A data and knowledge dual-driven self-organizing building change detection method

Abstract

The application discloses a data and knowledge double-driven self-organizing building change detection method, comprising the following steps: acquiring a building change detection data set, extracting a plurality of field features of a building from remote sensing images of two time points before and after each data set, and generating input data after feature enhancement; inputting the enhanced input into a self-organizing data and knowledge double-driven change detection network, searching for an optimal data and knowledge double-driven change detection network suitable for data characteristics; using the optimal data and knowledge double-driven change detection network to perform change detection on the data set, obtaining an initial change detection binary image, removing non-building area noise and filling building internal cavities, and simultaneously outputting a building change detection result. The application can effectively fuse building field prior knowledge, improve the detection accuracy of unlabelled data sets by using multi-data set collaboration and complementation, improve the edge definition and overall change detection accuracy of building changes, and has good multi-task scalability.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, specifically relating to a data- and knowledge-driven method for detecting self-organizing building changes. Background Technology

[0002] With the acceleration of urbanization, building change detection has significant application value in urban planning, illegal construction monitoring, and post-disaster assessment. Building change detection accurately extracts information on new construction, demolition, and renovation / expansion of buildings by comparing and analyzing remote sensing images of the same geographical area acquired at different times. However, the diverse forms of buildings, complex edge information, and interference from environmental factors such as seasonal changes and differences in lighting conditions pose considerable challenges to building change detection.

[0003] Existing deep learning change detection methods mainly suffer from the following problems: First, purely data-driven methods do not fully utilize the prior knowledge of the building's professional domain (such as morphological and structural features, brightness characteristics, and shadow characteristics), resulting in limited accuracy in recognizing building areas; Second, building edge information is difficult to detect accurately, and traditional methods generally suffer from missed detections and blurred edges in building edge areas; Third, existing methods usually design fixed-structure networks for a single fixed dataset, requiring retraining when faced with new datasets, resulting in weak generalization ability.

[0004] Therefore, there is an urgent need for a building change detection method that integrates prior knowledge from the building profession, enables collaborative learning across multiple datasets, and jointly optimizes edge detection tasks. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a data- and knowledge-driven method for detecting changes in self-organizing buildings, including: acquiring... For each building change detection dataset, multiple domain features of buildings are extracted from remote sensing images taken at two different time points. These extracted domain features are then fused with the original images to generate feature-enhanced input data. The enhanced input is fed into a self-organizing data knowledge dual-driven change detection network. An optimal data knowledge dual-driven change detection network that best suits the data characteristics is searched. During the evolutionary search process, this network achieves cross-dataset feature collaboration through a multi-channel, multi-scale feature interaction pool. It utilizes the labeled dataset to guide the unlabeled dataset, adaptively determining the building domain feature fusion location, cross-dataset interaction structure, and edge loss weights, fully leveraging the complementary advantages of multi-dataset joint training. The optimal data knowledge dual-driven change detection network is then used to... Change detection was performed on each dataset to obtain... An initial change detection binary map is generated, and connected component post-processing is performed on the binary map to remove noise in non-building areas and fill voids inside buildings, while outputting the result. This invention provides results for detecting changes in individual buildings. It effectively integrates prior knowledge in the building domain, leverages the synergistic complementarity of multiple datasets to improve the detection accuracy of unlabeled datasets, and simultaneously enhances the edge clarity and overall change detection accuracy of buildings, exhibiting good multi-task scalability.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Obtain For each building change detection dataset, multiple domain features of buildings are extracted from remote sensing images taken at two different times before and after each dataset. The extracted domain features of buildings are then fused with the original images to generate feature-enhanced input data. Step 2: The enhanced input data is fed into the self-organizing data knowledge dual-driven change detection network to search for the optimal data knowledge dual-driven change detection network that fits the data characteristics. During the evolutionary search process, the network achieves cross-dataset feature collaboration through a multi-channel, multi-scale feature interaction pool, uses the labeled dataset to guide the unlabeled dataset with knowledge, and adaptively determines the feature fusion position of the building domain, the cross-dataset interaction structure, and the edge loss weight. Step 3: Utilize the optimal data knowledge dual-drive change detection network to... Change detection was performed on each dataset to obtain... Initial change detection binary map; for The initial change detection binary graph undergoes connected component post-processing to remove noise from non-building areas and fill voids inside buildings, while simultaneously outputting... Results of building change detection.

[0007] Preferably, step 1 specifically comprises: Get A dataset of building change detection, including A positive integer greater than or equal to 1; each building change detection dataset contains two time points, one before and one after, for the same geographic area. and The captured remote sensing images are denoted as and ; For each dataset and The following four building-related features were extracted from the remote sensing images at two different time points: (1) Morphological Building Index (MBI); (2) The corner index Harris is calculated as follows: (4) (5) in Representation matrix The determinant, For matrix The square of the trace, , These represent the gradients in the horizontal and vertical directions of the image, respectively. For window weight function, This is an empirical coefficient; This indicates the size of the local calculation window; (3) Texture index PanTex; PanTex distinguishes between building and non-building areas by statistically analyzing the contrast texture features of an image in multiple directions; (4) Morphological Shadow Index (MSI); After extracting the above four types of building domain features and the Canny edge detection map, the features are fused and data augmented according to the following steps: Step 1-1, for the first A dataset of building change detection data, which is then used in... Five feature maps of time-lapse images , …, The following formula is used to combine the components to form... Multi-feature map at time step Similarly, for The five feature maps of the time-lapse image are superimposed to form... Multi-feature map at time step : (6) in, , …, These represent feature maps for the Morphological Building Index (MBI), Corner Index (Harris), Texture Index (PanTex), Morphological Shadow Index (MSI), and Canny Edge Detection Map, respectively. This indicates a pixel-by-pixel stitching operation; Steps 1-2: The integrated multi-feature map is superimposed pixel-by-pixel with the original image at the corresponding time point. While preserving the original information, the building-related features are enhanced to obtain the feature-enhanced image. and : (7) Steps 1-3: The final input of each subnetwork Feature Enhancement and The image is constructed by stitching together the two time-point images and the comprehensive multi-feature map: (8) in This indicates a pixel-by-pixel stitching operation.

[0008] Preferably, the remote sensing image is a high-resolution optical image or a SAR image.

[0009] Preferably, the empirical coefficient Take a value of 0.04 to 0.06.

