Remote sensing image segmentation method based on state space dual transformation and channel mixing

Abstract

The application discloses a remote sensing image segmentation method based on state space dual transformation and channel mixing. Firstly, dimension reduction and feature extraction are performed on the input image to generate reduced dimension features, which are further compressed to represent hidden states. These features are mapped into the hidden state space by using non-causal state space dual transformation, and the information interaction is optimized by channel mixing to generate hidden state features. Secondly, the reduced dimension features and the hidden state features are respectively reduced in dimension, then mixed in channels, and added together to fuse global information. Channel displacement is applied to enhance the feature sharing capability between different modalities, and after re-channel mixing and local spatial information extraction, the final enhanced features are generated. Finally, based on the enhanced features, the segmentation task is performed to output high-precision segmentation results. The application optimizes the calculation efficiency and feature fusion, and improves the precision and robustness of remote sensing image segmentation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, and particularly relates to a remote sensing image segmentation method based on state space dual transformation and channel mixing. BACKGROUND

[0002] Remote sensing image segmentation is a core task in many fields such as geographic information systems, environmental monitoring, and agricultural analysis. Traditional remote sensing image segmentation methods mainly rely on rule-based threshold segmentation, region growing, edge detection, and other means. These methods can achieve good segmentation results under certain conditions, but have limitations in complex background, multi-scale feature, and high-resolution data processing.

[0003] In recent years, the rapid development of deep learning technology, especially convolutional neural networks (CNN) and visual Transformers (ViT) and other models, has provided a new technical path for remote sensing image segmentation. However, existing deep learning methods still face problems such as high computational complexity, limited global information interaction, and insufficient feature expression capability, especially when processing large-scale remote sensing data, the demand for computing resources is huge, making it difficult to apply to resource-constrained edge devices or real-time processing tasks.

[0004] Therefore, how to provide a remote sensing image segmentation method that can improve segmentation accuracy while reducing computational overhead is a technical problem that needs to be solved by those skilled in the art. SUMMARY

[0005] The present application provides a remote sensing image segmentation method based on state space dual transformation and channel mixing, which combines state space modeling and hidden state mixing mechanism to optimize feature extraction and feature interaction process, making information transmission more efficient. On the other hand, it introduces feature channel mixing and displacement enhancement mechanism to improve cross-modal information fusion capability, making the segmentation result more accurate and robust.

[0006] The remote sensing image segmentation method based on state space dual transformation and channel mixing provided by the present application includes the following steps:

[0007] S1: Perform dimensionality reduction and feature extraction on the remote sensing input image through a series of convolution and linear transformation operations to obtain reduced features, compress the reduced features to generate hidden states, map the reduced features to the hidden state space using non-causal state space dual transformation, and perform channel mixing in the hidden state space to generate hidden state features;

[0008] S2: respectively dimension reduction on the dimension reduction feature and the hidden state feature, then channel mixing; add the dimension reduction feature and the hidden state feature to fuse the global information of both; after channel displacement of the global information, respectively channel displacement with the dimension reduction feature and the hidden state feature after channel mixing, and then extract local spatial information after channel mixing of the dimension reduction feature and the hidden state feature after channel displacement, and fuse to generate the final enhanced feature;

[0009] S3: based on the final enhanced feature, perform a segmentation task, and output a segmentation result.

[0010] Preferably, the step of dimension reduction on the input image in S1 comprises:

[0011] Receiving a remote sensing input image, generating a dimension-reduced feature map through four consecutive 3x3 convolution layers for downsampling processing.

[0012] Preferably, before the feature extraction on the remote sensing input image in S1, the following steps are further included:

[0013] Performing depth separable convolution processing on the dimension-reduced feature map, combining with the dimension-reduced feature map through residual connection, and obtaining an enhanced feature representation through normalization operation; performing feature extraction on the enhanced feature representation.

[0014] Preferably, the step of compressing the dimension reduction feature to generate a hidden state in S1 comprises:

[0015] Generating a projection matrix and a time step parameter through linear transformation of the dimension-reduced feature map, obtaining a discrete state through exponential transformation based on the projection matrix and the time step parameter, and generating a hidden state h using Hadamard multiplication combined with the convolution operation result of the projection matrix - .

