Remote sensing image segmentation method based on multi-scale gating bottleneck convolution scanning

Abstract

The invention discloses a remote sensing image segmentation method based on multi-scale gating bottleneck convolution scanning, and relates to the field of remote sensing image segmentation, and the method comprises the steps: inputting a remote sensing image, and extracting initial features through an initial convolution layer; carrying out multi-level down-sampling processing on the initial features through an encoder, and carrying out feature enhancement on each level by adopting a multi-scale gating bottleneck convolution scanning unit in sequence to obtain multi-level encoding features; introducing a semantic bridging attention unit at the deepest layer feature of the encoder to carry out global semantic modeling and residual fusion to obtain an enhanced deep layer semantic feature; performing up-sampling processing step by step in a decoder, and fusing the up-sampled feature with the coding feature of the corresponding level to obtain a multi-level fusion feature; and inputting the fusion feature output by the decoder into the segmentation head to obtain a prediction probability graph. According to the method, the segmentation continuity and category discrimination robustness of slender targets, complex boundaries and small-scale ground features in the remote sensing image are improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image segmentation, and more specifically to a remote sensing image segmentation method based on multi-scale gated bottleneck convolutional scanning. Background Technology

[0002] High-resolution remote sensing imagery has played a crucial role in fields such as land resource surveys, urban planning, agricultural monitoring, disaster prevention and mitigation, and environmental monitoring. Among these, remote sensing image segmentation aims to classify features such as roads, rivers, farmland, and buildings in images at the pixel level, and its accuracy and efficiency directly affect the reliability of subsequent analysis and decision-making.

[0003] Although the encoder-decoder architecture based on convolutional neural networks has made great progress in remote sensing image segmentation, traditional convolution operators are limited by local receptive fields and have difficulty effectively modeling long-range spatial dependencies between pixels. This often results in fragmentation, blurring, and structural distortion when segmenting slender targets with strong structural continuity (such as roads, rivers, and embankments) or complex features with weak boundaries. It is difficult to maintain the integrity and topological correctness of the feature outlines, which limits its practicality in high-precision application scenarios.

[0004] To capture a wider range of contextual information, attention mechanisms and Transformer structures have been introduced into segmentation models. However, these methods typically rely on a large number of parameters and intensive matrix operations, resulting in significant computational overhead and memory consumption. They are inefficient when processing large-format, high-resolution remote sensing images and cannot meet the actual deployment needs of resource-constrained or real-time requirements, presenting a prominent contradiction between lightweight and efficiency.

[0005] Furthermore, existing methods often use a single cross-entropy loss function for optimization during the training phase, which is not sensitive enough to the small-scale targets that are common in remote sensing images and the problem of severe imbalance in class distribution. This causes the model to easily ignore minority class samples during training, resulting in low recall rate for small targets, coarse segmentation boundaries, and overall model robustness and generalization ability being constrained.

[0006] Therefore, how to design a remote sensing image segmentation method based on multi-scale gated bottleneck convolutional scanning, effectively model the long-range structural dependence and global semantics in remote sensing images, and enhance the model's ability to segment slender targets, small-scale targets, and imbalanced scene types is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a remote sensing image segmentation method based on multi-scale gated bottleneck convolutional scanning, which aims to solve the common problems in high-resolution remote sensing image segmentation, such as fragmentation of slender targets, blurred boundaries, omission of small targets, and segmentation deviation caused by class imbalance. By constructing a lightweight encoder-decoder network that integrates multi-scale gated convolutional scanning, structure-aware scanning, and semantic bridging attention, high-precision and robust pixel-level segmentation of multi-scale and multi-morphological features can be achieved with limited computing resources.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A remote sensing image segmentation method based on multi-scale gated bottleneck convolutional scanning includes the following steps: S1. Input the remote sensing image and extract initial features through the initial convolutional layer; S2. The initial features are downsampled at multiple levels by the encoder, and each level is enhanced by a multi-scale gated bottleneck convolutional scanning unit to obtain multi-level encoded features. S3. At the deepest feature of the encoder, a semantic bridging attention unit is introduced to perform global semantic modeling and residual fusion to obtain enhanced deep semantic features. S4. In the decoder, upsampling is performed step by step, and the upsampled features are fused with the corresponding level's encoded features to obtain multi-level fused features. S5. Input the fused features output by the decoder into the segmentation head to obtain the predicted probability map.

