A remote sensing image segmentation method combining location extraction and pulse channel enhancement

Abstract

The application discloses a remote sensing image segmentation method combining positioning extraction and pulse channel enhancement, and relates to the field of remote sensing image segmentation, and comprises the following steps: using a positioning feature extraction network to extract spatial and position information of a remote sensing image, and generating a first enhanced feature map by fusing multi-level features; based on the first enhanced feature map, using a pulse channel feature enhancement network to perform multi-scale feature enhancement, and generating a second enhanced feature map; based on the second enhanced feature map, using a remote sensing image segmentation network to perform category prediction, and outputting a segmentation result of the remote sensing image. Through positioning feature extraction, pulse channel enhancement and compound loss optimization, the method can be applied to high-resolution remote sensing image processing and resource-limited environments, and the precision and efficiency of remote sensing image segmentation are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of remote sensing image segmentation, more particularly to a remote sensing image segmentation method combining positioning extraction and pulse channel enhancement. BACKGROUND

[0002] Remote sensing image segmentation is a core task in remote sensing technology, widely used in environmental monitoring, urban planning, agricultural resource management, disaster assessment and other fields. With the development of remote sensing technology, high-resolution images provide rich ground feature detail information, but also face challenges such as complex scenes, multi-scale target distribution and dynamic interference, which put higher requirements on the accuracy and efficiency of segmentation algorithms.

[0003] Traditional remote sensing image segmentation methods mostly rely on hand-designed features or shallow convolutional networks, which cannot fully extract fine spatial information from high-resolution images, especially in complex scenes, where edge details are easily lost. In addition, existing methods have poor adaptability to multi-scale targets and cannot effectively capture the features of different size ground objects. For example, small-scale targets and large-scale targets have significant differences in feature expression, which limits the segmentation accuracy.

[0004] In addition, advanced methods based on deep learning, including deep convolutional networks and Transformer models, have improved segmentation accuracy, but their model complexity is high, and the training and inference process consumes a large amount of computing resources, making it difficult to deploy to edge devices or real-time processing scenarios. At the same time, existing methods have limited perception of dynamic changes and lack efficient channel feature enhancement mechanisms, resulting in insufficient robustness of segmentation results under dynamic interference.

[0005] Therefore, how to design a remote sensing image segmentation method combining positioning extraction and pulse channel enhancement can effectively solve the defects of low segmentation accuracy, high resource consumption and insufficient dynamic response in the prior art, which is a problem that needs to be solved by those skilled in the art. SUMMARY

[0006] Therefore, the present application provides a remote sensing image segmentation method combining positioning extraction and pulse channel enhancement, which realizes accurate capture and enhancement of multi-scale features, significantly improves segmentation accuracy. At the same time, it greatly reduces the demand for computing resources, so that the method can run efficiently in a resource-limited environment while maintaining high accuracy, providing a flexible and powerful solution for remote sensing image segmentation in complex scenarios.

[0007] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0008] The present application provides a remote sensing image segmentation method combining positioning extraction and pulse channel enhancement, comprising the following steps:

[0009] S1, extracting spatial and position information of the remote sensing image by using a positioning feature extraction network to generate a first enhanced feature map fusing multi-level features;

[0010] S2, based on the first enhanced feature map, performing multi-scale feature enhancement by using a pulse channel feature enhancement network to generate a second enhanced feature map;

[0011] S3, based on the second enhanced feature map, performing category prediction by using a remote sensing image segmentation network to output a segmentation result of the remote sensing image.

[0012] Preferably, the S1 comprises:

[0013] S11, inputting the remote sensing image into a Mamba pre-training model to extract initial high-level semantic features

[0014] S12, performing low-level feature extraction and stabilization processing on the initial high-level semantic features by using a convolution layer and a normalization layer to obtain normalized features

[0015] S13, sequentially injecting position encoding, performing full-dimensional dynamic convolution and linear projection on the normalized features to generate intermediate features α, β and γ containing spatial perception;

[0016] S14, fusing the intermediate features α, β and γ and the normalized features by using spatial relationship modeling and skip link fusion to obtain fused features

[0017] S15, extracting complex spatial features by using a convolution layer and a windmill-shaped convolution layer on the fused features , and performing key position sensitivity enhancement by using a position trigger;

