A method and system for detecting polarization SAR changes based on Siamese attention complex convolutional neural networks

Abstract

This invention discloses a polarimetric SAR change detection method and system based on a Siamese attention complex convolutional neural network. It addresses the issues of poor quality difference maps generated by speckle noise in PolSAR images, and the problem that current real-number networks for SAR image change detection can disrupt the integrity of the complex structure of PolSAR data. This invention utilizes a Siamese network architecture to construct a more robust multi-scale feature difference map to suppress the influence of speckle noise. Furthermore, it combines a complex network to maintain the integrity of the PolSAR data structure and reduce polarimetric information loss. A corresponding complex attention module is also designed to enhance the identification of changed regions. Finally, the change detection results are obtained by decoding the multi-scale feature difference map. By using a Siamese attention complex convolutional neural network, the invention enhances the utilization of polarimetric information while suppressing the influence of speckle noise, thereby improving the accuracy of PolSAR change detection.

Description

Technical Field

[0001] This invention relates to the field of radar remote sensing image processing technology, and in particular to a polarization SAR change detection method and system based on a twin attention complex convolutional neural network. Background Technology

[0002] As society progresses, the demand for information on land surface changes is increasing. Traditional field surveys, due to their high cost and inefficiency, can no longer meet the requirements for efficient, rapid, and large-scale land surface change monitoring. Remote sensing technology, through satellite imaging of the Earth, can quickly conduct large-scale land surface monitoring, and its short revisit cycle allows for the acquisition of large amounts of data in a short time, greatly satisfying the basic needs of current land surface change monitoring. Optical remote sensing has many applications in change detection tasks; however, due to its passive sensing nature, it cannot achieve high-quality imaging in cloudy or rainy weather, limiting its use in certain scenarios.

[0003] Synthetic Aperture Radar (SAR) is an active remote sensing technology that acquires ground feature information by actively emitting electromagnetic waves and receiving echoes from ground objects. It can penetrate clouds and fog, offering advantages in all-weather, day-and-night operation, effectively compensating for the limitations of optical remote sensing. Currently, most SAR change detection research is based on single-polarization SAR imagery, identifying changes by the differences in intensity information of different ground features within the SAR imagery. However, when the intensities of different ground features are similar, confusion can easily occur, leading to low change detection accuracy. PolSAR, by emitting and receiving electromagnetic waves with different polarizations, can acquire echoes from different polarization channels of ground features, obtaining richer ground feature scattering information, thus better assisting in ground feature identification and interpretation.

[0004] In implementing this invention, the inventors discovered that statistical methods, such as hypothesis testing likelihood ratio methods, polarization likelihood ratio models, and polarization scattering difference methods, are still widely used in PolSAR change detection. However, PolSAR change detection methods based on statistical methods are mainly pixel-based, which can lead to the loss of spatial information. Convolutional neural network-based methods can better preserve spatial information; however, most current networks are real-number networks, primarily designed for change detection in single-polarization SAR images. PolSAR data is in complex form, and using real-number networks may result in the loss of some polarization information. Furthermore, speckle noise in SAR images has a significant impact on change detection results, making the suppression of speckle noise crucial.

[0005] In summary, given the existing problems of insufficient utilization of polarization information and low speckle noise suppression in PolSAR intelligent change detection methods, developing a change detection method that takes into account both spatial and polarization information is of great practical significance. Summary of the Invention

[0006] This invention provides a polarization SAR change detection method based on a Siamese attention complex convolutional neural network, which solves or at least partially solves the technical problems of low detection accuracy caused by the difficulty in fully utilizing polarization information and the low degree of speckle noise suppression in the prior art.

[0007] To address the aforementioned technical problems, the first aspect of this invention provides a polarization SAR change detection method based on a Siamese attention complex convolutional neural network, comprising:

[0008] The acquired PolSAR images from different time phases are preprocessed, including radiometric calibration, terrain correction, registration, T3 coherence matrix generation, and filtering.

