A remote sensing image semantic parsing method based on topology-frequency domain collaborative field perception network

Abstract

The present application relates to a kind of based on topological-frequency domain collaborative field sensing network's remote sensing image semantic analysis method, belong to remote sensing image semantic analysis technical field.The method includes: remote sensing image is initialized pretreatment, and image field representation is obtained;Topological-frequency domain collaborative field sensing encoder is constructed, and image field representation is carried out multi-level field decomposition and reorganization, and deep field representation is obtained;Field dynamic aggregation bottleneck layer is constructed, and deep field representation is carried out global field perception and sparse screening, and global optimization field feature is obtained;Field cognitive decoder is constructed, and global optimization field feature is inverted into pixel-level semantic field;Field cognitive loss function is constructed, and pixel-level semantic field is combined by back propagation training topological-frequency domain collaborative field sensing network, and the semantic field graph is output using trained topological-frequency domain collaborative field sensing network.It aims at solving the technical problems that existing technology has single cognitive dimension, feature fusion mean value, structure integrity loss and weak generalization ability.

Description

Technical Field

[0001] This invention relates to a method for semantic parsing of remote sensing images based on a topology-frequency domain cooperative field sensing network, belonging to the field of remote sensing image semantic parsing technology. Background Technology

[0002] Semantic segmentation of remote sensing images aims to assign semantic labels to each pixel in an image, serving as a bridge connecting pixel-level data with scene-level cognition. Traditional deep learning methods model this as a pixel classification problem, extracting multi-scale features through an encoder-decoder structure and relying on large-scale labeled data for supervised learning. However, existing methods have the following fundamental limitations:

[0003] First, the cognitive dimension is singular. Traditional convolution only focuses on the intensity changes of local neighborhood pixels, simplifying ground features into isolated sets of pixels. It cannot explicitly model the multi-dimensional ontological attributes of ground features as physical entities, such as the morphological structure (e.g., the continuity of roads, the closed contours of buildings) and the periodicity of texture (e.g., the regular texture of farmland, the undulating characteristics of water bodies). This results in a limited ability to interpret complex ground features.

[0004] Secondly, feature fusion tends to be homogenized. Skip connections simply and indiscriminately splice together the shallow details of the encoder and the deep semantics of the decoder. In essence, it is a homogenized reuse of features, without considering the differentiated needs of different categories for feature levels (such as urban buildings relying on edge details and forests relying on global textures), which inhibits the discriminative ability of features.

[0005] Third, the loss function is biased. The cross-entropy loss is calculated independently for each pixel, which cannot model the spatial relationship between pixels. This results in fragmented and discontinuous model prediction results, which contradicts the spatial coherence of real ground features.

[0006] Fourth, the training paradigm relies on strong supervision. Existing methods require massive amounts of dense annotations and have poor generalization ability across different regions and sensor data. Essentially, they rely on "memory" rather than "understanding".

[0007] Therefore, developing a remote sensing image processing method that can deconstruct image content from a multi-dimensional cognitive perspective and achieve deep semantic analysis has become a key requirement for overcoming existing technological bottlenecks. Summary of the Invention

[0008] The purpose of this invention is to provide a remote sensing image semantic parsing method based on topology-frequency domain collaborative field sensing network, which aims to solve the technical problems of existing technologies such as single cognitive dimension, averaged feature fusion, loss of structural integrity and weak generalization ability.

[0009] To achieve the above objectives, the technical solution of the present invention is: a method for semantic parsing of remote sensing images based on a topology-frequency domain cooperative field sensing network, comprising the following steps:

[0010] Step 1: Perform field initialization preprocessing on the remote sensing image to obtain the image field representation;

[0011] Step 2: Construct a topology-frequency domain collaborative field-aware encoder to perform multi-level field decomposition and recombination on the image field representation to obtain a deep field representation; wherein, the topology-frequency domain collaborative field-aware encoder includes a field feature decoupling module and a field feature recombination module. The field feature decoupling module decouples the image field representation into topological field components and frequency domain field components in parallel through topological projection basis and frequency domain projection basis. The field feature recombination module cross-fuses the topological field components and frequency domain field components through an inter-field attention mechanism to generate a single-level enhanced field representation. The single-level enhanced field representations are stacked level by level to obtain a deep field representation.

[0012] Step 3: Construct a field dynamic aggregation bottleneck layer to perform global field perception and sparse filtering on the deep field representation to obtain globally optimized field features; wherein, the field dynamic aggregation bottleneck layer includes a field global perception module and a field sparse filtering module. The field global perception module uses nonlocal attention to capture long-range dependencies between fields and generate field context vectors. The field sparse filtering module generates sparse gating masks based on the field context vectors and dynamically filters key field components.

[0013] Step 4: Construct a field cognitive decoder to invert the globally optimized field features into a pixel-level semantic field; wherein, the field cognitive decoder includes a field inversion module and a field boundary refinement module. The field inversion module dynamically fuses the decoder features with the corresponding encoder features through a field adaptive fusion mechanism, and the field boundary refinement module uses field gradient information to guide the localization of ground feature edges.

[0014] Step 5: The topology-frequency domain collaborative field-aware encoder, the field dynamic aggregation bottleneck layer, and the field cognition decoder are sequentially connected to form a topology-frequency domain collaborative field-aware network. A field cognition loss function is constructed, and the loss value is calculated using the pixel-level semantic field and the real label semantic field. The parameters of the topology-frequency domain collaborative field-aware network are iteratively optimized through backpropagation to obtain the trained topology-frequency domain collaborative field-aware network. The field cognition loss function includes field classification loss, field topology consistency loss, and field frequency domain consistency loss.

[0015] Step 6: Input the remote sensing image to be parsed into the trained topology-frequency domain collaborative field perception network, and output the semantic field map.

[0016] Optionally, Step 1 specifically includes:

[0017] Step 1.1: Perform input verification of the original remote sensing image. If the input remote sensing image is a single-channel grayscale image, convert it into an RGB three-channel image by channel copying; if the input remote sensing image is a non-square rectangular image, fill it with zeros to make it square.

[0018] Step 1.2: Scale the verified remote sensing images to a fixed size;

[0019] Step 1.3: After normalizing the size of the RGB three-channel image, the pixel values ​​are linearly mapped from [0,255] to the [0,1] interval, and then the image after linear mapping is normalized.

[0020] Step 1.4: The standardized image is considered as a composite field composed of multiple subfields of ground features. Each pixel (x, y) in the composite field is assigned a field attribute. The image field representation is constructed by convolving the standardized image with a pre-defined prior distribution of ground features. The expression is as follows:

[0021]

[0022] in, Let be the response value of the c-th type of land feature subfield at (x, y); For the standardized image, This is the field-aware convolution kernel corresponding to the c-th type of land cover; These are the learnable category weights; This represents the convolution operation. This represents the total number of land cover categories in the remote sensing scene; the convolution results of all land cover subfields are concatenated along the channel dimension to obtain a composite field feature map, completing the mapping from pixel space to field space.

[0023] Optionally, the topology-frequency domain collaborative field-aware encoder is composed of three levels of field-aware units stacked sequentially. Each level of field-aware unit includes a field feature decoupling module and a field feature reassembly module. The field feature decoupling module is a parallel dual-path field-aware convolution, where the dual paths are a topological path and a frequency domain path. Specifically:

[0024] In the topological path, the input image is first processed by a set of trainable orientation-sensitive convolutional kernels. Four sets of parallel convolutions are performed, and the results of the four branches are concatenated along the channel dimension to obtain the topological response map. Then, batch normalization (BN) and ReLU activation functions are sequentially applied to the topological response map. The number of channels is compressed using a 1×1 channel compressed convolution kernel. Finally, the topological attention weights are generated using the Sigmoid activation function. The topological field components are obtained by element-wise multiplying the input image with the topological attention weights. ;

[0025] In the frequency domain path, the input image is first processed. Each channel independently performs a two-dimensional fast Fourier transform, transforming the field from the spatial domain to the frequency domain to obtain the complex spectrum. Then, the amplitude spectrum is obtained by taking the modulus of the complex spectrum. A set of trainable band-sensitive convolutional kernels is used to perform parallel convolution on the amplitude spectrum. The four sets of convolutional kernels correspond to the low-frequency, mid-low-frequency, mid-high-frequency, and high-frequency bands, respectively, to obtain the frequency domain response map. Then, batch normalization (BN) and ReLU activation functions are sequentially applied to the frequency domain response map. The number of channels is compressed by a 1×1 channel compression convolutional kernel. Finally, the frequency domain attention weights are generated by the Sigmoid activation function and multiplied element-wise with the input image to obtain the frequency domain field components. ;

[0026] The field feature reconstruction module first reconstructs the topological field components. and frequency domain field components The concatenated features are obtained by concatenating along the channel dimension; the concatenated features are then convolved using a 1×1 fusion coefficient convolution kernel, and the adaptive fusion coefficients are calculated using the Sigmoid function. ;

[0027] For the input image Perform a 3×3 basic feature extraction convolution kernel to extract basic field features. And through two independent 1×1 channel projective convolution kernels, respectively and The topological field components are obtained by mapping from the input channel to the output channel. and frequency domain field components The final output is the enhanced field representation of the current-level unit. The expression is:

[0028]

[0029] Enhanced field characterization of each unit output Then, the feature map size is halved through convolution and input into the next level field perception unit; the enhanced field representations output by each level field perception unit are simultaneously saved as encoder skip connection features of the corresponding level.

