Methods, devices, equipment and media for generating remote sensing image change detection data

Abstract

This application discloses a method, apparatus, device, and medium for generating remote sensing image change detection data, belonging to the field of remote sensing image processing technology. The method for generating remote sensing image change detection data includes: acquiring an existing semantic label map; changing the label categories corresponding to at least some pixels of the existing semantic label map to obtain a synthetic semantic label map; determining change label data between the existing semantic label map and the synthetic semantic label map; generating a first remote sensing image corresponding to the synthetic semantic label map; and obtaining remote sensing image change detection data based on the first remote sensing image, a second remote sensing image corresponding to the existing semantic label map, and the change label data. This application eliminates the need for manual registration and annotation, obtaining a large amount of remote sensing image change detection data for model training at low cost.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a method, apparatus, device and medium for generating remote sensing image change detection data. Background Technology

[0002] Remote sensing image change detection utilizes multi-source remote sensing images covering the same surface area at different times to determine and analyze changes in land features within that area, including changes in the location, extent, properties, and state of features. Deep learning has also been applied in remote sensing image change detection; therefore, deep learning-based remote sensing image change detection requires a large amount of change detection data to train the neural network.

[0003] However, existing change detection data requires manual comparison of two remote sensing images, pixel by pixel, to mark the areas where changes have occurred. Different application scenarios require engineers with different specialties to analyze remote sensing images, resulting in a shortage of change detection data samples needed for neural network training. Summary of the Invention

[0004] The main objective of this application is to provide a method, apparatus, device, and medium for generating change detection data in remote sensing images, aiming to solve the technical problem of the scarcity of change detection data samples required for training existing neural networks.

[0005] To achieve the above objectives, this application provides a method for generating remote sensing image change detection data, the method comprising:

[0006] Obtain the existing semantic tag graph;

[0007] Change the label category of at least some pixels in the existing semantic label map to obtain a synthetic semantic label map;

[0008] Identify the changed label data between the existing semantic label graph and the synthesized semantic label graph;

[0009] Generate the first remote sensing image corresponding to the synthesized semantic label map;

[0010] Based on the first remote sensing image, the second remote sensing image corresponding to the existing semantic label map, and the change label data, remote sensing image change detection data is obtained.

[0011] In one possible embodiment of this application, changing the label category corresponding to at least some pixels of an existing semantic label map to obtain a synthetic semantic label map includes:

[0012] Identify the set of target connected components in an existing semantic label graph; wherein, the set of target connected components is the set of connected components corresponding to semantic labels of all categories or some categories;

[0013] Identify at least one connected component within the target set of connected components that needs to be modified;

[0014] Change the label category corresponding to each connected component to be changed to obtain a synthetic semantic label graph.

[0015] In one possible embodiment of this application, determining at least one connected component to be modified within the target connected component set includes:

[0016] Get the percentage of the changed area;

[0017] Based on the percentage of changed regions, obtain the number of connected components to be changed;

[0018] Based on the number of connected components to be changed, determine the connected components to be changed within the target set of connected components.

[0019] In one possible embodiment of this application, changing the label categories of each connected component to be changed to obtain a synthetic semantic label graph includes:

[0020] Based on the original label categories of the connected components to be changed, determine the target modification strategy;

[0021] Based on the target modification strategy, the label categories of each connected component to be modified are changed to obtain a synthetic semantic label graph.

[0022] In one possible embodiment of this application, generating a first remote sensing image corresponding to a synthetic semantic label map includes:

[0023] Determine the first temporal style information corresponding to the synthesized semantic tag map;

[0024] The first-phase style information and the synthesized semantic label map are input into the trained remote sensing image generative adversarial network to obtain the first remote sensing image.

[0025] In one possible embodiment of this application, before obtaining remote sensing image change detection data based on a first remote sensing image, a second remote sensing image corresponding to an existing semantic label map, and change label data, the method further includes:

[0026] Determine the second-phase style information corresponding to the existing semantic label map; the second-phase style information is different from the first-phase style information.

[0027] The second temporal style information and the existing semantic label map are input into the remote sensing image generative adversarial network to obtain the second remote sensing image.

[0028] In one possible embodiment of this application, determining the first temporal style information corresponding to the synthesized semantic tag map includes:

[0029] Acquire the first reference remote sensing image specified by the user;

[0030] Obtain the first specified temporal style information of the first reference remote sensing image;

[0031] Use the specified temporal style information as the first temporal style information; and / or

[0032] Determine the second-phase style information corresponding to the existing semantic label map, including:

[0033] Acquire a second reference remote sensing image as specified by the user;

[0034] Obtain the second specified temporal style information of the second reference remote sensing image;

[0035] Use the second specified phase style information as the second phase style information.

[0036] Secondly, this application also provides a remote sensing image change detection data generation device, comprising:

[0037] The data acquisition module acquires existing semantic tag graphs;

[0038] The data modification module changes the label category of at least some pixels in the existing semantic label map to obtain a synthetic semantic label map.

[0039] The change determination module is used to determine the changed label data between the existing semantic label graph and the synthesized semantic label graph;

[0040] The image generation module is used to generate the first remote sensing image corresponding to the synthesized semantic label map;

[0041] The sample generation module is used to obtain remote sensing image change detection data based on the first remote sensing image, the second remote sensing image corresponding to the existing semantic label map, and change label data.

[0042] In one possible embodiment of this application, the data modification module includes:

[0043] The connected component identification unit is used to identify the target connected component set in an existing semantic label graph; wherein, the target connected component set is the set of connected components corresponding to semantic labels of all categories or some categories;

[0044] A region determination unit is used to determine at least one connected component to be modified within the target connected component set.

[0045] The label category changing unit is used to change the label category corresponding to each connected component to be changed, thereby obtaining a synthetic semantic label graph.

[0046] In one possible embodiment of this application, the region determination unit is specifically used to obtain the percentage of the changed region; based on the percentage of the changed region, obtain the number of connected components to be changed; and based on the number of connected components to be changed, determine the connected components to be changed within the target connected component set.