[0010] Preferably, step 2 specifically comprises: Step 2-1: Construct an adaptive multi-dataset building change detection network framework driven by both data and knowledge and determine the search space; Based on the framework of the DFANet network-adaptive multi-data unsupervised remote sensing image change detection algorithm, a building domain knowledge embedding mechanism is introduced to expand the network structure search scope, and a data knowledge dual-driven adaptive multi-dataset building change detection network DFANet-DK is constructed. DFANet-DK consists of an input module, a multi-feature fusion module, a multi-dataset network module, an edge feature extraction module, and an output module. The search space of the data-knowledge dual-driven adaptive multi-dataset building change detection network consists of three parts: Subnet architecture encoding: 0 / 1 binary encoding is used, where 0 indicates that the network layer does not exist and 1 indicates that it exists. The initial number of layers for each subnet is set to a fixed value. ; Feature structure encoding: Two-digit encoding is used. The tens digit indicates whether multiple domain features of the building are included in the fusion of this layer, and the units digit indicates whether the output of this layer participates in cross-dataset interaction. Marginal loss function weights: real number encoding is used, and the search range is [0, 1.5]. Let the population size be... The initial number of layers in the candidate network is L, and the minimum number of layers in the search network is L. The minimum number of layers in the feature structure is ; Step 2-2: Generate initial data knowledge dual-driven change detection network population using a preset individual coding method; When the input is A dataset of building change detection with a population size of [number missing]. At that time, each candidate network individual The encoding structure is as follows: (9) (10) in, Indicates the candidate network population, Represents the encoding vector of a candidate network individual. Indicates the first The architecture encoding of the sub-network corresponding to each building dataset. Indicates the first Encoding of the feature structure of each dataset This represents the weight encoding of the edge loss function; No. Each dataset and Decomposed into: (11) in, Indicates the first The subnet architecture encoding of layer l in each task Indicates the first The feature structure encoding of the l-th layer in each task is a two-digit encoding method, and the specific decoding method is as follows: (12) in, Does the representation of multiple building features enter the first stage? For each task subnet at layer l: when the value is 2, it enters fusion; when the value is 1, it does not merge. Representing the Whether the output of the l-th layer of the task subnet participates in the multi-channel multi-scale feature interaction pool: when the value is 1, it enters the multi-channel multi-scale feature interaction pool; when the value is 0, it is ignored. Randomly generated according to the above individual coding method Each individual constitutes an initial population of the data knowledge dual-driven change detection network shown in equation (9), where each individual represents a possible architecture of the data knowledge dual-driven change detection network. Steps 2-3: Iteratively optimize the population using region-constrained self-selection genetic operations; Design region-constrained self-selection genetic operations, including crossover and mutation operations; Step 2-3-1: Cross operation; Crossover operations include three types: structural region crossover, feature region crossover, and two-region crossover. The implementation process is as follows: First, in the population The algorithm randomly selects two individuals as parents using a binary tournament algorithm. and ,Will and Align the beginning and end. and The subnet architecture coding region and feature structure coding region are automatically aligned; then a random number between 0 and 1 is randomly generated. ,when Less than the crossover probability If the condition for crossover is met, it means that the type of crossover operation needs to be further determined: Then randomly generate a number between 0 and 1. , representing the type of intersection; when When the value is less than 0.3, it indicates that structural regions intersect, respectively in and Within the subnet architecture coding area, randomly select a position as the intersection point, and swap the subnet architecture coding after the two parent intersection points; when When the value is between 0.3 and 0.6, it indicates that feature region crossing is performed, respectively in... and Randomly select a position within the feature structure encoding region as an intersection point and exchange the feature structure encodings; when When the value is greater than 0.6, it indicates a double-region intersection, requiring simultaneous... and Arbitrary positions are selected as intersection points in both the subnet architecture encoding region and the feature structure encoding region to achieve the cross-region operation; for the edge loss weights Directly to and of The average value is assigned to the offspring to ensure the continuity of the weight search. This ultimately yields the offspring individuals after the crossover. and ; Step 2-3-2: Mutation operation; Mutation operations also include three types: structural region mutation, characteristic region mutation, and two-region mutation, and the implementation process is as follows: First, in the population An individual is randomly selected as the parent. Then generate a random number between 0 and 1. ,when Less than the mutation probability This indicates that the mutation conditions are met, and further determination of the mutation type is needed: then randomly generate a random number between 0 and 1. ,when A value less than 0.3 indicates a structural region mutation. The subnet architecture coding region randomly selects a position as a mutation point, and inverts the corresponding bit encoding, i.e., changes 0 to 1 or 1 to 0; when A value between 0.3 and 0.6 indicates a mutation in the characteristic region. The feature structure encoding region randomly selects several positions as mutation points, and randomly generates new legal values ​​for the two-digit codes at the selected positions, i.e., the tens digit is 1 or 2, and the units digit is 0 or 1; when A value greater than 0.6 indicates a two-region mutation, while at... Arbitrary positions are selected as mutation points in both the subnet architecture encoding region and the feature structure encoding region to achieve dual-region mutation operation; for edge loss weights A new value is randomly generated within the range [0, 1.5] to replace the mutated offspring. ; Steps 2-4: Extract edge features and construct a deep edge-guided loss function to evaluate the fitness of each individual in the population; In the data-knowledge dual-driven change detection network, the edge feature extraction module simultaneously extracts change and edge information from the output features of the multi-channel, multi-scale feature interaction pool. Specifically, the output features of the multi-channel, multi-scale feature interaction pool are sequentially passed through a fully connected layer with a ReLU activation function, and then generated by two parallel 1×1 convolutional modules: a change probability map and an edge probability map. The change probability map is activated by a Softmax layer and output, corresponding to the change detection result, while the edge probability map is activated by a Softmax layer and output, corresponding to the building edge detection result. The ground truth edge labels are generated by processing the change detection ground truth map using the Canny algorithm. The deep edge-guided loss function is derived from the change detection loss. and edge detection loss The weighted combination results in a total loss of: (13) Among them, change detection loss for The sum of cross-entropy losses for each building change detection task: (14) in, For the first The actual category tags for each task For the first The predicted probability for each task The cross-entropy loss function; Edge detection loss Binary cross-entropy loss for building edge detection: (15) in, For pixels The edge true value (0 or 1). For pixels The edge prediction probability is summed over all pixels in the image. Perform calculations; Edge loss weights Automatically determined by evolutionary adaptive search, its physical meaning is the relative proportion of the loss contribution of the change detection task and the edge detection task during training, which is used to dynamically adjust the balance between change detection accuracy and building edge sharpness during training. Steps 2-5: Determine whether the current iteration count has reached the preset maximum iteration count; Step 2-6: If not, return to the genetic manipulation steps described in Step 2-3 to continue evolution; Steps 2-7: If yes, stop the iteration and select the individual with the highest fitness from each updated population. As the optimal individual.

[0011] Preferably, step 3 specifically comprises: Step 3-1: Connected Component Labeling; Perform connected component analysis on the change detection binary graph output by the network to identify connected regions in the image where all pixels have a value of 1, i.e., the changed regions, and label and count the area of ​​each connected region; Step 3-2: Noise Removal; Calculate the area of ​​each connected region and remove scattered noise connected regions whose area is lower than a preset minimum area threshold; Step 3-3: Filling voids; The internal void filling operation is performed on the preserved connected regions; taking advantage of the connectivity feature that the internal spaces of the building are interconnected and form a regular closed shape, the internal voids enclosed by the boundaries of each connected region, that is, the areas with an internal pixel value of 0, are filled with 1, so that the detection of the main area of ​​the building is smoother and more complete.

