A Remote Sensing Image Segmentation Method and System Based on Convolution-State Space Fusion and Position Triggers

Abstract

This invention discloses a remote sensing image segmentation method and system based on convolutional-state-space fusion and location triggers, comprising: preprocessing the original remote sensing image, patch embedding and spatial location encoding to convert large-size images into a structured representation of uniform scale while preserving spatial location information; decoupling the location-encoded patch vector sequence through the channels of multiple convolutional-state-space fusion units, extracting fine-grained edge and local structural features using convolutional branches, capturing long-range contextual information using state-space modeling branches, and then performing multi-scale feature fusion through a cross-modal attention mechanism, and introducing location triggers multiple times to dynamically strengthen the response regions of edges and small targets, outputting fused features; mapping to a classification probability map through adaptive pooling and fully connected layers, and constructing a supervised loss function to improve segmentation performance; this invention significantly improves the segmentation accuracy of small targets and boundaries while balancing computational efficiency and robustness.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and more specifically to a remote sensing image segmentation method and system based on convolution-state space fusion and position triggers. Background Technology

[0002] With the rapid development of remote sensing imaging technology and Geographic Information Systems (GIS), remote sensing images are increasingly widely used in fields such as urban planning, agricultural monitoring, disaster assessment, and land cover change analysis. Among these applications, remote sensing image segmentation, as a key step in achieving intelligent identification of ground features and regional semantic extraction, has become one of the core tasks in intelligent remote sensing analysis. Its goal is to assign semantic category labels to each pixel in high-resolution remote sensing images, thereby achieving accurate extraction of ground features.

[0003] However, due to the characteristics of remote sensing images, such as high resolution, strong texture, multi-source and multimodal (e.g., multispectral, panchromatic, SAR, etc.), and complex scenes, traditional image segmentation methods face many challenges in this field, such as blurred boundaries, easy neglect of small targets, and difficulty in fusing multi-source information.

[0004] In recent years, deep learning technologies, especially convolutional neural networks (CNNs) and Transformers, have made significant progress in image segmentation tasks. CNNs, with their powerful local perception capabilities and parameter sharing mechanisms, have become the mainstream model architecture for remote sensing image segmentation. However, due to their limited receptive field, they are unable to fully capture long-distance dependencies, resulting in certain limitations when dealing with large-scale targets, non-homogeneous objects, and complex spatial structures. On the other hand, the Transformer architecture can model global contextual information through its self-attention mechanism and has been gradually introduced into remote sensing image analysis in recent years. However, it has high computational costs, requires a large amount of data, and has insufficient expressive power when dealing with local details (such as small targets or edges).

[0005] In existing technologies, researchers have attempted to combine multiple mechanisms to improve the segmentation accuracy of remote sensing images, such as multi-scale feature fusion, attention mechanisms, pyramid structures, and residual connections.

[0006] However, these schemes still have the following shortcomings: (1) The feature extraction stage is not sensitive enough to local boundaries and small targets, which can easily lead to blurred boundaries and missed detection of small targets; (2) Multimodal fusion, such as the integration of multispectral and SAR, has information redundancy or semantic inconsistency problems; (3) Global modeling and local details conflict, and there is a lack of an efficient fusion mechanism to unify the modeling length dependence.

[0007] Therefore, how to solve the problems of boundary blurring, small target omission, and difficulty in multi-source information fusion in current remote sensing image segmentation while maintaining model computational efficiency and robustness is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of this, the present invention provides a remote sensing image segmentation method and system based on convolution-state space fusion and position triggering to solve some of the technical problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A remote sensing image segmentation method based on convolution-state space fusion and position triggers includes the following steps:

[0011] S1. The original remote sensing images are preprocessed, and large-size images are converted into a structured representation of a uniform scale through patch embedding and spatial location coding, while retaining spatial location information;

[0012] S2. The position-encoded patch vector sequence is decoupled through the channels of multiple convolutional-state-space fusion units (C-SSFU). Fine-grained edge and local structural features are extracted using the convolutional branch, and long-range context information is captured using the state-space modeling branch. The features from the two branches are fused into multi-scale features through a cross-modal attention mechanism. Position triggers are introduced multiple times to dynamically enhance the response regions of edges and small targets, and the final fused features are output.

