Global-local perception integrated remote sensing image semantic change detection system and method
Through the multi-branch coding structure and residual feature aggregation module of global-local perception integration, the problems of low efficiency and accuracy in remote sensing image change detection are solved, and efficient and precise recognition and detection of changing objects in complex scenes are achieved.
Patent Information
- Application Number
- CN202510682265.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-26
- Publication Date
- 2025-09-26
AI Technical Summary
Existing remote sensing image change detection methods are inefficient and inaccurate in identifying complex and ever-changing urban surface cover changes, making it difficult to meet the needs of fine identification. Traditional methods have problems of error accumulation and insufficient feature representation.
A multi-branch encoding structure with global-local perception integration is adopted, combined with residual and Transformer learning paths, a multi-scale residual feature aggregation module is designed to enhance information mining and feature fusion, and the semantic change information response module is used to improve detection accuracy.
It achieves efficient understanding of different changing objects in complex scenes, improves the level of refinement of object classification results in changing areas and the accuracy of detection results, and enhances feature cross-task transfer and interaction.
Smart Images
Figure CN120708044A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of remote sensing image processing, and in particular relates to a remote sensing image semantic change detection system and method based on global-local perception integration. Background Art
[0002] Amidst rapid urbanization, the Earth's surface is constantly changing. Change detection, a hot research topic in remote sensing, compares multi-temporal remote sensing images to detect changes in surface cover or features, providing crucial data support for applications such as urban planning and development, sustainable resource utilization, and environmental protection.
[0003] Traditional change detection research primarily focuses on binary change detection tasks, meaning that detection results are used only to determine the location of a change phenomenon and contain only two types of information: changed or unchanged. However, binary change detection results cannot effectively reflect the transformation of features in the changed region between two-temporal images, which, to a certain extent, limits our comprehensive understanding of surface change phenomena and fails to fully meet the demand for refined change detection results in current applications. In contrast, semantic change detection research, which has rapidly developed in recent years, can provide comprehensive information on the transformation of changed features, while precisely locating the changed region. Semantic change detection improves the interpretability of change detection results, providing a more powerful reference for research such as evaluating change drivers and predicting change development processes, significantly increasing the semantic richness and accuracy of change detection results.
[0004] Early research has primarily implemented semantic change detection through a step-by-step approach. This approach extracts spectral and texture features from remote sensing images, employs machine learning methods to classify land cover in multi-temporal images, and then compares classification results from different temporal images to identify change regions and features. Other studies have also employed change vector analysis to compare spectral differences between temporal images, then used threshold segmentation to identify change regions. Furthermore, these studies combined specialized knowledge or machine learning methods to detect feature transformations within these regions. This step-by-step approach to semantic change detection suffers from cumbersome steps and is prone to error accumulation after classification. To address this, some studies have treated different feature transformations as independent categories and employed direct classification to obtain semantic change detection results. However, due to the large number of categories and the significant imbalance in the proportion of different categories, classifier parameters struggle to converge effectively, resulting in low semantic change detection accuracy. Furthermore, the excessive number of training samples required also limits the further development of these methods.
[0005] Change information from high-resolution remote sensing images provides crucial data support for urban planning, resource management, and disaster warning. However, traditional remote sensing image change detection often relies on algebraic calculations, image transformations, and post-classification comparisons, resulting in low efficiency and accuracy, making it difficult to accurately identify complex and ever-changing urban land cover changes. Existing semantic change detection research employs a relatively simplistic approach to data modeling for dual-temporal remote sensing images, and the limited feature representations make it difficult to fully reflect the characteristics of the diverse shapes and sizes of changing objects in complex scenes. Furthermore, the logical associations between object classification and change detection tasks are insufficiently utilized, and the consistency of the output results from different task branches needs to be improved. Summary of the Invention
[0006] In view of this, the present invention aims to provide a remote sensing image semantic change detection system and method with global-local perception integration, which utilizes a multi-branch coding structure to fully extract global-local features, utilizes residual and Transformer learning paths to realize multi-level information mining of multi-temporal images in different ways, designs a multi-scale residual feature aggregation module to improve the model's ability to understand different changing objects in complex scenes, proposes a semantic change detection decoding structure with multi-task mutual reinforcement and information interaction, realizes efficient utilization of global-local multi-level features, improves the level of refinement of object classification results in changed areas, designs a semantic change information response module to fully utilize the logical associations between different tasks, and combines prior information to improve the accuracy and robustness of semantic change detection.
[0007] To achieve the above object, the technical solution created by the present invention is implemented as follows: A global-local perception integrated remote sensing image semantic change detection system includes a multi-level encoding subnetwork and a multi-level decoding subnetwork; wherein: the multi-level encoding subnetwork includes two local perception encoding branches and a global perception encoding branch; the dual-phase images are respectively input into the corresponding local perception encoding branches to obtain corresponding multi-level ground object encoding features; the channel cascade image of the dual-phase image is input into the global perception encoding branch to obtain multi-level change encoding features; the multi-level decoding subnetwork includes two classification perception decoding branches and a change perception decoding branch; the multi-level change encoding features and the multi-level ground object encoding features are input into the change perception decoding branch for decoding to obtain a binary change detection result; the multi-level change encoding features and the multi-level ground object encoding features are input into the corresponding classification perception decoding branch for decoding, and the output results of the two classification perception decoding branches are masked with the binary change detection results to obtain corresponding classification results of the change areas of the dual-phase image.
[0008] Furthermore, after the coded features of each level are processed by the corresponding MSFF module, the processed features are input into the multi-level decoding sub-network; in each MSFF module: the corresponding input coded features are subjected to feature extraction operations of multiple scales, the extracted features of different scales are spliced and fused, and the spliced and fused features are added to the corresponding elements of the coded features of each level to obtain the output features; in the process of feature extraction operation at each scale, the perception block of the corresponding scale is used to perceive the scale information; the output features of the perception block are subjected to the convolution operation of the convolution layer to obtain the extracted features of the corresponding scale; in the perception block with a scale of N×N, the input features are first subjected to a convolution operation with a convolution kernel size of 1×1; the features output by the convolution operation are first subjected to a convolution operation with a convolution kernel size of 1×N, and the output features are first subjected to a convolution operation with a convolution kernel size of N×1 to obtain the output features containing the corresponding scale information.