[0016] Preferably, the step of compressing the dimension reduction feature to generate a hidden state in S1 comprises:

[0017] S11: the enhanced feature of the dimension reduction feature After three linear transformations, a projection matrix and a time step parameter are generated respectively:

[0018]

[0019] In the formula, Linear represents a linear layer, ξ1 and ξ3 are projection matrices, and ξ2 is a time step parameter;

[0020] S12: using non-causal state space dual transformation for feature mapping, obtaining a discrete state through exponential transformation:

[0021]

[0022] wherein, is a hidden state dynamic coefficient, I is a remote sensing input image;

[0023] S13: the Hadamard multiplication is used to generate the hidden state h combining the convolution operation result of the projection matrix - :

[0024] h - = θ ⊙ DWConv(ξ).

[0025] Preferably, the step of mapping the reduced dimension feature to the hidden state space and performing the acausal state space dual transformation in the S1 includes:

[0026] S14: the enhanced feature φ is mapped to the hidden state space, and the acausal state space dual transformation is performed:

[0027]

[0028] wherein, h is the mapped hidden state, which is used to capture the long-range dependency;

[0029] S15: the channel mixing is performed on the mapped hidden state h and the projection matrix ξ3 in the hidden state space, to generate the hidden state feature h + :

[0030]

[0031] wherein, Sigmoid represents a sigmoid activation function.

[0032] Preferably, the S2 includes the following steps:

[0033] S21: the reduced dimension feature φ and the hidden state feature h + are respectively reduced in dimension through a 1x1 convolution layer, to obtain φ 1*1 , and the channel mixing is performed thereon;

[0034] S22: the channel mixed feature is respectively extracted through a 3x3 convolution layer to obtain local spatial information, to obtain φ 3*3 and

[0035] S23: the reduced dimension feature φ and the hidden state feature h + are added to fuse the global information of both, and then a 3x3 deep separable convolution layer and a residual connection are performed to obtain the feature φh :

[0036]

[0037] In the formula, This represents the matrix addition operation;

[0038] S24: For feature φ h After channel displacement, respectively with φ 3*3 and Add them together to get and

[0039] S25: Will and After channel mixing, each channel is passed through a 1×1 convolutional layer and then summed to obtain the final enhanced feature φ. + This data is then used as input for subsequent segmentation tasks.

[0040] Preferably, the channel displacement operation steps include: performing cyclic displacement on the input features in the channel dimension direction, that is, moving the channel forward or backward by a certain step size to adjust the arrangement order of the input features in the channel dimension.

[0041] Preferably, S3 includes the segmentation step of the segmentation model:

[0042] The final enhanced features are gradually restored to the resolution of the remote sensing input image based on a progressive upsampling strategy;

[0043] Feature extraction and classification are performed on the recovered feature map to obtain the predicted segmentation probability map P.

[0044] Preferably, step S3 includes a training step for the segmentation model:

[0045] The segmentation model is trained using a composite loss function that combines Dice Loss and cross-entropy loss.

[0046]

[0047]

[0048] In the formula, α is the weighting coefficient, i is the pixel index, and there are a total of N pixels. i For real labels, P i represents the predicted probability for the corresponding category, and ∈ represents a parameter that is less than a preset value.

[0049] The present invention has the following advantages over the prior art:

[0050] Computational efficiency and resource occupation optimization: The present application reduces the computational overhead in the feature transformation process through the hidden state mixed state space dual design, making the global information interaction more efficient. At the same time, the channel mixing and channel displacement mechanism is introduced, which avoids the high computational complexity brought by the traditional attention mechanism while ensuring the feature fusion effect. Therefore, when processing large-scale high-resolution remote sensing images, this method can significantly reduce the computational resource occupation and improve the inference speed, making it suitable for edge computing devices and cloud parallel processing tasks.

[0051] Feature fusion and segmentation accuracy improvement: The present application realizes efficient interaction of multiple features through channel mixing, channel displacement and depth separable convolution modules, so that different data can share information more fully and avoid information redundancy and feature independence problems. At the same time, in the decoding process, combined with skip connection and multi-level feature fusion, the ability to capture detailed information is enhanced, and the edge accuracy and robustness of the segmentation result are improved. Experiments show that this method has higher accuracy and generalization ability than traditional methods in remote sensing image segmentation tasks, especially in the differentiation of complex ground object classes.

