A remote sensing image segmentation method based on frequency domain global channel perception and cross-channel attention fusion

Abstract

The application discloses a remote sensing image segmentation method based on frequency domain global channel perception and cross-channel attention fusion, and relates to the technical field of remote sensing image processing; the method comprises the following steps: inputting remote sensing image features into a frequency domain, performing amplitude enhancement and phase channel transformation, inversely transforming the reconstructed frequency domain features back to a spatial domain, and obtaining frequency domain enhanced features; a cross-channel attention mechanism is applied to the frequency domain enhanced features, a channel dependency relationship is modeled through query, key and value interaction, and is fused with the frequency domain enhanced features to obtain fused features; the fused features are converted into a segmentation map, and model optimization is performed through a multi-task loss function. The method can ensure global consistency, take into account boundary fineness and small target detection, reduce computational complexity, and improve the practicability and engineering feasibility of the method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, more particularly to a remote sensing image segmentation method based on frequency domain global channel perception and cross-channel attention fusion, which is suitable for geographic information extraction, urban planning, environmental monitoring and other applications. BACKGROUND

[0002] At present, with the development of remote sensing technology, the means of obtaining high-resolution remote sensing images is becoming increasingly diversified, and the data scale and resolution are constantly improving. Remote sensing image segmentation, as a core technology in tasks such as geographic information extraction, urban planning, environmental monitoring, agricultural investigation, and disaster assessment, aims to accurately distinguish and label different ground object categories (such as buildings, roads, water bodies, farmland, forest land, etc.) in remote sensing images.

[0003] However, due to the large number of ground object categories, complex spatial structure, large spectral distribution difference, and often existing problems such as shadows, noise, and scale changes in remote sensing images, the segmentation task is of high difficulty. How to ensure global consistency while taking into account boundary fineness and small target detection has always been a research focus and difficulty in this field.

[0004] Existing remote sensing image segmentation methods are mostly based on deep convolutional neural networks (CNN) and encoder-decoder structures. Such methods have certain advantages in feature extraction and multi-scale fusion, but are limited by the local receptive field characteristics of convolution operators, often difficult to capture global dependencies across regions and long distances in remote sensing images, leading to poor performance in the segmentation of large-scale ground objects (such as urban clusters and rivers). At the same time, the convolution method has limited ability to depict slender structures (such as roads and river channels) and complex boundaries (such as building edges), and is prone to boundary blurring or breaking problems.

[0005] In recent years, the self-attention mechanism based on Transformer has been gradually applied to remote sensing image segmentation. This type of method can model long-range dependencies within the image by constructing similarity relationships between query (Query), key (Key), and value (Value), which helps to improve the global feature expression capability. However, due to the high resolution of remote sensing images, direct calculation of attention matrices in the spatial dimension often brings huge computational and storage overheads, making it difficult to efficiently apply in large-scale data. At the same time, Transformer still has certain deficiencies in modeling fine-grained boundaries, and the performance for small targets and edge regions is not ideal.

[0006] Besides spatial domain methods, frequency domain analysis methods have also gained increasing attention in recent years. Fourier transform can map images from the spatial domain to the frequency domain, thereby separating amplitude and phase information, facilitating the capture of global patterns and periodic features. In remote sensing images, amplitude typically corresponds to the overall structure and intensity distribution, while phase contains rich boundary and texture information. However, most existing frequency domain methods only utilize spectral enhancement or filtering to improve feature representation, lacking a design for modular integration with deep neural networks and difficulty in working in conjunction with attention mechanisms, thus limiting their practicality. Summary of the Invention

[0007] In view of this, the purpose of this invention is to address the shortcomings of existing remote sensing image segmentation methods in terms of global feature extraction and boundary detail characterization, and to provide a remote sensing image segmentation method based on frequency domain global channel perception and cross-channel attention fusion. This method can ensure global consistency while taking into account boundary fineness and small target detection, while reducing computational complexity and improving the practicality and engineering feasibility of the method.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention provides a remote sensing image segmentation method based on frequency domain global channel perception and cross-channel attention fusion, comprising the following steps:

[0010] (1) Frequency domain global channel sensing steps: Map the input remote sensing image features to the frequency domain, perform amplitude enhancement and phase channel transformation, reconstruct the frequency domain features, and then inversely transform back to the spatial domain to obtain the frequency domain enhanced features;

[0011] (2) Frequency domain-guided cross-channel attention fusion step: Apply a cross-channel attention mechanism to the frequency domain enhancement feature, model channel dependencies through query, key, and value interaction, and fuse with the frequency domain enhancement feature to obtain the fused feature;

[0012] (3) Segmentation prediction and loss optimization steps: The fused features are converted into a segmentation map and the model is optimized using a multi-task loss function.

