Remote sensing image semantic segmentation method based on frequency domain cooperation and geometric perception alignment

Abstract

The application belongs to the field of image semantic segmentation, and particularly discloses a remote sensing image semantic segmentation method based on frequency domain cooperation and geometric perception alignment, wherein a phase-amplitude cooperative frequency selection module (PAC-FSM) is designed at the encoder end, structural gating modulation of amplitude spectrum features is generated by using a phase spectrum, background noise is suppressed, and target boundaries are sharpened; at the decoder end, a lossless wavelet down-sampling module (WB-LDS) based on discrete wavelet transform is introduced, and a geometric perception frequency alignment module (GA-FAM) is combined, multi-scale strip convolution is used to capture anisotropic long-distance coherent geometric structure, and explicit alignment of the spatial domain and the frequency domain is realized; finally, high-frequency geometric features are injected into decoding features through adaptive scaling parameters. The application effectively alleviates the low-pass filtering effect of the state space model and the information loss problem of down-sampling, and improves the remote sensing image segmentation precision.

Description

Technical Field

[0001] This invention belongs to the field of image semantic segmentation, specifically relating to a remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment. Background Technology

[0002] With the rapid development of Earth observation technology, semantic segmentation of remote sensing images has gradually become one of the most critical application branches in this field. It plays a vital role in practical applications such as land use, urban planning, disaster assessment, and agricultural monitoring by performing pixel-level classification of high-resolution images acquired by satellite or airborne sensors.

[0003] Looking back at the development of semantic segmentation of remote sensing images, early research mainly relied on fully convolutional neural networks (FCNs) and DeepLab series convolutional neural networks (CNNs). CNNs offer high processing efficiency, but are limited by inherent inductive bias, making it difficult to effectively establish global long-range pixel dependencies. Subsequently, the introduction of the Transformer architecture broke the limitation of local receptive fields, demonstrating good performance in global context modeling through its self-attention mechanism. However, the self-attention mechanism has a quadratic computational complexity of O(L^2) with respect to sequence length. 2 (L is the sequence length), processing high-resolution remote sensing images leads to an exponential increase in memory overhead and computation time. Recently, the visual Mamba architecture based on the State Space Model (SSM) has successfully achieved global modeling with a computational complexity of O(L) by introducing a linear selective scanning mechanism. However, modern visual state space models still face two major algorithmic bottlenecks in dense prediction tasks in remote sensing:

[0004] 1. Low-pass filtering effect leads to overly smooth high-frequency physical boundaries: Mamba-based selective scanning mechanisms often exhibit effects similar to low-pass filters, excelling at modeling smooth, continuous regional semantics, but easily smoothing out abrupt physical boundaries and fine-grained textures. Traditional frequency domain enhancement methods typically directly modulate amplitude spectrum features while ignoring the spatial structure and edge contour information contained in the phase spectrum, easily leading to the introduction of unexpected visual artifacts during feature reconstruction.

[0005] 2. Multi-scale downsampling leads to permanent information loss and geometric topological breaks: In the multi-scale decoding and feature fusion stages, existing segmentation networks generally rely on pooling operations or stride convolutions. This lossy downsampling results in the permanent loss of high-frequency spatial details. Furthermore, although Mamba possesses a global receptive field, when explicitly capturing slender and coherent geometric priors (such as roads and rivers in remote sensing images), the lack of anisotropic structure modeling still leads to spatial topological breaks. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment. A phase-amplitude collaboration mechanism is introduced at the encoder end, and lossless wavelet downsampling and geometric perception multi-scale strip convolution are combined at the decoder end. This effectively alleviates the low-pass filtering effect and geometric topological breakage problem of the state space model, and achieves high-precision remote sensing image semantic segmentation.

[0007] The objective of this invention is achieved through the following technical solution: a semantic segmentation method for remote sensing images based on frequency domain collaboration and geometric perception alignment, comprising the following steps:

[0008] Step S1: Obtain the remote sensing image data to be segmented, input it into the encoder composed of a hierarchical feature extraction network, and extract multi-scale spatial semantic features at different levels.

