A SAR image change detection method based on convolution and width learning

Abstract

The application discloses a SAR image change detection method based on convolution and width learning, and builds a convolution and width learning network model, which comprises a cross feature extraction module and a double-width learning module. First and second time phase SAR images are acquired and composed into double time phase SAR images, the double time phase images are input into the cross feature extraction module for processing, and respective corresponding multilayer deep cross features are output. The double time phase SAR images and the respective corresponding multilayer deep cross features are spliced respectively, the cross features after splicing are received and processed by the double-width learning module, and the change map of the double time phase SAR images is predicted. The method combines the deep learning network and the width learning system to perform change detection. The cross feature extraction module can explore the correlation between the double time phase images without training when extracting features, and the double-width learning module can expand the double time state characteristics and analyze the differences, so that the method is compatible with the SAR image change detection task.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image change detection technology, and in particular to a SAR image change detection method based on convolution and width learning. Background Technology

[0002] Change detection utilizes two remote sensing images acquired in the same area to detect changes between two different points in time. Change detection technology provides crucial information about changes on the land surface and therefore has wide applications, such as environmental monitoring, forest protection, and natural disaster assessment. SAR (Synthetic Aperture Radar) is an active system that uses a narrowband antenna to measure the reflected signals from ground objects. Compared to optical images, SAR images avoid interference from sunlight and the atmosphere and can be acquired under all-weather conditions. Therefore, using SAR images for change detection research has become a hot topic in the field of remote sensing.

[0003] However, traditional algorithms extract shallow features through feature engineering, making it difficult to analyze the complex correlations within SAR images. Furthermore, SAR images inevitably suffer from distortion due to factors such as ground scattering characteristics, radar system variables, data processing methods, and meteorological conditions. Deep learning networks typically add feature modules to adapt to the complex problems faced in SAR image change detection, but these modules complicate the structure, increasing optimization time and requiring frequent adjustments in practical applications. Width learning, a machine learning system with a simple structure and robust learning capabilities, suffers from incomplete deep feature extraction because its extended feature weights are randomly initialized and it doesn't consider complex nonlinear representations. Therefore, for tasks with complex correlations between features, width learning systems have limited feature representation capabilities and struggle to detect changes in SAR images with complex information. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing SAR image change detection methods, this invention proposes a SAR image change detection method based on convolution and width learning, which includes the following steps:

[0005] S1. Construct a convolutional and width learning network model, which includes a cross-feature extraction module and a double-width learning module connected in sequence.

[0006] S2. Acquire the first temporal SAR image and the second temporal SAR image and combine them into a dual-temporal SAR image. Input the dual-temporal SAR image into the cross-feature extraction module for processing and output the multi-layer depth cross features corresponding to the first temporal SAR image and the second temporal SAR image, respectively.

[0007] S3. The first-phase SAR image, the second-phase SAR image, and their respective multi-layer depth cross features are stitched together to obtain the cross features corresponding to the first-phase SAR image and the second-phase SAR image, respectively.

[0008] S4. The dual-width learning module receives and processes the cross features corresponding to the first and second temporal SAR images to predict the change map of the dual temporal SAR images.

[0009] Preferably, the cross-feature extraction module in S1 includes several PCConv layers connected in sequence. The specific process of S2 is as follows:

[0010] S21. Input the first temporal SAR image and the second temporal SAR image into the cross-feature extraction module respectively;

[0011] S22. The first layer of several PCConv layers receives and processes the first and second temporal SAR images to obtain the first and second depth cross features of the first layer. The PCA analysis method is used to reduce the dimensionality of the first and second depth cross features of the first layer to obtain the dimensionality-reduced first and second depth cross features of the first layer.

[0012] S23. Other layers in several PCConv layers receive the first and second depth cross features after dimensionality reduction of the previous adjacent layer and process them using the same method as described in S22, thereby obtaining the first and second depth cross features of each layer, and the first and second depth cross features after dimensionality reduction of each layer.

[0013] S24. The last layer in a series of PCConv layers receives the first and second depth cross features after dimensionality reduction from the previous adjacent layer and processes them to obtain the first and second depth cross features of the last layer.

[0014] Preferably, in S22, the first layer of a plurality of PCConv layers receives and processes the first temporal SAR image and the second temporal SAR image to obtain the first depth cross feature and the second depth cross feature of the first layer, specifically including the following steps:

[0015] S221. The first temporal SAR image and the second temporal SAR image received by the first PCConv layer in a plurality of PCConv layers are respectively represented as the first input feature and the second input feature. The difference feature is calculated based on the first input feature and the second input feature. The first input feature, the second input feature and the difference feature are stitched together and converted into LAB mode to obtain the feature of the first PCConv layer in LAB mode.

[0016] S222. Input the features of the first PCConv layer in LAB mode into SLIC to perform pixel clustering and obtain k classes of pixels.

