Remote sensing image change detection method based on hybrid vision Mamba network

Abstract

The application discloses a remote sensing image change detection method based on a mixed vision Mamba network and belongs to the field of remote sensing image detection, and comprises the following steps: S1, acquiring double-time-phase remote sensing images, inputting the double-time-phase remote sensing images into a twin mixed encoder to obtain global features and local features; S2, inputting the global features and the local features into a global difference decoder and a local difference decoder respectively to obtain global difference features and local difference features; S3, inputting the global difference features and the local difference features into a double-branch fusion module to obtain fusion features; and S4, inputting the fusion features into a fusion decoder to obtain final prediction results. The twin mixed encoder based on a multi-mixed module is constructed, global and local information is effectively integrated by using the mixed module, and the problem that Mamba lacks local clues when processing a change detection task is relieved. The double-branch fusion module is proposed, and more discriminative change features are extracted for a detection task.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image detection, and in particular relates to a remote sensing image change detection method based on a hybrid vision Mamba network. Background Art

[0002] The goal of the change detection task is to detect meaningful changes in two remote sensing images acquired at the same location at different times. Change detection plays an important role in various fields, including urban planning, land cover analysis, disaster assessment, ecosystem monitoring, and resource management.

[0003] Advances in deep learning technology have brought a promising solution to the field of change detection. Existing models commonly use CNNs or Transformers as twin encoders to extract bi-temporal features. CNNs are very effective at extracting complex and correlated features, particularly local features, but often lack the ability to capture global dependencies. Transformers, through their self-attention mechanism, excel at processing complex spatial transformations and capturing long-range feature dependencies, helping to form comprehensive global representations. However, the complexity of image processing using Transformers scales quadratically with the length of the image patch. This results in significant computational cost, making them unsuitable for dense prediction tasks such as change detection. Recently, Mamba introduced time-varying parameters into the state-space model (SSM), enabling data-dependent global modeling with linear complexity, achieving significant success in natural language processing. The Mamba architecture was subsequently extended to computer vision and used in change detection, where it is considered an effective alternative to the Transformer. However, in dense prediction tasks such as change detection, local information still plays a crucial role in accurate detection. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the prior art, the remote sensing image change detection method based on hybrid vision Mamba network provided by the present invention solves the problem that the Mamba-based remote sensing change detection method ignores local information and is difficult to achieve fine detection in intensive prediction tasks.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a remote sensing image change detection method based on a hybrid vision Mamba network, comprising the following steps:

[0006] S1. Obtain a dual-temporal remote sensing image and input the dual-temporal remote sensing image into a twin hybrid encoder to obtain global features and local features.

[0007] S2. Input the global features and local features into the global difference decoder and the local difference decoder respectively to obtain global difference features and local difference features;

[0008] S3, inputting the global difference feature and the local difference feature into the dual-branch fusion module to obtain the fusion feature;

[0009] S4. Input the fusion features into the fusion decoder to obtain the final prediction result.

[0010] Further: in S1, the twin hybrid encoder includes a backbone module, a first hybrid module, a second hybrid module, and a third hybrid module connected in sequence;

[0011] Among them, the backbone module includes 3×3 convolutional layers connected to each other;

[0012] The first hybrid module, the second hybrid module, and the third hybrid module have the same structure, and all include parallel CNN branches and Mamba branches, and the CNN branches and the Mamba branches are connected through a dual-branch interaction module;

[0013] The Mamba branch is provided with a VSS submodule, the CNN branch is provided with a first local submodule and a second local submodule connected to each other, and the interaction module of the two branches includes a first interaction submodule and a second interaction submodule;

[0014] The output end of the first local submodule is connected to the input end of the VSS submodule through the first interaction submodule, and the output end of the VSS submodule is connected to the input end of the second local submodule through the second interaction submodule.