[0009] Step S2: In the deep feature path of the encoder, a phase-amplitude coordination module is introduced to convert the spatial features to the frequency domain and decouple them into amplitude spectrum and phase spectrum. The structural information extracted from the phase spectrum is used to modulate the amplitude spectrum features, and the frequency domain enhanced features are output after reconstruction.

[0010] Step S3: Input the multi-scale spatial semantic features into the decoder, introduce lossless wavelet downsampling in the decoding path, decompose the feature map into low-frequency components and high-frequency components, and peel off the pure high-frequency detail features while constructing multi-scale context features.

[0011] Step S4: Input the pure high-frequency detail features into the geometrically aware frequency alignment module, and use multi-scale anisotropic convolution branches to extract long-distance coherent geometric features to achieve explicit alignment of spatial domain and frequency domain features.

[0012] Step S5: Adaptively inject the aligned high-frequency geometric features into the decoded backbone features, combine multi-scale context fusion features to perform pixel-level classification, and output the semantic segmentation results of the remote sensing image.

[0013] The beneficial effects of this invention are as follows: This invention achieves high-precision semantic segmentation of remote sensing images by mitigating the low-pass filtering effect of the state-space model and the information loss problem of traditional downsampling. First, to address the low-pass filtering effect of the state-space model, this invention designs a phase-amplitude coordinated frequency selection module (PAC-FSM), which overcomes the limitations of traditional direct filtering. It uses the phase spectrum as a structure-guided gating mechanism to finely adjust the amplitude spectrum, sharpening the target boundary while suppressing background noise. Furthermore, wavelet transform downsampling based on discrete wavelet transform is designed at the decoder end, avoiding the information loss caused by traditional convolutional downsampling. Finally, this invention proposes a geometrically aware frequency alignment module (GA-FAM). Multi-scale strip convolution is introduced into the Mamba architecture to achieve alignment between the spatial and frequency domains, enhancing the model's ability to capture anisotropic and long-range coherent structures. Attached Figure Description

[0014] Figure 1 This is a diagram of the overall architecture of the remote sensing image semantic segmentation network (FG-RSSegMAN) based on the frequency domain and geometric perception state space model of this invention.

[0015] Figure 2 This is a structural diagram of the Phase-Amplitude Coordinated Frequency Selection Module (PAC-FSM) in this invention;

[0016] Figure 3 This is a diagram of the lossless wavelet downsampling (WB-LDS) structure in this invention;

[0017] Figure 4 This is a structural diagram of the geometrically sensed frequency alignment module (GA-FAM) in this invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0019] This invention improves upon the SegMAN model by introducing frequency domain and geometric perception mechanisms to enhance the accuracy of semantic segmentation of remote sensing images.

[0020] This invention, based on frequency domain collaboration and geometrically aware alignment, performs the following targeted improvements on SegMAN: Encoder side: It inherits the hierarchical coding structure of SegMAN (layers 1-4 of the encoder), retaining the LASS modules in the first two layers, but explicitly introduces the PAC-FSM module of this invention in parallel in the third and fourth layers for frequency domain enhancement. Decoder side: It replaces the traditional convolutional downsampling of SegMAN with the WB-LDS (lossless wavelet downsampling) of this invention, and combines it with GA-FAM (geometrically aware frequency alignment module) to extract long-range coherent geometric features. Specifically:

[0021] like Figure 1 As shown, the remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment includes the following steps:

[0022] Step S1: Obtain the remote sensing image data to be segmented, input it into the encoder composed of a hierarchical feature extraction network, and extract multi-scale spatial semantic features at different levels.

[0023] Step S2: In the deep feature path of the encoder, a phase-amplitude coordination module is introduced to convert the spatial features to the frequency domain and decouple them into amplitude spectrum and phase spectrum. The structural information extracted from the phase spectrum is used to modulate the amplitude spectrum features, and the frequency domain enhanced features are output after reconstruction.