[0018] S16, performing nonlinear conversion on the processing results of the windmill-shaped convolution layer and the position trigger by using a ReLU activation function, and fusing to obtain a first enhanced feature map

[0019] Preferably, the S14 comprises:

[0020] , multiplying the transposed and reshaped intermediate features α and β to obtain a spatial correlation matrix, and normalizing the spatial correlation matrix by using a softmax activation function to obtain a normalized spatial correlation matrix;

[0021] , multiplying the normalized spatial correlation matrix and the reshaped intermediate feature γ to obtain features enhanced in spatial information​​

[0022] the spatial information enhanced feature is reshaped and fused with the normalized feature to obtain a fused feature

[0023] Preferably, the position trigger comprises two convolution layers, an activation function and a full connection layer.

[0024] Preferably, the S2 comprises:

[0025] S21, inputting the first enhanced feature map into a first pulse neuron layer, a second pulse neuron layer and a convolution layer respectively, and reshaping to obtain features δ, ∈ and ζ;

[0026] S22, based on the features δ, ∈ and ζ, performing pulse feature weighting and spatial relationship modeling to obtain a feature map

[0027] S23, fusing the feature map with the first enhanced feature map to obtain a feature map

[0028] S24, performing channel trigger, dilated convolution layer to enhance channel response capability and expand receptive field, and applying depth separable convolution layer to perform depth separable convolution on the feature map

[0029] S25, fusing the processing results of the dilated convolution layer and the depth separable convolution layer to obtain a second enhanced feature map

[0030] Preferably, the S22 comprises:

[0031] multiplying the features ∈ and ζ and normalizing through a Softmax activation function to obtain relationship weights between channels;

[0032] multiplying the relationship weights with the feature δ and reshaping to obtain a feature map

[0033] Preferably, the channel trigger comprises a global average pooling layer, a full connection layer, a convolution layer and an activation function.

[0034] Preferably, in the S3, a remote sensing image segmentation network is used for class prediction, comprising:

[0035] inputting the second enhanced feature map into a convolution layer and an up-sampling layer to obtain a class prediction map ​

[0036] Class prediction map by Softmax activation function Class assignment is performed to obtain probability distribution of each pixel point And a prediction probability map is generated.

[0037] Preferably, the remote sensing image segmentation network adopts a composite loss function; the composite loss function is a weighted combination of a cross-entropy loss function and a Dice loss function, and is used to balance the pixel-level classification accuracy and the segmentation region overlap degree.

[0038] Preferably, in S3, further comprising:

[0039] The prediction probability map output by the remote sensing image segmentation network is thresholded to obtain a corresponding binary segmentation result.

[0040] According to the above technical solution, compared with the prior art, the technical solution of the present application has the following

[0041] Beneficial effects:

[0042] 1. The positioning feature extraction network extracts initial high-level semantic features through a pre-trained model, combines dynamic convolution, position coding and spatial relationship modeling, can dynamically adjust the size of the convolution kernel to adapt to different scale targets, and can retain low-level details and high-level semantic information through a skip link fusion. It significantly improves the model's perception ability in complex scenes, solves the problem of insufficient accuracy in multi-scale target segmentation, and is especially suitable for fine processing needs of high-resolution remote sensing images.

[0043] 2. The pulse channel feature enhancement network enhances the dynamic response capability by introducing a pulse neuron layer, and uses dilated convolution to expand the receptive field to capture long-distance spatial dependence. Combined with depth separable convolution to optimize the calculation efficiency, it can greatly reduce the calculation complexity while ensuring the feature enhancement effect. It can adapt to the real-time processing needs of edge computing devices and meet the efficient deployment in low storage environment and resource limited scenarios.

[0044] 3. In the remote sensing image segmentation network, the weighted combination of cross-entropy loss and Dice loss is used to optimize the class imbalance problem, and the adaptive threshold adjustment strategy is used to convert the probability map into a binary segmentation mask. It effectively alleviates the challenge of uneven distribution of ground object classes in remote sensing images, and improves the robustness and practicality of the segmentation result. BRIEF DESCRIPTION OF DRAWINGS

[0045] 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 only constitute a part of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on the provided drawings.

