Multi-dataset sar unsupervised change detection system based on self-organizing evolutionary architecture search

Abstract

The application discloses a kind of multi-dataset SAR unsupervised change detection systems based on self-organizing evolutionary architecture search, including dataset acquisition unit, optimal network individual screening unit and detection output unit, in which: dataset acquisition unit is used to obtain SAR remote sensing image dataset of different time phases;Optimal network individual screening unit is used to construct self-organizing search space, construct network population and screen optimal individual;Detection output unit is used to decode optimal individual into multi-dataset synergistic bifunctional network, which includes subnet module, multi-channel multi-scale feature interaction pool and output module;Subnet module is used to extract unique features of each SAR remote sensing image dataset, while screening multi-level features, to construct cross-dataset multi-scale features, input to multi-channel multi-scale feature interaction pool for fusion to generate shared features;After shared features and unique features are superimposed, the change detection result corresponding to each SAR remote sensing image dataset is obtained by using the output module classification.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a network-adaptive multi-data unsupervised remote sensing image change detection system. Background Technology

[0002] With the rapid development of remote sensing satellite technology, remote sensing images have been widely used in fields such as social monitoring and environmental governance. Among them, Synthetic Aperture Radar (SAR) images stand out due to their all-weather data acquisition capabilities; their imaging process is not limited by lighting, weather, or cloud cover, and they can maintain stable observations even in extreme weather conditions. This characteristic makes them irreplaceable in practical scenarios such as disaster emergency monitoring and dynamic resource surveys. In recent years, the rise of deep learning technology has driven the development of the field of remote sensing image change detection. Utilizing neural networks to extract multi-temporal image difference features and perform classification has become a research hotspot in this field.

[0003] However, existing deep learning-based SAR image change detection methods have significant limitations: these methods typically employ pre-fixed network structures, while different SAR datasets vary greatly in complexity, requiring different network depths and feature extraction capabilities, resulting in significant fluctuations in detection performance across multiple datasets for fixed structures; at the same time, relying on manually designed network structures is not only time-consuming and labor-intensive but also difficult to adapt to the differentiated needs of multiple scenarios, and existing methods do not fully explore the common features among multiple SAR datasets, further limiting the model's generalization ability and detection accuracy for complex change patterns.

[0004] These problems not only limit the efficiency of SAR image change detection technology in real-world scenarios, but also make it difficult to meet the application requirements of high precision and adaptability. Therefore, exploring a network architecture that can adaptively match data characteristics and efficiently utilize the common features of multiple datasets is of great theoretical research value and practical application significance for improving the stability and performance of SAR image change detection. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search, which aims to solve the problems in existing detection methods, such as the difficulty of adapting fixed network structures to the differentiated needs of multiple datasets, the time and effort required for manual network design, and the limited generalization ability due to insufficient mining of common features among multiple datasets.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search includes a dataset acquisition unit, an optimal network individual selection unit, and a detection output unit, wherein: The dataset acquisition unit is used to acquire SAR remote sensing image datasets from different time phases, and each SAR remote sensing image dataset corresponds to a different change detection task; The optimal network individual selection unit is used to construct a self-organizing search space that integrates structure and features, including a search subspace for subnet structure and a search subspace for feature interaction structure; an initial multi-data unsupervised change detection network population is generated based on the self-organizing search space; population evaluation, regional constraint genetic operations and population update are iteratively performed to select the optimal individual; The detection output unit is used to decode the optimal individual into a multi-dataset collaborative dual-function network. This network includes a subnet module, a multi-channel multi-scale feature interaction pool, and an output module. The subnet module is used to extract unique features from each SAR remote sensing image dataset. At the same time, based on feature interaction coding, the features output by the network layer corresponding to each subnet structure in the subnet module are filtered to form a series of multi-level features. The multi-level features corresponding to all SAR remote sensing image datasets are superimposed to form cross-dataset multi-scale features, which are then input into the multi-channel multi-scale feature interaction pool for fusion to generate shared features. The shared features and unique features are superimposed and input into the output module. After classification, the change detection results corresponding to each SAR remote sensing image dataset are obtained.

[0007] Furthermore, the self-organizing search space consists of a series of candidate network codes, each of which includes two parts: subnet structure coding and feature interaction coding. The search subspace of the subnet structure consists of all possible subnet structure encoding combinations, used to search for the corresponding optimal subnet structure for each SAR remote sensing image dataset; the search subspace of the feature interaction structure consists of all possible feature interaction encoding combinations, used to determine whether the features output by each network layer of the subnet structure participate in the interaction of the multi-channel multi-scale feature interaction pool; the number of feature interaction structure layers is consistent with the number of subnet structure layers. Subnet structure coding and feature interaction coding are concatenated to form a complete candidate network coding, which serves as an individual in the population.

[0008] Furthermore, each individual in the population represents a possible structure of a multi-data unsupervised remote sensing image change detection network. Individuals are encoded using 0-1 binary codes. The subnet structure code is located on the left side of the individual and is used to describe the subnet structure corresponding to each SAR remote sensing image dataset. Each bit in the code corresponds to a network layer in the subnet structure. A code of 1 indicates that the network layer exists, and a code of 0 indicates that the network layer does not exist. The feature interaction structure code is located on the right side of the individual and is used to describe the feature interaction method between subnet structures across datasets. Its length is consistent with the length of the subnet structure code. A code of 1 indicates that the feature output of the corresponding network layer in the subnet structure participates in the interaction process of the multi-channel multi-scale feature interaction pool, and a code of 0 indicates that the feature output of the network layer is discarded.

