Remote sensing image change detection method and device based on edge contour guidance

Abstract

The invention discloses a remote sensing image change detection method and device based on edge contour guidance. The method comprises the following steps: acquiring a remote sensing image; according to the remote sensing image, combining with a Canny operator to obtain a contour map; the remote sensing images comprise an early-stage remote sensing image and a late-stage remote sensing image which have a preset time interval; the profile diagrams comprise an early profile diagram and a late profile diagram; performing multi-time-step noise addition on the remote sensing image and the contour image, and performing preset multi-scale feature extraction in combination with a de-noising diffusion probability model to obtain multi-time-step feature images of the remote sensing image and the contour image under each scale; and performing enhanced fusion on the multi-time-step feature maps of the remote sensing map and the contour map under each scale, and sequentially fusing enhanced fusion results under each scale from small to large according to the scales to obtain a change detection result map.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a method and apparatus for detecting changes in remote sensing images based on edge contour guidance. Background Technology

[0002] Change detection algorithms accurately identify dynamic changes in land cover and the environment by comparing multi-temporal remote sensing images, playing a crucial role in areas such as urban expansion monitoring, disaster emergency response, ecological environment assessment, and agricultural management. They not only provide scientific basis for policy making and resource planning but also drive the automation and intelligent development of remote sensing technology, serving as a core technology for achieving dynamic updates of geographic information and supporting sustainable development decisions.

[0003] The core of change detection lies in accurately capturing dynamic changes on the Earth's surface. With the exponential growth of remote sensing data and the increasing complexity of application scenarios, traditional change detection methods are struggling to meet the dual challenges of high accuracy and robustness. Especially in complex scenarios such as changes in illumination, noise interference, and dynamic evolution of ground features, traditional methods often suffer from high false alarm rates and inaccurate boundary localization, becoming a key bottleneck restricting their practical application. Although existing neural network models have significantly improved detection performance through end-to-end feature learning, they still face two major challenges: First, insufficient scale perception, limiting their ability to capture subtle edge changes (such as road widening and vegetation boundary migration), especially prone to losing crucial geometric information under strong noise interference; second, lack of structural coherence, with generated change maps often exhibiting boundary breaks or discontinuities, leading to distortion of the topological relationship between the changed areas and the background. The root cause of these problems lies in the insufficient utilization of prior knowledge of spatial structure by existing models and their lack of ability to collaboratively model multi-scale features and the topological relationship of ground features.

[0004] Therefore, a new technical solution is urgently needed to address the technical problem of how to perform high-precision and robust remote sensing image change detection. Summary of the Invention

[0005] This invention provides a remote sensing image change detection method and apparatus based on edge contour guidance, which solves the technical problem of how to perform high-precision and high-robust remote sensing image change detection.

[0006] To achieve the above objectives, the present invention provides a remote sensing image change detection method based on edge contour guidance, comprising: Obtain remote sensing images; obtain contour maps by combining remote sensing images with the Canny operator; the remote sensing images include one early remote sensing image and one late remote sensing image with a preset time interval; the contour maps include one early contour map and one late contour map. Noise is added to the remote sensing image and the contour image at multiple time steps, and a pre-defined multi-scale feature extraction is performed in combination with a denoising diffusion probability model to obtain multi-time step feature maps of the remote sensing image and the contour image at each scale. The multi-timestep feature maps of remote sensing images and contour maps are enhanced and fused at each scale, and the enhancement and fusion results at each scale are fused sequentially from small to large to obtain the change detection result map.

[0007] Preferably, the enhancement and fusion of multi-timestep feature maps of remote sensing images and contour maps at various scales includes: Simultaneously at each scale, the multi-timestep feature maps of the remote sensing image and the contour image are stitched and fused along the channel dimension to obtain early and late augmented representations at each scale. Deep feature extraction and fusion of the early and late augmented representations are performed using two sets of consecutive convolutional layers and ReLU activation functions to obtain early and late feature representations at each scale. The absolute difference between the early and late feature representations is calculated to obtain the initial change feature map at each scale.

