Method and system for feeding back land utilization change based on land space-time model

Abstract

The invention relates to the technical field of natural resource monitoring and spatial information processing, in particular to a method and a system for feeding back land utilization change based on a land spatio-temporal model. The method comprises the following steps: deploying multi-source land monitoring equipment, carrying out collaborative data acquisition and standardization processing, and constructing a land space-time reference data set; performing triple mapping on the land space-time reference data set to obtain a land semantic association graph; constructing a land utilization knowledge graph based on the land semantic association graph; constructing a land change detection initial model by using the land utilization knowledge graph; meanwhile, in a high-frequency change scene, such as an urban and rural ecologic zone or an ecological sensitive area, a traditional model is slow in response to short-term land utilization disturbance, automatic adjustment cannot be carried out through deviation feedback between historical errors and model output, and the reliability of the model in actual application scenes such as resource regulation and control is limited.

Description

Technical Field

[0001] The present invention relates to the technical field of natural resource monitoring and spatial information processing, and in particular to a method and system for feeding back land use change based on a land spatiotemporal model. Background Art

[0002] Existing land-use change monitoring and analysis methods mostly rely on static remote sensing interpretation and annual statistical data, often failing to adequately capture the temporal evolution and spatial diffusion characteristics of land-use change. Traditional methods often employ single- or multi-period image change detection (e.g., post-classification comparison and difference analysis), failing to continuously model the dynamic drivers of land-use change. This results in a lack of timeliness and responsiveness in change prediction and risk assessment. Furthermore, existing land-use change monitoring systems rarely incorporate cyclical optimization structures based on model feedback mechanisms. Most methods employ a linear processing flow, from data acquisition, classification processing, change interpretation, to patch output, without error backtracking or structural fine-tuning of the initial model. This makes them ill-suited to land evolution processes under complex spatiotemporal conditions. Furthermore, current systems typically focus on patch-level spatial annotation and classification, neglecting the multi-scale spatial influencing mechanisms and semantic transformation patterns underlying land-use change. They lack the ability to jointly model factors such as slope, adjacency structure, and land-use semantics, making it difficult to support deep rule extraction and semantic constraint expression from multi-source data. At the same time, in high-frequency change scenarios, such as urban-rural transition zones or ecologically sensitive areas, traditional models are slow to respond to short-term land use disturbances and cannot automatically adjust through deviation feedback between historical errors and model outputs, limiting their reliability in practical application scenarios such as resource regulation. Summary of the Invention

[0003] Based on this, it is necessary to provide a method and system for feeding back land use changes based on a land spatiotemporal model to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a method for feeding back land use change based on a land spatiotemporal model is provided, the method comprising the following steps:

[0005] Step S1: Deploy multi-source land monitoring equipment, conduct collaborative data collection and standardization processing, and construct a land spatiotemporal benchmark dataset;

[0006] Step S2: Perform triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph; construct a land use knowledge graph based on the land semantic association graph; and construct an initial model for land change detection using the land use knowledge graph;

[0007] Step S3: Outputting a land change patch set based on the initial land change detection model; constructing a land evolution driving matrix based on the land change patch set; performing node prior distribution based on the land evolution driving matrix, and performing parameter error update to obtain backtracking parameters of the land optimization model; optimizing the initial land change detection model using the backtracking parameters of the land optimization model to obtain a land spatiotemporal optimization model;

[0008] Step S4: Output the land supply risk heat map based on the land spatiotemporal optimization model; evolve the holographic sand table according to the land supply risk heat map, and generate a land spatiotemporal change report.

[0009] The beneficial effect of the present invention lies in the construction of a spatiotemporal benchmark dataset of land covering spatial, temporal, and attribute information by deploying multi-source land monitoring equipment for collaborative data collection and standardized processing. This ensures data integrity, consistency, and comparability, providing a solid foundation for subsequent analysis. Subsequently, triple mapping is performed on the benchmark dataset to construct a land semantic association graph, which expresses land objects, their attributes, and their relationships in a structured form. This in turn constructs a land use knowledge graph, achieving semantic hierarchical integration and enhanced associations of land information. This graph structure supports the expression and reasoning of complex relationships, enhancing the semantic understanding capabilities of the initial model for land change detection. Based on the knowledge graph, the initial model outputs a set of land change patches, capturing subtle changes in land use through high-precision spatial resolution. Furthermore, a land evolution driving matrix is ​​constructed, comprehensively considering the multi-factor driving mechanisms of land change. Based on this matrix, node prior distributions and parameter errors are dynamically updated, enabling fine-tuning and backtracking of model parameters to optimize model prediction performance. The application of backtracking parameters in the land optimization model significantly improves the adaptability and accuracy of the initial model in both spatial and temporal dimensions, forming a spatiotemporal land optimization model. The land supply risk heat map generated by this optimization model quantitatively reflects the spatial distribution of potential risks in land resources and supports the intuitive identification of risk areas. Based on the evolution of the holographic sandbox of the risk heat map, through three-dimensional dynamic visualization technology, multi-dimensional data is integrated to provide decision makers with a multi-scale, multi-perspective display of land resource status. Ultimately, through the generated land spatiotemporal change report, the system integrates analysis results and risk assessment information to support scientific decision-making and continuous monitoring of land management, and enhance the intelligent level of land resource protection and utilization. The overall process realizes a closed loop from multi-source data collection, semantic association modeling, dynamic drive analysis to risk assessment and visualization, promoting the development of land management towards data-driven intelligent direction.