[0017] S223. Preset the size of the pseudo-random feature blocks, and extract k / 2 pseudo-random feature blocks from the first input features and the second input features received from the first PCConv layer, thereby obtaining k pseudo-random feature blocks;

[0018] S224. Randomly select one pixel from each of the k pixel classes, and use the positions of the k randomly selected pixels as the center positions of the k pseudo-random feature blocks.

[0019] S225. Using k pseudo-random feature blocks as convolution kernels, perform multidimensional convolution operations on the first input feature and the second input feature received by the first PCConv layer respectively, generating k first convolution features and second convolution features accordingly. Concatenate the k first convolution features and the second convolution features respectively to obtain the first depth cross feature and the second depth cross feature of the first PCConv layer.

[0020] Preferably, the features of the first PCConv layer in S221 in LAB mode can be specifically expressed by the following formula:

[0021]

[0022] in,

[0023] In the formula, This represents the features in LAB mode, and C(.) represents the concatenation function. Indicates the first input feature. Indicates the second input feature, The difference features between the first input features and the second input features are represented by T1, which is the first temporal SAR image, and T2 is the second temporal SAR image.

[0024] Preferably, the first depth cross feature and the second depth cross feature of the first PCConv layer in S225 can be specifically represented as follows:

[0025]

[0026]

[0027] in,

[0028]

[0029] In the formula, This represents the first depth cross feature of the first PCConv layer. The second depth cross feature of the first PCConv layer. For the l-th first convolutional feature, Let θ be the l-th second convolutional feature, and θ be the non-linear activation function. It is a convolution function. For the j-th dimension, the first input feature, For the j-th dimension, the second input feature, Let b be the kernel number of the l-th convolutional kernel in the j-th dimension, where j = 1, 2, ..., b, and b is the first input feature. Second input features The dimension of , k is the number of the first convolutional features and the second convolutional features.

[0030] Preferably, S3 can be expressed by the following formula:

[0031]

[0032]

[0033] In the formula, T1 is the first temporal SAR image, and T2 is the second temporal SAR image. This represents the first depth cross feature of the iPCConv layer. This represents the second deep cross feature of the i-th PCConv layer, where i = 1, 2, ..., r, and r is the total number of pseudo-random cross-convolutional layers. F t1 F represents the first cross feature corresponding to the first temporal SAR image. t2 This represents the second cross feature corresponding to the second temporal SAR image.

[0034] Preferably, the double-width learning module in S1 includes a double-width node generation unit and an output layer connected in sequence, and S4 specifically includes the following steps:

[0035] S41. Divide the first cross feature corresponding to the first temporal SAR image and the second cross feature corresponding to the second temporal SAR image into blocks respectively to obtain the first input data and the second input data accordingly.

[0036] S42. Input the first input data and the second input data into the dual-width node generation unit for processing, and output differential sparse feature nodes and differential enhancement nodes;

[0037] S43. Preset the real labels, and optimize the weights of the output layer by setting the optimization model based on the differential sparse feature nodes, differential enhancement nodes and real labels to obtain the output layer with optimized weights.

[0038] S44. Input the differential sparse feature nodes and differential enhancement nodes into the output layer after weight optimization to predict the change map of the dual-temporal SAR image.

[0039] Preferably, S42 specifically includes the following:

[0040] S421. Randomly set the first weight and the first bias, and calculate the first feature node corresponding to the first input data based on the first weight and the first bias;

[0041] S422. Set a preset number of iterations, input the first feature node and the first input data into the sparse encoder of the dual-width node generation unit, and use an alternating direction multiplier (ADMM) to iteratively optimize the sparse weights and biases within the preset number of iterations to obtain the optimized sparse weights and biases.

[0042] S423. Obtain the first sparse feature based on the first input data, the optimized sparse weights and biases, and obtain the second sparse feature based on the second input data, the optimized sparse weights and biases.

[0043] S424. Randomly set the second weight and the second bias, and use the second weight and the second bias to map the first sparse feature and the second sparse feature respectively, and generate the first enhancement node and the second enhancement node accordingly.

[0044] S425. The first sparse feature and the second sparse feature are analyzed by CVA analysis to obtain differential sparse feature nodes. The first enhancement node and the second enhancement node are analyzed by CVA analysis to obtain differential enhancement nodes.

[0045] Preferably, in S422, an alternating direction multiplier (ADMM) is used to iteratively optimize the sparse weights and biases within a preset number of iterations. The iterative optimization model can be specifically expressed as:

[0046]

[0047] in:

[0048] g(o) = η|o|1

[0049] In the formula, w represents the combination of sparse weights and sparse biases, w = [W p ,β p ], W p For sparse weights, β p For sparse bias, η is the convergence constraint coefficient, X t1 B is the first input data. t1 Let be the first feature node, o represent the parameter that restricts sparsity, and g(o) represent the function that restricts sparsity.