[0008] Preferably, the enhancement fusion results at each scale are sequentially fused from smallest to largest to obtain the change detection result map, including: At the minimum scale, the initial change feature map is adaptively calibrated to obtain the minimum scale output feature. At the remaining scales, the first input of each scale is sequentially added to the initial change feature map in ascending order of scale, and feature fusion and adaptive calibration are performed to obtain the output feature of each scale. The first input of each scale includes the result of upsampling the output feature of the previous scale twice. The output feature of the largest scale is sequentially processed by 3×3 convolution, ReLU activation function, and 3×3 convolution to obtain the change detection result map.

[0009] Preferably, adaptive calibration includes: The input features are subjected to adaptive calibration in both the channel dimension and the spatial dimension in parallel to obtain the first result and the second result, respectively. The first result and the second result are then added together to obtain the output of the adaptive calibration.

[0010] Preferably, adaptive calibration at the channel dimension includes: The input features are sequentially processed by global average pooling, fully connected network, sigmoid activation function, and element-wise multiplication with the input features to obtain the first result; The fully connected network processing includes cascaded dimensionality reduction and dimensionality increase parts; the dimensionality reduction part includes 1×1 convolution and ReLU activation function processing in sequence; the dimensionality increase part includes 1×1 convolution.

[0011] Preferably, adaptive calibration of spatial dimensions includes: The input features are sequentially processed by 1×1 convolution, Sigmoid activation function, and element-wise multiplication with the input features to obtain the second result.

[0012] Preferably, multi-time-step noise addition to remote sensing images and contour maps includes: Based on the preset noise scheduling, the remote sensing image and contour map are subjected to noise addition processing at a preset number of time steps: ; ; ; ; Among them, time step = , , … ; Maximum time step; Represents early remote sensing images; Represents late-stage remote sensing images; Represents an early outline drawing; Represents the late-stage outline; Indicates time step Noise-adding results to early remote sensing images; Indicates time step Noise-adding results to late-stage remote sensing images; Indicates time step Noise-adding results to the early contour map; Indicates time step Noise-adding results to late-stage remote sensing images; Indicates the standard Gaussian distribution Random noise tensor sampled in the middle; This represents the fidelity coefficient.

[0013] Preferably, a pre-defined multi-scale feature extraction is performed using a denoising diffusion probability model to obtain multi-timestep feature maps of remote sensing images and contour maps at each scale, including: The multi-timestep noise-adding results of the remote sensing image and the contour image are input into the U-Net module of the denoising diffusion probability model. The U-Net module performs downsampling and upsampling processes based on the input image. During the downsampling and upsampling processes, the output results of each scale of the downsampling process are input into the upsampling process of the same scale, and then the output of the previous scale in the upsampling process is concatenated before performing upsampling at the same scale. The output results of each scale of the upsampling process are saved. After processing the multi-timestep noise-adding results of the remote sensing image and the contour image respectively, the output results of each scale of the upsampling process are saved as multi-timestep feature maps of the remote sensing image and the contour image at each scale.

[0014] The present invention also provides a remote sensing image change detection device based on edge contour guidance, which is used in the method of the present invention, and the device is used to implement the method of the present invention.

[0015] The present invention has the following beneficial effects: The remote sensing image change detection method based on edge contour guidance of the present invention extracts multi-scale semantic features of dual-temporal remote sensing images and their contour maps under different levels of noise based on a pre-trained denoising diffusion probability model. It can capture progressive representations from fine texture details to robust high-level semantics. By performing feature fusion, feature enhancement, and change prediction on the multi-time step and multi-scale feature maps extracted after adding noise to the dual-temporal images and their contour maps, high-precision identification of changed regions can be achieved.

