Remote sensing image change detection method and system based on relation perception high-order interaction

Abstract

The invention provides a remote sensing image change detection method and system based on relation perception high-order interaction, and the method comprises the steps: firstly collecting remote sensing images of the same region in different periods, constructing a dual-temporal data set, employing an independent convolution encoder to extract single-temporal features, and keeping the original temporal-spatial information; the method comprises the following steps: constructing a spatial-temporal feature pyramid through a twin network structure, introducing a time relation perception module to perform modeling on space and channel dimensions, and capturing mutual relations between pixels and between channels by using a grouping attention mechanism and an affinity matrix calculation mode to realize deep fusion between features; the high-order space-time interaction module stacks a plurality of interaction units through a recursive structure, and enhances context information modeling in a spatial dimension, so that complex high-order feature interaction expression is realized, weighted cross entropy loss and weighted IoU loss are combined, the detection capability of the model on an edge region and a difficult-to-divide region is improved, and the detection efficiency is improved. And inputting a dual-time remote sensing image into the trained model, and outputting a high-precision change detection result.

Description

TECHNICAL FIELD

[0001] The application belongs to remote sensing image change detection, and particularly relates to a remote sensing image change detection method and system based on relationship perception high-order interaction. BACKGROUND

[0002] In the past few decades, convolutional neural networks have been widely used in change detection tasks due to their excellent local feature extraction capabilities. Most existing methods use a U-shaped encoder-decoder network structure for feature extraction, where the encoder is composed of early fusion or late fusion. Early fusion is to input the spliced double-time images, and late fusion is to extract features from the two images respectively. According to whether the weights are shared, late fusion can be divided into twin structure and pseudo-twin structure. Gradually, the encoder based on twin network gradually occupies the mainstream. This may be because the twin network encoder can capture difference information while retaining more respective features, which is beneficial to improve the change detection accuracy. Wang Biao et al. proposed a new type of remote sensing semantic change detection twin convolutional network in the Chinese patent with the publication number CN118470537A, which introduced a difference feature enhancement module and a feature selection interaction module to realize multi-dimensional deep fusion of change features, thereby enhancing the information integration capability between the features of multi-temporal remote sensing images. From the perspective of "perception-analysis-extraction", this method improves the detection accuracy and practicality. In order to further enhance the robustness and accuracy of the model in detecting changes, various attention mechanisms are introduced into the twin network. The principle of these attention mechanisms is not simply to add or increase features, but to reweight features by calculating the attention weight of each position or channel, so that the network can better utilize important information. Common attention mechanisms include channel attention, spatial attention, cross-attention mechanism, etc. The emergence of transformer-based architecture has greatly challenged the dominance of convolutional neural networks. Transformer-based change detection models are usually combined with convolutional neural networks, using parallel or serial design to take advantage of both to improve change detection performance. Liu Wei et al. constructed an end-to-end network architecture combining double-time twin network, Transformer module, multi-task head and spatio-temporal feature learning module in the Chinese patent with the publication number CN118397480A. This method takes advantage of multi-scale feature visual markers and the ability of Transformer to model long-distance dependencies to jointly perform semantic segmentation and change detection tasks, and achieves high-performance semantic change detection by fully utilizing the correlation between the two. In addition, some methods introduce attention mechanisms, graph neural networks, long short-term memory networks, etc. into transformers to further enhance detection performance.

[0003] In summary, although the deep learning-based method significantly improves the performance of change detection, there are still problems such as insufficient mining of time sequence relationship, insufficient context modeling, and the like, and it is difficult to comprehensively extract discriminative information of the change region. Meanwhile, most methods mainly focus on local or low-order feature relationship, and lack modeling ability for high-order spatio-temporal dependence relationship, which limits the detection effect in complex scenes. SUMMARY

[0004] In order to solve the problems of insufficient mining of time sequence relationship and insufficient modeling of high-order spatio-temporal dependence relationship in the prior art, the present application provides a remote sensing image change detection method based on relationship-aware high-order interaction.

