Change detection method fusing global difference attention and local difference feature conversion

Abstract

The invention relates to a change detection method fusing global difference attention and local difference feature conversion, and the method comprises the steps: obtaining a dual-temporal remote sensing image which comprises the remote sensing data of the same geographic region at two different time points; the dual-time-phase remote sensing image is input into the trained change detection network model, the changed spatial position and ground feature type in the dual-time-phase remote sensing image are obtained, and the change detection network model is an encoder-decoder architecture. By means of the method, high-precision, multi-scale and detail-rich change detection can be achieved, and the method has high robustness and a wide application range. According to the method for fusing global difference attention and local difference feature conversion, an efficient and accurate technical means is provided for remote sensing image change detection, and the method has important application value and wide application prospects.

Description

Technical Field

[0001] This invention relates to the field of remote sensing technology, and more specifically, to a change detection method that integrates global difference attention and local difference feature transformation. Background Technology

[0002] Current change detection methods mainly include image difference analysis, change vector analysis, machine learning-based methods such as support vector machines and random forests, as well as deep learning techniques. While these methods can extract changed regions, traditional pixel difference methods are susceptible to noise interference, feature-based machine learning methods have high computational costs when processing large-scale data, and are heavily reliant on data annotation and feature engineering. Deep learning methods offer high accuracy, but face challenges such as large data requirements, high computational resource consumption, and overfitting, especially in complex terrains and environments, where capturing subtle changes remains a challenge.

[0003] Current technologies lack robustness in processing multi-scale and multi-temporal data and suffer from insufficient model interpretability, limiting their effectiveness in practical applications. Future changes in change detection technology should optimize data preprocessing and noise suppression methods, improve the generalization ability of deep learning models, enhance computational efficiency, and increase model interpretability. Furthermore, combining multi-source data (such as high-resolution imagery and LiDAR data) with adaptive algorithms can further optimize detection accuracy, enabling it to play a greater role in environmental monitoring, disaster management, and land use change analysis. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a change detection method that integrates global difference attention and local difference feature transformation, in order to solve at least one of the above-mentioned technical problems.

[0005] In a first aspect, the technical solution of the present invention to solve the above-mentioned technical problems is as follows: a change detection method that integrates global difference attention and local difference feature transformation, the method comprising: Acquire dual-temporal remote sensing images, including remote sensing data of the same geographic area at two different time points; The two-temporal remote sensing images are input into the trained change detection network model to obtain the spatial location and land cover type of the changes in the two-temporal remote sensing images. The change detection network model is an encoder-decoder architecture. The encoder includes a feature extraction layer, and the decoder includes multiple cascaded DTN modules and deconvolution layers. During the encoding stage, multi-scale feature maps are extracted from dual-temporal remote sensing images through a feature extraction layer. These multi-scale feature maps include low-level feature maps and high-level feature maps. During the decoding stage, multiple DTN modules are used to progressively calculate the differences in the feature maps at multiple scales and fuse them to obtain a comprehensive difference feature map. Then, a deconvolution layer is used to upsample and enhance the details of the comprehensive difference feature map output by the last DTN module to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing image.

[0006] The beneficial effects of this invention are as follows: By acquiring dual-temporal remote sensing images and processing them using a trained encoder-decoder architecture change detection network model, the spatial location and land cover type of changes in dual-temporal remote sensing images can be accurately located. In the encoding stage, the feature extraction layer extracts multi-scale feature maps from the dual-temporal remote sensing images, providing rich feature information for subsequent processing. Low-level feature maps retain local image details, while high-level feature maps contain global semantic information, enabling the network to analyze and understand image features at different levels, laying a solid foundation for the change detection task. In the decoding stage, multiple cascaded DTN modules progressively perform difference calculations and feature fusion on the change features in the multi-scale feature maps to obtain a comprehensive difference feature map. This process achieves refined extraction and progressive enhancement of change features, effectively improving the expressive power of change features, enabling the network to more accurately capture change information between dual-temporal remote sensing images. Finally, by deconvolutional layers, the comprehensive differential feature map output by the last DTN module is upsampled and enhanced for detail, further restoring the spatial resolution of the feature map and highlighting the detailed information of the changed areas. This results in high-resolution and detailed change detection results, significantly improving the accuracy and completeness of change area identification. It provides an efficient and accurate technical means for remote sensing image change detection in fields such as environmental monitoring and ecological assessment, and has important application value and broad application prospects.

[0007] Based on the above technical solution, the present invention can be further improved as follows.

[0008] Furthermore, the aforementioned feature extraction layer is a Segformer backbone network, which extracts multi-scale feature maps from dual-temporal remote sensing images, including: Multi-scale feature maps are extracted from bi-temporal remote sensing images using the Transformer encoder in the Segformer backbone network.

[0009] Furthermore, a DFT layer module is connected between every two DTN modules, and the last DTN module is connected to the deconvolution layer. For each DFT layer module, the output feature map of the previous DTN module is used to perform feature enhancement, and the enhanced feature map is used as the input of the next DTN module connected to the DFT layer module.