[0009] Based on the comparison of the preprocessed PolSAR images from different time phases, a change label map is obtained;

[0010] The upper triangular matrix of the preprocessed biphase T3 matrix is ​​taken, and the imaginary part of the diagonal elements is set to 0. The resulting biphase complex matrix is ​​cropped with the same size as the corresponding change label map to obtain the sliced ​​dataset. The dataset is then randomly divided into training set and test set.

[0011] A Siamese attention complex convolutional network is constructed using the UNet encoding and decoding structure as the basic framework. The convolution, pooling, batch normalization and deconvolution in the network are designed with corresponding complex modules. A residual connection is established between the input tensor and the output tensor after network processing. A Siamese network structure is adopted in the encoder stage. The dual-temporal multi-scale features extracted in the Siamese encoder stage are subtracted to generate a multi-scale feature difference map. A complex attention module is designed and embedded into the decoder. The multi-scale feature difference map is fused by the decoder for decoding.

[0012] The Siamese attention complex convolutional network was trained using the training set, and a hybrid loss function was used to obtain the trained network model. The test set was then input into the trained network model to obtain the polarimetric SAR change detection results.

[0013] In one implementation, the filtering method in the preprocessing process employs refined Lee filtering.

[0014] In one implementation, a change label map is obtained based on the comparison of preprocessed PolSAR images from different time phases, including:

[0015] By comparing the pre-processed PolSAR images from the previous and subsequent time phases, the areas of change between the two time phases can be obtained.

[0016] The change label map is obtained by manually marking the change areas between the previous and subsequent time phases.

[0017] In one embodiment, the Siamese attention complex convolutional network includes: an initialization module, a complex downsampling module, a feature map difference module, a complex upsampling module, a complex attention module, and an output module. The initialization module initializes the various parameters in the network; the complex downsampling module reduces the feature map size to extract high-level features from the image; the feature map difference module calculates the feature difference map between the two downsampled images; the complex upsampling module restores the reduced feature map size to the original input size; the complex attention module enhances the feature recognition capability of the complex neural network; and the output module transforms the feature maps learned by the network into the final output.

[0018] In one implementation, the convolution, pooling, batch normalization, and deconvolution in the network are designed using corresponding complex modules, including:

[0019] The convolution module is designed using complex number modules, and the calculation method is as follows:

[0020] Z*x=(A*aB*b)+(A*b+a*B)i (1)

[0021] x is the input tensor, x = a + bi, Z is the convolution kernel, Z = A + Bi, a and b represent the real and imaginary parts of the complex number, A and B represent the real and imaginary parts of the convolution kernel, and i represents the imaginary unit;

[0022] The pooling layer is designed using a complex number module. Specifically, the maximum value in the max pooling layer is determined by comparing the complex number moduli and selecting the complex number with the largest moduli as the maximum pooling element to replace the entire region. The average pooling layer sums all the complex numbers in the feature map, then calculates the average of the real and imaginary parts and combines them to obtain the pooling average complex number, which is used to replace the entire region.

[0023] The batch normalization layer is designed using a complex number module. Specifically, the batch normalization layer in the complex number network calculates the mean and variance of each complex feature separately.

[0024] The complex activation function used is CLeakyReLU, and its mathematical expression is:

[0025]

[0026] in To take the real part of a complex number, This indicates taking the imaginary part of a complex number.

[0027] In one implementation, the processing of the complex attention module includes:

[0028] The input feature map F is processed by parallel global pooling modules to obtain the global pooled feature vector. The mathematical process is represented as follows:

[0029]

[0030] in, Let C represent the feature vector, H represent the image height, W represent the image width, CAvgpool(·) represent complex global average pooling, and CMaxpool(·) represent complex global max pooling. Indicates channel splicing;

[0031] Global pooling feature vector M p (F) The channel weight vector is obtained after two 1×1 convolutions. The mathematical process is represented as follows:

[0032]

[0033] in, This represents a 1×1 complex convolution with C′ output channels;

[0034] Channel weight vector M w (F) Normalized weighted sum is obtained The mathematical process is represented as follows:

[0035]

[0036] Where σ represents the Sigmoid function, and |·| represents the modulo operation. M′ represents the product of the numerical values ​​of corresponding elements. w (F) represents the normalized weight vector;

[0037] Based on the normalized weight vector M′ w (F) Obtain attention feature map The mathematical process is represented as follows:

[0038]

[0039] in To take the real part of a complex number, M′ represents taking the imaginary part of a complex number. w (F) First expand the dimensions to Then participate in feature map weighting;

[0040] A residual connection is established between the input tensor and the output tensor after network processing to obtain the final output feature map. The mathematical process is represented as follows:

[0041] F′=M c (F)+F (7).

