Land space purpose control compliance intelligent identification and auxiliary determination system

Abstract

The invention discloses a land space use control compliance intelligent identification and auxiliary determination system, and relates to the technical field of digital treatment. A data management module receives multi-source heterogeneous time sequence data to construct a dynamic graph structure substrate, and a probability scheduling module predicts global potential conflict probability by using a space-time graph neural network and performs calibration in combination with a historical conflict database; the graph enhancement module positions a key conflict path and a core node by reversely updating a conflict knowledge graph, and the reasoning execution module activates a rule engine to execute accurate matching in a region to be subjected to fine detection and outputs a deterministic conflict identifier; the fusion decision module finally fuses probabilistic prediction and rule judgment results to generate a mixed conflict list, and the feedback module sends the mixed conflict list to the execution terminal and optimizes the model based on feedback. According to the method, multi-level and self-adaptive identification of the land and space purpose conflict from wide-area prediction to accurate judgment is realized, and the supervision efficiency and the decision accuracy are improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of digital governance, in particular to a land space use control compliance intelligent identification and auxiliary judgment system. BACKGROUND

[0002] Land space use control is an important measure to ensure national food security, ecological security and sustainable development, and its core lies in the compliance review and supervision of various land use activities. In recent years, with the development of remote sensing, Internet of Things and other technologies, massive, multi-source and heterogeneous spatiotemporal data have provided the possibility for precise supervision and promoted the evolution of this field towards digitization and intelligentization.

[0003] The existing technical solutions mostly rely on the construction of rule engines and image recognition models to perform static comparison between the "current situation" and "planning" of the land space, thereby realizing conflict identification. These methods can effectively identify obvious and already existing violations under the condition of data specification and fixed scene, and have certain automatic processing capability. However, the land space is a dynamic and complex system, and various use conflicts are not isolated and static, but are correlated, transmitted and evolved in space over time. The limitation of the existing technology lies in that its analysis mode is essentially static and isolated, and it cannot dynamically perceive and predict the formation and propagation risk of potential conflicts from a global perspective. This leads to the fact that the system can only respond passively to the violations that have already occurred, and cannot provide early warning and intervention in the early stage of conflict germination or spread, resulting in a lag in regulatory behavior. At the same time, in the face of massive global data, there is a lack of an efficient mechanism to quickly focus on key areas from a macro-probabilistic level, which also causes waste of computing resources and bottleneck of regulatory efficiency. SUMMARY

[0004] In view of the deficiencies of the existing technology, the application provides a land space use control compliance intelligent identification and auxiliary judgment system.

[0005] In order to achieve the above-mentioned purpose, the technical scheme of the application is as follows:

[0006] The application discloses a land space use control compliance intelligent identification and auxiliary judgment system, which comprises:

[0007] A data management module is configured to receive multi-source heterogeneous time series data of a target area, construct a dynamic graph structure base, and initialize a conflict knowledge graph and a historical conflict database based on the graph structure base;

[0008] A probability scheduling module is configured to input the graph structure base into a pre-trained spatiotemporal graph neural network model, generate a probability distribution of global potential conflicts, and call a historical conflict list from the historical conflict database to generate a calibrated probability distribution.

[0009] a graph enhancement module, configured to output and reversely update to the conflict knowledge graph a part of the calibrated probability distribution with a probability value higher than a preset probability threshold;

[0010] a region positioning module, configured to locate a propagation path of a key conflict and a core node on the path by analyzing the updated conflict knowledge graph, and generate a region to be accurately inspected and a potential conflict probability of the region;

[0011] a reasoning execution module, configured to extract the region to be accurately inspected in real time, activate a rule engine, call a subset of regulation rules matched with characteristics of the core node and the propagation path, perform logical matching, and output a deterministic conflict identifier;

[0012] a fusion decision module, configured to fuse the deterministic conflict identifier and the potential conflict probability based on a fusion strategy, and generate a mixed conflict list;

[0013] a feedback module, configured to send the mixed conflict list to an execution terminal.

[0014] Compared with the prior art, the present application has the following beneficial effects:

[0015] 1. The present application can capture complex nonlinear spatiotemporal dependence between spatial entities, thereby generating a potential conflict probability distribution covering the whole domain, overcoming the limitation of traditional methods that can only identify static and explicit conflicts. The system can convert the probabilistic perception result under data driving into structured and inferable knowledge, laying a dynamic updated knowledge foundation for subsequent accurate positioning, and realizing the essential improvement of system cognition from static to dynamic and from retroactive to predictive.

[0016] 2. The present application can accurately identify the propagation path and core node of a key conflict, greatly reducing the overall computing cost of the system. The system can intelligently call a subset of regulation rules highly matched with the path and the core node, rather than traversing the entire global rule library, which makes the logical path of rule matching shorter and more targeted, and the "deterministic conflict identifier" output finally has higher accuracy and explainability.

[0017] 3. The present application improves the scientificity and reliability of auxiliary decision-making, and makes the entire system an intelligent agent that can balance the rigidity and flexibility of rules, and finally forms a closed loop through the feedback module, realizing the evolution of land space governance from one-way and static approval to two-way and dynamic intelligent supervision. BRIEF DESCRIPTION OF DRAWINGS

[0018] The disclosure of the present application will be described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present application. In the drawings, the same reference numerals are used to refer to the same components. Among them:

[0019] Figure 1 System module connection diagram of the present application;

[0020] Figure 2 System module flow chart of the present application;

[0021] Figure 3 Workflow diagram of the atlas enhancement module of the present application;

[0022] Figure 4 Workflow diagram of the region positioning module of the present application. DETAILED DESCRIPTION

[0023] It is easy to understand that, according to the technical solution of the present application, those skilled in the art can propose a plurality of structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the present application. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solution of the present application, and should not be regarded as the whole or as the limitation or restriction of the technical solution of the present application.

[0024] SUMMARY

[0025] In the prior art, land space use regulation compliance identification relies on manual interpretation or batch comparison of single rule engine, and it is difficult to balance regulatory efficiency and judgment accuracy. When facing wide-area and dynamically changing spatial data, the rule engine has huge global computing overhead, and the prediction model based on historical data is difficult to be directly used for law enforcement decision due to the lack of clear rule support. The existing system cannot effectively coordinate the macro risk perception ability and the micro rule decision ability, resulting in the disconnection between early warning and judgment when dealing with complex and evolving land space conflicts, and it is difficult to meet the demand for accurate, efficient and intelligent regulation.

