An artificial intelligence-based territorial space governance analysis method and system

Abstract

This invention discloses an artificial intelligence-based method and system for analyzing land spatial governance, belonging to the field of land spatial planning technology. The method includes: acquiring multi-source land spatial data; dividing the target area into several spatial units and constructing a spatiotemporal dynamic graph structure; extracting node features using a graph neural network and inputting them into a time-series prediction model to predict the future state of land space; comparing the predicted state with spatial control objectives to identify areas and time periods with conflict risks; for conflict areas, using a reverse reasoning process to deduce a sequence of intervention measures from the target state back to the target state; and outputting a recommended governance plan. This invention integrates state prediction with governance decision-making, realizing a shift from passive, reactive governance to proactive, forward-looking governance. It can accurately identify spatial conflict risks, comprehensively weigh the multidimensional costs of intervention measures, and generate highly feasible governance plans.

Description

Technical Field

[0001] This invention relates to the field of land and space planning technology, specifically to a land and space governance analysis method and system based on artificial intelligence. Background Technology

[0002] In recent years, the development of remote sensing monitoring technology, geographic information systems, and artificial intelligence methods has provided new technical support for the dynamic perception and intelligent analysis of national land space status. Currently, the technical methods for analyzing national land space governance can be mainly divided into the following categories: One category is the traditional analysis method based on human experience and planning indicators. Planners overlay and analyze current land use data with various control boundaries, combine their own experience to judge spatial conflict risks, and formulate corresponding land use adjustment plans. Another category is the spatial evolution simulation method based on mechanistic models, such as using cellular automata models or land use change simulation models, to simulate the changing trends of future national land space patterns by setting transformation rules. A third category is the analysis method based on machine learning, which uses classifiers or regression models to predict the potential for construction land expansion or ecological risks in specific areas, and provides corresponding control recommendations based on the prediction results.

[0003] However, most existing methods separate spatial state prediction and governance decision-making into two independent stages. The prediction stage only provides a one-way output of the future state, lacking the ability to deduce current intervention measures from the desired target state. This results in governance solutions that are often reactive and lack forward-looking and precise regulation. Secondly, existing technologies generally lack in-depth modeling of the complex relationships between spatial units, making it difficult to capture the connectivity impact of ecological corridors and the synergistic or restrictive relationships between functional units. This leads to the easy omission of potential risks caused by indirect transmission effects when identifying conflict risks. Thirdly, existing prediction models mostly adopt static structures, which cannot be efficiently updated with the acquisition of new monitoring data. The models also lack self-reflection and correction mechanisms for their own prediction biases, and the prediction accuracy gradually decreases over time. In addition, during the generation of intervention measures, existing methods often lack a comprehensive balance between economic costs, ecological disturbances, and social acceptance, and also lack differentiated handling mechanisms for different types of conflicts. As a result, the generated solutions face the problem of insufficient feasibility when implemented in practice. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide an artificial intelligence-based method and system for analyzing land and space governance, in order to solve the technical problems in the prior art where spatial state prediction and governance decision-making are separated and there is a lack of forward-looking regulatory capabilities to deduce intervention measures from the target state.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-based method for analyzing land spatial governance includes the following steps: acquiring multi-source land spatial data for a target area, including time-series remote sensing imagery, geographic national condition monitoring data, land use status data, and ecological protection red line data; dividing the target area into several spatial units and constructing a spatiotemporal dynamic graph structure, where each spatial unit is a node, and node attributes include the land use type, ecological function level, and development intensity index of the spatial unit; edges between nodes are established based on spatial adjacency and ecological corridor connectivity; and using a graph neural network to extract node features of each spatial unit, and inputting the extracted node features into a time-series prediction algorithm. The model, a time-series prediction model, predicts the state of national land space at multiple future time points by capturing the changing patterns of node features along the time axis. It compares the predicted future national land space state with pre-set spatial control targets to identify areas and time periods with potential conflict risks. For the identified areas with potential conflict risks, a reverse reasoning process is used to deduce the sequence of intervention measures needed for the current stage, starting from the expected future target state. This reverse reasoning process employs a heuristic search algorithm, prioritizing intervention measures with the least impact on surrounding units and the lowest implementation cost in each step of the backtracking. The derived sequence of intervention measures is then output as a recommended governance plan.

[0006] In the technical solution of the present invention, the time series prediction model preferably adopts an encoder-decoder architecture. The encoder encodes the node features within the historical time window into a context vector, and the decoder uses the context vector as the initial state to gradually generate prediction results for future time nodes. A spatial attention mechanism is introduced during the decoding process, so that when the time series prediction model predicts the future state of each spatial unit, it automatically pays attention to the historical change information of other spatial units that have ecological or functional associations with the current spatial unit, thereby improving the spatial consistency of the prediction.

[0007] In the technical solution of this invention, the cost function of the heuristic search algorithm in the reverse reasoning process comprehensively considers three dimensions: the economic cost of the intervention measures, the degree of ecological disturbance, and social acceptance. The degree of ecological disturbance is quantified by simulating the spillover effect of the intervention measures on the ecosystem services of surrounding spatial units, while social acceptance is estimated based on the historical success rate of similar intervention measures in similar areas. Through this multi-dimensional cost assessment, a balance is achieved between feasibility, economy, and eco-friendliness in the derived sequence of intervention measures.