[0010] Preferably, step S1 includes:

[0011] Step S11: acquiring Planet satellite image data; segmenting the Planet satellite image data according to a temporal and spatial resolution of 3 m / day, and performing NDVI vegetation index extraction processing to obtain cultivated land vegetation coverage intensity series data;

[0012] Step S12: deploying land vector monitoring data and collecting a land survey base map; performing topological verification on the land survey base map, and compressing redundant nodes using Douglas-Peucker with a tolerance of 0.5m to obtain a lightweight land vector base map;

[0013] Step S13: Setting the locator at the information base station and performing spatial density analysis of a grid unit of 100m×100m to obtain base station spatial density data; performing kernel density estimation on the base station spatial density data to obtain population distribution grid data;

[0014] Step S14: deploy soil moisture detectors and obtain soil moisture data from the soil moisture detectors; perform Kriging spatial interpolation on the soil moisture data and classify it according to a 1km grid to obtain a precipitation erosion factor;

[0015] Step S15: Fill missing values ​​in the cultivated land vegetation coverage intensity series data, lightweight land vector base map, population distribution raster data, and precipitation erosion factor, and perform time regularization processing to construct a land spatiotemporal benchmark dataset.

[0016] The present invention obtains Planet satellite image data, segments the image data based on its 3-meter spatial resolution and daily temporal resolution, and then extracts the NDVI vegetation index to quantify the temporal changes in the intensity of cultivated land vegetation cover, forming high-temporal and spatial resolution sequence data reflecting the dynamics of vegetation growth, providing continuous and detailed spatiotemporal information for the assessment of land ecological conditions. Secondly, land vector monitoring equipment is deployed and a land survey base map is collected. The geometric consistency and integrity of the spatial data are ensured through topological verification. The Douglas-Peucker algorithm is used to compress redundant nodes with a 0.5-meter tolerance, effectively reducing vector data redundancy, improving data storage and computing efficiency, and generating a lightweight land vector base map to facilitate subsequent spatial analysis and overlay calculations. In view of the spatial distribution characteristics of the population, based on the locator setting of the information base station, a spatial density analysis of 100-meter × 100-meter grid cells is carried out to obtain base station spatial density data. Then, the kernel density estimation technique is applied to convert the discrete base station density into spatially continuous population distribution raster data, achieving an accurate expression of the spatial distribution of the population in the scene. Soil moisture data is collected in real time by deployed detectors. Kriging spatial interpolation is used to spatially extrapolate the sampling point data. Combined with a 1-kilometer grid classification, precipitation erosion factors are calculated to reveal the spatial distribution of soil moisture and erosion risk. Finally, missing values ​​are uniformly filled in for the cultivated land vegetation cover intensity series data, lightweight land vector basemap, population distribution raster data, and precipitation erosion factors. Time warping is then used to synchronize the data time series, eliminating temporal inconsistencies between multi-source data and constructing a high-quality, temporally consistent land spatiotemporal benchmark dataset. This integrated dataset provides comprehensive and reliable spatiotemporal data support for subsequent land change analysis, risk assessment, and management decisions.

[0017] Preferably, step S2 includes:

[0018] Step S21: performing triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph;

[0019] Step S22: constructing a land use knowledge graph based on the land semantic association graph;

[0020] Step S23: Obtain historical land change patches; perform initial settings on the learning aggregator based on the land use knowledge graph, and use historical land change patches for training to obtain an initial model for land change detection.

[0021] This paper applies triple mapping to a spatiotemporal benchmark dataset of land, aiming to extract structured semantic relationships from the "land type-attribute-value" relationship in the raw spatiotemporal data and convert it into a node-edge structure suitable for graph data modeling, thereby forming a land semantic association graph. This graph not only preserves the spatial adjacency between features but also incorporates multidimensional information such as functional attributes, evolutionary trends, and temporal changes, establishing high-order semantic connections between land use elements. Next, a land use knowledge graph is constructed based on this semantic association graph. This graph, through methods such as entity alignment, semantic aggregation, and graph expansion, uniformly encodes land use categories, evolutionary paths, constraints, and external driving factors, enabling a graphical representation of complex land use knowledge. Furthermore, historical land change patch data is introduced as a real-world example of land change. By aligning the patch change information with the structural semantics in the knowledge graph, a semantic labeling system for training samples is constructed and used to initialize the learning aggregator. At this stage, the learning aggregator primarily aggregates and statistically analyzes the spatial-temporal-functional attributes of different semantic nodes and fuses features to learn the prior conditions and spatial configuration patterns for land change. During the training process, historical map data provides the model with the spatiotemporal evolution path of land use transfer, assists the model in grasping the evolution law, and drives the parameter learning process of the initial detection model together with the structural information in the semantic map. The final initial model for land change detection has the basic recognition ability of land use evolution in different periods and regions. This method constructs a detection model with semantic understanding ability and spatiotemporal recognition ability through the joint path of knowledge map and supervised training. Therefore, the present invention solves the problems of weak data association and insufficient semantic support in traditional land change modeling by fusing semantic association graph with historical map, and improves the model's learning efficiency of land evolution law and initial detection accuracy.

[0022] Preferably, step S23 includes:

[0023] Step S231: Obtain historical land change maps;

[0024] Step S232: Initially set the learning aggregator with an encoder depth of 4 and an initial learning rate of 0.001 for each node of the land use knowledge graph to obtain an initial model; upload the model gradient data for each node of the land use knowledge graph to obtain the quality data of each node;

[0025] Step S233: Using historical land change patches to train the initial model, and performing data quality weight calculation on each node quality data to obtain land analysis aggregation weight data;

[0026]

[0027] in Aggregate weight data for land analysis;

[0028] Step S234: constructing an initial model for land change detection based on the land analysis aggregation weight data.

