Remote sensing image segmentation method based on multi-scale wavelet transform and Mama

Abstract

The invention discloses a remote sensing image segmentation method based on multi-scale wavelet transform and Mama. The method comprises the following four steps: firstly, carrying out preprocessing and division on an ISPRS Potsdam data set and a Vaihingen data set; then constructing a segmentation network, wherein the network comprises a local detail extraction branch, a spatial semantic extraction branch, a cross-branch feature fusion part and a decoder; training and optimizing the network by using the training set; and finally, performing segmentation reasoning on a test image by using the trained model. According to the method, the high-frequency detail extraction capability of the image is enhanced through multi-scale wavelet transform, the long-distance dependency relationship is modeled by using the visual state space block, and effective fusion of the features is realized through the double-branch fusion module, so that the feature extraction capability and the semantic segmentation precision are improved, and meanwhile, the training efficiency and the stability are optimized.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a remote sensing image segmentation method based on multi-scale wavelet transform and Mamba. Background Technology

[0002] With the rapid development of aerospace and sensor technologies, the ability to acquire high-resolution remote sensing imagery has been revolutionized. These image data possess advantages such as wide coverage, high spatial resolution, and rich information on ground features, making them an indispensable source of information for fields such as urban planning, environmental monitoring, disaster assessment, and national defense. Correspondingly, the need for automated and accurate extraction of semantic information from these massive amounts of imagery (i.e., remote sensing image semantic segmentation) is becoming increasingly urgent. However, the inherent characteristics and complexity of high-resolution remote sensing imagery present a series of severe challenges when directly applying conventional natural image semantic segmentation methods, mainly in the following aspects: 1. Extreme Scene and Target Multi-Scale: High-resolution remote sensing imagery covers vast geographical areas, with scene complexity far exceeding that of conventional natural images. This is reflected not only in the huge differences in target size within the same scene (from farmland stretching for kilometers to vehicles occupying only a few pixels), but also in the drastically different texture features and spectral response characteristics of different land cover categories. Traditional single-scale feature extraction networks struggle to simultaneously capture the macroscopic information needed to maintain long-range context (such as river course) and the fine local features required to distinguish small targets (such as power lines and vehicles), resulting in a difficulty in achieving a balance between "seeing the forest" and "seeing the trees."

[0003] 2. Complex Structured Background Interference: Background "noise" in remote sensing imagery often possesses high structural and semantic deceptive qualities. For example, clouds and cloud shadows may form shapes and textures similar to real ground features (such as water bodies and buildings); building shadows can cause distortion and occlusion of local ground features. This interference is not random noise, and traditional filtering methods based on low-level color or texture are difficult to distinguish effectively, severely affecting segmentation accuracy and requiring algorithms with stronger contextual reasoning and feature completion capabilities.

[0004] 3. Challenges in Long-Range Spatial Dependency Modeling: The semantic understanding of many land cover categories strongly depends on the analysis of their large-scale spatial distribution patterns. For example, identifying "roads" requires understanding their linear connectivity network, while distinguishing "rivers" requires grasping their meandering morphology. This demands that the model possess a very large effective receptive field to capture long-range dependencies over a range of several kilometers. However, simply increasing the convolutional kernel or deepening the network layers leads to a dramatic increase in computational cost and optimization difficulties, contradicting the practical requirements for algorithm computational efficiency.

[0005] 4. Loss of small target features during downsampling: Existing mainstream semantic segmentation architectures based on encoder-decoder (such as U-Net) or feature pyramids generally rely on downsampling operations to expand the receptive field and reduce computational load. However, for small targets (such as small buildings, isolated trees, vehicles, etc.) that occupy only a few pixels but contain key semantic information and are abundant in remote sensing images, their weak feature signals are easily submerged or even completely lost during multiple downsampling processes, resulting in ineffective recovery during the decoding stage and causing missed detections.

[0006] 5. Stringent Requirements for Computational Efficiency and Real-Time Performance: For applications such as on-board processing and disaster emergency response, the algorithm needs to be capable of real-time or near real-time processing of massive imagery covering square kilometers. Existing models typically rely on high-performance GPU clusters for inference, making them difficult to deploy and run on resource-constrained hardware platforms such as onboard or airborne systems. How to significantly reduce the computational complexity and number of parameters of the model while maintaining high accuracy is a key bottleneck in realizing the practical application of the technology.