[0047] In one possible embodiment of this application, the data modification module specifically includes:

[0048] The strategy determination unit is used to determine the target modification strategy based on the original label category of the connected component to be changed.

[0049] The data modification unit is used to change the label category of each connected component to be modified based on the target modification strategy, and obtain a synthetic semantic label graph.

[0050] In one possible embodiment of this application, the image generation module specifically includes:

[0051] The first style determination unit is used to determine the first temporal style information corresponding to the synthesized semantic tag map;

[0052] The first image generation unit is used to input the first temporal style information and the synthesized semantic label map into the trained remote sensing image generative adversarial network to obtain the first remote sensing image.

[0053] In one possible embodiment of this application, the first style determination unit is specifically used to acquire a first reference remote sensing image determined by the user; obtain first specified temporal style information of the first reference remote sensing image; and use the specified temporal style information as the first temporal style information.

[0054] In one possible embodiment of this application, the apparatus further includes:

[0055] The second style determination unit is used to determine the second-phase style information corresponding to the existing semantic label map; the second-phase style information is different from the first-phase style information.

[0056] The second image generation unit is used to input the second temporal style information and the existing semantic label map into the remote sensing image generative adversarial network to obtain the second remote sensing image.

[0057] In one possible embodiment of this application, the second style determination unit is specifically used to acquire a second reference remote sensing image determined by the user; obtain second specified temporal style information of the second reference remote sensing image; and use the second specified temporal style information as the second temporal style information.

[0058] Thirdly, this application also provides a remote sensing image change detection data generation device, including: a processor, a memory, and a remote sensing image change detection data generation program stored in the memory, wherein the remote sensing image change detection data generation program is executed by the processor to implement the steps of the remote sensing image change detection data generation method of the first aspect.

[0059] Fourthly, this application also provides a computer-readable storage medium storing a remote sensing image change detection data generation program, which, when executed by a processor, implements the remote sensing image change detection data generation method as described in the first aspect.

[0060] This application proposes a method for generating change detection data for remote sensing images. By modifying and editing the label categories corresponding to at least some pixels of an existing semantic label map in the remote sensing image domain, a synthetic semantic label map and change label data between the two are obtained. Then, the obtained synthetic semantic label map is used to generate a first remote sensing image, i.e., a new remote sensing image. This new remote sensing image and the second remote sensing image corresponding to the real semantic segmentation data form a change pair. At this time, the change pair and the change label constitute the change detection data required for neural network training.

[0061] It is easy to see that, in the field of remote sensing imagery, semantic segmentation data does not require registration, and manually annotating pixel-level semantic labels is easier than for change detection data. As a result, there is a large amount of real semantic segmentation data that has been manually annotated. Therefore, this application modifies and edits the semantic label map of real semantic segmentation data and performs image generation operations to obtain change detection data. It does not require manual registration and annotation, thus obtaining a large amount of change detection data at low cost. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the structure of a remote sensing image change detection data generation device in the hardware operating environment involved in the embodiments of this application;

[0063] Figure 2 This is a flowchart illustrating the first embodiment of the remote sensing image change detection data generation method of this application;

[0064] Figure 3 This is a flowchart illustrating the second embodiment of the remote sensing image change detection data generation method of this application;

[0065] Figure 4 This is a schematic diagram illustrating the generation of phase style information in this application;

[0066] Figure 5 This is a schematic diagram of the overall processing flow of an example of this application;

[0067] Figure 6 for Figure 5 A detailed illustration of how to edit the Chinese label;

[0068] Figure 7 This application provides an example of a semantic label map A-seg and its corresponding first remote sensing image A.

[0069] Figure 8 This application provides an example of synthesizing a semantic label map B-seg and its corresponding second remote sensing image B.

[0070] Figure 9 This is Label-1, representing the change in label data between semantic label graph A-seg and the synthetic semantic label graph B-seg in one example of this application.

[0071] Figure 10 For another example in this application, the semantic label map D-seg and the corresponding first remote sensing image D;

[0072] Figure 11 In another example of this application, a semantically labeled image E-seg and a corresponding second remote sensing image E are synthesized;

[0073] Figure 12 Label-2 is the variational label data between semantic label graph D-seg and synthetic semantic label graph E-seg in another example of this application;

[0074] Figure 13 This is a schematic diagram of the module of the remote sensing image change detection data generation device of this application.

[0075] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0076] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0077] Remote sensing image change detection utilizes multi-source remote sensing images covering the same surface area at different times to determine and analyze changes in land features within that area, including changes in the location, extent, properties, and state of features. Deep learning has also been applied in remote sensing image change detection; therefore, deep learning-based remote sensing image change detection requires a large amount of change detection data to train the neural network.

[0078] Generally, acquiring remote sensing change detection data involves the following process: First, two remote sensing images of the same area at different times are acquired. Then, due to the differences in capture time and angle, the two images need to be georeferenced. This registration process requires manual identification of a certain number of identical feature points on the two images, which is very time-consuming. Finally, the two remote sensing images are manually compared pixel by pixel, and the areas where changes have occurred are marked to obtain change labels. Only then is change detection data, consisting of change labels and the two remote sensing images, obtained—essentially a change detection sample used during training.

[0079] Understandably, manual registration and annotation lead to a shortage of change detection data samples needed for neural network training, which means that obtaining a large amount of change detection data is costly.

[0080] To address this, this application provides a solution. Since semantic segmentation data does not require registration in the field of remote sensing imagery, and manually annotating pixel-level semantic labels is easier than for change detection data, there is a large amount of manually annotated real semantic segmentation data. Based on this, changes and image generation operations are performed on the semantic label map of the real semantic segmentation data to obtain change detection data. This eliminates the need for manual registration and annotation, thus allowing for the low-cost acquisition of a large amount of change detection data.

[0081] The inventive concept of this application is further illustrated below with reference to some specific embodiments.

