A remote sensing image semantic segmentation method and system based on an AsMamba importance perception architecture
By employing an importance-aware state-space model, the problems of class imbalance, scanning strategy limitations, and insufficient representation of key areas in remote sensing image segmentation were solved, achieving high-precision and efficient remote sensing image segmentation, particularly with significant segmentation results in rocky desertification monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF TECH
- Filing Date
- 2026-05-15
- Publication Date
- 2026-07-14
Smart Images
Figure CN122391648A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of remote sensing image processing and computer vision, specifically to a method and system for imbalance-like remote sensing image segmentation based on an importance-aware state-space model. Background Technology Semantic segmentation of remote sensing images is of great significance in land cover monitoring, resource surveys, and disaster assessment. Especially in the scenario of monitoring desertification in karst landforms, remote sensing images exhibit extremely high spatial heterogeneity, often containing a complex mixture of exposed rocks, shallow soil, and sparse vegetation.
[0002] In existing technologies, convolutional neural network-based methods are limited by local receptive fields, making it difficult to capture long-range spatial dependencies. They also fall short in distinguishing objects with similar spectral characteristics (such as bare soil and rocky desertification areas) or dealing with complex terrain occlusion, easily leading to boundary fragmentation. While Transformer-based methods possess global modeling capabilities, the computational complexity of their self-attention mechanism increases quadratically when processing high-resolution remote sensing images, resulting in extremely high training and deployment costs.
[0003] In recent years, state-space models have attracted attention due to their linear computational complexity and efficient sequence modeling capabilities. However, existing Mamba models still face the following challenges in remote sensing tasks: 1. Class Imbalance Issue: Target areas such as rocky desertification are usually scattered and fragmented, accounting for a very small percentage of the total sample. Conventional model optimization is prone to bias towards the majority, leading to serious missed detections of minority classes.
[0004] 2. Limitations of scanning strategy: The simple linear scan of the standard state-space model destroys the local spatial continuity of 2D images, resulting in the loss of fine-grained structural features.
[0005] 3. Insufficient representation of key regions: The uniform feature extraction weight allocation makes the model lack discriminative power when dealing with boundaries and class confusion transition areas, which easily leads to spurious responses. Summary of the Invention To address the aforementioned technical problems, this invention provides a method and system for imbalance-like remote sensing image segmentation based on an importance-aware state-space model. The method includes the following steps: S100: Obtain training set images and pixel-level labeled masks. During the network training phase, a complementary enhancement dynamic class balancing sampling strategy is adopted. The training samples are dynamically reweighted and then the sampled remote sensing images are input into the segmentation network. S200: The encoder of the segmentation network extracts features from the input remote sensing image. The encoder adopts the AsMamba base and captures the two-dimensional global long-range dependency features of the image through a snake bidirectional scanning strategy. It also dynamically fuses the state space features and identity mapping features by combining an adaptive importance gating mechanism to output an enhanced backbone feature representation. S300: The enhanced backbone feature representation is input into the decoder of the segmentation network for multi-scale feature aggregation, and a dynamic query value enhancement module is introduced at the decoder. By dynamically generating queries, modeling category relationships, and performing cross-attention interactions, the enhanced decoded features are output. S400: Based on the enhanced decoding features, predict the pixel-level category probability distribution, optimize the network parameters by combining the joint loss function, and after training, use the segmentation network to process the input remote sensing image to be tested, and output the final semantic segmentation result of the remote sensing image.
[0006] As a further aspect of the present invention, in step S100, the specific calculation process of the complementary enhancement dynamic category balancing sampling strategy is as follows: Calculate the basic sampling weights of each class in the training set samples. ; During model training, the precision and recall of the reference class on the validation set are calculated, and a complementary confidence metric reflecting the reliability of the reference class predictions is constructed. And potential regional coefficients that measure the size of unexplained regions ; Introducing a dynamic enhancement coefficient that decays with the number of iterations ; Calculate the final sampling weights The calculation formula is as follows: For the target minority class pixels; The proportion in the current sample For the first The set of categories contained in each sample This represents a minority category of targets, namely the rocky desertification category.