[0050] Preferably, in step S43, the weights of the output layer are optimized based on the differential sparse feature nodes, differential enhancement nodes, and the true label. The optimization model can be specifically expressed as follows:

[0051]

[0052] In the formula, Z d H represents the differential sparse feature nodes. d λ represents the differential augmentation node, λ represents the constraint parameter of the sum of squared weights, W represents the weights of the output layer, and Y represents the true label.

[0053] The aforementioned SAR image change detection method based on convolution and width learning constructs a convolution and width learning network model. This network model includes a cross-feature extraction module and a dual-width learning module connected sequentially. It acquires dual-temporal SAR images, inputs them into the cross-feature extraction module for processing, and outputs multi-layer depth cross-features corresponding to each of the dual-temporal SAR images. The dual-temporal SAR images and their corresponding multi-layer depth cross-features are then concatenated to obtain the respective cross-features of each dual-temporal SAR image. The dual-width learning module receives and processes the respective cross-features of the dual-temporal SAR images to predict the change map of the dual-temporal SAR images. This method proposes a unified framework for SAR image change detection, combining deep learning networks and width learning systems for change detection. The cross-feature extraction module extracts features without training, explores the correlation between dual-temporal images, and the dual-width learning module generates dual time nodes and analyzes their differences, making it compatible with SAR image change detection tasks. Attached Figure Description

[0054] Figure 1 This is a flowchart of a SAR image change detection method based on convolution and width learning in one embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of the overall structure of a convolutional and width learning network model in one embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the working process of a certain PCConv layer in the cross feature extraction module in one embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the working process of the dual-width learning module in one embodiment of the present invention. Detailed Implementation

[0058] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0059] See Figure 1 and Figure 2 , Figure 1 This is a flowchart of a SAR image change detection method based on convolution and width learning in one embodiment of the present invention. Figure 2 This is a schematic diagram of the overall structure of a convolutional and width learning network model in one embodiment of the present invention.

[0060] A SAR image change detection method based on convolution and width learning includes the following steps:

[0061] S1. Construct a convolutional and width learning network model, which includes a cross-feature extraction module and a double-width learning module connected in sequence.

[0062] S2. Acquire the first-phase SAR image and the second-phase SAR image and combine them into a dual-phase SAR image. Input the dual-phase SAR image into the cross-feature extraction module for processing, and output the multi-layer depth cross features corresponding to the first-phase SAR image and the second-phase SAR image respectively. and

[0063] S3. Concatenate the first-phase SAR image, the second-phase SAR image, and their corresponding multi-layer depth cross features to obtain the cross features F corresponding to the first-phase SAR image and the second-phase SAR image, respectively. t1 and F t2 ;

[0064] S4, The dual-width learning module receives the cross-features F corresponding to the first and second temporal SAR images, respectively. t1 and F t2 The data is then processed to predict the changes in the dual-temporal SAR images.

[0065] Specifically, the dual-temporal SAR image includes a first-temporal SAR image and a second-temporal SAR image. A cross-feature extraction module processes the dual-temporal SAR image to obtain multi-level deep cross-features corresponding to the dual-temporal SAR image. A dual-width learning module receives and processes the multi-level deep cross-features to obtain a change map of the dual-temporal SAR image. The cross-feature extraction module utilizes a multi-layer pseudo-random cross-convolutional layer (PCConv) to obtain multi-level deep cross-features of the dual-temporal image without any training process. The dual-width learning module receives the multi-level deep cross-features of the dual-temporal image and performs predictions of whether they are variable or invariant, enhancing the nonlinearity and robustness of the multi-level deep cross-features of the dual-temporal image.

[0066] In one embodiment, the cross-feature extraction module in S1 includes several PCConv layers connected in sequence. The specific process in S2 is as follows:

[0067] S21. Input the first temporal SAR image and the second temporal SAR image into the cross-feature extraction module respectively;

[0068] S22. The first layer of several PCConv layers receives and processes the first and second temporal SAR images to obtain the first depth cross features and the second depth cross features of the first layer. The PCA analysis method is used to perform dimensionality reduction processing on the first depth cross features and the second depth cross features of the first layer to obtain the dimensionality-reduced first depth cross features and the second depth cross features of the first layer.

[0069] Specifically, in S22, the first layer of several PCConv layers receives and processes the first temporal SAR image and the second temporal SAR image, thereby obtaining the first depth cross feature and the second depth cross feature of the first layer, which can be expressed by the formula:

[0070]

[0071]

[0072] In the formula, T1 and T2 are the first and second temporal SAR images, respectively, and PCConv1(.) represents the first pseudo-random cross-convolutional layer. and These are the first depth cross features and the second depth cross features of the first PCConv layer, respectively.

[0073] In S22, PCA analysis is used to reduce the dimensionality of the first and second depth cross features of the first layer, respectively, to obtain the dimensionality-reduced first and second depth cross features of the first layer, which can be expressed by the following formula:

[0074]

[0075]

[0076] In the formula, This refers to the first depth cross feature after dimensionality reduction of the first PCConv layer (i.e., the first depth cross feature after dimensionality reduction of the first layer). This is the second deep cross feature after dimensionality reduction of the first PCConv layer (that is, the second deep cross feature after dimensionality reduction of the first layer).