[0030] Optionally, the field dynamic aggregation bottleneck layer specifically refers to:

[0031] In the field-wide perception module, the input features are first processed by three independent 1×1 channel transformation kernels. Convolution is performed separately to generate a query matrix Q, a key matrix K, and a value matrix V. The transpose of Q is multiplied by the matrix product of K, and then the result is passed through a softmax function to obtain an attention map A. The attention map A is then multiplied by the transpose of V, and the feature map size is restored to obtain the global context vector, which is then compared with the input feature map. Perform residual summation to obtain the global perception features. ;

[0032] In the sparse screening module, the first step is to... Perform global average pooling, averaging all pixel values ​​for each channel to obtain the global descriptor, expressed as:

[0033]

[0034] in, For the c-th channel component of the global descriptor z, Globally perceptible features The value of the c-th channel at (x, y); H and W are the feature map dimensions;

[0035] The global descriptor is input into two consecutive fully connected layers. The first fully connected layer compresses the vector and processes it using the ReLU activation function. The second fully connected layer maps the vector back to its original dimension and generates a channel-gated vector G using the Sigmoid function, with each element taking values ​​in the range [0,1]. The gated vector G is then expanded using a broadcast operation to a dimension similar to the original vector. Given gate masks M of the same size, retain the first few values ​​of the gate mask with the largest values. One channel is used, and the mask values ​​of all other channels are set to zero; the final field dynamic aggregation bottleneck layer outputs the globally optimized field features as follows: .

[0036] Optionally, the field cognition decoder consists of three levels of decoding units that correspond one-to-one with the topology-frequency domain collaborative field-aware encoder levels. Each decoding unit includes a field inversion module and a field boundary refinement module, specifically:

[0037] In the in-situ inversion module, a 4×4 transposed convolution kernel is first used to process the input features of the current decoder. Perform upsampling and output the upsampled features. Upsampled features Jump connection features with encoder The concatenated features are obtained by concatenating along the channel dimension, then convolved using a 1×1 fusion weight convolution kernel, and finally the adaptive fusion weights are calculated using the Sigmoid function. ; Generate fusion features through field adaptive fusion The expression is:

[0038]

[0039] Fusion features The input 3×3 feature decoding convolution kernel performs two consecutive convolutions. After each convolution, batch normalization and ReLU activation functions are applied to complete the feature inversion of the current level decoding unit.

[0040] In the field boundary refinement module, the fusion features output by the field inversion module are... The spatial gradients in the x and y directions are calculated independently for each channel to obtain the boundary gradient map. The expression is:

[0041]

[0042] in, and These are the Sobel gradient operators in the x and y directions, respectively;

[0043] By performing convolution on the boundary gradient map using two consecutive 3×3 boundary feature extraction convolution kernels, detailed boundary features are extracted. Each convolution is followed by BN and ReLU activation functions to combine boundary features with fused features. Perform residual summation to complete boundary reinforcement; the expression is:

[0044]

[0045] in, These are the learnable boundary enhancement coefficients; The final output of the current decoding unit is input to the next decoding unit.

[0046] The features output from the last unit of the decoder are convolved by a 1×1 classification convolution kernel, and then activated by the Softmax activation function. The final output is a multi-channel land cover category prediction probability map. The sum of the values ​​of each pixel in C channels is 1, which represents the probability that the current pixel belongs to each type of land cover.

[0047] Optionally, the field cognition loss function is specifically:

[0048]

[0049] in, Let be the field cognition loss function. , As a balance factor, For field classification loss, For field topology consistency loss, This represents the loss of consistency in the field frequency domain;

[0050] The expression for the field classification loss is as follows:

[0051]

[0052] Where H and W are the feature map dimensions, This represents the total number of land cover categories in the remote sensing scene. Balance the weights for each category; For focusing parameters; Let (x, y) be the predicted probability that pixel (x, y) belongs to the c-th type of land cover. This is the true label of pixel (x, y), which is 1 if it belongs to class c, and 0 otherwise;

[0053] Among them, the field topology consistency loss first converts the real label Extended by one-hot encoding to Then compare the predicted probability map P with the one-hot encoded true label. The spatial gradient magnitude is calculated independently for each channel, and the expression is:

[0054]

[0055] in, To predict the gradient magnitude of the c-th channel of the probability map at pixel (x,y), Let c be the gradient magnitude of the c-th channel of the true label one-hot encoding at pixel (x,y).

[0056] Among them, the field frequency domain consistency loss affects the predicted probability map P and the one-hot encoded true label. Perform a two-dimensional FFT on each channel independently to obtain its own amplitude spectrum. and The expression for calculating the L2 norm squared difference for each channel and averaging it over all channels is:

[0057]

[0058] in, This represents a two-dimensional Fast Fourier Transform. The amplitude spectrum is obtained by taking the modulus of the complex spectrum. It is an L2 norm.

[0059] Optionally, during the training process, the topology-frequency domain cooperative field sensing network calculates a comprehensive index after each round of full training, expressed as:

[0060]

[0061] in, As a comprehensive indicator, For the average intersection-to-union ratio, the normalized topology loss and the normalized frequency domain loss are respectively derived from the field topology consistency loss. and field frequency domain consistency loss After being mapped to the [0,1] interval using minimum-maximum normalization, the result is as follows:

[0062] Statistical analysis of each round of validation sets , The historical minimum and maximum values ​​are mapped using the following expression:

[0063]

[0064] In the formula, , These represent the minimum and maximum values ​​of the topological consistency loss across all historical validation sets. , These are the minimum and maximum values ​​of the frequency domain consistency loss in each historical validation set round. The model weight with the highest comprehensive index in each validation set round is saved. An early stopping strategy is set so that training is terminated early when the comprehensive index of the validation set does not improve for a preset number of consecutive rounds.

[0065] The beneficial effects of this invention are:

[0066] (1) The cognitive dimension is upgraded from pixel intensity to dual field attributes of topology and frequency domain. This invention uses the field feature decoupling module in the topology-frequency domain collaborative field perception encoder to extract the spatial connectivity topology features and frequency domain texture periodic features of ground objects in parallel with direction-sensitive convolution and frequency band-sensitive convolution. It explicitly models the morphological structure and texture periodicity of ground objects as physical entities, overcomes the shortcomings of traditional convolution that only focuses on local pixel intensity and has a single cognitive dimension, and significantly improves the interpretation ability of structural ground objects such as roads and buildings and textured ground objects such as farmland and water bodies.

[0067] (2) Feature fusion is upgraded from indiscriminate splicing to on-demand adaptive fusion. The field feature recombination module and the decoder field inversion module of the present invention respectively perform weighted fusion of topological field, frequency field and encoder-decoder multi-scale features through adaptive fusion coefficients and adaptive fusion weights, so that different land features can call edge details or global textures as needed, overcome the feature meanification problem caused by simple splicing with skip connections, and enhance the feature discrimination ability.

[0068] (3) The structural integrity is improved from fragmented and discontinuous to continuous and consistent boundaries. This invention introduces field topology consistency loss, which uses the difference in distribution between the predicted probability map and the real label in the spatial gradient field as a constraint. Combined with the Sobel gradient guidance and residual enhancement of the field boundary refinement module, it forces the network to maintain the spatial structural continuity of ground objects, overcomes the problem of fragmented prediction results caused by independent calculation of cross-entropy for each pixel, and significantly reduces fragmentation misclassification.

[0069] (4) Cross-domain generalization ability is transformed from memorizing pixels to recognizing fields. This invention introduces field frequency domain consistency loss, which globally constrains the consistency of texture distribution between the predicted field and the real field in the frequency domain. This guides the network to learn the essential texture rules of ground objects rather than memorizing specific pixel distributions, effectively alleviating the problems of different objects with the same spectrum and different spectrums with the same object, and improving the model's generalization ability across different regions and different sensor data.

[0070] (5) The bottleneck layer combines global perception and sparse filtering to balance accuracy and efficiency. The field dynamic aggregation bottleneck layer of this invention captures long-range dependencies across regions through non-local attention, and uses SE-block channel gating in conjunction with Top-K sparsification (preserving the previous one). (Suppress background redundancy by using key channels) while ensuring global consistency across large scenes, reducing redundant computations and improving the expression efficiency of key field components. Attached Figure Description

[0071] Figure 1 This is an overall flowchart of the present invention;

[0072] Figure 2 This is a schematic diagram of the topology-frequency domain cooperative field sensing encoder of the present invention;

[0073] Figure 3 This is a schematic diagram of the field dynamic polymerization bottleneck layer structure of the present invention;

[0074] Figure 4 This is a schematic diagram of the field cognition decoder structure of the present invention;

[0075] Figure 5 This is a schematic diagram of remote sensing semantic segmentation according to the present invention. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0077] Example 1: As Figure 1 As shown, a remote sensing image semantic parsing method based on a topology-frequency domain collaborative field-aware network is presented. This method re-examines the problem of remote sensing image semantic parsing from the perspectives of cognitive science and physical fields. The core of this method lies in the three-layer design of topology-frequency domain collaborative field-aware encoding, field dynamic aggregation bottleneck screening, and field cognitive decoding reconstruction, which achieves deep parsing of remote sensing scenes using a field cognitive paradigm. The method includes the following steps:

[0078] Step 1: Perform field initialization preprocessing on the remote sensing image to obtain the image field representation;

[0079] It is important to understand that after acquiring the original remote sensing image, this embodiment sequentially performs basic preprocessing and field representation construction, mapping the image from pixel space to field space. Here, the "field" mentioned in this embodiment is a concept defined in this invention: it borrows the idea from physics that a "field" describes a continuously distributed physical quantity, and in this invention refers to a high-dimensional tensor representation in feature space that can express the continuous distribution patterns and physical properties of ground objects. Specifically, the image field corresponds to the feature tensor of the input image. The subfield of land features corresponds to the channel components in the feature map, used to characterize the spatial response intensity distribution of this type of land feature; the semantic field corresponds to the category probability distribution map of the final output of the network. .in, H represents the real number field; H and W represent the height and width of the feature map (or image) (in pixels), respectively. The 3 in the text represents the three color channels: R, G, and B; C represents the total number of land cover categories in the remote sensing scene.