[0007] As a further aspect of the present invention, in step S200, the specific implementation process of the serpentine bidirectional scanning strategy is as follows: For input features Expand horizontally row by row, and reverse the odd-numbered rows to construct a horizontal serpentine sequence. First, transpose the sequence vertically, then apply the same serpentine rules to construct a vertical serpentine sequence. ; Sequences in the two directions are modeled using independent state-space models (SSMs) respectively, and then modeled using inverse transformations. Restored to a two-dimensional feature map: The final output of the serpentine bidirectional state space branch is: .
[0008] As a further aspect of the present invention, in step S200, the specific implementation process of the adaptive importance gating mechanism is as follows: Global semantic weights are extracted using global average pooling and two layers of nonlinear mapping. Local spatial weights are extracted through position-wise convolution mapping. The global semantic weights and local spatial weights are multiplied element-wise to obtain the adaptive fusion coefficients. Using the adaptive fusion coefficient Features of the serpentine state space and identity mapping features Perform weighted fusion to output an enhanced backbone feature representation. .
[0009] As a further aspect of the present invention, in step S300, the dynamic query value enhancement module includes the following processing steps: Learnable base queries are expanded into candidate queries using linear mapping. And predict dynamic combination weights from input features. Constructing image-adaptive dynamic queries Introducing a learnable class relation matrix Enhance the relationships in dynamic queries to obtain enhanced relationship queries. The enhanced query and features are then interacted with through attention and projected back into the spatial domain. The enhanced features are then fused with the original input features using residuals to obtain the final enhanced features.
[0010] As a further aspect of the present invention, during the network parameter optimization process in step S400, a hard sample contrast loss is introduced for the query representation of the dynamic query value enhancement module. The specific calculation includes: performing L2 normalization on the dynamic query representation to obtain... And calculate the batch-level category center. ; Calculate the intra-class compactness loss that causes similar queries to cluster towards the category center. Calculate similarity based on category center With confusion Inter-class separability loss The hard sample contrast loss Defined as: in, For temperature coefficient, For the set of all distinct class pairs, These are boundary parameters.
[0011] As a further aspect of the present invention, in step S400, the joint loss function Loss due to the main decoder head Dice Dice loss of auxiliary decoder head and the hard sample contrast loss Together they form a whole, and their formula is: in Hyperparameters are used to balance the weights of various losses.
[0012] As a further aspect of the present invention, the present invention also provides a class-imbalanced remote sensing image segmentation system based on an importance-aware state-space model, comprising: Sampling module: acquires training set images and pixel-level labeled masks. During the network training phase, a complementary enhancement dynamic class balancing sampling strategy is used to dynamically reweight the training samples and input the sampled remote sensing images into the segmentation network. Segmentation Network Module: The encoder of the segmentation network extracts features from the input remote sensing image. The encoder captures the two-dimensional global long-range dependency features of the image through a serpentine bidirectional scanning strategy, and dynamically fuses the state space features and identity mapping features by combining an adaptive importance gating mechanism to output an enhanced backbone feature representation. Encoder module: Inputs the enhanced backbone feature representation into the decoder of the segmentation network for multi-scale feature aggregation, and introduces a dynamic query value enhancement module at the decoder end. By dynamically generating queries, modeling category relationships and performing cross-attention interaction, it outputs the enhanced decoded features. Decoder module: Based on the enhanced decoding features, it predicts the pixel-level category probability distribution, optimizes the network parameters by combining the joint loss function, and after training, it uses the segmentation network to process the input remote sensing image to be tested and outputs the final semantic segmentation result of the remote sensing image.