[0046] Figure 1 A remote sensing image segmentation process diagram combining positioning feature extraction and pulse channel enhancement is provided for the embodiments of the present application.

[0047] Figure 2 A first enhanced feature map generation process diagram using a positioning feature extraction network is provided for the embodiments of the present application.

[0048] Figure 3 A position trigger structure framework diagram is provided for the embodiments of the present application.

[0049] Figure 4 A second enhanced feature map generation process diagram using a pulse channel feature enhancement network is provided for the embodiments of the present application.

[0050] Figure 5 A channel trigger structure framework diagram is provided for the embodiments of the present application.

[0051] Figure 6 A remote sensing image segmentation method flowchart in an agricultural land classification scenario is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute a 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 skilled in the art without creative labor fall within the scope of protection of the present application.

[0053] Embodiment 1;

[0054] As shown in the drawings, Figure 1 The present embodiment provides a remote sensing image segmentation method combining positioning feature extraction and pulse channel enhancement, comprising the following steps:

[0055] S1, using a positioning feature extraction network to extract spatial and position information of a remote sensing image, and generating a first enhanced feature map fusing multi-level features;

[0056] S2, based on the first enhanced feature map, using a pulse channel feature enhancement network to perform multi-scale feature enhancement, and generating a second enhanced feature map;

[0057] S3. Based on the second enhanced feature map, a remote sensing image segmentation network is used to predict the category and output the segmentation result of the remote sensing image.

[0058] This method enhances the model's ability to perceive complex scenes by fusing dynamic convolution, position encoding, and spatial relationship modeling of the localization feature extraction network, thus solving the problem of insufficient accuracy in multi-scale target segmentation. The pulse channel feature enhancement network optimizes feature expression through spiking neuron layers and dilated convolution, achieving efficient computation and real-time processing. The remote sensing image segmentation network adopts a composite loss function and adaptive threshold adjustment, effectively alleviating the class imbalance problem and improving the robustness and practicality of the segmentation results. It is particularly suitable for the fine processing of high-resolution remote sensing images and efficient deployment in resource-constrained environments.

[0059] The following provides a further detailed explanation of each step in the above method:

[0060] like Figure 2 As shown, in this embodiment S1, a localization feature extraction network is used to extract spatial and location information from remote sensing images, generating a first enhanced feature map that fuses multi-level features; specifically, it includes:

[0061] S11. Input the remote sensing image into the Mamba pre-trained model and extract the initial high-level semantic features. The model's pre-training on remote sensing images enables it to fully capture global and local structural information, providing a highly robust semantic foundation for subsequent operations;

[0062] S12. Apply convolutional layers and normalization layers to the initial high-level semantic features. Low-level feature extraction and stabilization are performed to obtain normalized features.

[0063] S13. Normalized features Positional encoding is injected sequentially, full-dimensional dynamic convolution and linear projection are performed to generate intermediate features α, β and γ that contain spatial awareness;

[0064] S14. Utilize spatial relationship modeling and skip link fusion to integrate intermediate features α, β, and γ with the normalized features. Perform fusion to obtain the fused features

[0065] S15. Features after fusion The complex spatial features are extracted through the convolution layer and the windmill-shaped convolution layer, and the key position sensitivity is enhanced through the position triggerer; the windmill-shaped convolution layer and the position triggerer are introduced, the former extracts complex spatial features through a multi-directional convolution kernel, and the latter weights the key region through an activation function, thereby enhancing the sensitivity of the model to the edges and details of the ground objects, so that more accurate spatial feature modeling is realized in a complex scene;

[0066] S16, the processing results of the windmill-shaped convolution layer and the position triggerer are subjected to nonlinear conversion through a ReLU activation function and are fused to obtain a first enhanced feature map

[0067] It should be noted that the convolution layer in the embodiment is a standard convolution layer, which adopts a fixed-size 3*3 or 5*5 convolution kernel and a stride of 1 for extracting local spatial features.