[0082] The following explains the technologies involved in the embodiments of this application:

[0083] Generative Adversarial Networks (GANs) are deep learning models and one of the most promising unsupervised learning methods on complex distributions in recent years. The model produces reasonably good outputs through a game-like learning process between (at least) two modules: a generative model and a discriminative model. In original GAN ​​theory, it is not required that both G and D be neural networks; they only need to fit the corresponding generative and discriminative functions. However, in practice, deep neural networks are generally used as G and D. In this embodiment, the generative adversarial network is used to generate remote sensing images based on given semantic segmentation data.

[0084] Semantic segmentation is used to assign a semantic label category to each pixel in an image and associate the semantic label category with the pixel. In the field of remote sensing imagery, the manual annotation of semantic segmentation data is also pixel-level, that is, semantic label categories are assigned to different ground features pixel by pixel for a single remote sensing image.

[0085] The following embodiments of this application will describe the remote sensing image change detection equipment used in the technical implementation of this application:

[0086] like Figure 1As shown, the remote sensing image change detection data generation device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0087] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the remote sensing image change detection data generation device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0088] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a data storage module, a network communication module, a user interface module, and a remote sensing image change detection data generation program.

[0089] exist Figure 1 In the remote sensing image change detection data generation device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the remote sensing image change detection data generation device of this application can be set in the remote sensing image change detection data generation device. The remote sensing image change detection data generation device calls the remote sensing image change detection data generation program stored in the memory 1005 through the processor 1001 and executes the remote sensing image change detection data generation method provided in the embodiment of this application.

[0090] Based on, but not limited to, the above hardware structure, this application provides a first embodiment of a method for generating remote sensing image change detection data. (Refer to...) Figure 2 , Figure 2 A flowchart illustrating the first embodiment of the remote sensing image change detection data generation method is shown.

[0091] It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0092] In this embodiment, the method includes:

[0093] Step S100: Obtain the existing semantic tag map.

[0094] In this embodiment, the entity executing the remote sensing image change detection data is a remote sensing image change detection data generation device. This device can be a user equipment (UE) such as a mobile phone, smartphone, laptop, digital broadcast receiver, personal digital assistant (PDA), or tablet computer (PAD), or a handheld device, vehicle-mounted device, wearable device, computing device, or other processing device connected to a wireless modem, as well as a mobile station (MS). The remote sensing image change detection data is connected to the Internet to access publicly available databases in the field of remote sensing imagery.

[0095] The existing semantic label maps are semantic label maps from publicly available semantic segmentation data of remote sensing images in the field of remote sensing imagery. Understandably, since semantic segmentation data in the field of remote sensing imagery does not require registration, it is easier for professional engineers to manually annotate pixel-level semantic labels compared to change detection data. Therefore, there is a large amount of manually annotated, high-precision real semantic segmentation data. This large amount of manually annotated semantic segmentation data can be stored in one or more public databases in the field of remote sensing imagery, such as the multi-label category semantic segmentation database loveDA and the building semantic segmentation database xview2. Therefore, when the remote sensing image change detection data generation device executes step S100, it can obtain the aforementioned real semantic segmentation data from public databases via networks or other means, and extract the existing semantic label maps from it.

[0096] Step S200: Change the label category corresponding to at least some pixels of the existing semantic label map to obtain a synthetic semantic label map.

[0097] Step S300: Determine the changed label data between the existing semantic label graph and the synthesized semantic label graph.

[0098] It should be noted that when the remote sensing image change detection data generation device performs step S200, it can change the label category of the semantic label corresponding to some pixels to other label categories in the existing label categories of the existing semantic label map, or it can change the label category corresponding to some pixels to other label categories in the existing label categories of the application field to which the semantic label map belongs.

[0099] Understandably, in remote sensing imagery, semantic labels can be categorized based on the attributes of land features, such as buildings, farmland, grassland, and water bodies.

[0100] In one example, the existing semantic labels in the semantic label graph include building and background. In this case, the label category of pixels in a certain area can be changed from building to background.

[0101] Alternatively, in another example, if the existing semantic label map contains labels for buildings and background, the label category of the semantic labels corresponding to some pixels can be changed from buildings to water.

[0102] It is worth mentioning that in this embodiment, when performing step S200, the synthesized semantic label map is based on the existing semantic label map, and both always point to or reflect the same spatial region. This ensures that the resulting change detection data contains both changed and unchanged regions between the changed pairs. Thus, after modifying the existing semantic label map, the existing and synthesized semantic label maps contain both identical and different regions; the dataset consisting of these different regions is the change label data.

[0103] The specific format of the changed label data can be constructed as a binary changed label map or other formats. For example, on a black mask of the same size, the changed area can be segmented, and its pixel value can be set to 255 to obtain a black and white binary changed label map. Alternatively, the pixel value of the changed area can be determined according to the label category before and / or the label category after the change, to reflect the change in label category in more detail.

[0104] Understandably, since semantic segmentation assigns a semantic label category to each pixel in an image, the object corresponding to the changed label category during step S200 is the pixel. However, performing changes on pixels may result in a mixture of semantic labels from multiple categories within a single connected component, leading to poor realism. Therefore, to improve the realism of the final generated change detection data, as a specific implementation, step S300 specifically includes:

[0105] Step S310: Identify the set of target connected components in the existing semantic label graph.

[0106] The target connected component set is the set of connected components corresponding to the semantic labels of all or some categories.

[0107] A connected component is generally an image region consisting of foreground pixels that share the same pixel and are adjacent in position. Understandingly, a semantic label map includes the relationships between each pixel in an image and its corresponding label category. That is, in a semantic label map, the image has already been segmented into multiple closed and independent contour regions based on semantic segmentation results; each individual block-like region is a single connected component. Therefore, in an existing semantic label map, the image is segmented according to the semantic label category; pixels belonging to the same label category are segmented into the same connected component, pointing to a continuous region within the spatial area. In an existing semantic label map, the set of at least one connected component corresponding to the semantic labels of all label categories constitutes the target connected component set. Of course, if only a portion of the existing semantic label map is modified, then the target connected component set could be the set of at least one connected component corresponding to some semantic labels from all categories.