[0013] S3. By using adaptive pooling and fully connected layers, the fused features are mapped to the final pixel-level classification probability map, which is then compared with the real label map. A supervised loss function is constructed to improve the model's segmentation performance for minority classes and boundary regions, thus obtaining the final segmentation result.

[0014] Preferably, the preprocessing includes: geometric correction and registration, atmospheric correction and denoising, and spectral normalization or standardization.

[0015] Preferably, the specific content of patch embedding and spatial location encoding is as follows:

[0016] Divide the preprocessed image into non-overlapping patches;

[0017] Each small block is flattened into a one-dimensional vector, and then mapped to a D-dimensional feature space through a linear mapping layer;

[0018] Location encoding is added based on the spatial location information of each small block in the original image.

[0019] Preferably, the specific architecture of multiple convolutional-state-space fusion units (C-SSFUs) and multiple introduction of position triggers is as follows: the patch vector sequence with position encoding is sequentially input into two concatenated convolutional-state-space fusion units (C-SSFUs) in the first stage, and position triggers are introduced. After the first position trigger, two C-SSFUs, position triggers, four C-SSFUs, position triggers, and two more C-SSFUs are sequentially concatenated.

[0020] Preferably, the specific content of the Convolutional-State Space Fusion Unit (C-SSFU) is as follows:

[0021] S21. Input features Feature decoupling is performed using a channel grouping strategy to obtain the first feature. Second feature

[0022] S22. Convolutional Feature Enhancement Branch: First Feature Deep feature extraction is performed, using batch normalization layers, 3x3 convolutional layers, batch normalization layers, ReLU activation functions, 3x3 convolutional layers, batch normalization layers, ReLU activation functions, pointwise convolutional layers, and ReLU activation functions to obtain local geometric priors.

[0023] S23. State-space modeling branch: Second feature Long-range context modeling is performed, and features are obtained through layer normalization, linear layers, 3*3 depthwise separable convolutional layers, SiLU activation function, two-dimensional selective scan blocks, and layer normalization. At the same time, the second feature After passing through a linear layer and an activation function, we obtain... and After matrix multiplication, the global semantics are obtained through layer normalization.

[0024] S24. Multispectral collaborative modeling of bi-branch features is performed through cross-modal attention fusion:

[0025] Will and By splicing the pieces together, we can obtain the splicing features. Bimodal Interaction: Leveraging Local Geometric Priors of Convolutional Branches Global semantics for guiding state space branches Q is obtained after a linear layer. K is obtained after a linear layer; spectrally sensitive attention: Q is multiplied by K and then passed through the SiLU activation function, and... Multiplication yields characteristics

[0026] S25. Residual Feature Refinement: Constructing pathways to enhance ground feature characteristics. After the channel shuffle, it is compared with the initial input. Perform residual connection and output. This refers to the output features of the Convolutional-State Space Fusion Unit (C-SSFU).

[0027] Preferably, the specific details of introducing position triggers to dynamically enhance the response areas of edges and small targets are as follows:

[0028] Two-dimensional coordinate channels are added to the input feature map through coordinate convolutional layers to enhance spatial sensitivity; the GELU activation function introduces non-linearity to improve expressive power; conventional convolutional layers extract spatial context; a multi-head mechanism is adopted, using the features extracted from the convolution in the current position trigger as the query, and the global semantic feature map of the previous stage or higher layer as the key and value, and calculating attention weights through dot product to adaptively amplify the response of the edge and small target regions, while suppressing irrelevant background features. Finally, the weighted semantic information is fed back to the current position feature to achieve cross-scale information interaction and enhancement; finally, a position trigger mask is generated by convolutional layers and the Sigmoid architecture activation, which is used to perform pixel-level weighting of the feature map.

[0029] Preferably, the specific content of step S3 is as follows:

[0030] S31. Input the obtained fused feature map into the adaptive global pooling layer, perform average pooling in the spatial dimension, retain channel dimension information, and compress spatial redundancy;

[0031] S32. The pooled feature vectors are linearly mapped through a fully connected layer, and then the final segmentation probability map is obtained through an activation function, which is the probability value of each pixel belonging to each category.