[0012] Finally, output simultaneously Building change detection results , , … ,,correspond One input building change detection dataset.

[0013] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the above-described self-organizing building change detection method.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for detecting changes in self-organizing buildings.

[0015] A chip includes a processor for retrieving and running a computer program from a memory, causing a device equipped with the chip to perform the aforementioned self-organizing building change detection method.

[0016] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the above-described self-organizing building change detection method.

[0017] The beneficial effects of this invention are as follows: This invention overcomes the shortcomings of purely data-driven methods in utilizing domain-specific knowledge by constructing and injecting morphological building indexes (MBI), corner indexes (Harris), texture indices (PanTex), and morphological shadow indices (MSI) into the network, significantly improving the recognition accuracy and robustness of the main building area. Simultaneously, it models N building change detection datasets into a multi-task learning framework, achieving cross-dataset feature fusion and knowledge sharing through a multi-channel, multi-scale feature interaction pool. It utilizes labeled datasets to guide change detection knowledge on unlabeled datasets, effectively overcoming the accuracy degradation caused by insufficient sample labels in single datasets. Furthermore, it adaptively determines the existence of each layer in each sub-network, the location of building domain feature fusion, and whether the output features of each layer participate in cross-dataset interaction through evolutionary search, eliminating the need for manually designed fixed networks and automatically obtaining the optimal network architecture in multiple different building dataset scenarios. Finally, it designs a deep edge-guided loss function, combining change detection loss with building edge detection loss, guiding the network to simultaneously optimize change detection accuracy and building edge clarity during training, effectively solving the problems of blurred and missed detections of building edges. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the adaptive multi-dataset building change detection method driven by both data and knowledge in this invention. Figure 2 The overall framework diagram of the data knowledge-driven adaptive multi-dataset building change detection network provided in the embodiments of the present invention; Figure 3 These are example diagrams illustrating various building-related feature extraction methods according to embodiments of the present invention. Figure 3 (a) and Figure 3 (b) are two example remote sensing images; Figure 3 (a1) and Figure 3 (b1) is a feature map of the Morphological Building Index (MBI); Figure 3 (a2) and Figure 3 (b2) is the Harris corner detection map; Figure 3 (a3) and Figure 3 (b3) is the PanTex texture index feature map; Figure 3 (a4) and Figure 3 (b4) is the morphological shadow index (MSI) feature map; Figure 3 (a5) and Figure 3 (b5) is the Canny edge detection map.

[0019] Figure 4 This is an example diagram of the data augmentation result according to an embodiment of the present invention, wherein, Figure 4 (a) and Figure 4 (b) are the original images of the two examples; Figure 4 (a1) and Figure 4 (b1) is the MBI enhancement graph; Figure 4 (a2) and Figure 4 (b2) is the MSI enhancement plot; Figure 4 (a3) and Figure 4 (b3) is the Harris-enhanced plot; Figure 4 (a4) and Figure 4 (b4) is the combined enhancement diagram of MBI and MSI; Figure 4 (a5) and Figure 4 (b5) is the combined enhancement plot of MBI, MSI and Harris.

[0020] Figure 5 This is a schematic diagram of the individual encoding of a population in a data knowledge-driven dual-drive change detection network according to an embodiment of the present invention.

[0021] Figure 6 This is a structural diagram of the edge feature extraction module in an embodiment of the present invention.

[0022] Figure 7 This is a schematic diagram of the building connectivity characteristics according to an embodiment of the present invention.

[0023] Figure 8 This is a diagram illustrating the post-processing effect of an embodiment of the present invention, wherein... Figure 8 (a1) and Figure 8 (b1) is the original binary image of the change detection; Figure 8 (a2) and Figure 8 (b2) is the binary image after removing scattered noise points; Figure 8 (a3) and Figure 8 (b3) is the binary graph after the connected components are filled; Figure 8 (a4) and Figure 8 (b4) is a graph showing the final change detection results after simultaneous noise removal and void filling.

[0024] Figure 9This is an example image of the WHU building change detection dataset used in the experiments of this invention.

[0025] Figure 10 This is a diagram showing the change detection results of the comparative experiment in data A1 according to an embodiment of the present invention.

[0026] Figure 11 This is a diagram showing the change detection results of the comparative experiment in data A2 according to an embodiment of the present invention.

[0027] Figure 12 This is a diagram showing the change detection results of data B1 in the comparative experiment of an embodiment of the present invention.

[0028] Figure 13 This is a diagram showing the change detection results of data B2 in a comparative experiment according to an embodiment of the present invention. Figures 10 to 13 The sub-figures in the figure have the same meaning. Figure (a) shows the detection results of EVOCNN, Figure (b) shows the detection results of EVODNN, Figure (c) shows the detection results of SAFNet, Figure (d) shows the detection results of DFANet-1, Figure (e) shows the detection results of EGRCNN-1, Figure (f) shows the detection results of DFANet-DK-1, Figure (g) shows the detection results of SPNA, Figure (h) shows the detection results of MMOE, Figure (i) shows the detection results of HPS, Figure (j) shows the detection results of DFANet-2, Figure (t) shows the detection results of EGRCNN-2, and Figure (k) shows the detection results of DFANet-DK-2.

[0029] Figure 14 A line graph showing the experimental results of setting the weight parameters of the loss function in this embodiment of the invention, wherein, Figure 14 (a) is the loss convergence curve when the edge loss weight is 0. Figure 14 (b) shows the loss convergence curve when the edge loss weight is 0.5. Figure 14 (c) shows the loss convergence curve when the edge loss weight is 1.0. Figure 14 (d) is the loss convergence curve when the edge loss weight is 1.5.

[0030] Figure 15 This is a diagram illustrating the analysis results of evolutionary parameter settings in an embodiment of the present invention. Figure 15 (a) shows the effect curve of the iterative algebra gen on the Kappa coefficient. Figure 15 (b) shows the effect curve of population size pop on the Kappa coefficient. Figure 15 (c) is the curve showing the effect of mutation probability pm on change detection performance. Figure 15 (d) is the curve showing the effect of crossover probability pc on change detection performance. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] To address the aforementioned technical problems in existing technologies, this invention provides a data- and knowledge-driven method for detecting self-organizing building changes. The technical problem to be solved by this invention is achieved through the following technical solution: A data- and knowledge-driven method for detecting changes in self-organizing buildings includes the following steps: Get For each building change detection dataset, multiple domain features of buildings are extracted from remote sensing images taken at two different times before and after each dataset. The extracted domain features of buildings are then fused with the original images to generate feature-enhanced input data. The enhanced input is fed into the self-organizing data knowledge dual-driven change detection network, and the optimal data knowledge dual-driven change detection network that adapts to the data characteristics is obtained. In the process of evolutionary search, the network realizes cross-dataset feature collaboration through multi-channel multi-scale feature interaction pool, uses labeled datasets to guide unlabeled datasets with knowledge, adaptively determines the feature fusion position of building domain, cross-dataset interaction structure and edge loss weight, and gives full play to the complementary advantages of multi-dataset joint training. The optimal data knowledge dual-drive change detection network is used to... Change detection was performed on each dataset to obtain... An initial change detection binary map is generated; the binary map undergoes connected component post-processing to remove noise from non-building areas and fill voids inside buildings, while simultaneously outputting... Results of building change detection.