[0029] This method obtains historical land change data, which records plot boundaries, category transitions, and temporal information related to land use changes over time, as a supervised label data source for model training. Subsequently, based on the semantic structure and attribute information of each node in the land use knowledge graph, encoder parameters are set for each node input learning aggregator, with a depth of 4 and an initial learning rate of 0.001. This creates an initial model structure with sufficient expressiveness. Model gradient data is then uploaded, combining the topology and attribute characteristics of the knowledge graph. This upload process synchronizes the gradient changes of model parameter updates across nodes, thereby quantifying the responsiveness and update effectiveness of each node during training and quantifying it as node quality data. Next, the initial model is trained using historical land change data. The quality data of different nodes during training is incorporated into a weight evaluation mechanism to calculate land analysis aggregation weights. Specifically, during the aggregation process, each semantic node is assigned a weight coefficient that matches its historical performance, enabling data-driven semantic reinforcement selection. This aggregation weight not only reflects the node's individual contribution to the change detection task but also incorporates the semantic coupling effects of upstream and downstream nodes in the graph structure. Finally, the semantics of the model are enhanced and the structure is adjusted based on the aggregated weight data to complete the construction of the initial model for land change detection. This allows the model to have the ability to identify land use evolution while effectively avoiding the interference of low-quality nodes on the model training results. Starting from the data granularity, this method integrates semantic structure information and supervised training feedback to achieve parameter refinement and adaptive weight allocation in the initial stage of the model. Therefore, by introducing a dynamic weight aggregation mechanism based on graph node quality feedback, the present invention solves the problems of insufficient utilization of semantic structure and failure to distinguish node quality differences in the traditional model construction process, thereby improving the semantic adaptability and training stability of the initial model for land change detection.

[0030] Preferably, step S3 includes:

[0031] Step S31: outputting a land change patch set based on the initial land change detection model; performing buffer diffusion processing on the land change patch set to obtain a land evolution driving matrix, wherein the attenuation radius of the buffer diffusion processing is 1 km;

[0032] Step S32: performing node prior distribution based on the land evolution driving matrix and correcting misclassified spots to obtain backtracking parameters of the land optimization model;

[0033] Step S33: Optimize the initial land change detection model using the backtracking parameters of the land optimization model to obtain a land spatiotemporal optimization model.

[0034] The present invention uses a set of land change patches output by the initial model of land change detection. The patch data records the spatial change position, area, change type and other information of land use status in a specific time window, and is the basic data unit for land evolution analysis. A buffer diffusion process is performed on the patch set, that is, a buffer area with spatial diffusion characteristics is generated around the change patch, and the attenuation radius is set to 1km, which means that the change impact is the greatest in the center area of ​​the patch, and there is a decay trend within 1km outward. This diffusion process constructs a spatial influence weight field on the data to simulate the possible induced effects of land changes on surrounding areas, thereby generating a land evolution driving matrix. The matrix is ​​a weighted graph structure that reflects spatial adjacency, change intensity and propagation trend, and can capture the spatial diffusion path and direction of land changes. Subsequently, the node prior distribution is set based on the evolution driving matrix, and the spatial position and land type of the land change area are embedded in the model structure as prior information to improve the spatial adaptability and classification consistency of the model. At the same time, a misclassified patch correction mechanism is introduced to locate and correct land types misclassified in the initial model. By identifying the spatial isolation, semantic contradictions, or evolutionary irrationality of abnormal patches, misclassified labels are corrected, further improving the quality of training samples. The corrected patch results are used to generate model backtracking parameters, constructing a set of tuning parameters that reflect spatiotemporal behavior and classification accuracy feedback. Finally, these backtracking parameters are input into the initial model, and the model is optimized through parameter resetting, structural fine-tuning, and error backpropagation, forming a land spatiotemporal optimization model with dynamic evolutionary response and spatial consistency control capabilities. At the data level, this model integrates elements such as evolutionary trend driving, prior constraints, and error correction feedback, improving the spatial logic and temporal consistency of land change detection results. Therefore, through buffer diffusion modeling and backtracking parameter optimization strategies, the present invention solves the problems of missing evolutionary relationship modeling and accumulated classification errors in traditional land change detection, improving the spatial scalability and evolutionary adaptability of the model.

[0035] Preferably, step S32 includes:

[0036] Step S321: performing inverse distance weight analysis on the land evolution driving matrix to obtain land evolution promotion data;

[0037] Step S322: Obtain a land semantic association graph; perform node prior distribution based on land evolution promotion data, and use the land semantic association graph to define an adjacency matrix to obtain a land evolution adjacency matrix, wherein the weight of the spatial distance defined by the adjacency matrix is ​​less than 500m and the type is convertible is defined as 1;

[0038] Step S323: Obtain the drone inspection land map; use the drone inspection land map to conduct on-site verification of the land evolution adjacency matrix, and correct the misclassified spots to obtain the land optimization model backtracking parameters.

[0039] The present invention performs an inverse distance weight analysis on the land evolution driving matrix and uses spatial statistical methods to assign higher weights to areas closer to the center of the change patch, thereby obtaining land evolution promotion data. This data reflects the diffusion trend and impact intensity of land changes in spatial neighborhoods in the form of a grid or network, providing a quantitative basis for subsequent spatial association modeling. Subsequently, combined with the constructed land semantic association map, the semantic convertible relationship between land use types is introduced, and the spatial driving intensity and semantic adjacency relationship are used as conditional constraints to define the land evolution adjacency matrix. In the construction of the adjacency matrix, the weight between nodes with a spatial distance less than 500 meters and convertible types is set to 1, forming a graph structure with both spatial similarity and semantic connectivity to describe the evolution path and potential change possibility between land elements. This adjacency matrix not only reflects the static connection between spatial neighborhoods, but also incorporates the land use evolution logic and spatial structure drive, forming a spatiotemporal composite land change association network. Furthermore, drone patrol land patch data is introduced as a high-resolution, field-observed spatial information source to conduct on-site verification of the constructed land evolution adjacency matrix to determine whether there are semantic logic conflicts or spatial deviations in the evolution path. By comparing the patrol patches with the evolution prediction results, misclassified patches are identified and corrected, including spatial position adjustment, land class label repair and time series repositioning, and then the revised land optimization model backtracking parameters are generated at the data level to adjust the model's prediction logic and evolution path weights. This process integrates three types of heterogeneous data sources: spatial attenuation analysis, semantic graph modeling and drone field data feedback, forming a dynamic optimization mechanism with closed-loop verification capabilities. Therefore, the present invention solves the problems of rough spatial association, lack of semantic logic and insufficient field verification in traditional land change modeling by constructing a land evolution adjacency matrix that integrates spatial distance, semantic transformation and field verification, thereby improving the structural accuracy of the model and the reliability of spatial prediction.

