Semantic segmentation method for remote sensing images based on global dependency and local detail collaboration

Abstract

This invention discloses a remote sensing image semantic segmentation method based on global dependency and local detail collaboration, belonging to the field of remote sensing image segmentation technology. It includes: acquiring remote sensing images and preprocessing them to obtain a dataset; constructing a remote sensing image semantic segmentation model, including a ConvNeXt-based backbone network feature extraction module, an LA-Mamba decoding module, a multi-scale spatial feature fusion module, and a prediction output module connected sequentially; training the remote sensing image semantic segmentation model based on the dataset; acquiring the remote sensing image to be segmented, and performing semantic segmentation using the trained remote sensing image semantic segmentation model to obtain the segmentation result. This invention significantly improves segmentation performance while maintaining accuracy by designing an LA-Mamba module and a multi-scale spatial feature fusion module in the decoding stage to achieve complementary modeling of local details and global context.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image segmentation technology, and in particular to a remote sensing image semantic segmentation method based on global dependency and local detail collaboration. Background Technology

[0002] With significant advancements in remote sensing imaging technology, the spatial, spectral, and temporal resolution of remote sensing images has been improved, providing a robust data foundation for various remote sensing applications. Semantic segmentation of remote sensing images involves classifying each pixel in a remote sensing image into predefined land cover categories, such as buildings, roads, low vegetation, trees, water bodies, and vehicles, generating a high-precision semantic segmentation map. Compared to traditional classification methods based on the entire image or local patches, semantic segmentation provides more refined boundary and spatial distribution information, thus holding significant importance in applications such as urban planning, land use monitoring, agricultural yield estimation, and disaster emergency response.

[0003] With the continuous development of deep learning technology, deep learning-based methods have achieved significant success in the field of remote sensing image segmentation. Among existing methods, Convolutional Neural Networks (CNNs) were the earliest and most widely used models for remote sensing image segmentation. CNNs rely on local receptive fields and weight sharing mechanisms to efficiently extract local spatial features. However, due to the limited receptive field of convolutional kernels, even with multi-layer stacking to expand the receptive field, their global modeling ability remains insufficient, easily leading to inaccurate segmentation results for large-scale targets and long-distance dependent scenes.

[0004] The introduction of Transformer provides a new approach to semantic segmentation. Transformer's self-attention mechanism can capture dependencies between arbitrary locations globally, overcoming the limitations of CNN. However, the computational complexity of Transformer is quadratic with the input image size. High-resolution remote sensing images often reach thousands of pixels, and directly using Transformer leads to excessive memory consumption and slow computation speed, making it difficult to meet the practical needs of large-scale remote sensing image processing.

[0005] In recent years, state-space models, especially the Mamba architecture, have been increasingly introduced into visual tasks. Mamba achieves efficient modeling of long sequences through a selective state mechanism, possessing linear computational complexity and demonstrating great potential in visual modeling. While Mamba has alleviated some of the challenges in remote sensing segmentation, its ability to capture local edges, textures, and small targets is limited, easily leading to the loss of detailed information.

[0006] In summary, existing methods generally suffer from the following shortcomings: CNNs are limited to local modeling and lack global information. While Transformers possess global modeling capabilities, they are computationally expensive and poor at preserving details for high-resolution remote sensing images. State-space models are efficient in global modeling but lack the ability to represent local boundaries and small targets. Therefore, designing a remote sensing image semantic segmentation method that can preserve local details, efficiently establish global dependencies, and achieve high computational efficiency has become a pressing technical problem to be solved in this field. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a remote sensing image semantic segmentation method for smart cities that utilizes collaborative modeling of global dependencies and local details. This method employs an LA-Mamba module and a multi-scale spatial feature fusion module (MSFF) in the decoding stage to achieve complementary modeling of local details and global context, thereby significantly improving segmentation performance while maintaining accuracy.

