Remote sensing change detection method and system based on mixed attention Mama architecture

Abstract

The invention discloses a remote sensing change detection method and system based on a mixed attention Mama framework, and relates to the field of remote sensing image processing, and the method comprises the steps: inputting a multi-temporal remote sensing image pair into a weight sharing twin encoder, and extracting multi-scale features; performing staged fusion reconstruction through a decoder based on mixed attention Mama, performing parallel processing on text perception Mama and transpose attention through a mixed attention Mama block in each stage so as to cooperatively enhance local and global characterization, and inhibiting error propagation in combination with an interactive channel-space fusion block; and finally, a change probability graph is generated through the third reconstruction feature, and a binary change graph is output through argmax. According to the method, the linear calculation complexity is kept, long-range dependence modeling and fine-grained space structure sensing are considered, and the change detection precision and robustness are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing, and more specifically to a remote sensing change detection method and system based on a hybrid attention Mamba architecture. Background Technology

[0002] Remote sensing change detection is one of the core tasks of remote sensing information engineering, and it is widely used in major national demand scenarios such as disaster assessment, urban renewal, and dynamic monitoring of the ecological environment. In recent years, deep learning methods have become the mainstream technical approach: CNN-like methods, represented by FC-EF and FC-Siam-Diff, extract temporal difference features through Siamese convolutional structures, achieving good results on typical targets such as buildings; visual Transformer methods, represented by ChangeFormer, BIT, and SwinSUNet, introduce self-attention mechanisms to model long-range dependencies, significantly improving global consistency in complex scenes; and the recently emerging MambaBCD series of methods are based on the state-space model (SSM), using linear complexity state recursion to achieve efficient long sequence modeling, demonstrating excellent accuracy-efficiency balance on the LEVIR-CD and WHU-CD datasets.

[0003] However, existing technologies still have inherent limitations: CNNs are limited by their local receptive field, making it difficult to model long-distance semantic relationships across blocks or different object categories; while Transformers have global modeling capabilities, their self-attention computational complexity is O(N²), which leads to memory explosion and increased latency in inference of high-resolution remote sensing images (such as 1024×1024 and above); and while pure SSM architectures (such as MambaBCD) have the advantage of linear complexity, they ignore the two-dimensional spatial structure of the image due to one-dimensional serialization modeling, resulting in problems such as orientation sensitivity, blurred boundaries, and missed detection of small targets; in addition, in multi-stage encoder-decoder structures, feature upsampling layer by layer can easily lead to error accumulation, and the semantic differences between channels and spatial dimensions are not co-modeled, further weakening the positioning accuracy and robustness of changing boundaries. Summary of the Invention

[0004] The purpose of this invention is to provide a remote sensing change detection method and system based on a hybrid attention Mamba architecture, which can solve the technical problems of existing remote sensing change detection methods in high-resolution scenarios, such as weak long-range dependency modeling ability, high computational overhead, insufficient local structure modeling, and information loss and error accumulation in multi-stage fusion.

[0005] To achieve the above objectives, this application provides the following technical solution: The first aspect of this application provides a remote sensing change detection method based on a hybrid attention Mamba architecture, comprising the following steps: S1, input the multi-temporal remote sensing image pairs into a weight-shared twin encoder, and extract multi-scale features through the twin encoder; S2, input the multi-scale features into the decoder based on hybrid attention Mamba, and fuse and reconstruct the multi-scale features in stages through the decoder. In each stage, the first reconstructed features from the previous stage of the decoder and the second encoded features of the corresponding scale from the encoder are input into a hybrid attention Mamba block to obtain local enhanced features. The local enhanced features and the second encoded features are input into an interactive channel-space fusion block to obtain enhanced features. The enhanced features are upsampled and the third reconstructed features are output. S3. Based on the third reconstruction feature, a change probability map is obtained. Based on the change probability map, a binary change map is output to complete the change detection.