[0033] Example: like Figure 1 As shown, the data-knowledge dual-driven adaptive multi-dataset building change detection method provided in this embodiment of the invention may include the following steps: S1, obtain For each building change detection dataset, multiple domain features of buildings are extracted from remote sensing images taken at two different times before and after each dataset. The extracted domain features of buildings are then fused with the original images to generate feature-enhanced input data. This invention provides an embodiment for obtaining... A dataset of building change detection, including A positive integer greater than or equal to 1. Each building change detection dataset contains two time points, one before and one after, for the same geographic area. and The captured remote sensing images are denoted as and The remote sensing image can be a high-resolution optical image or a SAR image, depending on the actual application scenario.

[0034] Because remote sensing images of buildings exhibit high spectral variation, complex imaging conditions, and rich spatial information, and because factors such as changes in illumination and seasonal differences introduce numerous spurious variations between images at different times, data-driven methods alone are insufficient for accurately identifying specialized information such as structural features, brightness characteristics, and shadow characteristics of buildings. Therefore, this invention incorporates various building-specific features to enhance the original images, providing richer building discriminative information for subsequent networks. For example... Figure 3 As shown, for each dataset in and The following four building-related features were extracted from the remote sensing images at two different time points: (1) Morphological Building Index (MBI). MBI is an index that measures the bulkiness of urban buildings, sensitive to rectangular, T-shaped, and U-shaped building areas. MBI uses morphological operators to model the spectral and structural features of buildings, and in terms of brightness characteristics, it can detect the edges of buildings with high brightness, effectively distinguishing buildings from other land features. For example... Figure 3 (a1) and Figure 3 As shown in (b1), the rooftop area of ​​buildings in the MBI feature map shows the darkest color compared to other land features, which intuitively demonstrates the significant ability of MBI to distinguish building areas.

[0035] (2) Harris Corner Index (Harris). Harris corner detection is a key indicator for describing the geometric features of a building, suitable for detecting changes in the corners of the building's interior edges. Its calculation method is as follows: (4) (5) in Representation matrix The determinant, For matrix The square of the trace, This indicates the size of the local calculation window. , These represent the gradients in the horizontal and vertical directions of the image, respectively. For window weight function, The empirical coefficient is typically taken as 0.04 to 0.06. For example... Figure 3 (a2) and Figure 3As shown in (b2), the Harris corner detection map marks the building outline with a red box. When used in conjunction with the subsequent edge detection map, it can effectively increase the building outline information and maintain edge information while ensuring the complete detection of the main area.

[0036] (3) Texture Index (PanTex). PanTex reflects the spatial correlation between pixel grayscale values, distinguishing between buildings and non-building areas by statistically analyzing the contrast texture features of an image in multiple directions. Buildings typically exhibit high texture contrast in all directions, while linear structures in non-building areas (such as roads) only show high texture contrast in a specific direction, with lower contrast in other directions. By utilizing this multi-directional texture contrast difference, PanTex can effectively identify building areas and reduce false detections caused by features with similar textures to buildings, such as roads and concrete surfaces. Figure 3 (a3) and Figure 3 As shown in (b3), the building region in the PanTex feature map has a significantly stronger texture response.

[0037] (4) Morphological Shadow Index (MSI). MSI is complementary to MBI but in the opposite direction. It detects the shadowed areas at the edges of buildings, thus determining building boundaries. Because buildings have a certain height, their edges will produce obvious shadowed areas under specific lighting conditions. MSI utilizes this vertical dimensional information to capture building height-related features. For example... Figure 3 (a4) and Figure 3 As shown in (b4), the areas of large trees and buildings in the MSI feature map are the darkest. Combining MSI with MBI can more accurately locate the main body of the building and effectively reduce the interference of tall objects such as large trees on building detection.

[0038] In extracting the above four types of building domain features and Canny edge detection maps (such as...) Figure 4 (as shown in (a5)) After that, perform feature fusion and data augmentation processing according to the following steps: S11, for the first A dataset of building change detection data, which is then used in... Five feature maps of time-lapse images ,…, The components are stitched together according to the channel dimension to form... Multi-feature map at time step Similarly, for The five feature maps of the time-lapse image are stitched together along the channel dimension to form... Multi-feature map at time step : (6) S12, the multi-feature maps are concatenated with the original images at the corresponding time points along the channel dimension, enhancing building-related features while preserving the original information, to obtain the feature-enhanced image. and : (7) like Figure 4 As shown, after data augmentation, the recognizability of building areas in the image is significantly improved, and the building outlines and internal structural features are effectively enhanced.

[0039] S13, p. The final input of each subnetwork It is composed of two time-stitched images after feature enhancement and a comprehensive multi-feature map: (8) in This represents the channel stitching operation between images or feature maps at different times. It also includes the enhanced images before and after the feature enhancement, as well as the combined multi-feature map of the two time points, as input to the corresponding sub-network to improve the network's perception accuracy of building areas.

[0040] S2 feeds the enhanced input into the self-organizing data knowledge dual-driven change detection network and searches for the optimal data knowledge dual-driven change detection network that fits the data characteristics. During the evolutionary search process, the network achieves cross-dataset feature collaboration through a multi-channel, multi-scale feature interaction pool. It uses the labeled dataset to guide the unlabeled dataset with knowledge, adaptively determines the feature fusion position of the building domain, the cross-dataset interaction structure, and the edge loss weights, and fully leverages the complementary advantages of multi-dataset joint training.

[0041] S21. Construct an adaptive multi-dataset building change detection network framework driven by both data and knowledge, and determine the search space.

[0042] This invention, based on the framework of the Dual Functional Adaptive Search Network (DFANet), expands the network structure search scope by introducing a building domain knowledge embedding mechanism, thus constructing a data knowledge-driven adaptive multi-dataset building change detection network (DFANet-DK). Figure 2 As shown, DFANet-DK consists of an input module, a multi-feature fusion module, a multi-dataset network module, an edge feature extraction module, and an output module.

[0043] The search space of the data-knowledge dual-driven adaptive multi-dataset building change detection network consists of three parts: subnet architecture encoding, which uses 0 / 1 binary encoding, where 0 / 1 indicates the existence of each network layer, and the initial number of layers in each subnet is set to a fixed value. Feature structure encoding uses a two-digit code. The tens digit indicates whether multiple domain features of a building are included in the fusion of this layer, and the units digit indicates whether the output of this layer participates in cross-dataset interaction. Marginal loss function weights use real-number encoding, with a search range of [0, 1.5]. In this embodiment, the population size is pop, the initial number of layers in the candidate network is L, and the minimum number of layers in the search network is [missing value]. The minimum number of layers in the feature structure is .