[0052] In summary, the present application further improves the accuracy and robustness of remote sensing image segmentation under the premise of efficient computation, providing a more efficient and accurate solution for remote sensing data analysis. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0054] Figure 1 is a flowchart of the remote sensing image segmentation method based on state space dual transformation and channel mixing provided by the embodiment of the present application;

[0055] Figure 2 is a principle diagram of the hidden state mixed state space dual transformation provided by the embodiment of the present application;

[0056] Figure 3 is a principle diagram of the feature channel mixing and displacement enhancement provided by the embodiment of the present application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

[0058] The present application discloses a remote sensing image segmentation method based on state space dual transformation and channel mixing. In view of the problems of complex ground object categories, large scale changes, and difficult global information acquisition in remote sensing images, the present application combines state space models and hidden state mixing technology to realize efficient global information interaction. The feature channel mixing and displacement enhancement mechanism are used to optimize cross-modal information fusion to improve the segmentation effect of remote sensing images. Figure 1 As shown in the figure, the method comprises the following steps:

[0059] S1: The remote sensing input image is reduced in dimension and features are extracted through a series of convolution and linear transformation operations to obtain reduced dimension features. The reduced dimension features are compressed to generate hidden states. The reduced dimension features are mapped to the hidden state space by using the non-causal state space dual transformation. Channel mixing is performed in the hidden state space to generate hidden state features.

[0060] S2: The reduced dimension features and the hidden state features are mixed after being reduced in dimension. The global information of the reduced dimension features and the hidden state features is fused by adding them together. The global information is subjected to channel displacement, and then mixed with the channel mixed reduced dimension features and hidden state features. The channel displaced reduced dimension features and hidden state features are mixed again to extract local spatial information, and the final enhanced features are generated by fusion.

[0061] S3: Based on the final enhanced features, a segmentation task is performed to output a segmentation result.

[0062] In order to fully reflect the role of each step of the present application, the feature extraction stage of the encoder is described

[0063] In the feature extraction stage, an efficient hidden state mixing method is introduced to optimize the global information interaction of remote sensing images. First, the input image is reduced in dimension by a series of convolution operations to reduce the computational complexity and extract key features. Then, the state space transformation technology is used to map the original features to the hidden state space through linear transformation and state space dual mapping, so that the features can be interacted in a more compact representation.

[0064] In particular, the embodiment adopts a non-causal state space transformation method to effectively capture long-range dependencies and improve global information modeling capabilities. In addition, to further improve feature expression capabilities, a hidden state mixing strategy is proposed to mix channels within the state space, thereby reducing computational complexity and improving computational efficiency. This design can ensure efficient information interaction while optimizing the consumption of computational resources of the model, making it more suitable for large-scale remote sensing data processing.

[0065] To further enhance the feature expression capability of remote sensing images, the embodiment introduces a feature channel mixing and displacement enhancement mechanism to optimize the fusion process of different modal information. In the feature mixing process, convolutional dimensionality reduction is used to enable different source features to interact in a more compact manner. Next, a channel mixing strategy is introduced to divide channel groups and cross-exchange them, allowing different modal information to be more effectively fused and avoiding information redundancy and monotonous linear combination. Compared with traditional concatenation or weighted summation methods, the channel mixing method of the embodiment can promote efficient fusion of cross-modal features and improve segmentation accuracy and generalization ability.

[0066] In addition, to further enhance feature sharing capabilities, the embodiment provides a channel displacement-based enhancement mechanism. This method adjusts the arrangement of feature channels to enable more comprehensive interaction between different modal features, similar to a spatial attention mechanism, but without additional computational overhead. Through this operation, the multi-source information in remote sensing images can be more fully utilized to improve the final segmentation performance.

[0067] In one embodiment, as shown in Figure 2 the step of dimensionality reduction in S1 includes:

[0068] An original input image I is received, which has a size of HxWx3 (where H and W represent the height and width of the image, respectively, and 3 represents the number of RGB channels). To extract features, four consecutive 3x3 convolutional layers are used for downsampling, with a step size of 2 for each convolutional layer to reduce computational complexity and obtain key features. The final dimension-reduced feature map φ is obtained:

[0069] φ = DWConv(I)

[0070] This feature map provides preliminary information expression for subsequent state space conversion.

[0071] In one embodiment, as shown in Figure 2 before the feature extraction of the remote sensing input image in S1, the following steps are further included:

[0072] To further enhance the feature expression ability, the phi is processed by a depthwise separable convolution (DWConv), then connected with the original phi by residual connection, and the enhanced feature representation is obtained through a normalization layer (Layer Normalization)

[0073]

[0074] This process not only enhances the feature extraction ability, but also improves the stability and generalization ability of the network.