[0006] Furthermore, the multi-temporal remote sensing image pairs in S1 are and The twin encoder, based on the Visual Mamba architecture, extracts multi-scale features through four stages of downsampling and the Visual State Space module. and .

[0007] Furthermore, in S2, the hybrid attention-based Mamba decoder consists of four sequentially connected processing stages; For the j-th stage, j = 1, 2, 3, 4: The first reconstructed feature is the output of the (j-1)th processing stage. When j=1, the first reconstructed feature is the feature j=4 in the multi-scale feature pair after initialization processing. The second encoding feature is obtained by fusing the feature with index j in the multi-scale feature pair.

[0008] Furthermore, the specific process in S2 of inputting a set of features to be fused into a hybrid attention Mamba block for processing to obtain locally enhanced features includes: The set of features to be fused are concatenated and converted into a dense vector through an embedding layer; The dense vectors are input in parallel into a text-aware Mamba block and a transposed attention block for processing. The outputs of the text-aware Mamba block and the transposed attention block are integrated through a concatenation operation and a convolutional fusion layer to generate local enhanced features.

[0009] Furthermore, the specific process in S2 of inputting the local enhancement features and the second encoding features into the interactive channel-spatial fusion block for processing to obtain the enhancement features is as follows: The local enhancement feature is used as input feature F1, and the corresponding second encoded feature is used as input feature F2; The input features F1 and F2 are concatenated and then projected through pointwise convolution, batch normalization, and ReLU activation. Channel attention weights are generated through the channel feature enhancement module and used to recalibrate the projected features; Spatial structural differences are captured through a spatial feature enhancement module; The recalibrated features and the features after capturing spatial structure differences are aggregated by addition, and the enhanced features are output through residual connection.

[0010] Furthermore, the text-aware Mamba block uses a structure-aware state fusion method to capture local spatial dependencies, wherein multi-scale dilated convolution is used to enhance the combination of neighborhood states, and the dilation rate of the multi-scale dilated convolution is set to 1, 3, and 5.

[0011] Furthermore, the channel feature enhancement module specifically involves: applying global average pooling to generate channel descriptors, and generating the channel attention weights through two 1×1 convolutions and Sigmoid activation.

[0012] Furthermore, the spatial feature enhancement module specifically employs 3×3 depthwise convolution, batch normalization, and ReLU activation.

[0013] Furthermore, in S3, obtaining the change probability map based on the third reconstruction feature specifically involves passing the third reconstruction feature, which is the final output of the decoder, through a prediction head to obtain the change probability map. Based on the change probability map, the output of the binary change map is specifically achieved by performing an argmax operation on the change probability map to generate the binary change map.

[0014] A second aspect of this application provides a remote sensing change detection system based on a hybrid attention Mamba architecture, comprising: Feature extraction module: used to input multi-temporal remote sensing image pairs into a weight-shared twin encoder, and extract multi-scale features through the twin encoder; The fusion and reconstruction module is used to input multi-scale features into a hybrid attention Mamba-based decoder. The decoder fuses and reconstructs the multi-scale features in stages. In each stage, the first reconstructed features from the previous stage of the decoder and the second encoded features of the corresponding scale from the encoder are input into a hybrid attention Mamba block to obtain local enhanced features. The local enhanced features and the second encoded features are input into an interactive channel-space fusion block to obtain enhanced features. The enhanced features are upsampled to output a third reconstructed feature. Detection module: used to obtain a change probability map based on the third reconstructed feature, and output a binary change map based on the change probability map to complete change detection.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a remote sensing change detection method based on a hybrid attention Mamba architecture. It employs a weight-shared Siamese encoder to extract multi-scale features, constructs a hybrid attention Mamba-based decoder to achieve staged fusion and reconstruction, where each stage fuses the reconstructed features from the previous stage with the corresponding scale-encoded features. These are then processed by a hybrid attention Mamba block to generate local enhancement features, which are then combined with the encoded features and input into an interactive channel-spatial fusion block to generate further enhancement features. Finally, a third reconstructed feature is output via upsampling. This solves the problems of limited receptive field in CNNs, high computational overhead in Transformers, insufficient local modeling in SSM, and accumulation of multi-stage fusion errors in existing technologies. Therefore, while maintaining linear computational complexity, it simultaneously improves global context modeling capabilities and fine-grained local structure representation capabilities, significantly enhancing the accuracy of change boundary localization and the robustness of small target / weakly changing region identification. This results in change detection outperforming existing methods in key metrics such as Recall, Precision, and IoU. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a remote sensing change detection method based on a hybrid attention Mamba architecture in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the framework of a remote sensing change detection method based on hybrid attention Mamba in an embodiment of the present invention.