[0068] Further S14, comprising:

[0069] The intermediate features a and β are multiplied after being transposed and reshaped respectively to obtain a spatial correlation matrix, and the spatial correlation matrix is normalized through a softmax activation function to obtain a normalized spatial correlation matrix;

[0070] The normalized spatial correlation matrix is multiplied by the reshaped intermediate feature γ to obtain a spatial information enhanced feature

[0071] The spatial information enhanced feature is reshaped and fused with the normalized feature to obtain a fused feature

[0072] The position relationship is modeled through the spatial correlation matrix, the dynamic spatial dependency relationship of the intermediate features a, β and γ is explicitly coded through the transposition, reshaping and softmax normalization operations, and the difficulty of modeling the long-distance target correlation in the remote sensing image is solved. When fusing the normalized features, a skip link is used to effectively retain the detail information of the initial features, avoid the gradient vanishing problem in the training of the deep network, and further improve the stability of the feature fusion and the continuity of the segmentation result.

[0073] As shown in Figure 3 , the position triggerer comprises two convolution layers, an activation function and a fully connected layer. The local position features are extracted through the convolution layer, the global weight distribution is learned through the fully connected layer, and the response strength of the key position is dynamically adjusted through the activation function. The positioning ability of the model to the sparse distributed targets in the remote sensing image is enhanced.

[0074] As shown in Figure 4As shown, in the embodiment S2, based on the first enhanced feature map, multi-scale feature enhancement is performed by using the pulse channel feature enhancement network to generate a second enhanced feature map; specifically including:

[0075] S21, the first enhanced feature map are input into the first pulse neuron layer, the second pulse neuron layer and the convolution layer respectively, and are reshaped to obtain features δ, ∈ and ζ; the first enhanced feature map is processed by the pulse neuron layer, and the model for enhancing the time sequence characteristics of the pulse signal is used to adapt to dynamic changes, thereby solving the problem of insufficient response of the traditional convolution network to time sequence characteristics;

[0076] S22, based on the features δ, ∈ and ζ, pulse feature weighting and spatial relationship modeling are performed to obtain a feature map

[0077] S23, the feature map is combined with the first enhanced feature map by jump link fusion to obtain a feature map

[0078] S24, the feature map is processed by the channel trigger and the dilated convolution layer to enhance the channel response capability and expand the receptive field, and deep separable convolution is performed by applying the deep separable convolution layer; the dilated convolution and the deep separable convolution are combined, the former captures long-distance dependence by expanding the receptive field, and the latter reduces the amount of calculation by decomposing the convolution operation, thereby significantly reducing resource consumption while ensuring multi-scale feature extraction capability;

[0079] S25, the processing results of the dilated convolution layer and the deep separable convolution layer are fused to obtain a second enhanced feature map

[0080] Further, S22 includes: multiplying the features ∈ and ζ and normalizing them by the Softmax activation function to obtain the relationship weight between channels; multiplying the relationship weight and the feature δ and obtaining a feature map by reshaping

[0081] By normalizing the relationship weight between channels by Softmax and combining the weighted fusion of the feature δ, adaptive enhancement of key channels is realized, which aims to highlight the distinguishing features of different ground object categories in the remote sensing image, thereby improving the class distinction degree of the segmentation result in a complex scene.

[0082] As Figure 5As shown, the channel trigger includes a global average pooling layer, a fully connected layer, a convolution layer, and an activation function. The channel trigger includes a global average pooling layer and a fully connected layer, the former compresses global spatial information to generate a channel descriptor, and the latter learns the nonlinear relationship between channels, combines the convolution layer to dynamically adjust the channel weight, effectively suppresses the interference of redundant channels, and strengthens the model for high-frequency features in the remote sensing image.

[0083] In this embodiment S3, based on the second enhanced feature map, a remote sensing image segmentation network is used for class prediction to output a segmentation result of the remote sensing image.

[0084] In this step, first, a remote sensing image segmentation network is used for class prediction, including:

[0085] The second enhanced feature map is input into a convolution layer and an upsampling layer to obtain a class prediction map The class prediction map is activated by a Softmax activation function Class assignment is performed to obtain the probability distribution of each pixel point and a prediction probability map is generated. It restores the resolution of the feature map through the upsampling layer, combines the Softmax activation to generate the pixel-level probability distribution, and ensures that the segmentation result is consistent with the original image size.