[0009] Furthermore, the process for determining the optimal individual is as follows: The fitness of individuals within the population is evaluated by decoding each individual in the population into a multi-dataset collaborative dual-function network, and then training and testing on multiple SAR remote sensing image datasets. The sum of the cross-entropy loss of all SAR remote sensing image datasets is used as the fitness of each individual in the population. The binary tournament algorithm is used to select parents in the population, perform regionally constrained genetic operations to obtain offspring, and then add the offspring to the current population. Population renewal is performed based on individual fitness. Determine if the maximum number of iterations has been reached; if yes, select the individual with the lowest fitness value in the current population as the optimal individual; otherwise, start a new round of iterations based on the updated population.

[0010] Furthermore, region-constrained genetic operations include crossover operations and mutation operations. Crossover operations are divided into structural region crossover, characteristic region crossover, and two-region crossover, while mutation operations are divided into structural region mutation, characteristic region mutation, and two-region mutation. Two individuals are selected from the current population as parents based on the binary tournament algorithm. and Generate random numbers ,when Less than the crossover probability When the crossover operation is triggered, a crossover type random number is generated. ,according to Perform crossover operation to obtain offspring and Select one individual from the current population based on the binary tournament algorithm. As the parent generation; generate random numbers ,when Less than the mutation probability When the mutation operation is triggered, a random number of the mutation type is generated. ,according to Perform mutation operation to obtain offspring .

[0011] Furthermore, when At that time, the structure region crosses: and In the subnet structure coding region, one intersection point is randomly selected, and the coding segments after the intersection point are swapped, while the feature interaction coding region remains unchanged; when At that time, the feature region intersection is performed: and The feature interaction coding region randomly selects one intersection point, swaps the coding segments after the intersection point, and the subnet structure coding region remains unchanged; when At that time, perform dual-region crossover: respectively in and One intersection point is randomly selected between the subnet structure coding region and the feature interaction coding region, and the coding segments after the intersection points of the two regions are swapped. After the crossover operation, a offspring is generated. and ; Furthermore, when At that time, structural region mutation is performed: in the parent generation Randomly select one mutation point in the subnet structure coding region and flip the binary value of that point; when At that time, mutations occur in the characteristic region: in the parent generation Several mutation points are randomly selected from the feature interactive coding region, and the binary values ​​of each mutation point are flipped; when At that time, perform two-region mutation: in the parent generation and in the offspring. Several mutation points are randomly selected from both the subnet structure coding region and the feature cross-coding region. The binary values ​​of all mutation points are inverted. After the mutation operation, offspring are generated. .

[0012] Furthermore, the feature interaction coding part of the optimal individual is extracted and decoded. The features output by the network layer with a code of 1 are selected and sent to the multi-channel multi-scale feature interaction pool, while the features output by the network layer with a code of 0 are discarded. Then, a series of multi-level features are selected from the sub-network structure corresponding to each SAR remote sensing image dataset. A series of multi-level features selected from the same SAR remote sensing image dataset are superimposed to form Then, all SAR remote sensing image datasets Superposition yields multi-scale features across datasets ; Will A multi-channel, multi-scale feature interaction pool is input and non-linear transformations are performed to mine shared features across datasets; the multi-channel, multi-scale feature interaction pool includes multiple fully connected layers. Shared features respectively with the first The unique features output after extracting the corresponding subnet structure features from a SAR remote sensing image dataset The results are superimposed and input into the classification layer of the output module to obtain the final change detection result. A terminal device includes a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, it implements the steps of the dataset acquisition unit, the optimal network individual selection unit, and the detection output unit in the multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search.

[0013] A computer-readable storage medium storing a computer program; when executed by a processor, the computer program implements the steps of the dataset acquisition unit, the optimal network individual selection unit, and the detection output unit in the multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search.

[0014] Compared with the prior art, the present invention has the following technical features: This invention effectively overcomes the problems of uneven sample distribution and class imbalance in traditional single-dataset analysis by constructing a dynamic network architecture for multi-dataset collaboration. Its core advantage lies in the innovative implementation of dual adaptive search of "subnet architecture" and "feature interaction structure". While accurately preserving the unique discriminative features of each SAR remote sensing image dataset through independent subnets, it deeply mines shared information and complementary features across datasets by leveraging multi-channel, multi-scale feature interaction pools, significantly enhancing the comprehensiveness and discriminative power of feature expression. At the same time, with the addition of region-constrained self-selection genetic operations, this method avoids invalid iterations in evolutionary search, improves the search efficiency of the optimal network architecture, effectively maintains population diversity to avoid local optima, and has high task scalability. It can support collaborative analysis of three or more SAR datasets under unsupervised conditions, providing an efficient and feasible technical solution for multi-scene, large-scale remote sensing image change detection. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the method of the present invention; Figure 2 This is a diagram of the DFANet network structure used for change detection across multiple datasets; Figure 3 This is a schematic diagram of the integrated search for the optimal network structure based on structural features; Figure 4This is a schematic diagram of the adaptive search network initialization process; Figure 5 This is a schematic diagram of a region-constrained self-selection genetic operation; Figure 6 This is a schematic diagram of the self-selection crossover operation under region constraints; Figure 7 This is a schematic diagram of a self-selection mutation operator with regional constraints; Figure 8 This is a schematic diagram of the feature structure decoding process; Figure 9 This is a schematic diagram of a multi-channel, multi-scale feature interaction pool for change detection in remote sensing images; Figure 10 The datasets from Sardinia, Bern, and Ottawa used in the experiments of this invention's embodiments; Figure 11 This is the Yellow River series dataset used in the experiments of this invention embodiment; Figure 12 These are the experimental results of the comparison algorithm on a single dataset in the example experiments of this invention; Figure 13 This is the experimental result of the comparison algorithm on two datasets in the example experiment of this invention. Detailed Implementation