[0018] Figure 3 This is an experimental visualization analysis diagram of a remote sensing change detection method based on hybrid attention Mamba in an embodiment of the present invention. Detailed Implementation

[0019] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0020] See Figure 1 and Figure 2 This invention provides a remote sensing change detection method based on a hybrid attention Mamba architecture, comprising the following steps: Multi-source data for remote sensing change detection were collected and organized. The training set consisted of 985 pairs of high-resolution images from LEVIR-CD+ and building change images from WHU-CD, and was divided into training, validation, and test sets according to the official protocols of each dataset. All image pairs underwent uniform preprocessing. First, the original images were cropped or divided into 256×256 training blocks using a sliding window, ensuring that each training block contained at least some of the changed regions. Then, data augmentation operations, including random rotation, horizontal flipping, and vertical flipping, were applied to the training samples to enhance the model's robustness to viewpoint and geometric transformations, resulting in multi-temporal remote sensing image pairs.

[0021] S1, input the multi-temporal remote sensing image pairs into a weight-shared twin encoder, and extract multi-scale features through the twin encoder; In a more specific embodiment of the present invention, a multi-temporal remote sensing image refers to two registered optical remote sensing images of the same geographical area acquired at different times, denoted as It1 and It2, which have the same spatial resolution and band configuration. and Two temporal images and Simultaneously input into an encoder with the same structure and shared weights. In this process, the encoder generates multi-scale feature sets layer by layer through four downsampling operations. The encoder is based on the Visual Mamba architecture and contains four cascaded downsampling stages. Each stage consists of a downsampling layer (such as a 3×3 convolution with a stride of 2) and a Visual State Space module. The Visual State Space module realizes long-range dependency modeling based on the discretization of the continuous-time state space equation: the continuous form can be represented as Equation 1, and the discretized form adopts the matrix exponential form, as shown in Equation 2, thus obtaining the recursive form as shown in Equation 3. In order to improve the spatial locality modeling capability of two-dimensional images, the encoder introduces a cross-scan strategy in the state transfer, that is, injects two-dimensional neighborhood information into the state update by scanning in multiple directions (such as row scanning, column scanning or block local scanning), thereby maintaining the linear complexity of SSM while taking into account the local structure representation of the image.

[0022] S2, this invention constructs an end-to-end change detection network (Hybrid Attention MambaChange Detection, HAM-CD) consisting of a weight-sharing Siamese encoder and a decoder based on Hybrid Attention Mamba (HAM). Multi-scale features are input into the HAM-based decoder, which fuses and reconstructs the multi-scale features in stages. In each stage, a set of features to be fused is input into a HAM block to obtain local enhanced features, and the local enhanced features and the second encoded features are input into an interactive channel-space fusion block to obtain enhanced features. The enhanced features are then upsampled to output a third reconstructed feature. In a more specific embodiment of the invention, the hybrid attention Mamba-based decoder consists of four sequentially connected processing stages; For the j-th stage, j = 1, 2, 3, 4: Each stage includes a Hybrid Attention Mamba (HAM) block, an Interactive Channel-Spatial Fusion (ICSF) block, and an upsampling operation; The first reconstructed feature is the output of the (j-1)th processing stage. When j=1, the first reconstructed feature is the feature j=4 in the multi-scale feature pair after initialization processing. The second encoding feature is obtained by fusing the feature with index j in the multi-scale feature pair.