[0077] S23. Other layers in several PCConv layers receive the first and second depth cross features after dimensionality reduction from the previous adjacent layer and process them using the same method as described in S22, thereby obtaining the first and second depth cross features of each layer, and the first and second depth cross features after dimensionality reduction of each layer.

[0078] S24. The last layer in a series of PCConv layers receives and processes the first and second depth cross features after dimensionality reduction from the previous adjacent layer to obtain the first and second depth cross features of the last layer. Specifically, this can be represented as:

[0079]

[0080]

[0081] In the formula, This represents the first depth-cross feature of the r-th layer (the last layer). PCConv is the second depth cross feature of the r-th layer (i.e., the last layer). r (.) represents the r-th pseudo-random cross-convolutional layer. This is the first depth cross feature after dimensionality reduction at layer (r-1), which is also the first input feature at layer r. This is the second depth cross feature after dimensionality reduction at layer (r-1), which is also the second input feature at layer r, where r represents the total number of PCConv layers.

[0082] Specifically, the cross-feature extraction module receives the first and second temporal SAR images and extracts features layer by layer through several PCConv layers. PCA analysis is then used for feature dimensionality reduction to obtain the first and second deep cross features for each layer, as well as the dimensionality-reduced first and second deep cross features for each layer. It should be noted that the input to the first PCConv layer is the first and second temporal SAR images. The last PCConv layer receives the dimensionality-reduced first and second deep cross features from the previous adjacent PCConv layer and performs only feature extraction; it does not perform further dimensionality reduction.

[0083] In one embodiment, the first PCConv layer in S22 receives and processes a first temporal SAR image and a second temporal SAR image to obtain the first depth cross feature and the second depth cross feature of the first layer, specifically including the following steps:

[0084] S221. The first temporal SAR image and the second temporal SAR image received by the first PCConv layer in a plurality of PCConv layers are respectively represented as the first input feature and the second input feature. The difference feature is calculated based on the first input feature and the second input feature. The first input feature, the second input feature and the difference feature are concatenated together and converted into LAB mode to obtain the feature of the first PCConv layer in LAB mode.

[0085] The LAB color model is a color model established based on an international standard for color measurement developed by the Commission International e Eclairage (CIE) in 1931. It was improved and named in 1976. The LAB color model consists of three elements: lightness (L), and two color channels, A and B. A includes colors ranging from dark green (low lightness value) to gray (medium lightness value) to bright pink (high lightness value); B includes colors ranging from bright blue (low lightness value) to gray (medium lightness value) to yellow (high lightness value).

[0086] Furthermore, the specific features of the first PCConv layer in S221 under LAB mode can be expressed by the following formula:

[0087]

[0088] in,

[0089] In the formula, This represents the features in LAB mode, and C(.) represents the concatenation function. Indicates the first input feature. Indicates the second input feature, The difference features between the first input features and the second input features are represented by T1, which is the first temporal SAR image, and T2 is the second temporal SAR image.

[0090] S222. Input the features of the first PCConv layer in LAB mode into SLIC to perform pixel clustering and obtain k classes of pixels.

[0091] Specifically, the features of the first PCConv layer in LAB mode The input is fed into SLIC for pixel clustering. SLIC does not use a forced label connection module to reduce boundary adhesion and increase similarity within clusters. SLIC clusters pixels based on color distance d. c and spatial distance d s Calculate the distance D. The settings between each pixel point require calculating the distance D, and then divide the pixels into k classes based on the distance D:

[0092]

[0093] Where m is the compactness parameter, N is the number of pixels, D is the distance, and d c For color distance, d s denoted as spatial distance, and k as the number of pixel categories.

[0094] S223. Preset the size of the pseudo-random feature blocks, and extract k / 2 pseudo-random feature blocks from the first input features and the second input features received from the first PCConv layer, thereby obtaining k pseudo-random feature blocks.

[0095] Specifically, the size of the pseudo-random feature block is preset to s×s, starting from the first feature... Extract k / 2 pseudo-random feature blocks of size s×s from the second feature. Extract k / 2 pseudo-random feature blocks of size s×s, thus obtaining a total of k pseudo-random feature blocks B of size s×s.

[0096] S224. Randomly select one pixel from each of the k pixel classes, and use the positions of the k randomly selected pixels as the center positions of the k pseudo-random feature blocks. That is, the center position of the k pseudo-random feature blocks B is in the first feature block. Second feature The position of the selected k pixels is the same.

[0097] S225. Using k pseudo-random feature blocks as convolution kernels, perform multidimensional convolution operations on the first input feature and the second input feature received by the first PCConv layer respectively, generating k first convolution features and second convolution features accordingly. Concatenate the k first convolution features and the second convolution features respectively to obtain the first depth cross feature and the second depth cross feature of the first PCConv layer.