[0009] Furthermore, each local perception coding branch includes a shallow feature extraction unit, a first residual coding unit and a second residual coding unit; wherein: in the shallow feature extraction unit, the corresponding phase image is subjected to a convolution operation of the convolution layer to obtain the first-level ground object coding feature; in the first residual coding unit, the first-level ground object coding feature is subjected to a maximum pooling operation and then enters the ResNet module for feature coding to obtain the second-level ground object coding feature; in the second residual coding unit, the second-level ground object coding feature is continuously processed by no less than 2 ResNet modules, and the processed features are then processed by After the convolution operation of the convolution layer, the third-level feature coding features are obtained; the first-level feature coding features, the second-level feature coding features and the third-level feature coding features are processed by their respective MSFF modules and then enter the multi-level decoding sub-network; the ResNet module includes no less than two cascaded ResNet sub-modules. In each ResNet sub-module, the input features are sequentially subjected to the convolution operation of two convolution layers to obtain the first feature; at the same time, the input features are subjected to the convolution operation of the convolution layer to obtain the second feature; the output feature is obtained by adding the corresponding elements of the first feature and the second feature.
[0010] Furthermore, in the global perception coding branch: after the channel cascade image is processed by the three-level SwinT feature coding unit in sequence, each SwinT feature coding unit outputs the corresponding change coding feature; the first-level SwinT feature coding unit includes a block coding layer and a SwinT coding module; after the input feature is processed by the block coding layer, the processed feature is input into the SwinT coding module for attention feature extraction and feature coding, and the first-level change coding feature is obtained; the second-level SwinT feature coding unit and the third-level SwinT feature coding unit include a block merging layer and a SwinT coding module; after the input feature is processed by the block merging layer, the processed feature is input into the SwinT coding module for attention feature extraction and feature coding, and the second-level change coding feature and the third-level change coding feature are obtained accordingly. code features; the SwinT encoding module includes at least two cascaded SwinT encoding submodules; in each SwinT encoding submodule: the input features are sequentially processed by layer normalization and window-based multi-head self-attention operations, and the processed features are added to the corresponding elements of the input features to obtain the first attention features; the first attention features are sequentially processed by layer normalization and multi-layer perceptrons, and the processed features are added to the corresponding elements of the first attention features to obtain the second attention features; the second attention features are sequentially processed by layer normalization and displacement window-based multi-head self-attention operations, and the processed features are added to the corresponding elements of the second attention features to obtain the third attention features; the third attention features are sequentially processed by layer normalization and multi-layer perceptrons, and the processed features are added to the corresponding elements of the third attention features to obtain the change coding features.
[0011] Furthermore, in each classification-aware decoding branch: the last-level change coding feature is input into the decoding unit for first-level decoding processing to obtain the first-level decoding feature; the first-level decoding feature and the second-level change coding feature are concatenated in the channel dimension, and then input into the decoding unit for second-level decoding processing to obtain the second-level decoding feature; the second-level decoding feature and the first-level change coding feature, as well as the third-level feature coding feature processed by the MSFF module, are concatenated in the channel dimension, and then input into the decoding unit for third-level decoding processing to obtain the third-level decoding feature; the third-level decoding feature and the second-level feature coding feature processed by the MSFF module are concatenated in the channel dimension, and then input into the decoding unit for fourth-level decoding processing to obtain the fourth-level decoding feature; the fourth-level decoding feature and the first-level feature coding feature processed by the MSFF module are concatenated in the channel dimension, and then input into the decoding unit for fifth-level decoding processing to obtain the fifth-level decoding feature; the fifth-level decoding feature enters the softmax layer after the convolution operation, and is combined with the binary change detection result for classification mask processing to obtain the classification result of the change area.
[0012] Furthermore, in the change perception decoding branch, the last level change coding feature is input into the decoding unit for first level binary change decoding processing to obtain the first level binary change decoding feature; the first level binary change decoding feature and the second level change coding feature are connected in series in the channel dimension, and then input into the decoding unit for second level binary change decoding processing to obtain the second level binary change decoding feature; the second level binary change decoding feature and the two third level feature coding features processed by the MSFF module are input into the SCR module together for semantic change information extraction, and the extracted semantic feature and the first level binary change coding feature are connected in series in the channel dimension and then input into the SCR module. The third-level binary change decoding features are input into the decoding unit for third-level decoding processing to obtain the third-level binary change decoding features; the third-level binary change decoding features and the two second-level feature coding features processed by the MSFF module are jointly input into the SCR module for semantic change information extraction, and the extracted semantic features are then input into the decoding unit for fourth-level decoding processing to obtain the fourth-level binary change decoding features; the fourth-level binary change decoding features and the two first-level feature coding features processed by the MSFF module are jointly input into the SCR module for semantic change information extraction, and the extracted semantic features enter the sigmoid layer after the convolution operation to obtain the binary change detection results.
[0013] Furthermore, in the decoding unit: after performing a deconvolution operation on the input feature, a preliminary decoding feature is obtained; after performing two convolution operations of different scales on the preliminary decoding feature in sequence, it is added to the corresponding elements of the preliminary decoding feature to obtain the corresponding decoding feature.
[0014] Furthermore, in the SCR module: the two ground object coding features are channel-wise spliced, and then the change-response attention operation is performed on the spliced features; the operated features are respectively element-wise multiplied with the two ground object coding features, and then the corresponding elements of the two multiplied features are subtracted and the absolute values are calculated to obtain the attention features; the attention features are channel-wise spliced with the input binary change decoding features to obtain the semantic features; in the process of performing the change-response attention operation on the spliced features: the spliced features are simultaneously subjected to a convolution operation with a convolution kernel of 1×1, a maximum pooling operation, and an average pooling operation; the features after the three operations are sequentially subjected to channel splicing and convolution operations, and then the convolved features are subjected to a sigmoid operation to obtain the output features.
[0015] A global-local perception integrated remote sensing image semantic change detection method, comprising: S1: Obtain a dual-phase image dataset and preprocess the dual-phase image dataset to obtain a dual-phase image training set; S2: Using the dual-temporal image training set obtained in step S1, the global-local perception integrated remote sensing image semantic change detection system provided by the present invention is trained to obtain a semantic change detection model; S3: Based on the training results, adjust the hyperparameters during the training process and repeat step S2 until the optimal semantic change detection model is obtained; S4: Input the bi-temporal image to be detected into the optimal semantic change detection model obtained in step S3 to obtain a detection result.