[0008] To achieve the above objectives, this invention proposes a remote sensing image semantic segmentation method based on global dependency and local detail collaboration, comprising the following steps: (1) Acquire remote sensing images and preprocess them to obtain a dataset; (2) Construct a semantic segmentation model for remote sensing images, which includes a ConvNeXt-based backbone network feature extraction module, an LA-Mamba decoding module, a multi-scale spatial feature fusion module, and a prediction output module connected in sequence; (3) Based on the dataset, train the semantic segmentation model of the remote sensing image; (4) Obtain the remote sensing image to be segmented, and use the trained remote sensing image semantic segmentation model to perform semantic segmentation to obtain the segmentation result.

[0009] Furthermore, the preprocessing involves normalizing and resizing the remote sensing images to form a unified input tensor, and then generating more training samples through data augmentation to form a dataset.

[0010] Furthermore, the ConvNeXt-based backbone network feature extraction module includes four stages; The first stage uses a 4×4 convolution with a stride of 4 to achieve initial downsampling, and calculates the mean and variance of the feature vector at each spatial location along the channel dimension and normalizes them to obtain the input for the first stage. The second to fourth stages use a 2×2 convolution with a stride of 2 to further reduce the resolution and increase the number of channels. After four stages, four sets of feature maps with different resolutions and number of channels are obtained. Each stage consists of a preset number of ConvNeXt Blocks stacked sequentially, with the preset number determined by the network depth parameter.

[0011] Furthermore, the preset number of ConvNeXtBlocks in the four stages of the ConvNeXt-based backbone network feature extraction module are 3, 3, 27, and 3, respectively.

[0012] Furthermore, the ConvNeXt Block includes: passing the input feature map through a depthwise separable convolution; subsequently, the convolution output is rearranged and fed into a layer normalization layer; after normalization, the features enter a pointwise multilayer perceptron module; then, a learnable scaling factor is introduced at the output; finally, a residual connection is used to add the processed features to the original input features and output the result, as shown in the formula: in: For the input feature map, For depthwise separable convolution, For layer normalization, For point-by-point multilayer perceptron, It is a learnable scaling factor.

[0013] Furthermore, the LA-Mamba decoding module includes four stages, corresponding to the four stages of the ConvNeXt-based backbone network feature extraction module; each stage of the LA-Mamba decoding module consists of an LA-Mamba Block.

[0014] Furthermore, the LA-Mamba Block includes a local detail processing branch and a global feature enhancement branch; The local detail processing branch includes: The input features are divided into groups of size [value]. ps Each patch is then flattened into a sequence S; Layered normalization is applied within each patch, and query vectors are generated separately through linear mapping. Key vector Sum value vector The scaling dot product attention mechanism is used to calculate the correlation between local features, thereby obtaining the output result after local enhancement. ; Add the attention output to the residual of the attention output and the flattened sequence S; By performing a de-stitching operation, all local blocks are restored to complete feature maps, yielding the final output of the local detail processing branch. ; in: Representation layer normalization, Indicates reverse splicing; The dimension of the key vector; The global feature enhancement branch includes: Input features First, it goes through linear mapping, depthwise convolution, and... Activation is performed to obtain an intermediate representation with local enhancement, which is then input into a two-dimensional state-space model. Recursive modeling is performed in the four directions: horizontal, vertical, and reverse to capture long-range dependencies. The global features are then output and normalized by layers to obtain the global features. ; Input features Through linear mapping and Activation yields gating features. ; global features With gated features Element-wise multiplication is performed and mapped to enhanced global features through a linear layer. Simultaneously, a learnable scaling factor is introduced. The fusion results are dynamically adjusted to obtain enhanced features. ; Enhance features Input features The residuals are summed, and then passed sequentially through a local detail enhancement module and a channel attention module to obtain fused features. ; Introducing another learnable scaling factor , for fusion features Dynamic adjustments are made to obtain the final output of the global feature enhancement branch. ; in: It is a pointwise linear mapping; For depthwise separable convolution; for Activation function; For layer normalization; This is element-wise multiplication; A two-dimensional state space modeling module; It is a local detail enhancement module; CA It is a channel attention module; The results from the two branches are fused element by element to obtain the output of the LA-Mamba decoding module; in: This is the output of the LA-Mamba decoding module.