[0080] Step 1.1: Perform input verification of the original remote sensing image. The original remote sensing image adapted in this embodiment is an RGB three-channel digital image (RGB refers to the three primary color channels of red, green and blue). A single image is a square with a side length of 256 pixels to 1024 pixels. The pixel values ​​of the three RGB channels are independently selected, and the value range is [0,255]. The image bit depth is 8 bits. If the input remote sensing image is a single-channel grayscale image, it is converted into an RGB three-channel image by channel copying. If the input remote sensing image is a non-square rectangular image, it is filled with zeros to make it a square before proceeding to subsequent processing.

[0081] Step 1.2: Image Size Normalization. The verified original remote sensing images are uniformly scaled to a fixed size of 512×512 pixels. The scaling process uses a bicubic interpolation algorithm with an interpolation kernel size of 4×4. The interpolation formula is as follows:

[0082]

[0083] in, The pixel value at coordinates (x, y) in the target image is obtained through interpolation; x and y are the horizontal and vertical coordinates of the target pixel; i and j are the row and column indices (0 to 3) of the neighboring sampling points. These are the values ​​of the 16 pixels adjacent to the target pixel in the source image. The bicubic interpolation kernel function is expressed as:

[0084]

[0085]

[0086]

[0087] In the formula, the independent variable t represents the distance (in pixels) between the position to be interpolated and a certain neighboring sampling point.

[0088] Furthermore, the size adaptation rules are as follows: if the side length of the original image is less than 512 pixels, it is first enlarged proportionally to a short side of 512 pixels using the bicubic interpolation algorithm described above, and then the long side exceeding 512 pixels is cropped at the center; if the side length of the original image is greater than 512 pixels, the original image is first cropped at the center to a square with a side length of 512 pixels, and then the details are smoothed using bicubic interpolation to ensure that all input images are completely consistent in size.

[0089] Step 1.3: Color Space Normalization. First, the pixel values ​​of the normalized RGB three-channel image are linearly mapped from [0,255] to the [0,1] interval. The linear mapping formula is:

[0090]

[0091] in, This represents the original pixel value of the c channel at the (x,y) position in the image after size normalization. These are the pixel values ​​after linear mapping. Corresponding to Three channels. Then, standardization is performed on the linearly mapped image, setting the mean of the three channels to [0.485, 0.454, 0.406] and the standard deviation to [0.225, 0.224, 0.229]. The standardization formula is:

[0092]

[0093] in, This is the preset mean value for channel c. This is the preset standard deviation of channel C. The final standardized image pixel values ​​serve as input for the subsequent construction of the field representation.

[0094] Step 1.4: Field Representation Construction. The standardized image is considered as a composite field composed of multiple superimposed subfields of ground features, denoted as... ,in, This represents the subfield corresponding to the c-th type of land cover; in the composite field, each pixel (x, y) not only carries the original RGB intensity value, but is also assigned field attributes, i.e., the probability distribution of the pixel belonging to each subfield of land cover. The field representation is constructed through a convolution operation between a standardized image and a predefined prior distribution of ground features, expressed as:

[0095]

[0096] in, Let be the response value of the c-th type of land feature subfield at (x, y); For the standardized image (i.e. (The overall notation for the three channels) For the field-aware convolutional kernel (3×3) corresponding to the c-th type of land cover, the convolutional kernel has 3 input channels, 1 output channel, stride=1, padding=same, and the convolutional kernel weights are initialized using Kaiming uniform distribution (Kaiming initialization, also known as He initialization, is a well-known weight initialization method for ReLU-type activation functions). The learnable class weights are initially set to 1.0 and optimized synchronously during network training. This represents the convolution operation. The convolution results of all ground feature subfields are concatenated along the channel dimension to obtain a composite field feature map of size 512×512×C, completing the mapping from pixel space to field space, which serves as the input to the subsequent encoder. Here, "same padding" is a well-known padding method in convolutional neural networks, which involves padding the feature map edges with zeros to ensure that the spatial dimensions of the output feature map are consistent with the input. C should satisfy 1≤C≤32 and be strictly consistent with the number of output channels of the classification convolution at the end of the decoder; the number of input channels of the first-level field-aware unit of the encoder... That is, C, which is determined by the specific application scenario.

[0097] Step 2: Construct a topology-frequency domain collaborative field-aware encoder to perform multi-level field decomposition and recombination on the image field representation to obtain a deep field representation. The topology-frequency domain collaborative field-aware encoder includes a field feature decoupling module and a field feature recombination module. The field feature decoupling module decouples the image field representation into topological field components and frequency field components in parallel using a topological projection basis and a frequency projection basis. The field feature recombination module cross-fuses the topological field components and frequency field components through an inter-field attention mechanism to generate a single-level enhanced field representation. Multi-level field-aware units are stacked level by level to obtain a deep field representation.

[0098] Optionally, such as Figure 2As shown, the topology-frequency domain collaborative field-aware encoder consists of three stacked field-aware units. The channel configuration of the three units is as follows: Level 1: Cin = C (total number of categories), Cout = 32; Level 2: Cin = 64, Cout = 64; Level 3: Cin = 128, Cout = 128. Channel doubling occurs during inter-layer downsampling convolution (i.e., downsampling from Level 1 to Level 2: 32 → 64; downsampling from Level 2 to Level 3: 64 → 128; downsampling after Level 3 output: 128 → 128, compressing only the spatial size). The feature map size output by each unit is halved compared to the input, resulting in an enhanced field representation with a final output size of 64 × 64 × 128. Each field-aware unit includes a field feature decoupling module and a field feature reassembly module, the specific implementation details of which are as follows:

[0099] Optionally, the topological projection basis is composed of trainable direction-sensitive convolutional kernels that respond to the edge continuity and region closure of the image content; the frequency domain projection basis is composed of trainable frequency band-sensitive convolutional kernels that respond to the texture periodicity and frequency distribution of the image content in the Fourier frequency domain.

[0100] Step 2.1: Field Feature Decoupling Module: The core is parallel dual-path field-aware convolution, used to decouple the input image field... Synchronous decoupling into topological field components and frequency domain field components Input image The dimensions are H×W×Cin, where Cin is the number of input channels of the current level unit, and H and W are the height and width of the current feature map.

[0101] Step 2.1.1: Topological Path. The core of the topological path is to extract the spatially continuous topological features of ground features. The specific process is as follows: First, the input image is processed by a set of trainable orientation-sensitive convolutional kernels (3×3, 4 sets of orthogonal directional branches). Parallel convolution is performed, with four sets of convolutional kernels corresponding to the four orthogonal directions of 0°, 45°, 90°, and 135°. Each set of kernels has an input channel of Cin and an output channel of Cin (i.e., the number of input and output channels is the same for each direction of convolution). The stride is 1, and the padding is the same. The convolution results of the four branches are concatenated along the channel dimension to obtain a topological response map of size H×W×(4×Cin). Next, batch normalization (BN) and ReLU activation functions are applied sequentially to the topological response map. A 1×1 channel compression convolution kernel is then used to compress the number of channels from 4×Cin back to Cin. This kernel has 4×Cin input channels and Cin output channels, with a stride of 1 and padding of the same value. Finally, the topological attention weights are generated using the Sigmoid activation function. The dimensions are H×W×Cin, and the calculation formula is:

[0102]

[0103] in, It is the Sigmoid activation function. For the corresponding Orientation-sensitive convolution, , The input is a 1×1 channel compressed convolution (input 4×Cin, output Cin), and Concat represents the concatenation of channel dimensions; the topological field component is obtained by element-wise multiplication of the input features and the topological attention weights, as shown in the formula:

[0104]

[0105] in, This represents the element-wise multiplication operation of matrices. Size and input image Totally consistent.

[0106] Step 2.1.2: Frequency Domain Path. The core of the frequency domain path is to extract the frequency domain texture features of ground objects. The specific process is as follows: First, process the input image... Each channel independently performs a two-dimensional fast Fourier transform (2D-FFT) to transform the field from the spatial domain to the frequency domain, obtaining the complex spectrum. Size and input image Consistent; then, the amplitude spectrum is obtained by taking the modulus of the complex spectrum. The amplitude spectrum contains full-band information of ground texture. A set of trainable band-sensitive convolutional kernels (3×3, 4 band branches) are used to perform parallel convolution on the amplitude spectrum. The four convolutional kernels correspond to the low-frequency, mid-low-frequency, mid-high-frequency, and high-frequency bands, respectively. Each convolutional kernel has an input channel of Cin and an output channel of Cin, with a stride of 1 and padding of the same value. The convolutional results of the four branches are concatenated along the channel dimension to obtain a frequency domain response map of size H×W×(4×Cin). Next, batch normalization (BN) and ReLU activation functions are applied sequentially to the frequency domain response map. A 1×1 channel compression convolution kernel is then used to compress the number of channels from 4×Cin back to Cin. This kernel has 4×Cin input channels and Cin output channels, with a stride of 1 and padding of the same value. Finally, the frequency domain attention weights are generated using the Sigmoid activation function. The frequency domain field component is obtained by multiplying it element-wise with the input image. :

[0107]

[0108] in, Dimensions and Input Drawing Totally consistent.