[0044] S22, using a preset individual coding method to generate an initial data knowledge dual-driven change detection network population.

[0045] When the input is A dataset of building change detection with a population size of [number missing]. At that time, each candidate network individual The encoding structure is as follows: (9) (10) in, Indicates the candidate network population, Represents the encoding vector of a candidate network individual. Indicates the first The architecture encoding of the sub-network corresponding to each building dataset. Indicates the first Encoding of the feature structure of each dataset This represents the weight encoding of the edge loss function. Further, the... Each dataset and It can be broken down into: (11) in, Indicates the first The subnet architecture encoding of layer l in each task (0 / 1 binary, 0 indicates that the layer does not exist, 1 indicates that it exists). Indicates the first The feature structure encoding of the l-th layer in each task is a two-digit encoding method, and the specific decoding method is as follows: (12) in, Does the representation of multiple building features enter the first stage? For each task subnet at layer l: when the value is 2, it enters fusion; when the value is 1, it does not merge. Representing the Whether the output of layer l of a task subnet participates in the multi-channel, multi-scale feature interaction pool: a value of 1 indicates it enters the multi-channel, multi-scale feature interaction pool; a value of 0 indicates it is ignored. It is important to note that feature structure encoding is only meaningful when the corresponding subnet architecture encoding is 1 (i.e., this layer of the network exists). If the subnet architecture encoding is 0, the corresponding feature structure encoding automatically becomes invalid. Figure 5 As shown in Part Three.

[0046] Randomly generated according to the above encoding method Each individual constitutes the initial population of the data knowledge dual-driven change detection network shown in Equation (8), where each individual represents a possible architecture of the data knowledge dual-driven change detection network. These architectures differ in terms of the number of sub-network layers, the location of building feature fusion, the feature interaction method, and the edge loss weight.

[0047] With two tasks ( and The population initialization process is illustrated using the example of (e.g.) Figure 5 Part Two (as shown): Task The corresponding sub-network code is "110111", indicating that the fourth layer does not exist in the initial six-layer network, and the actual number of sub-network layers is five; Task The corresponding subnetwork code is "101110", indicating that the second and fifth layers do not exist in the initial six-layer network, and the actual number of subnetwork layers is four. Accordingly, the task... In the feature structure encoding (e.g., "21,11,10,11,10,11"), the position with a tens digit of 2 (first layer) indicates that multiple building features enter this layer for fusion, and the position with a units digit of 1 indicates that the features of this layer enter the multi-channel, multi-scale feature interaction pool to participate in the interaction; similarly, in the decoding task... Feature structure encoding. Edge loss weights. Randomly initialize within the range [0, 1.5].

[0048] S23 utilizes a region-constrained self-selection genetic operation to iteratively optimize the population.

[0049] Since the encoding of the data knowledge dual-driven change detection network includes three interrelated but functionally different encoding regions: subnet architecture region, feature structure region, and edge loss weight, the embodiments of the present invention design a region-constrained self-selection genetic operation, including crossover operation and mutation operation.

[0050] S231, cross operation.

[0051] Cross operations include three types: structural region cross, feature region cross, and two-region cross. The detailed implementation process is as follows: First, in the population The algorithm randomly selects two individuals as parents using a binary tournament algorithm. and ,Will and Align the beginning and end. and The subnet architecture coding region and feature structure coding region are then automatically aligned. Then, a random number between 0 and 1 is generated. ,when Less than the crossover probability If the condition for crossover is met, it means that the type of crossover operation needs to be further determined: Then randomly generate a number between 0 and 1. , representing the type of intersection. When When the value is less than 0.3, it indicates that structural regions intersect, respectively in and Within the subnet architecture coding area, randomly select a position as the intersection point, and swap the subnet architecture coding after the two parent intersection points; when When the value is between 0.3 and 0.6, it indicates that feature region crossing is performed, respectively in... and Randomly select a position within the feature structure encoding region as an intersection point and exchange the feature structure encodings; when When the value is greater than 0.6, it indicates a double-region intersection, requiring simultaneous... and Arbitrary positions are selected as intersection points in both the subnet architecture encoding region and the feature structure encoding region to achieve the cross-region operation; for the edge loss weights Directly to and of The average value is assigned to the offspring to ensure the continuity of the weight search. This ultimately yields the offspring individuals after the crossover. and .

[0052] S232, mutation operation.

[0053] Mutation operations also include three types: structural region mutation, characteristic region mutation, and two-region mutation. The detailed implementation process is as follows: First, in the population An individual is randomly selected as the parent. Then generate a random number between 0 and 1. ,when Less than the mutation probability This indicates that the mutation conditions are met, and further determination of the mutation type is needed: then randomly generate a random number between 0 and 1. ,when A value less than 0.3 indicates a structural region mutation. The subnet architecture coding region randomly selects a position as a mutation point, and inverts the corresponding bit encoding (0 becomes 1 or 1 becomes 0); when A value between 0.3 and 0.6 indicates a mutation in the characteristic region. The feature structure encoding region randomly selects several positions as mutation points, and randomly generates new legal values ​​for the two-digit codes at the selected positions (tens digit is 1 or 2, units digit is 0 or 1); when A value greater than 0.6 indicates a two-region mutation, while at... Arbitrary positions are selected as mutation points in both the subnet architecture encoding region and the feature structure encoding region to achieve dual-region mutation operations. For edge loss weights... A new value is randomly generated within the range [0, 1.5] and used to replace the mutated value. This ultimately yields the mutated offspring. .

[0054] S24. Extract edge features and construct a deep edge-guided loss function to evaluate the fitness of each individual in the population.

[0055] like Figure 6 As shown, the edge feature extraction module in the data-knowledge dual-driven change detection network simultaneously extracts change and edge information from the output features of the multi-channel, multi-scale feature interaction pool. Specifically, the output features of the multi-channel, multi-scale feature interaction pool are sequentially passed through a fully connected layer (with ReLU activation function), and then generated by two parallel 1×1 convolutional modules: a change probability map (output after activation by a Softmax layer, corresponding to the change detection result) and an edge probability map (output after activation by a Softmax layer, corresponding to the building edge detection result). The ground truth edge labels are generated by processing the change detection ground truth map using the Canny algorithm.

[0056] The depth edge guidance loss function is composed of change detection loss. and edge detection loss The weighted combination results in a total loss of: (13) Among them, change detection loss for The sum of cross-entropy losses for each building change detection task: (14) in, For the first The actual category tags for each task For the first The predicted probability for each task The cross-entropy loss function; Edge detection loss Binary cross-entropy loss for building edge detection: (15) in, For pixels The edge true value (0 or 1). For pixels The edge prediction probability is summed over all pixels in the image. Perform calculations; Edge loss weights Automatically determined by evolutionary adaptive search, its physical meaning is the relative proportion of the loss contribution of the change detection task and the edge detection task during training, which is used to dynamically adjust the balance between change detection accuracy and building edge sharpness during training.