[0015] Furthermore, the input to the multi-scale spatial feature fusion module is the feature values ​​output from the first, second, and third stages of the LA-Mamba decoding module. ; The three feature groups are processed using 3×3 convolutions to map them to the same number of channels. After aligning the spatial sizes, element-wise multiplication is performed to obtain the interactive features. ; The interactive features are combined with features from the other two levels to form { F 1, X 1, X 3}, their channel dimensions are unified through 1×1 convolution, and after interpolation to align spatial resolution, they are connected along the scale dimension to obtain ; All scale features are stacked along the channel dimension, and 3D convolution is introduced for modeling along the scale dimension. Batch normalization and nonlinear activation are then applied to enhance their expressive power. After completing the convolutional interactions between scales, 3D adaptive pooling is used to further compress the scale dimension, resulting in the final output of the multi-scale spatial feature fusion module. ; in: It is a 3D convolution; For batch normalization; for Activation function; This is a three-dimensional adaptive pooling method.

[0016] Furthermore, the prediction output module employs deconvolution to progressively upsample the output of the multi-scale spatial feature fusion module to the same spatial resolution as the input image. After scale restoration is completed, a 1×1 convolution is used to map the number of feature channels to the number of categories, thereby obtaining the category probability distribution map corresponding to each pixel.

[0017] Furthermore, the remote sensing image semantic segmentation model employs a composite loss function during training. ; in: Loss due to dice, For cross-entropy loss, and These represent the number of samples and the number of categories, respectively. and Represent the model's predicted probability and the true semantic label's... k Dimensional elements.

[0018] The beneficial effects of this invention are: The LA-Mamba module designed in this invention constructs a decoding structure, modeling detailed information and global dependencies in the local detail processing branch and the global feature enhancement branch respectively, achieving complementary fusion of the two types of features. The local detail processing branch can effectively capture local information such as building edges and small object textures, while the global feature enhancement branch models long-distance dependencies through selective scanning with linear complexity, thereby improving the global consistency and boundary clarity of the segmentation results while ensuring computational efficiency.

[0019] The multi-scale spatial feature fusion module (MSFF) designed in this invention enhances the complementary details between features at different scales through element-wise interaction, and then uses three-dimensional convolution to achieve cross-scale consistent modeling. It can comprehensively integrate shallow texture, edge information and deep semantic features, avoiding the problem of insufficient matching between shallow and deep features in traditional feature fusion methods, thereby significantly improving the segmentation accuracy and robustness of small targets and large targets in complex scenes.

[0020] Meanwhile, this invention employs a modular structural design. The encoder, LA-Mamba module, and multi-scale spatial feature fusion module based on the ConvNeXt backbone network can all be independently combined and expanded (without changing the overall network framework and feature propagation path, the model capacity and computational complexity can be expanded or pruned as needed by stacking or reducing the number of these modules and combining interface operations such as channel mapping and scale alignment). This allows the model to be flexibly adjusted according to different datasets and application scenarios, avoiding the problem of traditional convolutional kernels being fixed and unable to adapt to diverse input features. Therefore, this invention can improve the adaptability and scalability of the model while maintaining high efficiency. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the framework structure of the remote sensing image semantic segmentation model according to an embodiment of the present invention.

[0023] Figure 3This is a schematic diagram of the LA-Mamba Block structure in the decoding stage of an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of the Multi-Scale Spatial Feature Fusion (MSFF) module in an embodiment of the present invention.

[0025] Figure 5 This is a visualization comparison of different methods from the embodiments of the present invention on the Vaihingen dataset.

[0026] Figure 6 This is a visualization comparison of different methods from the embodiments of the present invention on the Potsdam dataset. Detailed Implementation

[0027] The technical solutions in 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] like Figure 1 As shown, this invention provides a remote sensing image semantic segmentation method based on global dependency and local detail collaboration, comprising the following steps: A remote sensing image semantic segmentation method based on global dependency and local detail collaboration is characterized by the following steps: S101. Acquire remote sensing images and perform preprocessing to obtain a dataset.

[0029] This invention employs the Vaihingen and Potsdam datasets to normalize the acquired remote sensing data, mapping pixel values ​​to the [0,1] interval and adjusting their size to form a unified input tensor. Then, through data augmentation, more training samples are generated to ensure the diversity of the input data and form a dataset.

