a predictive category of (2);
step 2.4: obtaining prediction categories of N slices according to the modes of the steps 2.1-2.3, comparing the prediction categories with corresponding real categories, and optimizing the classifier through back propagation, so that training of the encoder, the decoder and the classifier is completed, and a trained multi-index convolution self-encoder model is obtained and used for classifying PolSAR data to be classified.
Compared with the prior art, the invention has the beneficial effects that:
1. the PolSAR ground object fine classification method based on the multi-index convolution self-encoder effectively improves the integrity and the fineness of feature representation in the PolSAR ground object classification process, reduces the influence of individual extremum of images on the integral feature, and realizes the PolSAR ground object fine classification.
2. The invention provides the MI-SE module, solves the problem that in the traditional CAE network, the result of the feature map output by the convolution layer is simply and averagely input to the next layer, and performs effective weighting operation on the output of the convolution layer by utilizing the information in the feature map, thereby better extracting and characterizing the feature of the ground object. The characteristics carrying different information are weighted with different weights, and the more complete and fine characteristic information is obtained.
3. The invention provides three indexes for measuring the importance of different features, namely a first moment, a second central moment and a variation coefficient. The first moment can reflect the average value of the features, the second-order center distance can describe the fluctuation in each feature, the content of the edge and texture information can be reflected to a certain extent, and the variation coefficient can eliminate the influence of the average value on the dispersion, so that the fluctuation is reflected better. The three indexes complement each other, and each feature is weighted by different angles.
4. The convolution result is weighted through the three indexes, so that the difference in the ground object categories is obviously reduced, the difference between the ground object categories is increased, and in order to illustrate the capability of improving the separability between the different categories by the three different indexes, the t-distribution random neighborhood embedding algorithm is used for visualizing the middle characteristic diagram, and the result after 100 times of network iteration is shown in figure 3. So that the final softmax can more easily separate each category of ground objects.
Drawings
FIG. 1 is a MI-SE block structure of the invention;
FIG. 2 is a flow chart of the PolSAR image ground object target classification method of the present invention;
FIG. 3 is an illustration of a visual intermediate feature map of the present invention employing a t-distributed random neighborhood embedding algorithm;
FIG. 4 is an experimental data set of the invention San Francisco;
FIG. 5 is a graph of the classification results of the method of the present invention and other methods on the San Francisco dataset.
Detailed Description
In this embodiment, a multi-index compression and excitation (MI-SE) module shown in fig. 1 is added to a conventional CAE network, which proposes a multi-index convolutional self-encoder based classification network shown in fig. 2, so that the detail features and depth features of the PolSAR ground object target can be effectively considered, different attention degrees of different features can be realized when the model is guided to train, the completeness and fineness of feature characterization of the PolSAR ground object target are ensured, the gaps in various ground objects are reduced, and the gaps among various ground objects are increased, thereby obtaining a better classification result. Specifically, the PolSAR terrain fine classification method based on the multi-index convolution self-encoder comprises the following steps:
step 1: preprocessing and clipping operation of PolSAR data:
step 1.1: the scattering characteristic of any minimum resolution unit in the PolSAR data is represented by using a polarized scattering matrix S; according to the reciprocity principle and Pauli decomposition principle, converting any one polarized scattering matrix S in PolSAR data into a vectorWherein S is HH Representing the complex scattering coefficient when the transmitting polarization mode is horizontal polarization and the receiving polarization mode is horizontal polarization, S VV Representing the complex scattering coefficient when the transmitting polarization mode is vertical polarization and the receiving polarization mode is vertical polarization, S HV Representing the complex scattering coefficient when the transmitting polarization mode is vertical polarization and the receiving polarization mode is horizontal polarization;
step 1.2: based on the loss K, a polarization coherence matrix T is obtained using equation (1):
in the formula (1) (. Cndot. H Represents conjugate transpose, T ij Elements representing the ith row and jth column of the polarization coherence matrix T; and i, j=1, 2,3;