[0016] Furthermore, in step S2, the binary cross entropy loss function is used to calculate the loss change of the output of the change-aware decoding branch to obtain a binary loss value; the multi-class cross entropy loss is used to calculate the loss change of the output of the two classification-aware decoding branches to obtain the corresponding ground object classification loss value; the total loss function for training the semantic change detection system is obtained by the following formula: L SCD =L BCE +λ(L MCE1 +L MCE2 ); Among them, L SCD Represents the total loss function, L BCE Represents the binary loss value, L MCE1 and L MCE2 They represent two feature classification loss values respectively, and λ represents the weight.
[0017] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) The present invention creates a remote sensing image semantic change detection system and method with global-local perception integration, realizes a multi-branch encoding structure that fully extracts global-local features, utilizes a residual convolutional neural network and a SwinTransformer learning path to realize multi-level information mining of multi-temporal images in different ways, and designs a multi-scale residual feature aggregation module to improve the model's ability to understand different changing objects in complex scenes; (2) The present invention creates a remote sensing image semantic change detection system and method with global-local perception integration. The MSFF module can enhance the information fusion of dual-temporal global and local features, improve the level of refinement of the classification results of the changed area objects, and enhance the accuracy of the semantic change detection results. (3) The present invention creates a remote sensing image semantic change detection system and method with global-local perception integration. The SCR module can fully utilize the logical relationship between remote sensing image semantic segmentation and binary change detection tasks, realize cross-task transfer and interaction of image features, and combine prior information to improve the clarity and accuracy of change detection results. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] The accompanying drawings, which constitute part of the present invention, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings: Figure 1 A flowchart of a remote sensing image semantic change detection system with integrated global-local perception according to an embodiment of the present invention; Figure 2 A schematic diagram of each MSFF module described in an embodiment of the present invention; Figure 3 A schematic diagram of the ResNet submodule described in an embodiment of the present invention; Figure 4 A schematic diagram of the SwinT sub-coding module according to an embodiment of the present invention; Figure 5 A schematic diagram of a decoding unit according to an embodiment of the present invention; Figure 6 A schematic diagram of an SCR module according to an embodiment of the present invention; Figure 7 A schematic diagram of the attention-responsive operation of the change described in the embodiment of the present invention; Figure 8 The present invention provides a flowchart of a method for detecting semantic changes in remote sensing images using global-local perception integration according to an embodiment of the present invention. DETAILED DESCRIPTION
[0019] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.
[0020] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.
[0021] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second" and the like are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, features defined as "first", "second" and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.
[0022] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "installed," "connected," and "connected" should be understood in a broad sense. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to internal connections between two components. Those skilled in the art can understand the specific meanings of the above terms in the present invention based on specific circumstances.
[0023] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.
[0024] like Figure 1As shown, the global-local perception integrated remote sensing image semantic change detection system described in the embodiment of the present invention includes a multi-level encoding subnetwork and a multi-level decoding subnetwork. Among them, the multi-level encoding subnetwork includes two local perception encoding branches, and a global perception encoding branch. The dual-phase images are respectively input into the corresponding local perception encoding branches, and the corresponding multi-level ground object coding features are obtained respectively. The two local perception encoding branches share weight parameters, which are respectively used to extract the local information and spatial details of the original images of different phases, and optimize the response degree of the features to the image semantic categories layer by layer. The channel cascade image of the dual-phase image is input into the global perception encoding branch to obtain multi-level change coding features. The global perception encoding branch uses the channel cascade results of images of different phases as input data, and gives full play to the remote dependency capture capability to realize the global context information modeling of the dual-phase image. The multi-level decoding subnetwork includes two classification perception decoding branches, and a change perception decoding branch. The multi-level change-encoded features and the multi-level object-encoded features are input into the change-aware decoding branch for decoding, yielding binary change detection results. The multi-level change-encoded features and the multi-level object-encoded features are then decoded in the corresponding classification-aware decoding branch. The binary change detection results are then used to mask the outputs of the two classification-aware decoding branches, yielding classification results corresponding to the dual-temporal images. These three decoding branches effectively integrate the perceptual global and local features of the multi-level image. The two classification-aware decoding branches share decoding weights and utilize independent parameterized classification heads (i.e., their own outputs) to improve results. The change-aware decoding branch further enhances the model's focus on the changing region. Furthermore, masking the outputs of the two classification-aware decoding branches with the binary change detection results effectively captures information about the transformation of objects within the changing region.
[0025] Optical remote sensing imagery is complex and diverse, with widespread occurrences of the same object with different spectra and the same spectrum with different objects. Land objects are diverse in size and shape, and their types are plentiful. Given the limited receptive field of image features extracted by the local perception coding branch, smaller land objects may be missed in semantic change detection results, while the boundaries of complex land objects may be unclear and inaccurate. To overcome this limitation, the present invention incorporates an MSFF (Multi-Scale Residual Feature Aggregation) module between the multi-level encoding subnetwork and the multi-level decoding subnetwork to fully exploit the potential of residual coded features. Specifically, in some embodiments, the coded features of each level of land objects are processed by the corresponding MSFF module and then input into the multi-level decoding subnetwork. The input coded features are subjected to multiple feature extraction operations at different scales, and the extracted features at different scales are then concatenated and fused. The fused features are then added to the corresponding elements of the coded features of each level to generate output features, effectively enhancing the perception of multi-scale land objects.
[0026] During feature extraction at each scale, the corresponding scale's perception block is used to perceive scale information. The output features of the perception block are convolved in a convolutional layer to obtain the corresponding scale's extracted features. In an N×N perception block, the input features are first convolved with a 1×1 kernel. The output features of the convolution operation are then convolved with a 1×N kernel. Finally, the output features are convolved with an N×1 kernel to obtain the output features containing the corresponding scale information.