[0007] In summary, the core challenge of semantic segmentation of high-resolution remote sensing images lies in designing a deep learning model that can collaboratively model extreme multi-scale features, effectively resist structured interference, efficiently capture long-range context, and simultaneously remain sensitive to small targets while maintaining computational efficiency. Existing methods typically have limitations in one or more of these aspects, necessitating a comprehensive and innovative architecture to address these interconnected challenges. Summary of the Invention

[0008] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a remote sensing image segmentation method based on multi-scale wavelet transform and Mamba.

[0009] A remote sensing image segmentation method based on multi-scale wavelet transform and Mamba according to an embodiment of the present invention includes the following steps: S1. Data Processing: Obtain remote sensing image datasets including ISPRS Potsdam and ISPRS Vaihingen, and preprocess the image data, dividing them into training set, validation set and test set; S2. Constructing the network model: Building a segmented network named MSWT-Mamba; S3. Model Training: Input the training set into the MSWT-Mamba network for training. Calculate the loss function and perform backpropagation to update the network parameters to obtain the trained parameter model. S4. Model Inference: Input the test set into the trained parameter model and output the semantic segmentation result map of the target image; The segmentation network in step S2 includes at least the following sequentially connected components: The local detail extraction branch, based on a multi-level residual block structure, is used to extract multi-scale local spatial detail features from the input image. A spatial semantic extraction branch, which is based on a hierarchical structure including a multi-scale wavelet transform convolution module and a selective state space module, is used to model the global contextual semantic dependencies of an image. The cross-branch feature fusion part includes multiple dual-branch fusion modules, which are used to adaptively fuse local detail features and global semantic features at the corresponding level to generate fused features; The decoder section is based on an upsampling and skip connection architecture. It uses the fused features to gradually restore the spatial resolution and outputs the final semantic segmentation prediction map.

[0010] In some embodiments of the present invention, the preprocessing in step S1 includes normalization, data augmentation, pruning, and noise reduction, as detailed below: The image data is normalized, converting pixel values ​​to the [0, 1] interval. The calculation formula is as follows: , in, These are the original pixel values. The minimum pixel value in the image. The maximum pixel value in the image; Data augmentation of image data includes rotation transformation and / or horizontal flipping. Rotation transformation changes the image angle using a rotation matrix, which is: , in, The x-coordinate of the pixel before rotation. The pixel ordinate before rotation. The x-coordinate of the rotated pixel. The ordinate of the rotated pixel. The rotation angle is expressed in radians. Horizontal flipping is a process of mirroring an image vertically; the corresponding formula is: , in, For row indexes in the image, For column indexes in the image, The width of the image (i.e., the total number of columns). The image is in position after being horizontally flipped. Pixel value at that location, For the original image at position Pixel value at; Crop the image into fixed-size tiles; To perform Gaussian filtering for noise reduction on the image, the filter kernel function is: , in, This represents the horizontal coordinate offset of a point within the core relative to the center point. This represents the vertical coordinate offset of a point in the core relative to the center point.

[0011] In some embodiments of the present invention, the spatial semantic extraction branch in step S2 includes multiple processing stages, and the processing procedure of each stage includes: The input features are processed through two separate paths: The first path sequentially passes through the first normalization layer, the first linear transform layer, the multi-scale wavelet transform convolution module, the selective state space module, and the second normalization layer; The second path passes through the first normalization layer and the second linear transformation layer in sequence; After multiplying the outputs of the two paths, the result is passed through the third linear transformation layer and then residually connected with the original input of the branch to obtain the output features of this stage. The multi-scale wavelet transform convolution module is used to perform multi-level wavelet transform decomposition and reconstruction of input features by deploying convolution operations of different scales in parallel, so as to expand the feature receptive field; the selective state space module is based on the state space model and is used to perform global feature extraction and fusion.

[0012] In some embodiments of the present invention, the multi-scale wavelet transform convolution module performs multi-level wavelet transform decomposition on the input features by deploying convolution operations of three scales, namely 1×1, 3×3 and 5×5, in parallel to generate low-frequency and high-frequency components; after convolution processing of each component, the features are reconstructed by inverse wavelet transform; finally, the reconstructed features are fused with the original features and output.