[0032] S33. Compare the obtained segmentation probability map with the true label map provided when inputting into the neural network, and construct a joint loss function based on cross-entropy loss and Dice loss.

[0033] A remote sensing image segmentation system based on convolution-state space fusion and position triggers, based on the aforementioned remote sensing image segmentation method based on convolution-state space fusion and position triggers, includes a feature preprocessing module, a patch embedding and spatial location encoding module, a feature extraction and fusion module, and an output module;

[0034] The feature preprocessing module is used to preprocess the original remote sensing images;

[0035] The Patch embedding and spatial location coding module is used to convert large-size images into a structured representation of a uniform scale by using Patch embedding and spatial location coding on preprocessed remote sensing images, while preserving spatial location information.

[0036] The feature extraction and fusion module is used to decouple the position-encoded patch vector sequence through multiple channels of the Convolutional-State-Space Fusion Unit (C-SSFU). It uses the convolutional branch to extract fine-grained edge and local structural features, and the state-space modeling branch to capture long-range contextual information. The features of the two branches are fused into multi-scale features through a cross-modal attention mechanism, and position triggers are introduced multiple times to dynamically strengthen the response regions of edges and small targets, and the final fused features are output.

[0037] The output module is used to map the fused features into a final pixel-level classification probability map through adaptive pooling and fully connected layers, compare it with the real label map, construct a supervised loss function, improve the model's segmentation performance for minority classes and boundary regions, and obtain the final segmentation result.

[0038] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned remote sensing image segmentation method based on convolution-state space fusion and position triggers.

[0039] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the remote sensing image segmentation method based on convolution-state space fusion and position trigger.

[0040] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a remote sensing image segmentation method and system based on convolution-state space fusion and position triggering. While maintaining the model's computational efficiency, it effectively solves the problems of blurred boundaries, omission of small targets, and difficulty in multi-source information fusion that are common in current remote sensing image segmentation through the collaborative design of local convolution, state space modeling, and position-sensitive control. It has good application prospects and promotion value.

[0041] This invention significantly improves the segmentation accuracy of small targets and boundaries: by dynamically amplifying the response of edges and small targets through the position trigger module, and introducing residual detail reconstruction during the upsampling process, the class imbalance and boundary blurring problems are effectively alleviated; the convolution-state space fusion unit C-SSFU takes into account both local geometric details and long-range dependencies in the same module, making the model more sensitive to the segmentation of small features;

[0042] This invention balances computational efficiency and robustness: the design based on patch embedding and intensity grouping significantly reduces the complexity of global attention and convolution operations, making it suitable for high-resolution scenarios; multi-source preprocessing and progressive residual structure enhance the network's tolerance to interference such as atmospheric scattering, speckle noise, and cloud and fog occlusion. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0044] Figure 1 A schematic diagram of the remote sensing image segmentation method based on convolution-state space fusion and position trigger provided by the present invention;

[0045] Figure 2 This is a schematic diagram of the Convolution-State Space Fusion Unit (C-SSFU) provided by the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] This invention discloses a remote sensing image segmentation method based on convolution-state space fusion and position triggers, such as... Figure 1 This includes the following steps:

[0048] S1. The original remote sensing images are preprocessed, and large-size images are converted into a structured representation of a uniform scale through patch embedding and spatial location coding, while retaining spatial location information;

[0049] S2. The position-encoded patch vector sequence is decoupled through the channels of multiple convolutional-state-space fusion units (C-SSFU). Fine-grained edge and local structural features are extracted using the convolutional branch, and long-range context information is captured using the state-space modeling branch. The features from the two branches are fused into multi-scale features through a cross-modal attention mechanism. Position triggers are introduced multiple times to dynamically enhance the response regions of edges and small targets, and the final fused features are output.

[0050] S3. By using adaptive pooling and fully connected layers, the fused features are mapped to the final pixel-level classification probability map, which is then compared with the real label map. A supervised loss function is constructed to improve the model's segmentation performance for minority classes and boundary regions, thus obtaining the final segmentation result.

[0051] To further implement the above technical solutions, the preprocessing includes: geometric correction and registration, atmospheric correction and denoising, and spectral normalization or standardization.