[0013] Compared with the prior art, the present invention has achieved the following significant beneficial effects: (1) Significantly improves the overall segmentation accuracy of remote sensing images with complex scenes: This invention effectively solves the problem of complex background interference by fusing the AsMamba backbone with a dynamic query value enhancement mechanism. On a three-class UAV remote sensing image dataset containing background, rocky desertification, and vegetation, the average intersection-union ratio (mIoU) of the method of this invention reaches The average F1 score (mF1) reached Overall accuracy (OA) reached Compared to traditional classic convolutional networks (such as Deeplabv3+), mIoU, mF1, and OA are significantly improved. and Percentage points. Even compared to state-space model-based methods (such as RS3Mamba), mIoU is improved by [percentage missing]. percentage points.
[0014] (2) Significantly improves the accuracy of minority class identification under extreme class imbalance: Thanks to the synergistic effect of the CCBS sampling strategy and the contrast loss of difficult samples, this invention achieves a high Intersection over Union (IoU) ratio for the extremely difficult-to-identify and sporadically distributed minority class "rocky desertification". F1 score reached It surpasses existing mainstream methods. It effectively solves the problems of missed detections and false detections caused by the blurred boundaries of small patches and their easy confusion with the background / vegetation.
[0015] (3) A good balance between high accuracy and computational efficiency is achieved: The total number of parameters (Params) of the complete model in this invention is approximately The number of floating-point operations (FLOPs) is approximately Compared to the baseline model, it achieves significantly better performance with only a minimal increase in computational cost. The mIo improvement is demonstrated. It is proven that the bidirectional serpentine scanning strategy has extremely high efficiency in feature extraction and global modeling. Attached Figure Description
[0016] Figure 1 The flowchart of the overall network architecture of an imbalanced remote sensing image segmentation method based on an importance-aware state-space model provided by the present invention; Figure 2 This is a schematic diagram of the serpentine bidirectional scanning strategy in this invention; Figure 3 This is a schematic diagram of the internal structure of the dynamic query value enhancement module in this invention. Detailed Implementation
[0017] The technical solution of the present invention will be described in detail below with reference to specific mathematical formulas and embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.
[0018] An imbalance-like remote sensing image segmentation method and system based on an importance-aware state-space model, the method comprising the following steps: S100: Obtain training set images and pixel-level labeled masks. During the network training phase, a complementary enhancement dynamic class balancing sampling strategy is adopted. The training samples are dynamically reweighted and then the sampled remote sensing images are input into the segmentation network. S200: The encoder of the segmentation network extracts features from the input remote sensing image. The encoder adopts the AsMamba base and captures the two-dimensional global long-range dependency features of the image through a snake bidirectional scanning strategy. It also dynamically fuses the state space features and identity mapping features by combining an adaptive importance gating mechanism to output an enhanced backbone feature representation. S300: The enhanced backbone feature representation is input into the decoder of the segmentation network for multi-scale feature aggregation, and a dynamic query value enhancement module is introduced at the decoder. By dynamically generating queries, modeling category relationships, and performing cross-attention interactions, the enhanced decoded features are output. S400: Based on the enhanced decoding features, predict the pixel-level category probability distribution, optimize the network parameters by combining the joint loss function, and after training, use the segmentation network to process the input remote sensing image to be tested, and output the final semantic segmentation result of the remote sensing image.
[0019] As a further aspect of the present invention, in step S100, the specific calculation process of the complementary enhancement dynamic category balancing sampling strategy is as follows: Calculate the basic sampling weights of each class in the training set samples. ; During model training, the precision and recall of the reference class on the validation set are calculated, and a complementary confidence metric reflecting the reliability of the reference class predictions is constructed. And potential regional coefficients that measure the size of unexplained regions ; Introducing a dynamic enhancement coefficient that decays with the number of iterations ; Calculate the final sampling weights The calculation formula is as follows: For the target minority class pixels; The proportion in the current sample For the first The set of categories contained in each sample This represents a minority category of targets, namely the rocky desertification category.