[0108] It's worth noting that the target set of connected components can be determined based on factors such as the number of label categories in the existing semantic label graph, the distribution of individual connected components within it, or the relative size of individual connected components (e.g., the average or median of connected components corresponding to semantic labels across all label categories). Alternatively, it can be pre-configured by the user. For example, in one instance, if the existing semantic label graph contains only two label categories, the total number of connected components can correspond to only one label category. Alternatively, in another instance, if the existing semantic label graph contains seven label categories, the set of connected components corresponding to the semantic labels of all label categories except for one (e.g., background) can be used as the aforementioned set of connected components.

[0109] Step S320: Determine at least one connected component to be changed within the target connected component set.

[0110] When performing step S320, the connected components to be modified can be determined according to a preset selection strategy or randomly. Thus, the number of connected components to be modified can be only one, or it can include multiple connected components, each of which has its own semantic label.

[0111] In one example, if the user pre-configures a preset selection rule, that is, to change the connected components corresponding to all semantic labels of "bare land" at a certain location, then after selecting the corresponding semantic label graph, the connected components corresponding to "bare land" can be determined as the connected components to be changed.

[0112] In addition, as a specific implementation, when performing step S320, the percentage of the changed region can be obtained; based on the percentage of the changed region, the number of connected components to be changed can be obtained; and then based on the number of connected components to be changed, at least one connected component to be changed in the target connected component set can be determined.

[0113] The changed region ratio refers to the proportion of changed regions in the existing semantic label map within the entire semantic label map; that is, the ratio of the number of connected components to be changed to the total number of connected components in the target connected component set. Users can pre-configure the corresponding changed region ratio, so that the remote sensing image change detection data generation device can call the user-configured changed region ratio when executing step S320. Of course, if the user has not configured it, the previous changed region ratio from a historical task or a default value can be used. After obtaining the changed region ratio, the number of connected components to be changed can be calculated, and then a number of connected components to be changed can be randomly selected.

[0114] Step S330: Change the label category of each connected component to be changed to obtain a synthetic semantic label graph.

[0115] Once the connected components to be modified are identified, the semantic labels of all connected components to be modified can be changed to obtain a synthetic semantic label graph.

[0116] It is understandable that when performing step S330, if the connected component to be changed includes a single connected component, the label category of the semantic label corresponding to that single connected component can be changed. If the connected component to be changed includes multiple connected components, the label category of the semantic label corresponding to each connected component to be changed is changed to a predetermined label category.

[0117] It is worth mentioning that, when performing step S330, the label categories of each connected component to be modified can be changed according to the target modification strategy to obtain a synthetic semantic label graph, or the label categories of each connected component to be modified can be randomly changed to obtain a synthetic semantic label graph.

[0118] The target modification strategy can be determined in the following ways: (1) Determine the target modification strategy based on the original label categories of the connected components to be modified; (2) Change the label categories of each connected component to be modified based on the target modification strategy to obtain a synthetic semantic label graph.

[0119] The modification strategy can be a pre-configured tag category change mapping relationship, which includes the tag category before modification and the corresponding tag category after modification. Therefore, during step S330, the tag category change mapping relationship can be consulted based on the original tag category of the connected component to be modified, thereby determining the corresponding mapping relationship and identifying it as the target modification strategy. The tag category of the connected component to be modified is then changed according to the modified tag category recorded in the target modification strategy.

[0120] Alternatively, the modification strategy could involve a pre-configured label category grouping table. In this table, all semantic labels within the same label category group are modified to the same result semantic label from other label categories, and the same result semantic label differs between different label category groups. Therefore, during step S330, the label category grouping table can be consulted based on the original label categories of the connected component to be modified, thus determining the corresponding label category group. The mapping relationship corresponding to the label category group is then defined as the target modification strategy. Finally, the label categories of the connected component to be modified are changed according to the modified label categories recorded in the target modification strategy.

[0121] Alternatively, the modification strategy can be a combination of the above modification strategies. For example, in one example, the existing semantic label graph includes 8 label categories: A, B, C, D, E, F, and G. Among them, A, B, C, D, and E need to be modified. The modification strategy is as follows: A, B, and C are divided into the first group, and their corresponding result semantic label is D; D and E are divided into the second group, and the label category of their corresponding result semantic labels is not restricted and can be modified randomly.

[0122] Of course, the modified strategy can also be configured by the user in advance to other strategies.

[0123] It is easy to understand that by changing the region proportion and / or configuring the strategy mentioned above, the semantic label map can be edited in a controlled manner, and a large number of synthetic semantic label maps with better realism can be generated from a small number of existing semantic label maps, thereby providing richer change detection data samples.

[0124] Step S400: Generate the first remote sensing image corresponding to the synthesized semantic label map.

[0125] Step S500: Based on the first remote sensing image, the second remote sensing image corresponding to the existing semantic label map, and the change label data, obtain remote sensing image change detection data.

[0126] Since the synthesized semantic label map is not an existing semantic label map, it does not have a corresponding remote sensing image. Therefore, in order to construct the required change detection data, it is necessary to generate a first remote sensing image corresponding to the synthesized semantic label map. Understandably, as an option in this embodiment, an existing remote sensing image generation neural network model can be used to generate the first remote sensing image.

[0127] Alternatively, as another option in this embodiment, a new remote sensing image generation neural network model can be retrained using the existing semantic label map corresponding to the real remote sensing image. The synthesized semantic label map is then input into the corresponding remote sensing image generation neural network model to obtain the first remote sensing image. In this case, the obtained first remote sensing image is closer to the real remote sensing image, thus making the final remote sensing image change detection data more realistic.

[0128] Alternatively, as another option in this embodiment, a new remote sensing image generation neural network model can be retrained using the real remote sensing images corresponding to the aforementioned semantic label maps. Then, the synthesized semantic label map and the existing semantic label map are input into the remote sensing image generation neural network model, respectively, resulting in two remote sensing images output by the model: the first remote sensing image and the second remote sensing image. In this case, both the first and second remote sensing images are synthesized images, which improves the training effect of the remote sensing image change detection data composed of these two images and the change label data.