[0013] Further, step (1) includes:

[0014] (1.1) Input remote sensing image features Normalization is performed; the result is obtained after convolutional layers and nonlinear activation functions. ; After multilayer perceptron and Softmax normalization, preliminary channel attention weights are generated. Where R represents the set of real numbers, C represents the number of channels, and H and W represent the height and width of the feature map, respectively.

[0015] (1.2) will By two-dimensional fast Fourier transform, we get: ;

[0016] where, is the amplitude spectrum, is the phase spectrum; the enhanced amplitude is obtained by convolution and nonlinear transformation on the amplitude spectrum ; ;

[0017] (1.3) The channel fast Fourier transform is performed on the phase spectrum ; ; After convolution and nonlinear activation on A2 and P2, the complex frequency spectrum is reconstructed, and the enhanced channel weight ;

[0018] (1.4) Multiply the phase spectrum with the channel enhancement weight : ; and reconstruct the complex frequency spectrum with the enhanced amplitude : ; and the channel attention weight fusion, and recover to the time domain through inverse Fourier transform: .

[0019] Further, the step (1.3) reconstructs the complex frequency spectrum as follows:

[0020]

[0021] where, are the results of convolution and nonlinear activation on A2 and P2, respectively; e is the natural constant.

[0022] Further, the step (2) includes:

[0023] (2.1) Layer normalization is performed on the input remote sensing image feature F, and three independent linear layers are mapped to obtain the query Q, the key K and the value V;

[0024] (2.2) The enhanced query Q' is obtained by performing convolution, LeakyReLU activation, depth separable convolution and multilayer perception processing on the query Q; the global representation K' is obtained by performing global average pooling on the key K;

[0025] (2.3) The similarity distribution of the query and the key is calculated, and the attention output is obtained by acting on the value V vector;

[0026] (2.4) The attention output is fused with the frequency domain enhanced feature Additive fusion, get the final output fusion feature .

[0027] Further, in the step (2.3), the cross-channel attention output is calculated:

[0028]

[0029] wherein, denotes tensor dot product or broadcast multiplication, and d is a scaling factor.

[0030] Further, the step (3) comprises:

[0031] (3.1) performing convolution and normalization processing on the fusion feature to compress the channel dimension, and then mapping to the class space through 1×1 convolution;

[0032] (3.2) obtaining the class probability distribution through pixel-wise Softmax activation, and obtaining the final segmentation result through probability maximization;

[0033] (3.3) adopting a multi-task loss function, including cross-entropy loss, Dice loss and boundary perception loss;

[0034] (3.4) performing joint optimization based on weighted fusion of all loss functions.

[0035] Further, the step (3.2) comprises:

[0036] Through the pixel-wise Softmax activation function, the class probability distribution of each position is obtained:

[0037]

[0038] wherein, is a prediction probability tensor, denotes the probability of pixel (i,j) belonging to any class n; N denotes the number of classes of the remote sensing image; and m denotes an index value. denotes the unnormalized prediction value of the corresponding class n output by the deep neural network at the pixel position (i,j); denotes the unnormalized prediction value of the corresponding class m output by the deep neural network at the pixel position (i,j), which is used for the normalized sum in the denominator;

[0039]

[0040] denotes that the final segmentation result is obtained through probability maximization.

[0041] According to the technical solution described above, compared with the prior art, the present application has the following beneficial effects:

[0042] 1) Significantly improve global semantic consistency and boundary refinement capability.

[0043] The application separates the input features in the frequency domain by amplitude and phase and reconstructs the channel-level spectrum, enabling the model to capture long-distance dependencies and periodic structure information in remote sensing images, such as the overall distribution of coastlines, farmland patterns, and urban blocks. Meanwhile, the channel attention mechanism selectively highlights key channel features for segmentation results, and the cross-channel attention mechanism fuses the frequency domain enhancement results with channel interaction features, thereby achieving more detailed depiction of complex object boundaries and detail areas. Compared with existing methods that rely solely on spatial convolution or self-attention, the application can effectively improve boundary accuracy and the segmentation quality of thin structures (such as roads and rivers) while maintaining overall semantic consistency.