[0020] As a further aspect of the present invention, in step S200, the specific implementation process of the serpentine bidirectional scanning strategy is as follows: For input features Expand horizontally row by row, and reverse the odd-numbered rows to construct a horizontal serpentine sequence. First, transpose the sequence vertically, then apply the same serpentine rules to construct a vertical serpentine sequence. ; Sequences in the two directions are modeled using independent state-space models (SSMs) respectively, and then modeled using inverse transformations. Restored to a two-dimensional feature map: The final output of the serpentine bidirectional state space branch is: .
[0021] As a further aspect of the present invention, in step S200, the specific implementation process of the adaptive importance gating mechanism is as follows: Global semantic weights are extracted using global average pooling and two layers of nonlinear mapping. Local spatial weights are extracted through position-wise convolution mapping. The global semantic weights and local spatial weights are multiplied element-wise to obtain the adaptive fusion coefficients. Using the adaptive fusion coefficient Features of the serpentine state space and identity mapping features Perform weighted fusion to output an enhanced backbone feature representation. .
[0022] As a further aspect of the present invention, in step S300, the dynamic query value enhancement module includes the following processing steps: Learnable base queries are expanded into candidate queries using linear mapping. And predict dynamic combination weights from input features. Constructing image-adaptive dynamic queries Introducing a learnable class relation matrix Enhance the relationships in dynamic queries to obtain enhanced relationship queries. The enhanced query and features are then interacted with through attention and projected back into the spatial domain. The enhanced features are then fused with the original input features using residuals to obtain the final enhanced features.
[0023] As a further aspect of the present invention, during the network parameter optimization process in step S400, a hard sample contrast loss is introduced for the query representation of the dynamic query value enhancement module. The specific calculation includes: performing L2 normalization on the dynamic query representation to obtain... And calculate the batch-level category center. ; Calculate the intra-class compactness loss that causes similar queries to cluster towards the category center. Calculate similarity based on category center With confusion Inter-class separability loss The hard sample contrast loss Defined as: in, For temperature coefficient, For the set of all distinct class pairs, These are boundary parameters.
[0024] As a further aspect of the present invention, in step S400, the joint loss function Loss due to the main decoder head Dice Dice loss of auxiliary decoder head and the hard sample contrast loss Together they form a whole, and their formula is: in Hyperparameters are used to balance the weights of various losses.
[0025] As a further aspect of the present invention, the present invention also provides a class-imbalanced remote sensing image segmentation system based on an importance-aware state-space model, comprising: Sampling module: acquires training set images and pixel-level labeled masks. During the network training phase, a complementary enhancement dynamic class balancing sampling strategy is used to dynamically reweight the training samples and input the sampled remote sensing images into the segmentation network. Segmentation Network Module: The encoder of the segmentation network extracts features from the input remote sensing image. The encoder captures the two-dimensional global long-range dependency features of the image through a serpentine bidirectional scanning strategy, and dynamically fuses the state space features and identity mapping features by combining an adaptive importance gating mechanism to output an enhanced backbone feature representation. Encoder module: Inputs the enhanced backbone feature representation into the decoder of the segmentation network for multi-scale feature aggregation, and introduces a dynamic query value enhancement module at the decoder end. By dynamically generating queries, modeling category relationships and performing cross-attention interaction, it outputs the enhanced decoded features. Decoder module: Based on the enhanced decoding features, it predicts the pixel-level category probability distribution, optimizes the network parameters by combining the joint loss function, and after training, it uses the segmentation network to process the input remote sensing image to be tested and outputs the final semantic segmentation result of the remote sensing image.
[0026] Example 1: Detailed Explanation of the Core Algorithm Flow This embodiment details the implementation logic of each core module in the imbalanced remote sensing image segmentation method described above: 1. Complementary Enhancement Dynamic Class Balance Sampling: To address the scarcity of minority classes, this invention addresses the issue of basic sampling weights. Based on this, a performance feedback mechanism for reference classes is introduced.