[0013] In some embodiments of the present invention, the local detail extraction branch in step S2 includes multiple residual blocks, and the feature transformation process of the residual blocks is expressed as follows: , , in, express Convolution operations, BN represents batch normalization, and ReLU represents the activation function. Indicates the first Output characteristics of each residual block.

[0014] In some embodiments of the present invention, the cross-branch feature fusion part in step S2 includes multiple dual-branch fusion modules for fusing the first... Hierarchical local detail features and global semantic features Generate fusion features The calculation process of the dual-branch fusion module is as follows: , , , in, For layer normalization, For channel attention mechanism, For matrix multiplication, For adaptive cross-domain fusion blocks.

[0015] In some embodiments of the present invention, the calculation process of the adaptive cross-domain fusion block is as follows: For input features respectively and Perform depthwise separable convolution to generate the query vector. , key vector , value vector , Specifically: Calculate the multi-head attention between the two cross domains: , ; Will and After concatenation, the output is obtained through a 1×1 convolution. ; in, For depthwise separable convolution, This is a multi-head attention mechanism.

[0016] In some embodiments of the present invention, the decoder portion in step S2 is an architecture including skip connections, combining multiple layers of fused features with features from the upsampling process; during training, an auxiliary segmentation head is connected after multiple fused feature layers, and the total loss function is a weighted sum of the final output loss and each auxiliary loss: , in, , and Both are composite loss functions of Dice loss and cross-entropy loss.

[0017] The present invention also mentions an electronic device including a central processing unit and a storage module, wherein the storage module stores executable instructions that, when executed by the central processing unit, control the central processing unit to implement the process described above.

[0018] The present invention also mentions a computer-readable storage medium storing computer-programmable instructions thereon, which, when executed on a computing device, trigger the computing device to perform the operation steps of the method described above.

[0019] Compared with existing technologies, the remote sensing image segmentation method based on multi-scale wavelet transform and Mamba provided by this invention has the following advantages: First, this invention achieves efficient collaborative extraction and fusion of local details and global semantics through an innovative dual-branch parallel architecture. The local detail extraction branch effectively preserves and enhances the fine-grained spatial information of the image through a multi-layer residual structure, directly addressing the problem of easy loss of features for small targets. The spatial semantic extraction branch, by deeply fusing multi-scale wavelet transform (GWCM) with selective state space model (VSS / SS2D), expands the effective receptive field to model long-range dependencies (such as rivers and road networks) while enhancing the feature representation ability of different scale features (such as large-scale farmland and scattered vehicles) through multi-scale decomposition and processing in the wavelet domain. Experiments on public datasets show that this design can effectively improve the segmentation recall rate of small targets such as "low vegetation" and "cars" as well as complex structural features.

[0020] Second, the GWCM module proposed in this invention utilizes the multi-scale characteristics of wavelet transform, enabling shallow networks to indirectly cover a larger input region with logarithmically increasing parameters, effectively expanding the model's initial receptive field. Subsequently, the more concentrated and structured features, preprocessed by GWCM, are input into the VSS block based on the state-space model, leveraging the linear computational complexity advantage of the SS2D module to efficiently model global contextual relationships. This collaborative mechanism of "GWCM pre-enhancement + VSS efficient modeling" ensures a full understanding of the global semantic information of the image while avoiding parameter explosion and excessive computational burden, making it more suitable for real-time or near-real-time processing of square kilometer-scale high-resolution remote sensing images.

[0021] Third, the Adaptive Cross-Domain Fusion Block (ACFB) designed in the Cross-Branch Feature Fusion (DBFM) section of this invention introduces a multi-head attention mechanism across the cross-domains, enabling the model to dynamically evaluate and fuse features from local detail branches and global semantic branches. This strengthens feature inference and completion for regions with structural interference (such as cloud shadows and building shadows). Combining the complementary information provided by the two branches, the model can better overcome local feature distortion or loss caused by occlusion and shadows, improving the stability and accuracy of segmentation results under complex imaging conditions.