[0052] In this embodiment, geometric correction and registration are performed: using the positioning information of the remote sensing platform and ground control points (GCPs), geometric correction and registration are performed on multi-temporal images / multi-source images to ensure a one-to-one correspondence between image pixels.

[0053] Atmospheric correction and denoising: Perform atmospheric correction (e.g., 6S or DOS methods) on multispectral data to eliminate the effects of atmospheric scattering; apply Lee filters to synthetic aperture radar (SAR) data for speckle noise suppression;

[0054] Spectral normalization / standardization: Normalize the spectral values ​​of each band to [0,1] or a standard normal distribution to reduce the impact of spectral differences under different sensors and acquisition conditions on model training.

[0055] To further implement the above technical solution, the specific details of patch embedding and spatial location encoding are as follows:

[0056] To balance the global context and local texture features of high-resolution remote sensing images, the preprocessed images... Division There are P×P non-overlapping patches, where H and W are the spatial dimensions of the remote sensing image, i.e., height and width.

[0057]

[0058] Each small block x i Flattened into a one-dimensional vector It is then mapped to the D-dimensional feature space via a linear mapping layer (fully connected layer);

[0059]

[0060] Based on the spatial location information of each small patch in the original image, positional encoding is added.

[0061] In this embodiment, learnable parameters or sine / cosine functions can be used:

[0062]

[0063] Patch embedding and spatial location encoding help to: reduce computational complexity, as patching effectively reduces the computational burden of directly performing global convolution or attention on high-resolution images; and preserve local texture information, with each patch carrying rich local texture, boundary, and detail features of ground features, providing a foundation for subsequent segmentation.

[0064] To further implement the above technical solution, the specific architecture of multiple convolutional-state-space fusion units (C-SSFU) and multiple introduction of position triggers is as follows: the patch vector sequence with position encoding is sequentially input into the two concatenated convolutional-state-space fusion units (C-SSFU) in the first stage, and position triggers are introduced. After the first position trigger, two C-SSFUs, position triggers, four C-SSFUs, position triggers and two C-SSFUs are sequentially concatenated to output the final fused feature.

[0065] Each C-SSFU combines local convolutions to capture detailed boundaries with global state space mechanisms to model long-range dependencies, achieving multi-scale feature fusion and generating feature representations.

[0066] To further implement the above technical solutions, such as Figure 2 The specific content of the Convolutional-State Space Fusion Unit (C-SSFU) is as follows:

[0067] S21. Input features Feature decoupling is performed using a channel grouping strategy to obtain the first feature. Second feature

[0068] In this embodiment, H and W are the spatial dimensions of the remote sensing image, C is the number of channels, and the segmentation ratio is set to 1:1 to maintain computational symmetry.

[0069] S22. Convolutional Feature Enhancement Branch: First Feature Deep feature extraction is performed, using batch normalization layers, 3x3 convolutional layers, batch normalization layers, ReLU activation functions, 3x3 convolutional layers, batch normalization layers, ReLU activation functions, pointwise convolutional layers, and ReLU activation functions to obtain local geometric priors.

[0070] In this embodiment, a double 3×3 convolution is used to form an increasing receptive field structure, which is combined with batch normalization (BN) and ReLU activation function to enhance the ability to characterize the edges of ground features;

[0071] S23. State-space modeling branch: Second feature Long-range context modeling is performed, and features are obtained through layer normalization, linear layers, 3*3 depthwise separable convolutional layers, SiLU activation function, two-dimensional selective scan blocks, and layer normalization. At the same time, the second feature After passing through a linear layer and an activation function, we obtain... and After matrix multiplication, the global semantics are obtained through layer normalization.

[0072] S24. Multispectral collaborative modeling of bi-branch features is performed through cross-modal attention fusion:

[0073] Will and By splicing them together, we get Bimodal Interaction: Leveraging Local Geometric Priors of Convolutional Branches Global semantics for guiding state space branches Q is obtained after a linear layer. K is obtained after a linear layer; spectrally sensitive attention: Q is multiplied by K and then passed through the SiLU activation function, and... Multiplication yields characteristics

[0074] S25. Construct a feature enhancement pathway to refine residual features. After the channel shuffle, it is compared with the initial input. Perform residual connection and output. That is, the output features of the Convolutional-State-Space Fusion Unit (C-SSFU);

[0075] In this embodiment, the ground feature enhancement pathway includes: a channel shuffling strategy, which performs periodic permutation (permutation period T=4) on the stitched 2C channel to promote cross-dimensional interaction between multispectral bands and spatial features; progressive residuals, which use random depth dropout to improve the robustness of the model to noise such as cloud and fog obstruction; and identity mapping, where short-circuit connections of the original features preserve low-frequency information of continuous areas such as farmland and water bodies.