[0010] Furthermore, each DFT layer module includes a differential feature transformation network, a GDAM unit, an MLP unit, and an LN unit. Each DFT layer module is specifically used for: The differential feature transformation network extracts local variation features from the output feature map of the previous DTN module of the DFT layer module to obtain preliminary features; Global difference features are extracted from the preliminary features using GDAM units; Global differential features are mapped and processed using MLP units to obtain processed features; The enhanced feature map is obtained by standardizing the processed features using LN units.

[0011] Furthermore, each DTN module includes a differential feature calculation unit, a DVEN module, and a feature fusion unit. Each DTN module is specifically used for: The difference feature map is obtained by performing pixel-by-pixel difference calculation on the input feature map through the difference feature calculation unit. The input feature map includes the output feature map from the previous DTN module or the multi-scale feature map extracted by the encoder. The DVEN module is used to extract local variation features from the preliminary difference feature map to obtain a local difference feature map. The preliminary difference feature map and the local difference feature map are fused by the feature fusion unit to obtain the comprehensive difference feature map.

[0012] Furthermore, the aforementioned feature fusion unit fuses the preliminary difference feature map and the local difference feature map to obtain a comprehensive difference feature map, including: The feature fusion unit concatenates the initial difference feature map, the local difference feature map, and the input feature map along the channel dimension, and then performs dimensionality reduction and feature recombination through convolutional layers to form a comprehensive difference feature map.

[0013] Furthermore, the DVEN module is used to extract local variation features from the preliminary difference feature map to obtain a local difference feature map, including: The local variation features in the initial difference feature map are extracted by various difference convolution operations in the DVEN module, resulting in multiple local variation features. These multiple local variation features are then weighted and fused to obtain the local difference feature map.

[0014] Secondly, to solve the above-mentioned technical problems, the present invention also provides a change detection device that integrates global difference attention and local difference feature transformation, the device comprising: The acquisition module is used to acquire dual-temporal remote sensing images, including remote sensing data of the same geographic area at two different time points. The change detection module is used to input dual-temporal remote sensing images into a trained change detection network model to obtain the spatial location and land cover type of changes in the dual-temporal remote sensing images. The change detection network model is an encoder-decoder architecture. The encoder includes a feature extraction layer, and the decoder includes multiple cascaded DTN modules and deconvolution layers. During the encoding stage, multi-scale feature maps are extracted from dual-temporal remote sensing images through a feature extraction layer. These multi-scale feature maps include low-level feature maps and high-level feature maps. During the decoding stage, multiple DTN modules are used to progressively calculate the differences in the feature maps at multiple scales and fuse them to obtain a comprehensive difference feature map. Then, a deconvolution layer is used to upsample and enhance the details of the comprehensive difference feature map output by the last DTN module to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing image.

[0015] Thirdly, in order to solve the above-mentioned technical problems, the present invention also provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the change detection method of the present application that integrates global difference attention and local difference feature transformation.

[0016] Fourthly, in order to solve the above-mentioned technical problems, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the change detection method of the present application that integrates global difference attention and local difference feature transformation.

[0017] Additional aspects and advantages of this application will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of this application. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below.

[0019] Figure 1 This is a flowchart illustrating a change detection method that integrates global difference attention and local difference feature transformation, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a remote sensing image at one time point in a dual-temporal remote sensing image, provided as an embodiment of the present invention. Figure 3 This is a schematic diagram of a preprocessed remote sensing image provided in one embodiment of the present invention; Figure 4A schematic diagram of an overall model architecture provided in one embodiment of the present invention; Figure 5 This is a schematic diagram of a DVEN module provided in one embodiment of the present invention; Figure 6 This is a schematic diagram of a DFT layer module provided in one embodiment of the present invention; Figure 7 A schematic diagram of a GDAM module provided in one embodiment of the present invention; Figure 8 This is a schematic diagram of a change mask obtained after processing by a change detection network model, as provided in one embodiment of the present invention. Figure 9 This is a schematic diagram of a change detection device that integrates global difference attention and local difference feature transformation according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the structure of an electronic device provided in one embodiment of the present invention. Detailed Implementation

[0020] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0021] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0022] The solution provided in this invention can be applied to any application scenario requiring change detection. The solution provided in this invention can be executed by any electronic device, such as a user's terminal device, including at least one of the following: smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, smart TV, or smart in-vehicle device.

[0023] This invention provides a possible implementation, such as... Figure 1 The diagram shows a flowchart of a change detection method that integrates global difference attention and local difference feature transformation. This method can be executed by any electronic device, such as a terminal device, or jointly executed by a terminal device and a server. For ease of description, the method provided in this embodiment will be described below using a terminal device as the execution subject as an example. Figure 1 The flowchart shown indicates that the method may include the following steps: S10, acquire dual-temporal remote sensing images, including remote sensing data of the same geographic area at two different time points; S20, input the dual-temporal remote sensing images into the trained change detection network model to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing images. The change detection network model is an encoder-decoder architecture. The encoder includes a feature extraction layer, and the decoder includes multiple cascaded DTN modules and deconvolution layers. During the encoding stage, multi-scale feature maps are extracted from dual-temporal remote sensing images through a feature extraction layer. These multi-scale feature maps include low-level feature maps and high-level feature maps. During the decoding stage, multiple DTN modules are used to progressively calculate the differences in the feature maps at multiple scales and fuse them to obtain a comprehensive difference feature map. Then, a deconvolution layer is used to upsample and enhance the details of the comprehensive difference feature map output by the last DTN module to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing image.