[0044] 2) Improve robustness and computational efficiency, facilitating engineering deployment.

[0045] Traditional methods based on full-space self-attention often require the calculation of large-scale similarity matrices on high-resolution remote sensing images, resulting in huge computational and storage overheads, making it difficult to promote in practical applications. The application uses a channel-based attention mechanism and a frequency domain channel transformation strategy to focus on channel dimensions and low-dimensional representations in the frequency domain, significantly reducing computational complexity and memory consumption. Meanwhile, depth separable convolution and lightweight mapping layers are introduced in feature enhancement, effectively reducing the number of model parameters, enabling the method to maintain segmentation accuracy while having better running efficiency and robustness. This feature not only applies to diverse remote sensing image data, but also facilitates modular integration with existing backbone networks, thereby improving the feasibility and application value of actual engineering deployment. BRIEF DESCRIPTION OF DRAWINGS

[0046] 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 are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on the provided drawings.

[0047] Figure 1 Flowchart of the remote sensing image segmentation method based on frequency domain global channel perception and cross-channel attention fusion.

[0048] Figure 2 Principle diagram of the frequency domain global channel perception step.

[0049] Figure 3 Principle diagram of the frequency domain guided cross-channel attention fusion step. DETAILED DESCRIPTION

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Reference Figure 1 As shown in the figure, this invention discloses a remote sensing image segmentation method based on frequency domain global channel perception and cross-channel attention fusion, including the following steps:

[0052] (1) Frequency domain global channel sensing steps: Map the input remote sensing image features to the frequency domain, perform amplitude enhancement and phase channel transformation, reconstruct the frequency domain features, and then inversely transform back to the spatial domain to obtain the frequency domain enhanced features;

[0053] (2) Frequency domain-guided cross-channel attention fusion step: Apply a cross-channel attention mechanism to the frequency domain enhancement feature, model channel dependencies through query, key, and value interaction, and fuse with the frequency domain enhancement feature to obtain the fused feature;

[0054] (3) Segmentation prediction and loss optimization steps: The fused features are converted into a segmentation map and the model is optimized using a multi-task loss function.

[0055] This invention introduces a frequency-domain global channel sensing step and a frequency-domain guided cross-channel attention fusion step to achieve higher segmentation accuracy and stronger generalization ability for ground objects with multi-scale, complex textures, and slender boundaries in remote sensing images. Specifically, this invention aims to improve the class discrimination ability and boundary fidelity of remote sensing image segmentation in complex scenes, such as high-density towns, narrow roads, winding rivers, and small targets, especially enhancing the separability and localization accuracy of small targets and slender structures; reducing the sensitivity to noise, spectral variations, and class imbalance in training data, while taking into account computational complexity and the number of parameters, making the method easy to integrate with existing backbone networks and possessing good real-time performance and deployability in practical remote sensing applications.

[0056] The following is a detailed explanation of each of the above steps:

[0057] Step 1: Frequency Domain Global Channel Sensing;

[0058] like Figure 2As shown. This step aims to extract key information across space and channels from remote sensing image features through frequency domain global modeling and channel attention fusion mechanisms, thereby improving the ability to identify complex feature boundaries and detailed regions in remote sensing image segmentation tasks. The core idea of ​​this module is to map the input features to the frequency domain, extract its global features using Fourier transform, and then enhance them through nonlinear mapping and channel selection mechanisms. The specific process is as follows:

[0059] Step 1.1: Normalization and Initial Generation of Channel Weights. Assume the input features are...

[0060] Where R represents the set of real numbers, C represents the number of channels, and H and W represent the height and width of the feature map, respectively. First, F is normalized, either using LayerNorm or InstanceNorm; then, it passes through a convolutional layer Conv(·) and a non-linear activation function StarReLU σ(·), resulting in:

[0061]

[0062] Subsequently, After multilayer perceptron (MLP) and Softmax normalization, preliminary channel attention weights are generated: .