[0030] The data is divided into training and testing sets.

[0031] S102. Construct a semantic segmentation model for remote sensing images, which includes a ConvNeXt-based backbone network feature extraction module, an LA-Mamba decoding module, a multi-scale spatial feature fusion module, and a prediction output module connected in sequence.

[0032] like Figure 2As shown, the remote sensing image semantic segmentation model extracts multi-scale features based on the ConvNeXt backbone network. In the decoding stage, the LA-Mamba module performs parallel local detail processing and global feature enhancement branches to achieve joint representation of local details and global dependencies. Subsequently, the multi-scale spatial feature fusion (MSFF) module performs unified alignment and cross-scale fusion of features at different levels to generate feature representations that combine fine granularity and context. Finally, the model restores the original resolution through upsampling and convolutional classification heads, and outputs the final semantic segmentation result.

[0033] The ConvNeXt-based backbone network feature extraction module comprises four stages. The first stage uses a 4×4 convolution with a stride of 4 for initial downsampling, and calculates and normalizes the mean and variance of the feature vector at each spatial location along the channel dimension to obtain the first stage input. The second to fourth stages use 2×2 convolutions with a stride of 2 to further reduce resolution and increase the number of channels. After these four stages, four sets of feature maps with different resolutions and channel numbers are obtained. Each stage consists of a preset number of ConvNeXt Blocks stacked sequentially, the preset number determined by the network depth parameter. In this embodiment, the preset numbers of ConvNeXt Blocks in the four stages of the ConvNeXt-based backbone network feature extraction module are 3, 3, 27, and 3, respectively.

[0034] The internal workflow of ConvNeXt Block is as follows: The input feature map is processed through a depthwise separable convolution. The convolution output is then rearranged and fed into a normalization layer. After normalization, the features enter a pointwise multilayer perceptron module (equivalent to a 1×1 convolution). Next, a learnable scaling factor is introduced at the output. Finally, residual connections are used to add the processed features to the original input features, and the output is then calculated as follows: in: For the input feature map, For depthwise separable convolution, For layer normalization, For point-by-point multilayer perceptron, It is a learnable scaling factor.

[0035] Depthwise separable convolution: This operation can effectively extract local spatial structure information, such as building edges and road textures, while maintaining low computational cost; Layer normalization: To eliminate distribution differences between different channels and stabilize the training process; Learnable scaling factor: To dynamically adjust the contribution of residual branches, thereby enhancing representational capabilities while avoiding gradient instability; Residual connection: To ensure smooth information transmission as deep networks deepen, helping the model capture multi-level features and alleviate the gradient vanishing problem.

[0036] The LA-Mamba decoding module consists of four stages, corresponding to the four stages of the ConvNeXt-based backbone network feature extraction module; each stage of the LA-Mamba decoding module is composed of an LA-Mamba Block.

[0037] The LA-Mamba Block includes branches for local detail processing and global feature enhancement, such as... Figure 3 As shown, its core objective is to simultaneously capture local details and global dependencies in urban remote sensing images to enhance the expressive power of decoded features.

[0038] (1) such as Figure 3 As shown in (b), the local detail processing branch mainly focuses on neighborhood feature interactions and detail preservation, specifically including: (1.1) Divide the input features into patches of size ps, and then flatten each patch into a sequence S.

[0039] (1.2) Apply layer normalization processing within each patch and generate query vectors through linear mapping respectively. Key vector Sum value vector The scaling dot product attention mechanism is used to calculate the correlation between local features, thereby obtaining the output result after local enhancement. .

[0040] (1.3) Add the attention output to the residual of the flattened sequence S.

[0041] (1.4) By performing the inverse stitching operation, all local blocks are restored to complete feature maps, and the final output of the local detail processing branch is obtained. .

[0042] in: Representation layer normalization, Indicates reverse splicing; is the dimension of the key vector.