[0024] The method of this invention acquires dual-temporal remote sensing images and processes them using a trained encoder-decoder architecture change detection network model, enabling precise localization of changes in the spatial location and land cover types within the dual-temporal remote sensing images. In the encoding stage, the feature extraction layer extracts multi-scale feature maps from the dual-temporal remote sensing images, providing rich feature information for subsequent processing. Low-level feature maps preserve local image details, while high-level feature maps contain global semantic information, allowing the network to analyze and understand image features at different levels, laying a solid foundation for the change detection task. In the decoding stage, multiple cascaded DTN modules progressively perform difference calculations and feature fusion on the change features in the multi-scale feature maps to obtain a comprehensive difference feature map. This process achieves refined extraction and progressive enhancement of change features, effectively improving the expressive power of change features and enabling the network to more accurately capture change information between the dual-temporal remote sensing images. Finally, by deconvolutional layers, the comprehensive differential feature map output by the last DTN module is upsampled and enhanced for detail, further restoring the spatial resolution of the feature map and highlighting the detailed information of the changed areas. This results in high-resolution and detailed change detection results, significantly improving the accuracy and completeness of change area identification. It provides an efficient and accurate technical means for remote sensing image change detection in fields such as environmental monitoring and ecological assessment, and has important application value and broad application prospects.

[0025] The following specific embodiments further illustrate the solution of the present invention. In these embodiments, a change detection network model integrating global differential attention and local differential feature transformation is provided. This model combines a global differential attention mechanism and a local differential feature transformation network to achieve dynamic evaluation of the global semantic association and importance of changed regions. It boasts efficient linear time and space complexity and can capture global features across a wide range and multiple scales. The local differential feature transformation network, through the synergistic effect of various differential convolutions and traditional convolutions, accurately extracts local pixel-level change details and enhances differential representation. Through the synergistic effect of global and local features, the network of the present invention can grasp the semantic and structural correlation of changes at the global level while precisely capturing pixel-level change details at the local level, forming a complementary "global-local" feature representation. This provides more comprehensive and accurate feature support for change detection tasks, significantly improving the accuracy and completeness of changed region identification.

[0026] Based on this, the change detection method that integrates global difference attention and local difference feature transformation provided in this embodiment may include the following steps: S10, acquire dual-temporal remote sensing images, including remote sensing data of the same geographic area at two different time points; Two-temporal remote sensing imagery refers to remote sensing images of the same geographic area taken at two different points in time. This type of imagery data is typically used to monitor changes in a geographic area at different points in time, such as land use change, vegetation cover change, and urban expansion. Two-temporal remote sensing imagery includes remote sensing data of the same geographic area at two different points in time; for example, remote sensing image T1 at the first time point and remote sensing image T2 at the second time point. The two images cover the same geographic area but were taken at different times. By comparing the images from these two time points, changes in the area over a time span can be analyzed. Remote sensing images are typically taken by satellites, drones, or other remote sensing equipment. These images can be optical images (such as RGB images), multispectral images (containing multiple bands), hyperspectral images, or radar images, etc. See [example image here] for an example. Figure 2 The image shown is a remote sensing image at a specific point in time.

[0027] S20, input the dual-temporal remote sensing images into the trained change detection network model to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing images. The change detection network model is an encoder-decoder architecture. The encoder includes a feature extraction layer, and the decoder includes multiple cascaded DTN modules and deconvolution layers. During the encoding stage, multi-scale feature maps are extracted from dual-temporal remote sensing images through a feature extraction layer. These multi-scale feature maps include low-level feature maps and high-level feature maps. During the decoding stage, multiple DTN modules are used to progressively calculate the differences in the feature maps at multiple scales and fuse them to obtain a comprehensive difference feature map. Then, a deconvolution layer is used to upsample and enhance the details of the comprehensive difference feature map output by the last DTN module to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing image.

[0028] The change in spatial location refers to the specific geographic coordinates or regional location that changes within the remote sensing image. It indicates the specific location where the change occurred and is usually represented by pixel coordinates, geographic coordinates (latitude and longitude), or vectorized polygon boundaries.

[0029] Land feature type refers to the category of land features corresponding to an area that has undergone changes. Land features refer to natural or man-made objects on the Earth's surface, such as buildings, vegetation, water bodies, and roads. Land feature type change detection not only focuses on the location of the change but also on the transformation of land feature types before and after the change.

