Facility layout scheme generation method based on space-time knowledge graph

Abstract

The invention discloses a space-time knowledge graph-based facility layout scheme generation method. The method comprises the steps of S1, constructing an urban facility layout knowledge base; s2, acquiring a reference case place set of the target grid to be optimized; s3, extracting a facility significant co-occurrence mode in the reference case place set, and taking the facility significant co-occurrence mode as a recommended facility configuration mode of the target grid to be optimized; and S4, according to the feature vector of the city built environment of the target to-be-optimized grid obtained in the S1, inputting the feature vector and the recommended facility configuration mode of the target to-be-optimized grid obtained in the S3 into the constructed facility layout generation model, and obtaining an output result which is a facility layout scheme. According to the method, the facility distribution rule knowledge and the facility layout optimization intelligence are integrated, and the facility configuration recommendation and the facility layout optimization scheme can be automatically output for a specific land parcel.

Description

Technical Field

[0001] This invention relates to the field of urban planning technology, and in particular to a method for generating facility layout schemes based on spatiotemporal knowledge graphs. Background Technology

[0002] Currently, the optimization of public service facilities mainly relies on mathematical programming and spatial analysis methods, with typical methods including the p-median model, the maximum coverage model, and GIS-based spatial statistical methods. These methods allocate facility locations by optimizing travel distance, service radius, or coverage population. Some methods also introduce multi-objective optimization frameworks to balance fairness, efficiency, and cost constraints. At the technical implementation level, intelligent optimization techniques such as evolutionary algorithms, genetic algorithms, and simulated annealing have also been applied to facility layout problems, improving the model's solution capabilities. This research provides fundamental methodologies and practical examples for facility allocation.

[0003] However, existing facility optimization methods are mostly based on static population data or single spatial distribution data for analysis, lacking a unified expression and integration of multi-source heterogeneous data. They are difficult to simultaneously characterize key features such as the built environment, functional structure, and population vitality, resulting in insufficient accuracy and adaptability of facility layout schemes in complex urban environments, and a lack of dynamic intelligent planning methods.

[0004] Modern cities generate a large amount of multi-source data related to facilities, including remote sensing imagery, points of interest (POIs), population and GDP, travel activity, and road network data. These data come from different sources and are at different scales, thus requiring a unified technical framework for spatiotemporal alignment and modeling of multi-source data to extract multi-dimensional urban features and form a unified representation that can be used for retrieval, reasoning, and optimization. How to integrate multi-source data into a unified framework to form a spatiotemporal expression oriented towards decision-making is a core technical challenge for achieving intelligent facility layout. Although existing research has attempted to use deep learning or generative models for facility layout prediction, these generally lack integration with facility knowledge, easily generating results lacking spatial logic. Without constraints on graph structure consistency and benefit indicators, while generated schemes may possess formal diversity, they struggle to guarantee interpretability and planning rationality. Summary of the Invention

[0005] Purpose of the invention: To address the above-mentioned shortcomings, this invention proposes a facility layout scheme generation method based on spatiotemporal knowledge graphs, which integrates the knowledge of facility distribution patterns and the intelligent optimization of facility layout. It can automatically output facility configuration recommendations and facility layout optimization schemes for specific plots of land.

[0006] Technical solution: This invention provides a method for generating facility layout schemes based on spatiotemporal knowledge graphs, including: S1. Divide the learning area into grids and obtain the urban functional area division of each grid area. Construct a directed adjacency network of public service facility POIs in each urban functional area. Extract the feature vector of the urban built environment based on this network. Combine with multi-dimensional urban benefit indicators to construct an urban facility layout knowledge base. S2. Divide the target area to be optimized into a grid, obtain a set of several case locations, and based on the urban facility layout knowledge base constructed in S1, retrieve the reference case locations of the target grid to be optimized by searching based on the similarity of the feature vectors of the urban built environment of each grid area. S3. Extract significant co-occurrence patterns of facilities from the set of reference case sites, and use them as recommended facility configuration patterns for the target grid to be optimized; S4. Based on the feature vector of the urban built environment of the target grid to be optimized obtained in S1, input it and the recommended facility configuration mode of the target grid to be optimized obtained in S3 into the constructed facility layout generation model, and the output result is the facility layout scheme.

[0007] Specifically, in S1, the urban functional zone division of each grid area in the learning area is obtained, as follows: Public service facility POI data is acquired, and the public service facility POIs are reclassified based on the main functions of the city. Based on this, spatial corpus of each grid area in the learning area is constructed, and it is input into the word vector generation model to obtain vector representations of various public service facility POIs in each grid area. Based on this, regional vectors of each grid area are constructed and clustered to obtain the urban functional area division of each grid area in the learning area.

[0008] More specifically, the spatial corpus for each grid region of the learning area is constructed as follows: Based on the reclassification, within a certain grid area, each public service facility POI is traversed, and its neighboring public service facility POIs and their types are searched within a defined range. The interaction decay between each public service facility POI and its neighbor is calculated, and the POIs are sorted according to the decay order to obtain the corresponding public service facility POI sequence S. i ={t i1 ,t i2 ,…,t ij ,…,t iM}, where t ij The type of the j-th neighboring public service facility POI of the i-th public service facility POI is represented by M, and the number of neighboring public service facility POIs in the i-th public service facility POI is M. At the same time, the interaction attenuation between each public service facility POI and itself is calculated, thereby constructing the spatial corpus of the grid region.

[0009] Furthermore, the calculation of the interaction decay between each public service facility's POI and itself is as follows: ; Where d is the distance between two public service facility POIs; σ is a set scale parameter used to control the decay rate.

[0010] Furthermore, the word vector generation model adopts the Word2Vec model under the Skip-Gram architecture. For each public service facility POI corresponding to a public service facility POI sequence, within a set window width [-b, b], the type of a certain public service facility POI in its corresponding public service facility POI sequence is selected, i.e., the center word, and the types of other public service facility POIs are the context words. The optimization objective is to maximize the prediction probability of the center word for the context words, as follows: ; Wherein, logP(t) j+z |t k For a given type t of a public service facility POI j In the case of the central term, the type of POI (Point of Interest) of its adjacent public service facilities t j+z That is, the probability of the occurrence of context words.

[0011] More specifically, the construction of the region vector for each grid region is as follows: By statistically analyzing the TF-IDF values ​​of each type of public service facility POI within each grid area, and combining this with the vector representation of each type of public service facility POI, a region vector for each grid area is constructed, as follows: R a =(∑ a F at ×w at ) / (∑ t F at ); Among them, R a F is the region vector of the a-th grid region; at Let w be the TF-IDF value of the t-th type of public service facility POI in the a-th grid region; at The vector representation of the t-th type of public service facility POI in the a-th grid region is generated by the word vector generation model.

[0012] Specifically, in S1, the construction of a directed adjacency network of public service facility POIs within each urban functional area is as follows: Calculate the spatial radius of each public service facility POI within each urban functional area. Construct a two-dimensional Voronoi diagram for each type of public service facility POI to form multiple sets of directional Thiessen polygons. Starting from the public service facility POI at the center of each Thiessen polygon, retrieve other public service facility POIs within its range. Select public service facility POIs whose distance from the center public service facility POI is less than the aforementioned spatial radius to form multiple sets of adjacency pairs with the center public service facility POI, thereby obtaining spatial adjacency relationships and constructing a directed adjacency network of public service facility POIs.

[0013] Specifically, in S1, the extraction of the feature vector of the urban built environment is as follows: Urban remote sensing images are acquired and cropped based on urban built-up area grids to obtain a remote sensing image dataset. Local texture encoding and location encoding are used to obtain feature sequences containing spatial location information. These sequences are then input into a multi-layer sequence modeling module, and three feature propagation paths—forward scanning, reverse scanning, and random scanning—are fused to extract feature vectors of the urban built environment.

[0014] More specifically, the local texture encoding is as follows: After obtaining the remote sensing image dataset based on the grid clipping of the urban built-up area, the features of each image in the remote sensing image dataset are extracted in sequence. The image is divided into several image blocks with a set stride. Each image block is encoded using a two-dimensional convolution with a kernel size consistent with the size of the image block and a set stride, thereby obtaining the local texture vector of the image block. The location encoding is as follows: Construct the position index of the image patch based on its row and column indices in the image, and use it as the position embedding. The location embedding is added element-wise to the local texture vector to obtain the feature sequence containing spatial location information for the image patch. The feature sequences of all image patches are stacked to obtain the feature sequence containing spatial location information.

[0015] More specifically, based on the state-space model, a sequence modeling module with MambaMixer as the backbone is constructed to perform multi-level global context modeling on the local texture feature sequence containing spatial location information, thereby obtaining the multi-level sequence modeling module.

[0016] Furthermore, for the input local texture feature sequence containing spatial location information, a forward scan sequence, a reverse scan sequence, and a random scan sequence are constructed, and layer normalization and linear projection processing are performed on them respectively, mapping the features to two subspaces u. t ,v t and for u tApplying one-dimensional convolution for convolution enhancement, and enhancing the features after convolution enhancement. t Perform state-space recursive modeling, activate the output of the recursive modeling through a nonlinear function, and then combine it with v. t Gated fusion is performed, then the original dimension is restored through linear mapping, and residual connections are established with the input features. This is achieved through stacked L... Mamba The MambaMixer module of the layer processes the data and performs weighted fusion on the three sets of output features.

[0017] Specifically, in S1, the multidimensional urban benefit indicators include: Accessibility index: Calculated using cumulative opportunity accessibility, which is the number of public service point of interest (POI) accessible to the land parcel within the travel time threshold τ0; Diversity index: characterized by the number and distribution entropy of the types of public service points (POIs) accessible to the land parcel; Vitality Index: The vitality value of each public service facility's Point of Interest (POI) is used as the evaluation indicator; GDP.

[0018] Specifically, in S2, the cosine similarity calculation method is used to calculate the similarity of the feature vectors of the urban built environment.

[0019] Specifically, in S2, after selecting the set of case locations that meet the similarity threshold, the multidimensional urban benefit indicators in S1 can be combined to retain the set of case locations where all indicators are better than the target grid to be optimized, as the final set of reference case locations.

[0020] Specifically, in S3, the significant co-occurrence patterns of facilities extracted from the set of reference case locations are as follows: Let s be a source type node o Pointing to multiple target type nodes t e Based on different functional zones, different distance thresholds are selected to calculate weights, and the nodes are the Points of Interest (POIs) of public service facilities, specifically: ); Where, d r For source type node s o To target type node t e The actual distance, d e θ represents the distance threshold for the corresponding grid region; θ is the set smoothing parameter. The structural participation rate P is calculated after aggregating all source type nodes. R : ; Where the numerator is the sum of the average weights of the source type node connected to multiple target type nodes, and the denominator is... The number of source type nodes; Structural participation rate P R The type node group with a value ≥ξ is selected as a significantly co-occurring facility, where ξ is the selection threshold, thereby obtaining the significant co-occurrence patterns of facilities in the reference case set.

[0021] Specifically, in step S4, the facility layout generation model adopts a GAN model, and its loss function is a comprehensive loss function, as follows: ; Among them, L total Let L be the overall loss function, where α², β², and γ² are the weighting coefficients, respectively. GAN To combat the losses, L align To align the benefits with the losses, L struct For structural consistency loss, L fair To ensure fairness and constrain losses.

[0022] Beneficial effects:

[0023] This method, based on spatiotemporal knowledge embedding, combines remote sensing imagery, facility and activity data, and uses a spatiotemporal knowledge graph to achieve integrated modeling of multi-source data. Based on the graph and facility knowledge, it recommends facility configuration patterns and constructs an intelligent generation model for facility layout optimization. This invention integrates the knowledge-based understanding of facility distribution patterns and the intelligent optimization of facility layout, enabling automatic output of facility configuration recommendations and facility layout optimization schemes for specific plots of land. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in this invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of the facility layout scheme generation method based on spatiotemporal knowledge graph of the present invention.

[0026] Figure 2 Reclassification table for POIs of public service facilities.

[0027] Figure 3 This is an example diagram of a spatiotemporal knowledge graph architecture.

[0028] Figure 4 This is an example diagram of a spatiotemporal knowledge graph database.

[0029] Figure 5This is an example diagram of the architecture of a GAN model.

[0030] Figure 6 Example diagrams for extracting the patterns of facility combinations in different functional areas and scenarios.

[0031] Figure 7 This is an example diagram showing the output recommendations and layout results obtained using the method of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions and advantages of the present invention clearer, the present application will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0033] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of the present invention should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0034] The facility layout scheme generation method based on spatiotemporal knowledge graph of this invention is as follows: Figure 1 As shown, it includes: S1. Divide the learning area into grids to obtain the urban functional zones of each grid area. Construct a directed adjacency network of public service facility POIs in each urban functional zone. Extract the feature vector of the urban built environment based on this network. Combine with multi-dimensional urban benefit indicators to construct an urban facility layout knowledge base, i.e., a spatiotemporal knowledge graph.

[0035] Specifically, including: S11. Obtain POI data for public service facilities and reclassify the POIs for public service facilities based on the city's main functions.

[0036] In this invention, POI data for public service facilities can be extracted using existing map software such as Gaode Maps and Baidu Maps.

[0037] In this invention, the main functions of a city include residential, commercial, industrial, educational, or catering functions.

[0038] In this invention, by referring to tables and combining existing map software classifications of major urban functions (including original major categories and original subcategories), public service facility POIs can be reclassified into 10 major categories: education services, catering services, medical services, shopping services, living services, public transportation services, elderly care services, sports and entertainment services, culture and leisure services, and parks. Each major category has several subcategories. For specific reclassification details, please refer to [reference needed]. Figure 2 .

[0039] S12. Divide the learning area into grids. Based on the reclassification in S11, construct the spatial corpus of each grid area in the learning area. Input the corpus into the word vector generation model to obtain the vector representation of various public service facilities POIs in each grid area. Based on this, construct the regional vector of each grid area and perform clustering to obtain the urban functional area division of each grid area in the learning area.

[0040] In this invention, spatial corpora for each grid region of the learning area are constructed based on the reclassification in S11, as follows: Based on the reclassification in S11, within a certain grid area, each public service facility POI is traversed, and other public service facility POIs and their types are searched within a defined range centered on it, i.e., its neighboring public service facility POIs and their types. The interaction decay between each public service facility POI and its type is calculated, and the corresponding public service facility POI sequence S is obtained by sorting them according to the decay order. i ={t i1 ,t i2 ,…, t ij ,…,t iM}, where t ij Let M represent the type of the j-th neighboring public service facility POI of the i-th public service facility POI, and M be the number of neighboring public service facility POIs in the i-th public service facility POI. This is how the spatial corpus of the grid region is constructed.

[0041] In this embodiment, the type of public service facility POI is the middle class obtained in the reclassification of S11.

[0042] In this embodiment, a Gaussian distance decay weight w(d) is introduced to simulate the interaction decay between two public service facility POIs, as follows: ; Where d is the distance between two public service facility POIs; σ is a set scale parameter used to control the decay rate.

[0043] In this embodiment, the interaction attenuation between each public service facility POI and its neighboring public service facility POIs forms multiple spatial paths. The neighboring public service facility POIs and their types of all public service facility POIs can constitute a spatial corpus, which can reflect, within a set range, which public service facility POI types tend to co-occur and how this relationship attenuates with distance.

[0044] In this invention, the word vector generation model adopts the Word2Vec model, more specifically, the Word2Vec model under the Skip-Gram architecture. This model learns its semantic vector representation by predicting the types of neighboring public service facility POIs. The result obtained by this model makes the types of public service facility POIs with similar functional relationships or frequent co-occurrences closer in the vector space. For each public service facility POI's corresponding public service facility POI sequence, within a set window width [-b, b], the type of one public service facility POI in its corresponding public service facility POI sequence is selected, i.e., the center word, and the types of other public service facility POIs are the context words. The optimization objective of this model is to maximize the prediction probability of the center word for the context words, as follows: ; Wherein, logP(t) j+z |t k For a given type t of a public service facility POI j In the case of the central term, the type of POI (Point of Interest) of its adjacent public service facilities t j+z That is, the probability of the occurrence of context words.

[0045] In this invention, a region vector for each grid area is constructed. Specifically, this is done by statistically analyzing the TF-IDF values ​​of the types of public service facility POIs within each grid area, and combining these with the vector representations of various types of public service facility POIs, as follows: R a =(∑ a F at ×w at ) / (∑ t F at ); Among them, R a F is the region vector of the a-th grid region; at Let w be the TF-IDF value of the t-th type of public service facility POI in the a-th grid region; at Let be the vector representation of the t-th type of public service facility POI in the a-th grid region, generated by the aforementioned word vector generation model.

[0046] In this invention, K-Means clustering can be used, and the K value with the best contour coefficient is selected as the final clustering result, thereby completing the urban functional area division of each grid area of ​​the learning area.

[0047] In this invention, to specify the particular categories of each urban functional area, a facility density index and a specialization coefficient index can also be introduced. The facility density index is used to measure the "total quantity / intensity" of public facilities within a certain urban functional area, reflecting the overall richness of service resources in that urban functional area. The specialization coefficient index is used to measure the concentration of a specific type of facility within the urban functional area; specifically as follows: PD q =n q / A q ; EF q,t =(n q,t / n) / (n q / n); Among them, PD q Let n be the facility density of urban functional zone q. q A represents the number of public service facilities POIs within urban functional zone q. q Let q be the area of ​​the urban functional zone; EF q,t Let n be the specialization coefficient of urban functional zone q. q,t Let be the number of public service facility POIs of type t within urban functional area q, and n be the number of different types of public service facility POIs.

[0048] S13. Based on the urban functional area division of each grid area in the learning area obtained in S12, construct a directed adjacency network of public service facility POIs in each urban functional area.

[0049] Specifically, the spatial radius of each public service facility POI within each urban functional area is calculated. A two-dimensional Voronoi diagram is constructed for each type of public service facility POI, thereby forming multiple sets of directional Thiessen polygons. Taking the public service facility POI at the center of each Thiessen polygon as the starting point, other public service facility POIs within its range are retrieved. Public service facility POIs whose distance from the center public service facility POI is less than the aforementioned spatial radius are selected to form multiple sets of adjacency pairs with the center public service facility POI, thereby obtaining spatial adjacency relationships. Thus, a directed adjacency network of public service facility POIs can be constructed.

[0050] In this invention, adjacency pairs can be represented by triples {start_id, end_id, distance}, where start_id, end_id, and distance represent the public service facility POI at the center, the selected public service facility POI, and the distance between them, respectively. Multiple adjacency pairs can form a set of triples, which can be used as the spatial adjacency relationship of each urban functional area.

[0051] In this invention, the directed adjacency network of public service facility POIs is G=(V,E), where node V represents the type of public service facility POI and edge E represents the spatial adjacency relationship between public service facility POIs.

[0052] In this embodiment, the spatial radius of influence d0 of each public service facility POI within each urban functional area is calculated as follows: ; Where A is the area of ​​the urban functional zone, and n is the number of public service facilities (POIs) within the urban functional zone.

[0053] S14. Acquire urban remote sensing images and crop them based on the urban built-up area grid to obtain a remote sensing image dataset. Use local texture coding and location coding to obtain feature sequences containing spatial location information. Input these sequences into the multi-layer sequence modeling module and fuse three feature propagation paths: forward scanning, reverse scanning, and random scanning to extract feature vectors of the urban built environment.

[0054] In this invention, the spatial resolution of the acquired urban remote sensing images must meet the set requirements, such as being less than or equal to a set value. In this embodiment, the set value can be 2m.

[0055] In this invention, urban remote sensing images can be obtained using map software such as Google Maps, Gaode Maps, or Baidu Maps.

[0056] In this invention, the remote sensing image dataset obtained by cropping based on the urban built-up area grid can be cropped using GIS cropping tools.

[0057] In this invention, the local texture encoding is specifically as follows:

[0058] After obtaining a remote sensing image dataset based on urban built-up area grid clipping, the features of each image in the dataset are extracted sequentially. Specifically, each image can be divided into several image blocks of a set size and a set stride S, such as P×P. In this embodiment, P=16, resulting in L image blocks. Each image block is encoded using a two-dimensional convolution with a kernel size consistent with the image block size and a stride of P, thereby obtaining the local texture vector of the image block, realizing the mapping from local texture to vector space. The convolution operation on the image block can be represented as: ; Among them, W δ,c (u,v) are the weight parameters of the δ-th convolutional kernel at the pixel position (u,v) within the image patch, with RGB color channels c∈{R,G,B}. For the pixel value at pixel position (u,v) within the image patch, with RGB color channels c∈{R,G,B}; b δFor the bias term, y δ Let δ be the δ-th component of the output vector obtained by the convolution operation.

[0059] By performing a weighted summation of all RGB color channels and pixel positions in an image patch, linear encoding of the local texture pattern (roof, road, vegetation, etc.) of the image patch can be achieved, i.e., local texture encoding can be completed, thereby enabling the mapping from local texture to vector space.

[0060] In this embodiment, if the step size S=P, the divided image blocks will not overlap; if the step size S<P, the divided image blocks will overlap.

[0061] Since local texture coding divides an image into vector representations corresponding to unordered image blocks, this invention introduces positional coding to preserve the spatial location information of each image block in the original image, as follows: For the h-th image patch, based on its row and column index in the image (u h ,v h Construct its position index p h And for all locations, predefine or learn the location embedding, thus obtaining the location embedding e. pos (p h ).

[0062] The location embedding is added element-wise to the local texture vector to obtain a feature sequence λ containing spatial location information. i =e h +e pos (p h ), where e h This refers to the local texture vector of the image patch obtained from the aforementioned convolution. By stacking the feature sequences of all image patches, a local texture feature sequence x of length L is obtained. t =[λ1,λ2,…,λ L ] T This sequence serves as the input to the subsequent multi-layer sequence modeling module, enabling a structured transformation from a two-dimensional image to a one-dimensional feature sequence.

[0063] In this invention, by using position encoding, the subsequent multi-layer sequence modeling module can not only identify the local texture features of each image block, but also perceive its absolute and relative spatial position relationship in the whole image, which is beneficial for learning structural patterns such as street morphology, road direction, and functional area boundaries.

[0064] In this invention, a sequence modeling module with MambaMixer as the backbone can be constructed based on the state-space model (SSM) to perform multi-level global context modeling on the aforementioned local texture feature sequence containing spatial location information.

[0065] Specifically, the input local texture feature sequence of length L is subjected to layer normalization and linear projection processing, mapping the features to two subspaces u. t ,v t .

[0066] The linear projection process is as follows: ; Among them, u t v t ∈R d′ u t Used to construct dynamic sequence modeling paths, v t Used to provide gated branches to handle dynamic structure updates and gating adjustments in subsequent operations; d′ is the intermediate dimension; W in ∈R 2d′×D For learnable weights, x t The aforementioned local texture feature sequence;

[0067] To reflect the continuous spatial neighborhood relationships between image patches in remote sensing images, u can be used. t By applying one-dimensional convolution, this design can be viewed as constructing local receptive fields at the sequence level, enabling the model built by the multi-layer sequence modeling module to capture the changing trends of image structures such as roads, buildings, and land parcels within a local range. The one-dimensional convolution is represented as follows: ; in, t For u t Applying features enhanced by one-dimensional convolution, W conv (k) represents the weight of the convolution kernel at the relative displacement k, where K is the receptive field radius.

[0068] Features enhanced by convolution t By performing state-space recursive modeling, this invention, through such a design, enables the model to capture long-range structural dependencies across image patches without relying on self-attention, by explicitly maintaining the hidden states.

[0069] The recursive model can be represented as: ; Among them, SSM(⋅) uses a selective scanning strategy to achieve dynamic filtering of sequence directions and modeling of temporal correlations.

[0070] To enhance the model's response in structurally significant regions, the recursive output is activated by a nonlinear function ζ (such as SiLU) and then combined with the feature branch v. t Gating fusion is performed as follows: Yt =ζ(h t )⊙v t ; Here, ⊙ represents element-wise multiplication.

[0071] Through this design, the model built by the multi-layer sequence modeling module can adaptively control the degree of passage of features at different locations, thereby enhancing the model's ability to focus on key structures.

[0072] The representation Y after gating fusion t By restoring the original dimension through linear mapping and establishing residual connections with the input features, stable information transfer across layers is ensured, thereby improving the training efficiency of deep networks. Specifically: x t (l+1) =x t (l) W out Y t ; Where, x t (l) The local texture feature sequence input to the l-th layer is initially the aforementioned local texture feature sequence x. t W out ∈R D×d′ To output the projection matrix, x t (l+1) This is the local texture feature sequence output by the (l+1)th layer; By stacking L Mamba Layer (e.g., L) Mamba= The MambaMixer module (24 layers) described above can enhance the global contextual representation capability of the feature sequences corresponding to image patches in remote sensing images layer by layer, enabling the model to simultaneously capture spatial structure patterns at the plot, block, and district levels.

[0073] To further enhance the model's perception of spatial structure directionality, symmetry, and global invariance, this invention, based on MambaMixer, also designs a structural modeling mechanism with three feature propagation paths: forward, backward, and random scanning. Specifically, for the input sequence of a certain layer, such as x in the l-th layer... t (l) Construct three sequences with different orders: forward scan sequence x tf (l) The reverse scan sequence x is obtained by reversing the forward sequence. tr (l) The random scan sequence x is obtained by randomly rearranging the elements in the forward scan sequence. ts (l)The three sequences share the same set of MambaMixer parameters to ensure the comparability and structural consistency of features extracted under different scanning orders. The three sequences are then fed into the aforementioned MambaMixer module to obtain the corresponding output features H(f), H(r), and H(s), which are then weighted and fused as follows: ; Wherein, α1, β1, and γ1 are weights calculated by a global gating network, and α1,β1,γ1≥0, α1+β1+γ1=1; in this embodiment, they can be calculated using an MLP network and then processed using softmax.

[0074] Global average pooling is performed on the sequence features H after the aforementioned multi-path fusion to obtain the global structural feature vector of the remote sensing image. Based on the similarity retrieval of the feature vector, similar plots can be retrieved and case plots can be preliminarily screened.

[0075] S15. Obtain multi-dimensional urban benefit indicators.

[0076] Specifically, including: Accessibility index: Calculated using cumulative opportunity accessibility, which counts the number of Public Service Points (POIs) accessible to the site within a travel time threshold τ0 (e.g., 15 minutes), as detailed below: ; Among them, G di Q represents the number of Public Service Points (POIs) accessible to plot di. sj For the number of POIs for public service facilities, t ij The travel time from plot di to public service facility POIsj.

[0077] Diversity indicators: characterized by the number and distribution entropy of the types of public service points (POIs) accessible to the land parcel, as detailed below: ; ; Among them, D di p is the distribution entropy of the number of types of public service facility POIs accessible to plot di. t Let n be the proportion of public service facility POI of type t among all public service facility POIs in plot di. t Let t be the number of public service points (POIs) accessible to the plot.

[0078] Vitality index: The vitality value of each public service facility POI is used as the evaluation index; in this embodiment, it can be obtained using map software such as Google Maps, Gaode Maps or Baidu Maps, such as by downloading it through the Baidu Maps Smart Eye platform.

[0079] GDP: Download GDP data for each region from the national data website (or other publicly available data websites) as a macroeconomic support indicator.

[0080] S16. Obtain the interrelationships and spatial adjacency relationships between grids, plots, and public service facility POIs. Combine this with the corresponding attributes of grids, plots, and public service facility POIs obtained in S11-S15 to construct a spatiotemporal knowledge graph within the same spatiotemporal framework, thus obtaining the urban facility layout knowledge base.

[0081] In this invention, the relationships between grids, land parcels, and public service facility (POI) include: the spatial inclusion relationship between public service facility (POI) and grids, the spatial inclusion relationship between public service facility (POI) and land parcels, and the functional affiliation relationship between grids and land parcels.

[0082] In this invention, the spatial adjacency relationship between grids, plots, and public service facility (POI) includes POI spatial adjacency relationship and grid spatial adjacency relationship, which allows the knowledge base to fully consider complex spatial relationships.

[0083] This invention uses a unified spatial grid as a base and combines S11-S15 to construct a spatiotemporal knowledge graph of grid, public service facility POI, land parcel, image, and multi-dimensional attributes.

[0084] Specifically, please refer to Figure 3 The spatiotemporal knowledge graph includes the relationships and spatial adjacencies between the aforementioned grids, land parcels, and public service facility (POI) points of interest (POIs). It also includes the semantic attributes and spatial adjacency of POIs, the functional attributes of land parcels, and the built environment attributes, urban multidimensional benefit attributes, and spatial proximity of grids. Specifically, the semantic attributes of POIs, i.e., their type, can be obtained through reclassification in S11; their spatial adjacency can be obtained through the directed adjacency network in S13. The functional attributes of land parcels can be obtained through the urban functional zoning in S12 combined with their respective grid regions. The built environment attributes of grids represent the characteristics of land feature relationships and can be obtained through the feature vector of the urban built environment in S14. The urban multidimensional benefit attributes of grids, i.e., multidimensional urban benefit indicators, including facility accessibility indicators, facility diversity indicators, vitality value indicators, and GDP, can be obtained through S15. The spatial proximity of grids is directly provided by the adjacency relationships of grids in space; subsequent extraction of adjacent grid combinations is possible. Specific examples can be found in [reference needed]. Figure 4 .

[0085] By analyzing the interrelationships, spatial adjacencies, and corresponding attributes between grids, land parcels, and public service facility POIs, this invention constructs a spatiotemporal knowledge graph corresponding to the urban spatial data of "Grid—Public Service Facility POI—Land Parcel—Image—Multi-dimensional Attributes" based on a unified spatial grid, thus obtaining an urban facility layout knowledge base.

[0086] S2. Divide the target area to be optimized into a grid to obtain the target grid to be optimized, obtain a set of several case sites, and search based on the similarity of the built environment attributes of each grid area, i.e., the feature vector of the urban built environment, according to the urban facility layout knowledge base constructed in S1. Combine this with multi-dimensional urban benefit indicators for screening to obtain a set of reference case sites for the target grid to be optimized.

[0087] In this invention, the cosine similarity calculation method is used to calculate the similarity of feature vectors of urban built environments. Specifically, the similarity sim(f1, f2) of feature vectors of urban built environments between the target grid to be optimized and each set of case locations is calculated, as follows: ; Where f1 and f2 are the feature vectors of the urban built environment of the target grid to be optimized and a set of case locations, respectively.

[0088] In this invention, a similarity threshold is set, which in this embodiment can be set to φ=0.95. The set of cases that meet the similarity threshold in the set of all cases is selected as the set of reference cases.

[0089] In this invention, after selecting the set of case locations that meet the similarity threshold, further selection can be made. Specifically, by comparing the multidimensional urban benefit indicators in S15, the set of case locations whose indicators are all better than the target grid to be optimized can be retained as the final set of reference case locations.

[0090] S3. Extract significant co-occurrence patterns of facilities from the reference case set and use them as recommended facility configuration patterns for the target grid to be optimized.

[0091] The facility co-occurrence pattern is based on the adjacency network of public service facility points of interest (POIs) to measure the strength of co-occurrence edges between public service facility POIs. This invention introduces the structural participation rate P. R As a measure of pattern saliency, to obtain significant co-occurrence patterns of facilities in the reference case set, as follows:

[0092] Let s be a source type node o Pointing to multiple target type nodes t e Based on different functional zones, different distance thresholds are selected to calculate weights, and the nodes are the Points of Interest (POIs) of public service facilities, specifically: ); Where, d r For source type node s o To target type node t e The actual distance, d e θ is the distance threshold (i.e., service radius) for the corresponding grid area; θ is a smoothing parameter used to control the decay rate. When θ>1, the decay is slower, and when θ<1, the decay is faster. The structural participation rate P is calculated after aggregating all source type nodes. R : ; Where the numerator is the sum of the average weights of the source type node connected to multiple target type nodes, and the denominator is... The number of source type nodes; Structural participation rate P R Node groups with a type ≥ξ are selected as facilities with significant co-occurrence patterns, where ξ is the selection threshold, which can be set to 0.6 in this embodiment. That is, if P R If ≥ξ, then the corresponding type node group is regarded as a public service facility group that significantly co-occurs in the actual urban space, thereby obtaining the significant co-occurrence pattern of facilities in the reference case set.

[0093] S4. Construct a facility layout generation model. Based on the feature vector of the urban built environment of the target grid to be optimized obtained in S1, input the feature vector and the recommended facility configuration mode of the target grid to be optimized obtained in S3 into the facility layout generation model. The output result is the facility layout scheme.

[0094] In this invention, the facility layout generation model is as follows: Figure 5 As shown, a GAN model can be used, whose generator employs a structure combining fully connected layers and upsampling layers, and introduces coordinate convolutional layers to enhance spatial location awareness. Its inputs are a random noise vector, a built-up environment feature vector, and a facility knowledge condition vector. The facility knowledge condition vector includes significant co-occurrence patterns of facilities, distance thresholds, and structural participation rates. Its output is the true layout probability, i.e., a multi-channel facility layout probability map, where each channel corresponds to a facility type.

[0095] Continue to refer to Figure 5The discriminator of the facility layout generation model includes a multi-branch discriminator head: a pixel-level discriminator head, a full-image discriminator head, and a structure discriminator head. The pixel-level discriminator head is used to determine the authenticity of the facility layout pixel by pixel; the full-image discriminator head is used to evaluate the global rationality of the overall layout; and the structure discriminator head is used to measure the consistency between the generated layout and the knowledge graph structure in terms of graph edit distance and path preservation rate. The input to the discriminator is a probability map of real or generated facilities, a built environment feature vector, and a facility knowledge condition information vector. Its output includes a true / false probability and a type classification, where the type classification is used to perform semantic recognition of the facility type distribution in the input image.

[0096] In this invention, the loss function of the facility layout generation model adopts a comprehensive loss function to optimize realism, efficiency, and structural consistency, as detailed below: ; Among them, L total Let L be the overall loss function, where α², β², and γ² are the weighting coefficients, respectively. GAN To combat loss, a multi-head structure of the discriminator ensures realism at the pixel, full-image, and structural levels; L align To align the benefit loss, multiple benefit indicators are calculated for the generated results and compared with the target to calculate the difference; L struct For structural consistency loss, constraints are placed on the consistency of the generated results across the graph structure, including graph edit distance, co-occurrence edge coverage, link preservation rate, and distance distribution similarity; L fair To ensure fairness and constrain losses, minimum service coverage of facilities, lower bounds on accessibility for sensitive populations, and land use constraints are required.

[0097] The facility layout generation model of this invention employs adversarial training between its generator and discriminator. It distinguishes between real and fake generated layouts, improving training accuracy and ultimately outputting a realistic layout map. Simultaneously, it can output recommended facility type gaps for completion, such as "Food and Beverage Services - Chinese Restaurant," and further recommend structured facility patterns for completion, such as "Chinese Restaurant - Convenience Store, complete 2 within 100-300m." Combined with the facility layout map, it outputs interpretable results. This facility layout generation model encodes knowledge such as facility co-occurrence patterns and facility configuration pattern recommendations into conditional vectors and structural priors, while ensuring graph structure consistency. Consistency goals such as "graph matching," "path preservation," and "edge count alignment" are defined between the generator output and the target structure induced by the knowledge graph.

[0098] Compared with existing methods, this invention has significant technical advantages and application value. First, it enables deep fusion and unified representation of multi-source heterogeneous data, incorporating multi-dimensional information such as remote sensing imagery, POIs, population, GDP, road networks, and population activity into a spatiotemporal knowledge graph, thereby comprehensively depicting the coupling relationship between "built environment—functional requirements—facility supply." This unified modeling approach overcomes the limitations of fragmented and incomparable data in previous methods, making facility configuration transferable and comparable across regions and cities. Second, through the mining of facility co-occurrence patterns and functional chain analysis, this invention can automatically extract high-frequency facility combination patterns in different functional zones and typical urban environments, forming an interpretable and reusable facility knowledge base, providing knowledge-driven support for subsequent case retrieval and intelligent generation. This knowledge accumulation not only enhances the interpretability of the method but also provides scalability for cross-scenario applications.

[0099] In terms of facility configuration methods, this invention achieves a leap from traditional static optimization to dynamic intelligent configuration. By calculating and benchmarking the built environment characteristics and facility benefit indicators of the target plot, it can quickly retrieve reference cases with high similarity and better benefits, and output corresponding facility completion lists and structural optimization paths. This process significantly improves the scientific nature and efficiency of facility configuration, breaking through the limitations of relying on manual experience and single-step optimization. Simultaneously, in the facility layout generation process, this invention combines knowledge-driven approaches with generative models, introducing prior knowledge such as facility co-occurrence rules, functional chains, and distance thresholds into the Generative Adversarial Network (GAN). This ensures that the generated results not only closely resemble real-world cases in appearance but also logically and systematically conform to the requirements of urban spatial structure, such as... Figure 6 As shown. The final output can present the spatial distribution of public service facilities in the form of a probability graph, covering 10 major types of facilities. The generation process comprehensively considers multiple objectives such as facility accessibility, diversity, and vitality enhancement, ensuring that the generated scheme is both scientifically sound and meets fairness and policy constraints. Figure 7 As shown.

[0100] This invention possesses excellent scalability and practical value, enabling the recognition and intelligent optimization of urban facility configuration patterns. It can be applied to both macro-level regional planning and micro-level community or neighborhood renewal. In practice, this invention can provide government departments and planning agencies with intelligent tools to assist in facility layout, promoting the equalization and accessibility of public service systems.

[0101] Example 1: Planning of public service facilities in new urban areas; This invention can rapidly generate intelligent layout schemes for public service facilities based on the built environment characteristics, population distribution, and functional needs of a new development area. By integrating multi-source data such as remote sensing imagery, POIs, GDP, and population activity, the system can identify facility demand gaps in the new development area and, combined with facility co-occurrence patterns from typical cases, generate logically sound facility distribution schemes. This invention can replace traditional site selection methods that rely on manual experience and static indicators, providing an efficient and scientific reference for facility layout in new development areas.

[0102] Example 2: Renovation and facility improvement in old communities; This invention optimizes facility configuration in older residential communities and existing neighborhoods. By calculating the accessibility, diversity, and vitality indicators of existing community facilities, it identifies facility deficiencies in residents' daily lives. Based on this, the system generates corresponding facility improvement and optimization plans, outputting a list of facility gaps and a probability distribution map. This method is more intelligent and comprehensive than traditional methods relying on questionnaires and manual surveys, providing precise support for community renewal and improving residents' quality of life.

[0103] Example 3: Cross-city facility configuration benchmarking and relocation; This invention enables benchmarking and knowledge transfer of facility configurations between different cities based on spatiotemporal knowledge graphs. By comparing the built environment characteristics and facility benefit indicators of similar plots, the system can find better reference cases for the target city in terms of facility configuration and generate migration plans suitable for local conditions. This function can be applied to the formulation of regional coordinated development and cross-city planning policies, providing data-driven and intelligent support for the balanced development of public service facilities.

[0104] Example 4: Optimization of Emergency Service Facility Layout; This invention can generate optimized layout schemes for facilities such as emergency medical points and shelters based on dynamic population distribution and traffic accessibility information in the context of emergencies or emergency management. By combining facility co-occurrence patterns with travel time thresholds, the system can ensure that critical service facilities cover more people within a specified time, improving the city's emergency response capabilities during emergencies. This application can replace traditional site selection methods based on a single static indicator, providing technical support for urban safety and resilience construction.

[0105] Example 5: Equity Evaluation and Optimization of Public Service Facilities; This invention comprehensively assesses the differences in facility accessibility and service levels among different population groups by integrating indicators such as residents' travel time, facility type diversity, and population activity, and generates equitable optimization schemes based on this. By simulating the effect of different facility layouts on improving equity indicators, the system can provide the government with a scientific policy basis to promote the transformation of urban public service facilities from "sufficient quantity" to "equitable accessibility."

[0106] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of the invention as described above, which are not provided in the details for the sake of brevity.

[0107] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the protection scope of this invention.

Claims

1. A method for generating facility layout schemes based on spatiotemporal knowledge graphs, characterized in that, include: S1. Divide the learning area into grids and obtain the urban functional area division of each grid area. Construct a directed adjacency network of public service facility POIs in each urban functional area. Extract the feature vector of the urban built environment based on this network. Combine with multi-dimensional urban benefit indicators to construct an urban facility layout knowledge base. S2. Divide the target area to be optimized into a grid, obtain a set of several case locations, and based on the urban facility layout knowledge base constructed in S1, retrieve the reference case locations of the target grid to be optimized by searching based on the similarity of the feature vectors of the urban built environment of each grid area. S3. Extract significant co-occurrence patterns of facilities from the set of reference case sites, and use them as recommended facility configuration patterns for the target grid to be optimized; S4. Based on the feature vector of the urban built environment of the target grid to be optimized obtained in S1, input it and the recommended facility configuration mode of the target grid to be optimized obtained in S3 into the constructed facility layout generation model, and the output result is the facility layout scheme.

2. The facility layout scheme generation method according to claim 1, characterized in that, In step S1, the urban functional zone division of each grid area in the learning area is obtained, as follows: Public service facility POI data is acquired, and the public service facility POIs are reclassified based on the main functions of the city. Based on this, spatial corpus of each grid area in the learning area is constructed, and it is input into the word vector generation model to obtain vector representations of various public service facility POIs in each grid area. Based on this, regional vectors of each grid area are constructed and clustered to obtain the urban functional area division of each grid area in the learning area.

3. The facility layout scheme generation method according to claim 2, characterized in that, The spatial corpus for each grid region of the learning area is constructed as follows: Based on the reclassification, within a certain grid area, each public service facility POI is traversed, and its neighboring public service facility POIs and their types are searched within a defined range. The interaction decay between each public service facility POI and its neighbor is calculated, and the POIs are sorted according to the decay order to obtain the corresponding public service facility POI sequence S. i ={t i1 ,t i2 ,…,t ij ,…,t iM }, where t ij The type of the j-th neighboring public service facility POI of the i-th public service facility POI is represented by M, and the number of neighboring public service facility POIs in the i-th public service facility POI is M. At the same time, the interaction attenuation between each public service facility POI and itself is calculated, thereby constructing the spatial corpus of the grid region.

4. The facility layout scheme generation method according to claim 3, characterized in that, The calculation of the interaction attenuation between each public service facility's POI and its interaction is as follows: ； Where d is the distance between two public service facility POIs; σ is a set scale parameter used to control the decay rate.

5. The facility layout scheme generation method according to claim 3, characterized in that, The word vector generation model adopts the Word2Vec model under the Skip-Gram architecture. For each public service facility POI corresponding to a public service facility POI sequence, within a set window width [-b, b], the type of a certain public service facility POI in its corresponding public service facility POI sequence is selected, which is the center word, and the types of other public service facility POIs are the context words. The optimization objective is to maximize the prediction probability of the center word for the context words, as follows: ； Wherein, logP(t) j+z |t k For a given type t of a public service facility POI j In the case of the central term, the type of POI (Point of Interest) of its adjacent public service facilities t j+z That is, the probability of the occurrence of context words.

6. The facility layout scheme generation method according to claim 2, characterized in that, The construction of the region vector for each grid region is as follows: By statistically analyzing the TF-IDF values ​​of each type of public service facility POI within each grid area, and combining this with the vector representation of each type of public service facility POI, a region vector for each grid area is constructed, as follows: R a =(∑ a F at ×w at ) / (∑ t F at ); Among them, R a F is the region vector of the a-th grid region; at Let w be the TF-IDF value of the t-th type of public service facility POI in the a-th grid region; at The vector representation of the t-th type of public service facility POI in the a-th grid region is generated by the word vector generation model.

7. The facility layout scheme generation method according to claim 1, characterized in that, In step S1, the construction of a directed adjacency network of Public Service Facilities Points (POIs) within each urban functional area is as follows: Calculate the spatial radius of each public service facility POI within each urban functional area. Construct a two-dimensional Voronoi diagram for each type of public service facility POI to form multiple sets of directional Thiessen polygons. Starting from the public service facility POI at the center of each Thiessen polygon, retrieve other public service facility POIs within its range. Select public service facility POIs whose distance from the center public service facility POI is less than the aforementioned spatial radius to form multiple sets of adjacency pairs with the center public service facility POI, thereby obtaining spatial adjacency relationships and constructing a directed adjacency network of public service facility POIs.

8. The facility layout scheme generation method according to claim 1, characterized in that, In step S1, the extraction of the feature vector of the urban built environment is specifically as follows: Urban remote sensing images are acquired and cropped based on urban built-up area grids to obtain a remote sensing image dataset. Local texture encoding and location encoding are used to obtain feature sequences containing spatial location information. These sequences are then input into a multi-layer sequence modeling module, and three feature propagation paths—forward scanning, reverse scanning, and random scanning—are fused to extract feature vectors of the urban built environment.

9. The facility layout scheme generation method according to claim 8, characterized in that, The local texture encoding is specifically as follows: After obtaining the remote sensing image dataset based on the grid clipping of the urban built-up area, the features of each image in the remote sensing image dataset are extracted in sequence. The image is divided into several image blocks with a set stride. Each image block is encoded using a two-dimensional convolution with a kernel size consistent with the size of the image block and a set stride, thereby obtaining the local texture vector of the image block. The location encoding is as follows: Construct the position index of the image patch based on its row and column indices in the image, and use it as the position embedding. The location embedding is added element-wise to the local texture vector to obtain the feature sequence containing spatial location information for the image patch. The feature sequences of all image patches are stacked to obtain the feature sequence containing spatial location information.

10. The facility layout scheme generation method according to claim 8, characterized in that, Based on the state-space model, a sequence modeling module with MambaMixer as the backbone is constructed to perform multi-level global context modeling on the local texture feature sequence containing spatial location information, thus obtaining the multi-level sequence modeling module.

11. The facility layout scheme generation method according to claim 10, characterized in that, For the input local texture feature sequence containing spatial location information, a forward scan sequence, a reverse scan sequence, and a random scan sequence are constructed. Layer normalization and linear projection are then performed on these sequences to map the features to two subspaces u. t ,v t and for u t Applying one-dimensional convolution for convolution enhancement, and enhancing the features after convolution enhancement. t Perform state-space recursive modeling, activate the output of the recursive modeling through a nonlinear function, and then combine it with v. t Gated fusion is performed, then the original dimension is restored through linear mapping, and residual connections are established with the input features. This is achieved through stacked L... Mamba The MambaMixer module of the layer processes the data and performs weighted fusion on the three sets of output features.

12. The facility layout scheme generation method according to claim 1, characterized in that, In S1, the multidimensional urban benefit indicators include: Accessibility index: Calculated using cumulative opportunity accessibility, which is the number of public service point of interest (POI) accessible to the land parcel within the travel time threshold τ0; Diversity index: characterized by the number and distribution entropy of the types of public service points (POIs) accessible to the land parcel; Vitality Index: The vitality value of each public service facility's Point of Interest (POI) is used as the evaluation indicator; GDP.

13. The facility layout scheme generation method according to claim 1, characterized in that, In S2, the cosine similarity calculation method is used to calculate the similarity of the feature vectors of the urban built environment.

14. The facility layout scheme generation method according to claim 1, characterized in that, In step S2, after selecting the set of case locations that meet the similarity threshold, the multidimensional urban benefit indicators in step S1 can be combined to retain the set of case locations where all indicators are better than the target grid to be optimized, as the final set of reference case locations.

15. The facility layout scheme generation method according to claim 1, characterized in that, In S3, the extraction of significant co-occurrence patterns of facilities in the reference case set specifically includes: Let s be a source type node o Pointing to multiple target type nodes t e Based on different functional zones, different distance thresholds are selected to calculate weights, and the nodes are the Points of Interest (POIs) of public service facilities, specifically: )； Where, d r For source type node s o To target type node t e The actual distance, d e θ represents the distance threshold for the corresponding grid region; θ is the set smoothing parameter. The structural participation rate P is calculated after aggregating all source type nodes. R : ； Where the numerator is the sum of the average weights of the source type node connected to multiple target type nodes, and the denominator is... The number of source type nodes; Structural participation rate P R The type node group with a value ≥ξ is selected as a significantly co-occurring facility, where ξ is the selection threshold, thereby obtaining the significant co-occurrence patterns of facilities in the reference case set.

16. The facility layout scheme generation method according to claim 1, characterized in that, In step S4, the constructed facility layout generation model adopts a GAN model, and its loss function adopts a comprehensive loss function, as follows: ； Among them, L total Let L be the overall loss function, where α², β², and γ² are the weighting coefficients, respectively. GAN To combat the losses, L align To align the benefits with the losses, L struct For structural consistency loss, L fair To ensure fairness and constrain losses.

Images