[0027] Each MSFF module provided by the embodiment of the present invention is as follows Figure 2 As shown in the figure, the corresponding input feature encoding is subjected to feature extraction operations at four different scales, and the extracted features at four different scales are spliced and fused to obtain the output features. Specifically, each MSFF module performs feature extraction operations at scales of 1×1, 3×3, 5×5, and 7×7. Correspondingly, a convolution layer with a convolution kernel size of 1×1 is used for convolution operations, and perception blocks with scales of 3×3, 5×5, and 7×7 are used to perceive scale information. The output features of the three perception blocks are convolved with a convolution layer with a convolution kernel size of 3×3 to obtain the extracted features of the corresponding scale. The 3×3 convolution operation is used to mine the multi-scale information of the image. The extracted features of different scales are spliced and fused. Specifically, the extracted features of four different scales (i.e., the features after 1×1 convolution, and the features obtained after 3×3, 5×5, and 7×7 perception blocks and corresponding 3×3 convolution) are first channel-spliced, and then the spliced features are subjected to 1×1 convolution, batch normalization, and ReLU activation operations. At this point, the splicing and fusion of the extracted features of different scales are completed. The spliced and fused features are added to the corresponding elements of the ground feature coding features at each level to obtain the output features. The processing process of each MSFF module in the embodiment of the present invention can be expressed by the following formula: P1=Conv 1×1 (P); P2=Conv 3×3 (Aware 3×3 (P)); P3=Conv 3×3 (Aware 5×5 (P)); P4=Conv 3×3 (Aware 7×7 (P)); P msf =Conv 1×1 ([P1;P2;P3;P4])+P; Among them, P represents the input feature coding feature, P msfRepresents the output features of the MSFF module, P1, P2, P3 and P4 represent the extracted features of four different scales, Conv 1×1 and Conv 3×3 Represents the convolution operation of the convolution layer with convolution kernel of 1×1 and 3×3 respectively, Aware 3×3 、Aware 5×5 and Aware 7×7 denote the scale-aware operations of 3×3, 5×5, and 7×7 perceptual blocks, respectively, and [·;·] denotes channel concatenation.
[0028] Taking a 7×7 perceptual block as an example, the input features first undergo a convolution operation with a convolution kernel size of 1×1; the features output by the convolution operation first undergo a convolution operation with a convolution kernel size of 1×7, and then the output features first undergo a convolution operation with a convolution kernel size of 7×1 to obtain the output features containing the corresponding scale information. The above operation can be expressed as follows: Aware 7×7 (x)=Conv 7×1 (Conv 1×7 (Conv 1×1 (x))); Among them, x represents the input feature of the perception block with a scale of 7×7, Conv 1×7 and Conv 7×1 They represent the convolution operations of the convolutional layers with kernels of 1 × 7 and 7 × 1. After each convolution operation, batch normalization and ReLU activation operations are performed.
[0029] In some embodiments, each local perception coding branch includes a shallow feature extraction unit, a first residual coding unit, and a second residual coding unit. In the shallow feature extraction unit, the corresponding phase image is subjected to a convolution operation of the convolution layer to obtain a first-level feature coding feature. In the first residual coding unit, the first-level feature coding feature is subjected to a maximum pooling operation and then enters the ResNet module for feature coding to obtain a second-level feature coding feature. In the second residual coding unit, the second-level feature coding feature is processed by no less than two ResNet modules in succession, and the processed features are then subjected to a convolution operation of the convolution layer to obtain a third-level feature coding feature. The first-level feature coding feature, the second-level feature coding feature, and the third-level feature coding feature are respectively processed by their respective MSFF modules and then enter the multi-level decoding subnetwork.
[0030] In this embodiment of the present invention, each local perceptual coding branch is a modified ResNet34 network with pre-trained weights, in which the average pooling layer and fully connected layer at the end of the ResNet34 network are removed. At the same time, Layers 3 and 4 of ResNet34 no longer perform feature downsampling processing, and the corresponding residual connection branches are only used to adjust the channel dimension of the features. At the same time, convolution, batch normalization, and nonlinear activation operations are added after Layer 4 to further achieve feature optimization. Specifically, in each local perception coding branch, in the shallow feature extraction unit, the corresponding temporal image is subjected to the convolution operation of the convolution layer, and then batch normalization and ReLU activation operations are performed to obtain the first-level ground object coding features. The scale of the first-level ground object coding features is 1 / 2 of the scale of the input temporal image; in the first residual coding unit, the first-level ground object coding features are subjected to the maximum pooling operation (Max Pool) and then enter the ResNet module for feature encoding to obtain the second-level ground object coding features. The scale of the second-level ground object coding features is 1 / 4 of the scale of the input temporal image; in the second residual coding unit, the second-level ground object coding features are continuously processed by three ResNet modules, and the processed features are subjected to the convolution operation of the convolution layer, and then batch normalization and ReLU activation operations are performed to obtain the third-level ground object coding features. The scale of the third-level ground object coding features is 1 / 8 of the scale of the input temporal image. Each local perception coding branch can extract representative features of the dual-phase image. Such representative features can represent the local details of the image while retaining the original information. Moreover, these representative features can be used to realize the classification of ground objects in images of different phases. At the same time, comparing the consistency of features of different phases can provide rich detail references for the binary change detection task. Therefore, the extracted first-level ground object coding features, second-level ground object coding features and third-level ground object coding features are processed by their respective MSFF modules and then enter the multi-level decoding subnetwork.
[0031] In some embodiments, each ResNet module includes at least two cascaded ResNet sub-modules. In each ResNet sub-module, the input feature is convolved by two convolutional layers in sequence to obtain a first feature; at the same time, the input feature is convolved by the convolutional layer to obtain a second feature; the first feature and the second feature are added together by corresponding elements to obtain an output feature.
[0032] In the embodiment of the present invention, the ResNet module in the first residual coding unit includes 3 ResNet submodules, and the 3 ResNet modules in the second residual coding unit include 4, 6 and 3 ResNet submodules respectively. Each ResNet submodule is as follows Figure 3 As shown, the input feature F inFirst, after the convolution operation with a convolution kernel size of 3×3, batch normalization and ReLU activation operations, the convolution operation with a convolution kernel size of 3×3 and batch normalization operations are performed to obtain the first feature F1. At the same time, the input feature is subjected to the convolution operation of the convolution layer and then batch normalization operations to obtain the second feature F2. After the corresponding elements of the first feature F1 and the second feature F2 are added, the added features are activated by ReLU to obtain the output feature F out The processing of the ResNet module can be expressed as follows: F1=BN(Conv 3×3 (BR(Conv 3×3 (F in )))); F2=BN(Conv 1×1 (F in )); F out =ReLU(F1+F2); Among them, BN stands for batch normalization, and BR stands for batch normalization and ReLU activation operation.