[0016] This invention provides a multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search. By constructing a self-organizing search space integrating structure and features, and employing region-constrained genetic operations for evolutionary search, it generates optimal multi-dataset unsupervised change detection network individuals, which are then decoded into a multi-dataset collaborative dual-function network. Through a multi-channel, multi-scale feature interaction pool, this invention enables each dataset to retain its unique discriminative features while fully utilizing shared information and unique characteristics across multiple datasets. This effectively overcomes the problems of uneven sample distribution and class imbalance in single datasets, significantly improving change detection accuracy and model generalization ability.

[0017] like Figure 1 As shown, the system provided in this embodiment includes a dataset acquisition unit, an optimal network individual selection unit, and a detection output unit, wherein: S1, the dataset acquisition unit, is used to acquire SAR remote sensing image datasets.

[0018] Obtain N sets (N≥2) of synthetic aperture radar (SAR) remote sensing image datasets from different time phases; each dataset corresponds to a different change detection task, containing two or more SAR images of the same geographic area acquired at different times.

[0019] S2, the optimal network individual selection unit, is used to construct the search space and perform evolutionary search for the optimal network individual.

[0020] A self-organizing search space integrating structure and features is constructed, including a search subspace for subnet structure and a search subspace for feature interaction structure. An initial multi-data unsupervised change detection network population is generated based on the self-organizing search space. Population evaluation, regional constraint genetic operations and population update are performed iteratively to select the optimal multi-data unsupervised change detection network individual. like Figure 3 As shown, the evolutionary search process specifically includes the following sub-steps.

[0021] S2.1, Construct a self-organizing search space and initialize the population.

[0022] This step constructs a self-organizing search space that integrates structure and features, and generates an initial population of unsupervised remote sensing image change detection networks based on this self-organizing search space.

[0023] Specifically, the self-organizing search space defines the range of all possible structures that the multi-data unsupervised remote sensing image change detection network is allowed to explore; the multi-data unsupervised remote sensing image change detection network is the search object in the evolutionary search phase of this scheme, with each population individual representing a possible structure of the network. After evolutionary search and encoding / decoding, the final practically operable multi-dataset collaborative dual-function network (DFANet) is obtained; this network consists of a subnet module composed of N subnet structures corresponding one-to-one with the N SAR remote sensing image datasets obtained in S1, a multi-channel multi-scale feature interaction pool, and an output module (classification layer); each subnet structure specifically processes one SAR remote sensing image dataset and extracts the unique features of that dataset, the multi-channel multi-scale feature interaction pool receives the effective layer features of each subnet structure and fuses them to generate shared features, and the classification layer superimposes the shared features with the unique features of each subnet structure to output the change detection result.

[0024] The self-organizing search space consists of a series of candidate network codes, each of which includes two parts: subnet structure code and feature interaction code. The self-organizing search space is divided into a subnet structure search subspace and a feature interaction structure search subspace. The subnet structure search subspace is composed of all possible subnet structure code combinations and is used to search for the corresponding optimal subnet structure for each SAR remote sensing image dataset, i.e., the number of layers in the subnet and the state of each layer being retained or discarded. The initial number of subnet structures is set to 6 layers based on human experience, and the search range is 2-6 layers, so that the evolutionary search can adapt the optimal subnet structure to the data characteristics of different datasets.

[0025] The search subspace of the feature interaction structure consists of all possible combinations of feature interaction codes. It is used to search for the optimal way of cross-dataset feature interaction between subnets, that is, whether the features output by each network layer of each subnet structure participate in the interaction of the multi-channel multi-scale feature interaction pool. The number of feature interaction structure layers is consistent with the number of subnet structure layers, with a minimum of 1 layer, so that evolutionary search can determine the optimal way of interaction and sharing of similar information between multiple datasets.

[0026] The subnet structure encoding and feature interaction encoding of the search subspace of the subnet structure and the search subspace of the feature interaction structure are concatenated to form a complete candidate network encoding. Each candidate solution in the self-organizing search space corresponds to a possible combination of subnet structure and feature interaction structure. Subsequently, S2.2 evaluates the fitness of each combination on each SAR remote sensing image dataset obtained in S1 to search for the optimal individual network structure. Each network layer in the subnet structure can be, for example, a convolutional layer, and each subnet structure is composed of several stacked convolutional layers. Each layer of the feature interaction structure corresponds one-to-one with the convolutional layer in the subnet structure and is a binary marker structure. When the bit is 1, it indicates that the output feature of the corresponding convolutional layer is sent to the multi-channel multi-scale feature interaction pool to participate in cross-dataset fusion; when the bit is 0, it indicates that the output feature of the corresponding convolutional layer is discarded and does not participate in fusion.