[0009] Preferably, S2 includes: Initial features The downsampling features are obtained by processing the data through downsampling and multi-scale gated bottleneck convolutional scanning units, respectively. Enhanced features ; Initial features The downsampling features are obtained by processing the data through downsampling and multi-scale gated bottleneck convolutional scanning units, respectively. Enhanced features ; Initial features The downsampling features are obtained by processing the data through downsampling and multi-scale gated bottleneck convolutional scanning units, respectively. Enhanced features .

[0010] Preferably, in step S2, feature enhancement using a multi-scale gated bottleneck convolutional scanning unit includes: The input feature F is processed through two layers of light-rank bottleneck convolution and a single layer of light-rank bottleneck convolution, respectively. The outputs of the two convolutions are then multiplied element-wise to obtain the initial enhanced feature. ; For initial enhancement features The morphological enhancement features are obtained by performing additional gated bottleneck convolution processing and residual connection with the original input features F. ; Enhance morphological features Perform two nonlinear mappings, and multiply the outputs of both path element-wise to obtain the fused features. ; Fusion features Spatial attention weights are obtained by sequentially processing through linear layers, convolutional layers, and a softmax activation function, and then combined with morphological enhancement features. Element-wise multiplication is performed to obtain the fusion feature. ; Fusion features The final structure-enhanced features are output by processing through normalization and linear layers and then performing residual connections with themselves. .

[0011] Preferably, the two-way nonlinear mapping includes: In the first path, the first mapping feature is obtained through processing with a linear layer and the SiLU activation function. ; In the second path, the second mapping feature is obtained through linear layers, SiLU activation function, and 2D structure-aware scanning (SS2D) processing. . Preferably, the light-rank bottleneck convolution is composed of point convolution, depthwise convolution, and point convolution in sequence, and includes a normalization layer and a ReLU activation function.

[0012] Preferably, in step S3, introducing a semantic bridging attention unit for global semantic modeling and residual fusion includes: Downsampling features Perform multiple convolution mappings to obtain the query matrix Q, the key matrix K, and the value matrix V; Flatten the query matrix Q and the key matrix K into two-dimensional tensors, and calculate the attention weight matrix. , Indicates the dimension of the key vector; The attention features are obtained by weighted summation of the value matrix V. ; Attention features Perform upsampling processing and combine with deep features Residual connections yield enhanced semantic features. .

[0013] Preferably, S4 includes: Enhanced semantic features With corresponding coding features The fused features are obtained by adding elements one by one and then upsampling them. ; Fusion features With corresponding coding features The fused features are obtained by adding elements one by one and then upsampling them. ; Fusion features With corresponding coding features The multi-level fusion feature f is obtained by adding elements one by one and then upsampling.

[0014] Preferably, S5 includes: The fused features f are processed through multiple convolutional layers and the ReLU activation function to obtain high-dimensional features. ; High-dimensional features The predicted probability map P is obtained by mapping the data to C channels using a 1×1 convolutional layer and then normalizing the mapping result of each channel to a class probability using the Softmax function.

[0015] Preferably, it also includes using a composite loss function during training. :

[0016] in, , For adjustable weight parameters, For cross-entropy loss, This is a loss for Dice.

[0017] Preferably, the method further includes selecting the pixel index of the class with the highest probability from the predicted probability map P as the remote sensing image segmentation result. :

[0018] Where i and j represent the row index and column index of the pixel in the image, respectively, and c represents the category index.

[0019] As can be seen from the above technical solution, compared with the prior art, the technical solution of the present invention has the following beneficial effects: 1. By introducing multi-scale gated bottleneck convolutional scanning units in the encoding stage and utilizing the two-dimensional structure-aware scanning (SS2D) mechanism to establish spatial topological relationships between pixels along the horizontal, vertical and diagonal directions, the long-range structural dependencies of slender or tortuous features such as roads, rivers, and farmland boundaries in remote sensing images can be effectively captured and maintained. This significantly reduces the breakage and blurring of such targets in the segmentation results and improves the continuity and integrity of the boundaries.

[0020] 2. A lightweight gated bottleneck convolution consisting of point convolution and depthwise convolution is adopted as the basic operator, which reduces the number of model parameters and computational complexity. At the same time, semantic bridging attention units are introduced only at the bottleneck of the deepest layer of the encoder. By calculating global attention weights and residual fusion, the effective capture of contextual information of the whole image is achieved. While ensuring the efficiency of processing large-format remote sensing images, it alleviates the semantic gap problem in the encoding-decoding process and enhances the discrimination ability of complex ground feature boundary areas.