[0086] Further, the remote sensing image segmentation network adopts a composite loss function; the composite loss function is a weighted combination of a cross-entropy loss function and a Dice loss function, which is used to balance the pixel-level classification accuracy and the segmentation region overlap. In the composite loss function, the cross-entropy loss is used to optimize the pixel-level classification accuracy, and the weighted combination of the Dice loss is used to improve the region consistency by measuring the overlap, which solves the model bias problem caused by the class imbalance in the remote sensing image. Combined with the dynamic adjustment of the weight parameter, it can flexibly adapt to the needs of different data sets and enhance the generalization ability of the model.

[0087] Further, the prediction probability map output by the remote sensing image segmentation network is thresholded to obtain the corresponding binary segmentation result.

[0088] The embodiment proposes a remote sensing image segmentation method combining positioning extraction and pulse channel enhancement. The spatial and positional information of the remote sensing image is captured through the positioning feature extraction network, the multi-scale feature enhancement is performed through the pulse channel feature enhancement network, and finally the segmentation result is output through the remote sensing image segmentation network. This method combines advanced technologies such as dynamic convolution, position coding, pulse neuron layer, and composite loss function, effectively improves the accuracy and efficiency of remote sensing image segmentation, and is particularly suitable for fine processing of high-resolution images and efficient deployment in resource-constrained environments.

[0089] Embodiment 2;

[0090] In this embodiment, a remote sensing image segmentation method based on localization extraction and pulse channel enhancement is specifically applied to the agricultural land classification scenario.

[0091] Agricultural land classification is an important task for precision agriculture management, crop monitoring, and land resource assessment. High-resolution remote sensing images can provide detailed information such as farmland boundaries, crop types, and irrigation facilities. However, due to the large differences in farmland scale (such as small plots of farmland versus large-scale planting areas), complex crop textures, and seasonal variations, traditional segmentation methods are easily affected by dynamic interference, leading to blurred classification boundaries and category confusion.

[0092] like Figure 6 As shown, the specific steps of this method include:

[0093] 1) Agricultural remote sensing image preprocessing and localization feature extraction;

[0094] First, multispectral remote sensing images are input and subjected to radiometric correction, geometric registration, and normalization to preserve near-infrared and red-edge bands sensitive to vegetation, thereby enhancing the distinguishability of crop features. Next, the processed images are input into a Mamba pre-trained model to extract global farmland distribution and local plot boundary features.

[0095] For agricultural scenarios, prior information about farmland grids is introduced into the location encoding to enhance the model's ability to perceive the boundaries of regular farmland; and multi-scale convolution kernels are set in the full-dimensional dynamic convolution to capture the spatial continuity features of small plots of farmland texture and large-scale planting areas respectively.

[0096] 2) Pulse channel feature enhancement and multi-scale fusion;

[0097] In this step, a time decay factor is introduced into the spiking neuron layer to enhance the model's dynamic response to seasonal changes, taking into account the temporal characteristics of crop growth. Furthermore, channel triggers are used to strengthen the weights of vegetation index-related channels, suppressing interference from cloud cover or shaded areas.

[0098] In addition, depthwise separable convolution is used instead of standard convolution to reduce computational load while ensuring multi-scale feature extraction, thus adapting to the deployment needs of drone edge devices.

[0099] 3) Classification and post-processing of agricultural land;

[0100] In remote sensing image segmentation networks, adaptive weights are set for agricultural scene categories (such as wheat, corn, fallow land, and water bodies) to balance the problem of imbalanced samples. A boundary loss term is added to the composite loss function to enhance the segmentation accuracy of farmland boundaries.

[0101] Further, the output probability map is subjected to a morphological closing operation to fill small holes and smooth the boundaries; and in combination with the prior knowledge of the farmland vector map, abnormal areas in the segmentation result are corrected.

[0102] The remote sensing image segmentation method based on positioning extraction and pulse channel enhancement of the embodiment can be applied to agricultural resource survey, crop growth monitoring and precision irrigation planning, and is particularly suitable for high-precision and high-efficiency farmland plot segmentation and classification tasks in multispectral remote sensing images.