[0027] During training, the accuracy of the reference class on the validation set is calculated. and recall rate Define complementary confidence levels. Define the potential regional coefficient as a measure of the unexplained regional size. .
[0028] Introducing the number of iterations Attenuation dynamic enhancement factor This is to ensure a balance between early exploration and later stability. Ultimately, this is aimed at the... Sampling weights of each sample The calculation is as follows: in The proportion of pixels representing the target minority class in the current sample. This represents the set of categories contained in the current sample. This design avoids blind oversampling and increases the exposure rate of high-quality, difficult samples.
[0029] 2. Importance-aware state-space modeling: In the encoder section, the linear expansion of traditional SSMs disrupts the local space. This invention proposes a serpentine bidirectional scanning strategy: for two-dimensional input features... Expand horizontally row by row and reverse the odd-numbered rows to obtain After transposing vertically, a similar process is performed to obtain... After inverse transformation and addition of their respective SSM models, the results are obtained. This strategy achieves the best balance between preserving spatial continuity and computational efficiency compared to traditional 2D or 8D scanning.
[0030] An adaptive importance gating mechanism is then introduced to extract global semantic weights. and local spatial weights The final enhanced features are represented as For border and challenging areas, The value is relatively large.
[0031] 3. Dynamic query value augmentation and hard sample contrast learning: At the decoder, dynamic weights are predicted from input features. The basic query is built into an adaptive dynamic query. To enhance the separability between classes (especially the distinction between rocky desertification and bare soil), contrast constraints are introduced into the feature space.
[0032] For the L2 normalized query representation and batch category center Intraclass compactness loss for: Inter-class separability loss Combining confusion for: Final comparison of losses 4. Joint Loss Optimization: The network as a whole is trained using the total loss function. The Dice loss is directly optimized based on the degree of regional overlap.
[0033] Example 2: Details of Model Training and Parameter Setting To enable those skilled in the art to implement this invention, specific training environment and hyperparameter configurations are provided: Dataset preprocessing: using a resolution of RGB remote sensing image data of pixels. The applied data augmentation strategies include: scaling at... Random scaling between, size is Random cropping, with a probability of Random horizontal flip, with a probability of Random vertical flipping, optical distortion, and angle at which The probability between them is Random rotation.
[0034] Optimizer and Hyperparameters: Training was performed using the PyTorch framework on a single NVIDIA RTX 3090 GPU. The AdamW optimizer was used, with an initial learning rate of [missing value]. The backbone network learning rate multiplier is set to Set as The weight decay coefficient is Using the maximum norm as Gradient clipping. The total number of iterations is... Step forward The first step uses linear preheating, followed by cosine annealing. The batch size is set to [value missing]. .
[0035] Complementary Enhancement Dynamic Class Balance Sampling Parameter Settings: Initial Enhancement Coefficients =for The complementary weights are High-precision threshold is This mechanism is activated after the model has initially acquired discriminative capabilities, i.e., at the 8000th iteration.
[0036] Example 3: Comparative Experiment Description To further demonstrate the advancement, effectiveness, and outstanding substantive features of the method of this invention, this embodiment compares the performance of the method of this invention with existing mainstream remote sensing image semantic segmentation techniques on the same test set. The existing technologies used for comparison cover three major categories of mainstream architectures: traditional convolutional neural networks (DeepLabv3+, ABCNet), Transformer-based networks (Swin-Transformer, UnetFormer), and the latest state-space model-based networks (MF-Mamba, RS3Mamba).
[0037] 1. Overall performance evaluation The experiment used objective indicators such as mean Intersection over Union (mIoU), mean F1 score (mF1), and overall accuracy (OA) for quantitative evaluation. Experimental data show that the method of this invention achieved optimal results across all global and category evaluation indicators. Specifically, the method of this invention achieved a vegetation IoU of 72.90%, a rocky desertification IoU of 62.31%, a global mIoU of 67.61%, a mF1 score of 80.55%, and an overall accuracy (OA) of 82.25%.