[0098] Specifically, k pseudo-random feature blocks B are used as convolution kernels and shared. Multidimensional convolution operations are performed on the first input features and the second input features of the first PCConv layer, respectively, to generate k first convolution features and second convolution features accordingly:

[0099]

[0100]

[0101] In the formula, For the l-th first convolutional feature, Let θ be the l-th second convolutional feature, and θ be the non-linear activation function. It is a convolution function. The first input feature in the j-th dimension For the j-th dimension, the second input feature Let b be the kernel number of the l-th convolutional kernel in the j-th dimension, and b be the first input feature. Second input features The dimension of the convolution is j = 1, 2, ..., b, where k is the total number of convolutional features, which also corresponds to the total number of pseudo-random feature blocks B and the total number of convolutional kernels, respectively.

[0102] The k first convolutional features and second convolutional features are concatenated to obtain the first depth cross feature and the second depth cross feature of the first PCConv layer:

[0103]

[0104]

[0105] In the formula, This represents the first depth cross feature of the first PCConv layer. The second depth cross feature of the first PCConv layer. For the l-th first convolutional feature, Let l be the l-th second convolutional feature, and C(.) denote the concatenation function.

[0106] See Figure 3 , Figure 3 This is a schematic diagram of the working process of a certain PCConv layer in the cross feature extraction module in one embodiment of the present invention.

[0107] For ease of description, assume the cross-feature extraction module comprises r PCConv layers connected sequentially. The first and second depth cross features received by each PCConv layer after dimensionality reduction from the previous adjacent PCConv layer are both represented as the first input feature. Second input features For example, when the current layer is the first PCConv layer, the first PCConv layer directly uses the first temporal SAR image T1 and the second temporal SAR image T2 as the first input features. Second input features When the current layer is the second PCConv layer, the second PCConv layer receives the first depth cross feature after dimensionality reduction from the first PCConv layer. Second depth cross features And and As the first input feature of the current layer Second input features When the current layer is the third PCConv layer, the third PCConv layer receives the first depth cross feature after dimensionality reduction from the second PCConv layer. Second depth cross features And and As the first input feature of the current layer Second input features ...Through the above operations, each PCConv layer receives the first input feature... Second input features It can be represented as the following set:

[0108] Calculate the first feature of the input to each PCConv layer. Second feature Differences between characteristics The first feature Second feature and their differences Connected together and converted into features under the International Declaration Labs color mode (LAB). The characteristics of each PCConv layer in LAB mode can be expressed by the following formula:

[0109]

[0110] in,

[0111] In the formula, Indicates features in LAB mode, Indicates the first input feature. Indicates the second input feature, Let C represent the difference features between the first input features and the second input features, C represent the stitching function, T1 is the first temporal SAR image, and T2 is the second temporal SAR image. The first depth cross feature after dimensionality reduction of the (r-1)th PCConv layer. is the second depth cross feature after dimensionality reduction of the (r-1)th PCConv layer, where r is the total number of PCConv layers.

[0112] The features of each PCConv layer in LAB mode are input into SLIC for pixel clustering to obtain k classes of pixels. K / 2 pseudo-random feature blocks of a preset size are extracted from the first and second input features of each PCConv layer, resulting in k pseudo-random feature blocks. A pixel is randomly selected from each of the k classes of pixels, and the location of the k randomly selected pixels is used as the center position of the k pseudo-random feature blocks B. These k pseudo-random feature blocks are used as k convolution kernels to perform multi-dimensional convolution operations on the first and second input features received by each PCConv layer, generating k first and second convolution features. The k first and second convolution features are then concatenated to obtain the first and second depth cross features of each PCConv layer.

[0113] In one embodiment, S3 can be specifically expressed by the formula:

[0114]

[0115]

[0116] In the formula, T1 is the first temporal SAR image, and T2 is the second temporal SAR image. This represents the first depth cross feature of the iPCConv layer. This represents the second deep cross feature of the i-th PCConv layer, where i = 1, 2, ..., r, and r is the total number of pseudo-random cross-convolutional layers. F t1 F represents the first cross feature corresponding to the first temporal SAR image. t2 This represents the second cross feature corresponding to the second temporal SAR image.

[0117] Specifically, the first temporal SAR image T1 and the corresponding r-layer first depth cross-feature are stitched together to obtain the first cross-feature F corresponding to the first temporal SAR image. t1 The original image T2 of the second time phase and the corresponding second depth cross features of the r layer are concatenated to obtain the second cross features F corresponding to the second time phase SAR image. t2 .

[0118] In one embodiment, see Figure 4 The double-width learning module in S1 includes a double-width node generation unit and an output layer connected in sequence. S4 specifically includes the following steps:

[0119] S41, the first cross feature F corresponding to the first temporal SAR image. t1 The second cross feature F corresponding to the second temporal SAR image t2 Divide the data into blocks to obtain the first input data X. t1Second input data X t2 .