[0043] (2) As attached Figure 3 As shown in (a), the global feature enhancement branch consists of a Visual State Space Model (VSSM) module and a local detail enhancement module. Its core is to efficiently capture long-range dependencies through the selective scanning mechanism of the VSSM module. Specifically, it includes: (2.1) Input features First, it goes through linear mapping, depthwise convolution, and... Activation is performed to obtain an intermediate representation with local enhancement, which is then input into a two-dimensional state-space model. Recursive modeling is performed in the four directions: horizontal, vertical, and reverse to capture long-range dependencies. The global features are then output and normalized by layers to obtain the global features. .

[0044] (2.2) Input features Through linear mapping and Activation yields gating features. .

[0045] (2.3) Global features With gated features Element-wise multiplication is performed and mapped to enhanced global features through a linear layer. Simultaneously, a learnable scaling factor is introduced. The fusion results are dynamically adjusted to obtain enhanced features. .

[0046] (2.4) Enhanced features Input features The residuals are summed and then passed sequentially through the Local Detail Enhancement (LDE) module and the Channel Attention module. ) obtain fusion features .

[0047] (2.5) Introduce another learnable scaling factor , for fusion features Dynamic adjustments are made to obtain the final output of the global feature enhancement branch. .

[0048] in: It is a pointwise linear mapping; For depthwise separable convolution; for Activation function; For layer normalization; This is element-wise multiplication; A two-dimensional state space modeling module; It is a local detail enhancement module; It is a channel attention module; (3) The results of the two branches are fused element by element to obtain the output of the LA-Mamba decoding module; in: This is the output of the LA-Mamba decoding module.

[0049] The multi-scale spatial feature fusion module takes into account spatial details and semantic context information at different levels, providing a richer representation for subsequent upsampling and prediction outputs. Its input consists of the feature values ​​output from the first, second, and third stages of the LA-Mamba decoding module. (like Figure 2 As shown), the specific process is as follows: Figure 4 As shown: The three feature groups are processed using 3×3 convolutions to map them to the same number of channels. After aligning the spatial sizes, element-wise multiplication is performed to obtain the interactive features. .

[0050] The interactive features are combined with features from the other two layers to form {F1, X1, X3}. Their channel dimensions are unified using a 1×1 convolution. After interpolation to align with spatial resolution, they are concatenated along the scale dimension to obtain... .

[0051] All scale features are stacked along the channel dimension, and 3D convolution is introduced for modeling along the scale dimension. Batch normalization and nonlinear activation are then applied to enhance their expressive power. After completing the convolutional interactions between scales, 3D adaptive pooling is used to further compress the scale dimension, resulting in the final output of the multi-scale spatial feature fusion module. .

[0052] in: It is a 3D convolution; For batch normalization; for Activation function; This is a three-dimensional adaptive pooling method.

[0053] The prediction output module uses deconvolution to progressively upsample the output of the multi-scale spatial feature fusion module to the same spatial resolution as the input image. After scale restoration, a 1×1 convolution is used to map the number of feature channels to the number of categories, resulting in a category probability distribution map for each pixel.

[0054] S103. Based on the dataset, train the remote sensing image semantic segmentation model.

[0055] A composite loss function was used during the training of the remote sensing image semantic segmentation model. This method combines dice loss and cross-entropy loss to achieve a more balanced and stable model training.

[0056] in: Loss due to dice, For cross-entropy loss, and These represent the number of samples and the number of categories, respectively. and Represent the model's predicted probability and the true semantic label's... Dimensional elements.

[0057] S104. Obtain the remote sensing image to be segmented, and perform semantic segmentation using the trained remote sensing image semantic segmentation model to obtain the segmentation result.

[0058] The experimental setup for this embodiment of the invention is as follows: all experiments were implemented using the PyTorch framework on a single NVIDIA RTX A6000 GPU equipped with an Ubuntu 20.04 operating system. The optimizer was set to AdamW, with a base learning rate of 6e-4 and a weight decay of 2.5e-4. For the Vaihingen and Potsdam datasets, the AdamW optimizer was used for training, with each dataset undergoing 105 epochs and a batch size of 8.