[0038] 2. Compare the significant advancements in existing technologies. Compared to traditional convolutional networks: Compared to the classic DeepLabv3+ model, the method of this invention significantly improves mIoU and mF1 by 9.43 and 7.04 percentage points, respectively, and OA by 10.62 percentage points. This demonstrates that the present invention overcomes the limitation of traditional CNNs that rely solely on local receptive fields, and significantly improves the ability to recognize fine-grained details in complex backgrounds.
[0039] Compared to Transformer architectures: Compared to Swin-Transformer and UnetFormer, which have global modeling capabilities, the method of this invention improves mIoU by 6.79 and 4.61 percentage points, respectively. The data shows that this invention, while fully utilizing global context information, more effectively solves problems such as boundary ambiguity, extreme scale changes, and inter-class feature similarity, overcoming the deficiency of pure Transformer models in insufficient local structure recovery under complex background interference.
[0040] Compared to similar Mamba architectures: Even compared to the latest MF-Mamba and RS3Mamba, this invention still demonstrates an overwhelming advantage. Compared to the suboptimal RS3Mamba model (mIoU of 63.39% and OA of 81.05%), this invention further improves mIoU by 4.22 percentage points.
[0041] 3. A significant effect in addressing extreme class imbalances One of the core technical challenges of this invention lies in solving the problem of segmenting the "rocky desertification" minority class, which is characterized by its scattered distribution and high susceptibility to confusion. In terms of category-level metrics, this invention achieves IoU and F1 scores of 62.31% and 76.77%, respectively, for the most challenging rocky desertification class. Compared to the best-in-class RS3Mamba model, this invention further improves the IoU and F1 scores for the rocky desertification class by 2.24 and 1.72 percentage points, respectively. This demonstrates that this invention is not a simple application of conventional techniques, but rather a targeted and significant enhancement in the recognition performance of difficult categories.
[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for segmenting imbalance-like remote sensing images based on an importance-aware state-space model, characterized in that, Includes the following steps: S100: Acquire training set images and pixel-level labeled masks. During the network training phase, a complementary enhancement dynamic class balancing sampling strategy is used to dynamically reweight the training samples. The sampled remote sensing images are then input into the segmentation network. S200: The encoder of the segmentation network extracts features from the input remote sensing image. The encoder captures the two-dimensional global long-range dependency features of the image through a snake bidirectional scanning strategy, and dynamically fuses the state space features and identity mapping features by combining an adaptive importance gating mechanism to output an enhanced backbone feature representation. S300: The enhanced backbone feature representation is input into the decoder of the segmentation network for multi-scale feature aggregation, and a dynamic query value enhancement module is introduced at the decoder end. By dynamically generating queries, modeling category relationships and performing cross-attention interaction, the enhanced decoded features are output. S400: Based on the enhanced decoding features, predict the pixel-level category probability distribution, optimize the network parameters by combining the joint loss function, and after training, use the segmentation network to process the input remote sensing image to be tested, and output the final semantic segmentation result of the remote sensing image.
2. The imbalance-like remote sensing image segmentation method based on an importance-aware state-space model according to claim 1, characterized in that, In step S100, the specific calculation process of the complementary enhancement dynamic class balancing sampling strategy is as follows: Calculate the basic sampling weights of each class in the training set samples. ; During model training, the precision and recall of the reference class on the validation set are calculated, and a complementary confidence metric reflecting the reliability of the reference class predictions is constructed. and potential regional coefficients that measure the size of unexplained regions ; Introducing a dynamic enhancement coefficient that decays with the number of iterations ; Calculate the final sampling weights The calculation formula is as follows: in, For the target minority class pixels; The proportion in the current sample For the first The set of categories contained in each sample This represents a minority category of targets, namely the rocky desertification category.