[0027] An initial population of multi-data unsupervised change detection networks is generated based on a self-organizing search space. Each individual represents a possible structure of the multi-data unsupervised remote sensing image change detection network, which differs in the number of subnet layers, feature interaction methods, etc. Each multi-data unsupervised change detection network individual in the population uses 0-1 binary encoding, which includes two parts: subnet structure encoding and feature interaction encoding. The subnet structure encoding is located on the left side of the individual and is used to describe the subnet structure corresponding to each SAR remote sensing image dataset. The number of subnet structures corresponds to the number of change detection tasks; that is, the number of subnet structures equals the number of change detection tasks. The encoding length of the subnet structure is... The maximum possible number of layers in the subnet structure is set. Each bit in the encoding corresponds to a network layer in the subnet structure, with "1" indicating the presence of that layer and "0" indicating its absence. The feature interaction structure encoding is located to the right of the individual and describes the feature interaction methods between subnet structures across the dataset. Its length is the same as the subnet structure encoding length. To maintain consistency, the feature interactions of each subnet structure correspond to the number of layers in that subnet structure. A value of "1" indicates that the feature output of the corresponding network layer in the subnet structure participates in the interaction process of the multi-channel, multi-scale feature interaction pool, while a value of "0" indicates that the feature output of that network layer is discarded; the feature output is only meaningful if the network layer exists. The initial number of network layers for each subnet is set to be consistent, ensuring that all subnets have the same foundation at the starting point of the evolutionary process.

[0028] by Figure 4 Taking the dual-dataset task shown as an example, S1 obtains dataset one (corresponding to task). The corresponding subnet structure has five layers, and its subnet structure code is "110111". This indicates that in the six-layer network initialized in Dataset 1, the fourth layer is encoded as "0", meaning that this layer does not exist; Dataset 2 (corresponding to task The subnet structure corresponding to dataset 1 has four layers, and its subnet structure code is "101110". This means that in the six-layer network initialized in dataset 2, the second and fourth layers are encoded as "0", indicating that these two layers do not exist. The feature interaction structure also has a similar six-layer code: the feature interaction structure code corresponding to dataset 1 is "110101", indicating that the output features of the third and fifth layers are discarded and do not participate in subsequent interactions; the feature interaction structure code corresponding to dataset 2 is "100101", indicating that the output features of the second, third, and fifth layers are discarded and do not participate in subsequent interactions.

[0029] S2.2, Fitness Assessment.

[0030] Fitness evaluation is performed on individuals within the population. The fitness evaluation method is as follows: each individual in the population is decoded into a multi-dataset collaborative dual-function network DFANet, and then trained and tested on multiple SAR remote sensing image datasets. The test accuracy is used as the final individual fitness. That is, the sum of the cross-entropy losses of N SAR remote sensing image datasets is used as the fitness of each individual in the population. The lower the fitness value, the smaller the cross-entropy loss and the better the network model performance.

[0031] The fitness calculation formula is: ; in, Indicates the first in the population The fitness value of an individual indicates that the network loss is smaller and the individual's priority is higher. Indicates the first A dataset of SAR remote sensing images; Indicates the first The cross-entropy loss function corresponding to each individual.

[0032] S2.3, Region-constrained genetic operations.

[0033] The binary tournament algorithm is used to select parents from the population, perform region-constrained genetic operations to obtain offspring, and then add the offspring to the current population; for example... Figure 5 As shown, the region-constrained genetic operations include crossover operations and mutation operations; crossover operations are divided into structural region crossover, characteristic region crossover, and two-region crossover; mutation operations are divided into structural region mutation, characteristic region mutation, and two-region mutation.

[0034] Due to the dual-function nature of DFANet, ordinary crossover and mutation operators cannot simultaneously consider the structural and feature orientations of multiple tasks. Therefore, this solution requires separate processing of structure and features. Please refer to [link to relevant documentation]. Figure 6 and Figure 7 .

[0035] (1) The process of crossover operation is as follows: Two individuals are selected from the current population as parents based on the binary tournament algorithm. and (Run the binary tournament algorithm twice independently); generate random numbers. ,when Less than the crossover probability When the crossover operation is triggered, a crossover type random number is generated. ,according to Perform crossover operation to obtain offspring and The details are as follows: when At that time, the execution structure region crosses: and In the subnet structure coding region, one intersection point is randomly selected, and the coding segments after the intersection point are swapped, while the feature cross-coding regions remain unchanged, such as... Figure 6 As shown.

[0036] when At that time, the feature region crosses: and The feature interaction coding region randomly selects one intersection point, swaps the coding segments after the intersection point, and the subnet structure coding region remains unchanged.

[0037] when At that time, perform dual-region crossover: respectively in and One intersection point is randomly selected between the subnet structure coding region and the feature interaction coding region, and the coding segments after the intersection points of the two regions are swapped.

[0038] After the crossover operation, offspring are generated. and .

[0039] (2) The mutation operation process is as follows: Select one individual from the current population based on the binary tournament algorithm. As the parent generation; generate random numbers. ,when Less than the mutation probability When the mutation operation is triggered, a random number of the mutation type is generated. ,according to Perform mutation operation to obtain offspring The details are as follows: when At that time, structural region mutation is performed: in the parent generation One mutation point is randomly selected in the subnet structure coding region, and the binary value of that point is flipped.

[0040] when At that time, mutations occur in the characteristic region: in the parent generation. Several mutation points are randomly selected from the feature interactive coding region, and the binary values ​​of each mutation point are flipped.

[0041] when At that time, perform two-region mutation: in the parent generation and in the offspring generation respectively. Several mutation points are randomly selected in both the subnet structure coding region and the feature interaction coding region, and the binary values ​​of all mutation points are flipped.