[0015] Further: the VSS submodule includes a first normalization layer, a first linear layer, a depthwise convolution layer, an SS2D unit, and a second normalization layer connected in sequence, the output end of the second normalization layer is connected to the first input end of the first element multiplication, the input end of the first normalization layer serves as the input end of the VSS submodule, the output end of the first normalization layer is connected to the second input end of the first element multiplication through the second linear layer, the output end of the first element multiplication is connected to the first input end of the second element multiplication through the third linear layer, the output end of the first normalization layer is also connected to the second input end of the second element multiplication, and the output end of the second element multiplication serves as the output end of the VSS submodule;

[0016] The first local submodule and the second local submodule have the same structure, both including a reparameterized 3×3 convolution, a first 1×1 convolution, a GELU activation function and a second 1×1 convolution connected in sequence. The second 1×1 convolution is also additively connected to the first element, and the reparameterized 3×3 convolution is also additively connected to the first element.

[0017] Furthermore, the dual-temporal remote sensing image in S1 includes remote sensing images before and after the change. The remote sensing images before and after the change are input into the twin hybrid encoder to obtain global features and local features of the input image, wherein the global features include the first to third Mamba features, and the local features include the first to third CNN features. The method for the twin hybrid encoder to obtain output features based on the input image is specifically as follows:

[0018] S11. Input the input image into the backbone module, extract the main features in the input image, and downsample its size to 1 / 4 of the input image to obtain a first feature map;

[0019] S12. Input the first feature map into a first hybrid module to obtain a first CNN feature and a first Mamba feature;

[0020] S13. Input the first CNN feature and the first Mamba feature into the second mixing module and the third mixing module in sequence, output the second CNN feature and the second Mamba feature through the second mixing module, and output the third CNN feature and the third Mamba feature through the third mixing module.

[0021] Furthermore: in S13, the method in which the second hybrid module outputs the second CNN feature and the second Mamba feature is specifically:

[0022] S131, the first CNN feature is input into the first local submodule to obtain a second feature map;

[0023] S132. Perform 1×1 convolution, deformation, and normalization operations on the second feature map in sequence through the first interaction submodule to obtain a third feature map;

[0024] S133, adding the first Mamba feature to the third feature map and inputting the resultant sum into the VSS submodule to obtain a second Mamba feature;

[0025] S134. Perform deformation, 1×1 convolution, and normalization operations on the second Mamba feature in sequence through the second interaction submodule to obtain a fourth feature map;

[0026] S135. Add the fourth feature map to the second feature map and input the resultant to the second local submodule to obtain a second CNN feature.

[0027] Furthermore: in S2, the global differential decoder and the local differential decoder have the same structure, both including first to third DIFF modules and a linear fusion module, and the first to third DIFF modules are all connected to the linear fusion module;

[0028] The first to third DIFF modules have the same structure, including a first 3×3 convolutional layer and a second 3×3 convolutional layer connected to each other.

[0029] Further: S2 includes the following steps:

[0030] S21, inputting the global feature into the global difference decoder to obtain the global difference feature;

[0031] S22, inputting the local features into a local difference decoder to obtain local difference features;

[0032] The global differential decoder and the local differential decoder process input features in the same way, and S21 includes the following sub-steps:

[0033] S211, input the first to third Mamba features into the first to third DIFF modules respectively;

[0034] S212: Merge the results generated by the first to third DIFF modules through a linear fusion module to obtain first to third multi-scale differential features;

[0035] S213. Upsample the first to third multi-scale differential features to 1 / 4 of the size of the dual-temporal remote sensing image, concatenate the first to third multi-scale differential features along the channel size, and use 1×1 convolution to adjust the number of channels to obtain a global difference feature.