[0109] Step 2.2: The field feature recombination module implements adaptive cross-fusion of topological field components and frequency domain field components to generate enhanced field representations. The specific process is as follows: First, the topological field components... and frequency domain field components The concatenated features are obtained by concatenating along the channel dimension, resulting in a concatenated feature of size H×W×(2×Cin). The concatenated feature is then convolved using a 1×1 fusion coefficient convolution kernel, and the adaptive fusion coefficient is calculated using the Sigmoid function. (The fusion coefficient for encoder feature reconstruction), with dimensions H×W×Cout, is calculated using the following formula:

[0110]

[0111] in, The convolutional kernel uses 1×1 fusion coefficients, with 2×Cin input channels and Cout output channel, stride=1, and padding=same; simultaneously, the original input is... Perform a 3×3 basic feature extraction convolution kernel to extract basic field features. The convolutional kernel has Cin as its input channel and Cout as its output channel, with a stride of 1 and padding of the same value. Cout represents the number of output channels for the current-level field-sensing unit (32, 64, and 128 for the three levels, respectively). Before the fusion operation, two independent 1×1 channel projection convolutional kernels are used to separate the input channels. and Mapping from the Cin channel to the Cout channel:

[0112]

[0113] Input channel Cin, output channel Cout, stride=1, padding=same;

[0114]

[0115] Input channel Cin, output channel Cout, stride=1, padding=same;

[0116] In the formula, , These are the features obtained by projecting the topological field component and the frequency field component through a 1×1 channel (unified to the Cout channel for easy element-by-element fusion later). , These are the corresponding 1×1 channel projective convolution kernels. The final output is the enhanced field representation of the current level unit. (Dimensions H × W × Cout), the formula is:

[0117]

[0118] Enhanced field characterization of each unit output Then, a 2×2 downsampling convolution kernel is used to perform a convolution with a stride of 2, halving the feature map size before inputting it to the next-level field-aware unit. This downsampling convolution kernel has Cout as its input channel and the number of input channels for the next-level unit as its output channel, with a stride of 2 and padding of the same value. The enhanced field representations output by each level of the field-aware unit (before downsampling) are simultaneously saved as the corresponding level's encoder skip connection features for use by the corresponding level's field inversion module in the field cognition decoder. The skip connection feature sizes for the three levels of coding units are: Level 1 512×512×32, Level 2 256×256×64, and Level 3 128×128×128.

[0119] Step 3: Construct a field dynamic aggregation bottleneck layer to perform global field perception and sparse filtering on the deep field representation to obtain globally optimized field features; wherein, the field dynamic aggregation bottleneck layer includes a field global perception module and a field sparse filtering module. The field global perception module uses nonlocal attention to capture long-range dependencies between fields and generate field context vectors. The field sparse filtering module generates sparse gating masks based on the field context vectors and dynamically filters key field components.

[0120] Optionally, such as Figure 3 As shown, the bottleneck layer receives the 64×64×128 enhanced field representation output by the encoder, and outputs globally optimized field features through the field global perception module and the field sparsity filtering module. The specific implementation details are as follows:

[0121] Step 3.1: The global field perception module uses a non-local attention mechanism to capture the long-range dependencies between subfields of ground objects at different spatial locations. The specific process is as follows: First, the input features are processed by transforming the convolution kernels with three independent 1×1 channels. Perform convolutions separately to generate a query matrix Q, a key matrix K, and a value matrix V. All three convolution kernels have 128 input and output channels, a stride of 1, and padding of the same value. The dimensions of Q, K, and V are all 64×64×128. Flatten Q, K, and V into 128×4096 two-dimensional matrices, where 4096 represents the total number of pixels in the feature map (64×64). Calculate the matrix product of the transpose of Q and K, and then apply the Softmax function to obtain an attention map A of size 4096×4096. The formula is:

[0122]

[0123] in, This is the number of channels, used for scaling to avoid excessively large inner product values. The transpose of the attention map A is used as the matrix representation; the transposes of the attention maps A and V are multiplied by a matrix, and then the feature map size is restored to 64×64×128 to obtain the global field context vector, which is then compared with the original input features. Perform residual summation and output globally perceived features. This feature captures the cross-regional subfield associations of ground features in a large scene.

[0124] Step 3.2: Field Sparsity Filtering Module: Based on Globally Aware Features The key field component screening and background redundancy suppression are completed through the following process: First, the key field components are screened and background redundancy information is suppressed. Perform global average pooling, averaging all pixel values ​​for each channel to obtain a 1×1×128 dimension global descriptor. The formula is as follows:

[0125]

[0126] in, is the c-th channel component of the global descriptor z, that is, the pixel mean of the c-th channel across all spatial locations; Globally perceptible features The value of the c-th channel at (x, y); c is the channel index, H=64, W=64 are the feature map dimensions; the global descriptor is input into two consecutive fully connected layers (using a bottleneck structure of SE-block). The first fully connected layer compresses the 128-dimensional vector into a 16-dimensional vector (reduction rate r=8), and processes it with the ReLU activation function; the second fully connected layer maps the 16-dimensional vector back to a 128-dimensional vector, and generates a channel-gated vector G using the Sigmoid function, with each element ranging from [0, 1]; the gated vector G is expanded into a channel-gated vector using a broadcast operation. A 64×64×128 gated mask M of the same size; retain the first element with the largest value in the gated mask. One channel ( The mask value is set to 128 (adaptively adjusted based on the number of land cover categories), and all other channels have their mask values ​​set to zero. The final bottleneck layer outputs the globally optimized field features as follows:

[0127]

[0128] The output feature size is the same as the input, 64×64×128, and is used as the input for the subsequent decoder.

[0129] Step 4: Construct a field cognitive decoder to invert the globally optimized field features into a pixel-level semantic field; wherein, the field cognitive decoder includes a field inversion module and a field boundary refinement module. The field inversion module dynamically fuses the decoder features with the corresponding encoder features through a field adaptive fusion mechanism, and the field boundary refinement module uses field gradient information to guide the localization of ground feature edges.

[0130] Optionally, such as Figure 4 As shown, the field cognition decoder consists of three levels of decoding units that correspond one-to-one with the topology-frequency domain collaborative field perception encoder levels. The channel configuration of the three levels of units is as follows: Level 1: Cin=128 (bottleneck layer output), Cout=128, corresponding to encoder skip connection skip3 (size 128×128×128); Level 2: Cin=128, Cout=64, corresponding to skip connection skip2 (size 256×256×64); Level 3: Cin=64, Cout=32, corresponding to skip connection skip1 (size 512×512×32).

[0131] Each unit outputs a feature map whose size is double that of the input, resulting in a final output multi-channel prediction probability map of size 512×512×C, where C≤32. Each decoding unit includes a field inversion module and a field boundary refinement module, with specific implementation details as follows:

[0132] Step 4.1: The field inversion module is used to achieve adaptive fusion of feature upsampling and multi-scale features between the encoder and decoder. The specific process is as follows: Let the current decoder input features be... The dimensions are H×W×Cin; the corresponding skip connection feature of the encoder at the same scale is... The size is 2H×2W×Cout; firstly, a 4×4 transposed convolution kernel is used. Upsampling is performed on the transposed convolutional kernel with input channel Cin and output channel Cout, stride=2, padding=1, and the output size is 2H×2W×Cout. , with encoder skip connection features The dimensions are completely consistent; upsampled features Jump connection features with encoder The concatenation is performed along the channel dimension to obtain a 2H×2W×(2×Cout) concatenated feature. This feature is then convolved using a 1×1 fusion weight convolution kernel, and finally the adaptive fusion weights are calculated using the Sigmoid function. (The fusion weights for decoder field inversion), with dimensions of 2H×2W×Cout, are calculated using the following formula:

[0133]

[0134] in, The convolutional kernel is 1×1 with fused weights, the input channels are 2×Cout, the output channel is Cout, the stride is 1, and the padding is the same. Fused features are generated through field-adaptive fusion. The formula is:

[0135]

[0136] Fusion features The input 3×3 feature decoding convolution kernel performs two consecutive convolutions. After each convolution, batch normalization and ReLU activation functions are applied. The input and output channels of the convolution kernel are both Cout, the stride is Stride=1, and the padding is Padding=same, thus completing the feature inversion of the current level decoding unit.

[0137] Step 4.2: The field boundary refinement module is used to enhance the features of ground features and improve the accuracy of boundary positioning. The specific process is as follows: the fused features output by the field inversion module are refined. (Dimensions 2H×2W×Cout) Calculate the spatial gradients in the x and y directions independently for each channel to obtain the boundary gradient map. The formula is:

[0138]

[0139] in, and These are the Sobel gradient operators in the x and y directions, respectively. In the gradient graph, the larger the value, the higher the boundary probability of the corresponding ground feature. Size and Consistent; convolution is performed on the gradient map using two consecutive 3×3 boundary feature extraction convolution kernels to extract boundary detail features. Each convolution is followed by Batch Normalization (BN) and ReLU activation functions. The input and output channels of the convolution kernel are both Cout, the stride is 1, and the padding is the same. The boundary features are then added to the original fused features using residuals to complete the boundary enhancement. The formula is as follows:

[0140]

[0141] in, The learnable boundary enhancement coefficient is initially set to 0.1 and optimized synchronously with network training. The final output of the current decoding unit is input to the next decoding unit.

[0142] Furthermore, the features output from the last unit of the decoder are convolved by a 1×1 classification convolution kernel with 32 input channels and C output channels representing the number of land cover categories. The stride is 1 and the padding is the same. After passing through the Softmax activation function, the final output is a multi-channel land cover category prediction probability map with a size of 512×512×C. The sum of the values ​​of each pixel in the C channels is 1, representing the probability that the pixel belongs to a land cover category.

[0143] Step 5: The topology-frequency domain collaborative field-aware encoder, the field dynamic aggregation bottleneck layer, and the field cognition decoder are sequentially connected to form a topology-frequency domain collaborative field-aware network. A field cognition loss function is constructed, and the loss value is calculated using the pixel-level semantic field and the real label semantic field. The parameters of the topology-frequency domain collaborative field-aware network are iteratively optimized through backpropagation to obtain the trained topology-frequency domain collaborative field-aware network. The field cognition loss function includes field classification loss, field topology consistency loss, and field frequency domain consistency loss.