[0024] Step S3: Input the multi-scale spatial semantic features into the decoder, introduce lossless wavelet downsampling in the decoding path, decompose the feature map into low-frequency components and high-frequency components, and peel off the pure high-frequency detail features while constructing multi-scale context features.

[0025] Step S4: Input the pure high-frequency detail features into the geometrically aware frequency alignment module, and use multi-scale anisotropic convolution branches to extract long-distance coherent geometric features to achieve explicit alignment of spatial domain and frequency domain features.

[0026] Step S5: Adaptively inject the aligned high-frequency geometric features into the decoded backbone features, combine multi-scale context fusion features to perform pixel-level classification, and output the semantic segmentation results of the remote sensing image.

[0027] Combination Figures 1 to 4 The data flow and algorithm optimization process of the method of the present invention includes the following steps in detail:

[0028] In step S1, the remote sensing image to be segmented is acquired and input into an encoder composed of a hierarchical feature extraction network for shallow semantic feature extraction.

[0029] In the first layer of the encoder, the input remote sensing image is downsampled once and the LASS module, which integrates neighborhood attention and 2D selective scanning, to obtain the feature map F1 output by the first layer of the encoder.

[0030] In the second layer of the encoder, F1 obtains the feature map F2 of the second layer output of the encoder through a downsampling and LASS module.

[0031] In step S2, in the third and fourth layers of the deep encoder stage, the spatial domain feature map X is passed to the PAC-FSM module to perform frequency domain decoupling and enhancement: the processing procedure of the module for the input spatial feature map X is as follows:

[0032] (1) Fast Fourier Transform: The PAC-FSM module uses the existing 2D real-number fast Fourier transform (2D rFFT) to transform the input spatial feature map X into the frequency domain:

[0033] ;

[0034] Where H and W represent the height and width of the input spatial feature map X, respectively; h and w represent the row and column coordinate indices in the spatial domain, respectively; and u and v represent the vertical and horizontal spatial frequency coordinates in the two-dimensional frequency domain mapping, respectively. It is the frequency domain representation of complex numbers after compression due to the symmetry of real numbers.

[0035] (2) Separate the amplitude spectrum and phase spectrum, and use existing complex number calculation methods to... Strict decoupling into amplitude spectrum A and phase spectrum P:

[0036] ;

[0037] ;

[0038] Where Re and Im represent the real part and imaginary part of the complex number, respectively.

[0039] (3) Three-branch coordinated modulation: The PAC-FSM module has three branches, namely the global branch, the amplitude branch, and the phase-guided branch:

[0040] (a) Amplitude branch modulation: After the amplitude spectrum A is input into the amplitude branch, this invention uses a lightweight 3×3 depthwise separable convolution to extract features. Compared to the traditional 1×1 convolution, this mechanism can effectively capture the local frequency domain correlation between neighboring frequency points, thereby enhancing the model's ability to perceive the local frequency domain context.

[0041] ;

[0042] (b) Phase-guided branch modulation uses 1×1 convolutions to reduce the phase spectrum P to 1 / 4 of the original number of channels. This allows the network to extract the most essential features representing structural information from the input phase with a limited number of parameters. After the ReLU activation function, a 1×1 convolution is used to restore the phase spectrum P to its original number of channels, and finally, a Sigmoid activation function is used to generate the gating. This "bottleneck structure" of two convolutional layers combined with the intermediate ReLU has a much stronger expressive power than a single convolutional layer, allowing the model to learn complex, non-linear features.

[0043] ;

[0044] in This represents the Sigmoid activation function.

[0045] (c) Global branch modulation: After the amplitude spectrum A is input into the global branch, adaptive average pooling is first used to compress the high-frequency and low-frequency energies in the entire feature map into a single global feature. Then, the same two-layer convolution and sigmoid activation function as the phase-guided branch are used to control the enhancement intensity based on the overall tone of the image.

[0046] ;

[0047] (d) Calculate the enhanced amplitude spectrum by multiplying the outputs of the three branches together and then multiplying by a learnable parameter γ initialized to zero. Add this result to the original amplitude spectrum A to obtain the enhanced amplitude spectrum A. enhanced :

[0048] ;

[0049] in This is element-wise multiplication.