[0036] Further: in S3, the dual-branch fusion module includes a first dynamic feature enhancement module, a second dynamic feature enhancement module and a cross-branch fusion Mamba module, and the first dynamic feature enhancement module and the second dynamic feature enhancement module are both connected to the cross-branch fusion Mamba module;

[0037] S3 includes the following sub-steps:

[0038] S31, fusing the global difference feature with the local difference feature to obtain a coarse-grained fusion feature;

[0039] S32, inputting the global difference feature, the local difference feature, and the coarse-grained fusion feature into a first dynamic feature enhancement module to obtain a first enhanced feature;

[0040] S33, inputting the global difference feature, the local difference feature, and the coarse-grained fusion feature into a second dynamic feature enhancement module to obtain a second enhanced feature;

[0041] S34. Input the first enhanced feature and the second enhanced feature into the cross-branch fusion Mamba module to obtain a fusion feature.

[0042] Further: S32 includes the following sub-steps:

[0043] S321, subtracting the global difference feature from the local difference feature to obtain a first difference feature;

[0044] S322, after the first difference feature is processed through a pooling operation, a difference weight between the global difference feature and the local difference feature is calculated through dynamic difference perception attention, and the difference weight is used for amplifying feature difference of the coarse-grained fusion feature, so as to obtain a second difference feature;

[0045] S323, the global difference feature is processed through a learnable description convolution, a generated result, the second difference feature and the global difference feature are added, so as to obtain a first enhanced feature;

[0046] In S34, an expression of the fusion feature X is specifically:

[0047]

[0048] In the formula, ECA is a channel attention module, is an element-wise addition operation, D1 is the first enhanced feature, D2 is the second enhanced feature, LN is a standardization layer, H f is a mixed feature addition result, and an expression thereof is specifically:

[0049]

[0050] In the formula, H1 is the first mixed feature, and H2 is the second mixed feature, and expressions thereof are specifically:

[0051]

[0052] In the formula, SS2D is a selective scanning 2D unit, is an element-wise multiplication operation, SiLU is a kind of neural network activation function, Linear is a linear layer, and H is an initial mixed feature, and an expression thereof is specifically:

[0053]

[0054] In the formula, Dwc(·) is a deep convolution operation.

[0055] Further, S4 is specifically:

[0056] The fusion feature is upsampled by four times through transposed convolution, the upsampled result is input into a prediction head composed of convolution, a change mask with the same size as the dual-phase remote sensing image is generated, and the change mask is used as a final prediction result.

[0057] The present application has the following advantages:

[0058] (1) The present invention provides a remote sensing image change detection method based on a hybrid visual Mamba network, constructs a twin hybrid encoder based on multiple hybrid modules, the hybrid module consists of CNN and VMamba, and uses a dual-branch interaction module to gradually interactively couple the features of the two branches of the encoder, thereby maximally retaining the global context information and local features of the dual-phase image and alleviating the problem of Mamba's lack of local clues when processing change detection tasks.

[0059] (2) The present invention proposes a dual-branch fusion module, which includes a dynamic feature enhancement module for enhancing fine texture features and perceptual difference features, and a cross-branch fusion Mamba module for effectively exploring the correlation between global and local difference features. Through the synergistic effect of these two sub-modules, more discriminative change features are extracted for detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flow chart of the remote sensing image change detection method based on the hybrid vision Mamba network of the present invention.

[0061] Figure 2 This is a structural diagram of the remote sensing image change detection method based on the hybrid vision Mamba network.

[0062] Figure 3 Schematic diagram of the hybrid module.

[0063] Figure 4 This is a structural diagram of the VSS module.

[0064] Figure 5 Schematic diagram of the SS2D unit structure.

[0065] Figure 6 Schematic diagram of the dual-branch fusion module.

[0066] Figure 7 Schematic diagram of the first dynamic feature enhancement module.

[0067] Figure 8 Schematic diagram of LDC.

[0068] Figure 9 This is a schematic diagram of cross-branch fusion of Mamba modules. DETAILED DESCRIPTION

[0069] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0070] like Figure 1 As shown, in one embodiment of the present invention, a remote sensing image change detection method based on a hybrid vision Mamba network includes the following steps:

[0071] S1. Obtain a dual-temporal remote sensing image and input the dual-temporal remote sensing image into a twin hybrid encoder to obtain global features and local features.