[0016] The remote sensing image change detection device based on edge contour guidance of the present invention, used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0017] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the method flow of a preferred embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the U-Net module according to a preferred embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of multi-timestep feature map enhancement and fusion of remote sensing images and contour maps according to a preferred embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the output feature calculation process for each scale other than the minimum scale in a preferred embodiment of the present invention. Detailed Implementation

[0022] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0023] See Figure 1In a preferred embodiment of the present invention, a remote sensing image change detection method based on edge contour guidance is provided, comprising: Q1. Obtain remote sensing images; combine remote sensing images with the Canny operator to obtain contour maps; remote sensing images include one early remote sensing image and one late remote sensing image with a preset time interval; contour maps include one early contour map and one late contour map.

[0024] In a preferred embodiment of the present invention, Q1 specifically includes: Early remote sensing images I were extracted using the Canny operator. A and late-stage remote sensing image I B Early outline diagram E A and late-stage profile E B To capture key edge information of the image: ; ; in, This represents the Canny model.

[0025] Q2. Add noise to the remote sensing image and contour image at multiple time steps, and combine the noise reduction diffusion probability model to perform preset multi-scale feature extraction, so as to obtain multi-time step feature maps of the remote sensing image and contour image at each scale.

[0026] In a preferred embodiment of the present invention, multi-time-step noise addition to the remote sensing image and the contour image includes: Based on the preset noise scheduling, the remote sensing image and contour map are subjected to noise addition processing at a preset number of time steps: ; ; ; ; Among them, time step = , , … ; Maximum time step; Represents early remote sensing images; Represents late-stage remote sensing images; Represents an early outline drawing; Represents the late-stage outline; Indicates time step Noise-adding results to early remote sensing images; Indicates time step Noise-adding results to late-stage remote sensing images; Indicates time step Noise-adding results to the early contour map; Indicates time step Noise-adding results to late-stage remote sensing images; Indicates the standard Gaussian distribution Random noise tensor sampled in the middle; This represents the fidelity coefficient, which is generally determined based on a pre-set noise control table.

[0027] By adding noise at multiple time steps, it is possible to simulate a variety of noise conditions, from mild to severe, thereby enhancing the model's adaptability to various noise environments and accurately capturing subtle contour changes under strong noise interference.

[0028] In a preferred embodiment of the present invention, a preset multi-scale feature extraction is performed in conjunction with a denoising diffusion probability model to obtain multi-timestep feature maps of remote sensing images and contour maps at each scale, including: The multi-timestep noise-adding results of remote sensing images and contour maps are input into the U-Net module of the pre-trained denoising diffusion probability model. The U-Net module performs downsampling and upsampling processes based on the input images. During the downsampling and upsampling processes, the output results of each scale in the downsampling process are input into the upsampling process of the same scale, and then concatenated with the output of the previous scale in the upsampling process before performing upsampling at the same scale. The output results of each scale in the upsampling process are saved. When processing the multi-timestep noise-adding results of remote sensing images and contour maps, the weights of the U-Net module are shared.

[0029] After processing the multi-timestep noise-adding results of the remote sensing image and the contour image respectively, the output results of each scale of the upsampling process are saved as multi-timestep feature maps of the remote sensing image and the contour image at each scale.

[0030] In a preferred embodiment of the present invention, multi-scale feature extraction is performed at five scales, defined in descending order as scale 0, scale 1, scale 2, scale 3, and scale 4. In the U-Net module, for the k-th time step t... k The image, the downsampling process time step t k The features at scale 4 are first generated through the deep feature extraction block at the bottom of U-Net, capturing the global semantic information of the image; then, during the upsampling process, based on time step t... k Features at scale 4 are used to generate time steps t sequentially. k Scale 3, time step t k Scale 2, time step t k Scale 1 and time step t k The features at scale 0 correspond to feature maps of different resolutions. For , , and (t = t0, t1, t2…t) n The above process is repeated: ; ; ; ; in, This represents the pre-trained denoising diffusion probability model. Ultimately, it generates multi-timestep feature maps of remote sensing images and contour maps at various scales. , , and Time steps t = t0, t1, t2…t n ,scale = 0, 1, 2, 3, 4.

[0031] Through this multi-timestep, multi-scale feature extraction architecture, the model can capture progressive representations from fine texture details in the early diffusion steps to robust high-level semantics in the late diffusion steps, providing rich information for subsequent change detection tasks and enhancing the model's ability to perceive changes in complex scenes.