[0057] S25, determine whether the current iteration count has reached the preset maximum iteration count.

[0058] S26. If not, return to the genetic manipulation steps described in S23 and continue the evolution.

[0059] S27, if yes, stop the iteration and select the individual with the highest fitness from each updated population. As the optimal individual.

[0060] S3. Utilize the aforementioned optimal data knowledge dual-drive change detection network to perform change detection on all building change detection datasets, and obtain... An initial change detection binary image; for the... Each initial change detection binary image is post-processed using building connectivity features to remove misdetected pixels in non-building areas and fill in holes inside buildings, while simultaneously outputting... Results of building change detection.

[0061] like Figure 7 As shown, buildings typically exhibit relatively regular edge outlines, with interconnected internal spaces forming a unified connected region. Their exterior outlines are mostly square-shaped, representing a regular and closed shape. Based on these unique building characteristics, this embodiment of the invention, after outputting the change detection binary map from the network, introduces an additional post-processing module based on connected component analysis to enhance the results.

[0062] S31, Connected Component Labeling: Perform connected component analysis on the binary graph of change detection output by the network, identify all connected regions in the image with a pixel value of 1 (change region), and label and count the area of ​​each connected region. For example... Figure 8 As shown in (a1), the original binary image contains a large number of scattered misdetected pixels in non-building areas (such as road edges, parking lots, etc.) and void areas inside buildings caused by missed detections.

[0063] S32, Noise Removal: Calculate the area of ​​each connected region and remove scattered noise connected regions whose area is lower than a preset minimum area threshold. For example... Figure 8 As shown in (a2), after noise removal, the number of misdetected pixels in the non-building area is significantly reduced. In particular, the misdetection of typical non-building features such as straight lines at the road edge is completely eliminated, while the main building area is completely preserved.

[0064] S33, Hole Filling: Performs internal hole filling operation on the preserved connected regions. Utilizing the connectivity feature of interconnected interior spaces within a building, forming a regular closed shape, the internal holes (i.e., areas with an internal pixel value of 0) enclosed by the boundaries of each connected region are filled with 1, making the detection of the main building area smoother and more complete. For example... Figure 8 As shown in (a3), after the connected domain is filled, the detection inside the building is more thorough and the missed detection rate is significantly reduced.

[0065] It is understandable that, such as Figure 8 As shown in (a4), after simultaneously performing noise removal and hole filling operations, the false detection rate in the non-building areas of the final change detection result image is extremely low, while the accuracy of the main building area is high and the edges are intact, achieving a significant quality improvement compared to the original output image. The post-processing module is independent of the network training process and is executed when generating the final output, without affecting the model weight parameters.

[0066] Finally, output simultaneously Building change detection results , , … ,,correspond One input building change detection dataset.

[0067] To verify the effectiveness of the method provided in this embodiment of the invention, experimental data is used below. This embodiment uses the WHU building change detection dataset for experiments and compares the method of this invention with various change detection methods. The comparison methods applied to a single dataset include EVOCNN, EVODNN, SAFNet, DFANet-1, EGRCNN-1, and DFANet-DK-1; the comparison methods applied to two datasets include SPNA, MMOE, HPS, DFANet-2, EGRCNN-2, and DFANet-DK-2. DFANet-DK-1 represents the application of the method of this invention under single dataset conditions, and DFANet-DK-2 represents the application of the method of this invention under two dataset conditions. By comparing the detection result images and detection quantization indicators of the above methods on the WHU building change detection dataset, the effectiveness of the method of this embodiment of the invention is verified.

[0068] Figure 9 This is an example image of the homogeneous WHU building change detection dataset used in the experiments of this invention embodiment. The dataset consists of a pair of high-resolution aerial optical images (0.3 meters resolution) taken in the same area of ​​Christchurch, New Zealand in 2012 and 2016. The images are 32507×15354 pixels in size and were cropped into multiple non-overlapping image patches of 256×256 pixels in the experiment. This invention embodiment selects four images for verification, numbered A1, A2 (Group A) and B1, B2 (Group B), where A1 and B1 are set as unlabeled datasets, and A2 and B2 are set as labeled datasets (40% labeled). Figure 9 The images in each group are arranged from left to right as follows: Optical remote sensing images taken in 2012. Optical remote sensing images taken at the time (2016) and corresponding change detection ground truth maps (white areas represent building pixels that have changed, and black areas represent pixels that have not changed).

[0069] The experiment selected the following quantitative metrics to evaluate the algorithm performance: False Negative (FN), False Positive (FP), Overall Error (OE), Percentage of Correct Classification (PCC), and Kappa coefficient. The overall accuracy (PCC) can be calculated using the following formula: (16) Among them, True Positive (TP) is the number of correctly detected changed pixels, False Positive (FP) is the number of pixels that did not actually change but were incorrectly detected as changed, True Negative (TN) is the number of correctly detected unchanged pixels, and False Negative (FN) is the number of pixels that actually changed but were incorrectly detected as unchanged.

[0070] To further evaluate the change detection results, the Kappa coefficient is introduced. The Kappa coefficient can be calculated using the following formula: (17) PRE can be calculated using the following formula: (18) The Kappa coefficient is a statistical method for evaluating consistency. In the problem of building change detection, its consistency means whether the model's prediction results are the same as the actual building changes. The higher the value of the coefficient, the higher the accuracy of the model's detection.

[0071] Tables 1 and 2 show the quantitative indicators of change detection results for each method on dataset A and dataset B, respectively.

[0072] Table 1

[0073] Table 2

[0074]

[0075] As shown in Table 1, on dataset A, the method of this invention achieved good change detection results under both single and dual dataset conditions. For data A1, the PCC of DFANet-DK-1 was 0.9830, and the Kappa coefficient was 0.9122; for DFANet-DK-2, the PCC was 0.9878, and the Kappa coefficient was 0.9379. For data A2, the PCC of DFANet-DK-1 was 0.9278, and the Kappa coefficient was 0.8524; for DFANet-DK-2, the PCC was 0.9322, and the Kappa coefficient was 0.8627. (Combined...) Figure 10 and Figure 11It can be seen that some comparison methods have missed detections in the building edge area or produce false detections in non-building areas, while the method of the present invention can better maintain the integrity of the main building area and reduce the number of false detections in non-building areas. As shown in Table 2, the method of the present invention also has good detection performance on dataset B. For data B1, the PCC of DFANet-DK-2 is 0.9345, and the Kappa coefficient is 0.7231; for data B2, the PCC of DFANet-DK-2 is 0.9253, and the Kappa coefficient is 0.8059. Combined with... Figure 12 and Figure 13 It can be seen that the buildings in dataset B are significantly affected by factors such as lighting, roof texture, and blurred edges. Some comparison methods are prone to missing the main body of the buildings or having incomplete boundaries. However, the method of this invention can obtain a more complete area of ​​building changes and make the boundary of change detection clearer. (See Tables 1 and 2.) Figures 10 to 13 The results show that, compared with the single-dataset method, DFANet-DK-2 under the dual-dataset condition achieves better results on most metrics, indicating that the knowledge interaction and information complementarity between multiple building datasets can effectively enhance the network's ability to identify building edges and main areas, thereby improving the accuracy of building change detection.