[0042] After mutation, offspring are generated. ,like Figure 7 As shown.

[0043] S2.4, Population Update.

[0044] Population updates are performed based on individual fitness, retaining individuals with fitness values ​​below a preset fitness threshold for the next generation of the population.

[0045] S2.5, Determine the termination condition.

[0046] Determine if the maximum number of iterations has been reached; if the maximum number of iterations has been reached, select the individual with the lowest fitness value in the current population as the optimal individual (the optimal individual for the multi-data unsupervised change detection network); if the maximum number of iterations has not been reached, return to S2.2 to continue iterating.

[0047] S3, the detection output unit, is used to decode and perform change detection.

[0048] The optimal individual obtained in step S2 is decoded into a multi-dataset collaborative dual-function network. Unique features from each SAR remote sensing image dataset are extracted through sub-network modules. The selected cross-dataset multi-scale features are fused using a multi-channel, multi-scale feature interaction pool to generate shared features. These shared features are then superimposed with the unique features from each sub-network and input into a classification layer (e.g., a convolutional layer), and the output is... Change detection results corresponding to a SAR remote sensing image dataset.

[0049] In this invention, the mathematical expression of the entire Dual Functional Adaptive Search Network (DFANet) is as follows: ; ; ; in, Indicates the first A dataset of SAR remote sensing images; Indicates the first Change detection results corresponding to a dataset of SAR remote sensing images; Represents the dataset space, which is the set of all input SAR remote sensing image datasets; The result space represents the set of all possible change detection results output by the network. This represents DFANet, a multi-dataset collaborative dual-function network used for dataset change detection.

[0050] like Figure 2 As shown, the multi-dataset collaborative dual-function network corresponds to the multi-data unsupervised remote sensing image change detection network, including a sub-network module, a multi-channel multi-scale feature interaction pool, and an output module.

[0051] DFANet receives N SAR remote sensing image datasets. Then, each SAR remote sensing image dataset is input into its corresponding subnet structure for subsequent change detection.

[0052] The subnet module contains N subnet structures, each subnet specifically corresponding to a SAR change detection dataset, responsible for extracting unique features from each SAR remote sensing image dataset. This design aims to fully utilize the complementarity between different datasets while maintaining the uniqueness and specificity of processing each SAR remote sensing image dataset. Before the search process begins, to maintain data uniqueness and processing fairness, the initial number of network layers for each subnet structure is set to be consistent. This setting ensures that all subnet structures have the same foundation at the starting point of the evolutionary process, thereby enabling a more impartial evaluation of the effectiveness of subsequent evolutionary strategies.

[0053] During network evolution, some network structures are selectively retained to form more efficient subnet structures adapted to specific tasks; the network layers ultimately retained are... Figure 2 The solid lines represent the core components of the subnet, responsible for extracting and processing key features from the input SAR remote sensing image dataset. Meanwhile, the skipped or unselected network layers are represented by dashed lines. These layers were deemed unnecessary for the specific task during the evolution process, or their presence might degrade the model's performance or efficiency. Therefore, these layers are omitted in the final subnet structure.

[0054] The Multi-Level Feature Interaction Pool (MLFIP) is designed to process feature information from multiple SAR remote sensing image data at different scales and levels. During the evolution of the network, the MLFIP not only participates in the search of subnet structures, but also filters the output features of specific network layers in the subnet structure and inputs these features into the MLFIP to achieve feature exchange and fusion across datasets.

[0055] Specifically, suppose there is a subnet structure containing four layers. In this subnet structure, not all features from all layers are used for interaction. Instead, based on certain selection criteria (such as feature importance, information richness, etc.), only features from the first, third, and fourth layers are selected for interaction. These selected features are then input into MLFIP, which can process these features from different scales and levels and generate new and richer feature representations through complex interaction and fusion mechanisms.

[0056] Finally, the output module uses the classification layer to simultaneously generate change detection results for N SAR remote sensing image datasets. .

[0057] The network structure search parameters configured in this embodiment of the invention are shown in Table 1.

[0058] Table 1

[0059] See Figure 8 and Figure 9 The process of decoding the individual data points of the optimal multi-data unsupervised change detection network into a network structure and performing feature fusion is as follows: The first step involves extracting the feature interaction encoding portion from the individual components of the optimal multi-data unsupervised change detection network and decoding it. This decodes the input from each sub-network structure to determine whether the features output by each network layer participate in the interaction. Features output by network layers encoded as 1 are selected and sent to the multi-channel, multi-scale feature interaction pool, while features output by network layers encoded as 0 are discarded. In other words, a series of multi-level features are selected from the N sub-network structures originating from N SAR remote sensing image datasets. ;For example Figure 8 As shown, the part of the subnet structure encoded as "101110100101" that represents the interactive features is "100111", indicating that the features of the first, third and fourth layers of this subnet structure are selected and will be superimposed into the multi-channel multi-scale feature interaction pool for interaction and fusion operations.

[0060] A series of multi-level features selected from the same SAR remote sensing image dataset are superimposed to form Then, all SAR remote sensing image datasets Superimposed : ; in, Represents multi-scale features across datasets; Indicates the first The result of multi-level feature overlay of a SAR remote sensing image dataset; This indicates a feature overlay operation.