[0005] In order to achieve the above purpose, in a first aspect, the present application provides a remote sensing image change detection method based on relationship-aware high-order interaction, comprising the following steps: Collecting double-time-phase remote sensing images of the same area at different periods to form a double-time-phase data set; Based on a remote sensing image change detection model, inputting a double-time-phase remote sensing image pair in the double-time-phase data set, extracting single-time branch features and double-time-phase fusion features through a twin convolutional encoder, splicing the single-time branch features and the double-time-phase fusion features to obtain double-branch fusion features, and extracting high-order interaction features of spatial dimensions through recursive operation on the double-branch fusion features; Splicing the double-time-phase fusion features and the high-order interaction features, restoring the spatial resolution through cascaded upsampling, and finally obtaining the change detection result.

[0006] When extracting the double-time-phase fusion features through the twin convolutional encoder, the spatial relationship features and the channel relationship features are obtained through a time relationship perception module, the channel relationship features and the spatial relationship features are additively fused to obtain the double-time-phase fusion features.

[0007] Further, when obtaining the spatial relationship features through the time relationship perception module, the spatial relationship features are modeled based on a spatial relationship feature modeling module, which specifically comprises: For a given input feature , the spatial attention As is calculated:

[0008] wherein, and represent global max-pooling and global average-pooling operations respectively, represents concatenation along the channel dimension, represents a convolution operation, represents an activation function; The spatial relationship matrix of the double-time-phase remote sensing image pair is calculated, and the spatial attention is divided into grouping the sub-matrix to obtain a spatial relation affinity matrix :

[0009] trainable parameters through learning , and a full connection layer, to calculate the weighted coefficients of the input features and : :

[0010] input features and are weighted and fused to obtain spatial relation features : .

[0011] Further, a channel relation is obtained, and a channel relation modeling module includes: calculating channel attention Ac:

[0012] wherein, represents a global average pooling operation, represents a full connection layer; the channel attention is divided into r groups to obtain a sub-matrix with a size of , and the sub-matrix is used for relation calculation to obtain channel relation features .

[0013] Further, based on the high-order space-time interaction module, the spatial dimension high-order interaction features of the double-branch fusion features are extracted through a recursive operation. The high-order space-time interaction module includes four high-order space-time interaction units stacked in succession. In each high-order space-time interaction unit, a high-order interaction block is responsible for obtaining complex high-order interaction features, and a local dense set attention module is used to enhance the expression ability of local features. Both the high-order interaction block and the local dense set attention module include a residual connection and a normalization operation. A bidirectional gated recurrent unit is added after the high-order interaction block and the local dense set attention module.

[0014] Further, the spatial dimension high-order interaction features of the double-branch fusion features are extracted through a recursive operation, including: input features are expanded in the spatial dimension through two-dimensional convolution to generate new features . Next, the channels are divided into two parts, wherein the main features retain part of the information of the original features, and the auxiliary features ​Capture more context information, assist features Using deep convolution, get context features .

[0015] The context features Equally divided by channel Sub-feature, sub-feature calculation as follows:

[0016] Where, alpha is the scaling factor for stable training, each layer of recursive update feature, recursive generated feature is output by two-dimensional convolution, get high-order interaction features.

[0017] Further, the loss function of the remote sensing image change detection model is constructed as follows during training: The structure loss contains two parts: weighted binary cross entropy loss And weighted IoU loss ; In the decoder output part, the standard cross entropy loss is applied to measure the difference between the model prediction result and the real label; The total loss function is:

[0018] Where, The cross entropy loss weight of the decoder output is The structure loss weight of the time relationship perception module output is The structure loss weight of the high-order space-time interaction module output is

[0019] In the second aspect, the application provides a remote sensing image change detection system based on relationship perception high-order interaction, comprising a data acquisition module, a feature extraction module and a decoding module; the data acquisition module is used for collecting double-time remote sensing images of the same area in different periods to form a double-time data set; The feature extraction module is based on a remote sensing image change detection model, inputs a pair of double-time remote sensing images in the double-time data set, extracts single-time branch features and double-time fusion features through a twin convolutional encoder, splices the single-time branch features and the double-time fusion features to obtain double-branch fusion features, and extracts high-order interaction features in spatial dimension through recursive operation on the double-branch fusion features; The decoding module splices the double-time fusion features and the high-order interaction features, restores the spatial resolution through cascaded upsampling, and finally obtains the change detection result.