[0042] In one implementation, the hybrid loss function is:

[0043] Loss=α*DiceLoss+(1-α)*BCELoss (8)

[0044] Where DiceLoss represents the Dice coefficient loss, BCELoss represents the binary classification cross-entropy loss, DiceLoss is used to measure the overall similarity between the predicted image and the label image, and BCELoss is used to compare the similarity between individual pixels. The calculation expression is as follows:

[0045]

[0046] BCELoss=-[y k *logP k +(1-y k )*log(1-P k (10)

[0047] Where X and Y represent two sample sets, |X∩Y| represents the number of elements in the intersection of X and Y, and |X| and |Y| represent the number of elements in X and Y, respectively. k P represents the label value of pixel k. k This represents the output label probability corresponding to pixel k.

[0048] Based on the same inventive concept, a second aspect of the present invention provides a polarization SAR change detection system based on a Siamese attention complex convolutional neural network, comprising:

[0049] The image preprocessing module is used to preprocess the acquired PolSAR images from different time phases. The preprocessing includes radiometric calibration, terrain correction, registration, T3 coherence matrix generation, and filtering.

[0050] The change labeling module is used to obtain change label maps based on the comparison of preprocessed PolSAR images from different time phases.

[0051] The dataset construction module is used to take the upper triangular matrix of the preprocessed biphase T3 matrix and set the imaginary part of the diagonal elements to 0. The resulting biphase complex matrix is ​​cropped with the same size as the biphase complex matrix and the corresponding change label map to obtain the sliced ​​dataset. The dataset is then randomly divided into training set and test set.

[0052] The network construction module is used to build a Siamese attention complex convolutional network based on the UNet encoding and decoding structure. The convolution, pooling, batch normalization and deconvolution in the network are designed with corresponding complex modules. A residual connection is established between the input tensor and the output tensor after network processing. The Siamese network structure is used in the encoder stage to perform difference on the dual temporal multi-scale features extracted in the Siamese encoder stage to generate a multi-scale feature difference map. A complex attention module is designed and embedded into the decoder. The decoder fuses the multi-scale feature difference map for decoding.

[0053] The training and testing module is used to train the Siamese attention complex convolutional network using the training set and a hybrid loss function to obtain the trained network model. The test set is then input into the trained network model to obtain the polarimetric SAR change detection results.

[0054] Based on the same inventive concept, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the method described in the first aspect.

[0055] Based on the same inventive concept, a fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.

[0056] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:

[0057] This invention proposes a polarimetric SAR change detection method based on a Siamese attention complex convolutional neural network. The Siamese attention complex convolutional network is constructed using the UNet encoding and decoding structure as the basic framework. This network utilizes the Siamese network architecture to generate more robust multi-scale feature difference maps, which to some extent suppresses the influence of speckle noise. Furthermore, the use of a complex network maintains the integrity of the PolSAR data structure and enhances the utilization of polarimetric information. A complex attention module is also designed to strengthen the extraction of change regions, thus effectively improving the accuracy of PolSAR change detection. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1This is a flowchart illustrating the overall technical route of the polarimetric SAR change detection method in this embodiment of the invention.