[0030] Prior to S20, the method also includes: Preprocessing of dual-temporal remote sensing images includes radiometric correction, geometric registration, normalization, and tiling to eliminate spectral bias, align image spatial coordinates, standardize pixel values, and generate data in an appropriate format according to the requirements of the change detection network model.

[0031] As an example, see Figure 3 The preprocessed dual-temporal remote sensing image shown.

[0032] In the encoding stage, extracting multi-scale feature maps from dual-temporal remote sensing images through the feature extraction layer means extracting multi-scale feature maps from the remote sensing image T1 at the first time point and from the remote sensing image T2 at the second time point. Each multi-scale feature map includes a low-level feature map and a high-level feature map.

[0033] Low-level feature maps mainly contain detailed information such as edges and textures, while high-level feature maps contain deeper information such as semantic categories and spatial relationships. These extracted features will serve as a "unified feature base" for subsequent decoding, ensuring feature space alignment between the two temporal images and thus avoiding interpretation errors caused by differences in feature distribution. Optionally, the aforementioned feature extraction layer is a Segformer backbone network, which extracts multi-scale feature maps from dual-temporal remote sensing images, including: Multi-scale feature maps are extracted from dual-temporal remote sensing images using the Transformer encoder in the Segformer backbone network. The deep network of the Transformer encoder in Segformer achieves dimensionality reduction of features and aggregation of information, providing high-quality input features for the decoding end.

[0034] This invention constructs an end-to-end remote sensing change detection network model. In the decoding stage, the decoder consists of multiple DTN (Differential Feature Transformation Network) modules and deconvolution (DeConv) layers. By stacking multiple DTN modules, the change features in the remote sensing images are extracted step by step.

[0035] It should be noted that the output features of each DTN module can be called a comprehensive difference feature map, the difference being that they are determined based on different input features. The output features of the current DTN module serve as the input features of the next DTN module. For example, for the first DTN module, the input features could be multi-scale feature maps extracted from the remote sensing image T1 at the first time point and multi-scale feature maps extracted from the remote sensing image T2 at the second time point. The first DTN module is responsible for calculating the difference features between the two time point images. The input features of the second DTN module are the comprehensive difference feature map output by the first DTN module. The role of the second DTN module is to further extract and enhance the change features from the comprehensive difference feature map output by the first DTN module, gradually improving the ability to represent change features. Through the layer-by-layer processing of multiple DTN modules, the network can more comprehensively and accurately capture the change information in the dual-temporal remote sensing images, obtaining the comprehensive difference feature map output by the last DTN module. Finally, deconvolutional layers are used to further enhance spatial details, fusing low-level detail information with high-level semantic information.

[0036] Optionally, a DFT layer module (also known as a global difference attention module with residual modules) is connected between every two DTN modules. The last DTN module is connected to a deconvolutional layer. For each DFT layer module, feature enhancement is performed on the output feature map of the previous DTN module to obtain the enhanced feature map, which serves as the input to the next DTN module connected to the DFT layer module. The input features of the first DFT layer module are... Each DFT layer module enhances the global perception of changing regions, captures global change information across scales and time phases, and optimizes the training process. Finally, after multiple DTN modules and deconvolution operations, high-resolution and detailed change detection results are generated, providing accurate feature support for change detection tasks.

[0037] Optionally, each DTN module is a dual-temporal fusion network. The DTN modules in the decoder section are responsible for calculating and fusing the differences between the two temporal features. Each DTN module includes a difference feature calculation unit, a DVEN module (local difference feature transformation module), and a feature fusion unit. Each DTN module is specifically used for: The difference feature calculation unit performs pixel-by-pixel difference calculation on the input feature map to obtain a preliminary difference feature map. The input feature map may include only the output feature map from the previous DTN module, or it may include the output feature map from the previous DTN module and the multi-scale feature map extracted by the encoder. Optionally, the difference feature calculation unit can calculate the difference between the two temporal images by pixel-by-pixel subtraction to obtain a preliminary difference feature map.

[0038] The DVEN module extracts local variation features from the preliminary difference feature map to obtain a local difference feature map; that is, the DVEN module measures the differences between the two images in detail. The preliminary difference feature map and the local difference feature map are fused by the feature fusion unit to obtain the comprehensive difference feature map.

[0039] Optionally, the above-mentioned feature fusion unit fuses the preliminary difference feature map and the local difference feature map to obtain a comprehensive difference feature map, including: The feature fusion unit concatenates the initial difference feature map, the local difference feature map, and the input feature map of the DTN module along the channel dimension, and then performs dimensionality reduction and feature recombination through convolutional layers to form a comprehensive difference feature map.

[0040] Specifically, for the first DTN module, the input feature map of the first DTN module includes multi-scale feature maps extracted from the remote sensing image T1 at the first time point. And extract multi-scale feature maps from the remote sensing image T2 at the second time point. .