[0008] As a further improvement of this invention, after identifying areas and time periods with conflict risks, a conflict type discrimination step is included: classifying conflicts according to their causes into demand overload conflicts, spatial compression conflicts, and functional conflict conflicts. Demand overload conflicts refer to situations where the carrying capacity of a spatial unit exceeds the resource and environmental carrying capacity threshold; spatial compression conflicts refer to situations where ecological space is gradually encroached upon by construction land; and functional conflict conflicts refer to situations where the land use functions of adjacent spatial units mutually restrict each other. Different intervention measures are preferred for different types of conflicts during the reverse reasoning process. For example, ecological restoration or boundary control measures are prioritized for spatial compression conflicts, while land use function replacement or compatibility adjustment measures are prioritized for functional conflict conflicts, thereby achieving differentiated and precise governance.

[0009] As a further improvement of the present invention, it also includes a model introspection step: after each generated recommended governance scheme, the implementation effect of the recommended governance scheme in actual application is fed back to the time series prediction model. The time series prediction model compares the deviation between the predicted state and the actual state, and automatically adjusts the internal parameters of the time series prediction model according to the deviation value, so that the time series prediction model gradually corrects its cognitive bias in subsequent predictions. This mechanism enables the model to continuously learn from the actual application effect and avoids the prediction accuracy from decreasing over time.

[0010] In the technical solution of the present invention, the spatiotemporal dynamic graph structure is dynamically updated after each prediction. The update basis includes the newly added construction land change patches identified by remote sensing images, the ecological restoration project completion patches, and the land consolidation project acceptance patches. The affected nodes and edges are locally updated without rebuilding the entire graph through an incremental graph update algorithm, so as to keep the spatiotemporal dynamic graph structure synchronized with the real land space status and significantly reduce the computational cost of graph structure update.

[0011] In the technical solution of this invention, when the target area is divided into several spatial units, an adaptive grid partitioning strategy is adopted. That is, a high-resolution grid is used in ecologically sensitive areas and urban development boundaries, while a low-resolution grid is used in general agricultural areas and ecological buffer zones. Information transfer between grids of different resolutions is achieved by setting virtual boundary nodes. This strategy ensures the analysis accuracy of key areas while controlling the overall computational scale.

[0012] In the technical solution of the present invention, during the reverse reasoning process, when multiple intervention measures in the derived sequence of intervention measures conflict with the same spatial unit, a priority arbitration mechanism is introduced: according to the hierarchical order of ecological security priority, food security priority, and economic development priority, the execution order of conflicting intervention measures is automatically adjudicated or one of the intervention measures is selected for execution, and the abandoned intervention measures are recorded as alternative solutions and stored in the measure library for subsequent periods, thereby ensuring the internal consistency of the solution and retaining valuable alternative solutions.

[0013] This invention also provides an artificial intelligence-based land space governance analysis system for executing the above-described methods. The system includes: a data access module for acquiring and registering multi-source land space data; a spatiotemporal map construction module for dividing spatial units and constructing a spatiotemporal dynamic map structure; a prediction module, incorporating the aforementioned time-series prediction model, for predicting future land space status; a conflict identification module for comparing the predicted status with spatial control objectives to identify conflict areas and time periods; a reverse reasoning module for deriving a sequence of intervention measures from the target status; and an output module for outputting recommended governance solutions.

[0014] As a further improvement to the system, the system also includes a knowledge accumulation module, which records the combination of intervention measures tried in each reverse reasoning process and their corresponding implementation effects. When the similarity of the land use structure of the new area to be governed and the historical area exceeds a preset threshold, the system automatically retrieves the historically successful combination of intervention measures as the initial solution for reverse reasoning, thereby shortening the reasoning time and improving the reliability of the solution.

[0015] In summary, the present invention has the following main beneficial effects: This invention addresses the shortcomings of existing technologies, such as the disconnect between spatial state prediction and governance decision-making, and the lack of forward-looking regulatory capabilities to deduce intervention measures from the target state. By coupling the prediction model with a backward reasoning process, it achieves the reverse derivation of current intervention measures from the desired target state, transforming passive response-based governance into proactive and forward-looking governance. Furthermore, it addresses the problems of insufficient modeling of ecological and functional relationships between spatial units and the difficulty in identifying potential conflict risks caused by indirect transmission effects in existing technologies. By constructing a spatiotemporal dynamic graph structure and introducing a spatial attention mechanism, this invention fully captures the complex relationships between spatial units, significantly improving the comprehensiveness and accuracy of conflict risk identification. Finally, it addresses the lack of dynamic updates and self-correction in existing prediction models. To address the shortcomings of existing technologies that lead to a gradual decline in prediction accuracy due to positive mechanisms, this invention employs a model introspection step and an incremental graph dynamic update mechanism. This allows the model to continuously learn from actual implementation results and correct its own biases, effectively maintaining prediction accuracy. Furthermore, to address the deficiencies in existing technologies where intervention measures lack a comprehensive consideration of economic costs, ecological disturbances, and social acceptance, and are difficult to differentiate between different types of conflicts, this invention constructs a multi-dimensional cost function and introduces conflict type discrimination and priority arbitration mechanisms during the reverse reasoning process. This ensures that the generated governance solutions achieve a balance between feasibility, economy, and ecological friendliness, while simultaneously realizing differentiated and precise governance for different conflict causes, significantly improving the feasibility and overall benefits of the solutions. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall method of the present invention; Figure 2 This is a schematic diagram illustrating the reverse reasoning and conflict arbitration of the present invention; Figure 3 This is a diagram of the model introspection and dynamic graph update mechanism of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] The embodiments of the present invention will now be described.