[0026] In order to solve the above problems, the present application finds that there is a deep correlation between the probability distribution of potential conflicts and the structured knowledge of the knowledge graph, and the advantages can be complementary by constructing a closed loop of "probability prediction-knowledge update-rule fine judgment". It is found in the research process that the probability prediction based on spatio-temporal graph neural network can realize wide-area rapid screening but the accuracy is limited, and the judgment result based on rule engine is accurate but the calculation cost is high, so it is proposed to bridge the two through dynamic knowledge graph to realize the progressive processing idea from "wide-area early warning" to "accurate judgment". Further verification through system design, high-probability conflicts are injected into the knowledge graph to drive fine detection region positioning, forming a self-adaptive optimization closed loop of perception, cognition and decision cooperation.

[0027] After introducing the basic idea of the present application, the embodiments of the present application will be specifically introduced with reference to the drawings.

[0028] EMBODIMENT

[0029] As Figure 1 shown, the land space use regulation compliance intelligent identification and auxiliary judgment system comprises:

[0030] A data management module is configured to receive multi-source heterogeneous time-series data of a target area, construct a dynamic graph structure base, and initialize a conflict knowledge graph and a historical conflict database based on the graph structure base;

[0031] A probability scheduling module is configured to input the graph structure base into a pre-trained spatio-temporal graph neural network model, generate a probability distribution of global potential conflicts, and call a historical conflict list from the historical conflict database to generate a calibrated probability distribution;

[0032] A graph enhancement module is configured to output and update to the conflict knowledge graph in reverse the part of the calibrated probability distribution with a probability value higher than a preset probability threshold;

[0033] A region positioning module is configured to locate the propagation path of the key conflict and the core node on the path by analyzing the updated conflict knowledge graph, and generate a region to be accurately inspected and a potential conflict probability thereof;

[0034] An inference execution module is configured to extract the region to be accurately inspected in real time, activate a rule engine, call a regulation rule subset matched with the characteristics of the core node and the propagation path, perform logical matching, and output a deterministic conflict identifier;

[0035] A fusion decision module is configured to fuse the deterministic conflict identifier and the potential conflict probability based on a fusion strategy, and generate a mixed conflict list;

[0036] A feedback module is configured to send the mixed conflict list to an execution terminal.

[0037] As Figure 2 shown, the working principle and hardware structure of the present application are as follows: when the system is running, the data management module continuously receives multi-source heterogeneous time-series data of a target area through a satellite remote sensing data receiving station, an Internet of Things sensor network, and a government data interface. After the data enters the system, it is cleaned, standardized in format, and time-aligned by a data processing unit deployed on a server to form a standard data stream. Subsequently, the system uses a graph database construction engine to instantiate the fused data into a dynamic graph structure base with spatial entities as nodes and spatial relationships as edges based on a predefined land space ontology model, and adds a timestamp version to each graph element. The graph structure base is persistently stored in a distributed graph database, and based on this, the system initializes a conflict knowledge graph for storing entity conflict knowledge and a historical conflict database for recording historical cases.

[0038] The probability scheduling module is activated on the GPU computing nodes of the system, taking the graph structure base as input, and sending it into a pre-trained spatio-temporal graph neural network model for inference. The spatio-temporal graph neural network model captures spatial dependencies through its graph convolution layer and learns temporal evolution rules through its gated temporal convolution layer, and finally outputs a potential conflict probability distribution map covering the entire domain. To improve the robustness of the prediction, the probability scheduling module simultaneously calls a historical conflict list similar to the current graph structure topology pattern from the historical conflict database, and calibrates the initial probability distribution through a weighted fusion algorithm to generate a more reliable calibrated probability distribution.

[0039] The atlas enhancement module sets a preset probability threshold to filter the calibrated probability distribution, and converts high-risk conflict nodes and their spatial relationships with a probability value higher than the threshold into graph triples in the form of (subject, relationship, object) through a probability atlas mapping model. These triples are added to the conflict knowledge graph in real time, and the weights of the corresponding nodes are updated, thereby completing the reverse update of the knowledge graph, enabling it to evolve from a static knowledge base to a dynamic graph that reflects real-time dynamic risks.

[0040] Based on the updated conflict knowledge graph, the regional positioning module uses graph analysis algorithms such as betweenness centrality calculation to traverse the knowledge graph and identify core nodes in key hub positions in the network. Starting from these nodes, the module uses pathfinding algorithms such as random walk to find propagation paths connecting multiple core nodes. In this way, the system accurately locates the propagation path of key conflicts and the core nodes on the path, and delineates the areas that need to be carefully examined, along with their potential conflict probabilities.

[0041] After that, the inference execution module first extracts the vector boundary data of the above-mentioned areas to be carefully examined from the spatial database in real time. The built-in rule engine (a kind of expert system kernel) in the system is activated, and based on the land use type attributes of the core nodes and the conflict diffusion types represented by the propagation path, a highly matched subset of regulatory rules is intelligently selected from the complete global rule library. The rule engine performs logical matching on the spatial data within the subset of regulatory rules, and finally outputs a deterministic conflict identifier with legal effect.

[0042] The fusion decision module performs consistency checking and weighted calculation on the deterministic conflict identifier representing the determined facts and the potential conflict probability representing the risk warning based on the preset fusion strategy, generating a hybrid conflict list containing both explicit conclusions and risk levels.

[0043] Finally, the feedback module sends the hybrid conflict list to the enforcement terminal of law enforcement personnel or the superior supervision platform through the government network or dedicated communication link.

[0044] Through the above workflow, the application constructs a complete closed loop from macro risk perception to micro precise judgment, and then to result feedback, realizes the cooperation of data driving and knowledge guidance, can intelligently allocate computing resources, while ensuring full coverage, improves the efficiency and accuracy of compliance judgment in key areas, and provides reliable technical support for intelligent governance of territorial space.

[0045] The application further proposes that the specific steps of constructing a dynamic graph structure base by receiving multi-source heterogeneous time series data of the target area include:

[0046] After the system is started, the data management module continuously accesses multi-source heterogeneous time series data of the target area through a satellite data receiving station, Internet of Things sensor nodes deployed on site, and government data API interfaces. These data include, but are not limited to, time series remote sensing images, land use change vector data, construction project approval flow data, and environmental monitoring point time series data.