step 1.3: extracting six-dimensional eigenvectors f= [ a, B, C, D, E, F ] from the polarization coherence matrix T using equation (2):
in equation (2), span represents the total scattered power of all polarized channels, and span=t 11 +T 22 +T 33 A represents the decibel form of the total scattered power Span, B and C represent the 2 nd row and 2 nd column elements T, respectively 22 And line 3, column 3 element T 33 D, E and F represent three relative correlation coefficients;
step 1.4: normalizing the six-dimensional feature vector F to obtain a normalized six-dimensional feature vector F', so that preprocessed PolSAR data are formed by the normalized six-dimensional feature vectors of all minimum resolution units;
step 1.5: cutting the preprocessed PolSAR data into slices with the size of L multiplied by 6, thereby obtaining a slice set { s } of the PolSAR data 1 ,...,s n ,...,s N -wherein s n Represents the nth slice, N represents the total number of slices, n.epsilon.1, N]In this embodiment, the data is cut into slices of size l×l×6=15×15×6, the total number of slices being 0.5% of the amount of tag data in the data set;
step 2: constructing a multi-index convolutional self-encoder based classification network, comprising: an encoder, a decoder, and a classifier; and to aggregate the slices { s } 1 ,...,s n ,...,s N The input into the classification network,
step 2.1: the encoder consists of m convolution layers, m pooling layers, m MI-SE modules, m weighting modules and a full connection layer, in this embodiment, two convolution layers, two pooling layers, two MI-SE modules, two weighting modules and a full connection layer are constructed;
nth slice s n Inputting the characteristic diagram U into an encoder, and obtaining a characteristic diagram U with a scale of H multiplied by W multiplied by D after processing of a first convolution layer and a pooling layer, wherein H and W respectively represent the height and width of the characteristic diagram U, and D represents the channel number of the characteristic diagram U;
the profile U is input into a first MI-SE module and a multi-index is calculated using equation (3), comprising: first moment S 1 Second order central moment S 2 And coefficient of variation S 3 ：
In the formula (3), U (a, b, i) represents all elements of the a-th row and the b-th column of the characteristic diagram U;
will first moment S 1 Second order central moment S 2 And coefficient of variation S 3 Splicing the matrix into a multi-index matrix V with the size of 3 xD, and processing the multi-index matrix V by a convolution layer and a full connection layer in a first MI-SE module to obtain a weight matrix Y with the size of 1 xD;
the first weighting module multiplies the weight matrix Y with the feature map U to obtain a weighted feature map U' with the size of H multiplied by W multiplied by D;
the weighted characteristic diagram U' sequentially passes through the rest m-1 convolution layers, the pooling layer, the MI-SE module and the weighting module of the encoder, and finally the output characteristic passes through a full connection layer to obtain the characteristic vector alpha output by the encoder;
step 2.2: the decoder is formed by sequentially connecting k convolution layers and k up-sampling layers in a staggered manner;
the feature vector alpha is input into the decoder to obtain the nth slice s n Is a reconstructed slice of (a)Re-use of (4) to establish a counter-propagating loss function L MSE ：
In equation (4), MSE represents the minimum mean square error; p is the number of elements per slice,for the nth slice s n The c-th element in the slice reconstructed after decoder,>for the nth slice s n C element of (a);
encoder and decoder training with back propagation when the loss function L MSE When the minimum value is taken, training is stopped, and the optimal feature vector alpha output by the trained encoder is obtained * In this embodiment, three convolution layers and are provided in the decoderThree upsampling layers;
step 2.3: the classifier consists of t full-connection layers and one softmax layer;
optimal feature vector alpha * In the input classifier, firstly, reducing the dimension through t full-connection layers, then inputting the feature vector after the dimension reduction into a softmax layer, and calculating an nth slice s n The center element of (a) corresponds to the posterior probability of each category, and the category corresponding to the maximum posterior probability is selected as the nth slice s n In this embodiment, two fully connected layers and one softmax layer are built in the classifier;
step 2.4: obtaining prediction categories of N slices according to the modes of the steps 2.1-2.3, comparing the prediction categories with corresponding real categories, optimizing the classifier through back propagation, and training the encoder, the decoder and the classifier, so that a trained multi-index convolution self-encoder model is obtained and used for classifying PolSAR data to be classified.
So far, the PolSAR terrain fine classification method based on the multi-index convolution self-encoder is basically completed.
The effectiveness of the invention is further illustrated by the San Francisco dataset experiments below.