[0022] In summary, this invention systematically addresses the core challenges in high-resolution remote sensing image semantic segmentation, such as extreme multi-scale characteristics, complex background interference, difficulty in modeling long-range dependencies, easy loss of small targets, and high computational efficiency requirements, through three core technologies: "global modeling enhanced by multi-scale wavelet transform", "local detail preservation guaranteed by residual structure", and "adaptive cross-attention driven feature fusion". It has achieved significant improvements in segmentation accuracy, model efficiency, and robustness. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the remote sensing image segmentation method based on multi-scale wavelet transform and Mamba according to the present invention. Figure 2 This is the overall architecture diagram of the present invention; Figure 3 This is a schematic diagram of the GWCM-VSS structure; Figure 4 This is a schematic diagram of the Stem structure; Figure 5 This is a schematic diagram of the dual-branch fusion module; Figure 6 This is a schematic diagram of the adaptive cross-domain fusion block. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0025] To facilitate understanding, before introducing the embodiments of this disclosure, several terms involved in the embodiments of this disclosure will be explained as follows: The ISPRS Potsdam dataset is a benchmark dataset for semantic segmentation of high-resolution remote sensing images, provided by the International Society for Photogrammetry and Remote Sensing (ISPRS). The ISPRS Vaihingen dataset is a high-resolution remote sensing image dataset released by the International Society for Photogrammetry and Remote Sensing (ISPRS). It focuses on urban semantic segmentation tasks and aims to provide a standardized evaluation benchmark for remote sensing image analysis.

[0026] The following is for reference. Figures 1-6 The remote sensing image segmentation method based on multi-scale wavelet transform and Mamba according to embodiments of the present invention includes the following steps: S1. Data Processing: Obtain remote sensing image datasets including ISPRS Potsdam and ISPRS Vaihingen, and preprocess the image data, dividing them into training set, validation set and test set; S2. Constructing the network model: Building a segmented network named MSWT-Mamba; S3. Model Training: Input the training set into the MSWT-Mamba network for training. Calculate the loss function and perform backpropagation to update the network parameters to obtain the trained parameter model. S4. Model Inference: Input the test set into the trained parameter model and output the semantic segmentation result map of the target image; The segmentation network in step S2 includes at least the following sequentially connected components: The local detail extraction branch, based on a multi-level residual block structure, is used to extract multi-scale local spatial detail features from the input image. The spatial semantic extraction branch is based on a hierarchical structure that includes a multi-scale wavelet transform convolution module (GWCM) and a selective state space (SS2D) module, and is used to model the global contextual semantic dependencies of the image. The cross-branch feature fusion part includes multiple dual-branch fusion modules (DBFM), which are used to adaptively fuse local detail features and global semantic features at the corresponding level to generate fused features; The decoder section is based on an upsampling and skip connection architecture. It uses fused features to gradually restore the spatial resolution and outputs the final semantic segmentation prediction map.

[0027] In some embodiments of the present invention, the preprocessing in step S1 includes normalization, data augmentation, pruning, and noise reduction, as detailed below: To eliminate numerical differences between different images and improve model adaptability, image data can be normalized, converting pixel values ​​to the [0, 1] interval. The calculation formula is as follows: , in, These are the original pixel values. The minimum pixel value in the image. The maximum pixel value in the image. This is the normalized value.

[0028] Then, data augmentation is performed on the image data. Geometric transformations are used to simulate the changes in remote sensing images at different viewpoints and scales, enhancing the model's generalization ability. Specifically, this may include rotation transformations and / or horizontal flipping. Rotation transformations use a rotation matrix to change the image angle. The rotation matrix is: , in, The x-coordinate of the pixel before rotation. The pixel ordinate before rotation. The x-coordinate of the rotated pixel. The ordinate of the rotated pixel. The rotation angle is expressed in radians. Horizontal flipping is a process of mirroring an image vertically; the corresponding formula is: , in, For row indexes in the image, For column indexes in the image, The width of the image (i.e., the total number of columns). The image is in position after being horizontally flipped. Pixel value at that location, For the original image at position The pixel value at that location.

[0029] Next, the image is cropped into fixed-size tiles. For example, for... and Dataset cropping Pixel block.

[0030] Finally, the image is subjected to Gaussian filtering for noise reduction. The filter kernel function is: , in, This represents the horizontal coordinate offset of a point within the core relative to the center point. This represents the vertical coordinate offset of a point in the core relative to the center point.