[0072] S2. Input the global features and local features into the global difference decoder and the local difference decoder respectively to obtain global difference features and local difference features;

[0073] S3, inputting the global difference feature and the local difference feature into the dual-branch fusion module to obtain the fusion feature;

[0074] S4. Input the fusion features into the fusion decoder to obtain the final prediction result.

[0075] In this embodiment, the network structure of the present invention is as follows Figure 2 As shown, the twin hybrid encoder uses the remote sensing image I before the change pre And the changed remote sensing image I post As input, calculate I pre Corresponding CNN features and Mamba features and I post Corresponding CNN features and Mamba features i is the stage, and i=1,2,3.

[0076] In S1, the twin hybrid encoder includes a backbone module, a first hybrid module, a second hybrid module, and a third hybrid module connected in sequence;

[0077] Among them, the backbone module includes 3×3 convolutional layers connected to each other;

[0078] like Figure 3 As shown, the first hybrid module, the second hybrid module and the third hybrid module have the same structure, and all include parallel CNN branches and Mamba branches, and the CNN branches and the Mamba branches are connected through a dual-branch interaction module;

[0079] The Mamba branch is provided with a VSS submodule, the CNN branch is provided with a first local submodule and a second local submodule connected to each other, and the interaction module of the two branches includes a first interaction submodule and a second interaction submodule;

[0080] The output end of the first local submodule is connected to the input end of the VSS submodule through the first interaction submodule, and the output end of the VSS submodule is connected to the input end of the second local submodule through the second interaction submodule.

[0081] In this embodiment, in order to make full use of local features and global representation, the present invention designs the hybrid module as a concurrent structure, which includes a CNN branch, a Mamba branch, and a dual-branch interaction module (FIM).

[0082] like Figure 4 As shown, the VSS submodule includes a first normalization layer, a first linear layer, a depth convolution layer, an SS2D unit, and a second normalization layer connected in sequence. The output end of the second normalization layer is connected to the first input end of the first element multiplication, and the input end of the first normalization layer serves as the input end of the VSS submodule. The output end of the first normalization layer is connected to the second input end of the first element multiplication through the second linear layer. The output end of the first element multiplication is connected to the first input end of the second element multiplication through the third linear layer. The output end of the first normalization layer is also connected to the second input end of the second element multiplication, and the output end of the second element multiplication serves as the output end of the VSS submodule.

[0083] The first local submodule and the second local submodule have the same structure, both including a reparameterized 3×3 convolution, a first 1×1 convolution, a GELU activation function and a second 1×1 convolution connected in sequence. The second 1×1 convolution is also additively connected to the first element, and the reparameterized 3×3 convolution is also additively connected to the first element.

[0084] In this embodiment, the VSS submodule consists of a linear layer and a depthwise convolutional layer, followed by an SS2D unit. The structure of the SS2D unit is as follows: Figure 5 As shown in Figure 2, the SS2D unit flattens the input and processes it from four different directions: top left to bottom right, bottom right to top left, top right to bottom left, and bottom left to top. This allows for the extraction of long-range dependencies in multiple directions. For each direction, a selective scan state model is learned by learning parameters A, B, C, D, and Δ, and the results are combined to produce the final output of the SS2D unit.

[0085] The dual-temporal remote sensing images in S1 include remote sensing images before and after the change. The remote sensing images before and after the change are input into the twin hybrid encoder to obtain global features and local features of the input image. The global features include the first to third Mamba features, and the local features include the first to third CNN features. The method for the twin hybrid encoder to obtain output features based on the input image is as follows:

[0086] S11. Input the input image into the backbone module, extract the main features in the input image, and downsample its size to 1 / 4 of the input image to obtain a first feature map;

[0087] S12. Input the first feature map into a first hybrid module to obtain a first CNN feature and a first Mamba feature;

[0088] S13. Input the first CNN feature and the first Mamba feature into the second mixing module and the third mixing module in sequence, output the second CNN feature and the second Mamba feature through the second mixing module, and output the third CNN feature and the third Mamba feature through the third mixing module.