[0047] The multi-source heterogeneous time series data is cleaned and time-aligned to generate standardized data streams. The cleaning process includes: atmospheric correction and geometric registration of remote sensing images, topological inspection and repair of vector data, and outlier removal and missing value interpolation of sensor data. Subsequently, the system uses a dynamic time warping algorithm based on a sliding time window to time-align all data, ensuring that data from different sources are comparable under a unified time reference.

[0048] The standardized data stream is fused based on a predefined territorial space ontology model to generate a graph structure base with unified spatio-temporal coding; the territorial space ontology model is loaded at system initialization and stored in the knowledge base in the form of an OWL (Web Ontology Language) file, which clearly defines entity classes such as "farmland", "forest land", "construction land", and their relationships such as "adjacent to", "contained in", and "transformed from". The fusion engine (such as Apache Jena) uses this ontology model to perform semantic annotation and association on the standardized data stream, assigning a unified spatio-temporal code (e.g., using GeoSOT global subdivision grid coding) to each spatial entity (such as a graph patch), thereby generating a rudimentary graph structure base with a unified spatio-temporal framework.

[0049] According to the entities and relationships defined in the territorial space ontology model, the graph structure base is instantiated as nodes and edges. Specifically, in this process, each independent spatial entity (such as a planning plot or an ecological protection zone) is instantiated as a node, and the node attributes include its spatial location, land use type, area, etc.; spatial or logical relationships between entities (such as two plots being adjacent or a project being located within a certain area) are instantiated as edges connecting the nodes, and the edge attributes can include relationship type, strength (such as shared boundary length), etc.

[0050] To realize the dynamics of the graph structure, the system attaches a timestamp version (e.g., the UTC time of data acquisition) to each node and edge, and sets an update trigger condition, for example, when remote sensing detects that the surface changes exceed the preset threshold (such as the NDVI index changes more than ±0.1, or the surface cover type changes) or receives new approval data, an incremental update process is triggered. This process is realized through the version control mechanism of the graph database (such as Neo4j or Nebula Graph), ensuring that the graph structure base can continuously and accurately reflect the dynamic evolution of the land space.

[0051] The training of the spatio-temporal graph neural network model is completed offline. The training data comes from the historically accumulated, labeled graph structure base sequence and its corresponding conflict label verified by artificial verification. The spatio-temporal graph neural network model adopts a hybrid architecture based on message passing graph attention network (GAT) and gated recurrent unit (GRU). During the training process, the Adam optimizer is used, the learning rate is set to 1e-3, the batch size is 32, the training period is 100, and the binary cross-entropy loss function between the predicted conflict probability and the true label is minimized.

[0052] Through the above technical solutions, the dynamic graph structure base constructed by the present application not only provides high-quality, structured input for subsequent probability prediction and knowledge graph evolution, but also ensures the timeliness and accuracy of the system's perception of the state of the land space through its unified spatio-temporal coding and continuous updating mechanism. The base integrates multi-source heterogeneous data into an organic whole, enabling the machine to understand and calculate the complex relationships and dynamic interactions between spatial entities.

[0053] The present application further proposes that the specific steps of calling the historical conflict list from the historical conflict database for calibration to generate the calibrated probability distribution include:

[0054] By calculating the graph similarity, the system extracts the topological features of the current graph structure base, including but not limited to the degree distribution of nodes, the composition ratio of various spatial relationships (edges), and the graph-level representation vector generated by the graph embedding model (such as Node2Vec). Subsequently, the system uses cosine similarity as a measure to quickly retrieve the Top-K (K can be preset according to system resources, for example, 20) most similar historical graph patterns and their corresponding conflict records in the historical database. These conflict records are aggregated and encoded into a historical conflict pattern vector where N is the total number of spatial entities in the system, and the i-th component of the vector represents the frequency or average probability of the corresponding node i colliding in the similar historical pattern.

[0055] The historical conflict pattern vector is weightedly fused with the probability distribution of the global potential conflict to output the calibrated probability distribution. The weighting process is not fixed but dynamically determined by a lightweight calibration weight network. The network takes the average similarity score d between the current graph structure and the retrieved historical graph pattern as input and outputs the model prediction weight and the historical experience weight . The fusion formula is as follows:

[0056]

[0057] wherein, is the calibrated probability distribution. The calibration weight network can be a simple fully connected layer, and the weights and bias terms thereof are obtained through offline training. When training the network, the training set used is the same as that of the main model, and the target is to minimize the cross-entropy loss between the calibrated probability and the true conflict label. The training conditions are usually set as: learning rate 5e-4, using Adam optimizer, batch size 64, and training for 50 cycles. In this way, when the current situation is highly similar to the historical experience (high value of the average similarity score d), the system will give the historical conflict pattern vector a higher weight ; on the contrary, for novel or rare graph patterns, the prediction result of the model is relied on more (weight is higher). The weights and are normalized by the Softmax function to ensure .

[0058] Through the above calibration mechanism, the present application not only relies on the general rules learned by the model from the data, but also sensitively perceives and learns from the historical conflict patterns highly related to the current situation, thereby improving the accuracy and reliability of the potential conflict warning, especially for the conflict types that repeatedly occur in history or have specific topological characteristics, the false positive and false negative rates can be effectively reduced, and the probability prediction result is more valuable in practice.

[0059] As shown in Figure 3 , the present application further proposes that the specific steps of outputting and reversely updating the part of the calibrated probability distribution with a probability value higher than a preset probability threshold to the conflict knowledge graph include:

[0060] The calibrated probability distribution is input into the pre-stored probability graph mapping model. The mapping process is determined by a preset probability threshold trigger, which is determined by statistical methods based on the distribution characteristics of historical conflict data. For example, by calculating the percentile (e.g., 85th percentile) of the probability prediction value of the corresponding region before the occurrence of historical conflict events, the probability prediction value of the corresponding region before the occurrence of historical conflict events can be determined is set to 0.72. The system screens all nodes with probability values higher than this threshold to form a high-probability conflict node set , wherein is the value of the node in the calibrated probability distribution .