[0050] (4) Inverse Fourier Transform (iFFT): The spatial domain features are reconstructed in strict phase-preserving manner using the inverse Fourier transform.

[0051] ;

[0052] After completing the construction and processing of the PAC-FSM module, perform adaptive residual injection:

[0053] In the third and fourth layers of the encoder, the PAC-FSM is connected in parallel with the LASS Block and the global self-attention module, respectively;

[0054] The third layer of the encoder is constructed by parallel connection of PAC-FSM and LASS Block. The input features of the third layer are the feature map F2 output from the second layer of the encoder, and the output is a fused feature map. :

[0055] ;

[0056] in, This represents the original spatial feature map extracted by the LASS module in the third layer of the encoder; This represents the frequency domain feature mapping obtained after the input features pass through the PAC-FSM module in the third layer of the encoder;

[0057] The fourth layer of the encoder is constructed using a PAC-FSM in parallel with a global self-attention module; the input features of the fourth layer of the encoder are the output features of the third layer of the encoder. Output fused feature map :

[0058] ;

[0059] in, This represents the global spatial feature map extracted by the global self-attention module in the fourth layer of the encoder. This represents the frequency domain feature mapping obtained after the input features pass through the PAC-FSM module in the 4th layer of the encoder; The learnable parameters used to control the fusion process are initialized to 0. This design ensures that the model is equivalent to the original baseline model in the early stages of training, thereby guaranteeing extremely high training stability.

[0060] In frequency domain feature extraction, existing techniques typically only utilize amplitude spectrum or discrete cosine transform (DCT) to generate channel attention, ignoring the spatial geometric topology carried in the phase spectrum. This results in the model's inability to accurately locate complex, elongated feature boundaries in remote sensing images. Alternatively, data augmentation can be achieved by injecting random offsets into the phase spectrum, but directly modifying the phase spectrum easily disrupts spatial topological coherence, leading to decreased edge localization accuracy. Unlike these approaches, the Phase-Amplitude Coordinated Frequency Selection Module (PAC-FSM) proposed in this invention explicitly maintains the original phase spectrum without direct modification and uses the phase spectrum as a structure guide to selectively modulate the amplitude spectrum. This "phase-guided amplitude" mechanism avoids the risk of topological destruction caused by phase distortion while achieving precise enhancement of the physical contours of features. Furthermore, the PAC-FSM module utilizes inverse Fourier transform (iFFT) to reconstruct the polar coordinate space of frequency domain information. The amplitude spectrum enhanced by PAC-FSM... It provides a clearer representation of texture and energy distribution, while the original, unaltered phase spectrum P fully preserves the spatial topology of the ground features. By performing an inverse Fourier transform, this invention can losslessly map the frequency-selective and enhanced features back to the spatial domain, thereby outputting a feature map with more accurate semantic information and clearer geometric edges.

[0061] In step S3, the decoder receives multi-scale features F2, F3, and F4, and introduces the WB-LDS module into the multi-scale context construction path:

[0062] In step S3, the decoder receives the output features F2, F3, and F4 from the second, third, and fourth layers of the encoder, where F3 and F4 are the features from step S2. and The feature F obtained by fusing these three layers of features is then input into the WB-LDS module for processing. The processing includes:

[0063] Step S301: The spatial domain is transformed into the frequency domain. Using the Haar discrete wavelet transform, a signal processing technique well-known in the art, the spatial domain features are converted into four frequency domain components. A, H, V, and D represent the low-frequency component, the horizontal high-frequency component, the vertical high-frequency component, and the diagonal high-frequency component, respectively.

[0064] ;

[0065] Step S302: Decouple the high-frequency features and the main features. After concatenating the three high-frequency information components, perform feature representation learning on the high-frequency information using 1×1 convolution, batch normalization, and ReLU activation function to obtain the high-frequency features. Simultaneously, to preserve the complete features, in another branch, concatenate the four frequency domain components and perform the same operation as with the high-frequency information to obtain the main features.