San Francisco dataset PolSAR image target classification experiment:
1. experiment setting:
the experimental dataset was an L-band full polarized synthetic aperture radar image collected from san francisco by the air synthetic aperture radar platform of the united states space agency jet propulsion laboratory (Jet Propulsion Laboratory, JPL). The data set has a size of 1300 x 1200 and comprises five ground objects of low-density cities, high-density cities, developed cities, water bodies and vegetation. In order to input the data set into the multi-index convolutional self-encoder model, the size of the data slice was set to 15×15 in the experiment, and furthermore, the number of training samples in the data set was randomly selected to be 0.5% of the total samples. The dataset is shown in fig. 3 and the final classification result is shown in fig. 4. Fig. 3 (a) shows original data. Fig. 3 (b) shows the case where data can be separated when conventional CAE is used. Fig. 3 (c) shows the case where data can be separated after CAE is added to the first moment. Fig. 3 (d) shows the case where data can be separated after CAE has been added to the second order central moment. FIG. 3 (e) shows how the data can be separated after CAE is added to the coefficient of variation. FIG. 3 (f) shows the case where data can be separated when the method of the present invention is employed;
2. analysis of results:
the experiment in this example uses the Overall Accuracy (OA) and Kappa coefficient (Kappa coefficient) to quantitatively analyze the performance of the proposed method. To illustrate the superiority of the proposed methods, several common PolSAR image ground object classification methods are selected for comparison, and most of the methods are based on improvement of convolutional neural networks (convolutional neural networks, CNN), wherein the methods comprise methods based on complete local binary pattern fusion of CNN (completed local binary patterns feature integrated CNN, CLBP-CNN), combination of CNN and Markov random fields (combination of CNN and Markov random field, MRF-CNN), polarization compression and excitation networks (polarimetric squeeze-and-excitation network, PSE-NET), complex CNN (complex-valued CNN, CV-CNN), VGG-Net, residual neural networks (residual neural network, res-Net) and CAE. Batch processing size is set to be 64 in the training process of all methods, iteration times are 100, learning rate is 0.001, and optimization algorithm is Adam algorithm. The comparison results are shown in Table 1, wherein:
t and N in equation (5) represent the number of categories and the total number of samples in the classification task, c ii Corresponding to diagonal elements in the confusion matrix. C in formula (6) ij Corresponding to all elements in the confusion matrix.
TABLE 1 PolSAR terrain fine classification method Performance evaluation index based on Multi-index convolutional self-encoder
Analysis is performed in conjunction with fig. 5 and table 1, wherein (a) in fig. 5 represents an original label map, (b) in fig. 5-fig. 5 (h) represents a classification result of the comparison algorithm, and (i) in fig. 5 represents a classification result of the present invention; compared with other common PolSAR ground feature classification algorithms, the PolSAR ground feature fine classification algorithm based on the multi-index convolution self-encoder provided by the invention has better continuity in the classification result, namely fewer isolated points in the classification result image. In several other comparison methods, CLBP-CNN fuses description edge and texture feature CLBP, but the time cost spent in extracting features and training process in practical application is high due to the high CLBP data dimension. MRF-CNN is limited by the underlying network, and when the classification task is complex, it is difficult to remove the misclassified area, and even the classification performance is degraded again. PSE-NET increases the correlation between information through the proposed SE model, thus improving classification performance, but texture features are still lost, and more sample data is needed to participate in training due to the increase of parameters, otherwise the classification result is poor. The internal network parameters in CV-CNN also belong to a complex form, so the number of parameters that need to be trained is large. VGG-Net and Res-Net have very deep network depths, and also have very high requirements on the data volume of training samples, otherwise the final classification is less effective. The method of the invention embeds the newly proposed MI-SE module on the basis of CAE, effectively inhibits the interference information in the data set, and simultaneously carries out multi-index weighting on the characteristics extracted by the convolution and pooling layers, thereby greatly improving the characteristic separation degree. At the same time, three metrics are presented in the MI-SE module: the first moment may reflect the average of the features; the second order center distance may describe the fluctuations in each feature, reflecting the content of the edge and texture information to some extent; the coefficient of variation can eliminate the influence of the average on the dispersion, thereby better reflecting the fluctuation. The three indexes complement each other to realize weighting of each feature at different angles.
In conclusion, the PolSAR ground object fine classification method based on the multi-index convolution self-encoder has the characteristics of simple structure, accurate classification and high calculation efficiency, and has high application value in practical engineering.