[0032] In a preferred embodiment of the present invention, the U-Net module serves as the core feature extraction component, employing an encoder-decoder structure to achieve multi-scale feature fusion. Figure 2 The U-Net module consists of a downsampling path and an upsampling path, with feature transfer between them achieved through a skip connection mechanism. The input to the U-Net module... U-Net can be represented as a mapping function: .

[0033] Downsampling path of the U-Net module: In the preferred embodiment of this invention, multi-scale feature extraction is performed at five scales, and the downsampling path consists of five levels, used to progressively extract high-level semantic features and reduce spatial resolution. Except for the fifth level, each level consists of three residual blocks and one downsampling layer connected sequentially.

[0034] The first layer receives the input, processes it through three standard residual blocks, and then halves the spatial resolution using a downsampling layer. The second to fourth layers have similar structures, with each layer containing three standard residual blocks and one downsampling layer. The channel dimension expands with increasing layer depth.

[0035] Input in levels one through four The processing procedure can be represented as: ; in, Indicates hierarchy, Indicates levels one through four; Indicates a downsampling operation; Indicates the first The residual blocks of the layer without self-attention mechanism, m=1,2,3 represent the corresponding 3 residual blocks; Indicates the first Layer output; Indicates the first +1 layer input.

[0036] The fifth level, being the deepest, consists of three residual blocks integrating a self-attention mechanism. It does not include downsampling operations, maintaining the lowest spatial resolution to capture global contextual information. The input to the fifth level is... The processing procedure can be represented as: ; in, This represents a residual block with self-attention mechanism at the fifth level, where m=1,2,3 represents three corresponding residual blocks; This indicates the output of the fifth level.

[0037] The output features of each level in the downsampling path are saved for subsequent feature fusion with the upsampling path.

[0038] Upsampling path of the U-Net module: The upsampling path also consists of 5 levels, used to gradually restore spatial resolution and fuse multi-scale features.

[0039] The fifth layer receives the output of the fifth layer of the downsampling path as input. After processing through four residual blocks with self-attention mechanisms, it doubles the resolution through an upsampling layer. The structures of the fourth to second layers are similar, with each layer containing four standard residual blocks and one upsampling layer. The channel dimension decreases as the layer increases. The first layer consists of three standard residual blocks, with no upsampling operation, and the output restores the feature representation to the original resolution.

[0040] The inputs for the first to fourth levels are all composed of two parts of features: one part comes from the upsampled output of the previous level, and the other part comes from the output stored in the corresponding resolution scale level in the downsampled path.

[0041] In the first level, input The processing procedure is represented as follows: ; in, Indicates an upsampling operation; Indicates the first The residual blocks of the layer without self-attention mechanism; m=1,2,3 represent the corresponding 3 residual blocks; This indicates the output saved at the corresponding resolution scale level in the downsampling path; Indicates the first The output of the layer.

[0042] Input in levels two through four The processing procedure is represented as follows: ; in, Indicates hierarchy, Indicates levels two through four; Indicates an upsampling operation; Indicates the first The residual blocks of the layer without self-attention mechanism; m=1,2,3 represent the three residual blocks corresponding to the second to fourth levels; This indicates the output saved at the corresponding resolution scale level in the downsampling path; Indicates the first The output of the layer.

[0043] Fifth level input The processing procedure is represented as follows: ; in, Indicates an upsampling operation; This indicates the output of the fifth level; This represents the residual block with self-attention mechanism at the fifth level; m=1,2,3,4 represent the four residual blocks corresponding to the fifth level.

[0044] In a preferred embodiment of the present invention, the outputs of each layer of the upsampling path of the U-Net module are saved as the outputs of the U-Net module.

[0045] In a preferred embodiment of the present invention, multi-scale feature extraction, global-local information fusion, and time-series condition perception are used to effectively improve the accuracy and robustness of remote sensing image change detection, which is especially suitable for multi-scale change recognition tasks in complex scenarios.