[0063] Step 1.2: Frequency domain amplitude enhancement. [The following appears to be a separate, unrelated sentence:] ... Through two-dimensional Fast Fourier Transform (FFT):

[0064]

[0065] in, For amplitude spectrum, For the phase spectrum. For the amplitude spectrum. Convolution and nonlinear transformations are performed to enhance salient structures such as building boundaries and road textures in remote sensing images:

[0066]

[0067] Step 1.3: Phase Channel Transformation and Complex Spectrum Reconstruction. For the phase spectrum... Performing a channel fast Fourier transform (CFFT) yields:

[0068]

[0069] Convolution and non-linear activation are performed on A2 and P2 respectively:

[0070] ;

[0071] Then, complex spectrum reconstruction is performed:

[0072]

[0073] Where e refers to the natural constant e≈2.71828. The enhanced channel weights are then obtained through inverse fast Fourier transform (ICFFT).

[0074]

[0075] Step 1.4: Frequency Domain Fusion and Feature Update. The phase spectrum... With channel enhancement weights Multiplication:

[0076]

[0077] And then the enhanced amplitude Perform complex spectrum reconstruction:

[0078]

[0079] at last, With global weight The components are merged and then restored to the time domain via inverse Fourier transform:

[0080]

[0081] This output It preserves the global frequency domain information of remote sensing images and selectively highlights key land cover categories, such as farmland, water bodies, and buildings, through channel weights, thereby improving the accuracy and robustness of remote sensing image segmentation.

[0082] Step 2: Construct a frequency-domain guided cross-channel attention fusion mechanism.

[0083] like Figure 3 As shown, this step, building upon the first step of frequency domain global channel perception, introduces a cross-channel attention mechanism to achieve multi-dimensional interactive modeling of remote sensing image features. Through the mapping of queries, keys, and values, and frequency domain-guided weight updates, this module can more accurately characterize the spatial distribution relationships and category dependencies of complex features in remote sensing images.

[0084] Step 2.1: Input Feature Mapping. Let the input features be... First of all Perform layer normalization (LayerNorm) to ensure distribution consistency among different samples:

[0085]

[0086] The query, key, and value are then obtained using three independent linear layer mappings:

[0087]

[0088] in These are linear mapping matrices.

[0089] Step 2.2: Feature Enhancement of Query and Key. The query Q first passes through a convolutional layer, a LeakyReLU activation function, and a depthwise separable convolution (DWConv), and finally passes through a multilayer perceptron (MLP) to obtain the enhanced query vector Q′:

[0090]

[0091] Meanwhile, the global representation of key K is obtained through global average pooling (GAP):

[0092]

[0093] This operation enables K' to aggregate global contextual information in remote sensing images, making it particularly suitable for extracting statistical features of large-scale land features, such as urban clusters and rivers.

[0094] Step 2.3: Cross-channel attention calculation. Construct the interaction relationship between the query and the key, and apply it to the value vector V:

[0095]

[0096] in, This represents tensor dot product or broadcast multiplication, where d is a scaling factor (usually the channel dimension). This process is equivalent to a channelized form of self-attention, essentially using the similarity distribution of queries and keys to adjust the value vector, thereby enhancing cross-class dependencies in remote sensing image features. For example, in building-road interaction regions, this mechanism can guide the prediction of building outlines using road boundary features.

[0097] Step 2.4: Frequency Domain and Attention Fusion. To further improve the robustness of feature representation, a frequency domain global channel awareness step one is introduced. Finally, and By adding and fusing, we obtain the output features:

[0098]

[0099] This fusion approach retains the attention mechanism's ability to capture local dependencies while combining it with the frequency domain module's ability to model global patterns, thus enabling... In remote sensing image segmentation, it can simultaneously identify large-scale global structures, such as coastlines and field outlines, as well as fine boundaries, such as road intersections and river branches.

[0100] Step 3: Construct a segmentation prediction and loss optimization mechanism.

[0101] This step aims to integrate the features after frequency domain enhancement and cross-channel attention fusion. The results are converted into the final remote sensing image segmentation results, and the model is improved in terms of class discrimination, boundary fineness and global consistency through joint optimization of multiple loss functions.

[0102] Step 3.1: Feature compression and semantic mapping. Input features First, convolutional layers and normalization operations are applied to compress the channel dimension and aggregate contextual information:

[0103]

[0104] Subsequently, a 1×1 convolutional layer is used to map the features to the class space:

[0105]

[0106] Where N represents the number of categories in the remote sensing image, such as buildings, roads, water bodies, farmland, etc.