[0033] The present invention performs channel cascading on the dual-temporal images and inputs the channel cascade images into the global perception coding branch to fully explore the long-range dependencies of changing features within a single image and between different images. The present invention uses a tiny version of the Swin Transformer with lower computational complexity and faster inference speed to extract multi-level global features. Before the dual-temporal images are input into the Swin Transformer, the spatial size and number of channels are first adjusted using a patch embedding layer, and the original data is divided into non-overlapping image patches. Except for the first feature encoding stage, a patch merging layer is set before each stage to achieve feature resolution downsampling by merging adjacent patches. The Swin Transformer used uses the first three feature encoding stages to achieve multi-level global feature acquisition. Specifically, in the global perception coding branch, the channel cascade images are processed by three levels of SwinT feature encoding units in sequence, and each SwinT feature encoding unit outputs a corresponding change coding feature. The first-level SwinT feature encoding unit includes a block encoding layer and a SwinT encoding module; after the input features are block-encoded by the block encoding layer, the processed features are input into the SwinT encoding module for attention feature extraction and feature encoding, thereby obtaining the first-level change coding features. The second-level SwinT feature encoding unit and the third-level SwinT feature encoding unit include a block merging layer and the SwinT encoding module; after the input features are processed by the block merging layer, the processed features are input into the SwinT encoding module for attention feature extraction and feature encoding, thereby obtaining the second-level change coding features and the third-level change coding features, respectively. The SwinT encoding module includes at least two cascaded SwinT encoding submodules; in each SwinT encoding submodule, the input features are sequentially processed by layer normalization and window-based multi-head self-attention operation, and the processed features are added to the corresponding elements of the input features to obtain the first attention feature; the first attention feature is sequentially processed by layer normalization and multi-layer perceptron, and the processed features are added to the corresponding elements of the first attention feature to obtain the second attention feature; the second attention feature is sequentially processed by layer normalization and displacement window-based multi-head self-attention operation, and the processed features are added to the corresponding elements of the second attention feature to obtain the third attention feature; the third attention feature is sequentially processed by layer normalization and multi-layer perceptron, and the processed features are added to the corresponding elements of the third attention feature to obtain the change coding feature.
[0034] In an embodiment of the present invention, the input channel cascade image enters the first-level SwinT feature encoding unit, the SwinT encoding module in the first-level SwinT feature encoding unit includes two cascaded SwinT encoding submodules, and obtains the first-level change encoding feature, the size of the first-level change encoding feature is 1 / 8 of the input channel cascade image size; the first-level change encoding feature enters the second-level SwinT feature encoding unit, the SwinT encoding module in the second-level SwinT feature encoding unit includes two cascaded SwinT encoding submodules, and obtains the second-level change encoding feature, the size of the second-level change encoding feature is 1 / 16 of the input channel cascade image size; the second-level change encoding feature enters the third-level SwinT feature encoding unit, the SwinT encoding module in the third-level SwinT feature encoding unit includes six cascaded SwinT encoding submodules, and obtains the third-level change encoding feature, the size of the third-level change encoding feature is 1 / 32 of the input channel cascade image size. Each SwinT encoding submodule is as follows Figure 4 As shown, the input feature x l-1 Sequentially undergoes layer normalization LN and window-based multi-head self-attention operation W MSA After processing, the processed features are the same as the input features x l-1 Add the corresponding elements to get the first attention feature , which can be expressed as: ; First attention feature After being processed by layer normalization LN and multi-layer perceptron MLP, the processed features are consistent with the first attention features. Add the corresponding elements to get the second attention feature x l , which can be expressed as: ; Second attention feature x l Sequentially undergoes layer normalization LN and multi-head self-attention operation SW based on displacement window MSA After processing, the processed features correspond to the second attention features x l Add together to get the third attention feature , which can be expressed as: ; The third attention feature After being processed by layer normalization LN and multi-layer perceptron MLP, the processed features are consistent with the third attention feature Add the corresponding elements to get the change coding feature x l+1 It can be expressed as: .
[0035] Window-based multi-head self-attention operation W in the embodiment of the present invention MSA And the multi-head self-attention operation SW based on the displacement window MSA From the paper "Swin transformer: Hierarchical vision transformer using shifted windows" published at the ICCV 2021 conference.
[0036] In some embodiments, in each classification-aware decoding branch: the last-level change coding feature is input into the decoding unit for first-level decoding processing to obtain first-level decoding features; the first-level decoding feature and the second-level change coding feature are concatenated in the channel dimension, and then input into the decoding unit for second-level decoding processing to obtain second-level decoding features; the second-level decoding feature and the first-level change coding feature, as well as the third-level feature coding feature processed by the MSFF module, are concatenated in the channel dimension, and then input into the decoding unit for third-level decoding processing to obtain third-level decoding features; the third-level decoding feature and the second-level feature coding feature processed by the MSFF module are concatenated in the channel dimension, and then input into the decoding unit for fourth-level decoding processing to obtain fourth-level decoding features; the fourth-level decoding feature and the first-level feature coding feature processed by the MSFF module are concatenated in the channel dimension, and then input into the decoding unit for fifth-level decoding processing to obtain fifth-level decoding features; the fifth-level decoding feature enters the softmax layer after the convolution operation, and is combined with the binary change detection result for mask processing to obtain the classification result of the change area.
[0037] In order to optimize the model structure and improve the generalization ability of the model, the present invention designs a unified decoding unit for the decoding branches of different tasks. The decoding branch can further realize feature optimization, perform feature upsampling processing, and output pixel-level feature classification and binary change detection results of the original image size. In some embodiments, in the decoding unit, after the input features are deconvolved, preliminary decoding features are obtained; the present invention also introduces residual connections to reduce feature information loss, that is, the preliminary decoding features are sequentially subjected to two different scale convolution operations, and then added to the corresponding elements of the preliminary decoding features to obtain corresponding decoding features.