[0061] By the probability graph mapping model, the high-probability conflict nodes and the spatial relationships of the high-probability conflict nodes are converted into graph triples. For each node , the model generates a core assertion triple: , hasConflictRisk, "High"). The probability graph mapping model extracts edges (i.e., spatial relationships) connecting these high-probability nodes from the dynamic graph structure base. For any two high-probability nodes , if they have an edge in the original graph base, a relationship triple is generated: . This process connects discrete high-risk nodes through their inherent spatial relationships, and preliminarily reveals the potential conflict propagation chain.

[0062] The graph triples are added to the conflict knowledge graph, and the weights of the corresponding nodes are updated to complete the reverse update of the conflict knowledge graph. Each node in the knowledge graph has a dynamic weight , which is initialized to 1.0. When a new triple asserts that it has a high risk, its weight is updated as follows:

[0063]

[0064] wherein is the calibrated probability value of the node this time, is the learning rate parameter, which is usually set between 0.1 and 0.3 according to the requirement of the system for the speed of knowledge update, for example, it can be 0.2. This means that the higher the probability value of the conflict risk, the greater the improvement of the node weight, and the more important the node in the subsequent graph analysis (such as core node positioning).

[0065] The probabilistic graph mapping model itself is a rule-based and statistical combined model, and its "training" is actually a rule construction and optimization process. In the development stage, based on a large number of historical graph structure bases and subsequent real conflict data, the frequent "(node type A, spatial relationship R, node type B)" patterns and strong association rules of conflict occurrence are found out through association rule mining (such as Apriori algorithm), and these rules are coded into the model. The optimization process is to adjust the preset probability threshold and learning rate on a validation set to maximize the score of the knowledge graph in predicting future conflict events.

[0066] Through the above technical solutions, the potential and uncertain global risks perceived by the spatio-temporal graph neural network are effectively precipitated into the explicit, structured and deterministic knowledge that can be directly used by the rule engine in the conflict knowledge graph. This reverse updating mechanism enables the knowledge graph to absorb real-time situation awareness results and evolve itself, providing high-quality and high-timeliness input for the subsequent regional positioning module to accurately focus on the most threatening conflict core, and enhancing the system's continuous cognitive and active response capability to dynamically changing national space.

[0067] As shown in Figure 4 , the application further proposes that by analyzing the updated conflict knowledge graph, the specific steps of locating the propagation path of the key conflict and the core nodes on the path include: The system first extracts the global topological features of the updated conflict knowledge graph, calculates the betweenness centrality of each node in the updated conflict knowledge graph, and generates a node centrality list. Specifically, the betweenness centrality quantifies the importance of a node as a "bridge" in the network, and its calculation formula is: where denotes the betweenness centrality of node v, V is the set of all nodes in the graph, denotes the total number of shortest paths from node s to node t, and

[0069] denotes the number of these shortest paths passing through node v. The system efficiently calculates the betweenness values of all nodes through the Brandes algorithm to generate the node centrality list. Sort the node centrality list in descending order, and output the nodes with a centrality value exceeding a preset centrality threshold as the core node set . The preset centrality threshold is not a fixed value, but a dynamic adaptive method. For example, it can be set as a proportion of the maximum betweenness centrality of all nodes in the current graph, i.e. wherein, Max(Centrality betweenness of all nodes in current graph), scaling factor Obtained by analyzing historical data, usually set in the interval [0.6, 0.8], for example, take 0.7. This means that the system will filter out nodes with a betweenness centrality not lower than 70% of the highest node centrality. All nodes that meet this condition are aggregated into a core node set .

[0070] Subsequently, the system takes the nodes in the core node set as the starting point to perform a random walk algorithm (such as Personalized PageRank) to explore high-probability paths connecting multiple nodes in the set, and determines the high-probability paths as the key conflict propagation paths.

[0071] The transition probability of the walk is proportional to the weight associated with the edge of the graph (derived from the update of the previous graph enhancement module). The algorithm explores and evaluates paths connecting multiple core nodes in the set through multiple iterations of the walk, with the goal of finding high-probability paths that are frequently traversed and have high average weight of nodes on the path. The evaluation score of the path can be calculated by the following formula:

[0072]

[0073] where, is the weight of node u on the path, and L is the path length;

[0074] is the frequency of this path being visited during the random walk process. The system will ultimately determine the Top-N paths with an evaluation score exceeding a pre-set path threshold (which can be determined by a historical validation set) as the key conflict propagation paths.

[0075] The parameters of the random walk algorithm, such as walk step length and restart probability, can be considered as an offline optimization problem. During the model training phase, historical knowledge graph data and known conflict propagation paths are used as labels. Grid search or Bayesian optimization methods are used to maximize the score of the located paths compared to the true propagation paths, and the optimal parameter combination is found. Typical training conditions include: using a Bayesian optimization framework (such as Scikit-optimize) on a dataset containing thousands of historical graph instances, and finding the optimal solution after 50-100 iterations. In common optimization results, the restart probability is usually between 0.15 and 0.25.

[0076] By the technical solution, the application can intelligently penetrate complex network topology, identify key nodes and channels that truly control conflict information flow or risk diffusion, thereby liberating limited computing and regulatory resources from the heavy burden of "global coverage", accurately guiding to a few key areas and paths that have the greatest impact on the stability of the entire conflict network, improving the efficiency and pertinence of subsequent rule engine fine judgment, and providing a clear action roadmap for implementing precise and effective spatial intervention measures.

[0077] The application further proposes that the specific steps of calling the regulation rule subset matched with the characteristics of the core nodes and the propagation path include:

[0078] According to the land use type attribute of the core node, a first rule subset is filtered from the global rule library matched by the rule engine. Each core node in the knowledge graph has an explicit land use type attribute (such as "basic farmland", "urban construction land", "ecological protection zone", etc.). The system filters all rule clauses directly applicable to these land use types from the global rule library connected by the rule engine, which contains thousands of land space regulation rules, to form the first rule subset . For example, if the core node type is "basic farmland", the relevant rules related to "permanent basic farmland protection" and "non-agricultural non-food" are filtered out.

[0079] The system analyzes the topological and semantic features represented by the key conflict propagation path to determine the type of conflict diffusion it represents. The path features include the sequence of land use types of the nodes on the path, the types of spatial relationships (such as "adjacent to", "downstream direction", etc.). The system maps these features to predefined conflict diffusion types through a pre-trained path classification model, such as "adjacent pollution diffusion type", "upstream and downstream water competition type", "urban-rural fringe encroachment type", etc. According to the determined conflict diffusion type, the system again filters rule clauses from the global rule library that aim to prevent, control or solve this type of diffusion conflict, to form a second rule subset .