[0075] Multi-layer depthwise separable convolution (DWConv) and residual connection mechanism are adopted to further optimize information flow and feature representation ability. Depthwise separable convolution can effectively reduce the computational complexity while preserving local spatial information, and residual connection can maintain the information flow of the original features, avoid the problem of gradient disappearance, and improve the stability of the model.

[0076] In one embodiment, the step of generating a hidden state by compressing the reduced dimension feature in S1 includes:

[0077] The reduced dimension feature map is transformed by a linear transformation to generate a projection matrix and a time step parameter, a discrete state is obtained based on the projection matrix and the time step parameter through an exponential transformation, and the hidden state h is generated by Hadamard multiplication combined with the convolution operation result of the projection matrix - .

[0078] In one embodiment, as shown in Figure 2 the step of generating a hidden state by compressing the reduced dimension feature in S1 includes:

[0079] S11: The reduced dimension feature is enhanced to generate a feature After three linear transformations, a projection matrix and a time step parameter are generated respectively:

[0080]

[0081] In the formula, Linear represents a linear layer, xi, x3 is a projection matrix, and x2 is a time step parameter;

[0082] S12: The feature is mapped by using a non-causal state space dual transformation to obtain a discrete state through an exponential transformation:

[0083]

[0084] In the formula, is a hidden state dynamic coefficient, and I is a remote sensing input image;

[0085] S13: The Hadamard multiplication is used to generate the hidden state h based on the convolution operation result of the projection matrix - :

[0086] h - = θ ⊙ DWConv(ξ).

[0087] The key of this step is to migrate the traditional feature space transformation operation to the hidden state space, so that the computational complexity is reduced. This optimization is particularly suitable for high-resolution remote sensing image processing, which can effectively reduce memory occupation and improve computational efficiency.

[0088] In one embodiment, the step of mapping the reduced dimension feature to the hidden state space by using the acausal state space dual transformation in S1, performing channel mixing in the hidden state space, and generating the hidden state feature h includes:

[0089] S14: The enhanced feature φ is mapped to the hidden state space, and the acausal state space dual transformation is performed:

[0090]

[0091] In the formula, h is the mapped hidden state, which is used to capture long-range dependencies;

[0092] S15: The mapped hidden state h and the projection matrix ξ3 are mixed in the hidden state space to generate the hidden state feature h + :

[0093]

[0094] In the formula, Sigmoid represents the sigmoid activation function.

[0095] By performing channel mixing in the hidden state space, this embodiment effectively reduces the computational overhead of the SSD layer, optimizes the computational complexity while maintaining high expression capability. This method is particularly suitable for edge devices or cloud-based large-scale remote sensing data processing tasks.

[0096] In one embodiment, as shown in Figure 3 , S2 includes the following steps:

[0097] S21: The reduced dimension feature φ and the hidden state feature h are respectively reduced in dimension by a 1x1 convolution layer to obtain φ + , 1*1 , and the two are mixed in channel; Channel mixing connects the data streams of two different features, and has the following effects:

[0098] ​Cross-modal feature interaction: By shuffling and rearranging the channels of different modalities, information is effectively fused, avoiding information loss caused by separate processing.

[0099] Enhancing feature sharing ability: Breaking the feature independence of a single modality, enabling complementary learning, and improving segmentation accuracy.

[0100] Enhancing information expression: Compared with direct concatenation or addition operation, channel mixing mechanism can more effectively promote cross-modal feature flow and avoid information redundancy caused by monotonous linear combination.

[0101] In this embodiment, the specific operation steps of channel mixing include: dividing the channels into multiple groups (Group), and performing channel cross exchange within the group, so that the information of different channels can be more efficiently fused. The rearranged features are used for subsequent convolution calculation to ensure that the distribution of the mixed features has stronger representation ability.

[0102] S22: The features after channel mixing are respectively extracted by 3x3 convolution layers to obtain local spatial information 3*3 and This operation helps to enhance the edge, texture and structure information of different features to adapt to the complex ground object types of remote sensing images.

[0103] S23: Add the dimensionality reduction feature φ and the hidden state feature h + to fuse the global information of both, and then pass through a 3x3 depth separable convolution layer (extract local correlation) and a residual connection (not only preserve the original features, but also improve the training stability and avoid gradient vanishing) to obtain the feature h :

[0104]

[0105] wherein, represents matrix addition operation;

[0106] S24: After channel displacement of the feature h , add it with 3*3 and respectively to obtain and The introduction of channel displacement operation promotes deep information interaction between features, enabling more effective joint decision of multi-source data in remote sensing images, and enabling different modalities to share information more fully. The role of channel displacement:

[0107] Feature information transmission: Channel displacement mechanism enables features of different modalities to interact with each other rather than isolated calculation, thereby improving the fusion effect.