[0040] Preferably, step S33 includes:

[0041] Step S331: performing structural analysis of key fields on the backtracking parameters of the land optimization model to obtain land change deconstruction data;

[0042] Step S332: adjusting the data quality weight of the learning aggregator according to the land change deconstruction data, and outputting a model backtracking parameter set;

[0043] Step S333: perform dual-channel backpropagation optimization on the initial land change detection model according to the model backtracking parameter set, freeze the encoder, fine-tune the classification head parameters, and iterate the model optimization 50 times to obtain the land spatiotemporal optimization model.

[0044] This paper, through structured parsing of key fields in the backtracking parameters of a land optimization model, primarily transforms the heterogeneous optimization information generated during the initial modeling process into a unified data structure called "land change deconstruction data." This data includes model error sources, the distribution of misclassified regions, land type evolution trend indicators, and spatial adjacency error weights. Through field classification, semantic encoding, and standardized modeling methods, a data input interface is constructed that can directly control model parameters. Based on this data, the internal structure of the learning aggregator is weighted based on the land change deconstruction data. Specifically, this involves quantifying and weighting the data quality of different data nodes or graph substructures, eliminating the interference of low-confidence samples on the model gradient, and outputting a set of model backtracking parameters that characterizes the data confidence distribution. This parameter set not only reflects the data quality distribution pattern but also carries information about the model error feedback path, providing directional guidance for subsequent model updates. Furthermore, based on this parameter set, a dual-channel backpropagation optimization is performed on the initial land change detection model: one channel is used to update the classification weights under the semantic path, while the other channel focuses on error correction optimization of the spatial adjacency structure. Through the dual-channel division of labor back propagation path, the model's ability to integrate heterogeneous information sources is enhanced, while the direction and amplitude of learning deviations are controlled. In addition, the encoder part of the original model is frozen during the training process to ensure that its existing spatial feature extraction capability is not destroyed, and only the classification head is fine-tuned, thereby strengthening the classification logic while maintaining the stability of the overall model. The model training is set to iterate 50 times, aiming to fully converge the optimization parameters through multiple rounds of backtracking updates, and the final output land spatiotemporal optimization model has higher change recognition ability and structural generalization ability. Therefore, the present invention solves the problems of unclear error backtracking path, single optimization strategy and easy falling into local optimality in traditional land change models through the integration of structured analysis, weight adaptive adjustment and dual-channel optimization mechanism, thereby improving the response efficiency and accuracy stability of the model to complex land spatiotemporal evolution patterns.

[0045] Preferably, step S4 includes:

[0046] Step S41: outputting land spatiotemporal change data based on the land spatiotemporal optimization model; constructing a land supply risk heat map based on the land spatiotemporal change data;

[0047] Step S42: Evolving the holographic sand table based on the land supply risk heat map to obtain the national land space evolution sand table;

[0048] Step S43: Conduct full-process traceability analysis based on the national land space evolution sandbox, and construct a full-process report to obtain a land spatiotemporal change report.

[0049] This method extracts spatiotemporal land change data based on an optimized spatiotemporal land model. This data includes multiple structured fields, such as the location of land type changes, evolution paths, conversion intensity, and temporal evolution, forming a high-resolution dynamic change information set. Furthermore, by combining multidimensional weighting factors such as land use function, distribution of ecologically sensitive areas, cultivated land red lines, and construction intensity, this data is quantitatively analyzed through spatial overlay and weight calculation to construct a land supply risk heat map. This heat map uses a grid as the basic unit, overlaying multiple risk factors within each pixel. Normalization and risk index modeling are used to form a continuous spatially distributed risk gradient, enabling the spatial localization of land supply risks and the identification of high- and low-risk zones. Next, a holographic land evolution sand table is constructed based on this heat map. This sand table utilizes three-dimensional spatial modeling to visually present land change trends, risk levels, evolution trajectories, and spatial conflict zones. Topographic relief, construction density, and land use temporal animations are embedded in the sand table framework, creating an interactive, traceable, and comparable representation of land change scenarios. Finally, by systematically indexing and collating the high-risk areas, historical change paths, key change nodes and other contents involved in the land space evolution sandbox, a full-process land change traceability link is established, and a land spatiotemporal change report is generated. This report not only contains multi-dimensional analysis content such as map statistics tables, risk level distribution maps, and spatiotemporal evolution trend maps, but also embeds evolution path explanations and risk cause analysis texts to form a data product output with a clear structure and complete content. The overall process uses data-driven dynamic modeling and visualization integration to construct a closed-loop data chain from model output to risk presentation to result traceability. Therefore, the present invention solves the problems of single results, fuzzy risk distribution, and difficult to trace back the evolution process in traditional land evolution monitoring by integrating land spatiotemporal change data with a risk semantic expression system, thereby improving the spatial interpretability of land change results and the precision support capability of risk management.

[0050] Preferably, step S41 includes:

[0051] Step S411: outputting land spatiotemporal change data based on the land spatiotemporal optimization model;

[0052] Step S412: Obtaining a basic farmland protection vector; using the basic farmland protection vector to overlay the spatiotemporal land change data, and calculating the area ratio of the conflicting areas, to obtain a land conflict area map. If the judgment threshold for the conflicting area ratio exceeds a preset threshold of 15%, a red alert is issued.

[0053] Step S413: Use the land conflict area map to perform red-yellow risk area analysis to obtain a land supply risk heat map.