3. The imbalance-like remote sensing image segmentation method based on an importance-aware state-space model according to claim 1, characterized in that, In step S200, the specific implementation process of the serpentine bidirectional scanning strategy is as follows: For input features Expand horizontally row by row, and reverse the odd-numbered rows to construct a horizontal serpentine sequence. First, transpose the sequence vertically, then apply the same serpentine rules to construct a vertical serpentine sequence. ; Sequences in the two directions are modeled using independent state-space models (SSMs) respectively, and then modeled using inverse transformations. Restored to a two-dimensional feature map: The final output of the serpentine bidirectional state space branch is: 。 4. The imbalance-like remote sensing image segmentation method based on an importance-aware state-space model according to claim 3, characterized in that, In step S200, the specific implementation process of the adaptive importance gating mechanism is as follows: Global semantic weights are extracted using global average pooling and two layers of nonlinear mapping. Local spatial weights are extracted through position-wise convolution mapping. The global semantic weights and local spatial weights are multiplied element-wise to obtain the adaptive fusion coefficients. Using the adaptive fusion coefficient Features of the serpentine state space and identity mapping features Perform weighted fusion to output an enhanced backbone feature representation. 。 5. The imbalance-like remote sensing image segmentation method based on an importance-aware state-space model according to claim 1, characterized in that, In step S300, the dynamic query value enhancement module includes the following processing procedures: Learnable base queries are expanded into candidate queries using linear mapping. And predict dynamic combination weights from input features Constructing image-adaptive dynamic queries Introducing a learnable class relation matrix Enhance the relationships in dynamic queries to obtain enhanced relationship queries. The enhanced query and features are then interacted with through attention and projected back into the spatial domain. The enhanced features are then fused with the original input features using residuals to obtain the final enhanced features.
6. The imbalance-like remote sensing image segmentation method based on an importance-aware state-space model according to claim 5, characterized in that, In the network parameter optimization process of step S400, a hard sample contrast loss is introduced for the query representation of the dynamic query value enhancement module. The specific calculation includes: performing L2 normalization on the dynamic query representation to obtain... And calculate the batch-level category center. ; Calculate the intra-class compactness loss that causes similar queries to cluster towards the category center. Calculate similarity based on category center With confusion Inter-class separability loss The hard sample contrast loss Defined as: in, For temperature coefficient, For the set of all distinct class pairs, These are boundary parameters.
7. The imbalance-like remote sensing image segmentation method based on an importance-aware state-space model according to claim 6, characterized in that, In step S400, the joint loss function Loss due to the main decoder head Dice Dice loss of auxiliary decoder head and the hard sample contrast loss Together they form a whole, and their formula is: in Hyperparameters are used to balance the weights of various losses.
8. A class-imbalanced remote sensing image segmentation system based on an importance-aware state-space model, employing the method described in any one of claims 1-7, characterized in that, include: Sampling module: acquires training set images and pixel-level labeled masks. During the network training phase, a complementary enhancement dynamic class balancing sampling strategy is used to dynamically reweight the training samples and input the sampled remote sensing images into the segmentation network. Segmentation Network Module: The encoder of the segmentation network extracts features from the input remote sensing image. The encoder captures the two-dimensional global long-range dependency features of the image through a serpentine bidirectional scanning strategy, and dynamically fuses the state space features and identity mapping features by combining an adaptive importance gating mechanism to output an enhanced backbone feature representation. Encoder module: Inputs the enhanced backbone feature representation into the decoder of the segmentation network for multi-scale feature aggregation, and introduces a dynamic query value enhancement module at the decoder end. By dynamically generating queries, modeling category relationships and performing cross-attention interaction, it outputs the enhanced decoded features. Decoder module: Based on the enhanced decoding features, it predicts the pixel-level category probability distribution, optimizes the network parameters by combining the joint loss function, and after training, it uses the segmentation network to process the input remote sensing image to be tested and outputs the final semantic segmentation result of the remote sensing image.