[0023] S3. Based on the third reconstruction feature, a change probability map is obtained. Based on the change probability map, a binary change map is output to complete the change detection.

[0024] Through the above steps, this application achieves the following: The twin encoder, through weight sharing and the cross-scanning mechanism of Visual Mamba, simultaneously ensures temporal consistency and two-dimensional spatial integrity during the extraction of multi-scale features, solving the problems of limited receptive field in CNNs and orientation deviation in Mamba; The hybrid attention Mamba decoder, through parallel modeling of TAM and TAB, maintains linear computational complexity while ensuring that local enhanced features retain fine-grained structural information such as road edges and building outlines, and also contain semantic associations across farmland patches and blocks, overcoming the single-granularity defect of single models; The interactive channel-spatial fusion block highlights change-related semantic channels (such as NDVI difference channels) through channel recalibration, and suppresses pseudo-change responses caused by cloud shadows and shadows through the spatial enhancement module, and effectively blocks the propagation of errors along the decoding path by combining residual connections, improving the detection robustness of weakly changing targets (such as newly planted seedlings and small construction sites); Finally, the change probability map driven by the third reconstruction feature is significantly enhanced in terms of spatial consistency and class discriminative power, making the binary change map outperform existing methods in core indicators such as Recall, Precision, and IoU.

[0025] In a more specific embodiment of the present invention, the specific process of inputting a set of features to be fused into a hybrid attention Mamba block for processing to obtain local enhanced features in S2 includes: The set of features to be fused is concatenated and converted into a dense vector through an embedding layer; the set of features to be fused refers to the first reconstructed feature from the previous stage and the second encoded feature from the encoder at the corresponding scale, which are received by the decoder in the j-th stage.

[0026] The dense vectors are input in parallel into a text-aware Mamba block and a transposed attention block for processing. The Textual-aware Mamba Block (TAM) is a state-space model module improved for vision tasks, while the Transposed Attention Block (TAB) is a module that performs self-attention computation in the channel dimension.

[0027] The outputs of the text-aware Mamba block and the transposed attention block are integrated through a concatenation operation and a convolutional fusion layer to generate local enhanced features.

[0028] The splicing operation refers to combining the TAM output features along the channel dimension. With TAB output features Perform concatenation to form a joint feature. This operation preserves complete information about the modeling advantages of each model without any weighting or filtering.

[0029] The TAM module, based on the principle of Textual-Aware State Fusion (TASF), incorporates local convolution operators and uses multi-scale dilated convolutions to expand the receptive field and enhance the multi-scale combination capability of neighborhood information during the recursive process of the state space. Its discretized state update can be formalized as shown in Equation 4, and it captures fine-grained spatial structure through a neighborhood weighting strategy as shown in Equation 5. Meanwhile, the TAB module implements a transposed attention mechanism in the channel dimension to efficiently model long-range dependencies. TAB generates query Q, key K, and value V through pointwise and depthwise convolutions, respectively, and uses dot product-softmax in the channel direction to calculate the attention map to balance expressive power and computational efficiency.

[0030] Through the above steps, this application achieves the following: within the hybrid attention Mamba block, the outputs of the text-aware Mamba block and the transposed attention block are coordinated and fused by channel concatenation and nonlinear convolution in a parallel architecture. This avoids gradient dilution caused by simple addition and achieves adaptive weighting of feature importance through learnable weights, thus ensuring the dual advantages of local enhanced features in terms of structural fidelity and semantic richness.