[0080] Finally, instead of simply combining the two subsets, the system calculates their intersection, i.e. , and takes this intersection as the final called regulation rule subset. This intersection operation ensures that the called rules are not only specific to the specific land properties of the core area, but also consistent with the current conflict evolution pattern, achieving dual constraints on rule applicability in "static attributes" and "dynamic behaviors".

[0081] The path classification model is a key intelligent component, and its training process is as follows: the model usually adopts a graph neural network or a sequence model (such as LSTM) architecture, takes the embedded representation (fusing node attribute sequence and edge type sequence) of the conflict propagation path as input, and outputs the probability distribution of the conflict diffusion type. The model is supervised trained on a dataset containing a large number of historical conflict propagation paths and their artificially labeled types. During training, the classification cross-entropy loss function is used, the Adam optimizer is used, the initial learning rate is set to 3e-4, the batch size is 32, and the training period is until the validation set accuracy converges (usually about 100-150 rounds). The evaluation index of the model usually pursues an accuracy rate higher than 85% to ensure that the selected second rule subset has high relevance.

[0082] Through the above technical solution, the application realizes the high contextualization and intelligence of rule calling. Through the cross filtering of the two dimensions of “node attribute” and “path pattern”, the range of rules required to be loaded for each rule reasoning is narrowed, not only reducing the real-time computing load of the rule engine and shortening the response time, but more importantly, it excludes a large number of irrelevant rules for the current conflict situation, making the focus of rule matching more concentrated, thereby directly improving the accuracy and reliability of the deterministic conflict identification output, and providing a solid logical foundation for the final formation of high-quality regulatory decisions.

[0083] The application further proposes that the specific steps of fusing the deterministic conflict identification and the potential conflict probability based on the fusion strategy include:

[0084] The system pairs the deterministic conflict identification (a binary or multi-class discrete symbol result) output by the reasoning execution module with the corresponding potential conflict probability (a continuous scalar value) in the regional positioning module, and inputs them into the conflict knowledge graph. Here, the system performs consistency checking, which aims to evaluate the consistency degree of the rule determination result and the probability prediction result in the objective environment described by the graph. The checking process is performed by analyzing the path attributes connecting these conflict nodes, the states of the neighbor nodes, and the outcomes in similar situations in the history. For example, if the rule determines a conflict in a plot, but the potential conflict probability value is very low, but the knowledge graph shows that the plot is surrounded by multiple confirmed high-risk conflict nodes, then the system will judge it as “inconsistent”; otherwise, it is “consistent”.

[0085] Based on the result of the above consistency checking, the system dynamically calculates the weight value of the potential conflict probability in the fusion . The weight value is not fixed, but is dynamically generated by a lightweight weight calculation function. The function takes the consistency degree, the potential conflict probability value itself, and the authority of the node in the knowledge graph (such as node weight) as input. The weight value The calculation formula can be expressed as:

[0086]

[0087] Wherein: represents the quantitative score of consistency check (e.g. 1 for complete consistency, -1 for complete inconsistency, 0 for partial consistency);

[0088] is the potential conflict probability value;

[0089] is the weight of the node in the conflict knowledge graph;

[0090] a, b, c are adjustable parameters for controlling the influence strength of different factors on the final weight, which are determined through subsequent optimization process; Parameters a, b, c are determined through offline training. The training process uses historical data, including deterministic conflict identification, potential conflict probability, knowledge graph snapshot, and finally manually verified true conflict status. The training goal is to minimize the cross-entropy loss between the comprehensive confidence Z and the true state (converted to 0 / 1). The training conditions are usually set as: using Adam optimizer, learning rate 1e-3, batch size 64, training until the loss function converges (usually about 50 cycles);

[0091] e is a normalization factor to ensure the weight is within a reasonable range;

[0092] The function is used to map the calculation result to the interval (0, 1).

[0093] Finally, the system calculates the dynamic weight based on the calculated dynamic weight , and the potential conflict probability to output a comprehensive confidence Z: The comprehensive confidence Z constitutes the core measure of each conflict item in the mixed conflict list, which contains both the explicit conclusion of the rule engine and the risk assessment of the probability model, forming a continuous scale between 0 and 1, which more carefully reflects the possibility of conflict occurrence.

[0094] Through the technical solution, the application creatively realizes adaptive fusion of rule determination and probability prediction in different situations. The limitations of single dependence on rules (which may be rigid) or single dependence on model prediction (which may be uncertain) are overcome. The contributions of the two are intelligently balanced through the context awareness ability of the knowledge graph. When the two are consistent, the result is strengthened. When the two are inconsistent, the system can give higher weight to the more reliable party according to the specific situation, thereby outputting a more robust, reliable, and information-rich decision basis, and improving the intelligent level of the auxiliary decision system and the reliability in actual complex applications.

[0095] The application further proposes that the mixed conflict list is sent to the execution terminal while also including:

[0096] First, according to the predefined instruction mapping rule, a conflict type instruction mapping table stored in the system database is called. This conflict type instruction mapping table is predefined and configured by domain experts according to laws and regulations, department responsibilities and administrative processes during the system deployment stage, and it establishes a mapping relationship from "conflict type" to specific "instruction item".

[0097] Subsequently, the system traverses each conflict record in the mixed conflict list and matches the corresponding instruction item from the above conflict type instruction mapping table according to its specific conflict type (for example, "illegal occupation of basic farmland", "violation of construction within ecological protection red line").

[0098] Control operation: clearly specifies specific law enforcement or management actions, such as "on-site investigation", "order to stop work", "rectification within a certain period";

[0099] Responsible subject: indicates the institution or role responsible for executing the operation, such as "county natural resources bureau law enforcement brigade", "city ecological environment branch";

[0100] Instruction format: specifies the standardized template or data structure of the instruction to ensure the integrity and consistency of the information, for example, a JSON schema containing "location coordinates", "conflict description", "legal basis", "disposal time limit", etc.