[0076] To further implement the above technical solution, and addressing the common problems of edge misclassification and small target loss in remote sensing segmentation, a location trigger is introduced to dynamically enhance the response regions of edges and small targets. The specific details are as follows:

[0077] Two-dimensional coordinate channels are added to the input feature map through coordinate convolutional layers to enhance spatial sensitivity; the GELU activation function introduces non-linearity to improve expressive power; conventional convolutional layers extract spatial context; a multi-head mechanism is adopted, using the features extracted from the convolution in the current position trigger as the query, and the global semantic feature map from the previous stage or higher as the key and value. Attention weights are calculated through dot product to adaptively amplify the response in the regions where edges and small objects are located, while suppressing irrelevant background features. Finally, the weighted semantic information is fed back to the current position feature, realizing cross-scale information interaction and enhancement; finally, a position trigger mask M is generated through convolutional layers and a sigmoid architecture activation. pos ∈[0,1] H′×W′×1 It is used to perform pixel-level weighting on feature maps.

[0078] To further implement the above technical solution, the specific content of step S3 is as follows:

[0079] S31. Obtain the fused feature map The input to the adaptive global pooling layer is averaged in the spatial dimension, preserving channel dimension information and compressing spatial redundancy to adapt to changes in image feature scale for different input sizes;

[0080] S32. The pooled feature vectors are linearly mapped through a fully connected layer, and then the final segmentation probability map P is obtained through an activation function (such as Sigmoid or Softmax), which is the probability value of each pixel belonging to each category.

[0081] S33. Compare the obtained segmentation probability map P with the true label map Y provided when inputting into the neural network, and construct a joint loss function based on cross-entropy loss and Dice loss. To improve the model's sensitivity to small targets and boundary regions;

[0082] In this embodiment, the network output probability graph is P = {P} c，i，j The true label image is Y = {Y}. c，i，j}, Y c，i，j ∈{0,1}, where c=1,...,K, K represents the number of categories, (i,j) are pixel coordinates, and the joint loss function is:

[0083]

[0084] Where λ1 and λ2 are the weight coefficients of the two sub-losses, respectively. For pixel-level cross-entropy loss, The Dice coefficient loss is used to mitigate class imbalance.

[0085] Combining pixel-level cross-entropy loss and Dice loss:

[0086] Cross-entropy loss:

[0087]

[0088] Dice architecture loss:

[0089]

[0090] Where ε is a small constant to prevent division by zero.

[0091] A remote sensing image segmentation system based on convolution-state space fusion and position triggers, which is based on a remote sensing image segmentation method based on convolution-state space fusion and position triggers, includes a feature preprocessing module, a patch embedding and spatial location encoding module, a feature extraction and fusion module, and an output module;

[0092] The feature preprocessing module is used to preprocess the original remote sensing images;

[0093] The Patch embedding and spatial location coding module is used to convert large-size images into a structured representation of a uniform scale by using Patch embedding and spatial location coding on preprocessed remote sensing images, while preserving spatial location information.

[0094] The feature extraction and fusion module is used to decouple the position-encoded patch vector sequence through multiple channels of the Convolutional-State-Space Fusion Unit (C-SSFU). It uses the convolutional branch to extract fine-grained edge and local structural features, and the state-space modeling branch to capture long-range contextual information. The features of the two branches are fused into multi-scale features through a cross-modal attention mechanism, and position triggers are introduced multiple times to dynamically strengthen the response regions of edges and small targets, and the final fused features are output.

[0095] The output module is used to map the fused features into a final pixel-level classification probability map through adaptive pooling and fully connected layers, compare it with the real label map, construct a supervised loss function, improve the model's segmentation performance for minority classes and boundary regions, and obtain the final segmentation result.