[0021] 3. During the training phase, a composite loss function combining cross-entropy loss and Dice loss is adopted. While optimizing the overall classification accuracy, it focuses on strengthening the structural constraints on target boundary regions and small-scale targets. Combined with the multi-level feature fusion strategy of gradually integrating shallow details and deep semantics during the decoding phase, this alleviates the training bias problem caused by insufficient samples of a few categories, and improves the overall segmentation accuracy and stability of the model under different land cover scales and category distributions. Attached Figure Description

[0022] 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.

[0023] Figure 1 A flowchart of a remote sensing image segmentation method based on multi-scale gated bottleneck convolutional scanning provided in an embodiment of the present invention; Figure 2 This is a diagram of the encoder-decoder network framework provided in an embodiment of the present invention; Figure 3 This is a structural framework diagram of a multi-scale gated bottleneck convolutional scanning unit provided in an embodiment of the present invention; Figure 4 This is a framework diagram of a semantic bridging attention unit provided in an embodiment of the present invention. Detailed Implementation

[0024] 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.

[0025] like Figure 1 and Figure 2As shown, this embodiment provides a remote sensing image segmentation method based on multi-scale gated bottleneck convolutional scanning, including the following steps: S1. Input the remote sensing image and extract initial features through the initial convolutional layer; S2. The initial features are downsampled at multiple levels by the encoder, and each level is enhanced by a multi-scale gated bottleneck convolutional scanning unit to obtain multi-level encoded features. S3. At the deepest feature of the encoder, a semantic bridging attention unit is introduced to perform global semantic modeling and residual fusion to obtain enhanced deep semantic features. S4. In the decoder, upsampling is performed step by step, and the upsampled features are fused with the corresponding level's encoded features to obtain multi-level fused features. S5. Input the fused features output by the decoder into the segmentation head to obtain the predicted probability map.

[0026] This method adopts a progressive structure of encoding-bottleneck enhancement-decoding. It enhances the ability to model long-range structures of slender features and complex boundaries by using multi-scale gated bottleneck convolutional scanning units. It achieves cross-scale global semantic fusion with low computational overhead by utilizing lightweight gated convolution and bottleneck layer semantic attention mechanism. Furthermore, it improves the robustness and boundary integrity of the model for segmenting small targets and imbalanced scene in remote sensing images by combining composite loss and multi-level feature decoding strategy.

[0027] The following provides a further detailed explanation of each step in the above method; In this embodiment, S1, the input remote sensing image is processed by an initial convolutional layer to extract initial features; In this step, the remote sensing image before being input into the model needs to undergo one or more preprocessing operations, specifically including: normalizing the image pixel values ​​to eliminate imaging differences and unify the data scale; enhancing the overall or local contrast of the image through histogram equalization to highlight the outline and texture of weak boundary features; and applying data augmentation strategies including random rotation, flipping, scaling, and color jitter during the training phase to expand the diversity of training samples and improve the model's generalization ability to different imaging conditions and land feature morphologies. The standardized image tensor after the above processing will be used as the input to the subsequent initial convolutional layer, laying the foundation for feature extraction.

[0028] In this embodiment S2, the initial features are downsampled at multiple levels by the encoder, and each level is enhanced by a multi-scale gated bottleneck convolutional scanning unit to obtain multi-level encoded features; including: Initial features The downsampling features are obtained by processing the data through downsampling and multi-scale gated bottleneck convolutional scanning units, respectively. Enhanced features ; Initial features The downsampling features are obtained by processing the data through downsampling and multi-scale gated bottleneck convolutional scanning units, respectively. Enhanced features ; Initial features The downsampling features are obtained by processing the data through downsampling and multi-scale gated bottleneck convolutional scanning units, respectively. Enhanced features .

[0029] Downsampling processing gradually reduces the spatial resolution of the feature map by using stride convolution or pooling operations in the encoder, expanding the receptive field to capture high-level semantic information and reducing computational complexity. By downsampling step by step and alternating with multi-scale gated bottleneck convolutional scanning units, a pyramid-shaped feature hierarchy from high-resolution details to low-resolution semantics is constructed. The model can simultaneously capture ground features at different scales in remote sensing images, providing multi-scale representation capabilities for identifying large-scale uniform areas and small isolated targets.