[0038] The decoding unit of the embodiment of the present invention is as follows Figure 5 As shown, for the input feature F in After performing a deconvolution operation with a convolution kernel size of 2×2, the preliminary decoding feature F is obtained. up ; The preliminary decoding feature F up After performing convolution operations with kernel sizes of 3×3 and 1×1, the convolution kernel is combined with the initial decoding feature F up Add the corresponding elements to get the corresponding decoding feature F edMoreover, after each convolution operation, batch normalization and ReLU activation operations are performed on the convolution features. The above process can be expressed as follows: F up =BR(DeConv 2×2 (F in )) F ed =BR(Conv 1×1 (BRConv 3×3 (F up )))+F up ; Among them, DeConv 2×2 Represents a deconvolution operation with a convolution kernel size of 2×2.
[0039] Since the object classification and binary change detection results of dual-temporal remote sensing images are obtained from in-depth mining and analysis of the original image data. There is an inherent logical relationship between the object classification and binary change detection tasks. For example, the area with inconsistent dual-temporal image classification results should be the area that has changed, and the dual-temporal images in the unchanged area should have the same classification results. In order to fully enhance the promotion effect between related tasks, the present invention further designs an SCR (semantic change information response) module in different task information transmission stages. In some embodiments, in the change perception decoding branch, the last level of change coding features are input into the decoding unit for first-level binary change decoding processing to obtain first-level binary change decoding features; the first-level binary change decoding features and the second-level change coding features are concatenated in the channel dimension, and then input into the decoding unit for second-level binary change decoding processing to obtain second-level binary change decoding features; the second-level binary change decoding features and the two third-level object coding features processed by the MSFF module are jointly input into the SCR module for semantic change information extraction, and the extracted semantic features and the first-level change coding features are concatenated in the channel dimension and then input into the SCR module for semantic change information extraction. The third-level binary change decoding features are input into the decoding unit for third-level decoding processing to obtain the third-level binary change decoding features; the third-level binary change decoding features and the two second-level feature coding features processed by the MSFF module are jointly input into the SCR module for semantic change information extraction, and the extracted semantic features are then input into the decoding unit for fourth-level decoding processing to obtain the fourth-level binary change decoding features; the fourth-level binary change decoding features and the two first-level feature coding features processed by the MSFF module are jointly input into the SCR module for semantic change information extraction, and the extracted semantic features enter the sigmoid layer after the convolution operation to obtain the binary change detection results.
[0040] In some embodiments, in the SCR module: two ground feature coding features are channel-spliced, and then the spliced features are subjected to a change-response attention operation; the operated features are respectively element-multiplied with the two ground feature coding features, and then the two multiplied features are subtracted from each other and the absolute value is calculated to obtain the attention feature; the attention feature is channel-spliced with the input binary change decoding feature to obtain the semantic feature.
[0041] The SCR module in the embodiment of the present invention is as follows Figure 6 As shown, the two ground features are coded with F t1 and F t2 Perform channel splicing, and then perform change aware attention (CAA) operation on the spliced features. The features after CAA operation are respectively compared with the two ground feature coding features F t1 and F t2 Multiply the corresponding elements to get the feature F t1_att and F t2_att , perform the difference and take the absolute value processing on the two features to obtain the attention feature to increase the degree of distinction between the changing foreground and the unchanged background. The attention feature is combined with the input binary change decoding feature F cd Perform channel splicing to obtain semantic features F out , and the semantic features are used as input features of the next decoding unit to improve the model's ability to respond to semantically changing areas. The above process can be expressed as follows: F t1_att =F t1 ×CAA([F t1 ;F t2 ]); F t2_att =F t2 ×CAA([F t1 ;F t2 ]); F out =[F cd ;|F t1_att -F t2_att |]; In some embodiments, the process of performing a change-response attention operation on the spliced features includes: performing a convolution operation with a convolution kernel of 1×1, a maximum pooling operation, and an average pooling operation on the spliced features at the same time; performing channel splicing and convolution operations on the features after the three operations in sequence, and then performing a sigmoid operation on the convolved features to obtain output features.
[0042] In the embodiment of the present invention, the process of changing the response attention operation is as follows: Figure 7As shown in the figure, the features to be processed are simultaneously subjected to a 1×1 convolution operation, a maximum pooling operation MP (Max Pool), and an average pooling operation AP (AveragePool) to achieve feature importance encoding and improve the model's focus on the changed area. The features after the three operations are sequentially subjected to channel splicing and a 7×7 convolution operation. The convolved features are then subjected to a sigmoid operation. The pixel values of the resulting weight matrix are all between 0 and 1. Multiplying it with the dual-phase image features can achieve attention optimization in the spatial dimension, thereby improving the model's change detection results and obtaining the output features. The above process can be expressed as follows: CAA(x)=sigmoid(Conv 7×7 ([Conv 1×1 ;MP(x);AP(x)])); The formula for the sigmoid operation is: sigmoid(x)=1 / (1+e -x ).
[0043] The present invention also provides a global-local perception integrated remote sensing image semantic change detection method, such as Figure 8 As shown, including: S1: Obtain a dual-temporal image dataset and preprocess the dual-temporal image dataset to obtain a dual-temporal image training set.
[0044] In the examples of the present invention, the publicly available Landsat-SCD and SECOND datasets were selected for experiments. The Landsat-SCD dataset covers Tumushuke, Xinjiang, China. The Landsat-SCD dataset contains images with a spatial resolution of 30 meters from 1990 to 2020. Semantic variations occur between farmland, desert, buildings, and water. The dataset includes 2,385 pairs of original samples, each sized at 416×416 pixels. 1,435 pairs were used for model training, 475 pairs for validation, and 475 pairs for testing. The SECOND dataset covers urban areas of Hangzhou, Shanghai, and Chengdu, with image spatial resolutions ranging from 0.5 to 3 meters. It examines semantic variations between six ground objects: water, bare surfaces, low vegetation, trees, buildings, and sports fields. The 2,968 pairs of publicly available samples, each sized at 512×512 pixels, were divided into 2,078 pairs for training, 445 pairs for validation, and 445 pairs for testing.