[0108] Modal-aware enhancement: By adjusting the channel order, different modalities can capture each other's contextual information, similar to the idea of ShiftNet or SpatialShiftModule (SSM), achieving lightweight information interaction.

[0109] Computational efficiency: Compared with additional attention mechanisms or complex transformations, channel shifting does not introduce additional parameters, but only adjusts the arrangement of features, so it has lower computational overhead.

[0110] S25: After channel mixing, and After 1x1 convolution layer, add up to the final enhanced feature φ + , which carries multiple class fusion information, will be used as input for subsequent segmentation tasks, making the final segmentation result more robust and accurate.

[0111] In one embodiment, the operation steps of channel shifting include: performing cyclic shift on the input feature in the channel dimension direction, i.e. moving the channel forward or backward by a certain step, to adjust the arrangement order of the input feature in the channel dimension.

[0112] In one embodiment, to optimize the segmentation performance, S3 includes the training steps of the segmentation model, which ensures that the model takes into account both the class imbalance problem and the pixel-level classification accuracy:

[0113] S301: Train the segmentation model with a composite loss function combining DiceLoss (for handling class imbalance) and cross-entropy loss:

[0114]

[0115] where α is the weight coefficient, i is the pixel index, and there are N total pixel points. y i is the true label, P i is the predicted probability of the corresponding class, and ∈ is a small value to prevent the denominator from being zero.

[0116] S302: Use Adam optimizer for gradient update, and set the learning rate to an adaptive decay strategy (such as cosine annealing or poly decay) to ensure stable convergence of the model:

[0117]

[0118] where η is the current learning rate, is the loss gradient.

[0119] In one embodiment, the trained and optimized model directly performs forward propagation on the input image in the inference stage, and outputs a segmentation probability map P. Finally, the class label of each pixel point is obtained through the argmax operation, and the final segmentation result is obtained:

[0120]

[0121] In the formula, is the final pixel-level classification result.

[0122] So far, the method of the present application has realized the whole process from remote sensing image input to the final segmentation result.

[0123] In specific implementation, since φ + In the first two steps, dimension reduction and feature transformation are experienced, in order to restore the spatial scale matching the input image, the embodiment provides a step-by-step upsampling strategy based on the decoder.

[0124] The specific process is as follows:

[0125] S311: deconvolution upsampling: through a series of deconvolution operations, the features are gradually restored to the original resolution, and the step length is set to 2 each time to maintain the continuity of information flow.

[0126] S312: skip connection: in the decoding process, the U-Net structure is used to splice the shallow features φ generated in S1 with the features gradually upsampling to enhance local detail information and reduce the blurring effect of high-level features.

[0127] S313: fusion operation: in the upsampling process, a 3x3 convolution layer (with Batch Normalization and ReLU) is used to further extract the fused features to ensure the sufficiency of information expression.

[0128] S314: reduce the number of channels to the classification number C through a 1x1 convolution layer to obtain the predicted segmentation probability map P.

[0129] The above describes in detail the remote sensing image segmentation method provided by the present application based on state space dual transformation and channel mixing. The principles and implementation modes of the present application are described by specific examples in this paper. The above examples are only used to help understand the method and core idea of the present application. For those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the above description should not be understood as limiting the present application.

[0130] In this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the existence of additional elements of the same name in the process, method, article, or apparatus.

Claims

1. A remote sensing image segmentation method based on state-space dual transformation and channel mixing, comprising the following steps: S1: The remote sensing input image is reduced in dimension and features are extracted through a series of convolution and linear transformation operations to obtain the reduced dimension features. The reduced dimension features are compressed to generate the hidden state. The reduced dimension features are mapped to the hidden state space using the non-causal state space dual transformation. Channel mixing is performed in the hidden state space to generate the hidden state features. S2: After dimensionality reduction of the reduced features and the hidden state features respectively, channel mixing is performed; the reduced features and the hidden state features are added together to fuse their global information; the global information is channel shifted and then added to the channel-mixed reduced features and the hidden state features respectively; the added reduced features and the hidden state features are channel-mixed again to extract local spatial information and fuse to generate the final enhanced features; The operation steps for the channel displacement include: Perform a cyclic displacement on the input features in the channel dimension, that is, move the channels forward or backward by a certain step to adjust the arrangement order of the input features in the channel dimension; S3: Perform a segmentation task based on the final enhanced features and output the segmentation result.