[0101] After the matching of the instruction item is completed, the system generates the final instruction. It fuses the specific content of the item conflict in the mixed conflict list (such as location information, comprehensive confidence, associated entity) with the matched instruction item, and automatically attaches the spatiotemporal label inherited from the original data or generated by the regional positioning module (for example, the boundary coordinates of the conflict polygon, the effective time range). Finally, the system sends the generated spatial control instruction containing operational information to the corresponding execution terminal (such as the smart terminal of the law enforcement personnel or the business system of the relevant department) through the government affair special network or the mobile communication network.

[0102] The construction and maintenance of the conflict type instruction mapping table is a process of combining expert knowledge and historical data optimization. The "training" is reflected in the continuous optimization of the mapping relationship after the initial table is put into use and the instruction execution effect data collected by the feedback module. For example, a statistical-based method can be used to regularly analyze the correlation between different "control operations" and the final "disposal success rate" under a specific conflict type. If it is found that the success rate of a certain operation in a specific situation is continuously below a preset threshold (for example, for the "cross-regional pollution diffusion" conflict, the success rate of the "single local rectification" instruction is less than 60%), the system will prompt the administrator to adjust the mapping table and modify the operation of this conflict type to "joint cross-regional law enforcement". This process is usually set as an offline optimization task performed once every quarter or every half year.

[0103] Through the above technical solutions, the abstract conflict analysis conclusion is automatically converted into a specific and clear executable instruction containing the responsible subject and the operation process, reducing the manual interpretation and task allocation steps required in the traditional way, not only improving the speed of regulatory response, but also avoiding information ambiguity through the standardized instruction format, ensuring the scientificity, standardization and traceability of the control action, and finally forming a complete intelligent closed loop from situation awareness to decision execution.

[0104] The application further proposes that after the mixed conflict list is sent to the execution terminal, it further includes:

[0105] Through the government affair data interface or the mobile law enforcement terminal log, the instruction execution effect data returned by the execution terminal is collected. These instruction execution effect data are structured and at least contain the "disposal state" (such as "rectified", "not rectified", "disposal in progress") and "disposal result verification information" (such as on-site review image, monitoring data change) of each instruction.

[0106] The instruction execution effect data is compared with the prediction of the mixed conflict list to generate an effect difference signal. Specifically, the "treatment state" is compared with the "comprehensive confidence" or the original "deterministic conflict identifier" in the list. For example, if a region that is determined by the system to be a conflict with high confidence and an instruction is issued, and if it is confirmed through on-site verification that there is no conflict, it indicates that the system has a false positive; on the contrary, if a region that is not identified by the system or has low confidence is subsequently proven to have a conflict, it indicates that the system has a false negative. This comparison will generate a quantitative effect difference signal . The signal can be formally defined as the difference between the prediction and the true result:

[0107]

[0108] where N is the number of conflict items in this comparison, is the comprehensive confidence of conflict item i, is the true label converted from the execution effect data (1 represents that there is indeed a conflict, and 0 represents that there is no conflict), and L is a loss function, such as a cross-entropy loss function . The effect difference signal macroscopically reflects the deviation of the current model prediction performance from the actual situation.

[0109] Subsequently, the system uses the effect difference signal as a supervision signal to optimize the parameters of the spatio-temporal graph neural network model through a backpropagation algorithm. The optimization process does not modify the online model in the production environment in real time, but in an isolated training environment, the model is incrementally trained or fine-tuned using accumulated new effect data. The specific steps include: extracting the historical graph structure base sequence corresponding to the current effect data from the historical database as the input feature, using the newly verified effect data as the new training label, and integrating the effect difference signal into the loss function , calculating the gradient through backpropagation, and updating the model weights using an optimizer :

[0110]

[0111] where is the learning rate, which is usually set to a small value, such as 5e-5, in this incremental learning scenario, to avoid catastrophic forgetting of the original knowledge by new data. This optimization task is usually set as a periodic offline task (e.g., performed during low-load periods every week or every month), and when the model's performance on the validation set is evaluated, it is updated to the online probability scheduling module.

[0112] By the technical solution, a complete adaptive optimization closed loop from perception, decision, execution to learning is established, so that the system can learn from the actual effect of each regulatory action. This parameter optimization mechanism based on real feedback can make the spatio-temporal graph neural network model continuously adapt to the dynamic changes of the land space conflict mode, gradually correct its prediction deviation, and thus present higher prediction accuracy and reliability in long-term operation, realizing the autonomous improvement and sustainable development of the system intelligence level.

[0113] The application further proposes that after comparing the instruction execution effect data with the prediction of the mixed conflict list to generate an effect difference signal, the following is further included:

[0114] Based on the effect difference signal, a causal attribution analysis is performed on the reasoning chain triggering this determination to generate an attribution result. The causal attribution analysis traces back to the state of the initial graph structure base from the historical conflict list and identifies the key graph substructure of the conflict knowledge graph that leads to the final determination effect. Specifically, the system starts a backtracking analysis of the complete reasoning chain triggering this determination based on the determination deviation (whether false positive or false negative) indicated by the effect difference signal. The logic of this causal attribution analysis is to trace back from the mixed conflict list that produces this determination to the updated conflict knowledge graph state relied on by the regional positioning module, and further back to the state of the initial graph structure base used by the probability scheduling module. On this backtracking path, the system uses a graph structure difference analysis algorithm to identify the key graph substructure of the conflict knowledge graph that plays a decisive role in this determination. These substructures usually manifest as one or more closely related nodes and edges forming a pattern, for example, a specific "arable land-adjacent to-construction land-transformed from-forest land" relationship chain, which frequently appears in history and is highly related to the determination result.

[0115] According to the attribution result, the reference weight of the historical record associated with the key graph substructure in the historical conflict database is dynamically adjusted; wherein the adjustment strategy follows a clear reinforcement learning principle: for the key graph substructure that contributes greatly to the positive effect (i.e. prediction accuracy), the reference weight of the corresponding historical record in the database is adjusted upwards; otherwise, for the substructure that contributes greatly to the negative effect (i.e. prediction error), the reference weight of the corresponding historical record is adjusted downwards. The specific value of weight adjustment determined by the effect difference signal and the participation degree of the substructure in this reasoning together, and its calculation formula can be expressed as:

[0116]

[0117] wherein: the effect difference signal The value range of reflects the direction and degree of deviation between prediction and reality (for example, 0 when accurate, negative when false positive, and positive when false negative);

[0118] is the participation or influence factor of the key graph substructure in this reasoning chain, which can be quantified by calculating the importance (such as betweenness centrality) of the nodes and edges in this substructure in the path involved in this decision.