[0031] In some embodiments of the present invention, the spatial semantic extraction branch in step S2 includes multiple processing stages, and the processing procedure of each stage includes: The input features are processed through two separate paths: The first path sequentially passes through the first normalization layer, the first linear transform layer, the multi-scale wavelet transform convolution module, the selective state space module, and the second normalization layer; The second path passes through the first normalization layer and the second linear transformation layer in sequence; After multiplying the outputs of the two paths, the result is passed through the third linear transformation layer and then residually connected with the original input of the branch to obtain the output features of this stage. Among them, the multi-scale wavelet transform convolution module is used to perform multi-level wavelet transform decomposition and reconstruction of input features by deploying convolution operations of different scales in parallel, so as to expand the feature receptive field; the selective state space module is based on the state space model and is used to perform global feature extraction and fusion.

[0032] For example, the spatial semantic extraction branch can contain four stages for extracting global features. Its core components are GWCM (Multi-scale Wavelet Transform Convolutional Module) and SS2D (Selective State Space Module). The selective state space module models long-range dependencies, and its linear computational complexity is significantly better than the self-attention mechanism of the traditional Transformer. Specifically, the spatial semantic extraction branch includes a first normalization layer, a second normalization layer, a first linear transformation layer, a second linear transformation layer, a third linear transformation layer, GWCM, and SS2D. The input data undergoes two processing paths: the first path passes through the first normalization layer and the first linear transformation layer, then through GWCM and SS2D, and finally through the second normalization layer; the second path passes through only the first normalization layer and the second linear transformation layer. The results from these two paths are multiplied, then the multiplied result is passed through the third linear transformation layer, and finally, this result is added to the original input data to obtain the extracted features. The SS2D unit is responsible for using a state-space model to extract and fuse input features at a global level. The GWCM unit further expands the receptive field, improving SS2D's global control.

[0033] In some embodiments of the present invention, the multi-scale wavelet transform convolution module performs multi-level wavelet transform decomposition on the input features by deploying convolution operations of three scales (1×1, 3×3, and 5×5) in parallel, generating low-frequency and high-frequency components. After convolution processing of each component, the features are reconstructed through inverse wavelet transform. Finally, the reconstructed features are fused with the original features for output. The wavelet transform can use a Haar wavelet basis, with a decomposition level of 2-4, and each level is processed by a 3×3 or 5×5 convolution kernel.

[0034] Understandably, the spatial semantic extraction branch first divides the input image into blocks using Patch Embedding to generate... Then, through Patch Merging, GWCM, and VSS Block, downsampling is performed step by step, and the output feature maps are as follows: This design enables the model to overcome the receptive field limitations of convolution operations, fully grasp the overall semantic information of the image, and effectively solve the common problems of large-scale target recognition and complex scene understanding in remote sensing images.

[0035] In some embodiments of the present invention, the local detail extraction branch in step S2 includes multiple residual blocks, and the feature transformation process of the residual blocks is expressed as follows: , , in, express Convolution operations, BN represents batch normalization, and ReLU represents the activation function. Indicates the first Output characteristics of each residual block.

[0036] For example, the local detail extraction branch can contain four residual blocks, specifically responsible for extracting local feature details from the remote sensing image. Input remote sensing image (dimension 1). ,in, For batch size, (Image size).

[0037] Understandably, this branch first performs initial feature extraction: through the first convolutional layer. Convert the input Then, multi-scale feature extraction is performed: after four residual blocks, the output feature maps are as follows: The model employs four downsampling stages to progressively extract multi-scale features. The output feature map size and number of channels are halved and doubled sequentially at each stage, ultimately generating a feature pyramid rich in spatial information. This design ensures the model can capture multi-level visual features from fine-grained to coarse-grained levels, providing a solid local feature foundation for subsequent semantic segmentation.

[0038] In some embodiments of the present invention, the cross-branch feature fusion section in step S2 includes multiple dual-branch fusion modules for fusing the first... Hierarchical local detail features and global semantic features Generate fusion features The calculation process of the dual-branch fusion module is as follows: , , , in, For layer normalization, For channel attention mechanism, For matrix multiplication, For adaptive cross-domain fusion blocks.