[0046] Q3. Enhance and fuse the multi-timestep feature maps of the remote sensing image and the contour image at each scale, and then fuse the enhancement and fusion results at each scale in order from small to large to obtain the change detection result map.

[0047] See Figure 3 In a preferred embodiment of the present invention, enhancing and fusing the multi-timestep feature maps of the remote sensing image and the contour image at each scale includes: Simultaneously at various scales, the multi-timestep feature maps of remote sensing images and contour maps are stitched and fused along the channel dimension to obtain early augmented representations at each scale. and late enhancement ,include: ; ; in, This indicates a concatenation operation; dim=1 indicates concatenation along dimension 1. The resulting feature map shows enhanced information on edge contours, which helps improve the model's ability to perceive edge structures and lays the foundation for generating change detection results with high boundary clarity.

[0048] Early enhancement representations are performed using two consecutive sets of convolutional layers and the ReLU activation function. and late enhancement Deep feature extraction and fusion are performed to obtain early feature representations at various scales. and late-stage feature representation : ; ; in, Indicates activation function operation; This indicates that a convolution operation is performed using a convolution kernel of size 1×1; This indicates that a convolution operation is performed using a convolution kernel of size 3×3. and It contains advanced semantic information of phase A (early stage) and phase B (late stage) after nonlinear transformation and spatial information aggregation.

[0049] Computing early feature representations and late-stage feature representation The absolute difference is used to obtain the initial change feature maps of the enhanced fusion results at various scales. : ; in, This represents the absolute difference in dimension 1.

[0050] For the areas in the diagram that have not changed, and The corresponding values ​​should be very similar, with corresponding differences close to zero, and absolute differences... The corresponding part in the middle will also be a value close to zero; while for the region that has changed, its value in the middle will be close to zero. and The corresponding values ​​should have significant differences, resulting in an absolute difference. The corresponding part will be a relatively large value. Therefore, the calculation... and Feature map obtained from absolute difference Effectively highlight and The differences.

[0051] In a preferred embodiment of the present invention, the enhancement fusion results at each scale are sequentially fused from smallest to largest to obtain a change detection result map, including: At the smallest scale, the initial change feature map is adaptively calibrated to obtain the output feature at the smallest scale. At the remaining scales, in ascending order of scale, the first input of each scale is sequentially added to the initial change feature map for feature fusion and adaptive calibration to obtain the output feature at each scale. The first input of each scale includes the result of upsampling the output feature of the previous scale twice. The output feature at the largest scale is sequentially processed by 3×3 convolution, ReLU activation function, and 3×3 convolution to obtain the change detection result map.

[0052] In a preferred embodiment of the present invention, it is known that the first input exists at all scales except the smallest scale, and the first input includes the output features of the previous scale. Perform two upsampling operations to obtain change features with the same resolution as the current scale. : ; in, This indicates an upsampling operation.

[0053] See Figure 4 In a preferred embodiment of the present invention, the first input at each scale is sequentially added to the initial change feature map, and feature fusion and adaptive calibration are performed element-wise to obtain the output features at each scale, including: The first input Compared with the initial change feature map Perform feature fusion to obtain fused features : ; This process involves overlaying and integrating information from upsampling at a smaller scale, which contains more global semantic changes, with local detail changes at the current scale. Through this additive fusion, It retains the detail sensitivity of the current scale while incorporating global contextual information sampled from a smaller scale, achieving information complementarity of cross-scale features and effectively combining high-level semantic information with low-level detailed features.

[0054] Fusion features Adaptive calibration is performed by processing the data in parallel through a channel excitation module and a spatial excitation module. The outputs of the two modules are fused by addition to obtain the output features at each scale. : ; in, Indicates the channel excitation module; This represents the spatial excitation module.

[0055] In a preferred embodiment of the present invention, the parallel processing via the channel excitation module and the spatial excitation module includes: The input features are subjected to adaptive calibration in both the channel dimension and the spatial dimension in parallel to obtain the first result and the second result, respectively. The first result and the second result are then added together to obtain the output of the adaptive calibration.