[0119] is the learning rate parameter, which is used to control the magnitude of weight adjustment, and is usually set to a small value, such as 0.05, according to system stability requirements.

[0120] Subsequently, the weight of all historical records containing the key graph substructure in the historical conflict database is updated as follows:

[0121]

[0122] The weight is limited to a reasonable range, such as [0.5, 2.0], to avoid the excessive influence of individual cases.

[0123] The algorithm for identifying key graph substructures in causal attribution analysis, its "training" or tuning is reflected in the optimization of the graph similarity calculation threshold and the participation . This process is usually carried out during the system development phase, using a dataset containing a large number of decision cases annotated with causal relationships. By adjusting the similarity threshold of substructure matching (for example, the Jaccard similarity threshold from 0.7 to 0.9) through methods such as grid search, the consistency (Kappa coefficient) between attribution results and expert judgments is maximized. The optimization process may last for dozens of iterations until the parameter combination that best captures the causal relationship is found.

[0124] Through the above technical solutions, the application can automatically identify which historical experience patterns are reliable and have high predictive value, and which are prone to misjudgment, and through the dynamic adjustment of the weight, the more high-quality and reliable historical experience can play a greater influence in the future probability calibration and reasoning process. This mechanism makes the entire system not only a prediction tool, but also an intelligent agent that continuously learns from success and failure and evolves its internal knowledge system.

[0125] The following is a specific embodiment of the land space use regulation compliance intelligent identification and auxiliary decision system:

[0126] A county-level natural resources bureau applied a land space use regulation compliance intelligent identification and auxiliary judgment system in its jurisdiction to conduct dynamic supervision of the surrounding areas of the urban development boundary in the third quarter of 2023. The system accesses 0.5-meter resolution satellite remote sensing images (updated quarterly), land approval electronic certificate flow (real-time synchronization), and surface deformation data collected by Internet of Things monitoring equipment (sampling frequency 15 minutes / second) through multi-source data interfaces to build a dynamic graph structure base covering 1200 square kilometers. In the data preprocessing stage, the data management module performs atmospheric correction and geometric registration on the remote sensing images, and performs topological relationship verification on the approval data. The dynamic time warping algorithm (DTW) is used to unify the multi-source data to the reference time axis of September 15, 2023, generating an initial graph containing 832 spatial entity nodes and 1564 spatial relationship edges.

[0127] The probability scheduling module loads a pre-trained spatio-temporal graph neural network model (3 layers of graph convolution + 2 layers of gated temporal convolution) and outputs a global potential conflict probability distribution after inputting the graph structure base data. The model prediction shows that there is a high conflict risk in the northern industrial park edge area (initial probability 0.68). The system automatically calls the historical conflict database and retrieves the "industrial land illegal expansion encroaching on cultivated land" case that occurred in this area in 2021. The similarity of its historical conflict pattern vector to the current topological structure is 0.82. The probability distribution is calibrated through a weighted fusion algorithm: (where the historical pattern weight β = 0.7 and the model prediction weight α = 0.3) to generate the calibrated probability distribution.

[0128] The graph enhancement module sets a preset probability threshold = 0.72, and filters out 12 high-risk nodes (including 6 industrial plots, 4 cultivated land patches, and 2 ecological protection red line areas). Through the probability graph mapping model, it is converted into 18 groups of triples such as (industrial plot A, adjacent to, cultivated land B), (construction project C, located in, ecological protection red line D), and updates the conflict knowledge graph. The node weight is dynamically adjusted according to the formula , where γ = 0.2, and the cultivated land B node weight is increased from = 1.0 to = 1.0 + 0.2 × 0.85 = 1.17.

[0129] The regional positioning module calculates the node betweenness centrality of the updated graph and finds that the industrial plot numbered G-042 has a betweenness centrality of 0.63 (higher than the threshold value of 0.56), which is identified as a core node. The random walk algorithm is executed from this node as the starting point (restart probability 0.2), and after 500 iterations, the key propagation path is found: industrial plot G-042 → adjacent cultivated land → downstream irrigation canal → ecological wetland, with a path evaluation score = (1.17 + 1.08 + 0.95) / 3 x log (235 + 1) = 0.72 x 5.46 = 3.93 (higher than the path threshold 3.2), according to which a 15.2 square kilometer area to be refined is circled.

[0130] The reasoning execution module activates the rule engine, filters out the first rule subset such as the land management law and the ecological protection red line management method according to the core node "industrial land" attribute, matches the second rule subset in combination with the "adjacent encroachment type" conflict diffusion type, and finally calls 3 intersection rules. Through spatial logic matching, it is confirmed that the industrial plot G-042 has the behavior of exceeding the approved range of construction, and the deterministic conflict identifier (type: cultivated land encroachment, basis: land management law) is output.

[0131] The fusion decision module performs consistency checking and finds that the deterministic identifier is highly consistent with the potential conflict probability of 0.85 (consistency score C = 0.92). The comprehensive confidence is calculated according to the formula: , The mixed conflict list is generated, including 3 places with high confidence conflicts (Z > 0.85) and 2 places with medium confidence warnings (0.6 ≤ Z ≤ 0.85).

[0132] The feedback module automatically matches the conflict type instruction mapping table and sends a spatial control instruction to the county law enforcement brigade: "Carry out on-site investigation on the G-042 plot (responsible subject: law enforcement team one, time limit: before September 20, 2023)", and synchronously pushes the conflict area vector boundary, historical violation record and other attachments. After receiving the instruction through the mobile execution terminal, the law enforcement personnel complete the on-site check and upload the rectification results within 3 days. According to the feedback data, the system optimizes the model parameters through the back propagation algorithm, and the prediction accuracy of this type of conflict is improved from 82% to 89%.

[0133] In this supervision case, the system takes 4.5 hours from data collection to instruction issuance, which is 12 times more efficient than traditional manual inspection. By identifying the risk of industrial land expansion in advance, 3.2 mu of cultivated land is prevented from being illegally occupied, and relevant rectification measures maintain the compliance rate of regional ecological protection red line at 98.7%. Subsequent quarterly monitoring shows that the average response time of the system to similar conflicts is shortened to 2.8 hours, and the reference weight of the "industrial-cultivated land" conflict mode in the historical conflict database is dynamically adjusted to 0.81, achieving a coordinated improvement in supervision accuracy and efficiency.