[0019] This invention provides an artificial intelligence-based method for analyzing land and space governance. Its core concept lies in organically integrating land and space status prediction with governance decision-making. By constructing a spatiotemporal dynamic graph structure, it fully captures the complex relationships between spatial units. It uses a time-series prediction model to make forward-looking predictions of future states. Based on this, it adopts a reverse reasoning process to deduce current intervention measures from the expected target state. Through a series of mechanisms such as model self-reflection, dynamic graph updating, conflict type identification, multi-dimensional cost assessment, and priority arbitration, it ensures that the generated governance plan is forward-looking, accurate, feasible, and has optimal comprehensive benefits.

[0020] The technical solution of the present invention will be described in detail below with reference to specific implementation steps. For example... Figure 1 As shown, the overall flow of the method provided by this invention is as follows: Obtaining multi-source land space data: In implementing this invention, it is first necessary to acquire multi-source land space data for the target area. This data includes time-series remote sensing imagery, geographic national condition monitoring data, land use status data, and ecological protection red line data. Time-series remote sensing imagery refers to high-resolution optical images or synthetic aperture radar images covering different time points of the target area, used to extract information on changes in land cover types. Geographic national condition monitoring data includes basic geographic information such as topography, vegetation cover, and water system distribution. Land use status data originates from the annual change survey results of the natural resources department and includes attributes such as land use classification, ownership information, and area for each plot. Ecological protection red line data consists of boundary data for ecologically important areas and ecologically sensitive and fragile areas delineated according to law, serving as a key constraint for subsequent conflict identification.

[0021] Data preprocessing: In practice, after acquiring the aforementioned multi-source data, data preprocessing is required. Preprocessing includes coordinate system 1, spatial registration, data cleaning, and format conversion. Coordinate system 1 refers to unifying all data to the same coordinate reference system, such as using the CGCS2000 National Geodetic Coordinate System. Spatial registration involves geometrically correcting imagery and vector data from different sources and at different resolutions to ensure that pixels or patches at the same geographical location can be accurately superimposed in space. Data cleaning includes removing outliers, processing duplicate records, and correcting obvious logical errors. Format conversion converts data from different formats into a unified file format for easier subsequent processing.

[0022] Dividing space into units: After data preprocessing, the target area needs to be divided into several spatial units. Spatial units are the basic granularity for modeling and analysis in this invention. The division can be done using regular grid units, such as dividing the target area into 100m x 100m square grid units according to a preset size; or using administrative units, such as townships or streets as the basic units. In practical applications, to balance computational efficiency and analytical accuracy, an adaptive grid division strategy is preferred. This involves using high-resolution grids (e.g., 50m x 50m) in ecologically sensitive areas and urban development boundaries, and low-resolution grids (e.g., 200m x 200m) in general agricultural areas and ecological buffer zones. Information transfer between grids of different resolutions is achieved by setting virtual boundary nodes. Specifically, at the junction of high-resolution and low-resolution grids, a virtual boundary node is set as an information relay station. Information from the high-resolution grid is first aggregated to the virtual boundary node, and then transmitted from the virtual boundary node to the adjacent low-resolution grid, and vice versa, thereby achieving multi-scale information fusion.

[0023] Constructing a spatiotemporal dynamic graph structure: A spatiotemporal dynamic graph structure is constructed, with each spatial unit as a node. The attributes of each node include the land use type, ecological function level, and development intensity index of that spatial unit. Land use type refers to the land use classification of the unit, such as cultivated land, forest land, grassland, water area, construction land, and unused land. Ecological function level is a classification based on the results of ecosystem service value assessments; for example, after comprehensively assessing functions such as water conservation, soil retention, and biodiversity protection, it is divided into Level 1, Level 2, and Level 3. The development intensity index is a weighted composite value of the proportion of construction land within the unit, as well as indicators such as building density and plot ratio.

[0024] Edges between nodes are established based on spatial adjacency and ecological corridor connectivity. Spatial adjacency refers to two spatial units being geographically adjacent, i.e., sharing a boundary or corner. Ecological corridor connectivity refers to the existence of ecological flow channels between two spatial units. For example, if two forest patches are connected by a forest belt or water system, then there is ecological corridor connectivity between these two units. For two nodes with spatial adjacency, an edge is established between them, and the weight of the edge can be set to the shared length of the boundaries of the two units. For two nodes with ecological corridor connectivity, an edge is established between them, and the weight of the edge can be set to a combined value of the width, length, and vegetation cover of the ecological corridor.

[0025] Graph neural network feature extraction: After constructing the spatiotemporal dynamic graph structure, a graph neural network is used to extract the node features of each spatial unit. A graph neural network is a deep learning model specifically designed for processing graph-structured data, effectively capturing the dependencies between nodes. In this invention, a preferred architecture is a combination of graph convolutional networks and graph attention networks.

[0026] The role of graph convolutional layers is to aggregate neighborhood information for each node. Specifically, for each node, its own feature vector is weighted and summed with the feature vectors of all its neighbors, with the weights determined by the graph's adjacency matrix. After one layer of graph convolution, the new features of each node incorporate information from its direct neighbors; after multiple layers of graph convolution, the features of each node can incorporate information from multi-hop neighborhoods, thereby capturing a wider range of spatial relationships.