[0129] Thus, the second remote sensing image in this embodiment can be a real remote sensing image that can be found in an existing remote sensing database, or it can be a newly generated synthetic remote sensing image that is executed during the method steps provided in this embodiment.

[0130] After obtaining the first remote sensing image, the second remote sensing image, and the change label data, remote sensing image change detection data can be constructed, which can then be used as training samples for training the remote sensing image change detection neural network. It is understandable that the existing semantic label map, the synthesized semantic label map, the first remote sensing image, and the second remote sensing image are all of the same size.

[0131] It is easy to see that manual registration and annotation lead to a shortage of change detection data samples needed for neural network training. In the field of remote sensing imagery, segmentation data does not require registration, and manual annotation of pixel-level semantic labels is easier than that of change detection data. As a result, there is a large amount of real semantic segmentation data that has been manually annotated. Based on this, this embodiment modifies and edits the semantic label map of real semantic segmentation data and performs image generation operations to obtain change detection data. It does not require manual registration and annotation, thus obtaining a large amount of change detection data at low cost.

[0132] Based on the above embodiments, a second embodiment of the remote sensing image change detection data generation method of this application is proposed.

[0133] See Figure 3 In this embodiment, step S400 specifically includes:

[0134] Step S410: Determine the first temporal style information corresponding to the synthesized semantic tag map.

[0135] Step S420: Input the first temporal style information and the synthesized semantic label map into the trained remote sensing image generative adversarial network to obtain the first remote sensing image.

[0136] In the field of remote sensing imagery, remote sensing change detection typically involves two remote sensing images of the same spatial area taken at different times. Therefore, for the same spatial area, the same ground feature may not only change its label category over time, but more importantly, its color may change. This color variation can be caused by natural seasonal changes or human factors, resulting in a unique imaging style in the color scheme and spatial distribution within the entire area. Furthermore, changes in the hardware or imaging algorithms of the remote sensing camera over time can also lead to differences in the final image quality, which are also reflected in the imaging style. Therefore, this embodiment utilizes temporal style information to reflect the image style of the remote sensing image.

[0137] Thus, when performing step S400, the first phase style information for generating the first remote sensing image can be selected from multiple candidate phase style information.

[0138] Understandably, the first phase style information corresponding to the synthesized semantic label map can be determined randomly or based on a preset style filtering strategy.

[0139] For example, the preset style filtering strategy can be determined based on real remote sensing images with existing semantic label maps. Specifically, when executing step 410, the user-determined first reference remote sensing image is obtained, and then the first specified temporal style information of the first reference remote sensing image is obtained; finally, the first specified temporal style information is used as the first temporal style information.

[0140] In other words, the user can specify a first reference remote sensing image, which can be a remote sensing image publicly available in a database. Then, an image imaging style calculator can be used to calculate its imaging style, obtaining the first specified temporal style information. This is then used as the imaging style of the first remote sensing image to be generated in this embodiment. Thus, the first temporal style information is finally selected from the imaging styles of publicly available remote sensing images.

[0141] Alternatively, the preset style filtering strategy can be determined based on dimensions such as terrain style, season, or remote sensing satellite model. Specifically, during step 410, a style selection interface is displayed to the user, including a terrain style selection sub-interface, such as desert city style controls, water town style controls, frigid city style controls, Central Plains village style controls, and megacity style controls; a season selection sub-interface, such as spring controls, summer controls, rainy season controls, and welding season controls; and a remote sensing satellite selection sub-interface, etc. The user selects the corresponding controls according to their own needs, thereby obtaining the corresponding preset style filtering strategy. Based on the preset style filtering strategy, the first phase style information that meets the user's requirements is filtered from multiple candidate phase style information.

[0142] Alternatively, as one implementation method, multiple imaging styles can be learned from existing real remote sensing images to obtain the aforementioned multiple candidate temporal style information. Specifically, when training a remote sensing image generative adversarial network using multiple semantic segmentation samples, the actual temporal style information corresponding to each semantic segmentation sample is obtained. Then, the actual temporal style information is used as candidate temporal style information.

[0143] In one example, training a remote sensing image generative adversarial network using multiple semantic segmentation samples could specifically include:

[0144] Step A10: Input a semantic label map of size n×n×1 from the real semantic segmentation data as a semantic segmentation sample into the generative adversarial network (GAN) model to obtain a predicted image of size n×n×3. In one example, the predicted image can be an RGB three-channel image. Understandably, the expected output of the generative model is the real remote sensing image corresponding to the semantic label. Calculate the first error between the predicted image output by the GAN and the expected output.

[0145] Step A20: Concatenate the obtained n×n×3 predicted image and n×n×1 semantic label map to form an n×n×4 first feature map. Input the first feature map into the discriminative model of the generative adversarial network, and the discriminative model outputs a first discrimination result. It is easy to understand that the first discrimination result output by the discriminative model is a probability value between 0 and 1. Since the input image is the predicted image, the expected output of the discriminative model is false, i.e., 0. The second error is calculated by passing the first discrimination result and the expected output 0 through a loss function.

[0146] Step A30: The real remote sensing image corresponding to the real semantic segmentation data is concatenated with the n×n×1 semantic label map to form an n×n×4 second feature map, which is then input into the discriminant model. The discriminant model outputs a second discrimination result. The second discrimination result is also a probability value between 0 and 1. Since the input image is the real remote sensing image corresponding to the real semantic segmentation data, the expected output of the discriminant model is true, i.e., 1. The second discrimination result and the expected output 1 are used to calculate the third error through a loss function.

[0147] Step A40: Backpropagate the first error, second error, and third error to update the model parameters of the generative adversarial network.

[0148] Obtain new real semantic segmentation data from the real semantic segmentation dataset, and then repeat steps A10 to A20 until the generative model and the discriminative model of the generative adversarial network reach a balance state. The training is then complete, and the trained remote sensing image generative adversarial network is obtained.