[0144] Optionally, this embodiment designs a triple-field cognitive loss function to achieve joint supervision from three dimensions: pixel classification, topological structure, and frequency domain distribution. The total loss function formula is as follows:

[0145]

[0146] in, Let be the field cognition loss function. , As a balance factor, For field classification loss, The field topological consistency loss is used to penalize connectivity breaks in the predicted field. This is the frequency domain consistency loss, used to constrain the consistency of the frequency domain distribution between the predicted field and the true field; in this embodiment, Pick , Pick It can be adaptively adjusted according to the characteristics of the dataset; the detailed definitions of each loss term are as follows:

[0147] Step 5.1: The field classification loss adopts a weighted focus loss, which adaptively weights the sample size and classification difficulty of different land cover categories to solve the problems of imbalanced land cover categories and insufficient learning of difficult-to-classify samples in remote sensing images. The formula is as follows:

[0148]

[0149] Where H=512 and W=512 are the image dimensions; The category balancing weights are set to the reciprocal of the pixel percentage of each category, and their sum after normalization is C. To focus on the parameter, this embodiment uses 2, which is used to amplify the loss weight of hard-to-classify samples; Let (x, y) be the predicted probability that pixel (x, y) belongs to the c-th type of land cover. This is the true label (one-hot form) of pixel (x,y), which is 1 if it belongs to class c, and 0 otherwise.

[0150] Step 5.2: Field Topology Consistency Loss: This step constrains the consistency of the predicted results with the topological structure of the ground labels, strengthening the spatial continuity of the ground features. To ensure comparability of the loss terms across channels, the ground labels are first... Extended by one-hot encoding to Then compare the predicted probability map P with the one-hot encoded true label. The spatial gradient magnitude is calculated independently for each channel, using the following formula:

[0151]

[0152] in, To predict the gradient magnitude of the c-th channel of the probability map at pixel (x,y), The gradient magnitude of the c-th channel of the true label one-hot encoding at pixel (x,y) is used. This loss, through gradient constraints, ensures that the predicted land cover boundaries are consistent with the true topology, thus avoiding fragmentation of the classification results.

[0153] Step 5.3: Frequency Domain Consistency Loss: Introducing a global frequency domain constraint, the network compares the frequency domain distribution differences between the predicted and actual fields, guiding the network to learn the texture periodicity patterns of ground objects, and comparing the predicted probability map P with the one-hot encoded ground labels. Perform a two-dimensional FFT on each channel independently to obtain its own amplitude spectrum. The formula is: Calculate the L2 norm squared difference for each channel and then average it over all channels.

[0154]

[0155] in, This represents a two-dimensional Fast Fourier Transform. The amplitude spectrum is obtained by taking the modulus of the complex spectrum. Using the L2 norm, the field frequency domain consistency loss constraint ensures that the prediction results are consistent with the frequency domain texture distribution of the real labels, thus improving the classification accuracy of scenes with different objects in the same spectrum and different spectra in the same object.

[0156] Step 5.4: The topology-frequency domain collaborative field perception network is an end-to-end deep neural network consisting of a topology-frequency domain collaborative field perception encoder, a field dynamic aggregation bottleneck layer, and a field cognitive decoder connected end-to-end in sequence. The encoder is responsible for multi-level field decomposition and recombination, the bottleneck layer is responsible for global field perception and sparse filtering, and the decoder is responsible for semantic field inversion. The three components work together with the field cognitive loss function to complete the training. The specific training process is as follows:

[0157] The preprocessed training samples are input into the network, and the total loss is calculated. Then, the network parameters are iteratively optimized through backpropagation. The preprocessed training samples are specifically: training images that have been processed sequentially through Step 1.1-Step 1.3 to obtain 512×512×3 standardized images (the standardized images are then mapped to 512×512×C composite field feature maps through field representation in Step 1.4, which serve as the input to the first-level unit of the encoder). The specific training process and parameter configuration are as follows:

[0158] Step 5.4.1: Dataset Construction and Preprocessing. Collect remote sensing images of the target scene and corresponding pixel-level ground feature labels, and randomly divide them into training, validation, and test sets in a 7:2:1 ratio; all images and labels undergo size normalization processing as described in Step 1, uniformizing them to 512×512 pixels; the training set employs data augmentation strategies such as random horizontal flipping, vertical flipping, 90° rotation, brightness perturbation, and Gaussian noise injection to improve the model's generalization ability.

[0159] Step 5.4.2: Training Parameter Configuration. The optimizer used is AdamW, with an initial learning rate of 1e-4, a weight decay coefficient of 1e-5, an exponential decay rate β1 of the first moment estimate of 0.9, and an exponential decay rate β2 of the second moment estimate of 0.999. The batch size is set to 8, and the maximum number of training epochs is set to 200. A cosine annealing learning rate scheduling strategy is adopted, with a minimum learning rate of 1e-6 and an annealing period of 20 epochs.

[0160] Step 5.4.3: Model Convergence and Saving. During training, after each full training iteration, a comprehensive index is calculated on the validation set. The formula for the comprehensive index is:

[0161]

[0162] in, As a comprehensive indicator, For the average intersection-to-union ratio, the normalized topology loss and the normalized frequency domain loss are respectively derived from the field topology consistency loss. and field frequency domain consistency loss After mapping to the [0,1] interval using min-max normalization, the result is as follows:

[0163] During training, the validation set data for each round is statistically analyzed. , The historical minimum and maximum values ​​are mapped as follows:

[0164]

[0165] In the formula, , These represent the minimum and maximum values ​​of the topological consistency loss across all historical validation sets. , These represent the minimum and maximum values ​​of the frequency domain consistency loss on the validation sets in each historical round, respectively. The topology consistency score (TopoScore) is used. The FreqScore is used to calculate the frequency domain consistency score. The model weight with the highest overall metric on the validation set is saved for each round. An early stopping strategy is set to terminate training early when the overall metric on the validation set has not improved for 10 consecutive rounds, in order to avoid overfitting.

[0166] Step 6: Input the remote sensing image to be parsed into the trained topology-frequency domain collaborative field perception network, and output the semantic field map.

[0167] Optionally, the remote sensing image to be parsed undergoes preprocessing identical to that in the training phase, and is then input into the best-trained model to complete end-to-end semantic parsing. The specific process is as follows:

[0168] Step 6.1: Preprocess the image to be parsed by performing input validation, size normalization, and color space normalization in sequence to obtain a standardized image of 512×512×3.

[0169] Step 6.2: Input the standardized image into the trained optimal model. After forward propagation, the model outputs a 512×512×C multi-channel prediction probability map.

[0170] Step 6.3: Perform pixel-by-pixel category determination on the predicted probability map, take the land cover category corresponding to the maximum probability of each pixel in C channels as the final labeling result of the pixel, and generate a semantic field map with a size of 512×512, where each pixel value is the category number of the corresponding land cover;

[0171] Step 6.4: Restore the semantic field map to the original size of the image to be parsed by bicubic interpolation, thus completing the final semantic parsing of the remote sensing image.

[0172] It is understood that this embodiment constructs the overall architecture of the topology-frequency domain collaborative field sensing encoder, the field dynamic aggregation bottleneck layer, and the field cognitive decoder through Step 1-Step 6, as well as the collaborative mechanism of field feature decoupling-recombination, field dynamic aggregation, and field cognitive decoding.

[0173] The invention will be further illustrated by a specific example below.

[0174] Optionally, this example uses high-resolution remote sensing imagery as input, specifically an RGB three-channel remote sensing image of an urban scene with a spatial resolution of 0.06 m / pixel, an original image size of 1024×1024 pixels, a pixel value range of [0,255], and a bit depth of 8 bits. The total number of land cover categories is set to C=6 (buildings, roads, trees, cars, low vegetation, background).

[0175] S1: Topological-frequency domain field initialization preprocessing of remote sensing images

[0176] S1.1: Input verification of original remote sensing image: The original remote sensing image size is 1024×1024 pixels, RGB three channels, pixel value range [0,255], bit depth 8bit, which meets the input adaptation requirements of this invention and does not require channel conversion or zero padding.

[0177] S1.2: Image Size Normalization: The original image has a side length of 1024 pixels, which is greater than 512 pixels. Therefore, the original image is first cropped to a square with a side length of 512 pixels, and then the details are smoothed using a bicubic interpolation algorithm. The kernel function of the bicubic interpolation adopts the standard form, and the interpolation coefficient is -0.5. The final output is an RGB image with a size strictly of 512×512 pixels.

[0178] S1.3: Color Space Normalization: First, divide each pixel value of the size-normalized image by 255, linearly mapping from [0,255] to the [0,1] interval. Then, perform standardization processing on each channel separately: subtract the preset mean (red 0.485, green 0.454, blue 0.406), and then divide by the preset standard deviation (red 0.225, green 0.224, blue 0.229), to obtain a 512×512 pixel three-channel normalized image.

[0179] S1.4: Field Representation Construction: The normalized image is considered as a composite field composed of six subfields of land features. For each land feature, a 3×3 learnable field-aware convolutional kernel is used. This kernel has 3 input channels, 1 output channel, a stride of 1, and uses the same padding method. The kernel weights are initialized using a Kaiming uniform distribution. The normalized image is convolved with this kernel, and then multiplied by a learnable class weight (initial value 1.0) to obtain the subfield feature map for that land feature. The six subfield feature maps are concatenated along the channel dimension to obtain a composite field feature map of size 512×512×6, completing the mapping from pixel space to field space, which serves as the input to the first-level unit of the encoder.