[0059] The evaluation metrics used in this invention embodiment are: average F1-score (mF1), average intersection-over-union ratio (mIoU), and overall accuracy (OA). mIoU (average intersection-over-union ratio): measures the segmentation accuracy of the prediction result; F1 score comprehensively considers recall and precision, and is used to evaluate the balance and completeness of the prediction result; overall accuracy (OA) represents the proportion of correctly classified pixels to the total number of pixels.

[0060] in: k It is the number of target segmentation categories. This is a real example. It is a true counterexample. It is a false positive example. It represents the number of false negative pixels in the result.

[0061] Experimental Results and Analysis: Table 1 shows the results of comparative experiments between the embodiments of the present invention and various remote sensing image segmentation models on the Vaihingen dataset. As can be seen from Table 1, the method of the embodiments of the present invention achieves excellent performance on multiple evaluation metrics, with an overall accuracy (OA) of 91.93%, an average F1 score (mF1) of 91.80%, and an average intersection-over-union ratio (mIoU) of 85.01%, all superior to existing comparative methods. This indicates that the method of the embodiments of the present invention can more effectively capture multi-class features in complex urban scenes, significantly improving the accuracy and robustness of segmentation.

[0062] Table 1 Table 2 shows the experimental results of the embodiments of the present invention on the Potsdam dataset. It can be seen that the method of the embodiments of the present invention also performs outstandingly, with an overall accuracy (OA) of 92.24%, an average F1 value (mF1) of 93.54%, and an average intersection-union ratio (mIoU) of 88.07%. Compared with the comparison methods, it has achieved a leading advantage in multiple categories, especially in the segmentation accuracy of categories such as building, road, and car, which fully verifies the superiority of the present invention in large-scale high-resolution remote sensing image scenarios.

[0063] Table 2 Figure 5 Visual comparison charts of different methods from the embodiments of the present invention on the Vaihingen dataset are provided. Figure 6 A visualization comparison of the present invention and existing methods on the Potsdam dataset is shown, and the visualization results demonstrate the superior segmentation performance of the method in the embodiments of the present invention for complex scenes.

[0064] In summary, comparative experiments on the Vaihingen and Potsdam datasets demonstrate that the method of this invention exhibits higher segmentation accuracy and generalization ability across different scenarios and categories. It effectively addresses the problems of insufficient local detail extraction and inadequate global context modeling in existing methods, demonstrating high practical value and promising prospects for widespread application.

[0065] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A semantic segmentation method for remote sensing images based on global dependency and local detail collaboration, characterized in that, Includes the following steps: (1) Acquire remote sensing images and preprocess them to obtain a dataset; (2) Construct a semantic segmentation model for remote sensing images, which includes a ConvNeXt-based backbone network feature extraction module, an LA-Mamba decoding module, a multi-scale spatial feature fusion module, and a prediction output module connected in sequence; (3) Based on the dataset, train the semantic segmentation model of the remote sensing image; (4) Obtain the remote sensing image to be segmented, and use the trained remote sensing image semantic segmentation model to perform semantic segmentation to obtain the segmentation result.

2. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 1, characterized in that: The preprocessing involves normalizing and resizing the remote sensing images to form a unified input tensor, and then using data augmentation to generate more training samples to form a dataset.

3. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 1, characterized in that: The ConvNeXt-based backbone network feature extraction module includes four stages; The first stage uses a 4×4 convolution with a stride of 4 to achieve initial downsampling, and calculates the mean and variance of the feature vector at each spatial location along the channel dimension and normalizes them to obtain the input for the first stage. The second to fourth stages use a 2×2 convolution with a stride of 2 to further reduce the resolution and increase the number of channels. After four stages, four sets of feature maps with different resolutions and number of channels are obtained. Each stage consists of a preset number of ConvNeXt Blocks stacked sequentially, with the preset number determined by the network depth parameter.

4. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 3, characterized in that, The preset number of ConvNeXt Blocks in the four stages of the ConvNeXt-based backbone network feature extraction module are 3, 3, 27, and 3, respectively.

5. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 3, characterized in that, The ConvNeXt Block comprises: processing the input feature map through a depthwise separable convolution; subsequently, rearranging the convolution output and feeding it into a layer normalization layer; after normalization, the features enter a pointwise multilayer perceptron module; then, introducing a learnable scaling factor at the output; and finally, using residual connections, adding the processed features to the original input features and outputting the result, as shown in the formula: in: For the input feature map, For depthwise separable convolution, For layer normalization, For point-by-point multilayer perceptron, It is a learnable scaling factor.

6. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 3, characterized in that: The LA-Mamba decoding module comprises four stages, corresponding to the four stages of the ConvNeXt-based backbone network feature extraction module; each stage of the LA-Mamba decoding module consists of an LA-Mamba Block.

7. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 6, characterized in that: The LA-Mamba Block includes a local detail processing branch and a global feature enhancement branch; The local detail processing branch includes: The input features are divided into groups of size [value]. ps Each patch is then flattened into a sequence S; Layered normalization is applied within each patch, and query vectors are generated separately through linear mapping. Key vector Sum value vector The scaling dot product attention mechanism is used to calculate the correlation between local features, thereby obtaining the output result after local enhancement. ; Add the attention output to the residual of the attention output and the flattened sequence S; By performing a de-stitching operation, all local blocks are restored to complete feature maps, yielding the final output of the local detail processing branch. ; in: Representation layer normalization, Indicates reverse splicing; The dimension of the key vector; The global feature enhancement branch includes: Input features First, it goes through linear mapping, depthwise convolution, and... Activation is performed to obtain an intermediate representation with local enhancement, which is then input into a two-dimensional state-space model. Recursive modeling is performed in the four directions: horizontal, vertical, and reverse to capture long-range dependencies. The global features are then output and normalized by layers to obtain the global features. ; Input features Through linear mapping and Activation yields gating features. ; global features With gated features Element-wise multiplication is performed and mapped to enhanced global features through a linear layer. Simultaneously, a learnable scaling factor is introduced. The fusion results are dynamically adjusted to obtain enhanced features. ; Enhance features Input features The residuals are summed, and then passed sequentially through a local detail enhancement module and a channel attention module to obtain fused features. ; Introducing another learnable scaling factor , for fusion features Dynamic adjustments are made to obtain the final output of the global feature enhancement branch. ; in: It is a pointwise linear mapping; For depthwise separable convolution; for Activation function; For layer normalization; This is element-wise multiplication; A two-dimensional state space modeling module; It is a local detail enhancement module; CA It is a channel attention module; The results from the two branches are fused element by element to obtain the output of the LA-Mamba decoding module; in: This is the output of the LA-Mamba decoding module.

8. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 6, characterized in that: The input to the multi-scale spatial feature fusion module is the feature values ​​output from the first, second, and third stages of the LA-Mamba decoding module. ; The three feature groups are processed using 3×3 convolutions to map them to the same number of channels. After aligning the spatial sizes, element-wise multiplication is performed to obtain the interactive features. ; The interactive features are combined with features from the other two levels to form { F 1, X 1, X 3}, their channel dimensions are unified through 1×1 convolution, and after interpolation to align spatial resolution, they are connected along the scale dimension to obtain ; All scale features are stacked along the channel dimension, and 3D convolution is introduced for modeling along the scale dimension. Batch normalization and nonlinear activation are then applied to enhance their expressive power. After completing the convolutional interactions between scales, 3D adaptive pooling is used to further compress the scale dimension, resulting in the final output of the multi-scale spatial feature fusion module. ; in: It is a 3D convolution; For batch normalization; for Activation function; This is a three-dimensional adaptive pooling method.

9. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 1, characterized in that: The prediction output module uses deconvolution operation to gradually upsample the output of the multi-scale spatial feature fusion module to the same spatial resolution as the input image; After scale restoration, a 1×1 convolution is used to map the number of feature channels to the number of categories, resulting in a probability distribution map of the category corresponding to each pixel.

10. The remote sensing image semantic segmentation method based on global dependency and local detail collaboration according to claim 1, characterized in that: The semantic segmentation model for remote sensing images employs a composite loss function during training. ; in: Loss due to dice, For cross-entropy loss, and These represent the number of samples and the number of categories, respectively. and Represent the model's predicted probability and the true semantic label's... k Dimensional elements.

Images