[0076] To further verify the effectiveness of the deep edge-guided loss function in the change detection training process, four sets of control experiments were designed. Under the condition of keeping the network architecture, hyperparameters, and training strategies completely consistent, only the edge loss weights were changed. Implement differentiated control. For example... Figure 14 As shown, Figure 14 (a) is the convergence curve containing only change detection loss. It can be observed that the curve fluctuates greatly between 0 and 25 generations and takes about 100 generations to stabilize. Figure 14 (b) is the method of the present invention. The convergence curves show that convergence is very smooth from generation 0 to 25, and the loss value tends to stabilize around generation 25. This significantly accelerated the training convergence speed; Figure 14 (c) and Figure 14 (d) are respectively and The convergence curve shows that when the edge detection loss accounts for a higher proportion, the curve converges faster, but the overall change detection accuracy is still lower than that of the method of this invention. In summary, this invention adopts... The deep edge-guided loss function plays a strong role in maintaining the stability of network training and can achieve the optimal dynamic balance between change detection and edge detection tasks.

[0077] In this embodiment of the invention, the population size pop, the number of iterations gen, and the mutation probability were also verified. and crossover probability The impact of data-knowledge dual-driven adaptive multi-dataset building change detection method on performance was investigated using experiments on the homogeneous WHU dataset, group A. Figure 15 As shown in (a), the Kappa coefficient gradually increases with the number of iterations, but tends to stabilize after about 30 iterations. Therefore, the maximum number of iterations is set to gen=30. Figure 15 As shown in (b), the Kappa coefficient gradually increases with the increase of population size, reaching a peak at a population size of 50 and then stabilizing. Therefore, the population size is set to pop=50. Figure 15 (c) and Figure 15 As shown in (d), when the mutation probability =0.4 and crossover probability When the value is 0.5, all evaluation indicators achieve their optimal values, proving that when... and When set to 0.4 and 0.5 respectively, BEASNet achieves a good balance between exploring new network structures and maintaining the original performance.

[0078] In summary, the adaptive multi-dataset building change detection method driven by dual data knowledge provided in this embodiment of the invention can process multiple building change detection datasets simultaneously. Through the injection of building domain knowledge and cross-dataset feature collaboration, it significantly improves the completeness of building main area recognition, edge clarity, and anti-noise interference ability compared with existing technologies.

Claims

1. A data- and knowledge-driven method for detecting changes in self-organizing buildings, characterized in that, Includes the following steps: Step 1: Obtain For each building change detection dataset, multiple domain features of buildings are extracted from remote sensing images taken at two different times before and after each dataset. The extracted domain features of buildings are then fused with the original images to generate feature-enhanced input data. Step 2: The enhanced input data is fed into the self-organizing data knowledge dual-driven change detection network to search for the optimal data knowledge dual-driven change detection network that fits the data characteristics. In the evolutionary search process, the network achieves cross-dataset feature collaboration through a multi-channel, multi-scale feature interaction pool, uses the labeled dataset to guide the unlabeled dataset with knowledge, and adaptively determines the feature fusion position of the building domain, the cross-dataset interaction structure, and the edge loss weight. Step 3: Utilize the optimal data knowledge dual-drive change detection network to... Change detection was performed on each dataset to obtain... Initial change detection binary map; for The initial change detection binary graph undergoes connected component post-processing to remove noise from non-building areas and fill voids inside buildings, while simultaneously outputting... Results of building change detection.

2. The data- and knowledge-driven self-organizing building change detection method according to claim 1, characterized in that, Step 1 specifically involves: Get A dataset of building change detection, including A positive integer greater than or equal to 1; each building change detection dataset contains two time points, one before and one after, for the same geographic area. and The captured remote sensing images are denoted as and ; For each dataset and The following four building-related features were extracted from the remote sensing images at two different time points: (1) Morphological Building Index (MBI); (2) The corner index Harris is calculated as follows: (4) (5) in Representation matrix The determinant, For matrix The square of the trace, , These represent the gradients in the horizontal and vertical directions of the image, respectively. For window weight function, This is an empirical coefficient; This indicates the size of the local calculation window; (3) Texture index PanTex; PanTex distinguishes between building and non-building areas by statistically analyzing the contrast texture features of an image in multiple directions; (4) Morphological Shadow Index (MSI); After extracting the above four types of building domain features and the Canny edge detection map, the features are fused and data augmented according to the following steps: Step 1-1, for the first A dataset of building change detection data, which is then used in... Five feature maps of time-lapse images ,…, The following formula is used to combine the components to form... Multi-feature map at time step Similarly, for The five feature maps of the time-lapse image are superimposed to form... Multi-feature map at time step : (6) in, , …, These represent feature maps for the Morphological Building Index (MBI), Corner Index (Harris), Texture Index (PanTex), Morphological Shadow Index (MSI), and Canny Edge Detection Map, respectively. This indicates a pixel-by-pixel stitching operation; Steps 1-2: The integrated multi-feature map is superimposed pixel-by-pixel with the original image at the corresponding time point. While preserving the original information, the building-related features are enhanced to obtain the feature-enhanced image. and : (7) Steps 1-3: The final input of each subnetwork Feature Enhancement and The image is constructed by stitching together the two time-point images and the comprehensive multi-feature map: (8) in This indicates a pixel-by-pixel stitching operation.

3. The data- and knowledge-driven self-organizing building change detection method according to claim 2, characterized in that, The remote sensing image is a high-resolution optical image or a SAR image.

4. The data- and knowledge-driven self-organizing building change detection method according to claim 3, characterized in that, The empirical coefficient Take a value of 0.04 to 0.

06.