[0045] The preprocessing process in step S1 includes the division into training, validation, and test sets. During training, random flipping and rotation of training samples are introduced to enhance the samples. Specifically, the rand_rot90_MCD function, whose underlying function is np.rot90, is first called. Based on a random probability (50%), the input image and label are rotated 90 degrees counterclockwise. The rotated result is then passed to the rand_flip_MCD function, whose underlying function includes np.flip and a custom matrix transformation function. The image and label are randomly flipped based on a random probability (25% to keep the image as is, 25% to flip horizontally, 25% to flip vertically, and 25% to flip diagonally). Finally, the randomly rotated and flipped image and label are returned. This entire process increases data diversity through random transformations, which helps improve the model's generalization ability.
[0046] S2: Using the dual-temporal image training set obtained in step S1, the global-local perception integrated remote sensing image semantic change detection system provided by the present invention is trained to obtain a semantic change detection model.
[0047] The object classification and binary change detection results of the dual-temporal image are both pixel-level segmentation results. To optimize the network model parameters and effectively improve detection performance, the present invention selects a cross-entropy loss function to supervise the output results of different tasks. In some embodiments, the binary cross-entropy loss function is used to calculate the loss change of the output of the change-aware decoding branch to obtain the change detection loss value. The binary cross-entropy loss function is specifically: ; Among them, L BCE represents the binary loss value, is the ground truth value of the binary change of the corresponding pixel, which is 0 (unchanged) or 1 (changed); is the probability that the pixel is predicted to be a change category; The multi-category cross entropy loss is used to calculate the loss change of the output of the two classification perception decoding branches to obtain the corresponding ground object classification loss value. The multi-category cross entropy loss is specifically: ; Among them, L MCE It represents the loss value of feature classification, and N is the number of feature classification categories; is a sign function. When the true category of the changed pixel is n, is 1, otherwise 0; is the probability that the corresponding pixel is predicted to be the ground object category n; The total loss function L for training the semantic change detection system SCD It is the weighted sum of the binary loss and the object classification loss, that is, the total loss function is obtained by the following formula: L SCD =L BCE +λ(L MCE1 +L MCE2 ); Among them, L SCD Represents the total loss function, L MCE1 and L MCE2 They represent two feature classification loss values respectively, and λ represents the weight.
[0048] In this embodiment of the present invention, the training environment is as follows: the computer GPU is NVIDIA RTX4090, the operating system is Windows 11, the CPU processor is Intel(R) Core(TM) i9-14900K, the Python version is 3.9.18, the PyTorch version is 2.2.0, the CUDA version is 12.1, and the cuDNN version is 8.8.0.1.
[0049] S3: Based on the training results, adjust the hyperparameters during the training process and repeat step S2 until the optimal semantic change detection model is obtained.
[0050] In this embodiment of the present invention, the hyperparameters include the batch size of the sample set to 8, the epoch set to 60; Stochastic Gradient Descent (SGD) is used to optimize the model parameters, the weight decay is 0.0005, and the momentum is 0.9. The initial learning rate is set to 0.1, and the total_iter is 0.1×(1-iter / total_iter) 1.5 In order to improve the convergence performance of the model, iter is the training process of a batch and total_iter is the total number of training batches.
[0051] S4: Input the bi-temporal image to be detected into the optimal semantic change detection model obtained in step S3 to obtain a detection result.
[0052] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved. This is not limited herein.
[0053] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.
Claims
1. A global-local perception integrated remote sensing image semantic change detection system, characterized by: It includes a multi-stage encoding sub-network and a multi-stage decoding sub-network; among which: The multi-level coding subnetwork includes two local perception coding branches and a global perception coding branch; the dual-phase images are respectively input into the corresponding local perception coding branches to obtain corresponding multi-level ground feature coding features; the channel cascade image of the dual-phase image is input into the global perception coding branch to obtain multi-level change coding features; The multi-level decoding subnetwork includes two classification-aware decoding branches and a change-aware decoding branch; the multi-level change coding features and the multi-level ground object coding features are input into the change-aware decoding branch for decoding to obtain a binary change detection result; the multi-level change coding features and the multi-level ground object coding features are input into the corresponding classification-aware decoding branch for decoding, and the output results of the two classification-aware decoding branches are masked with the binary change detection result to obtain the corresponding classification results of the change areas of the dual-phase image.
2. The global-local perception integrated remote sensing image semantic change detection system according to claim 1, characterized in that: After the coded features of each level are processed by the corresponding MSFF module, the processed features are input into the multi-level decoding sub-network; in each MSFF module: Perform feature extraction operations at multiple scales on the corresponding input feature code, perform splicing and fusion operations on the extracted features at different scales, add the spliced and fused features to the corresponding elements of each level of feature code features to obtain the output features; During the feature extraction process at each scale, the perception block of the corresponding scale is used to perceive the scale information; After the output features of the perception block are subjected to a convolution operation by a convolution layer, extracted features of corresponding scales are obtained; In the perception block with a scale of N×N, the input features first undergo a convolution operation with a convolution kernel size of 1×1; the features output by the convolution operation first undergo a convolution operation with a convolution kernel size of 1×N, and then the output features first undergo a convolution operation with a convolution kernel size of N×1 to obtain the output features containing the corresponding scale information.
3. The global-local perception integrated remote sensing image semantic change detection system according to claim 2, characterized in that: Each local perceptual coding branch includes a shallow feature extraction unit, a first residual coding unit, and a second residual coding unit; wherein: In the shallow feature extraction unit, the corresponding time phase image is subjected to a convolution operation of the convolution layer to obtain the first-level ground feature coding feature; In the first residual coding unit, the first-level ground object coding features are subjected to a maximum pooling operation and then enter the ResNet module for feature coding to obtain second-level ground object coding features; In the second residual coding unit, the second-level feature coding features are processed by at least two ResNet modules in succession, and the processed features are further subjected to a convolution operation of a convolutional layer to obtain third-level feature coding features; The first-level feature coding features, the second-level feature coding features, and the third-level feature coding features are processed by their respective MSFF modules and then enter the multi-level decoding sub-network; The ResNet module includes at least two cascaded ResNet sub-modules. In each ResNet sub-module, the input feature is sequentially convolved through two convolutional layers to obtain a first feature. At the same time, the input feature is convolved through the convolutional layers to obtain a second feature. The first feature and the second feature are added together to obtain an output feature.