[0061] The second step is to The input is a multi-channel, multi-scale feature interaction pool, which undergoes a non-linear transformation to mine shared features across datasets. The structure of the multi-channel, multi-scale feature interaction pool is configured with five fully connected layers based on expert experience. Finally, the shared features are output after interactive processing. Contains similarity information from multiple datasets. It will return the classification layer of the corresponding task to assist in the final change detection task, as shown in the following formula:

[0062] in, This represents the mapping relationship of the multi-channel, multi-scale feature interaction pool.

[0063] The third step is to share features. respectively with the first The unique features output after extracting the corresponding subnet structure features from a SAR remote sensing image dataset (This feature is a high-level semantic feature output from the SAR remote sensing image dataset after processing by each network layer of the subnet structure, while the series of multi-level features selected by feature cross-coding are only some of the intermediate features.) This is superimposed and input into the final classification layer. The final change detection results are obtained from this; shared features This represents the similarity characteristics among multiple datasets, and features unique to each subnet. This represents the differential characteristics of each SAR remote sensing image dataset, through and The superposition of data sets effectively combines similar and different information among multiple datasets, enabling the multi-dataset collaborative dual-function network to select the most effective information among multiple datasets during network evolution, while also preserving the unique characteristics of each dataset, thus improving the accuracy of change detection. The output module generates N change detection results simultaneously. The formula for generating the change detection results is: ; ; in, Indicates the first Change detection results for each dataset; Represents the classification layer mapping function; This indicates a feature overlay operation; This indicates a constraint condition.

[0064] This invention designs a multi-channel, multi-scale feature interaction pool (MLFIP) that allows each task to retain the original classification layer output features while fully utilizing adaptive search for complementary information from different levels. MLFIP not only effectively compensates for features that may be lost during scale transformation but also successfully solves the complex problem of interaction between multiple datasets by integrating information from other tasks. This mechanism not only enhances the richness and diversity of features but also significantly improves DFANet's ability to understand and process complex data.

[0065] Example: To verify the effectiveness of the method provided in this embodiment of the invention, experimental data is presented below. Two different types of experiments—a comparative experiment and a task extension experiment—were specifically designed to verify the performance of DFANet.

[0066] Figure 10 This is a dataset display of Sardinia, Bern, and Ottawa; Figure 10 Images (a1)-(a3), (b1)-(b3), and (c1)-(c3) respectively show SAR satellite images of Sardinia, Bern, and Ottawa before and after the floods, as well as binary change reference maps (white areas in the images indicate that the pixels in the area have changed, and black areas indicate that the pixels in the area have not changed) obtained by combining local land information and expert knowledge. These are used to evaluate the change detection results of the algorithm.

[0067] Figure 11 This is a display of the Yellow River series datasets; Figure 11 (d1)-(d3), (e1)-(e3), (f1)-(f3), and (g1)-(g3) respectively show SAR satellite images of different areas in the Yellow River estuary region before and after the flood, as well as binary change reference maps from local geographic data and expert knowledge (white parts in the figure indicate that the pixels in the area have changed, and black parts indicate that the pixels in the area have not changed).

[0068] In this embodiment of the invention, the parameter settings for adaptive network structure search of the dataset are shown in Table 1 above.

[0069] The experiments selected the Kappa coefficient, pixel classification accuracy (PCC), false negatives (FN), false positives (FP), and overall error (OE) as the quantitative metrics to evaluate the algorithm performance. PCC can be calculated using the following formula:

[0070] The correct change count (True Positive, TP) is the number of correctly detected changed pixels. The false detection count (FP) is the number of pixels that did not actually change but were falsely detected as changing. The correct invariant count (True Negative, TN) is the number of correctly detected invariant pixels. The missed detection count (FN) is the number of pixels that actually changed but were falsely detected as unchanged.

[0071] To further evaluate the change detection results, the kappa coefficient is introduced. The kappa coefficient can be calculated using the following formula:

[0072] PRE can be calculated using the following formula:

[0073] The Kappa coefficient is a statistical method for evaluating consistency. In change detection problems, consistency means whether the model's predictions and the actual classification results are the same. The higher the value of this coefficient, the higher the classification accuracy achieved by the model.

[0074] OE It is the total number of errors. OE It can be calculated using the following formula:

[0075] OE The error comprehensively reflects the algorithm's level of missed detections and false detections; the smaller the error, the more reliable the detection result.

[0076] Table 2 shows the quantitative metrics for change detection using the method of this invention and several other algorithms on the Bern and Sardinian datasets.

[0077] Table 2 Quantitative Indicators of Change Detection Results in Bern and Sardinia

[0078] Figure 12 Figures (a)-(n) show the experimental results of the comparative algorithms on a single dataset. DFANet-1 demonstrated the best change detection results on both the Bern and Sardinian datasets, achieving PCC and Kappa coefficients of 99.65% and 84.35% on the Bern dataset, and 99.65% and 84.72% on the Sardinian dataset. Meanwhile, on the Bern dataset, the adaptive search methods EVODNN and EVOCNN achieved Kappa values ​​of 84.32% and 84.29%, respectively. Compared to the four algorithms with fixed network structures—SAFNet (84.15%), CapsNet (84.28%), TSPLR (84.28%), and M2HG (84.21%)—these improved metrics, demonstrating that adaptive network architecture is significant for improving change detection accuracy. Visually, Figure 12 Figures (b) and (g) show the detection results of the best algorithm EVODNN and the proposed DFANet-1, respectively. It can be observed that DFANet-1 produces fewer missed detections on the right side of the change region, clearer edges, and fewer missing regions.