[0030] like Figure 3 As shown, feature enhancement using multi-scale gated bottleneck convolutional scanning units includes: The input feature F is processed through two layers of light-rank bottleneck convolution and a single layer of light-rank bottleneck convolution, respectively. The outputs of the two convolutions are then multiplied element-wise to obtain the initial enhanced feature. ; For initial enhancement features The morphological enhancement features are obtained by performing additional gated bottleneck convolution processing and residual connection with the original input features F. ; Enhance morphological features Perform two nonlinear mappings, and multiply the outputs of both path element-wise to obtain the fused features. ; Fusion features Spatial attention weights are obtained by sequentially processing through linear layers, convolutional layers, and a softmax activation function, and then combined with morphological enhancement features. Element-wise multiplication is performed to obtain the fusion feature. ; Fusion features The final structure-enhanced features are output by processing through normalization and linear layers and then performing residual connections with themselves. ; Furthermore, performing two-way nonlinear mapping includes: In the first path, the first mapping feature is obtained through processing with a linear layer and the SiLU activation function. ; In the second path, the second mapping feature is obtained through linear layers, SiLU activation function, and 2D structure-aware scanning (SS2D) processing. ; The first path uses a linear layer and the SiLU activation function to perform nonlinear transformation of the channel dimension, enhancing the expressive power of the features; the second path introduces two-dimensional structure-aware scanning (SS2D), which explicitly models the directional spatial dependence between pixels by performing sequential scanning in horizontal, vertical, and diagonal directions. The element-wise fusion of the features from both paths makes the unit output contain both rich semantic information and precise structural information. Furthermore, the light-rank bottleneck convolution is composed of point convolution, depthwise convolution, and point convolution in sequence, and includes a normalization layer and a ReLU activation function.

[0031] The aforementioned multi-scale gated bottleneck convolutional scanning unit employs a series of operations, including lightweight convolution, residual enhancement, dual-path mapping fusion, spatial attention weighting, and residual output. First, it reduces computational burden by utilizing lightweight convolution. Then, it captures long-range spatial topology through a structure-aware path. Finally, it adaptively focuses on important regions using an attention mechanism. This design achieves a balance between receptive field expansion, structural continuity preservation, and feature discriminative power enhancement.

[0032] In this embodiment S3, at the deepest feature of the encoder, a semantic bridging attention unit is introduced to perform global semantic modeling and residual fusion to obtain enhanced deep semantic features. like Figure 4 As shown, the introduction of semantic bridging attention units for global semantic modeling and residual fusion includes: Downsampling features Perform multiple convolution mappings to obtain the query matrix Q, the key matrix K, and the value matrix V; Flatten the query matrix Q and the key matrix K into two-dimensional tensors, and calculate the attention weight matrix. , Indicates the dimension of the key vector; The attention features are obtained by weighted summation of the value matrix V. ; Attention features Perform upsampling processing and combine with deep features Residual connections yield enhanced semantic features. .

[0033] The semantic bridging attention unit operates on the deepest features of the encoder, generating Query, Key, and Value through convolution and calculating global attention weights to model the dependencies between pixels across the entire image. Finally, through upsampling and residual connections, the captured global contextual information is seamlessly fused back into the original feature map, effectively compensating for the semantic information loss and spatial relationship fading that may occur due to multiple downsampling.

[0034] In this embodiment, S4 involves progressive upsampling in the decoder, and then fusing the upsampled features with the corresponding level's encoded features to obtain multi-level fused features; including: Enhanced semantic features With corresponding coding features The fused features are obtained by adding elements one by one and then upsampling them. ; Fusion features With corresponding coding features The fused features are obtained by adding elements one by one and then upsampling them. ; Fusion features With corresponding coding features The multi-level fusion feature f is obtained by adding elements one by one and then upsampling.

[0035] Upsampling processing gradually restores the spatial resolution of the feature map in the decoder through methods such as deconvolution or bilinear interpolation, and performs skip connections or element-wise fusion with the features of the corresponding level in the encoder, thereby integrating shallow texture and deep semantics while reconstructing segmentation details, and achieving accurate modeling and complete restoration of multi-scale land cover structures in remote sensing images. By progressively upsampling deep semantic features and adding them element-wise with the corresponding enhanced shallow features in the encoder, cross-layer fusion of high-level semantics and low-level details is achieved, which helps to restore the spatial details and boundary clarity of the target while retaining rich category discrimination information.