[0149] Please see Figure 4 In this example, the remote sensing image generative adversarial network (GAN) also includes a style encoder, which calculates the style mask for each real remote sensing image input to the GAN, i.e., the actual temporal style information. During the generation of predicted images by the generative model, the corresponding actual temporal style information is also used as input in step A10 to control the style of the generated prediction network. Therefore, after training, the GAN not only calculates a relatively rich amount of realistic actual temporal style information, but the trained GAN can also be trained to produce remote sensing images with a style corresponding to and highly realistic than real remote sensing images based on the input semantic label map and the actual temporal style information. In one example, all style masks are stored in the GAN's style library, with each style mask named after a real remote sensing image, thus serving as candidate temporal style information.

[0150] Furthermore, as one implementation method, the second remote sensing image can be obtained in the following way:

[0151] Step B10: Determine the second-phase style information corresponding to the existing semantic tag map.

[0152] The style information of the second phase is different from that of the first phase.

[0153] Step B20: Input the second temporal style information and the existing semantic label map into the remote sensing image generative adversarial network to obtain the second remote sensing image.

[0154] The second temporal style information is used to control the style of the generated second remote sensing image, so that the imaging style of the final generated second remote sensing image is inconsistent with the imaging style of the first remote sensing image, thereby simulating the style differences between different phases of real dual-temporal remote sensing image data. How the second temporal style information is determined can be referred to how the first temporal style information is determined in the aforementioned embodiment, and will not be repeated in this embodiment.

[0155] Alternatively, the user can specify the second reference remote sensing image during step B10; obtain the second specified temporal style information of the second reference remote sensing image; and use the second specified temporal style information as the second temporal style information. In this way, an imaging style different from the first temporal style information is finally selected from the imaging styles of the publicly available remote sensing images.

[0156] As is easily understood, this embodiment sets different imaging styles for the generated first remote sensing image and the generated second remote sensing image to simulate the style differences between different phases of real dual-temporal remote sensing image data, and avoids the remote sensing image change detection neural network from falsely detecting non-ground objects in different phases of reality.

[0157] To enable those skilled in the art to better understand the scope of protection of the claims of this application, specific implementation examples in specific application scenarios are used to explain and illustrate the technical solutions described in the claims of this application. It should be understood that the following examples are only used to explain this application and are not intended to limit the scope of protection of the claims of this application.

[0158] Example 1: Building change detection is a fundamental and very important problem in the field of remote sensing image change detection. It is used to predict whether there are changes such as the construction or demolition of buildings in remote sensing images over multiple time periods.

[0159] Please see Figure 5 , Figure 5 This is a schematic diagram of the overall processing flow for Example 1. Please refer to [link / reference]. Figure 6 , Figure 6 This is a detailed diagram illustrating the semantic tag modification and editing steps in the overall processing workflow. Please refer to [link / reference]. Figure 7 The image on the left is the semantic label map A-seg, and the image on the right is the first remote sensing image A. Figure 8 In the image, the left side shows the synthetic semantic label image B-seg, and the right side shows the second remote sensing image B. Figure 9 Label-1 is the variable label data.

[0160] Please see Figure 5In this example, a generative adversarial network (GAN) is first trained using real semantic segmentation data from the publicly available building semantic segmentation database xview2, resulting in a trained remote sensing image GAN. During training, the actual temporal style information reflecting the imaging style of each real remote sensing image is obtained by calculating the style mask of the remote sensing image corresponding to the real semantic segmentation data. All actual temporal style information is stored in a style library. The input to the generative model of the GAN is the semantic label map and the actual temporal style information, and the output is the predicted synthesized image corresponding to the semantic label map.

[0161] Next, extract all semantic label images from the xview2 dataset as existing semantic label images. Each semantic label image contains only two label categories: background (black area, pixel value 0) and building (white area, pixel value 255). Take any semantic label image A-seg and edit its semantic labels. Specifically, first, use a connected component analysis algorithm to obtain the set of connected components C corresponding to all building semantic labels in semantic label image A-seg. Then, randomly select n (where n is a natural number, and the specific value can be configured in advance) independent building connected components from the connected component set C as the connected components to be modified. Next, change the label category of these selected connected components, changing their corresponding semantic labels from building to background, thus obtaining the synthesized semantic label image B-seg. Specifically, the pixel values ​​within these selected independent building connected components can be changed from 255 to 0.

[0162] Then, based on the changes between semantic label graph A-seg and semantic label graph B-seg, the changed label data Label-1 is obtained, such as... Figure 9 As shown. Specifically, on a black mask with all pixel values ​​of 0 and the same size as the semantic label map A-seg, the pixel values ​​of the corresponding regions where pixel values ​​change between semantic label map A-seg and semantic label map B-seg are changed from 0 to 255, resulting in the change label data Label-1. This reflects the difference between semantic label map A-seg and semantic label map B-seg. This change label data Label-1 is a binary change label.

[0163] Next, the first temporal style information S1 of the semantic label map A-seg and the second temporal style information S2 of the semantic label map B-seg are determined. Specifically, the user can select two real remote sensing images from the xview2 dataset and assign the styles of the two real remote sensing images to the first temporal style information S1 and the semantic label map B-seg, respectively, thus obtaining the first temporal style information S1 and the second temporal style information S2.

[0164] Finally, the generative model of the remote sensing image generative adversarial network is extracted. The semantic label map A-seg and the first temporal style information S1 are input into the generative model to obtain the first remote sensing image A. The semantic label map B-seg and the second temporal style information S2 are input into the generative model to obtain the second remote sensing image B. The first remote sensing image A, the second remote sensing image B, and the change label data Label-1 constitute a pair of building change detection data.

[0165] Example 2: Generating general change detection data using multi-label category semantic segmentation data

[0166] A general change detection neural network for remote sensing image change detection is used to extract areas where land cover types have changed between two sets of remote sensing images. The general change detection identifies a wide variety of change types, not only recognizing semantic changes but also further identifying the change label categories of changed patches.

[0167] Please see Figure 10 The image on the left is the semantic label map D-seg, and the image on the right is the first remote sensing image D. Figure 11 In the image, the left side shows the synthesized semantic label map E-seg, and the right side shows the second remote sensing image E. Figure 12 Label-2 is the variable label data.