[0041] Taking the first DTN module as an example, optionally, the first DTN module can be fused through Euclidean distance differences to obtain a comprehensive difference feature map. The calculation formula for the first DTN module is as follows: in, and These represent dual-temporal features, specifically multi-scale feature maps extracted from the remote sensing image T1 at the first time point and from the remote sensing image T2 at the second time point, respectively. It is a dual-temporal characteristic Euclidean distance calculation method. This is a comprehensive difference feature map calculated using Euclidean distance. This represents the regular convolution process.

[0042] The DTN dual-time fusion network designed using this invention can effectively capture local pixel-level change information, especially in terms of detail changes. The feature fusion unit concatenates the original image features and synthesized difference features along the channel dimension, and then performs dimensionality reduction and feature recombination through convolutional layers. This preserves the original image information while highlighting change cues, thus effectively capturing changes at different spatial resolutions. Optionally, the DVEN module is used to extract local variation features from the preliminary difference feature map to obtain a local difference feature map, including: The local variation features in the initial difference feature map are extracted by various difference convolution operations in the DVEN module, resulting in multiple local variation features. These multiple local variation features are then weighted and fused to obtain the local difference feature map.

[0043] The DVEN module is specifically designed for local difference feature transformation. It extracts and fuses local variation features from the image through various difference convolution operations (including difference convolution, horizontal difference convolution, vertical difference convolution, and adaptive difference convolution). These convolution operations capture differences in the image across different dimensions, and by adjusting the weights of the convolution results, the module finely processes the details of changes in different regions. Finally, an adaptive sigmoid function is used to weight and fuse the different convolution results, ensuring that changes in the image at different spatial scales are accurately captured.

[0044] Using the DVEN module designed in this invention, local change features of remote sensing images are extracted and fused through various differential convolution operations (such as differential convolution, horizontal and vertical differential convolution, etc.). By adjusting the weights of the convolution results, the change details of different regions are captured in detail. Furthermore, the image changes are accurately captured at different spatial scales through weighted fusion using an adaptive Sigmoid function.

[0045] Optionally, each DFT layer module includes a differential feature transformation network (also known as a DVEN module), a GDAM unit (also known as a global differential attention core control unit), an MLP unit (also known as a multilayer perceptron unit), and an LN unit (layer normalization unit). Each DFT layer module is specifically used for: The differential feature transformation network extracts local variation features from the output feature map of the previous DTN module of the DFT layer module to obtain preliminary features; Global difference features are extracted from the preliminary features using GDAM units; Global differential features are mapped and processed using MLP units to obtain processed features; The enhanced feature map is obtained by standardizing the processed features using LN units.

[0046] Input dual-temporal image features ( and The system receives input from the bottom layer (the first DTN module) and extracts local change features through the DVEN module, performing detail transformation and processing. Next, the features pass through the GDAM unit to extract global difference information and capture global changes between the two temporal images. Then, the features are mapped and processed through the MLP unit, and finally standardized through layer normalization (LN unit) to output the change detection results.

[0047] By utilizing the DFT layer module designed in this invention, and combining a global difference attention module with a local feature transformation network, subtle changes in two-phase images can be effectively captured. Through multi-level fusion of global and local information, the DFT layer enhances the model's ability to perceive changed regions, thereby improving the accuracy and robustness of changing region recognition and providing accurate and comprehensive feature support for change detection tasks.

[0048] Furthermore, the GDAM unit is the core control unit of the Global Difference Attention (DFT) layer, improving the accuracy and robustness of change detection through several key components. The TSSA module utilizes a spatiotemporal self-attention mechanism to enhance feature correlation, helping to capture cross-temporal and spatiotemporal change information. The DTN module calculates and fuses feature differences between two-phase images to extract local change features. A multi-head attention mechanism further fuses features, capturing the relationships between them. Residual connections alleviate the gradient vanishing problem, and normalization layers accelerate convergence and improve training stability. Finally, a feedforward network extracts high-level features, ensuring stable information flow.

[0049] The calculation formula for the GDAM unit is as follows: ; ; ; ; ; ; in, and These represent dual-temporal features, specifically multi-scale feature maps extracted from the remote sensing image T1 at the first time point and from the remote sensing image T2 at the second time point, respectively. It is the softmax function. , , , , , , , It is a trainable weight matrix. This represents the output feature map (comprehensive difference feature map) of the i-th DTN module. This represents global difference features.

[0050] Optionally, the TSSA module described above improves the performance of the self-attention mechanism by introducing statistical information (mean, variance) of the tokens. It extracts the statistical features of each token and combines them with the self-attention mechanism to calculate a weighted attention distribution, adjust the feature distribution, highlight key areas, and suppress redundant information.

[0051] The TSSA module designed in this invention preserves local structural information and strengthens global relationships in the processed features, thereby improving the ability of the change detection network to perceive important change regions.

[0052] To better illustrate and understand the principle of the method provided by this invention, the following description uses an optional specific embodiment to illustrate the solution of this invention. It should be noted that the specific implementation of each step in this specific embodiment should not be construed as a limitation of the solution of this invention. Other implementations that can be conceived by those skilled in the art based on the principle of the solution provided by this invention should also be considered within the scope of protection of this invention.