[0054] The present invention outputs land spatiotemporal change data based on a land spatiotemporal optimization model. This dataset records the spatial distribution and evolution trajectory of land use types in different periods in the form of raster or polygonal patches, and is accompanied by time tags and land class transfer information, with high structure and temporal continuity. On this basis, the statutory basic farmland protection vector boundary data is obtained. This data usually comes from the ownership demarcation results of the land department, and has administrative binding force and spatial accuracy guarantee. Subsequently, the land change patches and basic farmland vectors are spatially cross-superimposed using the geographic spatial overlay analysis method to extract the parts of the change patches that cross or encroach on the farmland protection boundary, and the conflict area is counted and the conflict ratio is calculated using the patch unit as the granularity. The ratio judgment uses the proportion of the overlapping parts of the change patches with the basic farmland as the core indicator, and the warning threshold is set to 15%. That is, if the conflict area in any change patch exceeds 15%, it will be marked as a red warning area in the layer, and below this value it will be divided into yellow or general risk areas. Through the above-mentioned conflict judgment and marking, a land conflict area map is formed. The map not only has location accuracy but also embeds risk level attribute fields. On this basis, a red-yellow risk zoning analysis was carried out, and factors such as conflict intensity, spatial concentration, and temporal trends were weighted and superimposed, and combined with geographical location sensitivity to construct a land supply risk heat map. This heat map shows the risk levels of different regions in the form of color gradients, achieving spatial focusing, risk expression, and auxiliary early warning of key land conflict areas. In the overall process, a complete chain from data identification to risk map output was constructed through high-precision model output, statutory boundary superposition, quantitative proportion judgment, and graded visual expression. Therefore, the present invention solves the problems of rough spatial boundaries, missing quantitative indicators, and vague risk expression in traditional land risk assessment by introducing basic farmland vectors to identify conflicts in land spatiotemporal change data and combining them with a graded early warning mechanism, thereby improving the scientific nature of land supply risk analysis and the accuracy of response.

[0055] In this specification, a system for feeding back land use changes based on a land spatiotemporal model is provided, which is used to execute the above-mentioned method for feeding back land use changes based on a land spatiotemporal model. The system for feeding back land use changes based on a land spatiotemporal model includes:

[0056] The land multi-source perception and benchmark construction module is used to deploy multi-source land monitoring equipment, conduct collaborative data collection and standardized processing, and build a land spatiotemporal benchmark dataset;

[0057] The semantic association graph construction and change model initialization module is used to perform triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph; construct a land use knowledge graph based on the land semantic association graph; and use the land use knowledge graph to construct an initial model for land change detection.

[0058] The land change evolution driving and model optimization module is used to output a land change patch set based on the initial land change detection model; construct a land evolution driving matrix based on the land change patch set; perform node prior distribution based on the land evolution driving matrix and update parameter errors to obtain the backtracking parameters of the land optimization model; and use the backtracking parameters of the land optimization model to optimize the initial land change detection model to obtain a land spatiotemporal optimization model.

[0059] The land risk assessment and sand table evolution display module is used to output the land supply risk heat map based on the land spatiotemporal optimization model; evolve the holographic sand table according to the land supply risk heat map, and generate a land spatiotemporal change report.

[0060] The beneficial effect of the present invention lies in the simultaneous collection of raw spatiotemporal data covering multiple dimensions, including topography, vegetation, soil, and hydrology, through multi-source land monitoring equipment. Through standardized processing, a unified format and highly consistent land spatiotemporal benchmark dataset is constructed, ensuring data integrity and comparability for subsequent analysis. Subsequently, using a triplet mapping method, the land spatiotemporal benchmark data is converted into a land semantic association graph, revealing the relationship network between different land use elements. This in turn constructs a land use knowledge graph, achieving a semantic and structured representation of land information. This knowledge graph not only enriches the multi-level attribute information of land use but also provides solid semantic support for the construction of the initial model for land change detection, enhancing the model's ability to identify dynamic land changes. Based on this, the initial model outputs a set of land change patches. By encoding the spatial distribution and characteristics of the change areas, a land evolution driving matrix is ​​formed. This matrix, through a dynamic update mechanism of node prior distributions and parameter errors, generates backtracking parameters for the land optimization model, enabling spatiotemporal feedback regulation of the land change process. The backtracking parameters are used to optimize the initial model to obtain a land spatiotemporal optimization model, improving the accuracy and timeliness of land change detection. Finally, the land supply risk heat map generated by the optimization model, combined with holographic sandbox technology, provides a multi-dimensional dynamic display of land spatial evolution and generates a systematic report on land temporal and spatial changes, supporting risk warnings and decision-making. The overall process achieves precise monitoring and risk assessment of land temporal and spatial changes through data integration, semantic expression, dynamic parameter optimization, and multi-dimensional visualization. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1A flowchart of the steps of a method for feedback of land use change based on land spatiotemporal model;

[0062] Figure 2 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0063] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0064] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of the present invention.

[0065] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0066] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0067] To achieve this, please refer to Figures 1 to 2 A method for feeding back land use change based on a land spatiotemporal model comprises the following steps:

[0068] Step S1: Deploy multi-source land monitoring equipment, conduct collaborative data collection and standardization processing, and construct a land spatiotemporal benchmark dataset;

[0069] Step S2: Perform triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph; construct a land use knowledge graph based on the land semantic association graph; and construct an initial model for land change detection using the land use knowledge graph;

[0070] Step S3: Outputting a land change patch set based on the initial land change detection model; constructing a land evolution driving matrix based on the land change patch set; performing node prior distribution based on the land evolution driving matrix, and performing parameter error update to obtain backtracking parameters of the land optimization model; optimizing the initial land change detection model using the backtracking parameters of the land optimization model to obtain a land spatiotemporal optimization model;

[0071] Step S4: Output the land supply risk heat map based on the land spatiotemporal optimization model; evolve the holographic sand table according to the land supply risk heat map, and generate a land spatiotemporal change report.