[0096] This invention introduces an end-to-end feature preprocessing module in the input stage to perform geometric correction, atmospheric correction, denoising, and spectral normalization on remote sensing images, effectively improving the usability and consistency of the original images. Subsequently, through patch embedding and spatial location encoding mechanisms, large-size images are converted into a structured representation of a uniform scale while retaining their spatial location information, reducing the difficulty of subsequent model processing.

[0097] In the feature extraction and fusion stage, a novel convolutional-state-space fusion unit (C-SSFU) was constructed. This module uses a channel decoupling design to extract fine-grained edge and local structural features using a convolutional branch and capture long-range contextual information using a state-space modeling branch. The features from the two branches are then fused in a multi-spectral manner through a cross-modal attention mechanism to achieve semantic enhancement of ground features. In addition, "position triggers" are introduced multiple times in the backbone structure to dynamically strengthen the response regions of edges and small targets through mechanisms such as coordinate convolution and multi-head cross-attention, thereby solving the problems of boundary blurring and small target loss.

[0098] In the output stage, adaptive pooling and fully connected layers are introduced to map the fused features into the final pixel-level classification probability map, and the joint loss function of cross-entropy and Dice is adopted to improve the model's segmentation performance for minority classes and boundary regions.

[0099] The overall architecture of this invention takes into account both global context modeling capabilities and enhanced perception of ground feature boundaries and details, making it suitable for high-precision semantic segmentation tasks of multi-source remote sensing images.

[0100] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a remote sensing image segmentation method based on convolution-state space fusion and position triggers.

[0101] A processing terminal includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements a remote sensing image segmentation method based on convolution-state space fusion and position triggers.

[0102] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the architectural differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0103] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image segmentation method based on convolution-state space fusion and position triggers, characterized in that, Includes the following steps: S1. The original remote sensing images are preprocessed, and large-size images are converted into a structured representation of a uniform scale through patch embedding and spatial location coding, while retaining spatial location information; S2. The position-encoded patch vector sequence is decoupled through the channels of multiple convolutional-state-space fusion units (C-SSFU). Fine-grained edge and local structural features are extracted using the convolutional branch, and long-range context information is captured using the state-space modeling branch. The features from the two branches are fused into multi-scale features through a cross-modal attention mechanism. Position triggers are introduced multiple times to dynamically enhance the response regions of edges and small targets, and the final fused features are output. S3. By using adaptive pooling and fully connected layers, the fused features are mapped to the final pixel-level classification probability map, which is then compared with the real label map. A supervised loss function is constructed to improve the model's segmentation performance for minority classes and boundary regions, thus obtaining the final segmentation result. The specific content of the Convolutional-State-Space Fusion Unit (C-SSFU) is as follows: S21. Input features A channel grouping strategy is used to decouple features and obtain the first feature. Second feature ; S22. Convolutional Feature Enhancement Branch: First characteristic Deep feature extraction is performed, followed by batch normalization layers and 3... 3 convolutional layers, batch normalization layers, ReLU activation function, 3 3 convolutional layers, batch normalization layers, ReLU activation function, pointwise convolutional layers, and ReLU activation function to obtain local geometric priors. ; S23. State-space modeling branch: Second feature Long-range context modeling was performed, including layer normalization, linear layers, and 3D modeling.

3. Depthwise separable convolutional layers, SiLU activation function, two-dimensional selective scan blocks, and layer normalization to obtain features. At the same time, the second feature After passing through a linear layer and an activation function, we obtain... , and After matrix multiplication, the global semantics are obtained through layer normalization. ; S24. Multispectral collaborative modeling of bi-branch features is performed through cross-modal attention fusion: Will and By splicing the pieces together, we can obtain the splicing features. Bimodal interaction: leveraging the local geometric priors of convolutional branches Global semantics guiding state space branches , Q is obtained after a linear layer. K is obtained after a linear layer; spectrally sensitive attention: Q is multiplied by K and then passed through the SiLU activation function, and... Multiplication yields characteristics ; S25. Residual Feature Refinement: Constructing pathways to enhance ground feature characteristics. After the channel shuffle, it is compared with the initial input. Perform residual connection and output. This refers to the output feature of the Convolutional-State-Space Fusion Unit (C-SSFU).