[0103] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0104] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image segmentation method combining localization extraction and pulse channel enhancement, characterized in that, Includes the following steps: S1. Using a localization feature extraction network, extract spatial and location information from remote sensing images to generate a first enhanced feature map that integrates multi-level features; S2. Based on the first enhanced feature map, multi-scale feature enhancement is performed using a pulse channel feature enhancement network to generate a second enhanced feature map; including: S21. The first enhanced feature map The first spiking neuron layer, the second spiking neuron layer, and the convolutional layer are input respectively, and the features δ, ∈, and ζ are obtained by reshaping. S22. Based on features δ, ∈, and ζ, perform pulse feature weighting and spatial relationship modeling to obtain feature maps. S23, Feature map Compared with the first enhanced feature map Perform skip link fusion to obtain feature maps S24. Feature Map Channel response capability and receptive field are enhanced by channel triggers and dilated convolutional layers, and depth-separable convolution is performed by applying depth-separable convolutional layers. S25. Fuse the results of the dilated convolutional layer and the depthwise separable convolutional layer to obtain the second enhanced feature map. S3. Based on the second enhanced feature map, a remote sensing image segmentation network is used to predict the category and output the segmentation result of the remote sensing image.

2. The remote sensing image segmentation method combining localization extraction and pulse channel enhancement according to claim 1, characterized in that, S1 includes: S11. Input the remote sensing image into the Mamba pre-trained model and extract the initial high-level semantic features. S12. Apply convolutional layers and normalization layers to the initial high-level semantic features. Low-level feature extraction and stabilization are performed to obtain normalized features. S13. Normalized features Positional encoding is injected sequentially, full-dimensional dynamic convolution and linear projection are performed to generate intermediate features α, β and γ that contain spatial awareness; S14. Utilize spatial relationship modeling and skip link fusion to integrate intermediate features α, β, and γ with the normalized features. Perform fusion to obtain the fused features S15. Features after fusion Complex spatial features are extracted through convolutional layers and windmill-shaped convolutional layers, and key location sensitivity is enhanced through location triggers. S16. The results of the windmill-shaped convolutional layer and the position trigger are processed by the ReLU activation function for nonlinear transformation and then fused to obtain the first enhanced feature map.

3. The remote sensing image segmentation method combining localization extraction and pulse channel enhancement according to claim 2, characterized in that, S14 includes: The intermediate features α and β are transposed and reshaped respectively, and then multiplied to obtain the spatial correlation matrix. The spatial correlation matrix is ​​then normalized using the softmax activation function to obtain the normalized spatial correlation matrix. Multiply the normalized spatial correlation matrix by the reshaped intermediate feature γ to obtain the feature after spatial information enhancement. Features enhanced with spatial information Reshape and compare with normalized features Perform fusion to obtain the fused features 4. The remote sensing image segmentation method combining localization extraction and pulse channel enhancement according to claim 2, characterized in that, The location trigger includes: two convolutional layers, an activation function, and a fully connected layer.

5. The remote sensing image segmentation method combining localization extraction and pulse channel enhancement according to claim 1, characterized in that, S22 includes: Multiply the features ∈ and ζ and normalize them using the Softmax activation function to obtain the relationship weights between channels; The relation weights are multiplied by the feature δ, and the feature map is obtained by reshaping.

6. The remote sensing image segmentation method combining localization extraction and pulse channel enhancement according to claim 1, characterized in that, The channel trigger includes: a global average pooling layer, a fully connected layer, a convolutional layer, and an activation function.

7. The remote sensing image segmentation method combining localization extraction and pulse channel enhancement according to claim 1, characterized in that, In step S3, category prediction is performed using a remote sensing image segmentation network, including: The second enhanced feature map is input into the convolutional layer and the upsampling layer to obtain the class prediction map. The category prediction map is obtained by using the Softmax activation function. Perform category assignment to obtain the probability distribution of each pixel. And generate a predicted probability map.

8. The remote sensing image segmentation method combining localization extraction and pulse channel enhancement according to claim 1, characterized in that, The remote sensing image segmentation network employs a composite loss function, which is a weighted combination of the cross-entropy loss function and the Dice loss function, used to balance pixel-level classification accuracy and segmentation region overlap.

9. The remote sensing image segmentation method combining localization extraction and pulse channel enhancement according to claim 1, characterized in that, S3 also includes: Thresholding is performed on the predicted probability map output by the remote sensing image segmentation network to obtain the corresponding binary segmentation result.

Images