[0072] In the embodiment of the present invention, reference Figure 1 FIG. 1 is a flow chart showing the steps of a method for feeding back land use changes based on a land spatiotemporal model according to the present invention. In this example, the method for feeding back land use changes based on a land spatiotemporal model comprises the following steps:

[0073] Step S1: Deploy multi-source land monitoring equipment, conduct collaborative data collection and standardization processing, and construct a land spatiotemporal benchmark dataset;

[0074] In this embodiment of the present invention, the deployed monitoring equipment includes high-resolution remote sensing image acquisition systems (such as multispectral and SAR satellite sensors), drone-mounted low-altitude photography equipment, ground-based automatic observation stations (for acquiring soil and meteorological parameters), and IoT sensing nodes (such as RFID and LoRa sensors). Each device synchronizes data acquisition within a unified time window to ensure spatiotemporal consistency. During the data aggregation process, spatial registration and time series alignment are first performed on each source data. A unified spatiotemporal reference system is established through the master control platform to construct a framework for raw land observation data with consistent spatial resolution and temporal continuity. Subsequently, a unified data standardization mapping model is used to reconstruct numerical values ​​and unify formats based on the sampling methods, sensor scales, and data structures of each data source, ensuring data compatibility and accessibility in the subsequent semantic modeling phase. During the standardization process, spatial interpolation algorithms, noise filtering mechanisms, and data integrity restoration methods are introduced to correct for void values, outliers, and blurred boundary areas. Furthermore, a multidimensional labeling system is constructed based on land classification coding rules, providing high-confidence, structured basic data support for the subsequent construction of semantic maps and change models. Ultimately, through the collaborative processing and semantically consistent mapping of multi-source observation data, a land spatiotemporal benchmark dataset is formed that integrates the spatial distribution of land features, temporal change trends, and multiple types of land attribute indicators, serving as the initial data basis for the entire system to carry out semantic mapping and evolutionary modeling.

[0075] Step S2: Perform triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph; construct a land use knowledge graph based on the land semantic association graph; and construct an initial model for land change detection using the land use knowledge graph;

[0076] In an embodiment of the present invention, a land spatiotemporal benchmark dataset is used as input, and entity recognition and relationship extraction techniques are used to perform structured expressions of key elements such as land class units, change events, spatial locations, timestamps, and land use attributes. Using the triple construction framework in natural language processing, the multi-source data is converted into a set of basic triples with semantic orientation according to the "subject-predicate-object" logical rules, such as "cultivated land-converted-to-construction land", "forest land-located-in-XX region", "cultivated land-change time-2022", and other structured data units. Subsequently, based on the connection relationships between high-frequency triples, an initial land semantic association graph is constructed. The graph uses land class entities as nodes and semantic relationships such as spatial conversion, use change, and functional association as edges. Graph database modeling methods such as Neo4j or RDF storage are used to achieve visualization and queryability of the semantic network. Furthermore, based on the semantic association graph, semantic fusion and concept clustering methods are used to synonymize and semantically aggregate heterogeneous concepts in multidimensional triples to construct a unified land use knowledge graph. This process incorporates an ontology model as a constraint framework to define a hierarchical structure of land types, land use patterns, change rules, and their interrelationships, enabling a systematic representation of land use knowledge. Once the knowledge graph is constructed, a reasoning mechanism based on graph neural networks (such as GCN and GAT) is employed to aggregate and propagate features of node attributes and graph structure, exploring potential evolutionary paths and temporal dependencies between land units. This outputs an initial inference model for land change detection. This model, based on the knowledge graph as input, integrates land attribute change sequences, graph structure semantic paths, and historical evolution patterns to generate a land change prediction representation with spatial identity, semantic causal structure, and temporal trends. This provides an initial semantic support structure with reasoning capabilities for subsequent patch extraction and evolutionary modeling.

[0077] Step S3: Outputting a land change patch set based on the initial land change detection model; constructing a land evolution driving matrix based on the land change patch set; performing node prior distribution based on the land evolution driving matrix, and performing parameter error update to obtain backtracking parameters of the land optimization model; optimizing the initial land change detection model using the backtracking parameters of the land optimization model to obtain a land spatiotemporal optimization model;

[0078] In this embodiment of the present invention, an initial model is used to identify regional changes in spatiotemporal benchmark data. A set of change patches is extracted based on changes in land class labels, sudden changes in coverage characteristics, and multi-temporal image difference calculation methods. This set of patches contains multidimensional information, including the spatial location of the change area, boundary morphology, change type, and occurrence time, forming a set of land change elements with high spatiotemporal resolution. Based on this, a land evolution driving matrix is ​​constructed by statistically analyzing the temporal transition paths, change frequencies, and land class transition patterns of the change patches. This matrix uses land class states as nodes and state transition probabilities as edge weights. Using Markov chains or evolutionary migration tensors, multi-period change modeling is performed, resulting in a structured data representation reflecting the dynamic evolution mechanisms between different land classes. Subsequently, based on the land evolution driving matrix, node prior distribution modeling techniques are introduced, combining the geographic environmental characteristics of the change patches and external driving factors (such as population density, transportation accessibility, and terrain slope). A Bayesian prior network is constructed to reflect the probability distribution of specific evolution paths, thereby modeling the uncertainty inherent in the evolutionary process. Furthermore, to improve the accuracy and adaptability of the inference model, parameter error analysis and backtracking mechanisms were used to iteratively optimize the transition probability matrix and prior distribution parameters in the model structure. This involved residual calculation, marginal distribution re-estimation, and maximum a posteriori estimation, generating a backtracking parameter set for the land optimization model. This backtracking parameter set was input as an external control into the initial change detection model. By adjusting the model structure, weight distribution, and semantic rule relationships, the model's structural readjustment and parameter relearning were completed, ultimately resulting in a spatiotemporal land optimization model with more accurate spatiotemporal responses and a clearer logic for land class evolution. This model will serve as the input foundation for subsequent steps, enabling higher-level risk reasoning and change visualization.

[0079] Step S4: Output the land supply risk heat map based on the land spatiotemporal optimization model; evolve the holographic sand table according to the land supply risk heat map, and generate a land spatiotemporal change report.