[0089] In S13, the method for the second hybrid module to output the second CNN feature and the second Mamba feature is specifically as follows:

[0090] S131, the first CNN feature is input into the first local submodule to obtain a second feature map;

[0091] S132. Perform 1×1 convolution, deformation, and normalization operations on the second feature map in sequence through the first interaction submodule to obtain a third feature map;

[0092] S133, adding the first Mamba feature to the third feature map and inputting the resultant sum into the VSS submodule to obtain a second Mamba feature;

[0093] S134. Perform deformation, 1×1 convolution, and normalization operations on the second Mamba feature in sequence through the second interaction submodule to obtain a fourth feature map;

[0094] S135. Add the fourth feature map to the second feature map and input the resultant to the second local submodule to obtain a second CNN feature.

[0095] In the embodiment, the CNN branch uses two local sub-modules to provide a fulcrum for the bidirectional interaction between the two branches, each local sub-module uses a re-parameterized 3x3 convolution to extract local features, the interaction sub-module includes a 1x1 convolution for adjusting the number of feature channels, and a series of deformation and normalization operations for solving the misalignment problem between the feature maps of the CNN branch and the patch embedding of the Mamba branch, considering the complementarity of global features and local features, the application gradually integrates the local features of the CNN branch into the embedding block using the interaction sub-module, and enriches the Mamba branch with local details. Similarly, the global context from the Mamba branch is fed back to the feature map to enhance the global perception ability of the CNN branch. During the interaction of the two branches, the CNN features and Mamba features are embedded with each other.

[0096] As Figure 2 shown, in S2, the structures of the global difference decoder and the local difference decoder are the same, and each includes first to third DIFF modules and a linear fusion module, the first to third DIFF modules are connected with the linear fusion module;

[0097] The first to third DIFF modules have the same structure and include a first 3x3 convolution layer and a second 3x3 convolution layer connected with each other.

[0098] S2 includes the following steps:

[0099] S21, input the global feature into the global difference decoder to obtain the global difference feature;

[0100] S22, input the local feature into the local difference decoder to obtain the local difference feature;

[0101] The global difference decoder and the local difference decoder have the same method for processing input features, so the working process of the local difference decoder will not be described again.

[0102] S21 includes the following steps:

[0103] S211, input the first to third Mamba features into the first to third DIFF modules respectively;

[0104] S212, merge the results generated by the first to third DIFF modules through the linear fusion module to obtain the first to third multi-scale difference features;

[0105] S213, upsample the first to third multi-scale difference features to 1 / 4 of the size of the dual temporal remote sensing image, connect the first to third multi-scale difference features along the channel dimension, and use a 1x1 convolution to adjust the number of channels to obtain the global difference feature.

[0106] As Figure 2As shown, in this embodiment, the present invention also inputs global difference features into the prediction head to generate a binary change map, representing the prediction results of the global difference decoder. Similarly, local difference features generated by the local difference decoder are input into the prediction head to generate a binary change map, representing the prediction results of the local difference decoder. By having each decoder generate a binary change map that is one-quarter the size of the input image and supervised by the change labels, the length of the gradient backpropagation path is effectively reduced, facilitating model training.

[0107] like Figure 6 As shown, in S3, the dual-branch fusion module includes a first dynamic feature enhancement module, a second dynamic feature enhancement module and a cross-branch fusion Mamba module, and the first dynamic feature enhancement module and the second dynamic feature enhancement module are both connected to the cross-branch fusion Mamba module;

[0108] S3 includes the following sub-steps:

[0109] S31, fusing the global difference feature with the local difference feature to obtain a coarse-grained fusion feature;

[0110] S32, inputting the global difference feature, the local difference feature, and the coarse-grained fusion feature into a first dynamic feature enhancement module to obtain a first enhanced feature;

[0111] S33, inputting the global difference feature, the local difference feature, and the coarse-grained fusion feature into a second dynamic feature enhancement module to obtain a second enhanced feature;

[0112] S34. Input the first enhanced feature and the second enhanced feature into the cross-branch fusion Mamba module to obtain a fusion feature.

