A SAR Image Change Detection Method with Collaborative Optimization of Network Parameters and Structure

Abstract

The present invention discloses a SAR image change detection method for collaborative optimization of network parameters and structure, including: obtaining denoised SAR images from two SAR images of the same location at different times; generating a similarity matrix by analyzing pixel neighborhood information based on neighborhood volatility; segmenting the similarity matrix to obtain a pseudo-label matrix; generating a sample set and selecting a training sample set; setting the convolutional neural network structure to traverse the space, using a multi-objective optimization algorithm to collaboratively optimize the network structure and weight parameters with the upper and lower significant surrogate models, seeking the optimal classification network model to obtain a set of Pareto fronts; selecting the solution at the inflection point to obtain the convolutional neural network structure and related parameters representing the optimal performance, retraining to obtain a change detection classification network; and inputting the sample set to obtain the change detection result. The present invention does not generate a difference map, introduces the idea of co-evolution to automatically and efficiently seek the optimal network structure and parameter configuration, can improve the detection accuracy, and reduce the computational complexity.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image change detection, and particularly relates to a SAR image change detection method for collaborative optimization of network parameters and structure. Background Art

[0002] Change detection is to quantitatively analyze and determine surface change characteristics and change processes from remote sensing data at the same location in different periods. Synthetic Aperture Radar (SAR) has the characteristics of all-weather, all-day, and high resolution, and has the advantage of being unaffected by weather compared with visible light, infrared sensors, etc., so it has been widely used in the field of remote sensing. The change detection technology of SAR images has been widely used in fields such as land surveying, disaster assessment, and building planning. Due to its special imaging mechanism, the inherent speckle noise of SAR images will affect the change detection process.

[0003] Traditional SAR image change detection techniques usually include three steps: preprocessing, generating a difference map, and analyzing the difference map. The preprocessing step mainly includes registration, geometric correction, and image denoising, etc. The step of generating a difference map is to compare two images at different times and generate a difference map. In the difference map, the pixels in the changed area and the pixels in the unchanged area will show obvious differences in gray scale. The step of analyzing the difference map is to analyze the obtained difference map, extract change information, and finally obtain the changed area and the unchanged area of the two images. Among them, the generation and analysis of the difference map are important steps in the SAR image transformation detection technology, and the quality of the difference map will directly affect the performance of change detection.

[0004] With the rapid development of deep learning technology, more and more change detection methods attempt to learn the feature information of SAR images by training deep neural networks, and classify the image feature information through a classifier to obtain the final change result. These methods can, to a certain extent, suppress the interference of speckle noise and improve the accuracy of change detection. For example, a Chinese patent with the publication number CN 111339827A and the title "SAR Image Change Detection Method Based on Multi-Region Convolutional Neural Network" discloses a change detection method. This method performs difference analysis on two SAR images to obtain a difference image; pre-classifies the difference image to obtain a constructed training data set and a test data set; sends the sample training data set into the proposed multi-region convolutional neural network for training; uses the trained network to test the test set, and then obtains the change detection result of the entire co-located multi-temporal SAR image. However, in such change detection methods based on deep neural networks, the structure and parameter settings of the neural network are both set based on empiricism, or through continuous manual attempts to select the optimal number of network layers, number of nodes, connection methods, and related parameters in simulation experiments to determine the final network. However, the artificially designed network framework not only has complex parameter tuning and serious time consumption, but also is difficult to find the architecture and parameter settings with the best performance, so it will have a certain adverse impact on the accuracy of change detection. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a SAR image change detection method for collaborative optimization of network parameters and structure. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] A SAR image change detection method for collaborative optimization of network parameters and structure, comprising:

[0007] Perform denoising processing on two SAR images obtained at different times at the same location to obtain denoised SAR images respectively;

[0008] Based on neighborhood volatility, perform pixel neighborhood information analysis on the two denoised SAR images to generate a similarity matrix;

[0009] Use a preset threshold segmentation algorithm to segment the similarity matrix to obtain a pseudo-label matrix; wherein, the element values in the pseudo-label matrix are 0 and 1, representing the unchanged class and the changed class in the two denoised SAR images respectively;

[0010] Use the two denoised SAR images to obtain a sample set, and select a part from the sample set as a training sample set using the pseudo-label matrix;

[0011] Set the structural traversal space of the convolutional neural network for change detection classification. Using the preset multi-objective optimization algorithm, adopt the upper and lower significant surrogate models to optimize the network structure and network weight parameters respectively, seek the optimal classification network model, and obtain a set of Pareto frontiers with high detection accuracy and low computational complexity as the solution set;

[0012] Select the solution at the inflection point of the Pareto frontier as the final solution, decode to obtain the convolutional neural network structure and related parameters representing the optimal performance, and use the obtained convolutional neural network structure and the training sample set to retrain the network to obtain the trained change detection classification network;

[0013] Input the sample set into the change detection classification network to obtain the change detection results of the two SAR images.

[0014] In an embodiment of the present invention, the pixel neighborhood information of the two denoised SAR images is analyzed based on neighborhood volatility to generate a similarity matrix, including:

[0015] Taking each pixel as the central pixel, using a rectangular sliding window of a preset size, traverse and calculate the neighborhood heterogeneity value for the two denoised SAR images at each pixel position to obtain the neighborhood heterogeneity function of the two denoised SAR images;

[0016] Construct a neighborhood dark pixel similarity function and a neighborhood bright pixel similarity function using the two denoised SAR images; wherein, the neighborhood dark pixel similarity function represents the proximity between the minimum central pixel gray value and the minimum neighborhood pixel gray value within the rectangular sliding window area at the same position of the two denoised SAR images; the neighborhood bright pixel similarity function represents the proximity between the maximum central pixel gray value and the maximum neighborhood pixel gray value within the rectangular sliding window area at the same position of the two denoised SAR images;

[0017] Construct a fluctuation parameter function using the two denoised SAR images; wherein, the fluctuation parameter function represents the proximity between the central pixel gray value and the neighborhood pixel gray value within the rectangular sliding window area at the same position of the two denoised SAR images;

[0018] For each pixel position, judge the relationship among the neighborhood heterogeneity function, the fluctuation parameter function, the neighborhood dark pixel similarity function, and the neighborhood bright pixel similarity function to generate the similarity matrix of the two denoised SAR images.

[0019] In an embodiment of the present invention, the expression of the neighborhood heterogeneity function includes:

[0020]

[0021] Among them, h(x) represents the neighborhood heterogeneity function; μ(x) represents the mean of the gray values of all pixels in the same rectangular sliding window area of the two denoised SAR images; σ(x) represents the variance value of the gray values of all pixels in the same rectangular sliding window area of the two denoised SAR images.