[0031] In another embodiment provided by the present invention, the specific process of inputting the local enhancement feature and the second coding feature into the Interactive Channel-Spatial Fusion (ICSF) block for processing to obtain the enhancement feature in step S2 is as follows: The local enhancement feature is used as input feature F1, and the corresponding second encoded feature is used as input feature F2; The input features F1 and F2 are concatenated and then projected through pointwise convolution, batch normalization, and ReLU activation. Channel attention weights are generated through the channel feature enhancement module and used to recalibrate the projected features; Spatial structural differences are captured through a spatial feature enhancement module; The recalibrated features and the features after capturing spatial structure differences are aggregated by addition, and the enhanced features are output through residual connection.

[0032] In a more specific embodiment of the present invention, the channel feature enhancement module specifically involves: applying global average pooling to generate channel descriptors, and generating the channel attention weights through two 1×1 convolutions and Sigmoid activation.

[0033] Specifically, ICSF receives input features F1 and F2, concatenates them along the channel dimension, and obtains the projected representation through 1×1 point convolution, batch normalization, and ReLU activation. Subsequently, a channel feature enhancement module is introduced. This module generates channel descriptors using global average pooling, and then generates channel attention weights through two 1×1 convolutions and sigmoid activation. ,right Channel recalibration is performed to highlight important semantic channels; in parallel, the spatial feature enhancement module uses 3×3 depthwise separable convolution combined with BN and ReLU to extract spatial structural difference features. Features after channel enhancement Spatial Enhancement Features The aggregate representation is obtained by adding elements one by one and projecting them using a 1×1 method. Finally, residual join is used to... Add and enhance features by activating the output. This fusion process ensures that cross-stage information can be selectively enhanced in the channel dimension while preserving detailed structure in the spatial dimension, thereby reducing the propagation of cascaded errors and improving the accuracy of the final segmentation boundary; In a more specific embodiment of the present invention, the text-aware Mamba block uses a structure-aware state fusion method to capture local spatial dependencies, wherein multi-scale dilated convolution is used to enhance the combination of neighborhood states, and the dilation rate of the multi-scale dilated convolution is set to 1, 3, and 5.

[0034] Based on the PyTorch framework, the HAM-CD network was trained end-to-end on GPU platforms such as the NVIDIA RTX 3090. A multi-scale supervision and composite loss mechanism was employed in the loss function, calculating supervision terms for the output of each stage of the decoder to enhance deep gradient propagation and thus alleviate the instability problem of deep training. The specific pixel-level loss was calculated using standard cross-entropy loss. The combined loss is shown in Equation 6, where cross-entropy is calculated according to pixel category, and Lovász-softmax is used to directly optimize the IoU metric to improve the small object detection performance under class imbalance. During training, a learning rate decay strategy and early stopping validation are used to prevent overfitting, and F1, IoU and other metrics are periodically evaluated on the validation set to select the optimal model weights.

[0035] In a more specific embodiment of the present invention, in step S3, obtaining the change probability map based on the third reconstruction feature specifically involves passing the third reconstruction feature finally output by the decoder through a prediction head to obtain the change probability map. Based on the change probability map, the output of the binary change map is specifically achieved by performing an argmax operation on the change probability map to generate the binary change map.

[0036] During the testing phase, the optimal model weights obtained during training are loaded, and a pair of full-resolution test images is input. The final change probability map is obtained through forward propagation, and then a binary change map is obtained by performing an argmax operation on the channel dimension. The network supports inputs of arbitrary size during inference and can process ultra-large images using a block sliding window approach to save GPU memory while ensuring detection integrity. Finally, the output results are compared with ground truth values, and evaluation metrics are calculated to verify the superiority of this invention in terms of accuracy and efficiency.

[0037] See Figure 3 The effects of the present invention will be further described below with reference to simulation experiments.

[0038] Simulation experimental conditions: The hardware platform for the simulation experiment of the invention: NVIDIA RTX 3090 GPU, paired with Intel(R) Core(TM) i7 series CPU and 32GB of memory.

[0039] Software platform: Win10 operating system, PyTorch 1.12.0.