[0113] In this embodiment, the dynamic feature enhancement module takes global and local difference features as input and performs coarse-grained fusion within the module to obtain coarse-grained fused features. These features are then subtracted from features in different branches to obtain difference features, enhancing the mapping of these difference features. These difference features are then merged with the original features, enriched with supplementary information from other branches. This process effectively extracts and amplifies complementary features and texture details inherent in the image, thereby improving overall fusion performance. This helps the model effectively capture subtle differences between input features, thereby improving the resolution and perception of different features and enhancing fusion performance.

[0114] like Figure 7 As shown, S32 includes the following sub-steps:

[0115] S321, subtracting the global difference feature from the local difference feature to obtain a first difference feature;

[0116] S322, after processing the first difference feature through a pooling operation, calculate the difference weight between the global difference feature and the local difference feature through dynamic difference perception attention, and use the difference weight to amplify the feature difference in a coarse-grained fusion feature to obtain a second difference feature; wherein the calculation of the difference weight can effectively amplify the feature difference;

[0117] S323: Process the global difference feature through a learnable descriptor convolution, and add the generated result, the second difference feature, and the global difference feature to obtain a first enhanced feature;

[0118] In this embodiment, the difference between the methods used by the first dynamic feature enhancement module and the second dynamic feature enhancement module for processing input features lies in the fact that the inputs to the LDC are different: the input to the LDC in the first dynamic feature enhancement module is global difference features, while the input to the LDC in the second dynamic feature enhancement module is local difference features. LDC uses learnable mask parameters and convolution operations to enhance texture processing of the input feature map. By adjusting the weights of the convolution kernel, it emphasizes texture information and enhances the model's perception of texture features.

[0119] like Figure 9 As shown, in S34, the expression of the fusion feature X is specifically obtained as follows:

[0120]

[0121] Where ECA is the channel attention module, is an element-by-element addition operation, D1 is the first enhanced feature, D2 is the second enhanced feature, LN is the normalization layer, H f is the result of adding mixed features, and its specific expression is:

[0122]

[0123] Where H1 is the first hybrid feature, H2 is the second hybrid feature, and the specific expression is:

[0124]

[0125] Where SS2D is the selective scanning 2D unit, is an element-by-element multiplication operation, SiLU is the neural network activation function, Linear is the linear layer, and H is the initial mixed feature. Its specific expression is:

[0126]

[0127] Where Dwc(·) is the depthwise convolution operation.

[0128] In this embodiment, the cross-branch integration Mamba module is as follows Figure 9As shown in Figure 2, the cross-branch fusion Mamba module takes D1 and D2 as input, performs fine-grained fusion and explores the information correlation between different branches to generate the initial mixed feature H, which is then input into the SS2D layer to capture the spatial long-distance dependency and generate H1 and H2. The channel attention module is used to reduce channel redundancy and obtain the final fused feature X.

[0129] S4 is specifically:

[0130] The fused features are upsampled four times through transposed convolution, and the upsampling results are input into the prediction head composed of convolution to generate a change mask with the same size as the dual-temporal remote sensing image, which is used as the final prediction result.

[0131] The present invention optimizes the predicted change mapping obtained by the three decoders through cross entropy loss, which is expressed as:

[0132]

[0133] In the formula, y represents the true value, Represents the predicted value, and N represents the number of pixels.

[0134] The beneficial effects of the present invention are as follows: the present invention provides a remote sensing image change detection method based on a hybrid visual Mamba network, constructs a twin hybrid encoder based on multiple hybrid modules, the hybrid module is composed of CNN and VMamba, and the features of the two branches of the encoder are gradually interactively coupled through a dual-branch interaction module, thereby maximally retaining the global context information and local features of the dual-phase image, and alleviating the problem of Mamba's lack of local clues when processing change detection tasks.