[0022] In an embodiment of the present invention, the expressions of the neighborhood dark pixel similarity function and the neighborhood bright pixel similarity function respectively include:

[0023]

[0024] Among them, f(x) represents the neighborhood dark pixel similarity function; g(x) represents the neighborhood bright pixel similarity function; I1(x) and I2(x) respectively represent the gray values of the central pixel x in the two denoised SAR images; I1(i) and I2(i) respectively represent the gray values of the neighborhood pixels other than the central pixel in the rectangular sliding window of the two denoised SAR images; Ω x , i≠x represents the set of neighborhood pixels other than x in the neighborhood of the central pixel x; i represents the position code of the neighborhood pixel; min() represents finding the minimum value; max() represents finding the maximum value.

[0025] In an embodiment of the present invention, the expression of the fluctuation parameter function includes:

[0026]

[0027] Among them, α(x) represents the fluctuation parameter function.

[0028] In an embodiment of the present invention, the expression of the similarity matrix of the two denoised SAR images includes:

[0029]

[0030] Among them, S ij represents the similarity matrix; ∧ represents taking the union.

[0031] In an embodiment of the present invention, obtaining a sample set by using the two denoised SAR images, and selecting a part from the sample set as a training sample set by using the pseudo-label matrix, includes:

[0032] Stack the two denoised SAR images according to the pixel alignment principle, and perform rasterization segmentation on the stacked image group to obtain a number of square tiles of the same size as the sample set;

[0033] For each tile, determine the larger value of the total number of two numerical values among the multiple element values in the pseudo-label matrix that match the pixel positions of the tile, and determine the element value corresponding to the larger value as the label category of the tile;

[0034] For each tile, count the number of neighboring tiles with the same label category as the tile among all neighboring tiles of the tile as the number of neighboring tiles with the same label category of the tile;

[0035] For each set of tiles corresponding to a label category, sort the tiles in descending order according to the number of neighboring tiles with the same label category, and select multiple tiles ranked at the front according to a preset ratio. All the tiles selected from the two label categories form a training sample set.

[0036] In an embodiment of the present invention, the structure traversal space of the convolutional neural network for change detection classification is set. Using a preset multi-objective optimization algorithm, upper and lower layer significant surrogate models are respectively used to optimize the network structure and network weight parameters, seeking an optimal classification network model, and obtaining a set of Pareto frontiers with high detection accuracy and low computational complexity as the solution set, including:

[0037] Define a two-layer optimization problem for the collaborative optimization of the convolutional neural network structure and parameters for SAR tile classification; wherein, the upper layer optimization is used to traverse the network structure system, and the lower layer optimization traverses the weights of the network for a given network structure system;

[0038] Perform multi-objective modeling for the collaborative optimization of the convolutional neural network structure and parameters; wherein, the objective functions include the correct classification rate PCC of change detection and the number of floating-point operations FLOPs as the computational complexity value;

[0039] Set the traversal space from four dimensions of the number of layers, convolution kernel size, dilation rate, and input resolution of the convolutional neural network, and decompose the convolutional neural network architecture into three connected stages. Each stage includes multiple convolutional layers for traversing the number of layers, and each convolutional layer adopts a bottleneck structure;

[0040] Construct upper and lower layer significant surrogate models respectively. Using the preset multi-objective optimization algorithm, the upper and lower layer significant surrogate models are respectively used to optimize the network structure and network weight parameters, seeking an optimal classification network model, and obtaining a set of Pareto frontiers with high detection accuracy and low computational complexity as the solution set; wherein, the upper layer surrogate model adopts an online learning algorithm, adaptively selects four types of accuracy prediction surrogate models, and seeks an architecture close to the current trade-off frontier in the traversal space; the lower layer surrogate model obtains initial weights through a super network model and fine-tunes the accuracy prediction surrogate model using weight sharing technology.

[0041] In one embodiment of the present invention, the bi-level optimization problem of collaborative optimization of the convolutional neural network structure and parameters for SAR patch classification is expressed as:

[0042]

[0043]

[0044] θ ∈ Ω θ , ω ∈ Ω ω

[0045] where the upper-level variable θ defines the structure of the candidate convolutional neural network, and the lower-level variable ω(θ) defines the relevant weights of the candidate convolutional neural network. represents the cross-entropy loss of the training data for a given architecture θ; constitutes m expected objectives; among them, the m expected objectives are divided into two groups. The first group f1~f k consists of objectives that depend on both the architecture and the weights; the second group f k+1 ~f m consists of objectives that depend only on the architecture; ω * (θ) represents the relevant weights of the convolutional neural network when the cross-entropy loss is minimized; Ω θ represents the set of structures of the candidate convolutional neural networks; Ω ω represents the set of relevant weights of the candidate convolutional neural networks.

[0046] In one embodiment of the present invention, the preset multi-objective optimization algorithm is the genetic algorithm NSGA-II based on fast non-dominated sorting;

[0047] Using the preset multi-objective optimization algorithm, the upper and lower significant surrogate models are used to optimize the network structure and network weight parameters respectively, seeking the optimal classification network model, and obtaining a set of Pareto fronts with high detection accuracy and low computational complexity as the solution set, including:

[0048] Based on the preset coding strategy, set the population size, and randomly generate the population size of coded individuals of the candidate convolutional neural network architectures; among them, under the preset coding strategy, the data of each coded individual is concatenated based on the coded data of the three stages; the coded data of each stage is, for the selection of the neural network architecture of this stage, encoded in the order of number of layers - convolutional kernel size - dilation rate - input resolution using an integer string, and padded with zeros to the string of the architecture with fewer layers to obtain a fixed-length code;

[0049] Use the population number of encoded individuals as the parent population, and evaluate the fitness of each individual in the parent population; decode each individual to obtain the structure of the candidate convolutional neural network, use the training sample set to train the network, and use the upper and lower significant surrogate models to optimize the structure and related weight parameters of the candidate convolutional neural network respectively; among them, four types of surrogate models are constructed in the upper layer, and an adaptive switching selection mechanism is used to adaptively select the best model through cross-validation; a super network model is constructed in the lower layer and trained according to the progressive contraction algorithm to obtain the weight parameters of the network, which are used as the warm start of the gradient descent algorithm; output the change detection accuracy rate and computational complexity of each trained candidate convolutional neural network as the fitness value of this individual;

[0050] Perform fast non-dominated sorting on the parent population according to the corresponding fitness values, and calculate the crowding degree to sort the individuals in the same Pareto rank according to the crowding degree;

[0051] Perform selection operation, crossover operation and mutation operation on the sorted parent population to obtain the offspring population; among them, the binary tournament method is used for the selection operation, the two-point crossover method is used for the crossover operation, and the polynomial mutation operator is used for the mutation operation;

[0052] Merge the sorted and selected parent population and the offspring population to obtain a merged population, and evaluate the fitness value of the merged population; among them, decode each individual in the merged population to obtain the structure of the candidate convolutional neural network, use the training sample set to train the network, and use the upper and lower significant surrogate models to optimize the structure and related weight parameters of the candidate convolutional neural network respectively, and obtain the change detection accuracy rate and computational complexity of the candidate network as the fitness value of this individual;

[0053] Perform fast non-dominated sorting on the merged population according to the corresponding fitness values, and calculate the crowding degree of the individuals in the merged population, and select the population number of individuals with small non-dominated ranks and large crowding degrees as the new parent population;

[0054] Repeat the operations of selection, crossover, mutation, merging, evaluation and sorting on the parent population until the maximum number of iterations is reached to obtain a set of Pareto fronts.