4. The global-local perception integrated remote sensing image semantic change detection system according to claim 2, characterized in that: In the global perceptual coding branch: After the channel cascade image is processed by three levels of SwinT feature encoding units in sequence, each SwinT feature encoding unit outputs a corresponding change coding feature; The first-level SwinT feature coding unit includes a block coding layer and a SwinT coding module; After the input features are processed by the block coding layer, the processed features are input into the SwinT coding module for attention feature extraction and feature coding to obtain the first-level change coding features; The second-level SwinT feature encoding unit and the third-level SwinT feature encoding unit include a block merging layer and the SwinT encoding module; The input features are processed by the block merging layer, and the processed features are input into the SwinT encoding module for attention feature extraction and feature encoding, thereby obtaining the second-level change encoding features and the third-level change encoding features; The SwinT encoding module includes at least two cascaded SwinT encoding submodules; in each SwinT encoding submodule: The input features are processed by layer normalization and window-based multi-head self-attention operation in sequence, and the processed features are added to the corresponding elements of the input features to obtain the first attention feature; The first attention feature is sequentially processed by layer normalization and a multi-layer perceptron, and the processed feature is added to the corresponding element of the first attention feature to obtain a second attention feature; The second attention feature is processed by layer normalization and a multi-head self-attention operation based on a shifted window in sequence, and the processed feature is added to the corresponding element of the second attention feature to obtain a third attention feature; The third attention feature is processed by layer normalization and multi-layer perceptron in sequence, and the processed feature is added to the corresponding element of the third attention feature to obtain the change coding feature.
5. The global-local perception integrated remote sensing image semantic change detection system according to claim 2, characterized in that: In each classification-aware decoding branch: The last level of the changed coding features are input into the decoding unit for first level decoding processing to obtain the first level decoding features; The first-level decoding features and the second-level change encoding features are concatenated in the channel dimension and then input into the decoding unit for second-level decoding processing to obtain the second-level decoding features; The second-level decoding features, the first-level change coding features, and the third-level feature coding features processed by the MSFF module are concatenated in the channel dimension and then input into the decoding unit for third-level decoding processing to obtain the third-level decoding features; The third-level decoding features and the second-level feature coding features processed by the MSFF module are concatenated in the channel dimension and then input into the decoding unit for fourth-level decoding processing to obtain the fourth-level decoding features; The fourth-level decoding features and the first-level feature coding features processed by the MSFF module are concatenated in the channel dimension and then input into the decoding unit for fifth-level decoding processing to obtain the fifth-level decoding features; The fifth-level decoding features enter the softmax layer after the convolution operation, and are combined with the binary change detection results to perform classification mask processing to obtain the classification results of the changed areas.
6. The global-local perception integrated remote sensing image semantic change detection system according to claim 5, characterized in that: In the change-aware decoding branch, The last level of change coding features are input into the decoding unit to perform first level binary change decoding processing to obtain first level binary change decoding features; The first-level binary change decoding features and the second-level change encoding features are concatenated in the channel dimension and then input into the decoding unit for second-level binary change decoding processing to obtain the second-level binary change decoding features; The second-level binary change decoding features and the two third-level feature coding features processed by the MSFF module are input into the SCR module for semantic change information extraction. The extracted semantic features are concatenated with the first-level binary change coding features in the channel dimension and then input into the decoding unit for third-level decoding processing to obtain the third-level binary change decoding features. The third-level binary change decoding features and the two second-level feature coding features processed by the MSFF module are input into the SCR module to extract semantic change information. The extracted semantic features are then input into the decoding unit for the fourth-level decoding process to obtain the fourth-level binary change decoding features. The fourth-level binary change decoding features and the two first-level feature coding features processed by the MSFF module are input into the SCR module to extract semantic change information. The extracted semantic features enter the sigmoid layer after the convolution operation to obtain the binary change detection results.
7. The global-local perception integrated remote sensing image semantic change detection system according to claim 6, characterized in that: In the decoding unit: after performing a deconvolution operation on the input feature, a preliminary decoding feature is obtained; after performing two convolution operations of different scales on the preliminary decoding feature in sequence, the preliminary decoding feature is added to the corresponding element of the preliminary decoding feature to obtain the corresponding decoding feature.
8. The global-local perception integrated remote sensing image semantic change detection system according to claim 6, characterized in that: In the SCR module, two feature coding features are channel-joined, and then a change response attention operation is performed on the joined features; the operated features are element-wise multiplied with the two feature coding features, and then the two multiplied features are element-wise subtracted and the absolute value is calculated to obtain an attention feature; the attention feature is channel-joined with the input binary change decoding feature to obtain the semantic feature; In the process of changing the response attention operation on the spliced features: the spliced features are simultaneously subjected to convolution operations with a convolution kernel of 1×1, maximum pooling operations, and average pooling operations; the features after the three operations are sequentially subjected to channel splicing and convolution operations, and then the convolved features are subjected to a sigmoid operation to obtain the output features.
9. A global-local perception integrated remote sensing image semantic change detection method, characterized in that: include: S1: Acquire a dual-phase image dataset, and preprocess the dual-phase image dataset to obtain a dual-phase image training set; S2: Using the dual-temporal image training set obtained in step S1, the global-local perception integrated remote sensing image semantic change detection system according to any one of claims 1 to 7 is trained to obtain a semantic change detection model; S3: Based on the training results, adjust the hyperparameters during the training process and repeat step S2 until the optimal semantic change detection model is obtained; S4: Input the bi-temporal image to be detected into the optimal semantic change detection model obtained in step S3 to obtain a detection result.
10. The method for detecting semantic changes in remote sensing images based on global-local perception integration according to claim 9, characterized in that: In step S2, Calculating a loss change of an output of the change-aware decoding branch using a binary cross entropy loss function to obtain a change detection loss value; The multi-category cross entropy loss is used to calculate the loss change of the output of the two classification perception decoding branches to obtain the corresponding ground object classification loss value; The total loss function for training the semantic change detection system is obtained by the following formula: L SCD =L BCE +λ(L MCE1 +L MCE2 ); Among them, L SCD Represents the total loss function, L BCE Represents the binary loss value, L MCE1 and L MCE2 They represent two feature classification loss values respectively, and λ represents the weight.