[0134] The technical scope of the present application is not limited to the content in the above description, and those skilled in the art can make various modifications and modifications to the above embodiments without departing from the technical idea of the present application, and these modifications and modifications should all belong to the protection scope of the present application.

Claims

1. A smart identification and auxiliary judgment system for compliance of land use control, characterized in that: include: The data management module is used to receive multi-source heterogeneous time-series data of the target area, construct a dynamic graph structure base, and initialize a conflict knowledge graph and a historical conflict database based on the graph structure base. The probability scheduling module is used to input the graph structure basis into a pre-trained spatiotemporal graph neural network model, generate a probability distribution of potential conflicts in the whole domain, and call the historical conflict list from the historical conflict database to calibrate and generate a calibrated probability distribution. The graph enhancement module is used to output the portion of the probability value in the calibrated probability distribution that is higher than a preset probability threshold and update it in reverse to the conflict knowledge graph. The regional positioning module is used to analyze the updated conflict knowledge graph to locate the propagation path of key conflicts and the core nodes on the path, and generate the regions to be finely inspected and their potential conflict probabilities. The inference execution module is used to extract the region to be inspected in real time, activate the rule engine, call a subset of control rules that match the characteristics of the core node and the propagation path, perform logical matching, and output a deterministic conflict identifier. The fusion decision module is used to fuse the deterministic conflict identifier with the potential conflict probability based on the fusion strategy to generate a hybrid conflict list; The feedback module is used to send the mixed conflict list to the execution terminal.

2. The intelligent identification and auxiliary judgment system for compliance of land use control as described in claim 1, characterized in that: The specific steps for receiving multi-source heterogeneous time-series data from the target region and constructing a dynamic graph structure basis include: The multi-source heterogeneous time-series data is cleaned and time-aligned to generate a standardized data stream; The standardized data stream is fused based on a predefined land space ontology model to generate the graph structure base with unified spatiotemporal coding; Based on the entities and relationships defined in the aforementioned land space ontology model, the graph structure base is instantiated as nodes and edges; The graph structure base is continuously updated by attaching a timestamp version to each node and edge.

3. The intelligent identification and auxiliary judgment system for compliance of land use control as described in claim 1, characterized in that: The specific steps for retrieving the historical conflict list from the historical conflict database to generate the calibrated probability distribution include: Query conflict records with similar topological patterns to the current graph structure base from the historical conflict database, and generate a historical conflict pattern vector; The historical conflict pattern vector is weighted and fused with the probability distribution of potential conflicts in the global domain to output the calibrated probability distribution.

4. The intelligent identification and auxiliary judgment system for compliance of land use control as described in claim 1, characterized in that: The specific steps for outputting the portion of the calibrated probability distribution with probability values ​​higher than a preset probability threshold and updating the conflict knowledge graph in reverse include: The calibrated probability distribution is input into a pre-stored probability map mapping model; The probabilistic graph mapping model transforms high-probability conflict nodes and their spatial relationships into graph triples. The graph triples are added to the conflict knowledge graph, and the weights of the corresponding nodes are updated to complete the reverse update of the conflict knowledge graph.

5. The intelligent identification and auxiliary judgment system for compliance of land use control as described in claim 1, characterized in that: By analyzing the updated conflict knowledge graph, the specific steps to locate the propagation path of key conflicts and the core nodes along the path include: Calculate the betweenness centrality of each node in the updated conflict knowledge graph and generate a list of node centralities; Sort the node centrality list in descending order, and output the core node set as the nodes whose centrality values ​​exceed the preset centrality threshold. Starting from a node in the core node set, a random walk algorithm is executed to explore high-probability paths connecting multiple nodes in the set, and these high-probability paths are identified as critical conflict propagation paths.

6. The intelligent identification and auxiliary judgment system for compliance of land use control as described in claim 1, characterized in that: The specific steps for invoking a subset of control rules that match the characteristics of the core node and the propagation path include: Based on the land use type attribute of the core node, a first subset of rules is selected from the global rule base matched by the rule engine; Based on the conflict propagation type represented by the propagation path, a second subset of rules is selected from the global rule base; The intersection of the first rule subset and the second rule subset is selected as the final set of control rules to be invoked.

7. The intelligent identification and auxiliary judgment system for compliance of land use control as described in claim 1, characterized in that: The specific steps for fusing the deterministic conflict identifier with the potential conflict probability based on the fusion strategy include: The deterministic conflict identifier and the corresponding potential conflict probability are input together into the conflict knowledge graph for consistency verification; Based on the verification results, the weight value of the potential conflict probability in the fusion is dynamically calculated; Based on the weight values, the deterministic conflict identifier and the potential conflict probability are weighted and calculated to output the overall confidence level.

8. The intelligent identification and auxiliary judgment system for compliance of land use control as described in claim 1, characterized in that: Sending the hybrid conflict list to the execution terminal also includes: According to the predefined instruction mapping rules, the conflict type instruction mapping table is invoked; Based on the conflict type of each item in the hybrid conflict list, the corresponding instruction item is matched from the conflict type instruction mapping table; the instruction item includes at least the control operation, the responsible entity, and the instruction format; The hybrid conflict list is combined with the spatiotemporal markers of the instruction items to generate a space control instruction, which is then sent to the execution terminal.

9. The intelligent identification and auxiliary judgment system for compliance of land use control according to claim 8, characterized in that: After sending the hybrid conflict list to the execution terminal, the process also includes: Collect the instruction execution effect data returned by the execution terminal; The execution effect data of the instructions is compared with the prediction of the mixed conflict list to generate an effect difference signal; Based on the effect difference signal, the parameters of the spatiotemporal graph neural network model are optimized using the backpropagation algorithm.

10. The intelligent identification and auxiliary judgment system for compliance of land use control according to claim 9, characterized in that: After comparing the instruction execution effect data with the predictions of the mixed conflict list to generate an effect difference signal, the method further includes: Based on the effect difference signal, causal attribution analysis is performed on the reasoning chain that triggered this judgment, and attribution results are generated. The causal attribution analysis traces back from the historical conflict list to the state of the initial graph structure base and identifies the key graph substructures of the conflict knowledge graph that lead to the final judgment effect. Based on the attribution results, the reference weights of historical records associated with the key graph substructure in the historical conflict database are dynamically adjusted.

Images