[0053] This invention proposes a change detection network that integrates global differential attention and local differential feature transformation. Through the synergistic effect of the global differential attention mechanism and the local differential feature transformation network, it accurately extracts change features at both global and local levels, improving the accuracy and robustness of change detection tasks and supporting environmental monitoring and ecological assessment. The method employs an encoder-decoder architecture: the encoder consists of a Segformer backbone network (feature extraction layer), responsible for extracting unified and multi-scale visual features from dual-temporal remote sensing images. The input to this encoder is dual-temporal (T1, T2) remote sensing images, and features are extracted from T1 and T2 images respectively to generate multi-scale feature maps. Low-level feature maps mainly contain detailed information such as edges and textures, while high-level feature maps contain deeper information such as semantic categories and spatial relationships. These extracted features serve as the "unified feature base" for subsequent decoding. The decoder consists of multiple DTN (Differential Feature Transformation Network) modules and deconvolution (DeConv) layers. By stacking multiple DTN modules, change features in the remote sensing images are extracted step by step. Each DTN module first processes the input features, extracts difference information at different scales, and then performs upsampling through deconvolution layers to restore the spatial resolution of the image and enhance the details of the changed areas.

[0054] Specifically: such as Figure 4 It is the overall model architecture diagram, which shows in detail the encoder-decoder architecture of the model.

[0055] Figure 5 , Figure 6 and Figure 7 These are schematic diagrams of each individual module in the overall model architecture diagram. Figure 5 It is the Local Differential Feature Transformation (DVEN) module. Figure 6 It is a global difference attention module (DFT layer). Figure 7 It is the core control unit of the global difference attention module (DFT layer). Figure 8 It is the change mask obtained after processing by the change detection network model.

[0056] The solution of the present invention has the following beneficial effects: 1. High-precision change detection: This invention, by fusing global differential attention and local differential feature transformation, can accurately extract change features at both the global and local levels. The global differential attention mechanism can capture global change information across scales and time phases, while the local differential feature transformation can precisely capture pixel-level change details. This complementary "global-local" feature representation significantly improves the accuracy and completeness of change region identification.

[0057] 2. Multi-scale feature fusion: During the encoding phase, feature extraction layers (such as the Transformer encoder in the Segformer backbone network) extract multi-scale feature maps from dual-temporal remote sensing images, including low-level and high-level feature maps. Low-level feature maps preserve local details of the image, while high-level feature maps contain global semantic information. Through the fusion of multi-scale features, the network can more comprehensively understand the changes in the image, thus providing a richer information foundation for change detection tasks.

[0058] 3. Gradual feature enhancement: The decoding stage employs multiple cascaded DTN modules and a global difference attention module (DFT layer) to progressively extract and enhance change features. Each DTN module generates a comprehensive difference feature map through difference calculation, local feature extraction, and feature fusion; each DFT layer further enhances the global perception capability of the feature map through a global difference attention mechanism. This layer-by-layer processing approach enables the network to progressively enhance its ability to represent change features, thereby more accurately capturing subtle changes between two-temporal remote sensing images.

[0059] 4. Enhanced detail and high-resolution output: By upsampling and detail enhancement of the comprehensive differential feature map output by the last DTN module through deconvolution layers, the spatial resolution of the feature map is restored, and detailed information in the changed areas is highlighted. This process generates high-resolution and detailed change detection results, providing high-quality output for change detection in remote sensing images in fields such as environmental monitoring and ecological assessment.

[0060] 5. High robustness: The network structure of this invention effectively alleviates the vanishing gradient problem, accelerates network convergence, and improves training stability by introducing multiple residual modules and layer normalization (LN) operations. This design makes the network more robust when processing complex image data and can adapt to different remote sensing image data and change detection tasks.

[0061] 6. Wide range of applications: This invention is applicable to various remote sensing image data, including optical images, multispectral images, hyperspectral images, and radar images. By adjusting network parameters and preprocessing steps, it can be flexibly applied to different geographical areas and monitoring tasks, such as land use change monitoring, vegetation cover change monitoring, disaster assessment, and urban planning.

[0062] 7. End-to-end solutions: This invention provides an end-to-end change detection network model that automates the entire process from inputting dual-temporal remote sensing imagery to outputting a change mask, requiring no manual intervention. This end-to-end solution significantly improves the efficiency and practicality of change detection while reducing labor costs.

[0063] The solution presented in this invention enables high-precision, multi-scale, and detail-rich change detection, exhibiting strong robustness and wide applicability. This method, which integrates global difference attention and local difference feature transformation, provides an efficient and accurate technical means for change detection in remote sensing images, possessing significant application value and broad application prospects.