[0168] First, a generative adversarial network (GAN) is trained using real multi-label semantic segmentation data from the publicly available multi-label semantic segmentation dataset loveDA, resulting in a trained remote sensing image GAN. During training, the actual temporal style information reflecting the imaging style of each multi-label real remote sensing image is obtained by calculating the style mask of the remote sensing image corresponding to the real multi-label semantic segmentation data. All actual temporal style information is stored in a style database. The input to the generative model of the GAN is the semantic label map and the actual temporal style information; the output is the predicted synthesized image corresponding to the semantic label map.

[0169] Next, extract all semantic label graphs from the loveDA database. These graphs contain seven semantic label categories: buildings, roads, water bodies, bare land, woodland, agricultural land, and background. Take any semantic label graph D-seg and use a connected component analysis algorithm to obtain the target connected component set F for all label categories in D-seg. Randomly select m (where m is a natural number, and its value can be pre-configured) independent contour regions from the target connected component set F as the connected components to be modified. The label category of the connected components to be modified can be any of the seven semantic label categories; there are no restrictions. Randomly change the semantic label category of the selected connected components to be modified to another semantic label from the seven categories to obtain the synthetic semantic label graph E-seg.

[0170] Then, based on the changes between the semantic label graph D-seg and the semantic label graph E-seg, the changed label data Label-2 is obtained, such as... Figure 12 As shown in the diagram. Specifically, on a black mask with all pixel values ​​of 0 and the same size as the semantic label map D-seg, areas where the feature label category changes are first marked. Then, the pixel values ​​of areas where the semantic label of a building changes to another label category are changed from 0 to 255. If it is a semantic label other than a building, the pixel value of the corresponding area is changed from 0 to 128. In this way, the resulting change label data Label-2 has three pixel values, representing the three change label categories: 0 represents no semantic label change, 128 represents a semantic change for non-building features, and 255 represents a semantic change for buildings. Compared to binary change labels, the change label data Label-2 in this example can display a more detailed classification of change label categories.

[0171] Next, the first temporal style information S3 of the semantic label map D-seg and the second temporal style information S4 of the semantic label map E-seg are determined. Specifically, the user can select two multi-label class real remote sensing images from the loveDA dataset, and assign the styles of the two multi-label class real remote sensing images to the first temporal style information S3 and the semantic label map E-seg, respectively, thus obtaining the first temporal style information S1 and the second temporal style information S4.

[0172] Finally, the generative model of the remote sensing image generative adversarial network is extracted. The semantic label map D-seg and the first temporal style information S1 are input into the generative model to obtain the first remote sensing image D. The semantic label map E-seg and the second temporal style information S2 are input into the generative model to obtain the second remote sensing image B. The first remote sensing image D, the second remote sensing image E, and the change label data Label-2 constitute a pair of building change detection data.

[0173] It is easy to see that the embodiments of this application utilize existing semantic segmentation data to generate change detection data, bypassing the problems of difficult registration and labeling of real dual-temporal change detection data, thereby obtaining a large number of change detection data with high realism. Moreover, as in the example above, by introducing temporal style information, style change control of the generated synthetic remote sensing image is achieved to simulate the temporal style changes of real change detection data.

[0174] Based on the same inventive concept, please refer to Figure 13 Secondly, this application also provides a remote sensing image change detection data generation device, including:

[0175] The data acquisition module acquires existing semantic tag graphs;

[0176] The data modification module changes the label category of at least some pixels in the existing semantic label map to obtain a synthetic semantic label map.

[0177] The change determination module is used to determine the changed label data between the existing semantic label graph and the synthesized semantic label graph;

[0178] The image generation module is used to generate the first remote sensing image corresponding to the synthesized semantic label map;

[0179] The sample generation module is used to obtain remote sensing image change detection data based on the first remote sensing image, the second remote sensing image corresponding to the existing semantic label map, and change label data.

[0180] In one possible embodiment of this application, the data modification module includes:

[0181] The connected component identification unit is used to identify the target connected component set in an existing semantic label graph; wherein, the target connected component set is the set of connected components corresponding to semantic labels of all categories or some categories;

[0182] A region determination unit is used to determine at least one connected component to be modified within the target connected component set.

[0183] The label category changing unit is used to change the label category corresponding to each connected component to be changed, thereby obtaining a synthetic semantic label graph.

[0184] In one possible embodiment of this application, the region determination unit is specifically used to obtain the percentage of the changed region; based on the percentage of the changed region, obtain the number of connected components to be changed; and based on the number of connected components to be changed, determine the connected components to be changed within the target connected component set.

[0185] In one possible embodiment of this application, the data modification module specifically includes:

[0186] The strategy determination unit is used to determine the target modification strategy based on the original label category of the connected component to be changed.

[0187] The data modification unit is used to change the label category of each connected component to be modified based on the target modification strategy, and obtain a synthetic semantic label graph.

[0188] In one possible embodiment of this application, the image generation module specifically includes:

[0189] The first style determination unit is used to determine the first temporal style information corresponding to the synthesized semantic tag map;

[0190] The first image generation unit is used to input the first temporal style information and the synthesized semantic label map into the trained remote sensing image generative adversarial network to obtain the first remote sensing image.

[0191] In one possible embodiment of this application, the first style determination unit is specifically used to acquire a first reference remote sensing image determined by the user; obtain first specified temporal style information of the first reference remote sensing image; and use the specified temporal style information as the first temporal style information.

[0192] In one possible embodiment of this application, the apparatus further includes:

[0193] The second style determination unit is used to determine the second-phase style information corresponding to the existing semantic label map; the second-phase style information is different from the first-phase style information.

[0194] The second image generation unit is used to input the second temporal style information and the existing semantic label map into the remote sensing image generative adversarial network to obtain the second remote sensing image.

[0195] In one possible embodiment of this application, the second style determination unit is specifically used to acquire a second reference remote sensing image determined by the user; obtain second specified temporal style information of the second reference remote sensing image; and use the second specified temporal style information as the second temporal style information.