[0055] In the solution provided by the embodiment of the present invention, the idea of co-evolution is introduced into the SAR image change detection classification network to seek the optimal convolutional neural network structure and parameter configuration, which can effectively improve the accuracy of change detection while taking into account the computational complexity of the convolutional neural network. And there is no need to generate a difference map, getting rid of the dependence of SAR image change detection technology on the quality of the difference map.

[0056] Moreover, in the embodiments of the present invention, for the two-layer optimization problem of seeking the optimal architecture and its related optimal weights, significant surrogate models are used in the upper and lower layers respectively. One adopts an online learning algorithm at the architecture level to improve the search efficiency, and the other at the weight level improves the training efficiency of gradient descent through hypernetwork fine-tuning.

[0057] The following will further elaborate on the present invention in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0058] Figure 1 It is a schematic flowchart of a SAR image change detection method for collaborative optimization of network parameters and structure provided by an embodiment of the present invention;

[0059] Figure 2 It is a schematic diagram of the positional relationship between pixels and a rectangular sliding window in two denoised SAR images according to an embodiment of the present invention;

[0060] Figure 3 It is a schematic diagram of the structure and principle of significant surrogate models in the upper and lower layers provided by an embodiment of the present invention;

[0061] Figure 4 It is a schematic diagram of a collaborative optimization coding strategy provided by an embodiment of the present invention;

[0062] Figure 5 It is the first group of SAR images and change reference diagrams for simulation experiments provided by an embodiment of the present invention;

[0063] Figure 6 It is the second group of SAR images and change reference diagrams for simulation experiments provided by an embodiment of the present invention;

[0064] Figure 7 For Figure 5 change detection simulation diagrams;

[0065] Figure 8 For Figure 6 change detection simulation diagrams. Detailed Embodiments

[0066] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0067] Please refer to Figure 1 , Figure 1It is a schematic flowchart of a method for collaborative optimization of network parameters and structure in SAR image change detection provided by an embodiment of the present invention, including the following steps:

[0068] S1. Denoise two SAR images acquired at the same location at different times to obtain denoised SAR images respectively.

[0069] During the imaging, transmission, conversion, or storage process of SAR images, they will be affected by various random interference signals, namely noise, which makes the images become rough, the quality decline, and the features be submerged. In order to reduce noise and restore the real picture, it is necessary to denoise the images. Any existing denoising technology can be adopted in the embodiment of the present invention. For example, mean filtering, median filtering, Lee filtering, Sigma filtering, Frost filtering, Gamma MAP filtering, threshold filtering based on wavelet decomposition, etc.

[0070] In an optional implementation manner, Wiener filtering can be used for the denoising process.

[0071] S2. Based on neighborhood volatility, perform pixel neighborhood information analysis on the two denoised SAR images to generate a similarity matrix.

[0072] In the embodiment of the present invention, a similarity matrix is obtained for the two denoised SAR images, and this similarity matrix reflects the similarity information of neighborhood pixels in the two denoised SAR images.

[0073] In an optional implementation manner, the step of performing pixel neighborhood information analysis on the two denoised SAR images based on neighborhood volatility to generate a similarity matrix includes the following steps:

[0074] S21. Taking each pixel as the central pixel, use a rectangular sliding window with a preset size to traverse and calculate the neighborhood heterogeneity value for the two denoised SAR images at each pixel position to obtain the neighborhood heterogeneity function of the two denoised SAR images.

[0075] In an optional implementation manner, the expression of the neighborhood heterogeneity function includes:

[0076]

[0077] Among them, h(x) represents the neighborhood heterogeneity function; μ(x) represents the mean of the gray values of all pixels in the same rectangular sliding window area of the two denoised SAR images; σ(x) represents the variance value of the gray values of all pixels in the same rectangular sliding window area of the two denoised SAR images.

[0078] In the embodiment of the present invention, the rectangular sliding window is a square with a preset size, and the size is at least 3×3. The following takes a 3×3 rectangular sliding window as an example to illustrate this step, and please combine Figure 2 for understanding. Figure 2 It is a schematic diagram of the positional relationship between pixels and the rectangular sliding window in two denoised SAR images according to the embodiment of the present invention.

[0079] Figure 2 The left and right figures in

[0080] are two denoised SAR images, where the dots represent pixels and the rectangular frames represent rectangular sliding windows. In the embodiment of the present invention, each pixel is sequentially used as the central pixel of the rectangular sliding window, and the rectangular sliding window is moved row by row and column by column to traverse in the denoised SAR image, and the traversal of the two denoised SAR images is synchronized. For each moving position of the rectangular sliding window, the 8 pixels around the central pixel are used as the neighborhood pixels of the central pixel.

[0081] As Figure 2 shown in the position of the rectangular sliding window, the values of μ(x) and σ(x) are respectively the mean and variance values of the gray values of a total of 18 pixels in the two rectangular sliding windows. Therefore, the neighborhood heterogeneity value for the position of the central pixel can be calculated using the ratio of the two. Then, the neighborhood heterogeneity function can be obtained from the neighborhood heterogeneity values of all pixel positions in the denoised SAR image.

[0082] It should be added that for the edge pixels of the denoised SAR image, there will be some blank pixels in the rectangular sliding window. The corresponding neighborhood heterogeneity value can be obtained by filling with a preset pixel gray value or using the gray values of the remaining pixels around the blank pixels. The specific method is not described in detail here.

[0083] S22. Construct a neighborhood dark pixel similarity function and a neighborhood bright pixel similarity function using the two denoised SAR images.

[0084] Among them, the neighborhood dark pixel similarity function represents the proximity between the minimum central pixel gray value and the minimum neighborhood pixel gray value in the rectangular sliding window area at the same position in the two denoised SAR images; the neighborhood bright pixel similarity function represents the proximity between the maximum central pixel gray value and the maximum neighborhood pixel gray value in the rectangular sliding window area at the same position in the two denoised SAR images.