[0060] Figure 2 This is a schematic diagram of the twin attention complex network structure in an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of complex convolution in an embodiment of the present invention;

[0062] Figure 4 This is a diagram of the sub-module structure in the network of this invention;

[0063] Figure 5 This is a structural diagram of the complex attention module. Detailed Implementation

[0064] To address the low accuracy of existing PolSAR intelligent change detection methods due to insufficient utilization of polarization information and low speckle noise suppression, this invention proposes a change detection method that considers both spatial and polarization information. The main concept is as follows:

[0065] By utilizing a Siamese network architecture, a more robust multi-scale feature difference map is constructed to suppress the influence of speckle noise. Furthermore, a complex network is incorporated to maintain the integrity of the PolSAR data structure and reduce polarization information loss. A corresponding complex attention module is also designed to enhance the identification of changed regions. Finally, change detection results are obtained by decoding the multi-scale feature difference map. Through the Siamese attention complex convolutional neural network, the utilization of polarization information is enhanced while suppressing the influence of speckle noise, thus improving the accuracy of PolSAR change detection.

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Example 1

[0068] This invention provides a polarization SAR change detection method based on a Siamese attention complex convolutional neural network, comprising:

[0069] The acquired PolSAR images from different time phases are preprocessed, including radiometric calibration, terrain correction, registration, T3 coherence matrix generation, and filtering.

[0070] Based on the comparison of the preprocessed PolSAR images from different time phases, a change label map is obtained;

[0071] The upper triangular matrix of the preprocessed biphase T3 matrix is ​​taken, and the imaginary part of the diagonal elements is set to 0. The resulting biphase complex matrix is ​​cropped with the same size as the corresponding change label map to obtain the sliced ​​dataset. The dataset is then randomly divided into training set and test set.

[0072] A Siamese attention complex convolutional network is constructed using the UNet encoding and decoding structure as the basic framework. The convolution, pooling, batch normalization and deconvolution in the network are designed with corresponding complex modules. A residual connection is established between the input tensor and the output tensor after network processing. A Siamese network structure is adopted in the encoder stage. The dual-temporal multi-scale features extracted in the Siamese encoder stage are subtracted to generate a multi-scale feature difference map. A complex attention module is designed and embedded into the decoder. The multi-scale feature difference map is fused by the decoder for decoding.

[0073] The Siamese attention complex convolutional network was trained using the training set, and a hybrid loss function was used to obtain the trained network model. The test set was then input into the trained network model to obtain the polarimetric SAR change detection results.

[0074] Please see Figure 1 This is a flowchart illustrating the overall technical route of the polarimetric SAR change detection method in this embodiment of the invention.

[0075] The main steps include preprocessing, dataset construction, model training, and prediction. Specifically, the PolSAR images from different time periods include the earlier and later PolSAR images, and the change label map is obtained by comparing the two.

[0076] Since the polarization coherence matrix T3 is a conjugate matrix, to avoid data redundancy, the upper triangular matrix is ​​taken from the preprocessed biphase T3 matrix (the first phase T3 matrix and the second phase T3 matrix), and the imaginary part of the diagonal elements is set to 0. The obtained biphase complex matrix and the corresponding change label map are cropped to the same size to obtain the sliced ​​dataset, and the dataset is randomly divided into training set and test set.

[0077] The polarization coherence matrix T3 is in the following form:

[0078]

[0079] Where H and V represent horizontal polarization and vertical polarization, respectively, and S HH S represents a polarization channel for horizontal transmission and horizontal reception. VVS represents the polarization channel for vertical transmission and vertical reception. HV This represents the polarization channel for vertical transmission and horizontal reception. * represents the conjugate operation, || represents the complex modulo operation, and <> represents the ensemble averaging operation.

[0080] Regarding Siamese attention complex convolutional networks, this invention uses the UNet encoding / decoding structure as the basic framework. The network's fundamental modules, such as convolution, pooling, batch normalization, and deconvolution, are detailed in the appendix. Figure 4 All models are designed using corresponding complex modules, and residual connections are established between the input tensor and the output tensor after network processing. A Siamese network structure is used in the encoder stage to interpolate the dual-temporal multi-scale features extracted in the Siamese encoder stage, generating a robust multi-scale feature difference map. A complex attention module is then designed and embedded into the decoder. Finally, the decoder fuses the multi-scale feature difference map for decoding. The residual connection between the input tensor and the output tensor after network processing aims to incorporate original input information and suppress the trend of information degradation acquired by deep networks. Since the off-diagonal elements of the polarimetric SAR coherence matrix are all complex numbers, a complex attention module is designed to improve the utilization of polarimetric information and enhance the feature recognition capability of the complex neural network. This makes the constructed network model more suitable for processing polarimetric SAR images, thereby suppressing speckle noise in SAR images and improving the detection accuracy of changing areas.