[0056] In a preferred embodiment of the present invention, the adaptive calibration of the channel dimension includes: The input features are sequentially processed through global average pooling, a fully connected network, a sigmoid activation function, and element-wise multiplication with the input features to obtain the first result. Global average pooling compresses the spatial information of each feature channel, generating a channel description vector; the fully connected network learns the dependencies between channels; the sigmoid activation function generates channel weights; and element-wise multiplication with the input features performs feature recalibration. The fully connected network processing includes cascaded dimensionality reduction and dimensionality increase parts. The dimensionality reduction part sequentially includes a 1×1 convolution and a ReLU activation function, where the 1×1 convolution reduces the number of image channels to half of the original number. The dimension-up component includes a 1×1 convolution, which restores the number of image channels back to the original number of channels.

[0057] Adaptive calibration of the channel dimension can automatically suppress irrelevant or interfering feature channels, while enhancing the response of feature channels related to changing regions, effectively improving the model's ability to express changing information.

[0058] In a preferred embodiment of the present invention, the adaptive calibration of the spatial dimension includes: The input features are sequentially subjected to 1×1 convolution, Sigmoid activation function processing, and element-wise multiplication with the input features to obtain the second result. Specifically, the 1×1 convolution is used to fuse information from all channels to generate a single-channel spatial feature map; the Sigmoid activation function is used to normalize the feature values ​​to spatial weights between 0 and 1; and the element-wise multiplication with the input features is used to enhance key spatial regions.

[0059] Adaptive calibration in the spatial dimension can effectively highlight the precise location of changes while suppressing responses from irrelevant background regions, significantly improving the model's perception accuracy and localization capability of structural change boundaries.

[0060] In a preferred embodiment of the present invention, the dual-path processing mechanism effectively combines the advantages of feature calibration in the channel dimension and the spatial dimension, providing a more accurate feature representation for changing regions, especially boundary details.

[0061] In a preferred embodiment of the present invention, by combining global semantic information and local structural details, the boundary clarity and localization accuracy of change detection results are significantly improved. This effectively avoids the boundary blurring and detail loss problems common in traditional methods, providing solid technical support for high-precision change detection. Furthermore, by introducing an edge contour feature guidance mechanism, the ability to perceive structural changes is significantly enhanced.

[0062] The remote sensing image change detection method based on edge contour guidance of the present invention extracts multi-scale semantic features of dual-temporal remote sensing images and their contour maps under different levels of noise based on a pre-trained denoising diffusion probability model. It can capture progressive representations from fine texture details to robust high-level semantics. By performing feature fusion, feature enhancement, and change prediction on the multi-time step and multi-scale feature maps extracted after adding noise to the dual-temporal images and their contour maps, high-precision identification of changed regions can be achieved.

[0063] In a preferred embodiment of the present invention, a remote sensing image change detection device based on edge contour guidance is also provided for use with the method of the present invention, and the device is used to implement the method of the present invention.

[0064] The remote sensing image change detection device based on edge contour guidance of the present invention, used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote sensing image change detection method based on edge contour guidance, characterized in that, include: Obtain a remote sensing image; obtain a contour map based on the remote sensing image and the Canny operator; the remote sensing image includes one early remote sensing image and one late remote sensing image with a preset time interval; the contour map includes one early contour map and one late contour map. The remote sensing image and the contour image are denoised at multiple time steps, and a preset multi-scale feature extraction is performed in combination with a denoising diffusion probability model to obtain multi-time step feature maps of the remote sensing image and the contour image at each scale. The multi-timestep feature maps of remote sensing images and contour maps are enhanced and fused at each scale, and the enhancement and fusion results at each scale are fused sequentially from small to large to obtain the change detection result map.