[0066] ;

[0067] ;

[0068] While existing wavelet downsampling techniques incorporate wavelet transform, they often directly use 1×1 convolution to channel-mix high-frequency and low-frequency components after feature decomposition. This mixing mechanism causes early overlap between high-frequency boundary features and low-frequency semantic features, preventing subsequent networks from independently and explicitly extracting and utilizing high-frequency geometric boundaries. In contrast, this invention designs a feature decomposer based on discrete wavelet transform, constructing a dual-stream architecture that decouples high-frequency features from the main features. Specifically, while submitting the mixed features encompassing the entire frequency domain to the subsequent state-space model for long-range context modeling, this invention explicitly extracts pure high-frequency features, transmitting them independently of the state-space module to the Geometric Aware Frequency Alignment (GA-FAM) module for geometric topology reconstruction. This decoupled dual-stream design not only avoids information loss during downsampling from max pooling and stride convolution but also mitigates the low-pass smoothing effect of the selective scanning mechanism of the state-space model on high-frequency features, resulting in sharper boundaries in the segmentation map.

[0069] In step S4, anisotropic geometric sensing and frequency domain alignment are performed based on the geometric sensing frequency alignment module (GA-FAM):

[0070] Pure high-frequency detail features decoupled from the WB-LDS module The input is fed in parallel to the geometry-aware frequency alignment module (GA-FAM):

[0071] Step S401: Multi-scale strip convolution geometric extraction. The strip convolution branch consists of parallel vertical and horizontal strip convolutions, which are then summed and subjected to batch normalization and nonlinearization of the ReLU activation function.

[0072] ;

[0073] in for High-frequency features obtained by mapping 1×1 convolution, batch normalization, and ReLU activation function;

[0074] Step S402: Feature fusion and dimensional alignment. The multi-scale strip features are concatenated with the original features in the channel dimension and fused to the target semantic dimension through 1×1 convolution, providing the decoder with features that have anisotropic perception capabilities.

[0075] ;

[0076] Where k represents the size of the strip convolution kernel.

[0077] The Geometric Awareness Frequency Alignment (GA-FAM) module achieves topological alignment between clean high-frequency features and backbone spatial semantic features by introducing an anisotropic strip convolution mechanism. Unlike traditional square convolution kernels, GA-FAM uses a large aspect ratio strip convolution to extend the convolution receptive field linearly along the horizontal and vertical directions. This allows for the establishment of long-distance spatial dependencies along the orientation of ground features and effectively limits the diffusion of the receptive field in orthogonal directions, thereby suppressing noise interference from irrelevant background areas on high-frequency edges. To achieve geometric feature alignment of multi-scale remotely sensed ground features, the GA-FAM module is configured with three sets of anisotropic strip convolution branches at different receptive field scales in parallel. The three branches respectively capture features of short-distance linear edges (such as building outlines), mid-distance geometric features (such as short streets), and cross-regional macroscopic anisotropic features (such as roads and rivers), enabling the feature map to possess multi-scale anisotropic spatial perception representation. Finally, the anisotropic features output by the three parallel branches are concatenated with the original features along the channel dimension and fused through 1×1 convolution to map to the target semantic dimension, achieving effective alignment between frequency domain features and spatial domain features.

[0078] Step S5 includes:

[0079] Step S501: Adaptive injection of high-frequency geometric features. Using an initialized learnable scaling parameter (preferably 0.1), the high-frequency geometric features output by the GA-FAM module are aligned. Dynamically fused into the backbone decoding feature map after upsampling and two-dimensional selective scanning (SS2D).

[0080] Step S502, Multi-scale Feature Concatenation and Classification: The context features from each level of the decoder are fused and fed into a classification head composed of a multilayer perceptron (MLP). The classification head is randomly initialized along with the entire network during the initial model construction phase.