[0027] The graph attention layer further enhances the flexibility of feature extraction. It employs a multi-head attention mechanism, calculating the attention weight between each node and each of its neighboring nodes. This weight reflects the importance of neighboring nodes to the current node. When calculating attention weights, not only the node's own feature vector is considered, but also the geographical distance, functional similarity, and ecological association between nodes. Geographical distance can be obtained by calculating the Euclidean distance from the center points of two spatial units; functional similarity can be obtained by comparing the similarity of attributes such as land use type and economic output structure between two units; ecological association can be obtained by the connectivity strength between two units in the ecological network. These three dimensions of information are weighted and fused to obtain the final attention weight, which is used to weighted aggregate the features of neighboring nodes. The multi-head attention mechanism refers to running multiple independent attention calculations in parallel, concatenating or averaging the results as the final output, thereby enhancing the model's expressive power and stability.

[0028] By alternating stacking of graph convolutional layers and graph attention layers, a deep spatial feature representation of each node is finally obtained. This representation fully integrates the node's own attributes, the attributes of neighboring nodes, and the spatial and ecological association information between nodes.

[0029] Time series forecasting models predict future states: The node features extracted by the graph neural network are input into a time-series prediction model, which is used to capture the changing patterns of node features over time and predict the state of national land space at multiple future time points.

[0030] In a preferred embodiment of the present invention, the time-series prediction model employs an encoder-decoder architecture. The encoder is a recurrent neural network or a long short-term memory network used to encode the node feature sequence within a historical time window into a fixed-length context vector. The length of the historical time window can be set according to the time span of the data and the prediction requirements, for example, selecting data from the past five years as the historical window. The encoder sequentially reads the node features at each time point, updates its internal state, and finally compresses all historical information into the context vector.

[0031] The decoder is also a recurrent neural network or long short-term memory network. It uses the context vector output by the encoder as its initial state and then progressively generates predictions for future time points. During the decoding process, the decoder outputs the predicted value at each time step and uses this predicted value as the input for the next time step, thus autoregressively generating the entire future time series. The number of time points for prediction can be set according to actual needs; for example, predicting the spatial status of the land over the next three years, with each year as a time point.

[0032] To improve the spatial consistency of predictions, a spatial attention mechanism is introduced during the decoding process. Specifically, when the decoder predicts the future state of a spatial unit, in addition to utilizing the unit's own historical information, it automatically pays attention to the historical changes of other spatial units that have ecological or functional connections with the current spatial unit through the spatial attention mechanism. The calculation method for spatial attention weights is similar to that of the aforementioned graph attention layer, determining the attention weights based on the geographical distance, functional similarity, and ecological connection between spatial units. In this way, if a neighboring unit has experienced significant expansion of construction land in the past few years, this change information will be transmitted to the prediction of the current unit through the spatial attention mechanism, making the prediction results more consistent with the spatial linkage patterns of national land spatial evolution.

[0033] Compared with space control objectives: By comparing the predicted future state of national land space with pre-set spatial control targets, areas and time periods with potential conflict risks are identified. These spatial control targets are derived from binding indicators issued by higher-level national land space plans, such as targets for arable land preservation, permanent basic farmland protection area, ecological protection red line area, urban development boundary scale, and carbon emission intensity per unit of GDP. The predicted future state is compared item by item with these targets. If the predicted state of a certain area at a certain time point exceeds the control target threshold, that area and time period are marked as having a conflict risk.

[0034] Identify areas and time periods where conflict risks exist: The above comparison clearly identifies which areas are at risk of violating control objectives at which future time points. The output of this step is a list of conflict areas and time periods, providing a foundation for subsequent conflict type identification and reverse reasoning.

[0035] Conflict type identification: After identifying the areas and time periods where conflict risks exist, the present invention further includes a conflict type discrimination step, which classifies conflicts into three types according to their causes: demand overload conflict, space compression conflict, and functional conflict.

[0036] Demand overload conflict refers to a situation where the carrying capacity of a spatial unit exceeds the resource and environmental carrying capacity threshold. For example, if the predicted urban construction land area of ​​a region exceeds the maximum urban construction land area allowed by the region's water resource carrying capacity, or if the predicted pollutant emissions of a region exceed the region's environmental capacity threshold, then it is classified as a demand overload conflict. Resource and environmental carrying capacity thresholds can be calculated using specialized assessment models, such as water resource carrying capacity evaluation models and atmospheric environmental capacity models.

[0037] Spatial compression conflict refers to the gradual encroachment of ecological space by construction land. Specifically, it involves analyzing changes in the ecological space boundary under predicted conditions. If the area of ​​ecological land within the ecological protection red line decreases, or if important ecological corridors outside the ecological protection red line are cut off by construction land, it is classified as a spatial compression conflict. This type of conflict typically manifests as a gradual expansion of construction land from the periphery of urban areas to the edge of ecological space, exhibiting clear spatial continuity characteristics.

[0038] Functional conflict refers to the mutual constraint of land use functions between adjacent spatial units. For example, if an area is planned for residential use, but its adjacent units are projected to be highly polluting industrial land, and there is a lack of effective protective buffer between the two, then it is considered a functional conflict. Another example is a high-standard farmland area, but its adjacent units are projected to be large-scale livestock farms. The pollutant emissions from these farms may cause non-point source pollution to the farmland, which also falls under the category of functional conflict.