[0180] S2: Topology-Frequency Domain Co-field Sensing Coding

[0181] Specifically, the channel configuration of the encoder's three-level units is as follows: Level 1: Cin=6, Cout=32, input 512×512×6, unit output 512×512×32 (→skip1); Level 2: Cin=64, Cout=64, input 256×256×64, unit output 256×256×64 (→skip2); Level 3: Cin=128, Cout=128, input 128×128×128, unit output 128×128×128 (→skip3). The channel transformation of the inter-layer downsampling convolution is: 32→64 (after Level 1), 64→128 (after Level 2), 128→128 (compressed to the bottleneck layer input after Level 3). Each field-sensing unit includes a field feature decoupling module and a field feature reconstructing module. The following is a detailed explanation using the first-level unit (input 512×512×6, unit output 512×512×32) as an example.

[0182] S2.1: Field feature decoupling: A parallel dual-path structure is adopted to synchronously decouple the input composite field into topological field components and frequency domain field components.

[0183] Specifically, the topology path uses four 3×3 orientation-sensitive convolutional kernels, corresponding to the four orthogonal directions of 0°, 45°, 90°, and 135°. Each kernel has 6 input and 6 output channels, a stride of 1, and the same padding method. The four convolutional results are concatenated along the channel dimension to obtain a 512×512×24 topology response map. Batch normalization and ReLU activation are then performed sequentially, followed by a 1×1 convolution to compress the number of channels from 24 back to 6. Finally, sigmoid activation is used to generate topology attention weights. Its dimensions are 512×512×6. Finally, the input features... (512×512×6) and Element-wise multiplication yields the topological field components. The dimensions are 512×512×6.

[0184] Specifically, the frequency domain path involves independently performing a two-dimensional fast Fourier transform on each channel of the input feature to transform the spatial domain field to the frequency domain, obtaining a complex spectrum. The magnitude spectrum (size 512×512×6) is then obtained by taking the modulus. Four sets of 3×3 band-sensitive convolutional kernels are used, corresponding to the low-frequency, mid-low-frequency, mid-high-frequency, and high-frequency bands, respectively. Each convolutional kernel has 6 input channels and 6 output channels, a stride of 1, and the same padding method. The four convolutional results are concatenated along the channel dimension to obtain a 512×512×24 frequency domain response map. Subsequently, batch normalization and ReLU activation are performed sequentially, followed by compression back to 6 channels via 1×1 convolution, and finally, frequency domain attention weights are generated through Sigmoid activation. The dimensions are 512×512×6. The input features are then compared with... Element-wise multiplication yields the frequency domain field components. The dimensions are 512×512×6.

[0185] S2.2: Field Feature Recombination: Recombining Topological Field Components (6 channels) and frequency domain field components (6 channels) are concatenated along the channel dimension to obtain a 512×512×12 concatenated feature (i.e., 2×Cin=2×6=12). A 1×1 convolutional kernel with fusion coefficients is used for convolution (12 input channels, 32 output channels, stride 1, same padding), and then the adaptive fusion coefficients are obtained through the Sigmoid function. The dimensions are 512×512×32. Meanwhile, the original input... Perform a 3×3 basic convolution (6 input channels, 32 output channels, stride 1, same padding) to extract basic field features. The dimensions are 512×512×32. Two independent 1×1 channel projective convolutions are used to... (6 channels) and (6 channels) mapped to 32 channels, resulting in and (All are 512×512×32). The final enhanced field characterization is obtained. The size is 512×512×32, which serves as the first-level skip connection feature skip1.

[0186] The enhanced field representation output from the first stage is passed through a 2×2 downsampling convolution (32 input channels, 64 output channels, stride 2, same padding) to obtain a feature map of size 256×256×64, which is then input to the second-stage field-aware unit. The second-stage unit outputs a 256×256×64 enhanced field representation (i.e., skip2), which is downsampled (64 input channels, 128 output channels, stride 2) to obtain a 128×128×128 feature map, which is then input to the third-stage field-aware unit. The third-stage unit outputs a 128×128×128 feature map (i.e., skip3), which is then compressed in size by a 2×2 downsampling convolution (128 input channels, 128 output channels, stride 2) to obtain a 64×64×128 deep enhanced field representation, which serves as the input to the bottleneck layer.

[0187] S3: Field Dynamic Aggregation Bottleneck Layer Processing

[0188] Optionally, the bottleneck layer receives the 64×64×128 enhanced field representation of the encoder's final output and passes it sequentially through the field global perception module and the field sparse filtering module.

[0189] S3.1: Field-wide Awareness: Employs a non-local attention mechanism. First, three independent 1×1 convolutional kernels (128 input channels, 128 output channels) are used to generate a query matrix Q, a key matrix K, and a value matrix V, each with a size of 64×64×128. Q, K, and V are then flattened into 128×4096 two-dimensional matrices (4096 being the total number of pixels). The product of the transpose of Q and K is calculated, and then divided by... The image is scaled and then normalized using the Softmax function to obtain a 4096×4096 attention map A. A is multiplied by the transpose of V, and then the image is restored to a 64×64×128 feature map, yielding the global context vector. Finally, this context vector is added to the original input features using residual calculation to output the global perceptual feature, which still has a size of 64×64×128.

[0190] S3.2: Field Sparsity Filtering: Global average pooling is performed on the global sensing features, averaging all pixel values ​​for each channel to obtain a 1×1×128 global descriptor. This descriptor is input into two consecutive fully connected layers (using an SE-block bottleneck structure). The first layer compresses the 128 dimensions to 16 dimensions and applies ReLU activation (reduction rate r=8). The second layer maps the 16 dimensions back to 128 dimensions and applies Sigmoid activation to generate channel-gated vectors, with each element taking values ​​in the range [0,1]. The gated vectors are expanded into a 64×64×128-dimensional gate mask through broadcasting, and a sparsity strategy is adopted to retain the maximum K channels, where K is 128 (i.e., all channels are retained; this can be adaptively adjusted according to the number of land cover categories). Finally, the global sensing features are multiplied element-wise with the gate mask to obtain the bottleneck layer output features, which still have a size of 64×64×128.

[0191] S4: Field Cognition Decoding and Semantic Field Inversion

[0192] Optionally, the channel configuration of the three-stage decoder unit is as follows: Stage 1: Cin=128, Cout=128, corresponding to skip3 (128×128×128); Stage 2: Cin=128, Cout=64, corresponding to skip2 (256×256×64); Stage 3: Cin=64, Cout=32, corresponding to skip1 (512×512×32). Each stage of the decoding unit includes a field inversion module and a field boundary refinement module. The first-stage decoding unit (input 64×64×128, output 128×128×128) is used as an example.

[0193] S4.1: Field Inversion: Current Decoder Input Features The dimensions are 64×64×128, corresponding to the same scale of the encoder's skip connection feature. The dimensions are 128×128×128. First, a 4×4 transposed convolution (128 input channels, 128 output channels, stride 2, padding 1) is applied to... Perform upsampling to obtain an upsampled feature of 128×128×128. ,and The dimensions are exactly the same. and Concatenating along the channel dimension yields a concatenated feature size of 128×128×256 (i.e., 2×Cout=2×128=256). This is then convolved using a 1×1 fusion weight convolution kernel (256 input channels, 128 output channels, stride 1, same padding), followed by a Sigmoid function to generate adaptive fusion weights. The dimensions are 128×128×128. Field-adaptive fusion is performed: upsampled features are... and Element-wise multiplication, skip connection feature and Multiply each element-wise, then add the two together to obtain the fused feature. The dimensions are 128×128×128. Finally, Input two consecutive 3×3 convolution kernels (both input and output channels are 128, stride is 1, padding is the same), each convolution is followed by batch normalization and ReLU activation to complete the feature inversion of the first-level decoding unit, and the output size is 128×128×128.

[0194] S4.2: Field Boundary Refinement: For the fused feature (128×128×128) output by the field inversion module, the spatial gradient is calculated independently for each channel using the Sobel gradient operator (3×3) in both the horizontal and vertical directions, resulting in two gradient maps. The absolute values ​​of the two gradient maps are then added together to obtain the boundary gradient map. (128×128×128). The boundary gradient map is input into two consecutive 3×3 boundary feature extraction convolution kernels (both input and output channels are 128, stride 1, padding is the same) to extract boundary detail features. Then, the boundary features are multiplied by a learnable boundary enhancement coefficient (initial value 0.1) and added to the original fused features using residuals to obtain the boundary-enhanced features with a size of 128×128×128, which serve as the final output of the first-level decoding unit.

[0195] Subsequent decoding layer and classification head: The second-level decoding unit takes 128×128×128 as input, performs a 4×4 transposed convolution (input 128, output 64, stride 2) → upsamples the feature to 256×256×64; concatenates it with skip2 (256×256×64) to obtain 256×256×128; fuses the weighted convolution (input 128, output 64); after fusion, performs two 3×3 convolutions (input and output 64); refines the boundary → output 256×256×64. The third-level decoding unit takes a 256×256×64 input, undergoes a 4×4 transposed convolution (64 inputs, 32 outputs, stride 2) → upsampled features of 512×512×32; concatenates with skip1 (512×512×32) to obtain 512×512×64; performs a weighted convolution (64 inputs, 32 outputs); after fusion, performs two 3×3 convolutions (32 inputs, 32 outputs); refines the boundaries → output 512×512×32. Classification Head: The final 512×512×32 features output by the decoder are convolved by a 1×1 classification convolution kernel (32 input channels, 6 output channels, stride 1, same padding), then pass through a Softmax activation function, outputting a multi-channel land cover category prediction probability map of size 512×512×6, where the sum of the probabilities of each pixel in the 6 channels is 1.