[0020] In a third aspect, the present application simultaneously provides a computer device, comprising a processor and a memory, the memory being used to store a computer executable program, the processor reading part or all of the computer executable program from the memory and executing, the processor executing part or all of the computer executable program to realize the remote sensing image change detection method based on relationship-aware high-order interaction.

[0021] A computer readable storage medium can also be provided, the computer readable storage medium storing a computer program, the computer program being executed by the processor to realize the remote sensing image change detection method based on relationship-aware high-order interaction.

[0022] The present application has the following advantages: The present application provides a remote sensing image change detection method based on relationship-aware high-order interaction, aiming to solve the problems of insufficient modeling ability for complex change regions, insufficient feature interaction and incomplete utilization of context information in the prior art. By introducing a time relationship-aware module, the model can fully exploit the spatial and channel relationships between double-time-phase images, realize more sufficient temporal information fusion, and effectively improve the discrimination ability of change detection. At the same time, the introduced high-order spatio-temporal interaction module performs deep feature modeling in the spatio-temporal dimension through a recursive high-order interaction mechanism, not only significantly enhancing the expression ability for complex spatial relationships, but also taking into account the modeling of global context information, improving the overall representation ability of the model. In addition, the loss function design fully considers the imbalance and uncertainty of the change region in the remote sensing image, and through the joint optimization of weighted binary cross-entropy and weighted IoU loss, the model pays more attention to the fine-grained regions that are difficult to detect, and improves the robustness of the model in terms of boundary blur and small target detection. Finally, the extensive experimental results on four remote sensing image change detection datasets verify the effectiveness of the method. In summary, the technical scheme of the present application effectively improves the precision and robustness of remote sensing image change detection through structured network design and multi-stage deep supervision strategy. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The figure is the overall framework of the relationship-aware high-order spatio-temporal interaction network in the embodiment of the present application.

[0024] Figure 2 The figure is the structure of the time relationship-aware module in the embodiment of the present application.

[0025] Figure 3 The figure is the structure of the high-order interaction unit in the high-order spatio-temporal interaction module in the embodiment of the present application.

[0026] Figure 4 The figure is the structure of the high-order interaction block in the high-order spatio-temporal interaction module in the embodiment of the present application.

[0027] Figure 5 This is a comparison chart of the detection results of different methods on the LEVIR-CD dataset in the embodiments of the present invention.

[0028] Figure 6 This is a comparison chart of the detection results of different methods on the WHU-CD dataset in the embodiments of the present invention.

[0029] Figure 7 This is a comparison chart of the detection results of different methods on the SYSU-CD dataset in the embodiments of the present invention.

[0030] Figure 8 This is a comparison chart of the detection results of different methods on the CDD dataset in the embodiments of the present invention. DETAILED DESCRIPTION

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

[0032] The structure of the relation-aware high-order spatiotemporal interaction network is as follows: Figure 1 Specifically, the relationship-aware high-order spatiotemporal interaction network includes a single-temporal branch and a dual-temporal branch. The single-temporal branch is used for independent forward propagation of single-temporal image features to maximize the preservation of original spatiotemporal information and ensure that key time series features are not overly blurred during the subsequent feature fusion process. The dual-temporal branch is used for interactive fusion and enhancement of dual-temporal features. By collaboratively processing the single-temporal branch features obtained from the single-temporal branch and the dual-temporal fusion features, it captures change information and improves sensitivity to changing areas.

[0033] At the shallow level, each convolutional block is connected to a temporal relationship perception module, which is used for preliminary interaction and fusion of bi-temporal features at an early stage. At the deep level, the model concatenates the features of the single-temporal branch and the dual-temporal branch to generate a joint feature containing rich spatiotemporal information as the final output of the temporal relationship perception module. This joint feature is then input into the high-order spatiotemporal interaction module. In the high-order spatiotemporal interaction module, four stacked high-order interaction units are processed, and a recursive mechanism is used to gradually refine the deep features, ultimately resulting in the output features of the high-order spatiotemporal interaction module.

[0034] In the decoding stage, the fused features of each layer of the dual-time branch are further spliced ​​with the high-order interaction features output by the high-order spatiotemporal interaction module, and then the spatial resolution is gradually restored through the cascade upsampler in the decoder to finally generate the change detection results.