[0120] S42, Transfer the first input data X t1 Second input data X t2 Input is processed by a double-width node generation unit, and output is differential sparse feature nodes Z. d and differential enhancement node H d .

[0121] Furthermore, S42 specifically includes the following:

[0122] S421. Randomly set the first weight and the first bias, and calculate the first input data X based on the first weight and the first bias. t1 The corresponding first feature node B t1 .

[0123] Specifically, the first feature node B t1 It can be represented as Among them, the p-th feature node Can

[0124] By randomly setting the first weight W b and the first bias β b To calculate:

[0125]

[0126] in, For the first feature node B t1 The p-th component, φ is a nonlinear activation function, W bp The first weight W b The p-th component, β bp For the first bias β b The p-th component, where n is the number of feature nodes.

[0127] S422, Preset the number of iterations, and set the first feature node B. t1 and the first input data X t1 In the input sparse encoder, an alternating direction multiplier (ADMM) is used to iteratively optimize the sparse weights and biases within a preset number of iterations to obtain the optimized sparse weights and biases.

[0128] Specifically, the first feature node and the first input data X t1 The data is fed into a sparse encoder to obtain sparse feature nodes. In the sparse encoder, an alternating direction multiplier (ADMM) is used to iteratively optimize the sparse weights and biases.

[0129] Furthermore, the iterative optimization model can be specifically expressed as:

[0130]

[0131] in:

[0132] g(o) = η|o|1

[0133] In the formula, w represents the combination of sparse weights and sparse biases, w = [W p ,β p ], W p For sparse weights, β p For sparse bias, η is the convergence constraint coefficient, X t1 B is the first input data. t1 η is the first feature node, o represents the parameter for restricting sparsity, and g(o) represents the function for restricting sparsity. In this embodiment, η can be specifically set to 0.001.

[0134] The above iterative optimization model can be solved through the following iterative steps:

[0135]

[0136] In the formula, c is the number of iterations, ρ represents the convergence constraint coefficient step size, and ρ > 0. For example, if the preset number of iterations c is 50, the optimized sparse weights and sparse biases are obtained after 50 iterations.

[0137] S423, Based on the first input data X t1 The optimized sparse weights and biases yield the first sparse feature Z. t1 According to the second input data X t2 The optimized sparse weights and biases yield the second sparse feature Z. t2 .

[0138] Specifically, based on the first input data X t1 The optimized sparse weights and biases yield the first sparse feature Z. t1 ,

[0139] Specifically, it can be expressed as:

[0140]

[0141] In the formula, Z represents the first sparse feature. t1 The e-th component, W pe Represents the sparse weight W p The e-th component, β pe Indicates sparse bias β p The e-th component.

[0142] Using the same method, we can determine the second input data X. t2 The optimized sparse weights and biases yield the second sparse feature Z. t2 This will not be elaborated upon here.

[0143] S424. Randomly set the second weight and bias, and use the second weight and bias to apply them to the first sparse feature Z. t1 Second sparse feature Z t2 Perform mapping to generate the first enhanced node H. t1 Second Enhancement Node H t2 .

[0144] Specifically, the second weight W is randomly set. g Second bias β g The first sparse feature Z is applied using a second weight and bias. t1 Perform mapping to generate the first enhanced node H. t1 ,

[0145]

[0146] In the formula, For the first enhancement node H t1 The s-th component, σ is the nonlinear activation function, q is the number of boosting nodes, and W gs The second weight W g The s-th component, β gs For the second bias β g The s-th component.

[0147] Using the same method, a second weight and bias can be applied to the second sparse feature Z. t2 Perform mapping to generate the second enhanced node H. t2 This will not be elaborated upon here.

[0148] S425. The first sparse feature Z is analyzed using the CVA method. t1 Second sparse feature Z t2 Analysis was performed to obtain the differential sparse feature nodes Z. d The first enhancement node H was analyzed using the CVA method. t1 Second Enhancement Node H t2 Analysis was performed to obtain the differential enhancement node H. d Specifically, it can be expressed as follows:

[0149] Z d =CVA(Z) t1 Z t2 )

[0150] H d =CVA(Ht1 H t2 )

[0151] In the formula, Z d For differential sparse feature nodes, Z t1 Z is the first sparsity feature. t2 As the second sparsity feature, H d For differential enhancement nodes, H t1 H is the first enhancement node. t2 This is the second enhancement node.

[0152] S43. Preset real labels, and then use the differential sparse feature nodes Z... d Differential enhancement node H d The model is optimized by setting the real labels, and the weights of the output layer are optimized to obtain the output layer with optimized weights.

[0153] Specifically, to minimize the error between the output layer's output and the pre-set true label Y, the weights of the output layer need to be optimized. Further, the optimization model in S43 can be expressed as:

[0154]

[0155] In the formula, Z d H represents the differential sparse feature nodes. d λ represents the differential augmentation node, λ represents the constraint parameter of the sum of squared weights, W represents the weights of the output layer, and Y represents the true label.