[0039] For example, the cross-branch feature fusion part can include four stages, each stage fusing local details to extract branch features through the Dual-Branch Fusion Module (DBFM). Spatial semantic extraction branch features Generate fusion features ,in, Indicates the scale level.

[0040] In some embodiments of the present invention, the calculation process of the adaptive cross-domain fusion block is as follows: For input features respectively and Perform depthwise separable convolution to generate the query vector. , key vector , value vector , Specifically: Calculate the multi-head attention between the two cross domains: , ; Will and After concatenation, the output is obtained through a 1×1 convolution. ; in, For depthwise separable convolution, This is a multi-head attention mechanism.

[0041] In some embodiments of the present invention, the decoder portion in step S2 is an architecture containing skip connections, combining the fused features from multiple levels with the features from the upsampling process; during training, an auxiliary segmentation head is connected after multiple fused feature layers, and the total loss function is the weighted sum of the final output loss and each auxiliary loss: , in, , and Both are composite loss functions of Dice loss and cross-entropy loss.

[0042] Specifically, the decoder adopts the classic UNetformer architecture, using skip connections to combine multi-scale features with the decoder's output to generate the final semantic segmentation prediction map. The figure shows the category inference for each pixel; these categories are predefined to distinguish various different types of surface material in the remote sensing image. Specifically, this is achieved in four feature fusion layers. The auxiliary segmentation head is then added; the total loss function is the weighted sum of the final output loss and the auxiliary loss. .

[0043] Understandably, the decoder gradually recovers the spatial resolution of the feature map through upsampling and skip connections. Finally, it outputs pixel-level classification results through convolutional layers, with the number of channels corresponding to the number of semantic segmentation categories.

[0044] In view of this, step S3 includes: inputting batch images Features are extracted through spatial semantic branch and local semantic branch respectively, and output respectively. and These features, in the dual-branch fusion module (DBFM), generate data that combines local details with global semantics. Features are used to gradually restore spatial resolution through upsampling and skip connection operations in the decoder, outputting a predicted segmentation map P. For loss calculation, cross-entropy loss is used to supervise the segmentation output, and network parameters are updated through backpropagation.

[0045] In summary, this invention addresses the core challenges in semantic segmentation of high-resolution remote sensing images, such as extreme multi-scale characteristics, difficulty in modeling long-range dependencies, easy loss of small target features, and high computational efficiency requirements. It innovatively proposes a dual-branch parallel architecture: enhancing local details through residual structures to protect small target information, efficiently modeling the global context and expanding the receptive field through the collaborative mechanism of multi-scale wavelet transform and selective state-space model, and combining an adaptive cross-domain feature fusion strategy. This effectively improves the segmentation accuracy, model running efficiency, and robustness to interference such as cloud shadow occlusion in complex scenes.

[0046] To further verify the feasibility and effectiveness of the method in this embodiment, experiments were conducted. Specifically, a network was built using Python, and test sets from the Potsdam and Vaihingen datasets were used to analyze the semantic segmentation performance of remote sensing images using the method in this embodiment. Two commonly used objective parameters for evaluating semantic segmentation methods were used as evaluation metrics: the mean F1 score (the number of correctly predicted pixels for each category in the segmented image) and the mean intersection over union (MIoU) ratio (the ratio of the intersection to the union of the segmented image and the label image) to evaluate the segmentation performance of the predicted semantic segmentation image.

[0047] The method of this embodiment is used to predict each remote sensing image in the Potsdam and Vaihingen datasets to obtain the predicted semantic segmentation image corresponding to each remote sensing image. The harmonic balance MF1 of the number of correctly predicted pixels for each category in the segmentation image, which reflects the semantic segmentation effect of the method of this invention, and the ratio MIoU of the intersection and union of the segmentation image and the label image are listed in Table 1 and Table 2.