[0039] Reverse reasoning process: For areas identified as having conflict risks, this invention employs a reverse reasoning process, starting from the desired future target state and working backward to derive the sequence of intervention measures needed for the current stage. For example... Figure 2As shown, the reverse reasoning process specifically includes: First, taking the desired target state as the starting point, a reverse search is conducted to find candidate intervention measures that can gradually bring the current state closer to the target state; in each backtracking step, a heuristic search algorithm is used to evaluate the comprehensive cost of each candidate measure (considering economic costs, ecological disturbance, and social acceptance), and the measure with the lowest cost is selected first; when multiple derived intervention measures conflict on the same spatial unit, a priority arbitration mechanism is triggered, and the execution order or the selected measures are automatically decided according to the hierarchical order of ecological security priority, food security priority, and economic development priority, ultimately generating a complete sequence of intervention measures.

[0040] The basic idea behind backward reasoning is to consider the future state of the conflict zone as an undesirable state and the space control objective as an desired state. Then, starting from the desired state, it works backward to deduce what measures are needed to transition from the current state to the desired state. This process is similar to backward search in path planning, but unlike forward planning, backward reasoning prioritizes the measures most critical to achieving the final goal, avoiding blind trial and error.

[0041] The reverse reasoning process employs a heuristic search algorithm. A heuristic search algorithm is a method that uses heuristic information within a search space to guide the search direction, effectively reducing the search space and improving search efficiency. In this invention, a state space is constructed, where each state represents a set of spatial units' land use type, development intensity, and ecological function configuration. The initial state is the current actual state, and the target state is the desired state that satisfies spatial control objectives. Starting from the target state, the reverse search considers which measures can be taken to gradually approach the target state from the current state. In each step of the reverse search, all possible intervention measures are considered, and the backtracking cost of each measure is evaluated.

[0042] Multi-dimensional cost assessment: The cost function of the heuristic search algorithm comprehensively considers three dimensions of intervention measures: economic cost, degree of ecological disturbance, and social acceptance. Economic cost refers to the funds required to implement the intervention measure, including land acquisition compensation, engineering construction and maintenance costs, and ecological restoration costs. The degree of ecological disturbance is quantified by simulating the spillover effects of the intervention measure on the ecosystem service functions of surrounding spatial units. Specifically, an ecosystem service function assessment model is established, the specific content of the intervention measure is input, and the impact on the ecosystem service value of surrounding units after implementation is simulated; the greater the impact, the higher the degree of ecological disturbance. Social acceptance is estimated based on the historical success rate of similar intervention measures in similar areas. Historical success rates can be obtained by reviewing local government land consolidation records and public participation feedback reports. If a certain type of measure has high public satisfaction and low resistance in past implementations, its social acceptance is high, which is reflected as a smaller cost term in the cost function.

[0043] The three dimensions mentioned above are weighted and summed to obtain the comprehensive cost of each intervention measure. The weights can be dynamically adjusted according to the actual situation; for example, when the ecological protection pressure is high at present, the weight of the degree of ecological disturbance can be increased. In each step of the backtracking, the intervention measure with the lowest comprehensive cost is selected first, thereby gradually deriving a complete sequence of intervention measures from the current state to the target state.

[0044] Generate a sequence of intervention measures: Through the aforementioned reverse reasoning process and multi-dimensional cost evaluation, a complete sequence of intervention measures from the current state to the target state is ultimately generated. Each intervention measure in this sequence corresponds to a specific spatial unit and specific implementation content.

[0045] Priority arbitration: During the reverse reasoning process, a situation may arise where multiple intervention measures are simultaneously derived for a given spatial unit, and these measures conflict with each other. For example, a spatial unit may be simultaneously required to implement two contradictory intervention measures: returning farmland to forest and developing construction land. To address this issue, this invention introduces a priority arbitration mechanism, such as... Figure 2 As shown.

[0046] The priority arbitration mechanism automatically adjudicates the order of execution of conflict intervention measures or selects one intervention measure to be implemented, following a hierarchical order of priority: ecological security, food security, and economic development. Specifically, it first determines whether the conflict measures involve ecological security bottom lines, such as intervention measures within ecological protection red lines. If so, ecological security-related measures are implemented first, while other measures are temporarily shelved or adjusted. If the conflict measures do not involve ecological security bottom lines, it further determines whether they involve food security bottom lines, such as intervention measures within permanent basic farmland. If so, food security-related measures are implemented first. If the conflict measures involve neither ecological nor food security bottom lines, the principle of prioritizing economic development is followed, selecting the measure that contributes more to economic development, or implementing the two measures in stages according to time sequence, implementing the measure that contributes more to economic development first, followed by the other measure.

[0047] For interventions that are discarded, they are not discarded as invalid information, but are recorded as alternatives and stored in the measure library for use in subsequent periods. For example, a measure that was not adopted in the current period due to priority may become the optimal measure after conditions change in the next period. In this case, it can be directly retrieved from the measure library, avoiding recalculation.

[0048] Output recommended governance solutions: The sequence of intervention measures determined after priority arbitration will be output as the recommended governance plan. The output may take the form of a written report, a table, or a visual map.