[0079] Figure 13 Figures (2a)-(2t) show the experimental results of the comparative algorithms on the dual datasets. DFANet-2 achieved the highest PCC exponent and Kappa coefficient on both the Bern and Sardinian datasets. Furthermore, the quantization metrics of the fixed-structure network MMOE are higher than EMCNNAS, HPS, and SPNA on both datasets. This is because SPNA and EMCNNAS are network architectures designed for complex hyperspectral images, which can overfit and lead to decreased accuracy when the input is changed to SAR images. HPS performs worse than MMOE because its hard parameter sharing mechanism is not suitable for all datasets.

[0080] Considering both single and dual dataset results, the detection performance on the single dataset is superior to that on the dual dataset. This is partly because the Bern dataset is relatively simple, resulting in similar performance across all methods in terms of PCC. On the other hand, the interaction structure of collaborative features in dual dataset methods is fixed, which may lead to a decrease in accuracy, especially compared to single-dataset adaptive methods like EVODNN and EVOCNN, where this decrease is more significant. DFANet, however, demonstrates the best performance compared to all other methods, indicating that it overcomes the limitation of knowledge on a single dataset and establishes reasonable and effective interactions across multiple datasets, ultimately improving change detection results.

[0081] To demonstrate the task scalability of the proposed DFANet, an expansion experiment from one task to four tasks was explored using a combination of two datasets (DFANet applied to four SAR remote sensing image data is called DFANet-4). Table 3 shows the two dataset combinations.

[0082] Table 3. Task Extension Experiment Dataset Combinations

[0083] Tables 4 and 5 respectively present the quantitative indicators of the change detection results of the task expansion experiment on dataset combinations I and II.

[0084] Table 4. Quantitative Indicators of Change Detection Results in Task Extension Experiment on Dataset Combination I

[0085] Table 5. Quantitative Indicators of Change Detection Results in Task Extension Experiment on Dataset Combination II

[0086] As shown in Table 4, DFANet-1 to DFANet-4 represent the application of the proposed DFANet on one to four SAR remote sensing image datasets. Taking the Sardinia dataset as an example, the kappa values ​​of DFANet-1, DFANet-2, DFANet-3, and DFANet-4 are 84.72%, 85.11%, 85.77%, and 86.73%, respectively. Clearly, DFANet-4 achieved the highest kappa coefficient among all four SAR remote sensing image datasets, and compared to DFANet-1, the kappa coefficient was improved by up to 2%. On other datasets, it can also be observed that the performance of collaborative datasets gradually improves with the increase in the number of collaborative datasets, demonstrating the effectiveness of multi-task collaboration. The research results verify that collaborative analysis among multiple datasets can fully utilize the hidden information between similar datasets, thereby improving the results of change detection.

[0087] Table 5 shows that DFANet-1 to DFANet-4 represent the applications of the proposed DFANet on one to four SAR remote sensing image datasets. Similar to the results of Experiment 1, DFANet-4 achieved the best change detection results on all four datasets, with Kappa values ​​of 84.97%, 94.54%, 77.79%, and 81.18%, respectively. On the Ottawa dataset, the Kappa coefficients of DFANet-1, DFANet-2, DFANet-3, and DFANet-4 increased with the number of tasks, reaching 93.97%, 94.00%, 94.23%, and 94.54%, respectively. This demonstrates that joint analysis of datasets can improve the accuracy of change detection.

[0088] In summary, the network-adaptive multi-data unsupervised remote sensing image change detection method provided in this invention not only preserves the unique characteristics of each SAR dataset during its evolution but also dynamically explores and discovers more reasonable and efficient feature interaction structures between datasets. Furthermore, it can be applied to three or more SAR image datasets, exhibiting high task scalability and adapting to change detection tasks of varying scales and complexities. Experiments on multiple SAR image datasets demonstrate that DFANet possesses efficient feature extraction and recognition capabilities. Moreover, the continuously improving change detection performance with increasing task numbers also proves the effectiveness of collaboration among multiple datasets.

[0089] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search, characterized in that, It includes a dataset acquisition unit, an optimal network individual selection unit, and a detection output unit, wherein: The dataset acquisition unit is used to acquire SAR remote sensing image datasets from different time phases, and each SAR remote sensing image dataset corresponds to a different change detection task; The optimal network individual selection unit is used to construct a self-organizing search space that integrates structure and features, including a search subspace for subnet structure and a search subspace for feature interaction structure; an initial multi-data unsupervised change detection network population is generated based on the self-organizing search space; population evaluation, regional constraint genetic operations and population update are performed iteratively to select the optimal individual; The detection output unit is used to decode the optimal individual into a multi-dataset collaborative dual-function network. This network includes a subnet module, a multi-channel multi-scale feature interaction pool, and an output module. The subnet module is used to extract unique features from each SAR remote sensing image dataset. At the same time, based on feature interaction coding, the features output by the network layer corresponding to each subnet structure in the subnet module are filtered to form a series of multi-level features. The multi-level features corresponding to all SAR remote sensing image datasets are superimposed to form cross-dataset multi-scale features, which are then input into the multi-channel multi-scale feature interaction pool for fusion to generate shared features. The shared features and unique features are superimposed and input into the output module. After classification, the change detection results corresponding to each SAR remote sensing image dataset are obtained.