[0135] The present invention proposes a dual-branch fusion module, which includes a dynamic feature enhancement module for enhancing fine texture features and perceptual difference features, and a cross-branch fusion Mamba module for effectively exploring the correlation between global and local difference features. Through the synergistic effect of these two sub-modules, more discriminative change features are extracted for detection tasks.

[0136] In the description of the application, it needs to be understood that the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the application. In addition, the terms "first", "second", "third" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implying the number of technical features indicated. Therefore, the features defined by "first", "second", "third" can explicitly or implicitly include one or more of the features.

Claims

1. A remote sensing image change detection method based on hybrid vision Mamba network is characterized by: The following steps are involved: S1. Obtain a dual-temporal remote sensing image and input the dual-temporal remote sensing image into a twin hybrid encoder to obtain global features and local features. S2. Input the global features and local features into the global difference decoder and the local difference decoder respectively to obtain global difference features and local difference features; S3, inputting the global difference feature and the local difference feature into the dual-branch fusion module to obtain the fusion feature; S4, input the fusion features into the fusion decoder to obtain the final prediction result; In S1, the twin hybrid encoder includes a backbone module, a first hybrid module, a second hybrid module, and a third hybrid module connected in sequence; Among them, the backbone module includes 3 × 3 convolutional layers connected to each other; The first hybrid module, the second hybrid module, and the third hybrid module have the same structure, and all include parallel CNN branches and Mamba branches, and the CNN branches and the Mamba branches are connected through a dual-branch interaction module; The Mamba branch is provided with a VSS submodule, the CNN branch is provided with a first local submodule and a second local submodule connected to each other, and the interaction module of the two branches includes a first interaction submodule and a second interaction submodule; The output end of the first local submodule is connected to the input end of the VSS submodule through the first interaction submodule, and the output end of the VSS submodule is connected to the input end of the second local submodule through the second interaction submodule; In S3, the dual-branch fusion module includes a first dynamic feature enhancement module, a second dynamic feature enhancement module and a cross-branch fusion Mamba module, and the first dynamic feature enhancement module and the second dynamic feature enhancement module are both connected to the cross-branch fusion Mamba module; S3 includes the following sub-steps: S31, fusing the global difference feature with the local difference feature to obtain a coarse-grained fusion feature; S32, inputting the global difference feature, the local difference feature, and the coarse-grained fusion feature into a first dynamic feature enhancement module to obtain a first enhanced feature; S33, inputting the global difference feature, the local difference feature, and the coarse-grained fusion feature into a second dynamic feature enhancement module to obtain a second enhanced feature; S34, inputting the first enhanced feature and the second enhanced feature into the cross-branch fusion Mamba module to obtain a fusion feature; S32 includes the following sub-steps: S321, subtracting the global difference feature from the local difference feature to obtain a first difference feature; S322, after the first difference feature is processed by the pooling operation, the difference weight between the global difference feature and the local difference feature is calculated by dynamic difference perception attention, and the difference weight is used for coarse-grained fusion feature amplification feature difference to obtain a second difference feature; S323: Process the global difference feature through a learnable descriptor convolution, and add the generated result, the second difference feature, and the global difference feature to obtain a first enhanced feature; In S34, the expression of the fusion feature X is obtained as follows: Where ECA is the channel attention module, is an element-by-element addition operation, D1 is the first enhanced feature, D2 is the second enhanced feature, LN is the normalization layer, H f is the result of adding mixed features, and its specific expression is: Where H1 is the first hybrid feature, H2 is the second hybrid feature, and the specific expression is: Where SS2D is the selective scanning 2D unit, is an element-by-element multiplication operation, SiLU is the neural network activation function, Linear is the linear layer, and H is the initial mixed feature. Its specific expression is: Where Dwc(·) is the depthwise convolution operation.