[0156] S44. Difference sparse feature nodes Z d and differential enhancement node H d The output layer with optimized weights is used to predict the changes in dual-temporal SAR images. Specifically, it can be expressed as:

[0157]

[0158] In the formula, Z represents the predicted changes in the dual-temporal SAR image. d H represents the differential sparse feature nodes. d This represents a differential enhancement node.

[0159] Since the classic BLS (Broad Learning System) lacks the ability to extend dual-temporal features and is incompatible with the change detection task in this technical solution, a dual-width learning module was designed, including a dual-width node generation unit and an output layer connected in sequence. This is to alleviate the burden on the first input data X. t1Due to the randomness and unpredictability of the data, sparse coding is used to create a simpler representation that is easier to capture internal information. The dual-width node generation unit shares sparse weights and biases, and sparse feature nodes and enhancement nodes are obtained in the dual branches respectively.

[0160] The aforementioned SAR image change detection method based on convolution and width learning constructs a convolution and width learning network model. This network model includes a cross-feature extraction module and a dual-width learning module connected sequentially. It acquires dual-temporal SAR images, inputs them into the cross-feature extraction module for processing, and outputs multi-layer depth cross-features corresponding to each of the dual-temporal SAR images. The dual-temporal SAR images and their corresponding multi-layer depth cross-features are then concatenated to obtain the respective cross-features of each dual-temporal SAR image. The dual-width learning module receives and processes the respective cross-features of the dual-temporal SAR images to predict the change map of the dual-temporal SAR images. This method proposes a unified framework for SAR image change detection, combining deep learning networks and width learning systems for change detection. The cross-feature extraction module extracts features without training, explores the correlation between dual-temporal images, and the dual-width learning module generates dual time nodes and analyzes their differences, making it compatible with SAR image change detection tasks.

[0161] The above provides a detailed description of a SAR image change detection method based on convolution and width learning provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the core ideas of this invention. It should be noted that those skilled in the art can make various improvements and modifications to this invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. A SAR image change detection method based on convolution and width learning, characterized in that, The method includes: S1. Construct a convolutional and width learning network model, which includes a cross-feature extraction module and a double-width learning module connected in sequence. S2. Acquire the first temporal SAR image and the second temporal SAR image and combine them into a dual-temporal SAR image. Input the dual-temporal SAR image into the cross-feature extraction module for processing and output the multi-layer depth cross features corresponding to the first temporal SAR image and the second temporal SAR image, respectively. S3. The first-phase SAR image, the second-phase SAR image, and their respective multi-layer depth cross features are stitched together to obtain the cross features corresponding to the first-phase SAR image and the second-phase SAR image, respectively. S4. The dual-width learning module receives and processes the cross features corresponding to the first and second temporal SAR images to predict the change map of the dual temporal SAR images.

2. The SAR image change detection method based on convolution and width learning as described in claim 1, characterized in that, The cross-feature extraction module in S1 consists of several PCConv layers connected in sequence. The specific process of S2 is as follows: S21. Input the first temporal SAR image and the second temporal SAR image into the cross-feature extraction module respectively; S22. The first layer of several PCConv layers receives and processes the first and second temporal SAR images to obtain the first and second depth cross features of the first layer. The PCA analysis method is used to reduce the dimensionality of the first and second depth cross features of the first layer to obtain the dimensionality-reduced first and second depth cross features of the first layer. S23. Other layers in several PCConv layers receive the first and second depth cross features after dimensionality reduction of the previous adjacent layer and process them using the same method as described in S22, thereby obtaining the first and second depth cross features of each layer, and the first and second depth cross features after dimensionality reduction of each layer. S24. The last layer in a series of PCConv layers receives the first and second depth cross features after dimensionality reduction from the previous adjacent layer and processes them to obtain the first and second depth cross features of the last layer.

3. The SAR image change detection method based on convolution and width learning as described in claim 2, characterized in that, In S22, the first layer of several PCConv layers receives and processes a first temporal SAR image and a second temporal SAR image to obtain the first depth cross feature and the second depth cross feature of the first layer. The specific steps include the following: S221. The first temporal SAR image and the second temporal SAR image received by the first PCConv layer in a plurality of PCConv layers are respectively represented as the first input feature and the second input feature. The difference feature is calculated based on the first input feature and the second input feature. The first input feature, the second input feature and the difference feature are stitched together and converted into LAB mode to obtain the feature of the first PCConv layer in LAB mode. S222. Input the features of the first PCConv layer in LAB mode into SLIC to perform pixel clustering and obtain k classes of pixels; S223. Preset the size of the pseudo-random feature blocks, and extract k / 2 pseudo-random feature blocks from the first input features and the second input features received from the first PCConv layer, thereby obtaining k pseudo-random feature blocks; S224. Randomly select one pixel from each of the k pixel classes, and use the positions of the k randomly selected pixels as the center positions of the k pseudo-random feature blocks. S225. Using k pseudo-random feature blocks as convolution kernels, perform multidimensional convolution operations on the first input feature and the second input feature received by the first PCConv layer respectively, generating k first convolution features and second convolution features accordingly. Concatenate the k first convolution features and the second convolution features respectively to obtain the first depth cross feature and the second depth cross feature of the first PCConv layer.