[0081] Step S503, Model Training and Inference: During the training phase, the network model, consisting of the encoder, PAC-FSM frequency domain collaborative module, WB-LDS lossless downsampling module, GA-FAM geometric perception module, and classification head, is jointly trained end-to-end. Using the labeled remote sensing image training set, the cross-entropy loss between the predicted results and the true labels is calculated, and the backpropagation algorithm is used to synchronously update the overall backbone feature extraction parameters and classification head parameters until the model converges, completing the construction and optimization of the segmentation model. In the prediction application phase, the unlabeled target remote sensing image to be segmented is input into the trained model. The low-resolution category probability prediction matrix output by the classification head is restored to the original resolution H×W through bilinear interpolation upsampling. Finally, the semantic category label prediction result corresponding to each pixel in the entire image is output, achieving high-precision remote sensing image ground feature extraction and semantic segmentation.

[0082] To reduce the interference of potential high-frequency noise on feature representation during the initial training phase, a learnable scaling parameter initialized to 0.1 is introduced. This parameter is then multiplied element-wise with the output of GA-FAM, and the residual is added to the upsampled SS2D output. As training progresses, the scaling parameter automatically adjusts based on the backpropagation of the loss function. This mechanism enables the network to adaptively adjust the proportion of high-frequency geometric information injection according to the complexity of different remote sensing scenes: in scenes with complex target boundaries, the parameter adaptively increases to enhance fine-grained edge perception; in scenes with flat backgrounds, the parameter maintains a lower weight to resist noise perturbations. This explicit residual structure allows the model to accurately capture fine-grained boundary details while maintaining macroscopic semantic consistency.

[0083] In embodiments of this application, a remote sensing image semantic segmentation model based on a frequency domain and geometric perception state space model is provided. This model is used to implement some or all of the semantic segmentation methods described above. The model construction process specifically includes:

[0084] (a) Remote sensing data preparation:

[0085] The present invention was experimentally verified on the LoveDA high-resolution remote sensing dataset to demonstrate the effectiveness and feasibility of the method: the LoveDA dataset contains 7 categories: Background, Building, Road Water, Barren, Forest, and Agriculture; the images were uniformly cropped to a size of 512×512 pixels.

[0086] (II) Model parameter settings and reasons:

[0087] 1) Number of training iterations: 40,000;

[0088] 2) Batch size: 16 (to maximize the use of GPU memory and speed up training);

[0089] 3) Input remote sensing image size: 512×512 (adapted to the dataset cropping size, balancing resolution and computational cost);

[0090] 4) Learning rate: 0.00006 (to avoid destroying pre-trained features);

[0091] 5) Optimizer: AdamW (AdamW combines Adam's fast convergence capability with SGD's good generalization capability).

[0092] 6) Loss function: CrossEntropyLoss (consistent with SegMAN);

[0093] (III) Model Evaluation:

[0094] In this method, IoU is selected as the evaluation index for intra-task categories, and mIoU is selected as the inter-task accuracy index to evaluate the performance of the present invention in the semantic segmentation task of remote sensing images.

[0095] (iv) Experimental results and comparisons:

[0096] The method of this invention differs from existing technologies (such as UNetFormer, RS). 3 Mamba, SegMAN, etc., were compared on the LoveDA dataset, with the main evaluation metrics including IoU (Intersection over Union), mIoU (mean Intersection over Union), and mF1 (mean F1 score). The semantic segmentation performance metrics are compared in Table 1.

[0097] Table 1 Comparison of Semantic Segmentation Performance Indicators

[0098]