[0085] Specifically, the expressions of the neighborhood dark pixel similarity function and the neighborhood bright pixel similarity function respectively include:

[0086]

[0087] Among them, f(x) represents the neighborhood dark pixel similarity function; g(x) represents the neighborhood bright pixel similarity function; I1(x) and I2(x) respectively represent the gray values of the central pixel x in the two denoised SAR images; I1(i) and I2(i) respectively represent the gray values of the neighborhood pixels other than the central pixel within the rectangular sliding window of the two denoised SAR images; Ω x , i≠x represents the set of neighborhood pixels other than x within the neighborhood of the central pixel x; i represents the position code of the neighborhood pixel; min() represents finding the minimum value; max() represents finding the maximum value.

[0088] Taking a certain value of the neighborhood dark pixel similarity function as an example, as Figure 2 shown, min{I1(x), I2(x)} represents the smaller of the gray values of the two central pixels within the two rectangular sliding windows. min{I1(i), I2(i)} represents the smaller of the gray values of the two neighborhood pixels at the same position i within the two rectangular sliding windows.

[0089] S23. Construct a fluctuation parameter function using the two denoised SAR images.

[0090] Among them, the fluctuation parameter function represents the proximity between the gray value of the central pixel and the gray value of the neighborhood pixel within the rectangular sliding window area at the same position of the two denoised SAR images.

[0091] Specifically, the expression of the fluctuation parameter function includes:

[0092]

[0093] Among them, α(x) represents the fluctuation parameter function.

[0094] As Figure 2 shown, the fluctuation parameter function represents the mean value obtained by summing the differences between the gray values of all neighborhood pixels and the corresponding central pixel gray values in the two denoised SAR images.

[0095] S24. For each pixel position, judge the relationship between the fluctuation parameter function, the neighborhood dark pixel similarity function, and the neighborhood bright pixel similarity function, and generate a similarity matrix of the two denoised SAR images.

[0096] For the element values in the similarity matrix, they are determined by the relationship between the fluctuation parameter function, the neighborhood dark pixel similarity function, and the neighborhood bright pixel similarity function at the corresponding pixel position.

[0097] Specifically, the expression of the similarity matrix of the two denoised SAR images includes:

[0098]

[0099] Among them, S ij represents the similarity matrix; ∧ represents taking the union.

[0100] It should be noted that the similarity matrix generated by the embodiments of the invention is not the same as the difference map in the prior art. The similarity matrix is only used to guide the generation of pseudo-labels for two SAR images, rather than being the object of subsequent change detection.

[0101] S3. Use a preset threshold segmentation algorithm to segment the similarity matrix to obtain a pseudo-label matrix.

[0102] In the embodiments of the present invention, the preset threshold segmentation algorithm can adopt any existing threshold segmentation algorithm, such as the Kittler&Illingworth (abbreviated as KI) threshold segmentation algorithm, etc.

[0103] For example, for an element in the similarity matrix, if it is less than the segmentation threshold in the KI threshold segmentation algorithm, the corresponding element value in the pseudo-label matrix is 0; otherwise, it is 1.

[0104] Among them, the element values of 0 and 1 in the pseudo-label matrix respectively represent the unchanged class and the changed class in the two denoised SAR images.

[0105] S4. Use the two denoised SAR images to obtain a sample set, and select some from the sample set as the training sample set by using the pseudo-label matrix.

[0106] In an optional implementation manner, this step includes the following steps:

[0107] S41. Stack the two denoised SAR images according to the pixel alignment principle, and perform rasterization segmentation on the stacked image group to obtain a number of square tiles of the same size as the sample set.

[0108] In the embodiments of the present invention, the size of the tile can be 2×2, etc., which is specifically set according to needs. It can be understood that each tile includes some pixels in the same area of the two denoised SAR images. Hereinafter, taking the tile size of 2×2 as an example, at this time, a tile contains 2×2×2 = 8 pixels. For the convenience of description, assume that the number of a number of tiles is M, and M is greater than 0.

[0109] S42. For each tile, determine the larger value of the total number of two numerical values among the multiple element values in the pseudo-label matrix that match the pixel position of the tile, and determine the element value corresponding to the larger value as the label category of the tile.

[0110] Specifically, for a tile, there are 4 corresponding elements in the pseudo-label matrix. Count the number of 0s and 1s in these 4 elements respectively, and select the larger one. For example, if the number of 0s is larger, then determine that the label category of this tile is 0.

[0111] It can be understood that after this step, several tiles are divided into two categories. One category has a label category of 0, and the other category has a label category of 1. Each label category contains multiple tiles, forming a tile set corresponding to this label category.

[0112] S43. For each tile, count the number of neighboring tiles with the same label category as this tile among all the neighboring tiles of this tile as the number of neighboring tiles with the same label for this tile.

[0113] Specifically, for a tile, the 8 adjacent tiles around it are its neighboring tiles. For this tile, the number of neighboring tiles with the same label category as it can be counted. For example, if the label category of this tile is 0, and 6 out of its 8 neighboring tiles have a label category of 0, then the number of neighboring tiles with the same label for this tile is 6.

[0114] Therefore, for each tile, the number of neighboring tiles with the same label can be obtained. Of course, it can be understood that for edge tiles, the number of their neighboring tiles is less than 8. Therefore, the number of neighboring tiles with the same label is relatively small and is not within the scope of consideration of the training sample set.

[0115] S44. For each tile set corresponding to a label category, sort the tiles in descending order according to the number of neighboring tiles with the same label, and select multiple tiles ranked in the front according to a preset ratio. All the tiles selected from the two label categories form the training sample set.

[0116] In the embodiment of the present invention, the preset ratio can be 1 / 4, etc., and can be selected according to needs.

[0117] Specifically, take the tiles in the tile sets with label categories of 0 and 1, and sort them in descending order according to the number of neighboring tiles with the same label respectively. In each sorted tile sequence, select the first 1 / 4 of the tiles. All the tiles selected from the two label categories form the training sample set.

[0118] S5. Set the structural traversal space of the convolutional neural network for change detection classification, use the preset multi-objective optimization algorithm, and adopt upper and lower two-layer significant surrogate models to optimize the network structure and network weight parameters respectively, seek the optimal classification network model, and obtain a set of Pareto fronts with high detection accuracy and low computational complexity as the solution set.

[0119] In an optional implementation manner, this step includes the following steps:

[0120] A1, define a two - layer optimization problem for the collaborative optimization of the convolutional neural network structure and parameters for SAR patch classification.

[0121] Among them, the upper - layer optimization is used to traverse the network structure system, and the lower - layer optimization traverses the weights of the network for a given network structure system.