[0081] In one implementation, the filtering method in the preprocessing process employs refined Lee filtering.

[0082] The refined Lee filter has a small computational load, good filtering effect, and can better preserve details such as points, lines and edges.

[0083] In one implementation, a change label map is obtained based on the comparison of preprocessed PolSAR images from different time phases, including:

[0084] By comparing the pre-processed PolSAR images from the previous and subsequent time phases, the areas of change between the two time phases can be obtained.

[0085] The change label map is obtained by manually marking the change areas between the previous and subsequent time phases.

[0086] In one embodiment, the Siamese attention complex convolutional network includes: an initialization module, a complex downsampling module, a feature map difference module, a complex upsampling module, a complex attention module, and an output module. The initialization module initializes the various parameters in the network; the complex downsampling module reduces the feature map size to extract high-level features from the image; the feature map difference module calculates the feature difference map between the two downsampled images; the complex upsampling module restores the reduced feature map size to the original input size; the complex attention module enhances the feature recognition capability of the complex neural network; and the output module transforms the feature maps learned by the network into the final output.

[0087] Please see Figure 2 This is a schematic diagram of the complex network structure of the Siamese attention network in an embodiment of the present invention. The network consists of an encoder and a decoder. The encoder uses a Siamese network to share weights, and the decoder embeds a complex attention module to fuse multi-scale feature difference maps for decoding. The Init Block is the initialization module; CDownsample is the complex downsampling module; Dif_Feature is the feature map difference module; CUpsamle is the complex upsampling module; the CSE module is the complex attention module; and the OutConv Block is the output module. Figure 4 This is a schematic diagram of the structure of each sub-module of the network in this invention. The basic modules include convolution, pooling, batch normalization, and deconvolution.

[0088] In one implementation, the convolution, pooling, batch normalization, and deconvolution in the network are designed using corresponding complex modules, including:

[0089] The convolution module is designed using complex number modules, and the calculation method is as follows:

[0090] Z*x=(A*aB*b)+(A*b+a*B)i (1)

[0091] x is the input tensor, x = a + bi, Z is the convolution kernel, Z = A + Bi, a and b represent the real and imaginary parts of the complex number, A and B represent the real and imaginary parts of the convolution kernel, and i represents the imaginary unit;

[0092] The pooling layer is designed using a complex number module. Specifically, the maximum value in the max pooling layer is determined by comparing the complex number moduli and selecting the complex number with the largest moduli as the maximum pooling element to replace the entire region. The average pooling layer sums all the complex numbers in the feature map, then calculates the average of the real and imaginary parts and combines them to obtain the pooling average complex number, which is used to replace the entire region.

[0093] The batch normalization layer is designed using a complex number module. Specifically, the batch normalization layer in the complex number network calculates the mean and variance of each complex feature separately.

[0094] The complex activation function used is CLeakyReLU, and its mathematical expression is:

[0095]

[0096] in To take the real part of a complex number, This indicates taking the imaginary part of a complex number.

[0097] In one implementation, the processing of the complex attention module includes:

[0098] The input feature map F is processed by parallel global pooling modules to obtain the global pooled feature vector. The mathematical process is represented as follows:

[0099]

[0100] in, Let C represent the feature vector, H represent the image height, W represent the image width, CAvgpool(·) represent complex global average pooling, and CMaxpool(·) represent complex global max pooling. Indicates channel splicing;

[0101] Global pooling feature vector M p (F) The channel weight vector is obtained after two 1×1 convolutions. The mathematical process is represented as follows:

[0102]

[0103] in, This represents a 1×1 complex convolution with C′ output channels;

[0104] Channel weight vector M w (F) Normalized weighted sum is obtained The mathematical process is represented as follows:

[0105]

[0106] Where σ represents the Sigmoid function, and |·| represents the modulo operation. M′ represents the product of the numerical values ​​of corresponding elements. w (F) represents the normalized weight vector;

[0107] Based on the normalized weight vector M′ w (F) Obtain attention feature map The mathematical process is represented as follows:

[0108]

[0109] in To take the real part of a complex number, M′ represents taking the imaginary part of a complex number. w (F) First expand the dimensions to Then participate in feature map weighting;

[0110] A residual connection is established between the input tensor and the output tensor after network processing to obtain the final output feature map. The mathematical process is represented as follows:

[0111] F′=M c (F)+F (7).

[0112] Specifically, please see Figure 3 This is a schematic diagram of complex convolution in an embodiment of the present invention.

[0113] The average pooling calculation rule in complex networks is consistent with that in real networks. It involves summing all complex numbers in the feature map, then averaging the real and imaginary parts separately, and combining them to obtain the pooled average complex number, which is then used to replace the entire region.

[0114] In a complex network, the Batch Normalization (BN) layer calculates the mean and variance for each complex feature. The mean of a complex feature is calculated in the same way as that of a real feature, using the arithmetic mean of all features. The variance of a complex feature is calculated by subtracting the complex conjugate of the mean from each feature and then averaging the sum of squares. Each complex feature is standardized by subtracting the mean from each feature and then dividing by the square root of the variance.

[0115] In the encoder stage, a Siamese network structure is adopted. The bi-temporal multi-scale features extracted in the Siamese encoder stage are interpolated to generate a robust multi-scale feature difference map. Then, a complex attention module is designed and embedded into the decoder. Finally, the multi-scale feature difference map is fused through the decoder for decoding. The structure diagram of the attention module is shown below. Figure 5 As shown.

[0116] In one implementation, the hybrid loss function is:

[0117] Loss=α*DiceLoss+(1-α)*BCELoss (8)

[0118] Where DiceLoss represents the Dice coefficient loss, BCELoss represents the binary classification cross-entropy loss, DiceLoss is used to measure the overall similarity between the predicted image and the label image, and BCELoss is used to compare the similarity between individual pixels. The calculation expression is as follows:

[0119]

[0120] BCELoss=-[y k *log P k +(1-y k )*log(1-P k (10)

[0121] Where X and Y represent two sample sets, |X∩Y| represents the number of elements in the intersection of X and Y, and |X| and |Y| represent the number of elements in X and Y, respectively. k P represents the label value of pixel k. k This represents the output label probability corresponding to pixel k.

[0122] Specifically, DiceLoss primarily measures the overall similarity between two samples, that is, the overall similarity between the predicted image and the label image, evaluating the prediction results from a macroscopic perspective. BCELoss, on the other hand, compares the similarity between individual pixels, comparing the prediction results from a microscopic perspective. Combining the two can suppress the influence of imbalances between changing and invariant pixels.

[0123] Example 2

[0124] Based on the same inventive concept, this embodiment discloses a polarization SAR change detection system based on a Siamese attention complex convolutional neural network, comprising:

[0125] The image preprocessing module is used to preprocess the acquired PolSAR images from different time phases. The preprocessing includes radiometric calibration, terrain correction, registration, T3 coherence matrix generation, and filtering.

[0126] The change labeling module is used to obtain change label maps based on the comparison of preprocessed PolSAR images from different time phases.

[0127] The dataset construction module is used to take the upper triangular matrix of the preprocessed biphase T3 matrix and set the imaginary part of the diagonal elements to 0. The resulting biphase complex matrix is ​​cropped with the same size as the biphase complex matrix and the corresponding change label map to obtain the sliced ​​dataset. The dataset is then randomly divided into training set and test set.

[0128] The network construction module is used to build a Siamese attention complex convolutional network based on the UNet encoding and decoding structure. The convolution, pooling, batch normalization and deconvolution in the network are designed with corresponding complex modules. A residual connection is established between the input tensor and the output tensor after network processing. The Siamese network structure is used in the encoder stage to perform difference on the dual temporal multi-scale features extracted in the Siamese encoder stage to generate a multi-scale feature difference map. A complex attention module is designed and embedded into the decoder. The decoder fuses the multi-scale feature difference map for decoding.

