Land block level planning scheme generation method based on grid sampling and machine learning

Abstract

The invention relates to the technical field of planning analysis, in particular to a plot-level planning scheme generation method based on grid sampling and machine learning, comprising the following steps: S1, performing initial grid sampling on a target plot to generate an initial sampling unit set; based on the initial sampling unit set, calculating a spatial heterogeneity index of each sampling unit through an identification model; s2, according to the heterogeneity index, adaptive dynamic sampling is carried out on a target land parcel, and an optimized sampling unit set is generated; and S3, inputting the optimized sampling unit set into a planning parameter prediction model, and generating a planning parameter of each optimized sampling unit. According to the method, the common problems of local conflicts, boundary jumping and functional breakage in the planning process are effectively solved, and the performability and coordination of the whole scheme are improved.

Description

Technical Field

[0001] This invention relates to the field of planning analysis technology, and in particular to a method for generating plot-level planning schemes based on grid sampling and machine learning. Background Technology

[0002] With the increasing demand for land spatial governance, integrated urban-rural development, and refined management of natural resources, land parcel-level spatial planning is gradually evolving from macro-indicator control to high-resolution, multi-parameter, and highly operable micro-planning. In order to more scientifically formulate land use layout, development intensity control, and ecological function zoning, it is urgent to construct a planning scheme generation method that is oriented towards the land parcel scale and has spatial intelligent perception and differentiated response capabilities.

[0003] In existing technologies, land use planning data collection mostly employs regular raster or fixed-scale point sampling methods, lacking analysis and adaptation to the spatial heterogeneity within the region. This easily leads to insufficient sampling in structurally complex areas and data redundancy in homogeneous areas, affecting sampling efficiency and hindering the spatial resolution of subsequent parameter predictions. Furthermore, most planning parameter prediction models still rely on independent calculations for each unit, failing to effectively incorporate spatial adjacency relationships, contextual features, and local structural continuity. This results in significant breakage in prediction results at boundary areas and insufficient coordination between regions, making it difficult to meet the practical requirements of "clear zoning, natural transitions, and functional integration." At the level of planning result output, traditional methods typically directly stitch unit prediction values ​​into an area layer, lacking dynamic adjustment and conflict coordination mechanisms. This easily leads to parameter overlap conflicts, uneven ecological development, or abrupt boundary changes, hindering policy implementation and spatial control guidance. Summary of the Invention

[0004] This invention provides a method for generating plot-level planning schemes based on grid sampling and machine learning. The plot-level planning method combines spatial heterogeneity identification, adaptive sampling, graph structure modeling and influence domain propagation fusion mechanism, which can not only improve data collection efficiency and regional expression ability, but also enhance the spatial consistency of parameter prediction and the coordination of planning results, thus meeting the technical requirements of the refined and intelligent development of modern spatial governance.

[0005] A method for generating plot-level planning schemes based on grid sampling and machine learning includes the following steps: S1: Initial grid sampling is performed on the target plot to generate an initial sampling unit set; based on the initial sampling unit set, the spatial heterogeneity index of each sampling unit is calculated through an identification model; wherein, the identification model takes the multidimensional features of the sampling unit as input and outputs a spatial heterogeneity index that characterizes the degree of feature fluctuation within the unit; S2: Based on the heterogeneity index, adaptive dynamic sampling is performed on the target plot to generate an optimized sampling unit set; wherein, adaptive dynamic sampling includes comparing the heterogeneity index with a preset threshold, and performing recursive encryption, local selective encryption or maintaining sparse sampling based on the comparison results. S3: Input the optimized sampling unit set into the planning parameter prediction model to generate planning parameters for each optimized sampling unit; based on the planning parameters, generate a plot-level planning scheme through a dynamic influence domain construction and propagation algorithm; wherein, the dynamic influence domain construction includes: taking each optimized sampling unit as the core, constructing an influence domain according to its planning parameter type and weight, and propagating and superimposing the influence domains of different planning parameters in space through a decay function to form a parameter intensity distribution map; the propagation algorithm iteratively optimizes and adjusts the range and intensity of the influence domain to make the boundary of the influence domain of adjacent units smoothly transition, and finally merges to generate a coordinated plot-level planning scheme.

[0006] Optionally, when performing initial grid sampling on the target plot, an adaptive grid division method based on the plot outline is adopted, and the basic grid size is dynamically adjusted according to the plot boundary curvature and area size to generate an initial sampling unit set covering the entire plot.

[0007] Optionally, the multidimensional features include elevation variation, vegetation cover, soil type, and infrastructure density; based on the initial set of sampling units, the spatial heterogeneity index of each sampling unit is calculated through a pre-trained heterogeneity identification model, specifically including: extracting a multidimensional feature vector for each sampling unit, wherein the multidimensional feature vector includes elevation variation coefficient, vegetation cover distribution variance, soil type diversity index, and infrastructure density.

[0008] Optionally, S1 further includes inputting the multidimensional feature vector into a pre-trained heterogeneity recognition model. The heterogeneity recognition model adopts a deep neural network structure, learns the complex correlation between features through multi-layer nonlinear transformation, and outputs a comprehensive spatial heterogeneity index. The calculation of the spatial heterogeneity index comprehensively considers the spatial distribution uniformity of features within the sampling unit, the fluctuation range of feature values, and the coupling relationship between different feature dimensions. The final spatial heterogeneity quantification value is obtained by weighted fusion of the degree of variation of each feature dimension.

[0009] Optionally, S2 includes partitioning the target land parcel using a multi-level threshold segmentation method based on the heterogeneity index, setting a first threshold and a second threshold, and dividing the land parcel into high heterogeneity regions, medium heterogeneity regions, and low heterogeneity regions.

[0010] Optionally, a quadtree recursive segmentation algorithm is used to refine the local mesh in the highly heterogeneous region, continuously dividing the basic mesh unit into sub-mesh until the heterogeneity index of the sub-mesh is lower than the first threshold or reaches the preset maximum segmentation level. In the heterogeneous region, a selective encryption strategy is adopted, and mesh refinement is performed only at a limited number of levels at the boundary of characteristic mutations; The initial sampling density is maintained in the low heterogeneity region, and a sparse sampling strategy is adopted.

[0011] Optionally, the adaptive dynamic sampling also includes a sampling quality feedback mechanism, which recalculates the heterogeneity index of each sampling unit after each round of grid encryption, dynamically adjusts the boundary and density of the encrypted region based on the recalculated heterogeneity index, and generates a spatially optimized sampling unit set through an iterative optimization process. The generation of the optimized sampling unit set also includes unit fusion, which involves merging multiple adjacent low heterogeneous sampling units into a representative sampling unit when their spatial feature similarity is higher than a first similarity threshold.

[0012] Optionally, the planning parameter prediction model is a planning parameter prediction model based on graph neural networks. The optimized sampling unit set is input into the planning parameter prediction model based on graph neural networks. The model treats each optimized sampling unit as a graph node, and the spatial adjacency relationship and feature similarity between units as edge connections. The neighborhood information is aggregated through a multi-layer message passing mechanism to generate planning parameters for each optimized sampling unit, including land use type, building density, plot ratio and greening rate.

[0013] Optionally, generating a plot-level planning scheme based on the planning parameters using a dynamic influence domain construction and propagation algorithm specifically includes: Taking each optimized sampling unit as the core, weights are assigned according to the type and importance of its planning parameters to construct an initial influence domain, in which the influence domain of ecological protection parameters is larger than that of development and construction parameters; An adaptive decay function is used to spatially propagate the influence domain of different planning parameters. The Gaussian kernel bandwidth of the decay function is dynamically adjusted according to the parameter type and the local heterogeneity index. In regions with high heterogeneity, a small bandwidth is used to preserve detailed features.

[0014] Optionally, the step of iteratively optimizing and adjusting the scope and intensity of the influence domain specifically includes: Calculate the gradient change at the boundary of adjacent influence domains, detect parameter conflict regions, and reallocate influence domain weights based on parameter priority rules; In each iteration, the overlapping influence domains are superimposed in intensity and conflict is resolved, with the influence domain of high-priority parameters having a dominant position in the conflict region; The propagation algorithm continues to run until the rate of change of the boundary of the adjacent influence domain is lower than the convergence threshold. Finally, a coordinated plot-level planning scheme is generated by spatial fusion of the parameter intensity distribution map. The scheme maintains the continuous spatial transition of parameters and functional coordination.

[0015] The beneficial effects of this invention are: This invention constructs a spatial heterogeneity index model based on multidimensional features of elevation variation, vegetation cover, soil type, and infrastructure density, enabling the quantification of the internal feature complexity of each sampling unit. Furthermore, it employs multi-level threshold partitioning, quadtree recursion, and boundary gradient awareness to locally densify high-heterogeneity regions, selectively refine at abrupt change boundaries, and maintain sparse sampling in low-heterogeneity regions. The grid structure is dynamically optimized by combining sampling quality feedback and feature similarity fusion mechanisms. This mechanism concentrates limited sampling resources on spatially complex regions, significantly reducing redundant data collection, enhancing local feature parsing capabilities, providing high-quality input for subsequent planning models, and improving sampling efficiency and expression accuracy.

[0016] This invention constructs optimized sampling units into a spatial topology graph, establishes edge weights based on spatial adjacency relationships and feature similarities between units, and integrates neighborhood information through a graph neural network structure in multi-layer message passing to output multi-dimensional planning parameters including land use type, building density, plot ratio, and greening rate. Compared to traditional prediction methods that use units as independent inputs, this invention enhances the model's understanding of spatial continuity, boundary transitions, and regional relationships, making it particularly suitable for site scenarios with heterogeneous structures and ambiguous regional transitions, and possessing stronger parameter inference capabilities and prediction consistency.

[0017] This invention proposes an influence domain construction mechanism based on the joint regulation of planning parameter type weights and spatial heterogeneity indices. It achieves adaptive diffusion of parameters in space through a Gaussian decay function and introduces conflict detection, priority decision-making, and multi-round iterative optimization processes to dynamically adjust the influence domain boundaries and intensity. Finally, all parameter influence maps are merged into a unified parameter intensity distribution map, outputting a plot-level planning scheme that maintains a balance between ecological and development control, natural functional zoning boundaries, and continuous spatial transitions. This mechanism effectively solves common problems in planning processes such as local conflicts, boundary jumps, and functional breaks, improving the overall feasibility and coordination of the plan. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the solution generation process according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the adaptive dynamic sampling method according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] like Figures 1-2 As shown, a method for generating plot-level planning schemes based on grid sampling and machine learning includes the following steps: S1: Initial grid sampling is performed on the target plot to generate an initial sampling unit set; based on the initial sampling unit set, the spatial heterogeneity index of each sampling unit is calculated through an identification model; wherein, the identification model takes the multidimensional features of the sampling unit as input and outputs a spatial heterogeneity index that characterizes the degree of feature fluctuation within the unit.

[0022] S11, Construction of initial grid sampling units: for the target plot Initial grid sampling is performed using an adaptive grid generation method based on land parcel contours, specifically including: Based on the curvature of the land parcel boundary With land area Dynamically adjust the base grid size Generate an initial set of sampling units covering the entire plot. ; The mesh size adjustment function is expressed as: ; in, Indicates the size of the basic mesh cell. This represents the boundary curvature adjustment coefficient. Indicates the location of the land parcel boundary. curvature at that point This represents the area adjustment coefficient. The first item represents the total area of ​​the land parcel. This indicates that the mesh density is adjusted based on the local curvature of the boundary. A larger curvature indicates a more complex boundary, and therefore the mesh should be smaller. The second term... This indicates that the area is being adjusted; the smaller the area, the better. The larger the grid size, the smaller the grid size, which facilitates fine sampling.

[0023] Parcel boundary curvature The following is how to obtain it: 1. Boundary extraction: First, boundary extraction is performed using parcel vector boundary data or remote sensing imagery to generate closed boundary curves; 2. Boundary Discretization: Discretize the boundary curve into a set of points with a fixed step size. ; 3. Curvature estimation: for arbitrary boundary points Calculate the change in tangent direction at its immediate and adjacent points, and estimate the curvature based on the three-point method, using the following formula: ;in, It is the angle between adjacent line segments, which can be obtained using the cross product of vectors.

[0024] Boundary curvature adjustment coefficient Used to measure the impact of boundary curvature on mesh size scaling. Typical values ​​range from 0.5 to 1.5. This indicates linear adjustment of curvature. For scenarios with drastic boundary changes (such as complex plots), a value of 1.2-1.5 can be used; for regular plots, it can be appropriately reduced to 0.5-0.8.

[0025] Area adjustment coefficient This is used to avoid generating excessively small grids in large plots of land; the typical range is 0.2-0.6. The larger the plot, the larger the cell grid tends to be. This can be used as an empirical default value, and can be appropriately reduced to 0.2 for micro-plots or slope edge treatment.

[0026] S12, Calculation of the spatial heterogeneity index: For the initial set of sampled cells Each sampling unit in , Perform the following processing: S121, Multidimensional Feature Vector Extraction: Extracting each sampling unit Multidimensional feature vectors Specifically, it includes: Elevation variation coefficient; obtained from digital elevation model (DEM) data in each sampling unit. Extract all elevation pixel values ​​from the DEM raster and calculate the ratio of its standard deviation to its mean, i.e., the coefficient of variation. The calculation steps are as follows: crop out the elevation pixel values ​​from the DEM raster. Calculate the standard deviation of elevation data for the region. with the mean , then use The results were obtained; : Variance of vegetation cover (NDVI) distribution; Normalized Difference Vegetation Index (NDVI) is calculated from remote sensing imagery in the sampling unit. Extract the NDVI values ​​of all pixels within the region and calculate their variance. Calculate NDVI = (NIR - Red) / (NIR + Red) using the red and near-infrared bands from the multispectral remote sensing data, and then calculate the variance of the NDVI pixel values ​​for that region. Soil type diversity index (Shannon entropy); the proportion of each soil type within a sampling unit is obtained through soil type distribution maps or databases (soil survey data), and the diversity is measured using the Shannon entropy formula. That is, statistical... Area ratio of different soil types Calculate the entropy value: A higher value indicates a greater diversity of soil types. Infrastructure density (number of facilities per unit area); This involves extracting vector data of infrastructure such as roads, buildings, and waterways from remote sensing imagery or GIS vector layers, and then performing this data in each sampling unit. The number or total length of facilities within the area is statistically analyzed and normalized to a density per unit area. The degree of construction intervention in the area can be calculated as total number of facilities / area or total facility length / area.

[0027] S122, Heterogeneity Identification Model Inference: Inferring Feature Vectors Input to the pre-trained heterogeneity recognition model This model is a deep neural network structure with multiple layers of nonlinear activation functions and feature fusion modules, and outputs a spatial heterogeneity index. ,Right now: .

[0028] The heterogeneity identification model is a deep learning model used to identify the degree of local heterogeneity from multidimensional parcel features. It adopts a deep neural network structure and includes the following components: Input layer: Receives the multidimensional feature vector fi extracted from each sampling unit, including elevation variation coefficient, NDVI distribution variance, soil diversity index, and infrastructure density.

[0029] Multi-layer hidden layers: These consist of several fully connected layers with activation functions. Each layer extracts and combines implicit relationships between different features through non-linear mapping. These layers can capture: Interactions between different features; Complex spatial patterns of characteristic fluctuations; Higher-order coupling structures (such as the patterns of joint occurrence of geomorphological changes and soil diversity).

[0030] Output layer: The final output is a scalar value Hi, which represents the spatial heterogeneity index of the sampling unit ci.

[0031] Furthermore, the heterogeneity index can be defined as follows: ;in, Representation of features In the sampling unit The degree of spatial variation within the sample unit is used to quantify the sampling unit. The degree of spatial non-uniformity of each feature dimension, reflecting the discreteness of each feature dimension, can be expressed as: ;in, This represents the standard deviation of the corresponding feature within the sampling unit. The average value of the corresponding features. To prevent small constants from being divided by zero, this process can be understood as: measuring whether each feature fluctuates drastically within the sampling unit. This represents a multi-feature coupling metric function, using principal component nonorthogonality to measure the correlation between features. It is used to assess whether there is significant correlation or cross-influence among four features within the same sampling unit, and is a cross-dimensional complexity measure. Specifically: in this unit... Within each sub-region, a corresponding four-dimensional feature vector (one four-dimensional vector per pixel) is extracted. Principal component analysis is then performed on the feature matrices of all these samples to extract the direction vectors of the first two principal components. Calculate the inner product of these two vectors as an index of nonorthogonality: If the angle between the directions of the principal components is close to 90°, it indicates that the features in each dimension are highly independent and have low coupling. A small value, especially if the included angle approaches 0° or 180°, indicates that the features are highly collinear and have strong coupling. A large value, in other words, indicates whether there is a high degree of coupling between topographic relief, vegetation distribution, soil diversity, and facility density. Indicates the first The weights of each feature dimension satisfy... , This represents the coupling degree adjustment coefficient.

[0032] Finally, the output contains a set of heterogeneity indices corresponding to each sampling unit: .

[0033] S2: Based on the heterogeneity index, adaptive dynamic sampling is performed on the target plot to generate an optimized sampling unit set; wherein, adaptive dynamic sampling includes comparing the heterogeneity index with a preset threshold, and performing recursive encryption, local selective encryption or maintaining sparse sampling based on the comparison results.

[0034] S2 calculates the set of spatial heterogeneity indices. Based on this, the target plot is divided into heterogeneous regions and subjected to adaptive dynamic sampling, specifically including the following: S21, Multi-level threshold segmentation and heterogeneous region partitioning: Set two heterogeneity index thresholds: First threshold Threshold for determining high heterogeneity; Second threshold Low heterogeneity threshold, satisfying .

[0035] For each sampling unit According to its spatial heterogeneity index The possible values ​​of are categorized as follows: .

[0036] The spatial heterogeneity index is obtained by weighting the standardized feature variability with the coupling degree of multiple features, and its value range is usually controlled within [0,1]. The threshold setting is based on the quantile principle, which sets the quantile points based on the statistical distribution: First threshold (high heterogeneity boundary): can be set as the 75th percentile of the heterogeneity index; The second threshold (low heterogeneity boundary) can be set to the 25th percentile.

[0037] A fixed threshold setting can also be used, which is suitable for large-scale processing or engineering deployment. A value of 0.6 indicates that regions with a heterogeneity index exceeding 0.6 are considered to have complex structures or drastic variations. A value of 0.3 indicates that regions with a heterogeneity index below 0.3 are structurally homogeneous and have stable characteristics.

[0038] S22, Adaptive mesh encryption strategy within the region: S221, High Heterogeneity Region - Quadtree Recursive Encryption: As mentioned above, by calculating the spatial heterogeneity index, the entire target plot has been divided into three types of heterogeneous regions: high, medium, and low. For high heterogeneous regions, a denser and more flexible grid sampling strategy is needed to ensure the capture of the complexity of the local spatial structure. Therefore, for all regions satisfying... The sampling units are refined using a quadtree recursive partitioning algorithm, and the initial grid units in highly heterogeneous regions are progressively refined layer by layer, as follows: 1. The initial grid cell is the root node, and each highly heterogeneous sampling cell... It is considered the root node of a quadtree; 2. If the current subgrid heterogeneity index And the current segmentation level If the area still has subdivision value, it should be further divided into four subgrids. Therefore, it is further divided into four subgrids. A subgrid is a smaller grid unit obtained by recursively dividing the high heterogeneity sampling unit through quadtree. The subgrid is a small unit generated by dividing the sampling unit in order to capture spatial changes more finely in the high heterogeneity sampling unit. It is a substructure of the original sampling unit. 3. If the heterogeneity index of the current subgrid satisfies or If the local features have become uniform and further refinement is no longer necessary, then the segmentation should be stopped.

[0039] 4. The recursion termination condition is: or ;in, The heterogeneity index of the current subgrid. For the current segmentation level, To preset the maximum segmentation level, this segmentation process is repeated for each sub-mesh that meets the conditions for continued segmentation until all sub-mesh meet the stopping conditions, forming a set of structurally adaptive non-uniform density sampling units.

[0040] A quadtree is a hierarchical spatial partitioning structure suitable for recursive subdivision of two-dimensional regions. It allows for localized densification in highly heterogeneous regions while maintaining coarse sampling in less heterogeneous regions, avoiding redundant computation and data burden caused by overly dense sampling across the entire region. For example, fine sampling can be performed in areas with complex terrain or interspersed vegetation, while automatically stopping partitioning in flat or contiguous farmland areas. The heterogeneity index comprehensively reflects the complexity of multiple factors such as local topography, vegetation, soil, and human activities. As a condition for further subdivision, it ensures that "only truly complex regions are encrypted," effectively improving sample representativeness and model training quality. This achieves adaptive termination based on "complexity convergence". Without setting a level cap, some regions may be over-segmented, especially highly heterogeneous regions caused by error noise. Setting a reasonable level cap can mitigate this. It can achieve a balance between accuracy and efficiency, preventing system performance bottlenecks.

[0041] S222, Heterogeneous Regions - Selective Encryption: For those satisfying Within the sampling unit, the overall characteristics of the area do not fluctuate significantly, but there may be local abrupt boundary changes, such as sudden changes in soil type, NDVI changes from woodland to bare land, or roads crossing farmland. Therefore, instead of uniformly refining the entire sampling unit, a selective refinement approach is adopted, with local refining only at abrupt boundary locations. Specifically, for each spatial location within the sampling unit... Calculate the features of each dimension Spatial gradient This is used to determine whether a point is at a mutation boundary. The mutation boundary region is defined as: ;in, Indicates the first Features at location Spatial gradient, For the first The mutation threshold of each feature, if Explanation of the first A feature undergoes a significant abrupt change at a certain location, which is marked as the abrupt change boundary region. Several sub-mesh cells are generated around the boundary region. Local refinement is performed only in these regions at a limited level (1–2 layers) to refine the sampling density, while the original mesh remains unchanged in the remaining regions. This results in a set of medium-heterogeneous region mesh partitioning results with dense distribution at the abrupt change boundary and sparse coverage in the stable region, in order to reduce redundant computation.

[0042] While the overall characteristics of moderately heterogeneous regions do not change drastically, they often contain structural turning points or boundary transition zones. Unified densification would lead to a large amount of worthless redundant data, while ignoring boundary details might miss crucial planning change information. Therefore, the design approach is: high-resolution local modeling + low-complexity maintenance across the entire region, i.e., local densification + overall restraint. Since moderately heterogeneous regions themselves are not complex, only the boundaries require focused processing. A limit of 1-2 layers of densification is set to preserve spatial resolution while controlling the number of samples. (Sudden change threshold) It is the threshold value for judging whether a feature has changed significantly at a certain location. It can use the historical distribution of the spatial gradient of each feature and set the 90th percentile as the mutation threshold, which means that only the first 10% of the regions with the most dramatic changes are judged as mutations.

[0043] S223, Low heterogeneity region - sparse sampling: For The sampling units indicate that the spatial feature structure within the region is relatively simple and uniform, with no significant interaction or spatial abrupt changes between features. Such regions are classified as low heterogeneity regions. Maintaining the initial sampling density unchanged and not performing densification operations is suitable for regions with relatively uniform regional features and slow changes. Low heterogeneity regions have highly consistent spatial structures, with relatively consistent landforms, vegetation, soil, or human intervention factors. Local feature fluctuations are very small, and there is a lack of strong coupling relationships between features. The overall spatial pattern is stable, and increasing the sampling density has limited help in improving accuracy. The initial grid design has been adaptively generated based on the plot boundaries and scale, and has strong spatial coverage rationality.

[0044] S23, after each round of mesh refinement, recalculate the heterogeneity index of all newly generated sampling cells. This constitutes a new set of heterogeneity indices. According to the new The encryption range and density of each region are dynamically adjusted to achieve self-optimization of sampling quality and adaptive evolution of region boundaries. The iterative process is repeated until one of the following convergence conditions is met: or reaching the maximum number of iterations. This formula is the termination condition in the adaptive dynamic sampling process, used to control whether the iterative encryption process converges or ends. The expression consists of two parts; satisfying either one stops the iteration. Heterogeneity index of all sampling units in the next iteration Compared to the previous iteration The results were compared one by one, and the maximum change in the heterogeneity index of all units was calculated. If this maximum change value is less than a very small threshold This indicates that the heterogeneous distribution has stabilized and no longer changes significantly, meaning the sampling results have converged, and the iteration can be terminated. If the number of iterations has reached the preset maximum value... Even if convergence is not achieved, the process is forcibly terminated to prevent it from getting stuck in infinite iteration or resource exhaustion. The maximum number of iterations, For the current iteration round, This represents the sampling stability threshold, which controls the maximum difference in the heterogeneity index allowed between two consecutive sampling rounds. The smaller the value, the stricter the requirement, which means that the sampling results should hardly change before stopping. The larger the value, the more lenient the requirement, which means that the sampling may stop earlier. With the heterogeneity index standardized to the range of [0,1], the recommended range is [0.01,0.05], and a specific value of 0.02 is acceptable, which means that a maximum of 2% heterogeneity fluctuation difference is allowed in each round as a criterion for judging whether the sample is approaching stability.

[0045] S24. In regions of low heterogeneity, the spatial features of multiple adjacent sampling units are often extremely similar. Continuing to retain them independently would cause data redundancy. Therefore, a feature similarity threshold is set to identify which sampling units are close enough to be merged. Specifically: for multiple continuously distributed low heterogeneous sampling units... If the similarity between their spatial feature vectors satisfies the following condition: Then they can be merged into a representative sampling unit. Where Sim is the feature similarity function. The first similarity threshold is 0.95. The eigenvector of the merged representative unit is a weighted average of the merged regions. This mechanism helps reduce redundant sampling in low-heterogeneity regions, improving sampling efficiency and system response speed.

[0046] S25 collects the refined subgrids, the original unchanged sampling units, and the merged representative units to form the final optimized sampling unit set.

[0047] S3: Input the optimized sampling unit set into the planning parameter prediction model to generate planning parameters for each optimized sampling unit; based on the planning parameters, generate a plot-level planning scheme through a dynamic influence domain construction and propagation algorithm; wherein, the dynamic influence domain construction includes: taking each optimized sampling unit as the core, constructing an influence domain according to its planning parameter type and weight, and propagating and superimposing the influence domains of different planning parameters in space through a decay function to form a parameter intensity distribution map; the propagation algorithm iteratively optimizes and adjusts the range and intensity of the influence domain to make the boundary of the influence domain of adjacent units smoothly transition, and finally merges to generate a coordinated plot-level planning scheme.

[0048] S31, Graph Neural Network Programming Parameter Prediction Model: The set of optimized sampling units is denoted as... Constructing a graph neural network with input graph ; where each optimized sampling unit Mapped to a node in the graph Each pair of adjacent sampling units If a spatial adjacency exists or the feature similarity is higher than the second similarity threshold, Then establish an edge Define the edge weights as follows: ; in, This indicates that the two units are spatially adjacent; otherwise, it is 0. Represents cosine similarity. Represents the weighting coefficients that control spatial relationships and feature similarity. Indicates the first Feature vector of each sampling unit Control whether to establish an edge between two sampling units; if the feature similarity between the two units is higher than 100%. This means they are considered related and can transmit information to each other, with a value range of [0.80, 0.95]. Smaller values ​​are preferred. Allowing for more edge connections, denser information propagation, and larger [scale / capacity] It places greater emphasis on highly similar connections, resulting in a sparse yet semantically accurate network.

[0049] Graph neural networks employ a multi-layer message passing mechanism, with each layer's node embedding updated as follows: ; in, Indicates the first Layer nodes Embedded representation, Indicates the first The weight matrix of the layer, This represents the activation function. For nodes The set of neighbors.

[0050] The final output is the prediction planning parameters for each sampling unit: ;in, This indicates land use type, including agriculture, construction, and green space. Indicates building density, Indicates floor area ratio, This indicates the greening rate.

[0051] The goal of S31 is to predict the land use planning parameters of each unit, such as land use type, building density, plot ratio, and greening rate, based on the optimized sampling units. To this end, a graph neural network model is adopted, which utilizes the spatial relationships and feature similarities between sampling units to construct a graph structure and perform multi-level message passing learning.

[0052] The main logic is as follows: 1. Treat each optimized sampling unit as a node in the graph and construct the graph structure. If two sampling units are spatially adjacent or their feature vectors are sufficiently similar, establish an edge between them. The weight of each edge is a weighted combination of spatial adjacency value and feature similarity. The larger the weight, the stronger the connection between the two units.

[0053] 2. Graph neural networks enable each node (sampling unit) to learn information from its neighboring nodes through a "message passing" mechanism. In each layer of the network, the sampling unit will weight and summarize the feature representations of its neighbors and update its own embedding vector. After multiple layers of transmission, the embedding vector of each sampling unit will integrate information from its neighborhood and reflect the spatial context features of the region.

[0054] 3. Finally, the embedding vector of each sampling unit is mapped to a specific planning parameter vector, which includes: Land use type (LU): indicates the classification of land use; Building density (BD): reflects development intensity; Floor Area Ratio (FAR): The ratio of total building area to ground area; Green coverage rate (GR): indicates the degree of ecological control.

[0055] Site planning parameters typically exhibit spatial continuity and regional correlation. Generative Neural Networks (GNNs) are well-suited for modeling such local correlations. Sampling units contain multi-dimensional features such as topography, vegetation, soil, and infrastructure, which interact in complex ways. GNNs can also automatically learn these relationships. For units lacking local information, GNNs can supplement information from neighboring nodes, thereby improving prediction accuracy and regional consistency.

[0056] S32, Dynamic Influence Domain Construction and Propagation Algorithm: S321, Influence Domain Construction: After predicting the planning parameters for each sampling unit, a complete plot-level planning scheme needs to be generated. The key is that the planning parameters of each sampling unit not only apply to itself but also influence its surrounding area. This influence gradually decays spatially and overlaps and coordinates with the influence of surrounding units. To achieve this influence propagation mechanism, an influence domain is constructed for each type of planning parameter k (including land use type LU, building density BD, floor area ratio FAR, and greening rate GR) for each sampling unit. The influence domain is defined as follows: at a certain location p, the effect of the parameters of the element on it is greater than a certain minimum influence threshold. The specific plan for the area is as follows: With each sampling unit With this as the core, based on its planning parameter type and importance coefficient Construct the parameter influence domain ,Right now: ;in, , To minimize the impact intensity threshold, greater impact weights are assigned to ecological protection parameters (greening rate, farmland conservation area). This results in a wider range of influence; however, the range of influence is relatively limited for development and construction parameters (building density, plot ratio). This determines how far the planning parameters of a given sampling unit propagate spatially. If the influence at a point is lower than... If the value is less than 0.05, it is considered that the point no longer belongs to the influence domain of the parameter. The value range is [0.05, 0.15], which means that 5%–15% of the influence intensity is used as the lower limit of the effective influence. If the value is too high, the influence domain may be too small, resulting in breaks and discontinuous boundaries between the influence areas. If the value is too low, the influence domain may be too large, resulting in redundant overlap and weakening the clarity of the boundaries between parameters.

[0057] The logical structure of the influence domain includes: Central sampling unit: Each sampling unit is an influence source, and its planning parameters are propagated to the surrounding area through spatial diffusion.

[0058] Parameter type weighting: Different planning parameters have different levels of influence and importance. Ecological parameters (greening rate) have high requirements for regional control, therefore their weighting is relatively high. If the weight is set too high, the influence range is wide; while development parameters (building density, plot ratio) mainly affect local areas, have a smaller weight, and have a limited scope of influence.

[0059] Influence is typically measured using a Gaussian decay function, gradually decreasing outwards from the center of the sampling unit. Influence domain Only the spatial range with an effective value greater than the threshold is retained to ensure that the propagation does not spread indefinitely.

[0060] S322, Attenuation Function Propagation Mechanism: The influence of each planning parameter diffuses in space with the parameter center as the kernel. Specifically, for a certain sampling unit... One type of its planning parameter will affect the surrounding area, in order to Centered on spatial location points As a variable, calculate the parameter at the point The intensity of the influence at a given point, expressed using a Gaussian decay kernel function, is as follows: This means: away from The closer, The smaller the value, the closer the influence value is to the maximum value. The farther away the distance, the more rapidly the influence value decreases with the exponential function, and the curve exhibits a bell-shaped distribution. The Gaussian kernel determines the width of the influence range. This represents any point within the plot. This represents the Euclidean distance from the point to the center of the sampling unit. This indicates the influence domain bandwidth parameter. The bandwidth determines how far the parameters of a sampling unit can propagate. A larger bandwidth results in a wider range of influence and a farther reach, while a smaller bandwidth results in a narrower range of influence, limiting its effect to its nearest neighbors. This scheme dynamically adjusts the bandwidth based on the spatial heterogeneity index of the unit, expressed as: ;in, Indicates the initial bandwidth. Indicates heterogeneity sensitivity factor, Indicates sampling unit Spatial heterogeneity index; in regions of high heterogeneity, bandwidth... The size is smaller to preserve spatial resolution; in stable regions, the range of influence is appropriately expanded.

[0061] That is to say: In regions with high spatial heterogeneity ( (Large) indicates that the local features are complex and drastic, so the detail resolution should be maintained and the bandwidth should be reduced to avoid excessive diffusion; In regions with low heterogeneity ( (Small) indicates a stable structure and consistent characteristics → bandwidth can be increased, promoting a smooth transition of planning parameters.

[0062] The purpose of S322 is to spatially diffuse the planning parameters of each sampling unit to influence the surrounding area, thereby forming a continuous and coordinated parameter distribution map across the entire plot. It's not simply about attaching the parameters of each unit to a fixed location, but rather simulating its influence on the neighborhood; the closer to the unit, the stronger the influence; the further away, the less the influence. Therefore, a Gaussian decay function is used to simulate this outward diffusion process.

[0063] S323, Iterative Optimization and Conflict Resolution: As mentioned above, each sampling unit has generated an influence domain based on its own planning parameters and spatial heterogeneity. Since the influence domains of different units may overlap spatially, conflicts may arise in some areas where multiple parameters "act simultaneously" but in inconsistent directions. Here, an iterative optimization mechanism is used to resolve conflicts between influence domains, ensuring that the final plot-level planning scheme has spatial coordination and functional continuity. Specifically, this includes optimizing all influence domains... Perform iterative propagation and boundary adjustment process: Calculate the intensity gradient at the boundary of the influence domain of adjacent sampling units, i.e., for each pair of adjacent sampling units. In the region where their influence domains overlap, compare the difference in the influence intensity of their respective parameters at that location: Where p and q are the corresponding positions of the two units on the boundary; if the gradient is greater than the set conflict threshold, it is considered a parameter conflict region. The conflict threshold determines whether the difference between two influence regions at a certain position is large enough to be considered a conflict. The value is 0.15. If it is less than 0.1, it is too sensitive and will misjudge the normal transition region as a conflict; if it is greater than 0.25, the real conflict region will be missed.

[0064] Within a conflict area, the impact of each parameter cannot be simply averaged. Instead, weight adjustments are made according to preset parameter priority rules. For example, if ecological parameters (greening rate) have higher priority than construction parameters (building density, plot ratio), then within the conflict area, the impact of ecological parameters is prioritized. The specific adjustment method is as follows: at each conflict point... The parameter with the strongest weighted influence among all participating parameters is retained as the dominant influence. That is, based on the preset parameter priority rules (ecological priority > construction priority), the influence weights of conflict zones are redistributed to ensure that high-priority planning parameters have a dominant position. ; After completing a conflict adjustment, the spatial distribution of each influence domain is recalculated. If the boundary of the influence domain changes very little between two consecutive rounds, it is considered to have reached stability. The criterion is that the maximum rate of change of the boundary of all adjacent units between the current round and the previous round is lower than the convergence threshold, and the iteration can be stopped. If the condition is not met, the conflict detection and adjustment will continue to be performed iteratively until the termination condition is met or the maximum number of iterations is reached.

[0065] S33, integrates the superposition results of the influence domains of all sampling unit parameters to generate a parameter intensity distribution map within the plot, which is the final spatial planning result for the entire plot. The specific execution method is as follows: Parameter overlay: For any point within the plot Traverse all sampling units The influence values ​​of various planning parameters (LU, BD, FAR, GR) at that point are summed up, i.e.: This indicates that the intensity of the planning parameters at each location is the result of the combined influence of multiple sampling units and multiple parameters.

[0066] Spatial functional labeling: After obtaining the complete parametric intensity distribution map, based on the maximum value decision principle, labeling is performed for each location point. Assign the most suitable land use or planning type. For example, if the influence of a certain point on the greening rate is much stronger than other parameters, then that point may be classified as a green space or ecological protection area.

[0067] Regional division: Based on the continuity of intensity distribution, the land parcel space is clustered or thresholded to form contiguous functional areas (residential land, commercial land, agricultural land, etc.), and output as layered data.

[0068] Because Gaussian diffusion, boundary conflict decoupling, and iterative fusion mechanisms were used in the early stages, the output scheme has the following characteristics: The spatial connections between the parameters are continuous (without abrupt changes); The boundaries between different functional areas are clear and the transitions are smooth; The conflict areas have been resolved through a priority control mechanism.

[0069] Based on the location of the maximum parameter intensity, spatial functions are labeled and regions are divided, and a plot-level planning scheme layer that maintains spatial continuity, functional coordination, and conflict optimization is output.

[0070] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0071] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating land parcel-level planning schemes based on grid sampling and machine learning, characterized in that, Includes the following steps: S1: Perform initial grid sampling on the target plot to generate an initial set of sampling cells; Based on the initial set of sampling units, the spatial heterogeneity index of each sampling unit is calculated through an identification model; wherein, the identification model takes the multidimensional features of the sampling unit as input and outputs a spatial heterogeneity index that characterizes the degree of feature fluctuation within the unit; S2: Based on the heterogeneity index, adaptive dynamic sampling is performed on the target plot to generate an optimized sampling unit set; wherein, adaptive dynamic sampling includes comparing the heterogeneity index with a preset threshold, and performing recursive encryption, local selective encryption or maintaining sparse sampling based on the comparison results. S3: Input the optimized sampling unit set into the planning parameter prediction model to generate planning parameters for each optimized sampling unit; based on the planning parameters, generate a plot-level planning scheme through a dynamic influence domain construction and propagation algorithm; wherein, the dynamic influence domain construction includes: taking each optimized sampling unit as the core, constructing an influence domain according to its planning parameter type and weight, and propagating and superimposing the influence domains of different planning parameters in space through a decay function to form a parameter intensity distribution map; the propagation algorithm iteratively optimizes and adjusts the range and intensity of the influence domain to make the boundary of the influence domain of adjacent units smoothly transition, and finally merges to generate a coordinated plot-level planning scheme.

2. The method for generating plot-level planning schemes based on grid sampling and machine learning according to claim 1, characterized in that, When performing initial grid sampling on the target plot, an adaptive grid division method based on the plot outline is adopted. The basic grid size is dynamically adjusted according to the plot boundary curvature and area size to generate an initial sampling unit set covering the entire plot.

3. The method for generating plot-level planning schemes based on grid sampling and machine learning according to claim 1, characterized in that, The multidimensional features include elevation variation, vegetation cover, soil type, and infrastructure density; Based on the initial set of sampling units, the spatial heterogeneity index of each sampling unit is calculated using a pre-trained heterogeneity identification model. Specifically, this includes extracting a multidimensional feature vector for each sampling unit, wherein the multidimensional feature vector includes the elevation variation coefficient, vegetation cover distribution variance, soil type diversity index, and infrastructure density.

4. The method for generating land parcel-level planning schemes based on grid sampling and machine learning according to claim 1, characterized in that, S1 further includes inputting the multidimensional feature vector into a pre-trained heterogeneity recognition model. The heterogeneity recognition model adopts a deep neural network structure, learns the complex correlation between features through multi-layer nonlinear transformation, and outputs a comprehensive spatial heterogeneity index. The calculation of the spatial heterogeneity index comprehensively considers the spatial distribution uniformity of features within the sampling unit, the fluctuation range of feature values, and the coupling relationship between different feature dimensions. The final spatial heterogeneity quantification value is obtained by weighted fusion of the degree of variation of each feature dimension.

5. The method for generating plot-level planning schemes based on grid sampling and machine learning according to claim 1, characterized in that, S2 includes partitioning the target land parcel according to the heterogeneity index using a multi-level threshold segmentation method, setting a first threshold and a second threshold, and dividing the land parcel into high heterogeneity region, medium heterogeneity region and low heterogeneity region.

6. The method for generating plot-level planning schemes based on grid sampling and machine learning according to claim 5, characterized in that, In the highly heterogeneous region, a quadtree recursive segmentation algorithm is used to refine the local mesh, continuously dividing the basic mesh unit into sub-mesh until the heterogeneity index of the sub-mesh is lower than the first threshold or reaches the preset maximum segmentation level. In the heterogeneous region, a selective encryption strategy is adopted, and mesh refinement is performed only at a limited number of levels at the boundary of characteristic mutations; The initial sampling density is maintained in the low heterogeneity region, and a sparse sampling strategy is adopted.

7. The method for generating plot-level planning schemes based on grid sampling and machine learning according to claim 1, characterized in that, The adaptive dynamic sampling also includes a sampling quality feedback mechanism. After each round of grid encryption, the heterogeneity index of each sampling unit is recalculated. The boundary and density of the encrypted region are dynamically adjusted according to the recalculated heterogeneity index. A spatially optimized sampling unit set is generated through an iterative optimization process. The generation of the optimized sampling unit set also includes unit fusion, which involves merging multiple adjacent low heterogeneous sampling units into a representative sampling unit when their spatial feature similarity is higher than a first similarity threshold.

8. The method for generating plot-level planning schemes based on grid sampling and machine learning according to claim 1, characterized in that, The planning parameter prediction model is a graph neural network-based planning parameter prediction model. The optimized sampling unit set is input into the graph neural network-based planning parameter prediction model. The model treats each optimized sampling unit as a graph node, and the spatial adjacency relationship and feature similarity between units as edge connections. The neighborhood information is aggregated through a multi-layer message passing mechanism to generate planning parameters for each optimized sampling unit, including land use type, building density, plot ratio and greening rate.

9. The method for generating plot-level planning schemes based on grid sampling and machine learning according to claim 1, characterized in that, Based on the aforementioned planning parameters, the generation of plot-level planning schemes through a dynamic influence domain construction and propagation algorithm specifically includes: Taking each optimized sampling unit as the core, weights are assigned according to the type and importance of its planning parameters to construct an initial influence domain, in which the influence domain of ecological protection parameters is larger than that of development and construction parameters; An adaptive decay function is used to spatially propagate the influence domain of different planning parameters. The Gaussian kernel bandwidth of the decay function is dynamically adjusted according to the parameter type and the local heterogeneity index. In regions with high heterogeneity, a small bandwidth is used to preserve detailed features.

10. The method for generating plot-level planning schemes based on grid sampling and machine learning according to claim 9, characterized in that, The iterative optimization and adjustment of the scope and intensity of the influence domain specifically includes: Calculate the gradient change at the boundary of adjacent influence domains, detect parameter conflict regions, and reallocate influence domain weights based on parameter priority rules; In each iteration, the overlapping influence domains are superimposed in intensity and conflict is resolved, with the influence domain of high-priority parameters having a dominant position in the conflict region; The propagation algorithm continues to run until the rate of change of the boundary of the adjacent influence domain is lower than the convergence threshold. Finally, a coordinated plot-level planning scheme is generated by spatial fusion of the parameter intensity distribution map. The scheme maintains the continuous spatial transition of parameters and functional coordination.

Images