[0122] Specifically, the two - layer optimization problem for the collaborative optimization of the convolutional neural network structure and parameters for SAR patch classification is expressed as:

[0123]

[0124] θ ∈ Ω θ ω ∈ Ω ω

[0125] Among them, the upper - layer variable θ defines the structure of the candidate convolutional neural network, and the lower - layer variable ω(θ) defines the relevant weights of the candidate convolutional neural network. Represents the cross - entropy loss of the training data for a given architecture θ; Constitute m expected objectives; among them, the m expected objectives are divided into two groups. The first group f1~f k Consists of objectives that depend on both the architecture and the weights; the second group f k+1 ~f m Consists of objectives that depend only on the architecture; ω * (θ) represents the relevant weights of the convolutional neural network when the cross - entropy loss is minimized; Ω θ Represents the set of structures of candidate convolutional neural networks; Ω ω Represents the set of relevant weights of candidate convolutional neural networks.

[0126] A2, perform multi - objective modeling for the collaborative optimization of the structure and parameters of the convolutional neural network.

[0127] In the actual design of the convolutional neural network for change detection classification, due to the limitations of computational time - consumption and over - fitting problems, it is expected to achieve accurate change detection by the network, and at the same time, the computational complexity of the network structure needs to be considered, that is, to seek a balance between the two objectives of detection accuracy and computational complexity. Therefore, two objective functions are set in the embodiments of the present invention.

[0128] Among them, the objective functions include the correct change detection rate PCC and the number of floating - point operations FLOPs as the value of computational complexity. Specifically:

[0129]

[0130] f(2)=FLOPs = 2H in W in (Cin K 2 +1)C out

[0131] Among them, f(1) and f(2) represent two objective functions; N is the total number of pixels in a SAR image; TN represents the number of pixels correctly classified; H in and W in represent the height and width of the input features respectively; C in and C out represent the number of channels of the input and output respectively; K represents the convolution kernel size.

[0132] A3, set the traversal space from four dimensions of the number of layers, convolution kernel size, dilation rate and input resolution of the convolutional neural network, and decompose the convolutional neural network architecture into three connected stages. Each stage includes multiple convolutional layers for traversing the number of layers, and each convolutional layer adopts a bottleneck structure.

[0133] As Figure 4 shown, the embodiment of the present invention decomposes the convolutional neural network architecture into three sequentially connected stages. Each stage includes multiple convolutional layers, and each convolutional layer can be designed with different sizes of convolution kernel sizes. As the number of network layers gradually deepens, the size of the output feature map gradually decreases, and the number of channels of the feature map gradually increases. Traverse the number of layers in each stage, and set the minimum number of convolutional layers in each stage to be two and the maximum to be four. Each layer adopts a bottleneck structure to search the dilation rate of the previous 1×1 convolution and the size of the deep convolutional kernel in this stage.

[0134] A4, construct upper and lower significant surrogate models respectively, use the preset multi-objective optimization algorithm, and use the upper and lower significant surrogate models to optimize the network structure and network weight parameters respectively, seek the optimal classification network model, and obtain a set of Pareto fronts with high detection accuracy and low computational complexity as the solution set.

[0135] Among them, the upper surrogate model adopts an online learning algorithm, adaptively selects four types of accuracy prediction surrogate models, and seeks an architecture close to the current trade-off front in the traversal space; the lower surrogate model obtains the initial weights through a super network model and fine-tunes the accuracy prediction surrogate model using the weight sharing technique.

[0136] Figure 3 is the structural and principle schematic diagram of the upper and lower significant surrogate models provided by the embodiment of the present invention.

[0137] To accelerate the upper-layer optimization, the embodiments of the present invention provide four different precision prediction surrogate models, namely, Multi-Layer Perceptron (MLP), Classification and Regression Tree (CART), Radial Basis Function (RBF), and Gaussian Process (GP). All four types of surrogate models are constructed in each iteration, and through an adaptive switching selection mechanism, the best model is adaptively selected by cross-validation. To accelerate the lower-layer optimization, a super network is constructed by obtaining the maximum value of the searched architecture hyperparameters. For example, each of the three blocks has four layers, the dilation rate is set to 4, and the convolutional kernel size of each layer is set to 5. Then, the progressive contraction algorithm is followed to train the super network. This process is performed once before the optimization search. During the search process, the weights inherited from the trained super network are used as the warm start for the gradient descent algorithm, that is, the initial weights of the stochastic gradient descent algorithm.

[0138] Through the above steps, a Pareto front of a set of network structures and parameters can be obtained.

[0139] Among them, the preset multi-objective optimization algorithm can be implemented by any existing multi-objective genetic algorithm, such as the Non-dominated Sorting Genetic Algorithm (NSGA), or can be any existing decomposition-based multi-objective optimization algorithm: such as the Multi-objective Evolutionary Algorithm Based on Decomposition (MOEA / D) algorithm, etc. Multi-objective optimization will obtain a set of non-dominated solutions in which multiple objectives are as optimal as possible at the same time, and this set can also be called the Pareto front.

[0140] In an optional implementation manner, the preset multi-objective optimization algorithm is the NSGA-II genetic algorithm based on fast non-dominated sorting.

[0141] Correspondingly, using the preset multi-objective optimization algorithm, the upper and lower layer significant surrogate models are used to optimize the network structure and network weight parameters respectively, and the optimal classification network model is sought, and a Pareto front with high detection accuracy and low computational complexity is obtained as the solution set, including:

[0142] B1. Based on the preset coding strategy, the population size is set, and the population size of encoded individuals of the candidate convolutional neural network architecture is randomly generated.

[0143] Among them, under the preset coding strategy, the data of each coding individual is obtained by concatenating the coding data of the three stages; the coding data of each stage is obtained by encoding the selection of the neural network architecture of this stage using an integer string in the order of number of layers - convolution kernel size - dilation rate - input resolution, and padding zeros into the string of the architecture with fewer layers to obtain a fixed-length coding.

[0144] Specifically, the embodiments of the present invention encode the candidate convolutional neural network architectures using a collaborative optimization coding strategy. Please refer to Figure 4 , Figure 4 which is a schematic diagram of the collaborative optimization coding strategy provided by the embodiments of the present invention.

[0145] As Figure 4 shown, an integer string is used to encode the architecture selection of each stage of the convolutional neural network in the order of "number of layers - convolution kernel size - dilation rate - input resolution", and a zero-padding strategy is adopted for the architecture string with fewer layers in order to form a fixed-length coding strategy for facilitating evolutionary computation. The three stages of the entire convolutional neural network adopt a unified coding form.

[0146] Among them, the population size is denoted by M.

[0147] B2. Use the population size of coding individuals as the parental population, evaluate the fitness of each individual in the parental population; decode each individual to obtain the structure of the candidate convolutional neural network, use the training sample set to perform network training, and use the upper and lower two-layer significant surrogate models to optimize the structure and related weight parameters of the candidate convolutional neural network respectively.