[0064] Based on and Figure 1 Based on the same principle as the method shown, this embodiment of the invention also provides a change detection device 20 that integrates global difference attention and local difference feature transformation, such as... Figure 9 As shown, the change detection device 20, which integrates global difference attention and local difference feature transformation, may include an acquisition module 210 and a change detection module 220, wherein: The acquisition module 210 is used to acquire dual-temporal remote sensing images, including remote sensing data of the same geographic area at two different time points; The change detection module 220 is used to input dual-temporal remote sensing images into a trained change detection network model to obtain the spatial location and land cover type of changes in the dual-temporal remote sensing images. The change detection network model is an encoder-decoder architecture. The encoder includes a feature extraction layer, and the decoder includes multiple cascaded DTN modules and deconvolution layers. During the encoding stage, multi-scale feature maps are extracted from dual-temporal remote sensing images through a feature extraction layer. These multi-scale feature maps include low-level feature maps and high-level feature maps. During the decoding stage, multiple DTN modules are used to progressively calculate the differences in the feature maps at multiple scales and fuse them to obtain a comprehensive difference feature map. Then, a deconvolution layer is used to upsample and enhance the details of the comprehensive difference feature map output by the last DTN module to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing image.

[0065] Optionally, the above feature extraction layer is a Segformer backbone network, and the feature extraction layer is specifically used for: Multi-scale feature maps are extracted from bi-temporal remote sensing images using the Transformer encoder in the Segformer backbone network.

[0066] Optionally, a DFT layer module is connected between every two DTN modules, and the last DTN module is connected to the deconvolution layer. For each DFT layer module, the output feature map of the previous DTN module is used to perform feature enhancement to obtain the enhanced feature map, which is used as the input of the next DTN module connected to the DFT layer module.

[0067] Optionally, each DFT layer module includes a differential feature transformation network, a GDAM unit, an MLP unit, and an LN unit, and each DFT layer module is specifically used for: The differential feature transformation network extracts local variation features from the output feature map of the previous DTN module of the DFT layer module to obtain preliminary features; Global difference features are extracted from the preliminary features using GDAM units; Global differential features are mapped and processed using MLP units to obtain processed features; The enhanced feature map is obtained by standardizing the processed features using LN units.

[0068] Optionally, each DTN module includes a differential feature calculation unit, a DVEN module, and a feature fusion unit. Each DTN module is specifically used for: The difference feature map is obtained by performing pixel-by-pixel difference calculation on the input feature map through the difference feature calculation unit; The DVEN module is used to extract local variation features from the preliminary difference feature map to obtain a local difference feature map. The preliminary difference feature map and the local difference feature map are fused by the feature fusion unit to obtain the comprehensive difference feature map.

[0069] Optionally, the feature fusion unit is specifically used for: The feature fusion unit concatenates the initial difference feature map, the local difference feature map, and the input feature map along the channel dimension, and then performs dimensionality reduction and feature recombination through convolutional layers to form a comprehensive difference feature map.

[0070] Optionally, the DVEN module is specifically used for: The local variation features in the initial difference feature map are extracted by various difference convolution operations in the DVEN module, resulting in multiple local variation features. These multiple local variation features are then weighted and fused to obtain the local difference feature map.

[0071] The change detection device that integrates global difference attention and local difference feature transformation in the embodiments of the present invention can execute the change detection method that integrates global difference attention and local difference feature transformation provided in the embodiments of the present invention. The implementation principle is similar. The actions performed by each module and unit in the change detection device that integrates global difference attention and local difference feature transformation in the embodiments of the present invention correspond to the steps in the change detection method that integrates global difference attention and local difference feature transformation in the embodiments of the present invention. For detailed functional descriptions of each module of the change detection device that integrates global difference attention and local difference feature transformation, please refer to the descriptions in the corresponding change detection methods that integrate global difference attention and local difference feature transformation shown above, which will not be repeated here.

[0072] The aforementioned change detection device that integrates global difference attention and local difference feature transformation can be a computer program (including program code) running on a computer device. For example, the change detection device that integrates global difference attention and local difference feature transformation is an application software. The device can be used to execute the corresponding steps in the method provided in the embodiments of the present invention.

[0073] In some embodiments, the change detection device that integrates global difference attention and local difference feature transformation provided in this invention can be implemented using a combination of hardware and software. As an example, the change detection device that integrates global difference attention and local difference feature transformation provided in this invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the change detection method that integrates global difference attention and local difference feature transformation provided in this invention. For example, the processor in the form of a hardware decoding processor can be one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0074] In other embodiments, the change detection device that integrates global difference attention and local difference feature transformation provided in this invention can be implemented in software. Figure 9 A change detection device that fuses global difference attention and local difference feature transformation, stored in a memory, is shown. It can be software in the form of programs and plug-ins, and includes a series of modules, including an acquisition module 210 and a change detection module 220, for implementing the change detection method that fuses global difference attention and local difference feature transformation provided in the embodiments of the present invention.

[0075] The modules described in the embodiments of the present invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0076] Based on the same principles as the methods shown in the embodiments of the present invention, the embodiments of the present invention also provide an electronic device, which may include, but is not limited to: a processor and a memory; the memory for storing computer programs; and the processor for executing the methods shown in any embodiment of the present invention by invoking the computer programs.

[0077] In one alternative embodiment, an electronic device is provided, such as Figure 10 As shown, Figure 10The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present invention.