[0049] Feedback on the implementation results of the plan: To ensure that the time series prediction model can maintain its prediction accuracy over the long term, this invention also includes a model introspection step. For example... Figure 3 As shown, the specific mechanism of this step is as follows: after each generated and implemented recommended governance scheme, the implementation effect of the scheme in actual application is fed back to the time series prediction model; the model compares the deviation between the predicted state and the actual state, and automatically adjusts its internal parameters according to the deviation value, so as to gradually correct its cognitive bias in subsequent predictions. Meanwhile, Figure 3 The process of updating the dynamic map is also shown: based on the newly added construction land change patches, ecological restoration project completion patches, and land consolidation project acceptance patches identified by remote sensing images, an incremental map update algorithm is used to locally update the affected nodes and edges without rebuilding the entire map, so that the spatiotemporal dynamic map structure keeps in sync with the real land space status.

[0050] Specifically, after each recommended governance plan is generated and implemented, the implementation effect of the plan in actual application is fed back to the time-series prediction model. The implementation effect includes the actual implementation of governance measures, the changes in the state of national land space after governance, and whether the expected goals have been achieved.

[0051] Model introspection and dynamic graph updates: The time-series forecasting model compares the observed actual state during implementation with the previously predicted state and calculates the deviation between the two. The deviation can be measured by metrics such as root mean square error and mean absolute error.

[0052] Based on the deviation value, the time series prediction model automatically adjusts its internal parameters, allowing the model to gradually correct its cognitive biases in subsequent predictions. Parameter adjustment can be achieved using incremental learning or online learning algorithms, that is, by using newly acquired observation data to slightly update the model parameters without completely retraining the model. This introspective mechanism enables the model to continuously learn from real-world application results, preventing prediction accuracy from declining over time.

[0053] Meanwhile, the spatiotemporal dynamic graph structure is also dynamically updated after each prediction. The update criteria include newly added construction land change patches identified from remote sensing imagery, completed ecological restoration project patches, and land consolidation project acceptance patches. These update criteria reflect the dynamic changes in the actual national land space. To reduce update overhead, this invention employs an incremental graph update algorithm, that is, updating only the affected nodes and edges locally without reconstructing the entire map. Specifically, for newly added construction land change patches, the spatial unit in which they are located is identified, and the land use type in the node attributes of that unit is updated to construction land, and the development intensity index is updated accordingly; for completed ecological restoration project patches, the spatial unit in which they are located is identified, and the ecological function level of that unit is upgraded accordingly; for land consolidation project acceptance patches, the land use type and attributes of the relevant units are updated according to the consolidation content. If the attribute of one of the two nodes connected by an edge changes significantly, the weight of that edge may also need to be recalculated. Through incremental updates, the spatiotemporal dynamic graph structure is kept synchronized with the actual national land space state, while significantly reducing the computational overhead of graph structure updates.

[0054] Preferences for differentiated intervention measures for different types of conflicts: To improve the targeting and effectiveness of intervention measures, this invention sets different intervention measure preferences for different types of conflicts.

[0055] For demand overload conflicts, since their essence lies in exceeding the carrying capacity threshold of resources and the environment, the preferred intervention measures are "reduction" and "efficiency improvement." Reduction measures include strictly controlling the scale of new construction land, orderly withdrawing illegal construction land within ecological red lines, and restricting the layout of high water-consuming and high-emission industries. Efficiency improvement measures include promoting the redevelopment of existing construction land, improving the efficiency of industrial land use, and promoting water-saving and energy-saving technologies. In the reverse reasoning process, for demand overload conflicts, heuristic search algorithms assign lower cost weights to reduction and efficiency improvement measures, making them more likely to be prioritized.

[0056] For spatial encroachment-type conflicts, since their essence is the gradual erosion of ecological space by construction land, the preferred intervention measures are "boundary delineation" and "restoration." Boundary delineation measures include strengthening the rigid constraints of ecological protection red lines, establishing ecological buffer zones between urban development boundaries and ecological spaces, and constructing ecological isolation zones. Restoration measures include implementing measures such as returning farmland to forests and grasslands, connecting ecological corridors, and ecological restoration of mines. In the reverse reasoning process, for spatial encroachment-type conflicts, heuristic search algorithms assign lower cost weights to boundary delineation and restoration measures, making them more likely to be prioritized.

[0057] For functional conflict types, since their essence lies in the mutual constraints of land use functions between adjacent units, the preferred intervention measures are "replacement" and "compatibility." Replacement measures refer to adjusting and replacing mutually restrictive land use types, such as relocating polluting enterprises near residential areas to industrial zones. Compatibility measures refer to using technological means to make the originally conflicting functions compatible, such as constructing protective forest belts between residential and industrial areas, implementing clean production transformation in industrial areas, and constructing ecological interception ditches between farmland and livestock farms. In the reverse reasoning process, for functional conflict types, heuristic search algorithms assign lower cost weights to replacement and compatibility measures, making them more likely to be prioritized.

[0058] Knowledge accumulation and reuse of historical experience: To further improve analysis efficiency, the artificial intelligence-based land space governance analysis system provided by this invention also includes a knowledge accumulation module. This module records the combination of intervention measures tried in each reverse reasoning process and their corresponding implementation effects. Specifically, the recorded content includes: the geographical location of the target area, land use structure characteristics, conflict type, sequence of intervention measures adopted, comprehensive costs, and actual effects after implementation.