[0080] In this embodiment of the present invention, a land supply risk factor system is established based on the land type distribution output by the spatiotemporal optimization model, combined with a land type conversion probability matrix, land evolution driving factors, and regional resource load indicators. This system includes, but is not limited to, high-dimensional indicators such as land scarcity, land use tension index, planning-development conflict intensity, ecological sensitivity level, and population-industry coupling pressure. Principal component analysis (PCA), entropy weighting, or fuzzy comprehensive weighting methods are used to normalize and weight these various indicators, constructing a multidimensional risk assessment function model and assigning a land supply risk score to each spatial unit. The scoring results are spatially mapped using a rasterized grid to form a land supply risk heat map. This heat map uses color coding for visualization and smoothes the spatial distribution using an interpolation algorithm, presenting a risk level distribution map with strong spatial continuity. Subsequently, to achieve an intuitive and interactive display of land supply trends, spatial three-dimensional modeling and time series reconstruction techniques are introduced. A holographic sandbox evolution model is constructed based on the land change time series data, evolutionary driving paths, and supply risk evolution trends in the optimization model. This model utilizes rule-based modeling and physical engine control methods. By binding geographic entities to their semantic tags, it enables a three-dimensional sand table to possess spatiotemporal linkage capabilities, enabling dynamic simulation of changes in land supply under specific strategies and demographic development scenarios. After the holographic sand table is constructed, the system extracts data snapshots from different time periods in a database-driven manner, comparing factors such as land type distribution, risk level, driving weights, and change trajectories to automatically generate a report on spatiotemporal land changes. Based on time series overlay analysis and spatial change trend inference, this report outputs regional land use structure adjustments, potential supply and demand conflict areas, the distribution of key change areas, and forecasts of future evolution trends, forming a comprehensive report that integrates text, charts, and maps.

[0081] Preferably, step S1 includes the following steps:

[0082] Step S11: acquiring Planet satellite image data; segmenting the Planet satellite image data according to a temporal and spatial resolution of 3 m / day, and performing NDVI vegetation index extraction processing to obtain cultivated land vegetation coverage intensity series data;

[0083] Step S12: deploying land vector monitoring data and collecting a land survey base map; performing topological verification on the land survey base map, and compressing redundant nodes using Douglas-Peucker with a tolerance of 0.5m to obtain a lightweight land vector base map;

[0084] Step S13: Setting the locator at the information base station and performing spatial density analysis of a grid unit of 100m×100m to obtain base station spatial density data; performing kernel density estimation on the base station spatial density data to obtain population distribution grid data;

[0085] Step S14: deploy soil moisture detectors and obtain soil moisture data from the soil moisture detectors; perform Kriging spatial interpolation on the soil moisture data and classify it according to a 1km grid to obtain a precipitation erosion factor;

[0086] Step S15: Fill missing values ​​in the cultivated land vegetation coverage intensity series data, lightweight land vector base map, population distribution raster data, and precipitation erosion factor, and perform time regularization processing to construct a land spatiotemporal benchmark dataset.

[0087] In this embodiment of the present invention, Planet satellite imagery data is acquired and segmented according to its original spatiotemporal resolution (3-meter spatial resolution, daily update frequency). Spectral index inversion is performed on multi-temporal image sequences using the standardized NDVI vegetation index calculation formula (NDVI = (NIR -Red) / (NIR + Red)), generating a high-frequency time-series vegetation cover intensity data sequence covering cultivated land areas. Next, a land vector monitoring data system is deployed to collect a high-precision land survey base map. Topological consistency checks are performed on the base map patch data to ensure seamless boundaries and closed polygon structures. Subsequently, the Douglas-Peucker line simplification algorithm is used to compress vector nodes at a set tolerance threshold of 0.5m, eliminating high-frequency noise and geometrically redundant points. Ultimately, a lightweight representation of land vector base map data is obtained. Simultaneously, population spatial information processing was conducted by setting locator coordinates at information base stations, performing 100m×100m grid division and spatial point density analysis to initially generate base station distribution density data. Kernel density estimation was then used to smooth this spatial point data and construct population distribution grid data reflecting the concentration of local population activity. Regarding the acquisition of ecological and environmental factors, soil moisture probes were deployed to continuously acquire soil moisture data for different spatial units. Kriging spatial interpolation was used to reconstruct the point data across the area, resulting in a 1km grid-based soil moisture spatial distribution map. This was then quantitatively transformed using regional precipitation factors to construct a precipitation erosion factor dataset. The above-mentioned multi-source data have significant differences in spatial resolution, time step, data structure and coordinate system. Therefore, in the data fusion stage, the nearest neighbor interpolation and temporal linear interpolation methods are first used to fill the missing values ​​to ensure the integrity of the data set; then all the data are time-warped to unify the time scale to daily or monthly scale to match the frequency of remote sensing data and the monitoring time series. Finally, a highly consistent land spatiotemporal benchmark data set is constructed, including cultivated land vegetation intensity series, simplified land patch boundary data, gridded population distribution and erosion factors, providing high-quality basic input for subsequent change detection and spatial analysis.

[0088] Preferably, step S2 includes the following steps:

[0089] Step S21: performing triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph;

[0090] Step S22: constructing a land use knowledge graph based on the land semantic association graph;

[0091] Step S23: Obtain historical land change patches; perform initial settings on the learning aggregator based on the land use knowledge graph, and use historical land change patches for training to obtain an initial model for land change detection.