[0107] Step 3.2: Probability Prediction and Segmentation Result Generation. The class probability distribution at each location is obtained by applying a pixel-by-pixel Softmax activation function:

[0108]

[0109] in To predict the probability tensor, This represents the probability that pixel (i,j) belongs to category n. N represents the number of categories in the remote sensing image; m represents the index value. This represents the unnormalized prediction (logit) of the corresponding class n at pixel position (i,j) output by the deep neural network, where n∈{1,2,…,N}. It's still a logit at position (i,j), but with index m, used for normalized summation of the denominator. The final segmentation result is obtained by maximizing the probability:

[0110]

[0111] Step 3.3: Multi-task loss function design. To ensure a balance in category differentiation, boundary details, and global consistency in remote sensing image segmentation results, this invention employs a multi-task loss function:

[0112] Cross-entropy loss (CE Loss) – emphasizes pixel-level classification accuracy:

[0113]

[0114] in, This is the one-hot representation of the real label.

[0115] Dice Loss – mitigates class imbalance and improves the recognition of small targets, such as roads and rivers.

[0116]

[0117] Boundary-aware loss (optional) – emphasizes narrow boundaries such as roads and rivers in remote sensing images:

[0118]

[0119] Here, ∇ represents the gradient operator, used to compare the consistency between the predicted and the true segmentation boundary.

[0120] Step 3.4: Final Loss and Optimization Objective. The final total loss function adopts a weighted fusion form:

[0121]

[0122] in, These are the weight hyperparameters. Minimize... The model can achieve a balance between global accuracy and local boundary refinement, thus obtaining better remote sensing image segmentation results.

[0123] Example: Building and road segmentation in high-resolution remote sensing imagery:

[0124] Using remote sensing imagery of a certain urban area at a resolution of 0.5 m as experimental data, the image size is 1024×1024, and the number of categories N=4, corresponding to "buildings, roads, water bodies, and farmland". The segmentation method proposed in this invention, based on frequency domain global channel perception and cross-channel attention fusion, is employed. The specific process is as follows:

[0125] 1. Frequency Domain Global Channel Awareness:

[0126] The input image is convolved and normalized, then mapped to the frequency domain to obtain the amplitude and phase spectra. Convolution and nonlinear transformations are applied to the amplitude spectrum to enhance edge and texture structures; a channel Fast Fourier Transform (CFFT) is performed on the phase spectrum, and the complex spectrum is reconstructed to obtain enhanced channel weights. These are then combined with the enhanced amplitude spectrum and returned to the spatial domain via an inverse Fourier transform to obtain the frequency domain enhanced feature map.

[0127] 2. Cross-channel attention fusion:

[0128] The frequency domain enhanced feature map is layer-normalized to obtain the query Q, key K, and value V. The query is processed by convolution, LeakyReLU, depthwise separable convolution, and multilayer perceptron to obtain Q′, and the key is processed by global average pooling to obtain K′. The similarity distribution between Q′ and K′ is used to adjust the value vector V, forming the attention output. This output is added to and fused with the frequency domain enhanced features to obtain the final fused feature.

[0129] 3. Segmentation Prediction and Loss Optimization:

[0130] The fused features are convolutionally and normalized to compress the channel dimensions, then mapped to a 4-class space via a 1×1 convolution. For each pixel location (i,j), the class probability distribution is obtained through pixel-wise softmax. The category with the highest probability is taken as the final prediction result. .

[0131] During the training phase, a weighted combination of cross-entropy loss, Dice loss, and boundary-aware loss is used to jointly optimize the model parameters.

[0132] This invention proposes a remote sensing image segmentation method combining frequency domain enhancement and cross-channel attention fusion, which can be integrated into existing deep learning backbone networks for remote sensing image segmentation tasks. First, this invention designs a frequency domain global channel awareness mechanism, mapping input features to the frequency domain. Through amplitude enhancement and phase channel transformation, combined with complex spectrum reconstruction and channel weight adjustment, it achieves effective capture of global information and selective highlighting of key information. This not only enhances the global consistency of large-scale features in remote sensing images but also improves the ability to identify key categories through the channel selection mechanism.

[0133] Secondly, this invention proposes a frequency-domain guided cross-channel attention fusion mechanism. Based on frequency domain enhancement, it introduces interactive modeling of queries, keys, and values, organically combining global context with channel dependencies to achieve collaborative expression of multi-dimensional features. This mechanism can effectively improve the class distinction and boundary delineation accuracy of complex features, and is particularly suitable for scenarios in remote sensing imagery where large-scale structures and fine-grained boundaries coexist.