[0040] The simulation experiments of this invention include two remote sensing change detection benchmark datasets: the LEVIR-CD+ dataset, which contains 985 pairs of high-resolution images with a resolution of 0.5 meters per pixel and a size of 1024×1024, focusing on building change detection; and the WHU-CD dataset, which covers urban building changes with an image resolution of 0.3 meters per pixel, and is divided into training and test sets according to the official protocol.

[0041] Simulation content and result analysis: The simulation experiments of this invention compare the proposed HAM-CD method with existing representative change detection methods, including CNN-based, Transformer-based, and SSM-based models, to comprehensively evaluate its performance.

[0042] To verify the simulation results of this invention, we used five evaluation metrics—recall, precision, overall accuracy (OA), F1-score, and intersection over union (IoU)—to observe the experimental performance of this invention and three existing technologies on test samples. Recall and precision reflect the model's ability to correctly identify changing pixels and the accuracy of predicting changing regions. OA reflects the proportion of all correctly classified pixels. The F1-score balances the model's precision and recall, comprehensively evaluating the model's overall performance in classifying positive and negative samples. IoU measures the overlap between the predicted ground truth and the ground truth. Higher values ​​for these five metrics indicate better detection and classification performance.

[0043] The comparison results of all objective evaluation indicators of the present invention and the existing advanced methods are shown in Table 1: Table 1 is the evaluation table of objective evaluation values ​​of the present invention and the comparison methods.

[0044] Table 1

[0045] As shown in Table 1, the five metrics of this invention—Recall, Precision, OA, F1-score, and IoU—are superior to existing methods on the LEVIR-CD dataset, and most of these metrics are also superior to existing methods on the WHU-CD dataset.

[0046] Figure 3 This is a visualization analysis diagram of the method proposed in this invention, which compares and shows the variable binary changes detected by different comparison methods on three benchmark datasets. Figure 3 As can be seen from the above, compared with other methods, the method proposed in this invention can detect more accurate binary change maps and has a higher degree of matching with the real change maps.

[0047] In one embodiment of the present invention, a remote sensing change detection system based on a hybrid attention Mamba architecture is provided, comprising: Feature extraction module: used to input multi-temporal remote sensing image pairs into a weight-shared twin encoder, and extract multi-scale features through the twin encoder; The fusion and reconstruction module is used to input multi-scale features into a hybrid attention Mamba-based decoder. The decoder fuses and reconstructs the multi-scale features in stages. In each stage, the first reconstructed features from the previous stage of the decoder and the second encoded features of the corresponding scale from the encoder are input into a hybrid attention Mamba block to obtain local enhanced features. The local enhanced features and the second encoded features are input into an interactive channel-space fusion block to obtain enhanced features. The enhanced features are upsampled to output a third reconstructed feature. Detection module: used to obtain a change probability map based on the third reconstructed feature, and output a binary change map based on the change probability map to complete change detection.

[0048] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or basic characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention.

[0049] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A remote sensing change detection method based on a hybrid attention Mamba architecture, characterized in that, Includes the following steps: S1, input the multi-temporal remote sensing image pairs into a weight-shared twin encoder, and extract multi-scale features through the twin encoder; S2, input the multi-scale features into the decoder based on hybrid attention Mamba, and fuse and reconstruct the multi-scale features in stages through the decoder. In each stage, the first reconstructed features from the previous stage of the decoder and the second encoded features of the corresponding scale from the encoder are input into a hybrid attention Mamba block to obtain local enhanced features. The local enhanced features and the second encoded features are input into an interactive channel-space fusion block to obtain enhanced features. The enhanced features are upsampled and the third reconstructed features are output. S3. Based on the third reconstruction feature, a change probability map is obtained. Based on the change probability map, a binary change map is output to complete the change detection.

2. The remote sensing change detection method based on a hybrid attention Mamba architecture according to claim 1, characterized in that, The multi-temporal remote sensing image pairs in S1 are and The twin encoder, based on the VisualMamba architecture, extracts multi-scale features through four stages of downsampling and the Visual State Space module. and .