[0129] The training and testing module is used to train the Siamese attention complex convolutional network using the training set and a hybrid loss function to obtain the trained network model. The test set is then input into the trained network model to obtain the polarimetric SAR change detection results.

[0130] Since the system described in Embodiment 2 of this invention is the same system used to implement the polarization SAR change detection method based on Siamese attention complex convolutional neural network in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this system based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All systems used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0131] Example 3

[0132] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed, implements the method described in Embodiment 1.

[0133] Since the computer-readable storage medium described in Embodiment 3 of this invention is the same computer-readable storage medium used in implementing the polarization SAR change detection method based on Siamese attention complex convolutional neural network in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer-readable storage medium based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer-readable storage media used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0134] Example 4

[0135] Based on the same inventive concept, this application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method in Embodiment 1.

[0136] Since the computer device described in Embodiment 4 of this invention is the same computer device used to implement the polarization SAR change detection method based on Siamese attention complex convolutional neural network in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0137] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0138] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0139] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A polarimetric SAR change detection method based on twin attention complex convolutional neural network, characterized in that, include: The obtained PolSAR images of the before and after time phases are preprocessed, wherein the preprocessing includes radiation calibration, terrain correction, registration, coherence matrix generation and filtering; Based on the comparison of the preprocessed PolSAR images from different time phases, a change label map is obtained; The preprocessed two-phase The matrix is ​​taken as an upper triangular matrix, and the imaginary part of the diagonal elements is set to 0. The resulting biphase complex matrix is ​​cropped with the same size as the biphase complex matrix and the corresponding change label map to obtain the sliced ​​dataset. The dataset is then randomly divided into training set and test set. A Siamese attention complex convolutional network is constructed using the UNet encoding and decoding structure as the basic framework. The convolution, pooling, batch normalization and deconvolution in the network are designed with corresponding complex modules. A residual connection is established between the input tensor and the output tensor after network processing. A Siamese network structure is adopted in the encoder stage. The dual-temporal multi-scale features extracted in the Siamese encoder stage are subtracted to generate a multi-scale feature difference map. A complex attention module is designed and embedded into the decoder. The multi-scale feature difference map is fused by the decoder for decoding. The Siamese attention complex convolutional network was trained using the training set, and a hybrid loss function was used to obtain the trained network model. The test set was then input into the trained network model to obtain the polarimetric SAR change detection results. The processing steps of the complex attention module include: Input feature map The global pooling feature vector is obtained through parallel global pooling modules. The mathematical process is represented as follows: （3） in, Represents the eigenvector. C Indicates the number of channels. H Indicates the height of the image. W Indicates the width of the image. This represents global average pooling of complex numbers. This represents global max pooling of complex numbers. Indicates channel splicing; Global pooling feature vectors The channel weight vector is obtained after two 1×1 convolutions. The mathematical process is represented as follows: （4） in, This indicates that the number of output channels is 1×1 complex convolution; Channel weight vector Normalized weighted average The mathematical process is represented as follows: （5） in express function, This represents the modulo operation. This indicates that the values ​​of the corresponding elements are multiplied. This represents the normalized weight vector; Based on normalized weight vector Obtaining attention feature maps The mathematical process is represented as follows: （6） in To take the real part of a complex number, This indicates taking the imaginary part of the complex number. First, expand the dimensions to Then participate in feature map weighting; A residual connection is established between the input tensor and the output tensor after network processing to obtain the final output feature map. The mathematical process is represented as follows: （7）。 2. The polarization SAR change detection method based on Siamese attention complex convolutional neural network as described in claim 1, characterized in that, The filtering method used in the preprocessing process is refined Lee filtering.