[0048] The experimental results on the Vaihingen dataset (Table 1) show that the MSWT-Mamba method achieves high F1 scores and IoU values ​​for major land cover categories such as buildings, impervious surfaces, and trees. The segmentation accuracy for buildings (Bui.) is particularly outstanding, with an F1 score of 97.01% and an IoU of 94.24%, indicating that the model possesses strong recognition and segmentation capabilities for targets with clear structures and large scales. Furthermore, for typical small targets such as vehicles (Car), the model still achieves an F1 score of 89.92% and an IoU of 81.64%, significantly outperforming traditional methods. This verifies that the proposed method enhances high-frequency detail extraction through multi-scale wavelet transform and effectively alleviates the problem of feature loss during downsampling for small targets by combining local detail branches. The segmentation accuracy for the low-lying vegetation (Low.) category is slightly lower due to its complex texture and blurred boundaries, but the F1 score of 80.86% and IoU of 67.88% still remain at a high level, indicating that the model still has strong discriminative ability when distinguishing difficult categories. Overall, the model achieved an MF1 score of 90.36% and an MIoU of 82.82% on the Vaihingen dataset, demonstrating excellent comprehensive segmentation performance in urban scenes.

[0049] On the more complex and category-rich Potsdam dataset (Table 2), the model performs best in categories with significant long-range structural features, such as roads and water bodies, achieving F1 scores of 74.45% and 76.88%, respectively, and IoU scores of 59.20% and 62.00%, respectively. This is attributed to the efficient modeling capability of the Selective State Space Module (SS2D) for global context dependencies. The building (Bui.) category also maintains high segmentation accuracy (F1: 75.12%, IoU: 60.10%). However, in categories such as background, bare soil, and agricultural land, the model's segmentation accuracy decreases due to factors such as uneven sample distribution, large texture variations, and susceptibility to shadow occlusion. Nevertheless, the model still achieves a mF1 score of 67.38% and a mIoU of 51.54% on the Potsdam dataset, demonstrating strong competitiveness and robustness in such complex multi-class scenes.

[0050] Overall, the experimental results fully verify the effectiveness of the MSWT-Mamba method proposed in this invention. It extracts local details and global semantics collaboratively through a dual-branch architecture, expands the receptive field and enhances detail preservation by using multi-scale wavelet transform, combines a selective state-space model to model long-range dependencies with linear complexity, and achieves feature complementarity and enhancement through an adaptive cross-domain fusion mechanism. This systematically improves the semantic segmentation accuracy and stability of high-resolution remote sensing images under challenges such as extreme multi-scale, complex backgrounds and long-range dependencies. The present invention also discloses an electronic device, including a central processing unit and a storage module, wherein the storage module stores executable instructions, and when the executable instructions are run by the central processing unit, they control the central processing unit to implement the process of the method described above.

[0051] The present invention also discloses a computer-readable storage medium storing computer-programmable instructions thereon, which, when executed on a computing device, trigger the computing device to perform the operation steps of the method described above.

[0052] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A remote sensing image segmentation method based on multi-scale wavelet transform and Mamba, characterized in that, Includes the following steps: S1. Data Processing: Obtain remote sensing image datasets including ISPRS Potsdam and ISPRS Vaihingen, and preprocess the image data, dividing them into training set, validation set and test set; S2. Constructing the network model: Building a segmented network named MSWT-Mamba; S3. Model Training: Input the training set into the segmentation network for training, and update the network parameters by calculating the loss function and performing backpropagation to obtain the trained parameter model. S4. Model Inference: Input the test set into the trained parameter model and output the semantic segmentation result map of the target image; The segmentation network in step S2 includes at least the following sequentially connected components: The local detail extraction branch, based on a multi-level residual block structure, is used to extract multi-scale local spatial detail features from the input image. A spatial semantic extraction branch, which is based on a hierarchical structure including a multi-scale wavelet transform convolution module and a selective state space module, is used to model the global contextual semantic dependencies of an image. The cross-branch feature fusion part includes multiple dual-branch fusion modules, which are used to adaptively fuse local detail features and global semantic features at the corresponding level to generate fused features; The decoder section is based on an upsampling and skip connection architecture. It uses the fused features to gradually restore the spatial resolution and outputs the final semantic segmentation prediction map.