[0035] like Figure 2 As shown, the temporal relationship perception module processes the dual-temporal remote sensing image pair. Specifically, for a given input feature , the spatial attention As is calculated as:

[0036] where, and denote global max pooling and global average pooling operation, respectively, denotes concatenation along channel dimension, denotes convolution operation, denotes activation function.

[0037] To further establish the spatial relationship between the two temporal remote sensing image pairs, the spatial relationship matrix is calculated, and the spatial attention is divided by grouping operation. Specifically, the spatial attention is divided into groups, obtaining sub-matrix. Grouping operation not only can reduce the computational complexity, making each sub-matrix easier to process, but also can enhance the modeling ability of local features, prompting the model to better capture the relationship between different spatial regions.

[0038] Subsequently, the spatial relationship affinity matrix is obtained by matrix multiplication:

[0039] Next, the weighted coefficients , and the full connection layer are calculated by learning the trainable parameters and :

[0040] Finally, the input features and are weighted and fused to obtain the spatial relationship features . The process can be represented as: .

[0041] For channel relationship modeling, the calculation process of channel attention Ac is as follows:

[0042] where, denotes global average pooling operation, denotes full connection layer.

[0043] Similarly, the channel attention is divided into r groups, obtaining sub-matrix with size , and the relationship is calculated by using the sub-matrix to obtain the channel relationship features .

[0044] ​​Finally, the spatial relationship features and the channel relationship features are additively fused to obtain final output features:

[0045] The spatial relationship modeling enables the model to focus on important spatial features in local regions, avoiding the redundancy of global calculation, while the channel relationship modeling can better capture the interdependence between different channels. Through the mutual fusion of spatial and channel relationships, the temporal relationship perception module can effectively capture the spatial and channel relationships between the double-time input features, achieving more sufficient feature fusion.

[0046] The output result processed by the temporal relationship perception module is sent to the high-order spatio-temporal interaction module for processing. The high-order spatio-temporal interaction module includes four high-order spatio-temporal interaction units stacked in sequence, as shown in reference Figure 3 In each high-order spatio-temporal interaction unit: The high-order interaction block is responsible for obtaining complex high-order interaction features. The local dense set-to-set attention module can enhance the expression ability of local features. In addition, each sub-module includes a residual connection and a normalization operation, and a bidirectional gated recurrent unit is added after the high-order interaction module and the local dense set-to-set attention module to dynamically adjust feature updates and improve sequence modeling capability.

[0047] The main goal of the high-order interaction module is to realize the extraction of high-order interaction features in the spatial dimension through recursive operation. As shown in reference Figure 4 , the input feature is expanded in the spatial dimension through two-dimensional convolution to generate a new feature . Next, the channels are divided into two parts, where the main feature retains part of the information of the original feature, and the auxiliary feature is used to capture more context information. The auxiliary feature is subjected to deep convolution to obtain context feature , whose calculation expression is as follows:

[0048] Next, the context feature is further divided into sub-features by channel, calculated as follows:

[0049] where α is a scaling coefficient used to stabilize training. Each layer recursively updates the feature through the following recursive formula:

[0050] where is the first recursive features are used to match the linear transformation of the channel dimension, represents element-wise multiplication, For the In this way, the high-order interaction module can model the complex neighborhood relationship in the spatial dimension layer by layer, while recursively accumulating the interaction information between high-order features. Finally, the recursively generated features After two-dimensional convolution output, the extraction of high-order interactive features is completed.

[0051] The high-order interaction module takes into account both computational efficiency and modeling capabilities. By introducing a recursive high-order interaction mechanism, it achieves deep feature mining in the spatial dimension. It can not only efficiently extract features with complex interaction relationships in space, but also maintain the ability to model the global context.

[0052] Next, the output of the high-order interaction module is reconstructed and concatenated with the dual-phase fusion features output by the temporal relationship perception module. The decoder gradually performs cascade upsampling to restore the spatial resolution, and the segmentation head finally outputs the change detection results.

[0053] In order to verify the effect of the present invention, the following verification experiment was carried out.