[0148] Among them, four types of surrogate models are constructed in the upper layer, and an adaptive switching selection mechanism is used to adaptively select the best model through cross-validation; a supernetwork model is constructed in the lower layer and trained according to the progressive shrinking algorithm to obtain the weight parameters of the network, which are used as the warm start of the gradient descent algorithm; output the change detection accuracy rate and computational complexity of each trained candidate convolutional neural network as the fitness value of this individual.

[0149] The change detection accuracy rate and computational complexity values are obtained using existing calculation methods and will not be elaborated here.

[0150] B3. Perform fast non-dominated sorting on the parental population according to the corresponding fitness values, and calculate the crowding degree to sort the individuals in the same Pareto rank according to the crowding degree.

[0151] B4. Perform selection operation, crossover operation and mutation operation on the sorted parental population to obtain the offspring population.

[0152] Specifically, the selection operation uses the binary tournament method, that is, in the parental population, pairs of individuals are selected to generate M pairs of parents. Then, crossover and mutation operations are performed on the M pairs of parents. Among them, the crossover operation uses the two-point crossover method, and the mutation operation uses the polynomial mutation operator. A random integer value within 0 to 9 is generated, and the probabilities of crossover and mutation are 0.8 and 0.05 respectively, so as to generate M new individuals to form the offspring population.

[0153] B5. Merge the sorted and selected parental population and the offspring population to obtain a merged population, and evaluate the fitness value of the merged population.

[0154] The number of individuals in the merged population is 2M.

[0155] Among them, each individual in the merged population is decoded to obtain the structure of the candidate convolutional neural network, the network is trained using the training sample set, and the upper and lower significant surrogate models are used to optimize the structure and related weight parameters of the candidate convolutional neural network respectively, so as to obtain the change detection accuracy and computational complexity of the candidate network as the fitness value of this individual.

[0156] B6. Perform fast non-dominated sorting on the merged population according to the corresponding fitness value, and calculate the crowding degree of the individuals in the merged population. Select the individuals with a small non-dominated rank and a large crowding degree in the population quantity as the new parental population.

[0157] B7. Repeat the operations of selection, crossover, mutation, merging, evaluation and sorting on the parental population until the maximum number of iterations is reached to obtain a set of Pareto fronts.

[0158] Specifically, repeat steps B4 - B6 until the maximum number of iterations is reached, such as 100 times, etc., to obtain a set of Pareto fronts with high detection accuracy and low computational complexity.

[0159] For the specific execution process and related concepts of NSGA-II, please refer to the prior art and will not be elaborated here.

[0160] S6. Select the solution at the inflection point of the Pareto front as the final solution, decode it to obtain the convolutional neural network structure and related parameters representing the optimal performance, and use the obtained convolutional neural network structure and the training sample set to re-train the network to obtain the trained change detection classification network.

[0161] For the solution process of the final solution, please refer to the relevant knowledge of the Pareto front for understanding and will not be elaborated in detail here.

[0162] In the embodiments of the present invention, the trained change detection classification network is automatically obtained through the collaborative optimization of the network structure and parameters. Compared with the manually designed network, it has the advantages of fast optimization and better performance.

[0163] S7. Input the sample set into the change detection classification network to obtain the change detection result of the two SAR images.

[0164] Specifically, input all the tiles obtained in S41 into the change detection classification network, and output the change detection result of the two SAR images according to the classification result.

[0165] In the solution provided by the embodiments of the present invention, the idea of co-evolution is introduced into the SAR image change detection classification network to seek the optimal convolutional neural network structure and parameter configuration. While effectively improving the accuracy of change detection, it can take into account the computational complexity of the convolutional neural network. And there is no need to generate a difference map, getting rid of the dependence of the SAR image change detection technology on the quality of the difference map.

[0166] Moreover, for the two-layer optimization problem of seeking the best architecture and its related best weights in the embodiments of the present invention, significant surrogate models are used in the upper and lower layers respectively. One adopts an online learning algorithm at the architecture level to improve the search efficiency, and the other, at the weight level, improves the training efficiency of gradient descent through hypernetwork fine-tuning.

[0167] To verify the effect of the method in the embodiments of the present invention, the following is illustrated in combination with simulation experiments.

[0168] (I) Evaluation indicators

[0169] For the simulation experiment, qualitative and quantitative analyses are used to evaluate the algorithm performance. The main evaluation indicators used in the quantitative analysis are:

[0170] ① False detection number FP: Compare the change detection results obtained by different methods with the change reference map. The number of pixels that belong to the unchanged class in the change reference map but belong to the changed class in the simulation experiment result map is called the false detection number.

[0171] ② Missed detection number FN: Compare the change detection results obtained by different methods with the change reference map. The number of pixels that belong to the changed class in the change detection reference map but belong to the unchanged class in the simulation experiment result map is called the missed detection number.

[0172] ③ Overall error rate OE: The percentage of the sum of the false detection number and the missed detection number in the total number of pixels.

[0173] ④ KC coefficient for measuring the consistency between the simulation experiment result map and the change reference Figure 1 map:

[0174]

[0175] Among them, PCC represents the probability of correctly classifying pixels, and PRE represents the expected consistency ratio.

[0176] (2) Contents of simulation experiments

[0177] The existing method is used to conduct simulation experiments on different SAR image datasets. The method of the embodiment of the present invention is compared with two existing relatively advanced neural network-based change detection technologies, and simulation experiments are conducted on different SAR image datasets. In the comparison experiment, "DNN" is proposed in the paper "Change Detection in Synthetic Aperture Radar Images Based on Deep Neural Networks"; "CAE" is proposed in the paper "Deep learning and superpixel feature extraction based on contractive autoencoder for change detection in SAR images".

[0178] The first SAR image dataset used in the experiment of the embodiment of the present invention is the Ottawa dataset. It represents a part (290x350 pixels) of two SAR images over the city of Ottawa. These images were taken by the RADARSAT SAR sensor in May 1997 and August 1997 respectively. This dataset reflects the river channel changes in the Ottawa area affected by the rainy season image in 1997. The Ottawa dataset is as Figure 5 , Figure 5 (c) is Figure 5 (a) and Figure 5 (b)'s change reference diagram, where the white area represents the changed area and the black area represents the unchanged area.

[0179] The second set of images used in the simulation is the Sardinia dataset, including images of the Sardinia region in April 1999 and images of the Sardinia region in May 1999. The size of the two images is 412x300 pixels. These data show the changes in lakes and water bodies in the Sardinia region. Figure 6 Respectively show two SAR images and the reference image; Figure 6 (c) is Figure 6 (a) and Figure 6 (b)'s change reference diagram, where the white area represents the changed area and the black area represents the unchanged area.