4. The SAR image change detection method based on convolution and width learning as described in claim 3, characterized in that, The specific features of the first PCConv layer in S221 under LAB mode can be expressed by the following formula: in, In the formula, This represents the features in LAB mode, and C(.) represents the concatenation function. Indicates the first input feature. Indicates the second input feature, The difference features between the first input features and the second input features are represented by T1, which is the first temporal SAR image, and T2 is the second temporal SAR image.

5. The SAR image change detection method based on convolution and width learning as described in claim 4, characterized in that, The first depth cross feature and the second depth cross feature of the first PCConv layer in S225 can be specifically represented as follows: in, In the formula, F1 t1 For the first depth cross feature of the first PCConv layer, F1 t2 This represents the second depth cross feature of the first PCConv layer. For the l-th first convolutional feature, Let θ be the l-th second convolutional feature, and θ be the non-linear activation function. It is a convolution function. For the j-th dimension, the first input feature, For the j-th dimension, the second input feature, Let b be the kernel number of the l-th convolutional kernel in the j-th dimension, where j = 1, 2, ..., b, and b is the first input feature. Second input features The dimension of , k is the number of the first convolutional features and the second convolutional features.

6. In the SAR image change detection method based on convolution and width learning as described in claim 5, S3 can be specifically expressed by the formula: In the formula, T1 is the first temporal SAR image, T2 is the second temporal SAR image, and F i t1 F represents the first depth cross feature of the iPCConv layer. i t2 This represents the second deep cross feature of the i-th PCConv layer, where i = 1, 2, ..., r, and r is the total number of pseudo-random cross-convolutional layers. F t1 F represents the first cross feature corresponding to the first temporal SAR image. t2 This represents the second cross feature corresponding to the second temporal SAR image.

7. The SAR image change detection method based on convolution and width learning as described in claim 6, characterized in that, The double-width learning module in S1 includes a double-width node generation unit and an output layer connected in sequence. S4 specifically includes the following steps: S41. Divide the first cross feature corresponding to the first temporal SAR image and the second cross feature corresponding to the second temporal SAR image into blocks respectively to obtain the first input data and the second input data accordingly. S42. Input the first input data and the second input data into the dual-width node generation unit for processing, and output differential sparse feature nodes and differential enhancement nodes; S43. Preset the real labels, and optimize the weights of the output layer by setting the optimization model based on the differential sparse feature nodes, differential enhancement nodes and real labels to obtain the output layer with optimized weights. S44. Input the differential sparse feature nodes and differential enhancement nodes into the output layer after weight optimization to predict the change map of the dual-temporal SAR image.

8. The SAR image change detection method based on convolution and width learning as described in claim 7, characterized in that, S42 specifically includes the following: S421. Randomly set the first weight and the first bias, and calculate the first feature node corresponding to the first input data based on the first weight and the first bias; S422. Set a preset number of iterations, input the first feature node and the first input data into the sparse encoder of the dual-width node generation unit, and use an alternating direction multiplier (ADMM) to iteratively optimize the sparse weights and biases within the preset number of iterations to obtain the optimized sparse weights and biases. S423. Obtain the first sparse feature based on the first input data, the optimized sparse weights and biases, and obtain the second sparse feature based on the second input data, the optimized sparse weights and biases. S424. Randomly set the second weight and the second bias, and use the second weight and the second bias to map the first sparse feature and the second sparse feature respectively, and generate the first enhancement node and the second enhancement node accordingly. S425. The first sparse feature and the second sparse feature are analyzed by CVA analysis to obtain differential sparse feature nodes. The first enhancement node and the second enhancement node are analyzed by CVA analysis to obtain differential enhancement nodes.

9. The SAR image change detection method based on convolution and width learning as described in claim 8, characterized in that, In S422, an alternating direction multiplier (ADMM) is used to iteratively optimize the sparse weights and biases within a preset number of iterations. The iterative optimization model can be specifically expressed as: in: g(o) = η|o|1 In the formula, w represents the combination of sparse weights and sparse biases, w = [W p ,β p ], W p For sparse weights, β p For sparse bias, η is the convergence constraint coefficient, X t1 B is the first input data. t1 Let be the first feature node, o represent the parameter that restricts sparsity, and g(o) represent the function that restricts sparsity.

10. The SAR image change detection method based on convolution and width learning as described in claim 7, characterized in that, In S43, the weights of the output layer are optimized based on the differential sparse feature nodes, differential enhancement nodes, and the true label. The optimization model can be specifically represented as follows: In the formula, Z d H represents the differential sparse feature nodes. d λ represents the differential augmentation node, λ represents the constraint parameter of the sum of squared weights, W represents the weights of the output layer, and Y represents the true label.

Images