3. The remote sensing change detection method based on a hybrid attention Mamba architecture according to claim 1, characterized in that, In S2, the decoder based on hybrid attention Mamba consists of four sequentially connected processing stages; For the j-th stage, j = 1, 2, 3, 4: The first reconstructed feature is the output of the (j-1)th processing stage. When j=1, the first reconstructed feature is the feature j=4 in the multi-scale feature pair after initialization processing. The second encoding feature is obtained by fusing the feature with index j in the multi-scale feature pair.

4. The remote sensing change detection method based on a hybrid attention Mamba architecture according to claim 3, characterized in that, The specific process in S2 of inputting a set of features to be fused into a hybrid attention Mamba block to obtain locally enhanced features includes: The set of features to be fused are concatenated and converted into a dense vector through an embedding layer; The dense vectors are input in parallel into a text-aware Mamba block and a transposed attention block for processing. The outputs of the text-aware Mamba block and the transposed attention block are integrated through a concatenation operation and a convolutional fusion layer to generate local enhanced features.

5. The remote sensing change detection method based on a hybrid attention Mamba architecture according to claim 1, characterized in that, The specific process in S2 where the local enhancement features and the second coding features are input into the interactive channel-spatial fusion block for processing to obtain the enhancement features is as follows: The local enhancement feature is used as input feature F1, and the corresponding second encoded feature is used as input feature F2; The input features F1 and F2 are concatenated and then projected through pointwise convolution, batch normalization, and ReLU activation. Channel attention weights are generated through the channel feature enhancement module and used to recalibrate the projected features; Spatial structural differences are captured through a spatial feature enhancement module; The recalibrated features and the features after capturing spatial structure differences are aggregated by addition, and the enhanced features are output through residual connection.

6. The remote sensing change detection method based on a hybrid attention Mamba architecture according to claim 4, characterized in that, The text-aware Mamba block uses a structure-aware state fusion method to capture local spatial dependencies, wherein multi-scale dilated convolution is used to enhance the combination of neighborhood states, and the dilation rate of the multi-scale dilated convolution is set to 1, 3, and 5.

7. The remote sensing change detection method based on a hybrid attention Mamba architecture according to claim 5, characterized in that, The channel feature enhancement module specifically involves: applying global average pooling to generate channel descriptors, and generating the channel attention weights through two 1×1 convolutions and Sigmoid activation.

8. The remote sensing change detection method based on a hybrid attention Mamba architecture according to claim 5, characterized in that, The spatial feature enhancement module specifically uses 3×3 depthwise convolution, batch normalization, and ReLU activation.

9. The remote sensing change detection method based on a hybrid attention Mamba architecture according to claim 1, characterized in that, In S3, obtaining the change probability map based on the third reconstruction feature specifically involves passing the third reconstruction feature, which is the final output of the decoder, through a prediction head to obtain the change probability map. Based on the change probability map, the output of the binary change map is specifically achieved by performing an argmax operation on the change probability map to generate the binary change map.

10. A remote sensing change detection system based on a hybrid attention Mamba architecture, characterized in that, include: Feature extraction module: used to input multi-temporal remote sensing image pairs into a weight-shared twin encoder, and extract multi-scale features through the twin encoder; The fusion and reconstruction module is used to input multi-scale features into a hybrid attention Mamba-based decoder. The decoder fuses and reconstructs the multi-scale features in stages. In each stage, the first reconstructed features from the previous stage of the decoder and the second encoded features of the corresponding scale from the encoder are input into a hybrid attention Mamba block to obtain local enhanced features. The local enhanced features and the second encoded features are input into an interactive channel-space fusion block to obtain enhanced features. The enhanced features are upsampled to output a third reconstructed feature. Detection module: used to obtain a change probability map based on the third reconstructed feature, and output a binary change map based on the change probability map to complete change detection.

Images