[0054] In the present application, experiments were conducted using four data sets: LEVIR-CD, WHU-CD, SYSU-CD, and CDD. LEVIR-CD contains 637 pairs of high-resolution remote sensing images (1024x1024, 0.5m / pixel), with 31333 change instances labeled, divided into 445 pairs of training set, 64 pairs of validation set, and 128 pairs of test set, after cropping, a total of 7120 training blocks, 1024 validation blocks, and 2048 test blocks were generated. The WHU-CD dataset consists of two aerial images taken in 2012 and 2016 respectively, with a resolution of 0.3m / pixel and an image size of 32507x15354, focusing on post-earthquake building change detection, after cropping and removing non-change areas, 7120 pairs of training images, 1024 pairs of validation images, and 2048 pairs of test images were obtained. The SYSU-CD dataset covers changes in a certain area from 2007 to 2014, containing a total of 20000 pairs of 256x256 size images, covering urban construction, suburban expansion, vegetation and road changes, etc. divided into 12000 training set, 4000 pairs of validation set and 4000 pairs of test set. The CDD dataset focuses on real and synthetic seasonal changes, containing 16000 pairs of 256x256 images, covering artificial changes, natural changes and mixed changes, divided into 10000 pairs of training, 3000 pairs of validation and 3000 pairs of test, providing rich benchmark resources for remote sensing image change detection. Five common evaluation indicators in change detection are used to measure the performance of the proposed model, including precision (Precision), recall (Recall), overall accuracy (OA), F1 score (F1), and intersection over union (IoU).

[0055] Figure 5 、 Figure 6 、 Figure 7 and Figure 8 , Table 1, Table 2, Table 3 and Table 4 are the detection results of three data sets on different methods. Among them, the comparison algorithms include FC-EF, IFNet, STANet, CDNet, BiT, ChangeFormer, SNUNet, P2V, SEIFNet, MFINet and SAF, etc.

[0056] Table 1 Evaluation index of LEVIR-CD dataset detection results

[0057] Table 2 Evaluation index of WHU-CD dataset detection results

[0058] Table 3 Evaluation index of SYSU-CD dataset detection results

[0059] Table 4 CDD dataset detection result evaluation index

[0060] On the LEVIR-CD dataset, the method of the present application achieved an F1 Score of 91.31 and an IoU of 84.01, significantly better than other comparative algorithms. Traditional methods such as FC-EF and IFNet have certain limitations in shallow feature processing and feature fusion methods, and cannot fully capture the spatio-temporal change information of dual-time remote sensing images. While the methods of STANet and SNUNet, although to some extent, improve the feature expression ability, but the ability of modeling the fine-grained features of the change area is still insufficient, which is easy to miss or misreport in complex scenes. The method of the present application has a significant advantage in spatio-temporal feature modeling and adaptive fusion capability, and can fully capture the spatio-temporal change information of dual-time remote sensing images. On the WHU-CD dataset, the F1 Score of the method of the present application is 91.18, and the IoU is 83.79. The change type of this dataset is mainly building reconstruction, and the background change is complex. In this dataset, the method of the present application performs relatively outstanding in suppressing pseudo changes. The model of the present application effectively filters irrelevant background changes by introducing a time relationship perception module (TRAM), thereby suppressing pseudo changes. On the SYSU-CD dataset, the F1 Score of the method of the present application is 79.93, and the IoU is 66.57, and the performance is also better than other comparative algorithms. This dataset contains multiple change types, including urban building construction, vegetation change, etc., and previous methods often have obvious shortcomings in capturing fine-grained changes, especially in complex urban expansion or road construction scenarios. The model of the present application can maintain high robustness in multiple change types through adaptive learning of the TRAM module. On the CDD dataset, the F1 Score of the method of the present application is 96.89, and the IoU is 93.96, and the performance is also outstanding. In contrast, the method of the present application introduces a time relationship perception module, models the spatial and channel relationship, realizes the adaptive fusion of dual-time image features, and enhances the model's ability to perceive subtle changes; the HOSTIM module further excavates the deep relationship of spatio-temporal features through a recursive high-order interaction mechanism, ensuring that the model captures complex change information while retaining rich contextual features.