3. The polarization SAR change detection method based on Siamese attention complex convolutional neural network as described in claim 1, characterized in that, Based on the comparison of the preprocessed PolSAR images from different time phases, a change label map is obtained, including: By comparing the pre-processed PolSAR images from the previous and subsequent time phases, the areas of change between the two time phases can be obtained. The change label map is obtained by manually marking the change areas between the previous and subsequent time phases.

4. The polarization SAR change detection method based on Siamese attention complex convolutional neural network as described in claim 1, characterized in that, The Siamese Attention Complex Convolutional Network comprises an initialization module, a complex downsampling module, a feature map differencing module, a complex upsampling module, a complex attention module, and an output module. The initialization module initializes the network parameters; the complex downsampling module reduces the feature map size to extract high-level features from the image; the feature map differencing module calculates the feature difference map between the two downsampled images; the complex upsampling module restores the reduced feature map size to its original input size; the complex attention module enhances the feature recognition capability of the complex neural network; and the output module transforms the learned feature maps into the final output.

5. The polarization SAR change detection method based on Siamese attention complex convolutional neural network as described in claim 1, characterized in that, The convolution, pooling, batch normalization, and deconvolution functions in the network are designed using corresponding complex number modules, including: The convolution module is designed using complex number modules, and the calculation method is as follows: （1） For the input tensor, , For convolution kernel, , a and b These represent the real and imaginary parts of a complex number, respectively. A and B These represent the real and imaginary parts of the convolution kernel, respectively. i Represents the imaginary unit; The pooling layer is designed using a complex number module. Specifically, the maximum value in the max pooling layer is determined by comparing the complex number moduli and selecting the complex number with the largest moduli as the maximum pooling element to replace the entire region. The average pooling layer sums all the complex numbers in the feature map, then calculates the average of the real and imaginary parts and combines them to obtain the pooling average complex number, which is used to replace the entire region. The batch normalization layer is designed using a complex number module. Specifically, the batch normalization layer in the complex number network calculates the mean and variance of each complex feature separately. Complex activation function uses Its mathematical expression is: （2） in To take the real part of a complex number, This indicates taking the imaginary part of a complex number.

6. The polarization SAR change detection method based on Siamese attention complex convolutional neural network as described in claim 1, characterized in that, The hybrid loss function is: （8） in This represents the Dice coefficient loss. This represents the cross-entropy loss in binary classification. Used to measure the overall similarity between the predicted image and the label image. The similarity between individual pixels is calculated using the following expression: （9） （10） in X and Y They represent two sample sets respectively. express and The number of elements in the intersection between them. and They represent , The number of elements in the middle. Represents pixels k The tag value, This represents a pixel. k The corresponding output label probability.

7. A polarization SAR change detection system based on a Siamese attention complex convolutional neural network, characterized in that, Based on the method described in claim 1, it includes: The image preprocessing module is used to preprocess the acquired PolSAR images from different time phases. Preprocessing includes radiometric calibration, terrain correction, registration, and coherence matrix adjustment. Generation and filtering; The change labeling module is used to obtain change label maps based on the comparison of preprocessed PolSAR images from different time phases. The dataset building module is used to process the preprocessed biphase data. The matrix is ​​taken as an upper triangular matrix, and the imaginary part of the diagonal elements is set to 0. The resulting biphase complex matrix is ​​cropped with the same size as the biphase complex matrix and the corresponding change label map to obtain the sliced ​​dataset. The dataset is then randomly divided into training set and test set. The network construction module is used to build a Siamese attention complex convolutional network based on the UNet encoding and decoding structure. The convolution, pooling, batch normalization and deconvolution in the network are designed with corresponding complex modules. A residual connection is established between the input tensor and the output tensor after network processing. The Siamese network structure is used in the encoder stage to perform difference on the dual temporal multi-scale features extracted in the Siamese encoder stage to generate a multi-scale feature difference map. A complex attention module is designed and embedded into the decoder. The decoder fuses the multi-scale feature difference map for decoding. The training and testing module is used to train the Siamese attention complex convolutional network using the training set and a hybrid loss function to obtain the trained network model. The test set is then input into the trained network model to obtain the polarimetric SAR change detection results.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed, it implements the method as described in any one of claims 1 to 6.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

Images