[0099] As shown in Table 1, on the LoveDA dataset with cross-urban / rural scenarios, the semantic segmentation method provided in this invention improves the mIoU by 1.7% compared to the baseline model SegMAN, and is significantly higher than other comparative models. This verifies the effectiveness of the improvement strategy of this invention, demonstrates the generalization and superiority of this invention, and has broad practical application scenarios.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any equivalent substitutions, minor adjustments to mathematical formula parameters, or equivalent changes to the model computation flow made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the algorithmic logic, feature fusion scheme, and inventive concept of the present invention, should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A semantic segmentation method for remote sensing images based on frequency domain collaboration and geometric perception alignment, characterized in that: Includes the following steps: Step S1: Obtain the remote sensing image data to be segmented, input it into the encoder composed of a hierarchical feature extraction network, and extract multi-scale spatial semantic features at different levels. Step S2: In the deep feature path of the encoder, a phase-amplitude coordination module is introduced to convert the spatial features to the frequency domain and decouple them into amplitude spectrum and phase spectrum. The structural information extracted from the phase spectrum is used to modulate the amplitude spectrum features, and the frequency domain enhanced features are output after reconstruction. Step S3: Input the multi-scale spatial semantic features into the decoder, introduce lossless wavelet downsampling in the decoding path, decompose the feature map into low-frequency components and high-frequency components, and peel off the pure high-frequency detail features while constructing multi-scale context features. Step S4: Input the pure high-frequency detail features into the geometrically aware frequency alignment module, and use multi-scale anisotropic convolution branches to extract long-distance coherent geometric features to achieve explicit alignment of spatial domain and frequency domain features. Step S5: Adaptively inject the aligned high-frequency geometric features into the decoded backbone features, combine multi-scale context fusion features to perform pixel-level classification, and output the semantic segmentation results of the remote sensing image.

2. The remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment according to claim 1, characterized in that: Step S1 includes: The remote sensing image to be segmented is acquired and input into an encoder composed of a hierarchical feature extraction network for shallow semantic feature extraction: In the first layer of the encoder, the input remote sensing image is downsampled once and the LASS module, which integrates neighborhood attention and 2D selective scanning, to obtain the feature map F1 output by the first layer of the encoder. In the second layer of the encoder, F1 obtains the feature map F2 of the second layer output of the encoder through a downsampling and LASS module.

3. The remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment according to claim 2, characterized in that: In step S2, the PAC-FSM module is introduced to perform frequency domain decoupling and enhancement on the spatial domain feature map X. The PAC-FSM module refers to the phase-amplitude coordination module, and the processing procedure of this module on the input spatial feature map X is as follows: (1) Fast Fourier Transform: The spatial feature map X of the input is transformed to the frequency domain using a 2D real-number fast Fourier transform. ； Where H and W represent the height and width of the input spatial feature map X, respectively; h and w represent the row and column coordinate indices in the spatial domain, respectively; and u and v represent the vertical and horizontal spatial frequency coordinates in the two-dimensional frequency domain mapping, respectively. This is due to the compression of the complex number frequency domain representation by the symmetry of real numbers; (2) Separating the amplitude spectrum and phase spectrum: Using complex number calculation methods, the amplitude spectrum and phase spectrum are separated. Strict decoupling into amplitude spectrum A and phase spectrum P: ； ； Where Re and Im represent the real and imaginary parts of the complex number, respectively; (3) Three-branch coordinated modulation: Modulation is performed through global branch, amplitude branch and phase-guided branch to obtain the enhanced amplitude spectrum. ; (4) Inverse Fourier Transform: The spatial domain features are reconstructed in strict phase-preserving manner using the inverse Fourier transform. 。 4. The remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment according to claim 3, characterized in that: The three-branch coordinated modulation includes: (a) Amplitude branch modulation: After the amplitude spectrum A is fed into the amplitude branch, a lightweight 3×3 depthwise separable convolution is used to extract features, capturing the local frequency domain correlation between neighboring frequency points, thus enhancing the model's ability to perceive the local frequency domain context: ； (b) Phase-guided branch modulation: A 1×1 convolution is used to reduce the dimensionality of the phase spectrum P to 1 / 4 of the original number of channels, extracting the most core features that represent structural information from the input phase; after passing through the ReLU activation function, a 1×1 convolution is used to restore the phase spectrum P to the original number of channels, and finally the Sigmoid activation function is used to generate the gating: ； in This represents the Sigmoid activation function; (c) Global Branch Modulation: After the amplitude spectrum A is input into the global branch, adaptive average pooling is first used to compress the high-frequency and low-frequency energies in the entire feature map into a global feature; then, the same two-layer convolution and sigmoid activation function as the phase-guided branch are used to control the enhancement intensity according to the tone of the entire image. ； (d) Calculate the enhanced amplitude spectrum by multiplying the outputs of the three branches together and then multiplying by a learnable parameter γ initialized to zero. Add this result to the original amplitude spectrum A to obtain the enhanced amplitude spectrum. : ； in This is element-wise multiplication.