2. The remote sensing image segmentation method based on convolution-state space fusion and position triggers according to claim 1, characterized in that, The preprocessing includes geometric correction and registration, atmospheric correction and denoising, and spectral normalization or standardization.

3. The remote sensing image segmentation method based on convolution-state space fusion and position triggers according to claim 1, characterized in that, The specific details of patch embedding and spatial location encoding are as follows: Divide the preprocessed image into non-overlapping patches; Each small block is flattened into a one-dimensional vector, and then mapped to a D-dimensional feature space through a linear mapping layer; Location encoding is added based on the spatial location information of each small block in the original image.

4. The remote sensing image segmentation method based on convolution-state space fusion and position triggers according to claim 1, characterized in that, The specific architecture of multiple convolutional-state-space fusion units (C-SSFUs) and multiple introduction of position triggers is as follows: The patch vector sequence with position encoding is sequentially input into two concatenated convolutional-state-space fusion units (C-SSFUs) in the first stage, and position triggers are introduced. After the first position trigger, two C-SSFUs, position triggers, four C-SSFUs, position triggers, and two more C-SSFUs are sequentially concatenated.

5. The remote sensing image segmentation method based on convolution-state space fusion and position triggers according to claim 1, characterized in that, The specific details of introducing position triggers to dynamically enhance the response areas of edges and small targets are as follows: Two-dimensional coordinate channels are added to the input feature map through coordinate convolutional layers to enhance spatial sensitivity; the GELU activation function introduces non-linearity to improve expressive power; and conventional convolutional layers extract spatial context. A multi-head mechanism is adopted, using the features extracted from the convolution in the current position trigger as the query and the global semantic feature map of the previous stage or higher layer as the key and value. Attention weights are calculated by dot product, which adaptively amplifies the response of the edge and small target regions, while suppressing irrelevant background features. Finally, the weighted semantic information is fed back to the current position feature, realizing cross-scale information interaction and enhancement. Finally, a position trigger mask is generated by activation through convolutional layers and sigmoid architecture, which is used to perform pixel-level weighting on the feature map.

6. The remote sensing image segmentation method based on convolution-state space fusion and position triggers according to claim 1, characterized in that, The specific content of step S3 is as follows: S31. Input the obtained fused feature map into the adaptive global pooling layer, perform average pooling in the spatial dimension, retain channel dimension information, and compress spatial redundancy; S32. The pooled feature vectors are linearly mapped through a fully connected layer, and then the final segmentation probability map is obtained through an activation function, which is the probability value of each pixel belonging to each category. S33. Compare the obtained segmentation probability map with the true label map provided when inputting into the neural network, and construct a joint loss function based on cross-entropy loss and Dice loss. .

7. A remote sensing image segmentation system based on convolution-state space fusion and position triggers, characterized in that, A remote sensing image segmentation method based on convolution-state space fusion and position triggering according to any one of claims 1-6 includes a feature preprocessing module, a patch embedding and spatial location encoding module, a feature extraction and fusion module, and an output module. The feature preprocessing module is used to preprocess the original remote sensing images; The Patch embedding and spatial location coding module is used to convert large-size images into a structured representation of a uniform scale by using Patch embedding and spatial location coding on preprocessed remote sensing images, while preserving spatial location information. The feature extraction and fusion module is used to decouple the position-encoded patch vector sequence through multiple channels of the Convolutional-State-Space Fusion Unit (C-SSFU). It uses the convolutional branch to extract fine-grained edge and local structural features, and the state-space modeling branch to capture long-range contextual information. The features of the two branches are fused into multi-scale features through a cross-modal attention mechanism, and position triggers are introduced multiple times to dynamically strengthen the response regions of edges and small targets, and the final fused features are output. The output module is used to map the fused features into a final pixel-level classification probability map through adaptive pooling and fully connected layers, compare it with the real label map, construct a supervised loss function, improve the model's segmentation performance for minority classes and boundary regions, and obtain the final segmentation result.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the remote sensing image segmentation method based on convolution-state space fusion and position trigger as described in any one of claims 1-6.

9. A processing terminal, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements a remote sensing image segmentation method based on convolution-state space fusion and position trigger as described in any one of claims 1-6.

Images