2. The remote sensing image change detection method based on hybrid vision Mamba network according to claim 1 is characterized in that: The VSS submodule includes a first normalization layer, a first linear layer, a depthwise convolution layer, an SS2D unit, and a second normalization layer connected in sequence. The output end of the second normalization layer is connected to the first input end of the first element multiplication. The input end of the first normalization layer serves as the input end of the VSS submodule. The output end of the first normalization layer is connected to the second input end of the first element multiplication through the second linear layer. The output end of the first element multiplication is connected to the first input end of the second element multiplication through the third linear layer. The output end of the first normalization layer is also connected to the second input end of the second element multiplication. The output end of the second element multiplication serves as the output end of the VSS submodule. The first local submodule and the second local submodule have the same structure, both including a reparameterized 3 × 3 convolution, a first 1 × 1 convolution, a GELU activation function, and a second 1 × 1 convolution connected in sequence. The second 1 × 1 convolution is also additively connected to the first element, and the reparameterized 3 × 3 convolution is also additively connected to the first element.

3. The remote sensing image change detection method based on hybrid vision Mamba network according to claim 2 is characterized in that: The dual-temporal remote sensing images in S1 include remote sensing images before and after the change. The remote sensing images before and after the change are input into the twin hybrid encoder to obtain the global features and local features of the input image. The global features include the first to third Mamba features, and the local features include the first to third CNN features. The method for the twin hybrid encoder to obtain the output features based on the input image is as follows: S11. Input the input image into the backbone module, extract the main features in the input image, and downsample its size to 1 / 4 of the input image to obtain a first feature map; S12. Input the first feature map into a first hybrid module to obtain a first CNN feature and a first Mamba feature; S13. Input the first CNN feature and the first Mamba feature into the second mixing module and the third mixing module in sequence, output the second CNN feature and the second Mamba feature through the second mixing module, and output the third CNN feature and the third Mamba feature through the third mixing module.

4. The remote sensing image change detection method based on hybrid vision Mamba network according to claim 3 is characterized in that: In S13, the method for the second hybrid module to output the second CNN feature and the second Mamba feature is specifically as follows: S131, the first CNN feature is input into the first local submodule to obtain a second feature map; S132. Perform 1×1 convolution, deformation, and normalization operations on the second feature map in sequence through the first interaction submodule to obtain a third feature map; S133, adding the first Mamba feature to the third feature map and inputting the resultant sum into the VSS submodule to obtain a second Mamba feature; S134. Perform deformation, 1×1 convolution, and normalization operations on the second Mamba feature in sequence through the second interaction submodule to obtain a fourth feature map. S135. Add the fourth feature map to the second feature map and input the resultant to the second local submodule to obtain a second CNN feature.

5. The remote sensing image change detection method based on hybrid vision Mamba network according to claim 3 is characterized in that: In S2, the global differential decoder and the local differential decoder have the same structure, both including the first to third DIFF modules and a linear fusion module, and the first to third DIFF modules are all connected to the linear fusion module; The first to third DIFF modules have the same structure, including a first 3×3 convolutional layer and a second 3×3 convolutional layer connected to each other.

6. The remote sensing image change detection method based on hybrid vision Mamba network according to claim 5 is characterized in that: S2 includes the following steps: S21, inputting the global feature into the global difference decoder to obtain the global difference feature; S22, inputting the local features into a local difference decoder to obtain local difference features; The global differential decoder and the local differential decoder process input features in the same way, and S21 includes the following sub-steps: S211, input the first to third Mamba features into the first to third DIFF modules respectively; S212, merging the results generated by the first to third DIFF modules through a linear fusion module to obtain first to third multi-scale differential features; S213. Upsample the first to third multi-scale differential features to 1 / 4 of the size of the dual-temporal remote sensing image, concatenate the first to third multi-scale differential features along the channel size, and use 1×1 convolution to adjust the number of channels to obtain a global difference feature.

7. The remote sensing image change detection method based on hybrid vision Mamba network according to claim 1 is characterized in that: S4 is specifically: The fused features are upsampled four times through transposed convolution, and the upsampling results are input into the prediction head composed of convolution to generate a change mask with the same size as the dual-temporal remote sensing image, which is used as the final prediction result.

Images