[0061] Another optional embodiment, based on the technical concept of the method described in the present application, the present application provides a remote sensing image change detection system based on relationship perception high-order interaction, including a data acquisition module, a feature extraction module and a decoding module; the data acquisition module is used to collect dual-time remote sensing images of the same area in different periods to form a dual-time dataset; The feature extraction module inputs a pair of double-time remote sensing images in a double-time data set based on a remote sensing image change detection model, extracts single-time branch features and double-time fusion features through a twin convolutional encoder, splices the single-time branch features and the double-time fusion features to obtain double-branch fusion features, and extracts high-order interaction features of spatial dimensions through recursive operation on the double-branch fusion features. The decoding module splices the double-time fusion features and the high-order interaction features, restores the spatial resolution through cascaded upsampling, and finally obtains the change detection result.

[0062] In another aspect, the present application provides a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and the computer program is executable by a processor to implement the remote sensing image change detection method based on relationship-aware high-order interaction.

[0063] The present application also provides a computer device, which comprises a processor and a memory, the memory being used to store a computer executable program, and the processor reading the computer executable program from the memory and executing, and the processor being capable of implementing the remote sensing image change detection method based on relationship-aware high-order interaction when executing the computer executable program.

[0064] The computer device can be a notebook computer, a desktop computer or a workstation.

[0065] The processor can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC) or a ready-to-program field programmable gate array (FPGA).

[0066] The memory can be an internal storage unit of a notebook computer, a desktop computer or a workstation, such as a memory or a hard disk, or can be an external storage unit, such as a mobile hard disk or a flash card.

[0067] The computer-readable storage medium can include computer storage medium and communication medium. The computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. The computer-readable storage medium can include read-only memory (ROM), random access memory (RAM), solid state disk (SSD) or optical disk, etc. Among them, the random access memory can include resistance random access memory (ReRAM) and dynamic random access memory (DRAM).

[0068] So far, the embodiments of the present application have been described in detail in conjunction with the accompanying drawings. It should be noted that the implementation not shown or described in the drawings or the text of the specification is known to those skilled in the art, and is not described in detail. In addition, the definition of each component described above is not limited to the various specific structures, shapes or methods mentioned in the embodiments, and can be simply changed or replaced by those skilled in the art.

[0069] In summary, the present application acquires remote sensing images of the same area at different times, constructs a double-time-phase data set, and uses an independent convolutional encoder to extract single-time-phase features while preserving their original spatio-temporal information. A spatio-temporal feature pyramid is constructed through a twin network structure, a time relationship perception module is introduced to model the spatial and channel dimensions respectively, and a grouping attention mechanism and affinity matrix calculation method are used to capture the mutual relationship between pixels and channels, realizing deep fusion between features. Further, the high-order spatio-temporal interaction module stacks multiple interaction units through a recursive structure, strengthens the modeling of context information in the spatial dimension, and thus realizes complex high-order feature interaction expression. In terms of optimization, a structure loss function is proposed, combining weighted cross-entropy loss and weighted IoU loss to improve the detection capability of the model for edge regions and difficult regions, and introducing standard cross-entropy as auxiliary supervision at the output end of the decoder. Finally, the trained model can input double-time remote sensing images to predict and output high-precision change detection results.

Claims

1. A remote sensing image change detection method based on relationship-aware high-order interaction, characterized in that: The following steps are involved: Collect dual-temporal remote sensing images of the same area at different times to form a dual-temporal dataset; Based on the remote sensing image change detection model, a pair of dual-temporal remote sensing images from a dual-temporal dataset is input. Single-temporal branch features and dual-temporal fusion features are extracted through a twin convolutional encoder. The single-temporal branch features and the dual-temporal fusion features are concatenated to obtain dual-branch fusion features. High-order interaction features in the spatial dimension are extracted from the dual-branch fusion features through recursive operations. The dual-temporal fusion features are spliced ​​with high-order interaction features, and the spatial resolution is restored through cascade upsampling to finally obtain the change detection results.

2. The remote sensing image change detection method based on relationship-aware high-order interaction according to claim 1 is characterized in that: When extracting dual-phase fusion features through the twin convolutional encoder, the spatial relationship features and channel relationship features are obtained through the time relationship perception module, and the channel relationship features and spatial relationship features are additively fused to obtain dual-phase fusion features.