[0180] Simulation 1. Use the method of the embodiment of the present invention and the existing method for Figure 5Perform change detection simulation on the first set of Ottawa datasets shown, and the results are as Figure 7 shown. Among them, Figure 7 (a) is the result diagram simulated by the DNN method, Figure 7 (b) is the result diagram simulated by the CAE method, Figure 7 (c) is the result diagram simulated by the method of the embodiment of the present invention. Analyze the change detection simulation experimental data of the first set of SAR images shown, and the quantitative evaluation analysis obtained is shown in Table 1. Figure 7 shown.

[0181] Combined with Figure 7 and Table 1, it can be seen that although the change detection map obtained by the DNN method can detect the main change areas, there are many white noise points in this method, resulting in a high FN value and limited classification accuracy. For the change detection map detected by the CAE method, the FN value is the lowest, but some areas such as the reduced lake area in the upper left corner are not detected, resulting in a high FP. In contrast, the method of the embodiment of the present invention can detect most of the change areas of the image, and there are fewer noise points in the detection result, the total error rate OE is the lowest, and the KC index is the highest. Therefore, the embodiment of the present invention can obtain better change detection results and has higher change detection accuracy.

[0182] Table 1 Quantitative evaluation of the change detection results of the SAR images of the Ottawa dataset

[0183]

[0184] Simulation 2. Use the existing method and the method of the embodiment of the present invention to perform change detection on the Figure 6 second set of images shown, and the results are as Figure 8 shown. Among them, Figure 7 (a) is the result diagram simulated by the DNN method, Figure 7 (b) is the result diagram simulated by the CAE method, Figure 7 (c) is the result diagram simulated by the present invention. Analyze the change detection simulation experimental data of the second set of SAR images shown, and the quantitative evaluation analysis obtained is shown in Table 2. Figure 8 shown.

[0185] Combined with Figure 8As can be seen from Table 2, although the change detection map obtained by the DNN method can detect the main change regions, there are many white noise points in this method, resulting in a high FN value and limited classification accuracy. For the change detection map obtained by the CAE method, the FN value is the lowest and the white noise points are the fewest, but some details are lost. For example, the small area of the shrinking lake in the lower right corner is not well detected, resulting in a high FP. In contrast, the method of the embodiment of the present invention can detect most of the change regions in the image, and there are fewer noise points in the detection result. The total error rate OE is the lowest and the KC index is the highest, with higher change detection accuracy, indicating that the method of the embodiment of the present invention has better change detection performance.

[0186] Quantitative Evaluation of Change Detection Results of SAR Images in Sardinia Dataset

[0187]

[0188] From the above simulation experiment analysis, it can be seen that for the problem of SAR image change detection, the method of the embodiment of the present invention gets rid of the dependence on the difference map, adopts the idea of collaborative optimization, seeks the optimal structure system and weight parameters of the convolutional neural network, improves the classification performance of the neural network while taking into account the computational complexity, and has higher change detection accuracy, which is better than the currently widely used methods.

[0189] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including a..." does not exclude the existence of additional identical elements in the process, method, article or device including the element.

[0190] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for SAR image change detection with collaborative optimization of network parameters and structure, characterized in that, Including: Performing denoising processing on two SAR images acquired at the same location at different times to respectively obtain denoised SAR images; Based on neighborhood volatility, performing pixel neighborhood information analysis on the two denoised SAR images to generate a similarity matrix; Using a preset threshold segmentation algorithm to segment the similarity matrix to obtain a pseudo-label matrix; wherein, the element values in the pseudo-label matrix are 0 and 1, respectively representing the unchanged class and the changed class in the two denoised SAR images; Obtaining a sample set by using the two denoised SAR images, and selecting a part from the sample set as a training sample set by using the pseudo-label matrix; Setting a structural traversal space for a convolutional neural network used for change detection classification, using a preset multi-objective optimization algorithm, and respectively optimizing the network structure and network weight parameters by using upper and lower layer significant surrogate models, seeking an optimal classification network model, and obtaining a set of Pareto fronts with high detection accuracy and low computational complexity as the solution set, including: Defining a two-layer optimization problem for the collaborative optimization of the convolutional neural network structure and parameters for SAR patch classification; wherein, the upper layer optimization is used to traverse the network structure system, and the lower layer optimization traverses the weights of the network for a given network structure system; performing multi-objective modeling for the collaborative optimization of the convolutional neural network structure and parameters; wherein, the objective functions include the correct classification rate PCC of change detection and the number of floating-point operations FLOPs as the computational complexity value; setting a traversal space from four dimensions of the number of layers, convolutional kernel size, dilation rate, and input resolution of the convolutional neural network, and decomposing the convolutional neural network architecture into three connected stages, each stage including multiple convolutional layers for traversing the number of layers, and each convolutional layer adopts a bottleneck structure; respectively constructing upper and lower layer significant surrogate models, using the preset multi-objective optimization algorithm, and respectively optimizing the network structure and network weight parameters by using the upper and lower layer significant surrogate models, seeking an optimal classification network model, and obtaining a set of Pareto fronts with high detection accuracy and low computational complexity as the solution set; wherein, the upper layer surrogate model adopts an online learning algorithm, adaptively selects four types of accuracy prediction surrogate models, and seeks an architecture close to the current trade-off front in the traversal space; the lower layer surrogate model obtains initial weights through a supernet model, and uses the weight sharing technology to fine-tune the accuracy prediction surrogate model; Selecting the solution at the inflection point of the Pareto front as the final solution, decoding to obtain the convolutional neural network structure and related parameters representing the optimal performance, and re-training the network by using the obtained convolutional neural network structure and the training sample set to obtain a trained change detection classification network; Inputting the sample set into the change detection classification network to obtain the change detection results of the two SAR images; 2. The SAR image change detection method for collaborative optimization of network parameters and structure according to claim 1, characterized in that, The above-mentioned based on neighborhood volatility, performing pixel neighborhood information analysis on the two denoised SAR images to generate a similarity matrix, including: Taking each pixel as the central pixel, using a rectangular sliding window with a preset size, traversing and calculating the neighborhood heterogeneity values of the two denoised SAR images at each pixel position to obtain the neighborhood heterogeneity function of the two denoised SAR images; Construct a neighborhood dark pixel similarity function and a neighborhood bright pixel similarity function using the two denoised SAR images; wherein, the neighborhood dark pixel similarity function represents the proximity between the minimum central pixel gray value and the minimum neighborhood pixel gray value within the rectangular sliding window area at the same position in the two denoised SAR images; the neighborhood bright pixel similarity function represents the proximity between the maximum central pixel gray value and the maximum neighborhood pixel gray value within the rectangular sliding window area at the same position in the two denoised SAR images; Construct a fluctuation parameter function using the two denoised SAR images; wherein, the fluctuation parameter function represents the proximity between the central pixel gray value and the neighborhood pixel gray value within the rectangular sliding window area at the same position in the two denoised SAR images; For each pixel position, judge the relationship among the neighborhood heterogeneity function, the fluctuation parameter function, the neighborhood dark pixel similarity function, and the neighborhood bright pixel similarity function to generate a similarity matrix of the two denoised SAR images.