[0036] In this embodiment, S5, the fused features output from the decoder are input into the segmentation head to obtain a predicted probability map; including: The fused features f are processed through multiple convolutional layers and the ReLU activation function to obtain high-dimensional features. ; High-dimensional features The predicted probability map P is obtained by mapping the data to C channels using a 1×1 convolutional layer and then normalizing the mapping result of each channel to a class probability using the Softmax function.

[0037] Furthermore, a composite loss function is used during training. :

[0038] in, , For adjustable weight parameters, For cross-entropy loss, This is a loss for Dice.

[0039] By combining the cross-entropy loss (L_CE), which focuses on the overall pixel classification accuracy, with the Dice loss (L_Dice), which focuses on the structural integrity of the target region (especially small targets), this loss function can simultaneously drive the model to optimize the global classification accuracy and the local boundary fitting quality, effectively mitigating the impact of the common class imbalance problem in remote sensing images on the segmentation performance of small targets. Furthermore, the network optimizes all learnable parameters through backpropagation. Specifically, it uses gradient descent or its variants (such as the Adam optimizer) to calculate the gradient of the loss and updates the weights of the encoder, multi-scale gated bottleneck convolutional scanning unit, semantic bridging attention unit, and segmentation head to minimize the difference between the predicted result P and the true label Y.

[0040] Furthermore, it also includes selecting the pixel index of the class with the highest probability as the remote sensing image segmentation result from the predicted probability map P. :

[0041] Where i and j represent the row index and column index of the pixel in the image, respectively, and c represents the category index; By comparing the probabilities of all categories at each pixel location and selecting the category index with the highest probability as the predicted label for that pixel, the mapping from continuous probability values ​​to discrete classification results is completed.

[0042] The remote sensing image segmentation method based on multi-scale gated bottleneck convolutional scanning in this embodiment, when practically applied to large-format high-resolution remote sensing image segmentation, works as follows: First, the pre-processed remote sensing image is input and downsampled at multiple levels by an encoder. Each level embeds a multi-scale gated bottleneck convolutional scanning unit. This unit uses lightweight gated convolution to extract local features and establishes long-range topological relationships between pixels in the horizontal, vertical and diagonal directions through two-dimensional structure-aware scanning SS2D. In this way, while gradually extracting deep semantics, the structural continuity of slender features such as roads and rivers is explicitly maintained. Subsequently, a semantic bridging attention unit is introduced at the deepest feature of the encoder. By calculating global attention weights and fusion with residuals, the model’s ability to understand the contextual information of the whole image is enhanced, effectively alleviating the problem of semantic information dilution caused by downsampling. During the decoding stage, by progressively upsampling and fusing with the enhanced features of the corresponding level of the encoder, the deep semantics and shallow details are complemented, and finally the pixel-level category probability map is output by the segmentation head. This workflow is specifically designed to address the characteristics of diverse feature scales, complex boundaries, numerous small targets, and uneven categories in remote sensing imagery. By organically combining lightweight local feature extraction, explicit long-range structure modeling, centralized global attention, and multi-level feature fusion, it improves the integrity of feature boundaries, the continuity of slender targets, and the detection rate of small-scale targets in the segmentation results while ensuring model computational efficiency and deployment feasibility. It can effectively support practical application scenarios with stringent requirements for segmentation accuracy and reliability, such as land resource surveys, urban fine planning, and disaster assessment.

[0043] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0044] 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 multi-scale gated bottleneck convolution scanning, characterized in that, The method comprises the following steps: S1, inputting a remote sensing image, and extracting initial features through an initial convolutional layer; S2, performing multi-level down-sampling processing on the initial features through an encoder, and performing feature enhancement on each level through a multi-scale gated bottleneck convolution scanning unit in sequence to obtain multi-level encoded features; S3, introducing a semantic bridging attention unit at the deepest layer of the features of the encoder to perform global semantic modeling and residual fusion, and obtaining enhanced deep semantic features; S4, performing gradual up-sampling processing in the decoder, and fusing the features after the up-sampling processing with the encoded features of the corresponding level to obtain multi-level fused features; S5, inputting the fused features output by the decoder into a segmentation head to obtain a prediction probability map.