[0134] Finally, this invention achieves a balance between category accuracy, boundary details, and global consistency through joint optimization using a multi-task loss function, thereby obtaining superior segmentation results.

[0135] In summary, this invention innovatively introduces a frequency domain global modeling and cross-channel attention fusion mechanism based on traditional convolution and attention methods. This not only solves the shortcomings of existing methods in global feature extraction and boundary detail characterization, but also takes into account computational efficiency and practical engineering feasibility, and has strong innovative value and application prospects.

[0136] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0137] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A remote sensing image segmentation method based on frequency domain global channel perception and cross-channel attention fusion, characterized in that, Includes the following steps: (1) Frequency domain global channel sensing steps: Map the input remote sensing image features to the frequency domain, perform amplitude enhancement and phase channel transformation, reconstruct the frequency domain features, and then inversely transform back to the spatial domain to obtain the frequency domain enhanced features; (2) Frequency domain-guided cross-channel attention fusion step: Apply a cross-channel attention mechanism to the frequency domain enhancement feature, model channel dependencies through query, key, and value interaction, and fuse with the frequency domain enhancement feature to obtain the fused feature; (3) Segmentation prediction and loss optimization steps: The fused features are converted into a segmentation map, and the model is optimized using a multi-task loss function; Step (1) includes: (1.1) Input remote sensing image features Normalization is performed; the result is obtained after convolutional layers and nonlinear activation functions. ; After multilayer perceptron and Softmax normalization, preliminary channel attention weights are generated. Where R represents the set of real numbers, C represents the number of channels, and H and W represent the height and width of the feature map, respectively. (1.2) will The following can be obtained through two-dimensional fast Fourier transform: ; in, For amplitude spectrum, For the phase spectrum; for the amplitude spectrum The enhancement amplitude is obtained by performing convolution and nonlinear transformation: ; (1.3) Phase spectrum Performing a channel fast Fourier transform yields: ; respectively A 2 and P 2. After convolution and nonlinear activation, the complex spectrum is reconstructed, and then enhanced channel weights are obtained through channel fast Fourier transform. ; (1.4) Phase spectrum With channel enhancement weights Multiplication: ; and then the enhanced amplitude Perform complex spectrum reconstruction: ; With channel attention weights The components are merged and then restored to the time domain via inverse Fourier transform: ; Step (2) includes: (2.1) Perform layer normalization on the input remote sensing image features F, and obtain the query through three independent linear layer mappings. Q ,key K Sum V ; (2.2) For queries Q Enhanced queries are obtained by performing convolution, LeakyReLU activation, depthwise separable convolution, and multilayer perceptron processing. Q ′;Key K Global average pooling is performed to obtain the global representation. K ′; (2.3) Calculate the similarity distribution between the query and the key, applying it to the value. V Vectors yield attention output ; (2.4) Outputting attention Frequency domain enhancement features The features are added together and fused to obtain the final output fused feature. ; In step (2.3), the cross-channel attention output is calculated: in, This represents tensor dot product or broadcast multiplication. d This is the scaling factor.

2. The method according to claim 1, characterized in that, The complex spectrum is reconstructed in step (1.3) as follows: in, They are respectively for A 2 and P 2. The result after convolution and nonlinear activation; e is the natural constant.

3. The method according to claim 1, characterized in that, Step (3) includes: (3.1) The fused features are convolved and normalized to compress the channel dimension, and then mapped to the category space through 1×1 convolution; (3.2) The class probability distribution is obtained by pixel-wise Softmax activation, and the final segmentation result is obtained by maximizing the probability; (3.3) A multi-task loss function is adopted, including cross-entropy loss, Dice loss and boundary-aware loss; (3.4) Joint optimization is performed based on the weighted fusion of all loss functions.

4. The method according to claim 3, characterized in that, Step (3.2) includes: The class probability distribution at each location is obtained by applying a pixel-by-pixel softmax activation function: in, To predict the probability tensor, This represents the probability that pixel (i,j) belongs to any category n; N represents the number of categories in the remote sensing image; m represents the index value; This represents the unnormalized prediction value of the corresponding class n output by the deep neural network at pixel position (i,j); This represents the unnormalized prediction value of the corresponding class m output by the deep neural network at pixel position (i,j), which is used for the normalized summation of the denominator; This means that the final segmentation result is obtained by maximizing the probability.

Images