5. The data- and knowledge-driven self-organizing building change detection method according to claim 2, characterized in that, Step 2 specifically involves: Step 2-1: Construct an adaptive multi-dataset building change detection network framework driven by both data and knowledge and determine the search space; Based on the framework of the DFANet network-adaptive multi-data unsupervised remote sensing image change detection algorithm, a building domain knowledge embedding mechanism is introduced to expand the network structure search scope, and a data knowledge dual-driven adaptive multi-dataset building change detection network DFANet-DK is constructed. DFANet-DK consists of an input module, a multi-feature fusion module, a multi-dataset network module, an edge feature extraction module, and an output module. The search space of the data-knowledge dual-driven adaptive multi-dataset building change detection network consists of three parts: Subnet architecture encoding: 0 / 1 binary encoding is used, where 0 indicates that the network layer does not exist and 1 indicates that it exists. The initial number of layers for each subnet is set to a fixed value. ; Feature structure encoding: Two-digit encoding is used. The tens digit indicates whether multiple domain features of the building are included in the fusion of this layer, and the units digit indicates whether the output of this layer participates in cross-dataset interaction. Marginal loss function weights: real number encoding is used, and the search range is [0, 1.5]. Let the population size be... The initial number of layers in the candidate network is L, and the minimum number of layers in the search network is L. The minimum number of layers in the feature structure is ; Step 2-2: Generate initial data knowledge dual-driven change detection network population using a preset individual coding method; When the input is A dataset of building change detection with a population size of [number missing]. At that time, each candidate network individual The encoding structure is as follows: (9) (10) in, Indicates the candidate network population, Represents the encoding vector of a candidate network individual. Indicates the first The architecture encoding of the sub-network corresponding to each building dataset. Indicates the first Encoding of the feature structure of each dataset This represents the weight encoding of the edge loss function; No. Each dataset and Decomposed into: (11) in, Indicates the first The subnet architecture encoding of layer l in each task Indicates the first The feature structure encoding of the l-th layer in each task is a two-digit encoding method, and the specific decoding method is as follows: (12) in, Does the representation of multiple building features enter the first stage? For each task subnet at layer l: when the value is 2, it enters fusion; when the value is 1, it does not merge. Representing the Whether the output of the l-th layer of the task subnet participates in the multi-channel multi-scale feature interaction pool: when the value is 1, it enters the multi-channel multi-scale feature interaction pool; when the value is 0, it is ignored. Randomly generated according to the above individual coding method Each individual constitutes an initial population of the data knowledge dual-driven change detection network shown in equation (9), where each individual represents a possible architecture of the data knowledge dual-driven change detection network. Steps 2-3: Iteratively optimize the population using region-constrained self-selection genetic operations; Design region-constrained self-selection genetic operations, including crossover and mutation operations; Step 2-3-1: Cross operation; Crossover operations include three types: structural region crossover, feature region crossover, and two-region crossover. The implementation process is as follows: First, in the population The algorithm randomly selects two individuals as parents using a binary tournament algorithm. and ,Will and Align the beginning and end. and The subnet architecture coding region and feature structure coding region are automatically aligned; then a random number between 0 and 1 is randomly generated. ,when Less than the crossover probability If the condition for crossover is met, it means that the type of crossover operation needs to be further determined: Then randomly generate a number between 0 and 1. , representing the type of intersection; when When the value is less than 0.3, it indicates that structural regions intersect, respectively in and Within the subnet architecture coding area, randomly select a position as the intersection point, and swap the subnet architecture coding after the two parent intersection points; when When the value is between 0.3 and 0.6, it indicates that feature region crossing is performed, respectively in... and Randomly select a position within the feature structure encoding region as an intersection point and exchange the feature structure encodings; when When the value is greater than 0.6, it indicates a double-region intersection, requiring simultaneous... and Arbitrary positions are selected as intersection points in both the subnet architecture encoding region and the feature structure encoding region to achieve the cross-region operation; for the edge loss weights Directly to and of The average value is assigned to the offspring to ensure the continuity of the weight search, ultimately yielding the offspring individuals after crossover. and ; Step 2-3-2: Mutation operation; Mutation operations also include three types: structural region mutation, characteristic region mutation, and two-region mutation, and the implementation process is as follows: First, in the population An individual is randomly selected as the parent. Then generate a random number between 0 and 1. ,when Less than the mutation probability This indicates that the mutation condition is met, and further determination of the mutation type is needed: then randomly generate a random number between 0 and 1. ,when A value less than 0.3 indicates a structural region mutation. The subnet architecture coding region randomly selects a position as a mutation point, and inverts the corresponding bit encoding, i.e., changes 0 to 1 or 1 to 0; when A value between 0.3 and 0.6 indicates a mutation in the characteristic region. The feature structure encoding region randomly selects several positions as mutation points, and randomly generates new legal values ​​for the two-digit codes at the selected positions, i.e., the tens digit is 1 or 2, and the units digit is 0 or 1; when A value greater than 0.6 indicates a two-region mutation, and simultaneously... Arbitrary positions are selected as mutation points in both the subnet architecture encoding region and the feature structure encoding region to achieve dual-region mutation operation; for edge loss weights A new value is randomly generated within the range [0, 1.5] to replace the mutated offspring. ; Steps 2-4: Extract edge features and construct a deep edge-guided loss function to evaluate the fitness of each individual in the population; In the data-knowledge dual-driven change detection network, the edge feature extraction module simultaneously extracts change and edge information from the output features of the multi-channel, multi-scale feature interaction pool. Specifically, the output features of the multi-channel, multi-scale feature interaction pool are sequentially passed through a fully connected layer with a ReLU activation function, and then generated by two parallel 1×1 convolutional modules: a change probability map and an edge probability map. The change probability map is activated by a Softmax layer and output, corresponding to the change detection result, while the edge probability map is activated by a Softmax layer and output, corresponding to the building edge detection result. The ground truth edge labels are generated by processing the change detection ground truth map using the Canny algorithm. The deep edge-guided loss function is derived from the change detection loss. and edge detection loss The weighted combination results in a total loss of: (13) Among them, change detection loss for The sum of cross-entropy losses for each building change detection task: (14) in, For the first The actual category tags for each task For the first The predicted probability for each task The cross-entropy loss function; Edge detection loss Binary cross-entropy loss for building edge detection: (15) in, For pixels The edge true value (0 or 1). For pixels The edge prediction probability is summed over all pixels in the image. Perform calculations; Edge loss weights Automatically determined by evolutionary adaptive search, its physical meaning is the relative proportion of the loss contribution of the change detection task and the edge detection task during training, which is used to dynamically adjust the balance between change detection accuracy and building edge sharpness during training. Steps 2-5: Determine whether the current iteration count has reached the preset maximum iteration count; Step 2-6: If not, return to the genetic manipulation steps described in Step 2-3 to continue evolution; Steps 2-7: If yes, stop the iteration and select the individual with the highest fitness from each updated population. As the optimal individual.

6. The data- and knowledge-driven self-organizing building change detection method according to claim 5, characterized in that, Step 3 specifically involves: Step 3-1: Connected Component Labeling; Perform connected component analysis on the change detection binary graph output by the network to identify connected regions in the image where all pixels have a value of 1, i.e., the changed regions, and label and count the area of ​​each connected region; Step 3-2: Noise Removal; Calculate the area of ​​each connected region and remove scattered noise connected regions whose area is lower than a preset minimum area threshold; Step 3-3: Filling voids; Perform internal hole filling operation on the preserved connected regions; take advantage of the connectivity feature that the internal spaces of the building are interconnected and form a regular closed shape as a whole, fill the internal holes enclosed by the boundaries of each connected region, that is, the areas with an internal pixel value of 0, with 1, so that the detection of the main area of ​​the building is smoother and more complete. Finally, output simultaneously Building change detection results , , … ,,correspond One input building change detection dataset.

7. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 6.

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

9. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the method as described in any one of claims 1 to 6.

Images