[0078] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0079] Bus 4002 may include a pathway for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 10 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0080] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0081] The memory 4003 stores application code (computer program) for executing the present invention, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0082] Among these, electronic devices can also be terminal devices. Figure 10 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0083] This invention provides a computer-readable storage medium storing a computer program that, when run on a computer, enables the computer to execute the corresponding content in the aforementioned method embodiments.

[0084] According to another aspect of the present invention, a computer program product or computer program is also provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various embodiments described above.

[0085] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0086] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0087] The computer-readable storage medium provided in this invention can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0088] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0089] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A change detection method integrating global differential attention and local differential feature transformation, characterized in that, include: Acquire dual-temporal remote sensing images, including remote sensing data of the same geographic area at two different time points; The dual-temporal remote sensing images are input into a trained change detection network model to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing images. The change detection network model is an encoder-decoder architecture. The encoder includes a feature extraction layer, and the decoder includes multiple cascaded DTN modules and a deconvolution layer. During the encoding stage, multi-scale feature maps are extracted from the dual-temporal remote sensing images through the feature extraction layer. The multi-scale feature maps include low-level feature maps and high-level feature maps. During the decoding stage, multiple DTN modules progressively perform difference calculations and feature fusion on the changing features in the multi-scale feature map to obtain a comprehensive difference feature map. The deconvolution layer then upsamples and enhances the details of the comprehensive difference feature map output by the last DTN module among the multiple DTN modules to obtain the changing spatial location and land cover type in the dual-temporal remote sensing image.

2. The method according to claim 1, characterized in that, The feature extraction layer is a Segformer backbone network. The extraction of multi-scale feature maps from the dual-temporal remote sensing images through this feature extraction layer includes: Multi-scale feature maps are extracted from the dual-temporal remote sensing images using the Transformer encoder in the Segformer backbone network.

3. The method according to claim 1, characterized in that, A DFTlayer module is connected between every two DTN modules, and the last DTN module is connected to the deconvolution layer. For each DFT layer module, feature enhancement is performed on the output feature map of the previous DTN module to obtain an enhanced feature map, which is then used as the input of the next DTN module connected to the DFT layer module.

4. The method according to claim 3, characterized in that, Each DFT layer module includes a differential feature transformation network, a GDAM unit, an MLP unit, and an LN unit. Each DFT layer module is specifically used for: The differential feature transformation network extracts local change features from the output feature map of the previous DTN module of the DFT layer module to obtain preliminary features. Global difference features are extracted from the preliminary features using the GDAM unit; The global difference features are mapped and processed by the MLP unit to obtain the processed features. The processed features are standardized using the LN unit to obtain an enhanced feature map.

5. The method according to any one of claims 1 to 4, characterized in that, Each DTN module includes a differential feature calculation unit, a DVEN module, and a feature fusion unit. Each DTN module is specifically used for: The difference feature calculation unit performs pixel-by-pixel difference calculation on the input feature map to obtain a preliminary difference feature map. The input feature map includes the output feature map from the previous DTN module or the multi-scale feature map extracted by the encoder. The DVEN module extracts local variation features from the preliminary difference feature map to obtain a local difference feature map. The feature fusion unit fuses the preliminary difference feature map and the local difference feature map to obtain a comprehensive difference feature map.

6. The method according to claim 5, characterized in that, The step of fusing the preliminary difference feature map and the local difference feature map through the feature fusion unit to obtain a comprehensive difference feature map includes: The feature fusion unit concatenates the preliminary difference feature map, the local difference feature map, and the input feature map along the channel dimension, and then performs dimensionality reduction and feature recombination through a convolutional layer to form a comprehensive difference feature map.

7. The method according to claim 5, characterized in that, The step of extracting local variation features from the preliminary difference feature map using the DVEN module to obtain a local difference feature map includes: The local variation features in the preliminary difference feature map are extracted by various difference convolution operations in the DVEN module to obtain multiple local variation features. The multiple local variation features are then weighted and fused to obtain a local difference feature map.

8. A change detection device integrating global difference attention and local difference feature transformation, characterized in that, include: The acquisition module is used to acquire dual-temporal remote sensing images, including remote sensing data of the same geographic area at two different time points. The change detection module is used to input the dual-temporal remote sensing images into a trained change detection network model to obtain the spatial location and land cover type of the changes in the dual-temporal remote sensing images. The change detection network model is an encoder-decoder architecture. The encoder includes a feature extraction layer, and the decoder includes multiple cascaded DTN modules and a deconvolution layer. During the encoding stage, multi-scale feature maps are extracted from the dual-temporal remote sensing images through the feature extraction layer. The multi-scale feature maps include low-level feature maps and high-level feature maps. During the decoding stage, multiple DTN modules progressively perform difference calculations and feature fusion on the changing features in the multi-scale feature map to obtain a comprehensive difference feature map. The deconvolution layer then upsamples and enhances the details of the comprehensive difference feature map output by the last DTN module among the multiple DTN modules to obtain the changing spatial location and land cover type in the dual-temporal remote sensing image.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-7.

Images