3. The remote sensing image change detection method based on relationship-aware high-order interaction according to claim 2 is characterized in that: The spatial relationship features are obtained through the temporal relationship perception module and based on the spatial relationship feature modeling module, specifically including: For a given input feature , calculate the spatial attention As: in, and Represent the global maximum pooling and global average pooling operations respectively, represents splicing along the channel dimension, represents the convolution operation, express Activation function; Calculate the spatial relationship matrix of the dual-temporal remote sensing image pair and use grouping operation to allocate spatial attention Divided into groups, obtain a submatrix, and based on the submatrix, obtain the spatial relationship affinity matrix through matrix multiplication : Through the learned trainable parameters 、 And the fully connected layer, calculate the input features and The weighting coefficient and : Input features and Perform weighted fusion to obtain spatial relationship features : 。 4. The remote sensing image change detection method based on relationship-aware high-order interaction according to claim 2 is characterized in that: Obtain channel relationships based on the channel relationship modeling module, including: Calculate channel attention Ac: in, represents the global average pooling operation, represents a fully connected layer; Divide the channel attention into r groups, and get a size of The sub-matrix is ​​used to calculate the relationship and obtain the channel relationship feature .

5. The remote sensing image change detection method based on relationship-aware high-order interaction according to claim 1 is characterized in that: Based on the high-order spatiotemporal interaction module, the dual-branch fusion features are recursively operated to extract high-order interaction features in the spatial dimension. The high-order spatiotemporal interaction module contains four continuously stacked high-order spatiotemporal interaction units. In each high-order spatiotemporal interaction unit, the high-order interaction block is responsible for obtaining complex high-order interaction features. The local dense synthetic attention module is used to enhance the expressive power of local features. Both the high-order interaction block and the local dense synthetic attention module include residual connections and normalization operations; a bidirectional gated recurrent unit is added after the high-order interaction block and the local dense synthetic attention module.

6. The remote sensing image change detection method based on relationship-aware high-order interaction according to claim 5 is characterized in that: The high-order interactive features of the spatial dimension are extracted by recursive operation of the dual-branch fusion feature, including: input features Expand the spatial dimension through two-dimensional convolution to generate new features ,Next, the channel is divided into two parts, where the main feature Retain part of the original feature information, auxiliary features Capture more contextual information for auxiliary features Use deep convolution to obtain contextual features ; context features Divided into channels sub-features, and the sub-features are calculated as follows: Among them, α is the scaling factor used to stabilize training. Each layer recursively updates the features, and the recursively generated features are output by two-dimensional convolution to obtain high-order interactive features.

7. The remote sensing image change detection method based on relationship-aware high-order interaction according to claim 1 is characterized in that: The loss function for remote sensing image change detection model training is constructed as follows: The structural loss consists of two parts: weighted binary cross entropy loss and weighted IoU loss ; At the decoder output, a standard cross-entropy loss is applied to measure the difference between the model prediction and the true label. The total loss function is: in, is the weight of the cross entropy loss of the decoder output, is the structural loss weight output by the temporal relationship perception module, It is the structural loss weight output by the high-order spatiotemporal interaction module.

8. A remote sensing image change detection system based on relationship-aware high-order interaction, characterized in that: It includes a data acquisition module, a feature extraction module and a decoding module; the data acquisition module is used to collect dual-phase remote sensing images of the same area at different times to form a dual-phase data set; The feature extraction module is based on the remote sensing image change detection model. It inputs a pair of dual-temporal remote sensing images from the dual-temporal dataset and extracts single-temporal branch features and dual-temporal fusion features through a twin convolutional encoder. The single-temporal branch features and the dual-temporal fusion features are then concatenated to obtain dual-branch fusion features. The dual-branch fusion features are then recursively operated to extract high-order interaction features in the spatial dimension. The decoding module concatenates the dual-temporal fusion features with the high-order interaction features, restores the spatial resolution through cascade upsampling, and finally obtains the change detection results.

9. A computer device, characterized in that: It includes a processor and a memory, the memory is used to store computer executable programs, the processor reads part or all of the computer executable programs from the memory and executes them, and when the processor executes part or all of the computer executable programs, it can implement the remote sensing image change detection method based on relationship perception high-order interaction as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the remote sensing image change detection method based on relationship-aware high-order interaction as described in any one of claims 1 to 7.

Images