2. The remote sensing image segmentation method based on multi-scale wavelet transform and Mamba as described in claim 1, characterized in that, The preprocessing in step S1 includes normalization, data augmentation, pruning, and noise reduction, as detailed below: The image data is normalized, converting pixel values ​​to the [0, 1] interval. The calculation formula is as follows: ， in, These are the original pixel values. The minimum pixel value in the image. The maximum pixel value in the image; Data augmentation of image data includes rotation transformation and / or horizontal flipping. Rotation transformation changes the image angle using a rotation matrix, which is: ， in, The x-coordinate of the pixel before rotation. The pixel ordinate before rotation. The x-coordinate of the rotated pixel. The ordinate of the rotated pixel. The rotation angle is expressed in radians. Horizontal flipping is a process of mirroring an image vertically; the corresponding formula is: ， in, For row indexes in the image, For column indexes in the image, The width of the image (i.e., the total number of columns). The image is in position after being horizontally flipped. Pixel value at that location, For the original image at position Pixel value at; Crop the image into fixed-size tiles; To perform Gaussian filtering for noise reduction on the image, the filter kernel function is: ， in, This represents the horizontal coordinate offset of a point within the core relative to the center point. This represents the vertical coordinate offset of a point in the core relative to the center point.

3. The remote sensing image segmentation method based on multi-scale wavelet transform and Mamba as described in claim 1, characterized in that, The spatial semantic extraction branch in step S2 includes multiple processing stages, and the processing procedure of each stage includes: The input features are processed through two separate paths: The first path sequentially passes through the first normalization layer, the first linear transform layer, the multi-scale wavelet transform convolution module, the selective state space module, and the second normalization layer; The second path passes through the first normalization layer and the second linear transformation layer in sequence; After multiplying the outputs of the two paths, the result is passed through the third linear transformation layer and then residually connected with the original input of the branch to obtain the output features of this stage. The multi-scale wavelet transform convolution module is used to perform multi-level wavelet transform decomposition and reconstruction of input features by deploying convolution operations of different scales in parallel, so as to expand the feature receptive field; the selective state space module is based on the state space model and is used to perform global feature extraction and fusion.

4. The remote sensing image segmentation method based on multi-scale wavelet transform and Mamba according to claim 3, characterized in that, The multi-scale wavelet transform convolution module performs multi-level wavelet transform decomposition on the input features by deploying convolution operations of three scales, 1×1, 3×3 and 5×5, in parallel to generate low-frequency and high-frequency components. After convolution processing of each component, the features are reconstructed by inverse wavelet transform. Finally, the reconstructed features are fused with the original features for output.

5. The remote sensing image segmentation method based on multi-scale wavelet transform and Mamba according to claim 1, characterized in that, The local detail extraction branch in step S2 contains multiple residual blocks, and the feature transformation process of the residual blocks is expressed as follows: ， ， in, express Convolution operations, BN represents batch normalization, and ReLU represents the activation function. Indicates the first Output characteristics of each residual block.

6. The remote sensing image segmentation method based on multi-scale wavelet transform and Mamba according to claim 1, characterized in that, The cross-branch feature fusion part in step S2 includes multiple dual-branch fusion modules for fusing the first... Hierarchical local detail features and global semantic features Generate fusion features The calculation process of the dual-branch fusion module is as follows: ， ， ， in, For layer normalization, For channel attention mechanism, For matrix multiplication, For adaptive cross-domain fusion blocks.

7. The remote sensing image segmentation method based on multi-scale wavelet transform and Mamba according to claim 6, characterized in that, The calculation process of the adaptive cross-domain fusion block is as follows: For input features respectively and Perform depthwise separable convolution to generate the query vector. , key vector , value vector , Specifically: Calculate the multi-head attention between the two cross domains: , ; Will and After concatenation, the output is obtained through a 1×1 convolution. ; in, For depthwise separable convolution, This is a multi-head attention mechanism.

8. The remote sensing image segmentation method based on multi-scale wavelet transform and Mamba according to claim 1, characterized in that, The decoder portion in step S2 has an architecture that includes skip connections, combining the fused features from multiple layers with the features from the upsampling process; during training, an auxiliary segmentation head is connected after multiple fused feature layers, and the total loss function is the weighted sum of the final output loss and each auxiliary loss: ， in, , and Both are composite loss functions of Dice loss and cross-entropy loss.

9. An electronic device comprising a central processing unit and a storage module, characterized in that, The storage module stores executable instructions, which, when executed by the central processing unit, control the central processing unit to implement the process of the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having stored thereon computer-programmable instructions, characterized in that, When the programmable instructions are executed on a computing device, they trigger the computing device to perform the operation steps of the method as described in any one of claims 1 to 8.

Images