5. The remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment according to claim 3, characterized in that: Step S2 includes: Residual adaptive injection: PAC-FSM is connected in parallel with LASS Block and global self-attention module in the third and fourth layers of the encoder, respectively; The third layer of the encoder is constructed by parallel connection of PAC-FSM and LASS Block. The input features of the third layer are the feature map F2 output from the second layer of the encoder, and the output is a fused feature map. : ； in, This represents the original spatial feature map extracted by the LASS module in the third layer of the encoder; This represents the frequency domain feature mapping obtained after the input features pass through the PAC-FSM module in the third layer of the encoder; The fourth layer of the encoder is constructed using a PAC-FSM in parallel with a global self-attention module; the input features of the fourth layer of the encoder are the output features of the third layer of the encoder. Output fused feature map : ； in, This represents the global spatial feature map extracted by the global self-attention module in the fourth layer of the encoder. This represents the frequency domain feature mapping obtained after the input features pass through the PAC-FSM module in the 4th layer of the encoder; These are learnable parameters used to control the fusion process, initialized to 0.

6. The remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment according to claim 5, characterized in that: In step S3, the decoder receives the output features F2, F3, and F4 from the second, third, and fourth layers of the encoder, where F3 and F4 are the features from step S2. and The feature F obtained by fusing these three layers of features is then input into the WB-LDS module for processing. The WB-LDS module refers to lossless wavelet downsampling, and the processing includes: Step S301: Transform from the spatial domain to the frequency domain. Use Haar discrete wavelet transform to convert the spatial domain features into four frequency domain components A, H, V, and D, representing the low-frequency component, horizontal high-frequency component, vertical high-frequency component, and diagonal high-frequency component, respectively. ； Step S302: Decoupling high-frequency features and main features: After concatenating the three high-frequency information, feature representation learning is performed on the high-frequency information through 1×1 convolution, batch normalization and ReLU activation function to obtain high-frequency features; Meanwhile, in another branch, the four frequency domain components are concatenated and then subjected to the same operation as for the high-frequency information to obtain the main features: ； 。 7. The remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment according to claim 6, characterized in that: In step S4, the clean high-frequency detail features decoupled from the WB-LDS module are... The data is input in parallel to the GA-FAM module, which refers to the geometric sensing frequency alignment module. The processing includes: Step S401: Multi-scale strip convolution geometry extraction: The strip convolution branch is composed of parallel strip convolutions in the vertical direction and strip convolutions in the horizontal direction. After addition, batch normalization and nonlinearization of the ReLU activation function are performed. ； in for High-frequency features obtained by mapping 1×1 convolution, batch normalization, and ReLU activation function; Step S402: Feature Fusion and Dimension Alignment: The multi-scale strip features are concatenated with the original features along the channel dimension, and then fused and mapped to the target semantic dimension through a 1×1 convolution, providing the decoder with features that have anisotropic perception capabilities. ； Where k represents the size of the strip convolution kernel.

8. The remote sensing image semantic segmentation method based on frequency domain collaboration and geometric perception alignment according to claim 7, characterized in that: Step S5 includes: Step S501, Adaptive Injection of High-Frequency Geometric Features: Using the initialized learnable scaling parameters, the high-frequency geometric features output by the GA-FAM module are adaptively injected. Dynamically fused into the backbone decoding feature map that has undergone upsampling and two-dimensional selective scanning; Step S502: Multi-scale feature splicing and classification, fusing the contextual features of each level of the decoder and feeding them into the classification head composed of a multi-layer perceptron; Step S503, Model Training and Inference: In the model building stage, pixel-level labeled data is used to perform end-to-end training and parameter optimization on the entire network, including the classification head; in the inference prediction stage, the low-resolution prediction results output by the classification head are restored to the original resolution H×W through bilinear interpolation upsampling, and finally the category label prediction matrix corresponding to each pixel of the remote sensing image to be predicted is output.

Images