[0059] When a new area to be addressed needs analysis, the knowledge accumulation module first calculates the land use structure similarity between the new area and historical areas. Land use structure similarity can be calculated by comparing indicators such as the proportion of various land use types, spatial pattern characteristics, and ecological function level distribution between the two areas, for example, using cosine similarity or Euclidean distance. When the similarity exceeds a preset threshold, a combination of historically successful intervention measures is automatically retrieved as the initial solution for reverse inference. This means that in the reverse search process for a new area, the heuristic search algorithm no longer starts from scratch, but uses a combination of historically successful measures as the initial solution, and then performs local optimization and adjustments based on this, thereby significantly shortening the inference time. Simultaneously, since the feasibility of historically successful measures has been verified in practice, using them as the initial solution also improves the reliability of the generated solutions for the new area.

[0060] After retrieving historical measures, fine-tuning is necessary through intervention experiments in a small number of target areas. This is because even with similar land use structures, two regions may still differ in terms of topography, climate, and socioeconomic conditions. Fine-tuning involves selecting a small number of representative spatial units within the target area, implementing the core measures from the historical measures combination, observing the actual response, and adjusting the parameters in the measures combination appropriately based on the response results to better adapt it to local conditions.

[0061] The specific composition of the system: To implement the above method, this invention also provides an artificial intelligence-based land space governance analysis system. This system includes a data access module, a spatiotemporal map construction module, a prediction module, a conflict identification module, a reverse reasoning module, an output module, and a knowledge accumulation module.

[0062] The data access module is used to acquire and register multi-source land spatial data. This module provides a standardized data interface, enabling access to data sources of different formats and origins, and automatically performing coordinate transformation, format unification, and data verification.

[0063] The spatiotemporal graph construction module is used to divide spatial units and construct a spatiotemporal dynamic graph structure. This module integrates an adaptive grid partitioning algorithm and a graph construction algorithm, which can automatically determine the optimal grid partitioning scheme based on the spatial characteristics of the input data, calculate the spatial adjacency relationships between nodes and the connectivity of ecological corridors, and generate a spatiotemporal dynamic graph structure.

[0064] The prediction module incorporates a time-series prediction model to forecast future national spatial states. This module includes the implementation of the encoder-decoder architecture, the computation of the spatial attention mechanism, and the execution of the model introspection step. The prediction module can take the current spatiotemporal dynamic graph structure as input and output predicted states for multiple future time points.

[0065] The conflict identification module compares the predicted state with the spatial control objectives to identify conflict areas and time periods. This module stores the spatial control objectives issued by higher authorities and can automatically compare them item by item, and call the conflict type discrimination logic to classify the identified conflicts into demand overload type, space compression type, or functional conflict type.

[0066] The reverse reasoning module is used to deduce the sequence of interventions from the target state. This module implements a heuristic search algorithm and includes a built-in multi-dimensional cost evaluation function and intervention preference settings for different types of conflict. The module also contains priority arbitration logic to handle conflicts between interventions.

[0067] The output module is used to output recommended governance solutions. The output module can present the recommended governance solutions in the form of text reports, tabular lists, or visual maps, including specific intervention measures, implementation sequence, expected effects, and comprehensive cost assessment for each spatial unit.

[0068] The knowledge accumulation module records the combinations of interventions tried during each reverse reasoning process and their corresponding effects, and automatically performs similarity matching and knowledge transfer when analyzing new regions. This module enables the system to continuously learn and accumulate experience, and the system's analysis efficiency continuously improves with increased usage.

[0069] In practical applications, the technical solution of this invention can be deployed on the computing platform of natural resources authorities or land and space planning units. Operators first import multi-source land and space data of the target area into the system, which automatically completes preprocessing and spatial unit division. Subsequently, the system automatically constructs a spatiotemporal dynamic graph structure, extracts features using graph neural networks, and predicts the land and space status at multiple future time points using a time-series prediction model. The system compares the prediction results with the control targets, automatically identifies areas and time periods with conflict risks, and determines the conflict type. Next, the system initiates a reverse reasoning process, combining intervention measures preferences for different types of conflicts to deduce the sequence of intervention measures from the target state backwards. During the derivation process, the system calls the knowledge accumulation module to search for historical successful cases with similar land use structures. If found, the system retrieves the combination of measures as the initial solution; otherwise, it starts searching from scratch. After the derivation is complete, the system outputs a recommended governance plan, including specific intervention measures, implementation sequence, and expected effects. After the plan is implemented, the system obtains feedback on the implementation effect, performs model self-reflection steps and dynamic graph updates, and continuously optimizes subsequent prediction accuracy and analysis results.

[0070] Although embodiments of the present invention have been shown and described, these specific embodiments are merely explanations of the invention and are not intended to limit it. The specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. After reading this specification, those skilled in the art may make modifications, substitutions, and variations to the embodiments as needed without departing from the principles and spirit of the invention, but such modifications, substitutions, and variations are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A land space governance analysis method based on artificial intelligence, characterized in that, Includes the following steps: Acquire multi-source territorial spatial data for the target area; The target area is divided into several spatial units to construct a spatiotemporal dynamic graph structure; The node features of each spatial unit are extracted using graph neural networks, and the extracted node features are input into a time series prediction model to predict the state of national land space at multiple future time points. By comparing the predicted future state of national land space with pre-set spatial control targets, areas and time periods with potential conflict risks can be identified. For areas identified as having conflict risks, a reverse reasoning process is used to deduce the sequence of intervention measures that need to be implemented in the current stage by starting from the expected future target state and working backward. The derived sequence of intervention measures will be output as a recommended governance plan.