2. The remote sensing image change detection method based on edge contour guidance according to claim 1, characterized in that, Enhancement and fusion of multi-timestep feature maps from remote sensing images and contour maps at various scales includes: Simultaneously at each scale, the multi-timestep feature maps of the remote sensing image and the contour image are stitched and fused along the channel dimension to obtain early and late augmented representations at each scale. Deep feature extraction and fusion are then performed on the early and late augmented representations using two consecutive convolutional layers and ReLU activation functions, respectively, to obtain early and late feature representations at each scale. The absolute difference between the early and late feature representations is calculated to obtain the initial change feature map at each scale.

3. The remote sensing image change detection method based on edge contour guidance according to claim 2, characterized in that, The enhancement fusion results at each scale are sequentially fused from smallest to largest to obtain the change detection result map, including: At the minimum scale, the initial change feature map is adaptively calibrated to obtain the minimum scale output feature. At the remaining scales, the first input of each scale is sequentially added to the initial change feature map in ascending order of scale, and feature fusion and adaptive calibration are performed element-wise to obtain the output feature of each scale. The first input of each scale includes the result of upsampling the output feature of the previous scale twice. The output feature of the largest scale is sequentially processed by 3×3 convolution, ReLU activation function, and 3×3 convolution to obtain the change detection result map.

4. The remote sensing image change detection method based on edge contour guidance according to claim 3, characterized in that, Adaptive calibration includes: The input features are subjected to adaptive calibration in both the channel dimension and the spatial dimension in parallel to obtain the first result and the second result, respectively. The first result and the second result are then added together to obtain the output of the adaptive calibration.

5. The remote sensing image change detection method based on edge contour guidance according to claim 4, characterized in that, Adaptive calibration at the channel dimension includes: The input features are sequentially processed by global average pooling, fully connected network, sigmoid activation function, and element-wise multiplication with the input features to obtain the first result; The fully connected network processing includes cascaded dimensionality reduction and dimensionality increase parts; the dimensionality reduction part includes 1×1 convolution and ReLU activation function processing in sequence; the dimensionality increase part includes 1×1 convolution.

6. The remote sensing image change detection method based on edge contour guidance according to claim 5, characterized in that, Adaptive calibration of spatial dimensions includes: The input features are sequentially processed by 1×1 convolution, Sigmoid activation function, and element-wise multiplication with the input features to obtain the second result.

7. The remote sensing image change detection method based on edge contour guidance according to claim 6, characterized in that, Adding noise to the remote sensing image and the contour image at multiple time steps includes: According to the preset noise scheduling, the remote sensing image and the contour image are subjected to noise addition processing for a preset number of time steps: ； ； ； ； Among them, time step = , , … ; Maximum time step; Represents early remote sensing images; Represents late-stage remote sensing images; Represents an early outline drawing; Represents the late-stage outline; Indicates time step Noise-adding results to early remote sensing images; Indicates time step Noise-adding results to late-stage remote sensing images; Indicates time step Noise-adding results to the early contour map; Indicates time step Noise-adding results to late-stage remote sensing images; Indicates the standard Gaussian distribution Random noise tensor sampled in the middle; This represents the fidelity coefficient.

8. The remote sensing image change detection method based on edge contour guidance according to claim 7, characterized in that, Combining a denoising diffusion probability model, a pre-defined multi-scale feature extraction was performed to obtain multi-timestep feature maps of remote sensing images and contour maps at each scale, including: The multi-timestep noise-adding results of the remote sensing image and the contour image are respectively input into the U-Net module of the denoising diffusion probability model; the U-Net module performs downsampling and upsampling processes based on the input image; in the downsampling and upsampling processes, the output results of each scale of the downsampling process are input into the upsampling process of the same scale, and are concatenated with the output of the previous scale in the upsampling process before performing upsampling at the same scale; the output results of each scale of the upsampling process are saved; After processing the multi-timestep noise-adding results of the remote sensing image and the contour image respectively, the output results of each scale of the upsampling process are the multi-timestep feature maps of the remote sensing image and the contour image at each scale.

9. A remote sensing image change detection device based on edge contour guidance, used in the method described in any one of claims 1 to 8, characterized in that, The apparatus is used to implement the method according to any one of claims 1 to 8.

Images