3. The SAR image change detection method for collaborative optimization of network parameters and structure according to claim 2, characterized in that The expression of the neighborhood heterogeneity function includes: where h(x) represents the neighborhood heterogeneity function; μ(x) represents the mean of all pixel gray values within the same rectangular sliding window area in the two denoised SAR images; σ(x) represents the variance value of all pixel gray values within the same rectangular sliding window area in the two denoised SAR images.

4. The SAR image change detection method for collaborative optimization of network parameters and structure according to claim 3, characterized in that The expressions of the neighborhood dark pixel similarity function and the neighborhood bright pixel similarity function respectively include: Among them, f(x) represents the neighborhood dark pixel similarity function; g(x) represents the neighborhood bright pixel similarity function; I1(x) and I2(x) respectively represent the gray values of the central pixel x in the two denoised SAR images; I1(i) and I2(i) respectively represent the gray values of the neighborhood pixels other than the central pixel within the rectangular sliding window of the two denoised SAR images; Ω x , i≠x represents the set of neighborhood pixels other than x within the neighborhood of the central pixel x; i represents the position code of the neighborhood pixel; min() represents finding the minimum value; max() represents finding the maximum value.

5. The SAR image change detection method for collaborative optimization of network parameters and structure according to claim 4, characterized in that, The expression of the fluctuation parameter function includes: where α(x) represents the fluctuation parameter function.

6. The SAR image change detection method for collaborative optimization of network parameters and structure according to claim 5, characterized in that, The expression of the similarity matrix of the two denoised SAR images includes: Among them, S ij represents the similarity matrix; ∧ represents taking the union.

7. The SAR image change detection method for collaborative optimization of network parameters and structure according to claim 1, characterized in that, The process of obtaining a sample set using the two denoised SAR images and selecting a part from the sample set as a training sample set using the pseudo-label matrix includes: Stack the two denoised SAR images according to the pixel alignment principle, and perform rasterization segmentation on the stacked image group to obtain a number of square tiles of the same size as the sample set; For each tile, determine the larger value of the total number of occurrences of two numerical values among the multiple element values in the pseudo-label matrix that match the pixel position of the tile, and determine the element value corresponding to the larger value as the label category of the tile; For each tile, count the number of neighboring tiles with the same label category as the tile among all neighboring tiles of the tile as the number of neighboring tiles with the same label category of the tile; For each set of tiles corresponding to a label category, sort the tiles in descending order according to the number of neighboring tiles with the same label category, and select a number of tiles ranked at the front according to a preset ratio. All the tiles selected from the two label categories constitute the training sample set.

8. The SAR image change detection method for collaborative optimization of network parameters and structure according to claim 7, wherein The double-layer optimization problem of collaborative optimization of the convolutional neural network structure and parameters for SAR tile classification is expressed as: θ ∈ Ω θ ,ω ∈ Ω ω Among them, the upper-layer variable θ defines the structure of the candidate convolutional neural network, and the lower-layer variable ω(θ) defines the relevant weights of the candidate convolutional neural network; represents the cross-entropy loss of the training data for a given architecture θ; constitutes m desired objectives; among them, the m desired objectives are divided into two groups. The first group f1 to f k consists of objectives that depend on both the architecture and the weights; the second group f k+1 to f m consists of objectives that depend only on the architecture; ω * (θ) represents the relevant weights of the convolutional neural network when the cross-entropy loss is minimized; Ω θ represents the set of structures of the candidate convolutional neural networks; Ω ω represents the set of relevant weights of the candidate convolutional neural networks.

9. The SAR image change detection method for collaborative optimization of network parameters and structure according to claim 1 or 8, characterized in that The preset multi-objective optimization algorithm is the genetic algorithm NSGA-II based on fast non-dominated sorting; Using the preset multi-objective optimization algorithm, the upper and lower two-layer significant surrogate models are used to optimize the network structure and network weight parameters respectively, seeking the optimal classification network model, and obtaining a set of Pareto frontiers with high detection accuracy and low computational complexity as the solution set, including: Based on the preset coding strategy, set the population size, and randomly generate the population size of encoded individuals of the candidate convolutional neural network architecture; wherein, under the preset coding strategy, the data of each encoded individual is concatenated based on the encoded data of the three stages; the encoded data of each stage is that for the selection of the neural network architecture of this stage, an integer string is used to encode in the order of number of layers - convolutional kernel size - dilation rate - input resolution, and zeros are padded into the string of the architecture with fewer layers to obtain a fixed-length encoding; Taking the population size of encoded individuals as the parent population, evaluate the fitness of each individual in the parent population; decode each individual to obtain the structure of the candidate convolutional neural network, use the training sample set for network training, and use the upper and lower two-layer significant surrogate models to optimize the structure and related weight parameters of the candidate convolutional neural network respectively; among them, four types of surrogate models are constructed in the upper layer, and an adaptive switching selection mechanism is used to adaptively select the best model through cross-validation; a supernet model is constructed in the lower layer and trained according to the progressive shrinking algorithm to obtain the weight parameters of the network, which are used as the warm start of the gradient descent algorithm; output the change detection accuracy and computational complexity of each trained candidate convolutional neural network as the fitness value of this individual; Perform fast non-dominated sorting on the parent population according to the corresponding fitness values, and calculate the crowding degree to sort the individuals in the same Pareto rank according to the crowding degree; Perform selection operation, crossover operation and mutation operation on the sorted parent population to obtain the offspring population; wherein, the binary tournament method is used for the selection operation, the two-point crossover method is used for the crossover operation, and the polynomial mutation operator is used for the mutation operation; Merge the sorted and selected parent population and the offspring population to obtain the merged population, and evaluate the fitness value of the merged population; among them, decode each individual in the merged population to obtain the structure of the candidate convolutional neural network, use the training sample set for network training, and use the upper and lower two-layer significant surrogate models to optimize the structure and related weight parameters of the candidate convolutional neural network respectively, and obtain the change detection accuracy and computational complexity of the candidate network as the fitness value of this individual; Perform fast non-dominated sorting on the merged population according to the corresponding fitness values, and calculate the crowding degree of the individuals in the merged population, and select the population size of individuals with small non-dominated ranks and large crowding degrees as the new parent population; Repeat the operations of selection, crossover, mutation, merging, evaluation and sorting on the parent population until the maximum number of iterations is reached, and obtain a set of Pareto frontiers.

Images