[0196] S5: Definition of Field Cognition Loss Function and Network Training

[0197] S5.1: Loss Function Definition: This example designs a triple-field cognitive loss function, jointly supervised from three dimensions: pixel classification, topological structure, and frequency domain distribution. The true label Y is expanded using one-hot encoding to... The data is strictly aligned with the P-dimensional prediction probability map of the 6 channels. The total loss is the classification loss plus 0.8 times the topological loss plus 0.5 times the frequency domain loss. The field classification loss uses a weighted focal loss. For the 6 land cover classes, the pixel proportion of each class is first calculated, its reciprocal is taken, and normalized to obtain the class balance weight. Focus parameters A value of 2 is used to amplify the loss weights for hard-to-classify samples. The weighted cross-entropy between the true one-hot label and the predicted probability is calculated pixel by pixel and multiplied by [the value]. Item. Field topology consistency loss first applies to P (6 channels) and (6 channels) Calculate the spatial gradient magnitude independently for each channel to obtain Then, the absolute differences between the two values ​​are calculated by pixel and channel, and the average is taken. This loss constrains the predicted ground feature boundaries to be consistent with the actual topology, avoiding classification fragmentation. The field frequency domain consistency loss is applied to P and Each channel independently performs a two-dimensional fast Fourier transform, takes the amplitude spectrum (size 512×512), calculates the sum of the squared differences of all pixels between the two amplitude spectra, divides it by the total number of pixels, and finally averages the results of the six channels. The loss-constrained prediction results are consistent with the frequency domain texture distribution of the real label.

[0198] S5.2: Network Training: This example uses the Potsdam high-resolution remote sensing dataset released by the International Society for Photogrammetry and Remote Sensing (ISPRS) as an example for model training. The dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio, with 1050 images in the training set, 300 in the validation set, and 150 in the test set. All images and labels are normalized to 512×512 pixels as described in Step 1. Data augmentation is used during training: random horizontal flipping, vertical flipping, 90° rotation, brightness perturbation (±0.2), and Gaussian noise injection (standard deviation 0.01). Training parameter configuration: The optimizer is AdamW, with an initial learning rate of 1e-4, a weight decay coefficient of 1e-5, an exponential decay rate of 0.9 for the first moment estimate, and an exponential decay rate of 0.999 for the second moment estimate; the batch size is set to 8, and the maximum number of training epochs is set to 200; a cosine annealing learning rate scheduling strategy is used, with a minimum learning rate of 1e-6 and an annealing cycle of 20 epochs. Model convergence and storage: After each round of full training, a comprehensive index is calculated on the validation set. Save the model weights with the highest overall validation set metric in each round. Implement an early stopping strategy: terminate training early if the overall validation set metric does not improve for 10 consecutive rounds to avoid overfitting.

[0199] S6: Remote Sensing Image Field Analysis Inference and Semantic Field Map Output

[0200] Optionally, in this step, the image to be parsed is input into the trained network to complete end-to-end semantic parsing and output a semantic field map. The complete process of remote sensing image field parsing inference in this step is as follows: Figure 5 As shown: Figure 5 (a) is the original remote sensing image input. Figure 5 (b) is the multi-channel probability map of the model's forward inference output. Figure 5 (c) is a visualization of the semantic field generated by the multi-channel probability graph. Figure 5 (d) is the final output remote sensing semantic segmentation map; Figure 5 (a) to Figure 5 (d) The processing inputs and results corresponding to S6.1 to S6.5 below are used to intuitively demonstrate the step-by-step parsing process of this invention from the original remote sensing image to the semantic segmentation result. The specific implementation is as follows:

[0201] S6.1: Image Preprocessing: Perform the same preprocessing as in the training phase (input validation, size normalization, color space normalization) on the remote sensing image to be analyzed, making it a standardized image with a size of 512×512×3. The original remote sensing image to be analyzed is shown below. Figure 5 As shown in (a), this is a three-channel RGB remote sensing image of an urban scene.

[0202] S6.2: Forward Inference: The standardized image is input into the trained network model, and a multi-channel prediction probability map of size 512×512×6 is obtained through forward propagation. Each pixel in this probability map has 6 probability values, representing the likelihood that the pixel belongs to 6 land cover categories, and the sum of the 6 values ​​is 1. The visualization result of this multi-channel prediction probability map is as follows. Figure 5 As shown in (b), the brighter the gray area in the figure, the higher the probability that the location belongs to the corresponding land cover category.

[0203] S6.3: Pixel-by-Pixel Classification and Segmentation Region Extraction: Perform pixel-by-pixel classification on the predicted probability map, taking the land cover category corresponding to the maximum probability value of each pixel across the six channels, and generating a semantic field map of size 512×512. The visualization result of the generated semantic field map is shown below. Figure 5As shown in (c), the figure uses continuously distributed grayscale to visually reflect the spatial response and attribution of each feature subfield. In the semantic field map, the value of each pixel is its feature category number (0-5). Next, connected component analysis is used to divide the pixel set of the same category into several independent segmentation regions. For each segmentation region, its contour polygon is extracted (implemented using a boundary tracing algorithm in image processing), and the following information is recorded: the sequence of contour polygon vertices (pixel coordinates arranged in clockwise or counterclockwise order), the circumscribed rectangle (coordinates of the upper left corner and lower right corner, used for fast localization), and the region category (such as buildings, roads, etc.).

[0204] S6.4: Generate a segmentation coordinate map: Based on the information of each segmented region described above, generate a segmentation coordinate map, which describes the region distribution of the entire image in list form. An example is shown below:

[0205] Category: Building, Region 1 outline vertex sequence: (125,82), (348,78), (352,218), (118,222), (120,150), (126,148) form a closed region. Number of vertices: 6, circumscribed rectangle: top left (118,78), bottom right (352,222), circumscribed size 234×144.

[0206] Category: Trees, Region 1 outline vertex sequence: (52,52), (98,48), (102,148), (48,152), (50,100) form a closed region. Number of vertices: 5, circumscribed rectangle: top left (48,48), bottom right (102,152), circumscribed size 54×104.

[0207] S6.5: Generating the Final Semantic Segmentation Image: Using the segmentation coordinate map, each segmented region is mapped back to the original image, and the regions are filled according to category to generate the final semantic segmentation image. The generated final remote sensing semantic segmentation image is shown below. Figure 5 As shown in (d), the different grayscale (or color) areas in the image correspond to six types of land features: buildings, roads, trees, cars, low vegetation, and background. If the original image size is larger than... The semantic field map can be restored to its original size using bicubic interpolation. The final output image displays the region for each land cover category. (Comparison) Figure 5 (a) to Figure 5 (d) As can be seen, the present invention can process the input remote sensing raw image ( Figure 5 (a) is sequentially analyzed into a multi-channel probability map ( Figure 5 (b) Semantic field visualization diagram ( Figure 5 (c) and the final semantic segmentation map ( Figure 5(d) The obtained segmentation results are semantically clear, have continuous boundaries, and have complete ground feature structures, which intuitively verifies the effectiveness of the topology-frequency domain collaborative field perception method of the present invention in remote sensing semantic analysis of complex urban scenes.

[0208] In summary, this invention breaks through the traditional pixel classification paradigm, treating images as composite physical fields composed of different ground feature subfields: It constructs a topology-frequency domain collaborative field-aware encoder, simultaneously extracting the topological connectivity and frequency domain fluctuation patterns of ground features through field feature decoupling and recombination; it constructs a field dynamic aggregation bottleneck layer to achieve global field perception and sparse filtering; it constructs a field cognition decoder, retrieving the semantic field through field adaptive fusion and boundary refinement; and it designs a field cognition loss function, introducing topological consistency loss and frequency domain consistency loss constraints. This invention enables the network to leap from memorizing pixels to recognizing fields, significantly improving the resolution accuracy, structural integrity, and generalization ability of complex ground features. It provides a paradigm shift technology path for remote sensing image interpretation from image processing to scene cognition, applicable to interpreting the category attributes, spatial morphology, and relationships of ground features from high-resolution remote sensing images, and suitable for scenarios such as smart cities, precision agriculture, and ecological monitoring.

[0209] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for semantic parsing of remote sensing images based on a topology-frequency domain cooperative field sensing network, characterized in that, The method includes the following steps: Step 1: Perform field initialization preprocessing on the remote sensing image to obtain the image field representation; Step 2: Construct a topology-frequency domain collaborative field-aware encoder to perform multi-level field decomposition and recombination on the image field representation to obtain a deep field representation; wherein, the topology-frequency domain collaborative field-aware encoder includes a field feature decoupling module and a field feature recombination module. The field feature decoupling module decouples the image field representation into topological field components and frequency domain field components in parallel through topological projection basis and frequency domain projection basis. The field feature recombination module cross-fuses the topological field components and frequency domain field components through an inter-field attention mechanism to generate a single-level enhanced field representation. The single-level enhanced field representations are stacked level by level to obtain a deep field representation. Step 3: Construct a field dynamic aggregation bottleneck layer to perform global field perception and sparse filtering on the deep field representation to obtain globally optimized field features; wherein, the field dynamic aggregation bottleneck layer includes a field global perception module and a field sparse filtering module. The field global perception module uses nonlocal attention to capture long-range dependencies between fields and generate field context vectors. The field sparse filtering module generates sparse gating masks based on the field context vectors and dynamically filters key field components. Step 4: Construct a field cognitive decoder to invert the globally optimized field features into a pixel-level semantic field; wherein, the field cognitive decoder includes a field inversion module and a field boundary refinement module. The field inversion module dynamically fuses the decoder features with the corresponding encoder features through a field adaptive fusion mechanism, and the field boundary refinement module uses field gradient information to guide the localization of ground feature edges. Step 5: The topology-frequency domain collaborative field-aware encoder, the field dynamic aggregation bottleneck layer, and the field cognition decoder are sequentially connected to form a topology-frequency domain collaborative field-aware network. A field cognition loss function is constructed, and the loss value is calculated using the pixel-level semantic field and the real label semantic field. The parameters of the topology-frequency domain collaborative field-aware network are iteratively optimized through backpropagation to obtain the trained topology-frequency domain collaborative field-aware network. The field cognition loss function includes field classification loss, field topology consistency loss, and field frequency domain consistency loss. Step 6: Input the remote sensing image to be parsed into the trained topology-frequency domain collaborative field perception network, and output the semantic field map.

2. The remote sensing image semantic parsing method based on a topology-frequency domain cooperative field sensing network according to claim 1, characterized in that, Step 1 specifically refers to: Step 1.1: Perform input verification of the original remote sensing image. If the input remote sensing image is a single-channel grayscale image, convert it into an RGB three-channel image by channel copying; if the input remote sensing image is a non-square rectangular image, fill it with zeros to make it square. Step 1.2: Scale the verified remote sensing images to a fixed size; Step 1.3: After normalizing the size of the RGB three-channel image, the pixel values ​​are linearly mapped from [0,255] to the [0,1] interval, and then the image after linear mapping is normalized. Step 1.4: The standardized image is considered as a composite field composed of multiple subfields of ground features. Each pixel (x, y) in the composite field is assigned a field attribute. The image field representation is constructed by convolving the standardized image with a pre-defined prior distribution of ground features. The expression is as follows: ； in, Let be the response value of the c-th type of land feature subfield at (x, y); For the standardized image, This is the field-aware convolution kernel corresponding to the c-th type of land cover; These are the learnable category weights; This represents the convolution operation. This represents the total number of land cover categories in the remote sensing scene; the convolution results of all land cover subfields are concatenated along the channel dimension to obtain a composite field feature map, completing the mapping from pixel space to field space.

3. The remote sensing image semantic parsing method based on a topology-frequency domain cooperative field sensing network according to claim 1, characterized in that, The topology-frequency domain collaborative field-aware encoder is composed of three stacked field-aware units. Each field-aware unit includes a field feature decoupling module and a field feature recombination module. The field feature decoupling module is a parallel dual-path field-aware convolution, where the dual paths are a topological path and a frequency domain path. Specifically: In the topological path, the input image is first processed by a set of trainable orientation-sensitive convolutional kernels. Four sets of parallel convolutions are performed, and the results of the four branches are concatenated along the channel dimension to obtain the topological response map. Then, batch normalization (BN) and ReLU activation functions are sequentially applied to the topological response map. The number of channels is compressed using a 1×1 channel compressed convolution kernel. Finally, the topological attention weights are generated using the Sigmoid activation function. The topological field components are obtained by element-wise multiplying the input image with the topological attention weights. ; In the frequency domain path, the input image is first processed. Each channel independently performs a two-dimensional fast Fourier transform, transforming the field from the spatial domain to the frequency domain to obtain the complex spectrum. Then, the amplitude spectrum is obtained by taking the modulus of the complex spectrum. A set of trainable band-sensitive convolutional kernels is used to perform parallel convolution on the amplitude spectrum. The four sets of convolutional kernels correspond to the low-frequency, mid-low-frequency, mid-high-frequency, and high-frequency bands, respectively, to obtain the frequency domain response map. Then, batch normalization (BN) and ReLU activation functions are sequentially applied to the frequency domain response map. The number of channels is compressed by a 1×1 channel compression convolutional kernel. Finally, the frequency domain attention weights are generated by the Sigmoid activation function and multiplied element-wise with the input image to obtain the frequency domain field components. ; The field feature reconstruction module first reconstructs the topological field components. and frequency domain field components The concatenated features are obtained by concatenating along the channel dimension; the concatenated features are then convolved using a 1×1 fusion coefficient convolution kernel, and the adaptive fusion coefficients are calculated using the Sigmoid function. ; For the input image Perform a 3×3 basic feature extraction convolution kernel to extract basic field features. And through two independent 1×1 channel projective convolution kernels, respectively and The topological field components are obtained by mapping from the input channel to the output channel. and frequency domain field components The final output is the enhanced field representation of the current-level unit. The expression is: ； Enhanced field characterization of each unit output Then, the feature map size is halved through convolution and input into the next level field perception unit; the enhanced field representations output by each level field perception unit are simultaneously saved as encoder skip connection features of the corresponding level.

4. The remote sensing image semantic parsing method based on topology-frequency domain cooperative field sensing network according to claim 1, characterized in that, The specific field dynamic aggregation bottleneck layer is: In the field-wide perception module, the input features are first processed by three independent 1×1 channel transformation kernels. Convolution is performed separately to generate a query matrix Q, a key matrix K, and a value matrix V. The transpose of Q is multiplied by the matrix product of K, and then the result is passed through a softmax function to obtain an attention map A. The attention map A is then multiplied by the transpose of V, and the feature map size is restored to obtain the global context vector, which is then compared with the input feature map. Perform residual summation to obtain the global perception features. ; In the sparse screening module, the first step is to... Perform global average pooling, averaging all pixel values ​​for each channel to obtain the global descriptor, expressed as: ； in, For the c-th channel component of the global descriptor z, Globally perceptible features The value of the c-th channel at (x, y); H and W are the feature map dimensions; The global descriptor is input into two consecutive fully connected layers. The first fully connected layer compresses the vector and processes it using the ReLU activation function. The second fully connected layer maps the vector back to its original dimension and generates a channel-gated vector G using the Sigmoid function, with each element taking values ​​in the range [0,1]. The gated vector G is then expanded using a broadcast operation to a dimension similar to the original vector. Given gate masks M of the same size, retain the first few values ​​of the gate mask with the largest values. One channel is used, and the mask values ​​of all other channels are set to zero; the final field dynamic aggregation bottleneck layer outputs the globally optimized field features as follows: .

5. The remote sensing image semantic parsing method based on a topology-frequency domain cooperative field sensing network according to claim 1, characterized in that, The field cognition decoder consists of three levels of decoding units that correspond one-to-one with the levels of the topology-frequency domain collaborative field perception encoder. Each level of the decoding unit includes a field inversion module and a field boundary refinement module, specifically: In the in-situ inversion module, a 4×4 transposed convolution kernel is first used to process the input features of the current decoder. Perform upsampling and output the upsampled features. Upsampled features Jump connection features with encoder The concatenated features are obtained by concatenating along the channel dimension, then convolved using a 1×1 fusion weight convolution kernel, and finally the adaptive fusion weights are calculated using the Sigmoid function. ; Generate fusion features through field-adaptive fusion The expression is: ； Fusion features The input 3×3 feature decoding convolution kernel performs two consecutive convolutions. After each convolution, batch normalization and ReLU activation functions are applied to complete the feature inversion of the current level decoding unit. In the field boundary refinement module, the fusion features output by the field inversion module are... The spatial gradients in the x and y directions are calculated independently for each channel to obtain the boundary gradient map. The expression is: ； in, and These are the Sobel gradient operators in the x and y directions, respectively; By performing convolution on the boundary gradient map using two consecutive 3×3 boundary feature extraction convolution kernels, detailed boundary features are extracted. Each convolution is followed by BN and ReLU activation functions to combine boundary features with fused features. Perform residual summation to complete boundary reinforcement; the expression is: ； in, These are the learnable boundary enhancement coefficients; The final output of the current decoding unit is input to the next decoding unit. The features output from the last unit of the decoder are convolved by a 1×1 classification convolution kernel, and then activated by the Softmax activation function. The final output is a multi-channel land cover category prediction probability map. The sum of the values ​​of each pixel in C channels is 1, which represents the probability that the current pixel belongs to each type of land cover.

6. The remote sensing image semantic parsing method based on topology-frequency domain cooperative field sensing network according to claim 1, characterized in that, The field cognition loss function is specifically as follows: ； in, Let be the field cognition loss function. , As a balance factor, For field classification loss, For field topology consistency loss, This represents the loss of consistency in the field frequency domain; The expression for the field classification loss is as follows: ； Where H and W are the feature map dimensions, This represents the total number of land cover categories in the remote sensing scene. Balance the weights for each category; For focusing parameters; Let (x, y) be the predicted probability that pixel (x, y) belongs to the c-th type of land cover. This is the true label of pixel (x, y), which is 1 if it belongs to class c, and 0 otherwise; Among them, the field topology consistency loss first converts the real label Extended by one-hot encoding to Then compare the predicted probability map P with the one-hot encoded true label. The spatial gradient magnitude is calculated independently for each channel, and the expression is: ； in, To predict the gradient magnitude of the c-th channel of the probability map at pixel (x,y), Let c be the gradient magnitude of the c-th channel of the true label one-hot encoding at pixel (x,y). Among them, the field frequency domain consistency loss affects the predicted probability map P and the one-hot encoded true label. Perform a two-dimensional FFT on each channel independently to obtain its own amplitude spectrum. and The expression for calculating the L2 norm squared difference for each channel and averaging it over all channels is: ； in, This represents a two-dimensional Fast Fourier Transform. The amplitude spectrum is obtained by taking the modulus of the complex spectrum. It is an L2 norm.

7. The remote sensing image semantic parsing method based on a topology-frequency domain cooperative field sensing network according to claim 6, characterized in that, During the training process, the topology-frequency domain collaborative field sensing network calculates a comprehensive index after each round of full training, expressed as: ； in, As a comprehensive indicator, For the average intersection-to-union ratio, the normalized topology loss and the normalized frequency domain loss are respectively derived from the field topology consistency loss. and field frequency domain consistency loss After being mapped to the [0,1] interval using minimum-maximum normalization, the result is as follows: Statistical analysis of each round of validation sets , The historical minimum and maximum values ​​are mapped using the following expression: ； ； In the formula, , These represent the minimum and maximum values ​​of the topological consistency loss across all historical validation sets. , These are the minimum and maximum values ​​of the frequency domain consistency loss in each historical validation set round. The model weight with the highest comprehensive index in each validation set round is saved. An early stopping strategy is set so that training is terminated early when the comprehensive index of the validation set does not improve for a preset number of consecutive rounds.

Images