2. The remote sensing image segmentation method based on state-space dual transformation and channel hybridization according to claim 1, characterized in that, The steps for dimensionality reduction of the input image in S1 include: The system receives remote sensing input images and performs downsampling processing through four consecutive 3×3 convolutional layers to generate dimensionality-reduced feature maps.

3. The remote sensing image segmentation method based on state-space dual transformation and channel hybridization according to claim 1, characterized in that, Before feature extraction from the remote sensing input image in S1, the following steps are also included: The depthwise separable convolution is performed on the dimensionality-reduced feature map, and then combined with the dimensionality-reduced feature map through residual connection. After normalization, the enhanced feature representation is obtained; feature extraction is performed on the enhanced feature representation.

4. The remote sensing image segmentation method based on state-space dual transformation and channel hybridization according to claim 1, characterized in that, The step in S1 to compress the reduced-dimensional features to generate the hidden state includes: The dimensionality-reduced feature map is transformed linearly to generate a projection matrix and a time step parameter. Based on the projection matrix and the time step parameter, an exponential transformation is performed to obtain the discrete state. Finally, the hidden state h is generated using Hadamard multiplication, combining the result of the convolution operation on the projection matrix. - .

5. The remote sensing image segmentation method based on state-space dual transformation and channel hybridization according to claim 4, characterized in that, The step in S1 to compress the reduced-dimensional features to generate the hidden state includes: S11: The enhanced features of the dimensionality reduction features After three linear transformations, the projection matrix and time step parameters are generated respectively: In the formula, Linear represents a linear layer, ξ1 and ξ3 are projection matrices, and ξ2 is the time step parameter; S12: Feature mapping is performed using a non-causal state-space dual transformation, and discrete states are obtained through exponential transformation. In the formula, Here, I represents the hidden state dynamic coefficients, and I represents the remote sensing input image. S13: Combine the results of the convolution operation on the projection matrix with Hadamard multiplication to generate the hidden state h. - :

6. The remote sensing image segmentation method based on state-space dual transformation and channel hybridization according to claim 5, characterized in that, In S1, the steps of mapping the dimensionality-reduced features to the hidden state space using a non-causal state-space dual transformation, and then performing channel mixing within the hidden state space to generate hidden state features include: S14: Features enhanced by the dimensionality reduction features Map to the hidden state space and perform a noncausal state space dual transformation: In the formula, h is the mapped hidden state, used to capture remote dependencies; S15: Perform channel mixing on the mapped hidden state h and the projection matrix ξ3 in the hidden state space to generate the hidden state feature h. + : In the formula, Sigmoid represents the sigmoid activation function.

7. The remote sensing image segmentation method based on state-space dual transformation and channel hybridization according to claim 1, characterized in that, S2 includes the following steps: S21: The dimensionality reduction features are... and the hidden state feature h + Dimensionality reduction was achieved by using 1×1 convolutional layers respectively, resulting in And then the two are mixed in channels; S22: The features after channel blending are respectively processed through 3×3 convolutional layers to extract local spatial information, resulting in... and S23: The dimensionality reduction features are... and the hidden state feature h + The two are added together to fuse their global information, and then passed through a 3×3 depthwise separable convolutional layer and residual connections to obtain features. In the formula, This represents the matrix addition operation; S24: Features After channel displacement, respectively with and Add them together to get and S25: Will and After channel mixing, each channel is passed through a 1×1 convolutional layer and then summed to obtain the final enhanced feature. And it serves as the input for subsequent segmentation tasks.

8. The remote sensing image segmentation method based on state-space dual transformation and channel hybridization according to claim 1, characterized in that, S3 includes the segmentation steps of the segmentation model: The final enhanced features are gradually restored to the resolution of the remote sensing input image based on a progressive upsampling strategy; Feature extraction and classification are performed on the recovered feature map to obtain the predicted segmentation probability map P.

9. The remote sensing image segmentation method based on state-space dual transformation and channel hybridization according to claim 1, characterized in that, S3 includes the training steps for the segmentation model: The segmentation model is trained using a composite loss function that combines Dice Loss and cross-entropy loss. In the formula, α is the weighting coefficient, i is the pixel index, there are a total of N pixels, and y i For real labels, P i ε represents the predicted probability for the corresponding category, and ε is a parameter that is less than a preset value.

Images