[0196] It should be noted that the various implementation methods of the remote sensing image change detection data generation device in this embodiment and the technical effects they achieve can be referred to the various implementation methods of the remote sensing image change detection data generation method in the foregoing embodiments, and will not be repeated here.

[0197] Furthermore, embodiments of this application also propose a computer storage medium storing a remote sensing image change detection data generation program. When executed by a processor, the remote sensing image change detection data generation program implements the steps of the remote sensing image change detection data generation method described above. Therefore, it will not be repeated here. Additionally, the beneficial effects of using the same method will not be repeated here either. For technical details not disclosed in the computer-readable storage medium embodiments of this application, please refer to the description of the method embodiments of this application. As an example, program instructions can be deployed to execute on a single computing device, or on multiple computing devices located at one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0198] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0199] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided in this application, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0200] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0201] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for generating remote sensing image change detection data, characterized in that, The method includes: Obtain the existing semantic tag graph; By changing the label categories corresponding to at least some pixels of the existing semantic label map, a synthetic semantic label map is obtained; Determine the changed label data between the existing semantic label map and the synthesized semantic label map; Determine the first temporal style information corresponding to the synthesized semantic tag map; The first temporal style information and the synthesized semantic label map are input into the trained remote sensing image generative adversarial network to obtain the first remote sensing image; Determine the second temporal style information corresponding to the existing semantic tag map; the second temporal style information is different from the first temporal style information. The second temporal style information and the existing semantic label map are input into the remote sensing image generative adversarial network to obtain the second remote sensing image; Based on the first remote sensing image, the second remote sensing image, and the change label data, remote sensing image change detection data is obtained.

2. The method for generating remote sensing image change detection data according to claim 1, characterized in that, The step of changing the label category corresponding to at least some pixels of the existing semantic label map to obtain a synthetic semantic label map includes: Identify the target connected component set in the existing semantic label graph; the target connected component set is the set of connected components corresponding to semantic labels of all categories or some categories; Identify at least one connected component to be modified within the target connected component set; Change the label category corresponding to each of the connected components to be changed to obtain the synthesized semantic label graph.

3. The method for generating remote sensing image change detection data according to claim 2, characterized in that, Determining at least one connected component to be modified within the target connected component set includes: Get the percentage of the changed area; Based on the percentage of the changed region, the number of connected components to be changed is obtained; Based on the number of connected components to be changed, determine the connected components to be changed within the target connected component set.

4. The method for generating remote sensing image change detection data according to claim 3, characterized in that, The step of changing the label categories of each of the connected components to be changed to obtain the synthetic semantic label graph includes: Based on the original label category of the connected component to be changed, determine the target modification strategy; Based on the target modification strategy, the label categories of each of the connected components to be modified are changed to obtain the synthetic semantic label graph.

5. The method for generating remote sensing image change detection data according to claim 1, characterized in that, Determining the first temporal style information corresponding to the synthesized semantic tag map includes: Acquire the first reference remote sensing image specified by the user; Obtain the first specified temporal style information of the first reference remote sensing image; Use the specified temporal style information as the first temporal style information; and / or Determining the second temporal style information corresponding to the existing semantic tag map includes: Acquire a second reference remote sensing image as specified by the user; Obtain the second specified temporal style information of the second reference remote sensing image; Use the second specified temporal style information as the second temporal style information.

6. A remote sensing image change detection data generation device, characterized in that, include: The data acquisition module acquires existing semantic tag graphs; The data modification module changes the label category corresponding to at least some pixels of the existing semantic label map to obtain a synthetic semantic label map; The change determination module is used to determine the changed label data between the existing semantic label map and the synthesized semantic label map; The image generation module is used to generate the first remote sensing image corresponding to the synthesized semantic label map; The sample generation module is used to obtain remote sensing image change detection data based on the first remote sensing image, the second remote sensing image corresponding to the existing semantic label map, and the change label data; The image generation module specifically includes: The first style determination unit is used to determine the first temporal style information corresponding to the synthesized semantic tag map; The first image generation unit is used to input the first temporal style information and the synthetic semantic label map into the trained remote sensing image generative adversarial network to obtain the first remote sensing image. The device further includes: The second style determination unit is used to determine the second temporal style information corresponding to the existing semantic tag map; the second temporal style information is different from the first temporal style information. The second image generation unit is used to input the second temporal style information and the existing semantic label map into the remote sensing image generation adversarial network to obtain the second remote sensing image.

7. The remote sensing image change detection data generation device according to claim 6, characterized in that, The data modification module includes: A connected component identification unit is used to identify the target connected component set in the existing semantic label graph; wherein, the target connected component set is the set of connected components corresponding to semantic labels of all categories or some categories; A region determination unit is used to determine at least one connected component to be modified within the target connected component set; The label category changing unit is used to change the label category corresponding to each of the connected components to be changed, thereby obtaining the synthetic semantic label graph. Specifically, the region determination unit is used to: obtain the percentage of regions to be changed; obtain the number of connected components to be changed based on the percentage of regions to be changed; determine the connected components to be changed within the target connected component set based on the number of connected components to be changed; and / or The data modification module specifically includes: The strategy determination unit is used to determine the target modification strategy based on the original label category of the connected component to be changed; The data modification unit is configured to change the label category of each of the connected components to be modified based on the target modification strategy, thereby obtaining the synthetic semantic label graph; and / or The first style determination unit is specifically configured to acquire a first reference remote sensing image determined by the user; obtain first specified temporal style information of the first reference remote sensing image; use the specified temporal style information as the first temporal style information; and / or The second style determination unit is specifically used to acquire a second reference remote sensing image determined by the user; obtain second specified temporal style information of the second reference remote sensing image; and use the second specified temporal style information as the second temporal style information.

8. A remote sensing image change detection data generation device, characterized in that, include: A processor, a memory, and a remote sensing image change detection data generation program stored in the memory, wherein the remote sensing image change detection data generation program is executed by the processor to implement the steps of the remote sensing image change detection data generation method as described in any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a remote sensing image change detection data generation program, which, when executed by a processor, implements the steps of the remote sensing image change detection data generation method as described in any one of claims 1 to 5.

Images