Claims (1)
1. A PolSAR ground object fine classification method based on a multi-index convolution self-encoder is characterized by comprising the following steps:
step 1: preprocessing and clipping operation of PolSAR data:
step 1.1: the scattering characteristic of any minimum resolution unit in the PolSAR data is represented by using a polarized scattering matrix S; according to the reciprocity principle and Pauli decomposition principle, converting any one polarized scattering matrix S in PolSAR data into a vectorWherein S is HH Representing the complex scattering coefficient when the transmitting polarization mode is horizontal polarization and the receiving polarization mode is horizontal polarization, S VV Representing the complex scattering coefficient when the transmitting polarization mode is vertical polarization and the receiving polarization mode is vertical polarization, S HV Representing the complex scattering coefficient when the transmitting polarization mode is vertical polarization and the receiving polarization mode is horizontal polarization;
step 1.2: based on the loss K, a polarization coherence matrix T is obtained using equation (1):
in the formula (1) (. Cndot. H Represents conjugate transpose, T ij Elements representing the ith row and jth column of the polarization coherence matrix T; and i, j=1, 2,3;
step 1.3: extracting a six-dimensional eigenvector f= [ a, B, C, D, E, F ] from the polarization coherence matrix T using formula (2):
in equation (2), span represents the total scattered power of all polarized channels, and span=t 11 +T 22 +T 33 A represents the decibel form of the total scattered power Span, B and C represent the 2 nd row and 2 nd column elements T, respectively 22 And line 3, column 3 element T 33 D, E and F represent three relative correlation coefficients;
step 1.4: normalizing the six-dimensional feature vector F to obtain a normalized six-dimensional feature vector F', so that preprocessed PolSAR data are formed by the normalized six-dimensional feature vectors of all minimum resolution units;
step 1.5: cutting the preprocessed PolSAR data into slices with the size of L multiplied by 6, thereby obtaining a slice set { s } of the PolSAR data 1 ,...,s n ,...,s N -wherein s n Represents an nth slice, N represents the total number of slices; n is E [1, N]；
Step 2: constructing a multi-index convolutional self-encoder based classification network, comprising: an encoder, a decoder, and a classifier; and aggregating the slices { s } 1 ,...,s n ,...,s N -input into the classification network;
step 2.1: the encoder consists of m convolution layers, m pooling layers, m MI-SE modules, m weighting modules and a full connection layer;
the nth slice s n Inputting the characteristic diagram U into the encoder, and obtaining a characteristic diagram U with a scale of H multiplied by W multiplied by D after the treatment of a first convolution layer and a pooling layer, wherein H and W respectively represent the height and the width of the characteristic diagram U, and D represents the channel number of the characteristic diagram U;
the feature map U is input into a first MI-SE module, and a multi-index is calculated by using a formula (3), and the method comprises the following steps: first moment S 1 Second order central moment S 2 And coefficient of variation S 3 ：
In the formula (3), U (a, b, i) represents all elements of the a-th row and the b-th column of the characteristic diagram U;
the first moment S 1 Second order central moment S 2 And coefficient of variation S 3 Splicing the matrix into a multi-index matrix V with the size of 3 xD, and processing the multi-index matrix V by a convolution layer and a full connection layer in the first MI-SE module to obtain a weight matrix Y with the size of 1 xD;
the first weighting module multiplies the weight matrix Y with the feature map U to obtain a weighted feature map U' with the size of H multiplied by W multiplied by D;
the weighted feature map U' sequentially passes through the rest m-1 convolution layers, the pooling layer, the MI-SE module and the weighting module of the encoder, and finally the output features pass through a full connection layer to obtain the feature vector alpha output by the encoder;
step 2.2: the decoder is formed by sequentially connecting k convolution layers and k up-sampling layers in a staggered manner;
the feature vector alpha is input into a decoder to obtain an nth slice s n Is a reconstructed slice of (a)Re-use of (4) to establish a counter-propagating loss function L MSE ：
In equation (4), MSE represents the minimum mean square error; p is the number of elements per slice,for the nth slice s n The c-th element in the slice reconstructed after decoder,>for the nth slice s n C element of (a);
the encoder and decoder are trained using back propagation whenThe loss function L MSE When the minimum value is taken, training is stopped, and the optimal feature vector alpha output by the trained encoder is obtained * ；
Step 2.3: the classifier consists of t full-connection layers and one softmax layer;
the optimal feature vector alpha * In the input classifier, firstly, reducing the dimension through t full-connection layers, then inputting the feature vector after the dimension reduction into a softmax layer, and calculating an nth slice s n The center element of (a) corresponds to the posterior probability of each category, and the category corresponding to the maximum posterior probability is selected as the nth slice s n Is a predictive category of (2);