2. The multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search according to claim 1, characterized in that, The self-organizing search space consists of a series of candidate network codes, each of which includes two parts: subnet structure coding and feature interaction coding. The search subspace of the subnet structure consists of all possible subnet structure encoding combinations, used to search for the corresponding optimal subnet structure for each SAR remote sensing image dataset; the search subspace of the feature interaction structure consists of all possible feature interaction encoding combinations, used to determine whether the features output by each network layer of the subnet structure participate in the interaction of the multi-channel multi-scale feature interaction pool; the number of feature interaction structure layers is consistent with the number of subnet structure layers. Subnet structure coding and feature interaction coding are concatenated to form a complete candidate network coding, which serves as an individual in the population.

3. The multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search according to claim 1, characterized in that, Each individual in the population represents a possible structure of a multi-data unsupervised remote sensing image change detection network. Individuals are encoded using 0-1 binary codes. The subnet structure code is located to the left of the individual and is used to describe the subnet structure corresponding to each SAR remote sensing image dataset. Each bit in the code corresponds to a network layer in the subnet structure. A code of 1 indicates that the network layer exists, and a code of 0 indicates that the network layer does not exist. The feature interaction structure code is located to the right of the individual and is used to describe the feature interaction method between subnet structures across datasets. Its length is consistent with the length of the subnet structure code. A code of 1 indicates that the feature output of the corresponding network layer in the subnet structure participates in the interaction process of the multi-channel multi-scale feature interaction pool, and a code of 0 indicates that the feature output of the network layer is discarded.

4. The multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search according to claim 1, characterized in that, The process for determining the optimal individual is as follows: The fitness of individuals within the population is evaluated by decoding each individual in the population into a multi-dataset collaborative dual-function network, and then training and testing on multiple SAR remote sensing image datasets. The sum of the cross-entropy loss of all SAR remote sensing image datasets is used as the fitness of each individual in the population. The binary tournament algorithm is used to select parents in the population, perform regionally constrained genetic operations to obtain offspring, and then add the offspring to the current population. Population renewal is performed based on individual fitness. Determine if the maximum number of iterations has been reached; if yes, select the individual with the lowest fitness value in the current population as the optimal individual; otherwise, start a new round of iterations based on the updated population.

5. The multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search according to claim 1, characterized in that, Region-constrained genetic operations include crossover operations and mutation operations. Crossover operations are divided into structural region crossover, characteristic region crossover, and two-region crossover, while mutation operations are divided into structural region mutation, characteristic region mutation, and two-region mutation. Two individuals are selected from the current population as parents based on the binary tournament algorithm. and ; Generate random numbers ,when Less than the crossover probability When the crossover operation is triggered, a crossover type random number is generated. ,according to Perform crossover operation to obtain offspring and ; Select one individual from the current population based on the binary tournament algorithm. As the parent generation; generate random numbers ,when Less than the mutation probability When the mutation operation is triggered, a random number of the mutation type is generated. ,according to Perform mutation operation to obtain offspring .

6. The multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search according to claim 5, characterized in that, when At that time, the execution structure region crosses: and In the subnet structure coding region, a random intersection point is selected, and the coding segments after the intersection point are swapped, while the feature interaction coding region remains unchanged; when At that time, the feature region crosses: and The feature interaction coding region randomly selects one intersection point, swaps the coding segments after the intersection point, and the subnet structure coding region remains unchanged; when At that time, perform dual-region crossover: respectively in and One intersection point is randomly selected between the subnet structure coding region and the feature interaction coding region, and the coding segments after the intersection points of the two regions are swapped. After the crossover operation, a offspring is generated. and .

7. The multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search according to claim 5, characterized in that, when At that time, structural region mutation is performed: in the parent generation Randomly select one mutation point in the subnet structure coding region and flip the binary value of that point; when At that time, mutations occur in the characteristic region: in the parent generation. Several mutation points are randomly selected from the feature interactive coding region, and the binary values ​​of each mutation point are flipped; when At that time, perform two-region mutation: in the parent generation and in the offspring generation respectively. Several mutation points are randomly selected from both the subnet structure coding region and the feature cross-coding region. The binary values ​​of all mutation points are inverted. After the mutation operation, offspring are generated. .

8. The multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search according to claim 1, characterized in that, The feature interaction encoding part of the optimal individual is extracted and decoded. The features output by the network layer with encoding 1 are selected and sent to the multi-channel multi-scale feature interaction pool, while the features output by the network layer with encoding 0 are discarded. Then, a series of multi-level features are selected from the sub-network structure corresponding to each SAR remote sensing image dataset. A series of multi-level features selected from the same SAR remote sensing image dataset are superimposed to form Then, all SAR remote sensing image datasets Superposition yields multi-scale features across datasets ; Will A multi-channel, multi-scale feature interaction pool is input and non-linear transformations are performed to mine shared features across datasets; the multi-channel, multi-scale feature interaction pool includes multiple fully connected layers. Shared features respectively with the first The unique features output after extracting the corresponding subnet structure features from a SAR remote sensing image dataset The results are superimposed and input into the classification layer of the output module to obtain the final change detection result.

9. A terminal device, comprising a processor, a memory, and a computer program stored in the memory; characterized in that, When the processor executes the computer program, it implements the steps of the dataset acquisition unit, the optimal network individual selection unit, and the detection output unit in the multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search as described in any one of claims 1-8.

10. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the steps of the dataset acquisition unit, the optimal network individual selection unit, and the detection output unit in the multi-dataset SAR unsupervised change detection system based on self-organizing evolutionary architecture search as described in any one of claims 1-8.

Images