2. The remote sensing image segmentation method based on multi-scale gating bottleneck convolution scanning according to claim 1, characterized in that, The S2 comprises: The initial features are obtained by scanning the input image through a plurality of convolutional layers The down-sampling processed features are obtained by scanning the input image through a down-sampling and a multi-scale gated bottleneck convolution respectively The enhanced features are obtained by scanning the input image through a plurality of convolutional layers ​ The initial features are obtained by processing the input image through a convolutional neural network (CNN) to obtain a plurality of initial features The down-sampling processed features are obtained by processing the scanning unit through down-sampling and multi-scale gated bottleneck convolution scanning unit processing, respectively , enhanced features ; The initial features are obtained by processing the input image through a convolutional neural network (CNN) to obtain a plurality of feature maps The down-sampling processed features are obtained by processing the scanning unit through down-sampling and multi-scale gated bottleneck convolution respectively , enhanced features . 3.The method of claim 1, wherein, In the S2, the feature enhancement through the multi-scale gated bottleneck convolution scanning unit comprises: The input feature F is processed by two layers of light-rank bottleneck convolution and a single layer of light-rank bottleneck convolution respectively, and the outputs of the two are multiplied element by element to obtain an initial enhanced feature ; To the initial enhanced feature Through the additional gating bottleneck convolution processing, and residual connection with the original input feature F, the morphology enhanced feature is obtained ; morphological enhancement features performing two-way nonlinear mapping, element-wise multiplying the outputs of both to obtain fused features ; The fusion feature is obtained by sequentially processing the linear layer, the convolution layer and a Softmax activation function The spatial attention weight is obtained by sequentially processing the linear layer, the convolution layer and a Softmax activation function Element-wise multiplication is performed to obtain the fusion feature ; The fused features The final structure-enhanced features are output by processing through a normalization layer and a linear layer and performing residual connection with itself .

4. The remote sensing image segmentation method based on multi-scale gated bottleneck convolution scanning according to claim 3, characterized in that, The two-way nonlinear mapping comprises: In the first path, the first mapping feature is obtained through linear layer and SiLU activation function processing ; In the second path, a second mapping feature is obtained through a linear layer, a SiLU activation function, and a two-dimensional structure perception scanning SS2D processing .

5. The method of claim 3, wherein the method is characterized by, The light rank bottleneck convolution is sequentially composed of a point convolution, a depth convolution, and a point convolution, and comprises a normalization layer and a ReLU activation function.

6. The method of claim 1, wherein, In the S3, the introduction of the semantic bridging attention unit for global semantic modeling and residual fusion comprises: Down-sampling processed features Multiple convolution mappings are performed to obtain a query matrix Q, a key matrix K, and a value matrix V. The query matrix Q and the key matrix K are flattened into two-dimensional tensors, and an attention weight matrix is calculated , denotes the dimension of the key vector; performing weighted sum on the value matrix V to obtain an attention feature ; Attention feature Up-sampling is performed, and the deep feature Residual connection is performed to obtain an enhanced semantic feature .

7. The method of claim 1, wherein, The S4 comprises: enhanced semantic feature corresponding encoded feature element-wise addition and up-sampling to obtain the fused feature ; Fusing features with corresponding encoded features element-wise addition and up-sampling to obtain fused features ; fused features corresponding encoded features element-wise addition and up-sampling to obtain multi-level fused features f.

8. The method of claim 1, wherein, The S5 comprises: The fusion feature f is processed through a plurality of convolution layers and a ReLU activation function to obtain a high-dimensional feature ; map high-dimensional features mapped to C channels by a 1x1 convolutional layer, and the mapping result of each channel is normalized to a class probability by a softmax function to obtain a prediction probability map P.

9. The remote sensing image segmentation method based on multi-scale gated bottleneck convolution scanning according to claim 1, characterized in that, Also included is employing a composite loss function during training : wherein, , is an adjustable weight parameter, is a cross-entropy loss, is a Dice loss.

10. The remote sensing image segmentation method based on multi-scale gated bottleneck convolution scanning according to claim 1, characterized in that, Also included is selecting the pixel index of the maximum probability class as the remote sensing image segmentation result for the predicted probability map P : Wherein, i, j respectively represent the row index and column index of the pixels in the image, and c represents the category index.

Images