2. The artificial intelligence-based land space governance analysis method according to claim 1, characterized in that, The time-series prediction model adopts an encoder-decoder architecture. The encoder encodes the node features within the historical time window into a context vector, and the decoder uses the context vector as the initial state to gradually generate prediction results for future time nodes. During the decoding process, a spatial attention mechanism is introduced, enabling the time-series prediction model to automatically pay attention to the historical change information of other spatial units that have ecological or functional connections with the current spatial unit when predicting the future state of each spatial unit. The time-series prediction model predicts the state of national land space at multiple future time nodes by capturing the changing patterns of node features on the time axis.

3. The artificial intelligence-based land space governance analysis method according to claim 1, characterized in that, In the reverse reasoning process, the cost function of the heuristic search algorithm comprehensively considers three dimensions: the economic cost of the intervention, the degree of ecological disturbance, and social acceptance. The degree of ecological disturbance is quantified by simulating the spillover effect of the intervention on the ecological service functions of surrounding spatial units, and social acceptance is estimated based on the historical success rate of similar interventions in similar areas. The reverse reasoning process uses a heuristic search algorithm to prioritize the intervention with the least impact on surrounding units and the lowest implementation cost in each backtracking step.

4. The artificial intelligence-based land space governance analysis method according to claim 1, characterized in that, After identifying the areas and time periods with conflict risks, a conflict type identification step is also included: the conflicts are classified according to their causes into demand overload conflicts, space compression conflicts, and functional conflicts. The demand overload type conflict refers to the carrying capacity of a spatial unit exceeding the resource and environmental carrying capacity threshold; the spatial compression type conflict refers to the gradual encroachment of ecological space by construction land; and the functional conflict type conflict refers to the mutual constraint of land use functions between adjacent spatial units. Different intervention measures are preferred in the reverse reasoning process for different types of conflicts. The multi-source land space data includes time-series remote sensing images, geographic national condition monitoring data, land use status data, and ecological protection red line data.

5. The artificial intelligence-based land space governance analysis method according to claim 1, characterized in that, After the recommended governance solution is output, a model self-reflection step is also included: the implementation effect of the recommended governance solution in actual application is fed back to the time series prediction model, the time series prediction model compares the deviation between the predicted state and the actual state, and automatically adjusts the internal parameters of the time series prediction model according to the deviation value, so that the time series prediction model gradually corrects its own cognitive bias in subsequent predictions.

6. The artificial intelligence-based land space governance analysis method according to claim 1, characterized in that, The spatiotemporal dynamic graph structure is dynamically updated after each prediction. The update is based on newly added construction land change patches identified by remote sensing imagery, completed ecological restoration project patches, and land consolidation project acceptance patches. An incremental graph update algorithm is used to locally update affected nodes and edges without reconstructing the entire map, keeping the spatiotemporal dynamic graph structure synchronized with the actual land space status. Each spatial unit is a node, and the node attributes include the land use type, ecological function level, and development intensity index of the spatial unit. The edges between nodes are established based on spatial adjacency and ecological corridor connectivity.

7. The artificial intelligence-based land space governance analysis method according to claim 1, characterized in that, When the target area is divided into several spatial units, an adaptive grid partitioning strategy is adopted, that is, a high-resolution grid is used in ecologically sensitive areas and urban development boundaries, while a low-resolution grid is used in general agricultural areas and ecological buffer zones. Information transmission between grids of different resolutions is achieved by setting virtual boundary nodes.

8. The artificial intelligence-based land space governance analysis method according to claim 1, characterized in that, In the reverse reasoning process, when multiple intervention measures in the derived sequence conflict with the same spatial unit, a priority arbitration mechanism is introduced: according to the hierarchical order of ecological security priority, food security priority, and economic development priority, the execution order of conflicting intervention measures is automatically adjudicated or one of the intervention measures is selected for execution, and the abandoned intervention measures are recorded as alternative solutions and stored in the measure library for subsequent periods.

9. An artificial intelligence-based land space governance analysis system, used to execute the artificial intelligence-based land space governance analysis method according to any one of claims 1 to 8, characterized in that, include: The data access module is used to acquire and register multi-source land and space data; The spatiotemporal graph construction module is used to divide spatial units and construct a spatiotemporal dynamic graph structure; The prediction module, which incorporates the time-series prediction model, is used to predict the future state of national land space. The conflict identification module is used to compare the predicted state with the spatial control target to identify conflict areas and time periods; The reverse reasoning module is used to deduce the sequence of intervention measures from the target state. The output module is used to output recommended governance solutions.

10. A land space governance analysis system based on artificial intelligence according to claim 9, characterized in that, It also includes a knowledge accumulation module, which records the combination of intervention measures tried in each reverse reasoning process and their corresponding implementation effects. When the land use structure similarity between the new area to be governed and the historical area exceeds a preset threshold, the module automatically retrieves the historically successful combination of intervention measures as the initial solution for reverse reasoning, thereby shortening the reasoning time and improving the reliability of the solution.

Images