step 2.4: obtaining prediction categories of N slices according to the modes of the steps 2.1-2.3, comparing the prediction categories with corresponding real categories, and optimizing the classifier through back propagation, so that training of the encoder, the decoder and the classifier is completed, and a trained multi-index convolution self-encoder model is obtained and used for classifying PolSAR data to be classified.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111319995.4A CN114037896B (en) | 2021-11-09 | 2021-11-09 | PolSAR ground object fine classification method based on multi-index convolution self-encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111319995.4A CN114037896B (en) | 2021-11-09 | 2021-11-09 | PolSAR ground object fine classification method based on multi-index convolution self-encoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114037896A CN114037896A (en) | 2022-02-11 |
| CN114037896B true CN114037896B (en) | 2024-02-13 |
Family
ID=80143590
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111319995.4A Active CN114037896B (en) | 2021-11-09 | 2021-11-09 | PolSAR ground object fine classification method based on multi-index convolution self-encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114037896B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020162834A1 (en) * | 2019-02-08 | 2020-08-13 | Singapore Health Services Pte Ltd | Method and system for classification and visualisation of 3d images |
| CN112434628A (en) * | 2020-11-30 | 2021-03-02 | 西安理工大学 | Small sample polarization SAR image classification method based on active learning and collaborative representation |
| CN112966779A (en) * | 2021-03-29 | 2021-06-15 | 安徽大学 | PolSAR image semi-supervised classification method |
| CN113240040A (en) * | 2021-05-27 | 2021-08-10 | 西安理工大学 | Polarized SAR image classification method based on channel attention depth network |
| CN113392871A (en) * | 2021-04-06 | 2021-09-14 | 北京化工大学 | Polarized SAR terrain classification method based on scattering mechanism multichannel expansion convolutional neural network |
-
2021
- 2021-11-09 CN CN202111319995.4A patent/CN114037896B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020162834A1 (en) * | 2019-02-08 | 2020-08-13 | Singapore Health Services Pte Ltd | Method and system for classification and visualisation of 3d images |
| CN112434628A (en) * | 2020-11-30 | 2021-03-02 | 西安理工大学 | Small sample polarization SAR image classification method based on active learning and collaborative representation |
| CN112966779A (en) * | 2021-03-29 | 2021-06-15 | 安徽大学 | PolSAR image semi-supervised classification method |
| CN113392871A (en) * | 2021-04-06 | 2021-09-14 | 北京化工大学 | Polarized SAR terrain classification method based on scattering mechanism multichannel expansion convolutional neural network |
| CN113240040A (en) * | 2021-05-27 | 2021-08-10 | 西安理工大学 | Polarized SAR image classification method based on channel attention depth network |
Non-Patent Citations (1)
| Title |
|---|
| 全卷积网络和条件随机场相结合的全极化SAR土地覆盖分类;赵泉华;谢凯浪;王光辉;李玉;;测绘学报;20200115(01);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114037896A (en) | 2022-02-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110135267B (en) | Large-scene SAR image fine target detection method | |
| CN110533631B (en) | SAR image change detection method based on pyramid pooling twin network | |
| CN109214452B (en) | HRRP target identification method based on attention depth bidirectional cyclic neural network | |
| CN109190491B (en) | Sea ice classification method for residual convolutional neural network SAR (synthetic Aperture Radar) image | |
| CN113240040B (en) | Polarized SAR image classification method based on channel attention depth network | |
| CN106355151A (en) | Recognition method, based on deep belief network, of three-dimensional SAR images | |
| CN101369316B (en) | Image characteristics extraction method based on global and local structure amalgamation | |
| Yang et al. | Aa-transunet: Attention augmented transunet for nowcasting tasks | |
| CN111639587B (en) | Hyperspectral image classification method based on multi-scale spectrum space convolution neural network | |
| CN108564115A (en) | Semi-supervised polarization SAR terrain classification method based on full convolution GAN | |
| CN111709487B (en) | Underwater multi-source acoustic image substrate classification method and system based on decision-level fusion | |
| CN104252625A (en) | Sample adaptive multi-feature weighted remote sensing image method | |
| CN110674858A (en) | Traffic jam detection method based on space-time association and big data mining | |
| CN113343563A (en) | Landslide susceptibility evaluation method based on automatic sample selection and surface deformation rate | |
| CN103366371A (en) | K distribution and texture feature-based SAR (Synthetic Aperture Radar) image segmentation method | |
| CN115222994A (en) | Hyperspectral image classification method based on hybrid spectrum network and multi-head self-attention mechanism | |
| CN114913434B (en) | High-resolution remote sensing image change detection method based on global relation reasoning | |
| CN106446935A (en) | Kernel sparse representation and spatial constraint-based polarimetric SAR image classification method | |
| CN114821340A (en) | Land utilization classification method and system | |
| CN115909078A (en) | Ship classification method based on HRRP and SAR data feature level fusion | |
| CN116168235A (en) | Hyperspectral image classification method based on double-branch attention network | |
| CN116580322A (en) | Unmanned aerial vehicle infrared small target detection method under ground background | |
| Ferreira et al. | Ship detection in sar images using convolutional variational autoencoders | |
| CN110516727A (en) | Hyperspectral image classification method based on FPGA depth edge filter | |
| CN102073875B (en) | Sparse representation-based background clutter quantification method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |