A method for evaluating territorial spatial planning based on deep learning

Abstract

The present invention belongs to the technical field of spatial planning, and specifically refers to a method for evaluating national territorial spatial planning based on deep learning. The method constructs a scene graph structure to generate an initial scene graph, uses an iterative planner to update the scene graph and record feedback. The scene graph structure can visually present the relationships between planning objects, facilitating the analysis of complex national territorial spatial layouts; extracts the physical features of the current scene graph, deploys a planning evaluation model to generate a label distribution, determines evaluation indicators to calculate the spatial planning loss, optimizes the planning evaluation model, deeply mines the information of the current scene graph, helps to accurately grasp the relationship characteristics between objects, provides accurate evaluation for spatial planning, comprehensively considers planning constraints, and provides reliable decision-making support; adopts a dynamic update mechanism to timely discover spatial planning problems and improve the rationality and feasibility of spatial planning.

Description

Technical Field

[0001] The present invention belongs to the field of spatial planning technology, and specifically refers to a national land space planning evaluation method based on deep learning. Background Art

[0002] National land space planning refers to the spatial and temporal arrangements for the development and protection of national land space in a certain area. It is the basic basis for various development, protection and construction activities. The current national land space planning has the following shortcomings:

[0003] The current national land space planning process is difficult to grasp the complex spatial relationship. In the regional development plan, it is impossible to clearly show the mutual connection and influence between different functional areas, and it is difficult to adjust the plan in time according to the actual development situation, which may cause confusion in the layout of functional areas and uncoordinated infrastructure construction. For example, there is a lack of reasonable transportation connection and supporting facilities planning between industrial areas and residential areas;

[0004] The current national land space planning is unable to deeply analyze the spatial layout characteristics, and thus cannot accurately evaluate the planning effect, resulting in the discovery of low land use efficiency and ecological environmental damage after the planning scheme is implemented. In the planning process, the coordinated development of economy, society and ecology cannot be fully considered, and the excessive pursuit of economic growth and neglect of ecological environmental protection have led to over-exploitation of land resources.

[0005] The current national land space planning lacks a long-term strategic layout and only meets current short-term needs, which is not conducive to sustainable development. Especially in transportation planning, future urban expansion and changes in transportation patterns are not taken into account, resulting in planning strategies that cannot adapt to development needs and require frequent transformation and expansion, resulting in waste of resources. In addition, planning strategies lack flexibility and agility and are unable to cope with uncertainties and changes. Summary of the invention

[0006] In view of the above problems, the present invention provides a land space planning evaluation method based on deep learning. The method constructs a scene graph structure, generates an initial scene graph, uses an iterative planner to update the scene graph and record feedback. The scene graph structure can intuitively present the relationship between planning objects, which is convenient for analyzing complex land space layout; extracts the physical characteristics of the current scene graph, deploys a planning evaluation model to generate label distribution, determines evaluation indicators to calculate spatial planning losses, optimizes the planning evaluation model, and deeply mines the current scene graph information, which helps to accurately grasp the relationship characteristics between objects, provide accurate evaluation for spatial planning, comprehensively consider planning constraints, improve model robustness and generalization ability, and provide reliable decision support; adopts a dynamic update mechanism to timely discover spatial planning problems and improve the rationality and feasibility of spatial planning.

[0007] A land space planning evaluation method based on deep learning specifically includes the following steps:

[0008] Step S1: Data collection and preprocessing: Collect the planning data of the territorial spatial planning, perform preprocessing operations on the planning data of the territorial spatial planning to obtain preprocessed data, and generate a global dictionary;

[0009] Step S2: Scenario graph construction and update: Construct an initial scenario graph, generate a planning action sequence and planning constraints, execute the planning action sequence while observing the planning constraints, convert the initial scenario graph into the current scenario graph, and generate feedback information on the planning action sequence;

[0010] Step S3: Deployment of the planning evaluation model: Extract the physical features of the current scenario graph, construct a planning evaluation model, obtain the label distribution of the current scenario graph based on the physical features, determine the planning evaluation indicators, and further calculate the spatial planning loss. By minimizing the spatial planning loss, train and optimize the planning evaluation model;

[0011] Step S4: Application of spatial planning evaluation: Use the planning evaluation model to conduct land use evaluation, traffic condition evaluation, and socioeconomic evaluation, and generate an evaluation result report.

[0012] Furthermore, Step S1 includes the following steps:

[0013] Step S11: Collect geospatial data: Collect high-resolution satellite images, and use the satellite images to provide land use type information and topographic and geomorphic information. The land use type information includes construction land, agricultural land, and ecological land, and the topographic and geomorphic information includes altitude, slope, and aspect;

[0014] Step S12: Collect socioeconomic data: Collect population distribution data, economic development indicators, and traffic network distribution information. The economic development indicators include GDP and industrial structure data, and the traffic network distribution information includes traffic flow data;

[0015] Step S13: Collect planning document data: Collect the planning objectives and spatial layout plans of the territorial spatial planning. The planning objectives include the proportion of the ecological protection red line and the control target of the construction land scale, and the spatial layout plan includes functional zoning and industrial layout;

[0016] Step S14: Preprocess geospatial data: Use ground control points to correct the image into the correct geographic coordinate system through polynomial fitting, perform normalization processing on the geospatial data, and map the value range to a specific interval;

[0017] Step S15: Preprocess socioeconomic data: Use the K-nearest neighbor algorithm to fill in the missing values of the socioeconomic data;

[0018] Step S16: Preprocess the planning file data: Extract the key indicators and time nodes in the planning objectives through natural language processing technology, and vectorize the spatial layout plan.

[0019] Step S17: Construct a global dictionary: Merge the names extracted from the planning data of the territorial spatial planning to form an initial set of names, and delete the duplicate names to generate a global dictionary.

[0020] Furthermore, step S15 includes the following steps:

[0021] Step S151: Process missing values: For the missing traffic data of some sections in the traffic network distribution information, based on the traffic data and flow data of the surrounding sections, use the K-nearest neighbor algorithm to predict the missing traffic data values.

[0022] Step S152: Normalization processing: Convert the data into the form of a standard normal distribution.

[0023] Step S153: Multi-source data fusion: For the population distribution data, economic development indicators, and traffic network distribution information, perform fusion through weighted summation and feature splicing methods.

[0024] Furthermore, step S2 includes the following steps:

[0025] Step S21: Define the scene graph: According to the preprocessed data, construct the planning scene graph structure as G = {V, E}, where V is the node representing the objects in the planning scene, E is the edge describing the relationship between the objects, and construct a relationship set R representing the basic spatial planning relationship.

[0026] Step S22: Generate the scene graph: Use the scene graph generator to receive the preprocessed data, relationship set, and global dictionary according to the planning scene graph structure, and generate an initial scene graph, where the initial scene graph includes N objects.

[0027] Step S23: Define the planning task: According to the initial scene graph, design a spatial planner based on the Transformer technology to generate a planning action sequence and planning constraints. The planning constraints set the preconditions that must be met before each planning action is executed and the postconditions that need to be met after execution.

[0028] Step S24: Update the scene graph: Execute the planning action sequence while observing the planning constraints, convert the initial scene graph into the current scene graph, and generate feedback information of the planning action sequence. Design an iterative planner, set an end marker, and execute the planning action sequence until the end marker is obtained.

[0029] Furthermore, step S3 includes the following steps:

[0030] Step S31: Extract physical features: The roads on the side in the current scene map are divided into three categories: ordinary roads, arterial roads, and expressways. Collect the road perimeter configuration, traffic-related POIs, land use types, and building outlines to describe the physical features of the current scene map from four dimensions. The physical features include the degree of population flow, transportation facility construction, road topology, and spatial richness;

[0031] Step S32: Project the scene map: Construct a probabilistic encoder and a deterministic decoder, project the current scene map of the national territorial space planning into the feature dimension through a multi-layer perceptron, initialize N query subsets, and capture the spatial layout context information;

[0032] Step S33: Build a planning evaluation model: The planning evaluation model includes a prior encoder, a posterior encoder, a decoder, a re-encoder, a convolutional encoder, and a planner. The prior encoder, posterior encoder, decoder, and re-encoder contain transformer layers, repeated self-attention blocks, cross-attention blocks, and multi-layer perceptrons. The planner includes two cross-attention blocks, and comprehensively consider the spatial layout context information provided by the current scene map using the planning evaluation model;

[0033] Step S34: Calculate the parameters of the posterior distribution: Use the posterior encoder to calculate the parameters of the posterior distribution, and output the planning evaluation information through an additional cross-attention block;

[0034] Step S35: Calculate the parameters of the prior distribution: Use the prior encoder through the transformer layer, set the initial hypothesis, iteratively adjust the spatial layout context information using the initial hypothesis for noise detection, capture the occlusion relationship between objects and the repeated detection relationship of the same object, input the residual of the initial hypothesis into the multi-layer perceptron, estimate the mean vector and diagonal covariance matrix of the prior distribution, calculate the parameters of the prior distribution, and describe the error of the planning evaluation information;

[0035] Step S36: Feature information fusion: Use the decoder to generate a set of proxy bounding boxes for the error of the planning evaluation information, and perform feature information fusion on the initial hypothesis and the error of the planning evaluation information through the repeated self-attention block and the cross-attention block to form the fused features;

[0036] Step S37: Generate the label distribution: Use the re-encoder to re-encode the fused features to form the encoded features. At the same time, obtain the physical features of the current scene map from the convolutional encoder. Use the planner through two cross-attention blocks to respectively focus on the encoded features and the physical features to obtain the label distribution of the current scene map;

[0037] Step S38: Optimize the planning evaluation model: Determine the planning evaluation index according to the label distribution of the current scene map, further calculate the spatial planning loss, train and optimize the planning evaluation model, and combine the physical features of the current scene map to update the spatial layout context information.

[0038] Furthermore, step S38 includes the following steps:

[0039] Step S381: Determine the planning evaluation index: Use the K-means clustering algorithm to aggregate the label distribution of the current scene graph, extract typical scenes for ordinary roads, main roads, and expressways respectively, and determine the scene labels of each typical scene. Take the scene labels as the planning evaluation index;

[0040] Step S382: Calculate the spatial planning loss: Add a prior regularization incentive to the planning evaluation index, and combine the prior distribution. Infer the prior distribution of the current scene graph through the probability encoder, and calculate the spatial planning loss. The specific formula used is as follows:

[0041] ;

[0042] Among them, represents the spatial planning loss, represents the intensity coefficient for controlling the prior regularization incentive, represents the number of hypotheses of the processing object, n represents, z represents the latent variable, represents the prior distribution, represents a set of latent variables of the planning error, S represents a set of vectors describing N objects in the current scene graph, M represents the rasterized current scene graph, represents the land use coefficient, i and j represent the serial numbers of the current scene graph, represents the relative proportion of each planning building type in the i-th current scene graph, represents the economic development equilibrium coefficient, represents the population factor in the i-th current scene graph, represents the traffic coefficient, represents the economic activity intensity in the j-th current scene graph, represents the distance decay coefficient, represents the traffic impedance between the i-th current scene graph and the j-th current scene graph after distance decay;

[0043] Step S383: Optimize the planning evaluation model: Use the consistency distillation method to train the planning evaluation model. Given the current scene graph, generate noise samples by adding noise and discretization, perform backpropagation training to minimize the mean square distance, add sine position information embedding to combine with the physical features of the current scene graph, and update the spatial layout context information.

[0044] The beneficial effects achieved by the present invention are as follows:

[0045] (1) The method fuses multi-source data, providing a rich, accurate, and interrelated data foundation for territorial spatial planning. The fusion of different types of data comprehensively reflects the current situation and goals of the planning area, making planning decisions more scientific. Combining land use types with population distribution and economic development indicators helps to reasonably determine the functional positioning of different regions and achieve the optimal allocation of land resources;

[0046] (2) The method constructs a scene graph structure and defines a set of relationships to represent basic spatial planning relationships. An initial scene graph is generated by a scene graph generator, generating a planning action sequence and constraints. The iterative planner updates the scene graph by executing the action sequence while adhering to the constraints, visually presenting various elements and their relationships in territorial spatial planning in a graph structure, facilitating the analysis and understanding of complex spatial layouts. The dynamic update mechanism makes the planning process visual and traceable, enabling the timely discovery of problems and conflicts in planning decisions. In urban renewal planning, it clearly shows the relationships between buildings and surrounding transportation and public service facilities. Through the execution and feedback of the planning action sequence, it optimizes planning decisions such as building layout adjustment and traffic route optimization, improving the rationality and feasibility of the planning;

[0047] (3) The method projects the scene graph into the feature dimension through a multi-layer perceptron, calculates the posterior distribution parameters and prior distribution parameters by the collaborative work of each component, performs feature information fusion and label distribution generation, can deeply mine the spatial layout context information in the scene graph, accurately capture the complex relationships and features between objects, can better consider the relationships between road topologies, transportation facility construction, and population flow levels, thus providing more accurate evaluation and optimization suggestions for spatial planning. Combining features such as land use types and building outlines, it more reasonably judges the rationality and development trends of land use, helping to formulate scientific land use planning strategies;

[0048] (4) The method more comprehensively considers various goals and constraints in territorial spatial planning, avoiding a single factor dominating the optimization direction of the model. The introduction of the land use coefficient prompts the model to focus on the rationality of the land use structure. The use of the consistency distillation method and noise processing improves the robustness and generalization ability of the model, enabling it to better adapt to different planning scenarios and data changes, providing more reliable evaluation and decision-making support for territorial spatial planning. Brief Description of the Drawings

[0049] Figure 1 It is a flowchart of an evaluation method for territorial spatial planning based on deep learning proposed by the present invention;

[0050] Figure 2 It is a flowchart of step S3 proposed by the present invention. Detailed Embodiment

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention.

[0052] Example 1: Refer to Figure 1 , this embodiment provides a method for evaluating territorial spatial planning based on deep learning, specifically including the following steps:

[0053] Step S1: Data collection and preprocessing: Collect the planning data of territorial spatial planning, perform preprocessing operations on the planning data of territorial spatial planning to obtain preprocessed data, and generate a global dictionary;

[0054] Step S2: Scenario graph construction and update: Construct an initial scenario graph, generate a planning action sequence and planning constraints, execute the planning action sequence while observing the planning constraints, convert the initial scenario graph into the current scenario graph, and generate feedback information of the planning action sequence;

[0055] Step S3: Deployment of the planning evaluation model: Extract the physical features of the current scenario graph, construct a planning evaluation model, obtain the label distribution of the current scenario graph according to the physical features, determine the planning evaluation indicators, and further calculate the spatial planning loss. By minimizing the spatial planning loss, train and optimize the planning evaluation model;

[0056] Step S4: Application of spatial planning evaluation: Use the planning evaluation model to conduct land use evaluation, traffic condition evaluation and social and economic evaluation, and generate an evaluation result report.

[0057] Example 2: This embodiment is based on the above embodiment, and step S1 includes the following steps:

[0058] Step S11: Collect geospatial data: Collect high-resolution satellite images, and use the satellite images to provide land use type information and topographic and geomorphic information. The land use type information includes construction land, agricultural land and ecological land, and the topographic and geomorphic information includes altitude, slope and aspect;

[0059] Step S12: Collect social and economic data: Collect population distribution data, economic development indicators and traffic network distribution information. The economic development indicators include GDP and industrial structure data, and the traffic network distribution information includes traffic flow data;

[0060] Step S13: Collect planning document data: Collect the planning objectives and spatial layout plans of territorial spatial planning. The planning objectives include the proportion of the ecological protection red line and the control target of the construction land scale, and the spatial layout plan includes functional zoning and industrial layout;

[0061] Step S14: Preprocess geospatial data: Using ground control points, correct the image to the correct geographic coordinate system through polynomial fitting method, normalize the geospatial data, and map the value range to a specific interval [0,1];

[0062] Step S15: Preprocess socioeconomic data: Adopt the K-nearest neighbor algorithm to fill in the missing values. For example, for economic data, convert it into the form of a standard normal distribution with a mean of 0 and a standard deviation of 1 to eliminate the influence of the dimension between different indicators;

[0063] Step S16: Preprocess the planning file data: Through natural language processing technology, extract the key indicators and time nodes in the planning objectives, and vectorize the spatial layout plan;

[0064] Step S17: Construct a global dictionary: Merge the names extracted from the relevant data of the territorial spatial planning to form an initial name set, and delete the duplicate names to generate a global dictionary.

[0065] Example 3: This example is based on the above example, and step S15 includes the following steps:

[0066] Step S151: Process missing values: For the missing traffic flow data in some sections of the traffic network distribution information, based on the traffic flow data and traffic flow data of the surrounding sections, use the K-nearest neighbor algorithm to predict the missing traffic flow data values;

[0067] Step S152: Normalization processing: Convert the data into the form of a standard normal distribution;

[0068] Step S153: Multi-source data fusion: For population distribution data, economic development indicators, and traffic network distribution information, perform fusion through weighted summation and feature splicing methods.

[0069] Example 4: This example is based on the above example, and step S2 includes the following steps:

[0070] Step S21: Define the scene graph: According to the preprocessed data, construct the planning scene graph structure as G={V,E}, where V is the node representing the objects in the planning scene, E is the edge describing the relationship between the objects, and construct the relationship set as R representing the basic spatial planning relationship;

[0071] Step S22: Generate the scene graph: Use the scene graph generator, according to the planning scene graph structure, receive the preprocessed data, relationship set, and global dictionary, and generate an initial scene graph. The initial scene graph includes 128 objects;

[0072] Step S23: Define the planning task: Based on the initial scene graph, design a spatial planner using Transformer technology to generate a sequence of planning actions and planning constraints. The planning constraints set the preconditions that must be met before each planning action is executed and the postconditions that need to be satisfied after execution.

[0073] Step S24: Update the scene graph: Execute the sequence of planning actions while adhering to the planning constraints, convert the initial scene graph into the current scene graph, and generate feedback information for the sequence of planning actions. Design an iterative planner that receives the initial scene graph, the current scene graph, the planning constraints, and the feedback information for the sequence of planning actions, set an end marker, and execute the sequence of planning actions until the end marker is obtained to get the current scene graph, which reflects the spatial layout context information.

[0074] If a planning action violates the planning constraints, increment the error count and output the feedback information until the end marker is obtained.

[0075] If a planning action passes the planning constraint verification, update the state of the current scene graph to reflect the execution of each planning action. For the removal operation of an object, remove the edge representing the initial support relationship and create a new edge representing the new support relationship, and correspondingly modify the current scene graph of the plan, and continue to verify the next planning action in the sequence of planning actions until the end marker is obtained to get the current scene graph, which reflects the spatial layout context information.

[0076] Example 5: Refer to Figure 2 , this example is based on the above example, and step S3 includes the following steps:

[0077] Step S31: Extract physical features: Extract the roads in the current scene graph and classify them into three categories: ordinary roads, main roads, and expressways. Collect the configurations around the roads, traffic-related POIs, land use types, and building outlines to describe the physical features of the current scene graph from four dimensions. The physical features include the degree of population flow, transportation facility construction, road topology, and spatial richness.

[0078] Traffic-related POIs include subway stations, bus stops, and parking lots.

[0079] Land use types include green spaces, administrative and public services, commercial and business facilities, education areas, and industrial areas.

[0080] Building outlines include shape and height.

[0081] Road topology includes road betweenness centrality, spatial integration, angle selection, and intersection density.

[0082] Spatial richness includes land use types and building outlines.

[0083] Step S32: Project the scenario graph: Construct a probabilistic encoder and a deterministic decoder, project the current scenario graph of the territorial space planning into the feature dimension through a multi-layer perceptron, initialize 128 query subsets, and capture the spatial layout context information;

[0084] Step S33: Construct a planning evaluation model: The planning evaluation model includes a prior encoder, a posterior encoder, a decoder, a re-encoder, a convolutional encoder, and a planner. The prior encoder, posterior encoder, decoder, and re-encoder contain transformer layers, repeated self-attention blocks, cross-attention blocks, and multi-layer perceptrons. The planner includes two cross-attention blocks, and comprehensively consider the spatial layout context information provided by the current scenario graph using the planning evaluation model;

[0085] Step S34: Calculate the parameters of the posterior distribution: Use the posterior encoder to calculate the parameters of the posterior distribution, output the planning evaluation information through an additional cross-attention block, and obtain the spatial detection features;

[0086] Step S35: Calculate the parameters of the prior distribution: Use the prior encoder through the transformer layer, set the initial hypothesis, iteratively adjust the spatial layout context information for noise detection using the initial hypothesis, capture the occlusion relationship between objects and the repeated detection relationship of the same object, input the residual of the initial hypothesis into the multi-layer perceptron, estimate the mean vector and diagonal covariance matrix of the prior distribution, and calculate the parameters of the prior distribution to describe the error of the planning evaluation information;

[0087] Step S36: Feature information fusion: Use the decoder to generate a set of proxy bounding boxes for the error of the planning evaluation information, and perform feature information fusion on the initial hypothesis and the error of the planning evaluation information through repeated self-attention blocks and cross-attention blocks to form the fused features;

[0088] Step S37: Generate the label distribution: Use the re-encoder to re-encode the fused features to form the encoded features, and at the same time obtain the physical features of the current scenario graph from the convolutional encoder. Use the planner through two cross-attention blocks to focus on the encoded features and physical features respectively to obtain the label distribution of the current scenario graph;

[0089] Step S38: Optimize the planning evaluation model: Determine the planning evaluation index according to the label distribution of the current scenario graph, further calculate the territorial space planning loss, train and optimize the planning evaluation model, and update the spatial layout context information in combination with the physical features of the current scenario graph.

[0090] Example 6: This example is based on the above example, and step S38 includes the following steps:

[0091] Step S381: Determine the planning evaluation metrics: Use the K-means clustering algorithm to aggregate the label distribution of the current scene graph. Extract 4, 3, and 3 typical scenes for ordinary roads, arterial roads, and expressways respectively, and determine the top 15 scene labels for each typical scene. Use the scene labels as the planning evaluation metrics.

[0092] Step S382: Calculate the spatial planning loss: Add a prior regularization incentive to the planning evaluation metrics, and combine with the prior distribution. Infer the prior distribution of the current scene graph through the probability encoder, and calculate the spatial planning loss. The specific formula used is as follows:

[0093] ;

[0094] where, represents the spatial planning loss, represents the intensity coefficient controlling the prior regularization incentive, represents the number of assumptions of the processing object, n represents, z represents the latent variable, represents the prior distribution, represents a set of latent variables of the planning error, S represents a set of vectors describing N objects in the current scene graph, M represents the rasterized current scene graph, represents the land use coefficient, i and j represent the serial numbers of the current scene graph, represents the relative proportion of each planning building type in the i-th current scene graph, represents the economic development equilibrium coefficient, represents the population factor in the i-th current scene graph, represents the traffic coefficient, represents the economic activity intensity in the j-th current scene graph, represents the distance decay coefficient, represents the traffic impedance between the i-th current scene graph and the j-th current scene graph after distance decay;

[0095] Step S383: Optimize the planning evaluation model: Initialize all to the mean. Use the consistency distillation method to train the planning evaluation model. Given the current scene graph data points, generate noise samples by adding noise and discretization, and perform backpropagation training to minimize the mean square distance. Add sine position information embedding to combine with the physical features of the current scene graph, and update the spatial layout context information.

[0096] The above describes the present invention and its implementation manners. This description is not restrictive. If those of ordinary skill in the art are inspired by it and, without departing from the gist of the present invention, creatively design similar structural manners and embodiments to this technical solution, they shall fall within the protection scope of the present invention.

Claims

1. A method for evaluating territorial spatial planning based on deep learning, characterized in that, It includes the following steps: Step S1: Data collection and preprocessing: Collect the planning data of the territorial spatial planning, perform preprocessing operations on the planning data of the territorial spatial planning to obtain preprocessed data, and generate a global dictionary; Step S2: Scenario graph construction and update: Combine the preprocessed data and the global dictionary to construct an initial scenario graph, generate a planning action sequence and planning constraints. During the process of executing the planning action sequence according to the planning constraints, convert the initial scenario graph into the current scenario graph, and generate feedback information of the planning action sequence; Step S3: Deployment of the planning evaluation model: Extract the physical features of the current scenario graph, construct a planning evaluation model, obtain the label distribution of the current scenario graph according to the physical features, determine the planning evaluation index, calculate the spatial planning loss, and train and optimize the planning evaluation model by minimizing the spatial planning loss; Step S4: Spatial planning evaluation application: Use the planning evaluation model for evaluation and generate an evaluation result report; Step S3 includes the following steps: Step S31: Extract physical features; Step S32: Project the scenario graph: Construct a probabilistic encoder and a deterministic decoder, project the current scenario graph into the feature dimension through a multi-layer perceptron, initialize the query subset, and capture the spatial layout context information; Step S33: Construct the planning evaluation model: Use the planning evaluation model to comprehensively consider the spatial layout context information provided by the current scenario graph; Step S34: Calculate the posterior distribution parameters: Use the posterior encoder to calculate the parameters of the posterior distribution, output the planning evaluation information through an additional cross-attention block, and obtain the spatial detection features; Step S35: Calculate the prior distribution parameters: Use the prior encoder through the transformer layer, set the initial hypothesis, use the initial hypothesis to iteratively adjust the spatial layout context information for noise detection, input the residual of the initial hypothesis into the multi-layer perceptron, estimate the mean vector and diagonal covariance matrix of the prior distribution, and calculate the parameters of the prior distribution to describe the error of the planning evaluation information; Step S36: Feature information fusion: Perform feature information fusion on the initial hypothesis and the error of the planning evaluation information to form fused features; Step S37: Generate the label distribution: Re-encode the fused features to form encoded features, obtain the physical features of the current scenario graph, and get the label distribution of the current scenario graph; Step S38: Optimize the planning evaluation model: According to the label distribution of the current scenario graph, determine the planning evaluation index, calculate the spatial planning loss, adopt the prior regularization encouragement and consistency distillation method to train and optimize the planning evaluation model, and update the spatial layout context information.

2. The method for evaluating territorial spatial planning based on deep learning according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: Define the scenario graph: According to the preprocessed data, construct the planning scenario graph structure and relationship set; Step S22: Generate the scenario graph: Use the scenario graph generator, according to the planning scenario graph structure, receive the preprocessed data, relationship set and global dictionary, and generate the initial scenario graph; Step S23: Define the planning task: According to the initial scenario graph, design a spatial planner based on the Transformer technology, generate a planning action sequence and planning constraints, and the planning constraints set preconditions and postconditions for each planning action; Step S24: Update the scene graph: Execute the planned action sequence in compliance with the planning constraints, convert the initial scene graph into the current scene graph, generate feedback information for the planned action sequence, set the end marker, and execute the planned action sequence until the end marker is obtained to get the current scene graph.

3. The method for evaluating territorial spatial planning based on deep learning according to claim 2, wherein: The extraction of physical features in step S31 specifically includes: extracting the road edges within the current scene graph and describing the physical features of the current scene graph from four dimensions; The feature information fusion in step S36 specifically includes: using a decoder to perform feature information fusion on the initial hypothesis and the error of the planning evaluation information through repeated self-attention blocks and cross-attention blocks to form fused features; The generation of the label distribution in step S37 specifically includes: using a re-encoder to re-encode the fused features to form encoded features, and at the same time obtaining the physical features of the current scene graph from the convolutional encoder, and using a planner to focus on the encoded features and the physical features respectively through two cross-attention blocks to obtain the label distribution of the current scene graph.

4. The method for evaluating territorial spatial planning based on deep learning according to claim 3, wherein: Step S38 includes the following steps: Step S381: Determine the planning evaluation metrics: Aggregate the label distribution of the current scene graph, extract typical scenes for the road edges respectively, determine the scene labels for each typical scene, and use the scene labels as the planning evaluation metrics; Step S382: Calculate the spatial planning loss: Add prior regularization encouragement to the planning evaluation metrics, and combine with the prior distribution to calculate the spatial planning loss; Step S383: Optimize the planning evaluation model: Adopt the consistency distillation method to train the planning evaluation model. Given the current scene graph, generate noise samples by adding noise and discretization, perform backpropagation training, add sine position information embedding and combine with the physical features of the current scene graph to update the spatial layout context information.

5. The method for evaluating territorial spatial planning based on deep learning according to claim 3, characterized in that: The planning evaluation model includes a prior encoder, a posterior encoder, a decoder, a re-encoder, a convolutional encoder, and a planner. The prior encoder, the posterior encoder, the decoder, and the re-encoder contain transformer layers, repeated self-attention blocks, cross-attention blocks, and multi-layer perceptrons. The planner includes two cross-attention blocks.

Images