[0092] In an embodiment of the present invention, a land spatiotemporal benchmark dataset is used as input, and a semantic triple mapping method is used to convert the land unit attributes, spatial relationships, and time labels therein into a structured knowledge representation, and encode them according to the "subject-predicate-object" rule, such as "cultivated land-located-region A", "forest land-change time-2020", "construction land-adjacent-cultivated land", etc. Through entity recognition and relationship extraction, data such as land types, spatial locations, change events, and use attributes are semantically organized, and a land semantic association graph is constructed in the form of a graph database. On this basis, the node attributes, edge weights, and topological structure in the semantic graph are utilized, and through ontology modeling and concept hierarchy alignment, similar semantic entities are aggregated, constrained, and associated to form a land use knowledge graph containing land type definitions, use categories, spatiotemporal relationships, and transformation rules. The graph has a complete land type classification system and attribute evolution path, and a semantically consistent graph index structure is constructed through relationship graph traversal and hierarchical reasoning. Subsequently, historical land change patch data was introduced. This dataset, derived from land use monitoring or remote sensing interpretation results from different years, includes fields such as patch boundaries, land types before and after the change, change time, and external driving factors, along with clear labels for change types. To model change detection capabilities, the initial parameters of the learning aggregator are set based on semantic feature paths extracted from the knowledge graph, including node embedding method, number of graph convolutional propagation layers, and neighborhood sampling strategy. Common methods include GCN, GraphSAGE, or GAT architectures. The historical patch data is fed into a graph neural network as training samples, with node change labels as the supervision target. The aggregator's parameter structure is continuously optimized through backpropagation. During training, the spatial similarity, semantic relevance, and historical evolution paths between nodes are jointly learned. The final output is an initial graph neural network model for land change detection. This model has the ability to predict the change trends of unknown nodes in the graph, providing a well-structured and parameter-convergent initial judgment basis for subsequent change patch extraction and evolution modeling.

[0093] It is particularly important that step S22 includes:

[0094] Step S221: performing rule conversion on the land semantic association graph to obtain a land semantic RDF triple set;

[0095] Step S222: Defining spatial and temporal rules based on the DEM slope based on the land semantic RDF triple set to obtain a land spatiotemporal rule assertion table;

[0096] Step S223: Perform graph lightweight design based on the land spatiotemporal rule assertion table to obtain a land use knowledge graph.

[0097] In an embodiment of the present invention, the land semantic association graph is subjected to rule conversion, that is, the nodes (such as land class name, functional type, spatial location) and edges (such as adjacent relationships, evolutionary paths, and changes in use) in the original graph structure are converted into standardized RDF triples (Resource Description Framework), forming a "subject-predicate-object" ternary semantic structure, thereby obtaining a set of land semantic RDF triples. This conversion process provides the original graph with the data exchange and reasoning foundation of the semantic network, and supports subsequent interconnection processing with other geographic ontologies and spatiotemporal ontologies. Then, based on this RDF semantic structure, the slope attribute derived from elevation data (DEM) is introduced as a spatial constraint factor for land use change to construct a spatial rule definition; at the same time, the temporal rules are constructed in combination with the temporal information of land change (such as interannual change frequency and seasonal fluctuations). Spatial rules, such as "Areas with a slope greater than 15° may not be converted to farmland," and temporal rules, such as "Forest land that has not undergone a change in use for three consecutive years is marked as stable," are converted into assertive data entries through formal logic. This leads to the construction of a spatiotemporal land rule assertion table, which is structured with a triple index, rule number, spatiotemporal constraint fields, and applicable scope, forming a rule data layer for knowledge reasoning and graph constraints. Finally, combining the land semantic triple set with the spatiotemporal rule assertion table, a lightweight design is implemented in the construction of the land use knowledge graph. This includes merging redundant nodes, pruning low-frequency relationships, normalizing edge weights, and sparse encoding of rule paths to reduce graph size, improve query efficiency, and enhance semantic computability. The resulting knowledge graph not only retains core information about land use structure and evolution, but also possesses rule-driven reasoning and visualization capabilities. Through standardized representation, rule modeling, and graph structure optimization, this process achieves the integrated representation of semantic, spatial, and temporal information within the graph structure.

[0098] Preferably, step S23 includes the following steps:

[0099] Step S231: Obtain historical land change maps;

[0100] Step S232: Initially set the learning aggregator with an encoder depth of 4 and an initial learning rate of 0.001 for each node of the land use knowledge graph to obtain an initial model; upload the model gradient data for each node of the land use knowledge graph to obtain the quality data of each node;

[0101] Step S233: Using historical land change patches to train the initial model, and performing data quality weight calculation on each node quality data to obtain land analysis aggregation weight data;

[0102]

[0103] in Aggregate weight data for land analysis;

[0104] Step S234: constructing an initial model for land change detection based on the land analysis aggregation weight data.

[0105] In an embodiment of the present invention, historical land change patches are used as supervision sample inputs, wherein the patches contain basic attributes such as land type before and after the change, time label, spatial boundary and interpretation confidence, providing a standard reference for subsequent model training. On this basis, a land use knowledge graph is constructed, and each node in the graph is layered according to a coding depth of 4. The aggregator structure in the graph neural network is initialized with a learning rate of 0.001 to form an initial node representation propagation mechanism. This initial model topologically unfolds the graph structure before training, and inputs the multi-dimensional semantic attributes of each node (such as land type, adjacency, historical change path, etc.) into the aggregation function to achieve local propagation of features between nodes and generation of embedding vectors. In order to improve the model's sensitivity to node differences and the quality control of training data, a node quality weight mechanism is introduced to quantify the confidence level of each historical patch from two dimensions: cloud coverage of remote sensing images and field verification error. The specific weight calculation formula is: , the training samples are weighted and distributed through this formula to ensure that the influence of high-confidence samples during model training is significantly higher than that of low-quality data. Subsequently, the weight is used to adjust the gradient backpropagation weight of the node embedding representation during the training process, forming a weighted aggregated node embedding representation and obtaining the aggregated weight data for land analysis. Finally, based on the node features after fusion weight adjustment, the output layer structure of the graph neural network is reconstructed to enhance the model's ability to discriminate land type changes, and an initial model for land change detection is generated based on the training results. The model has the ability to semantically predict the change trends of unlabeled nodes in the new input graph structure, and serves as an important basis for subsequent evolutionary modeling and optimization.

[0106] As an example of the present invention, refer to Figure 2 As shown, in this example, step S3 includes:

[0107] Step S31: outputting a land change patch set based on the initial land change detection model; performing buffer diffusion processing on the land change patch set to obtain a land evolution driving matrix, wherein the attenuation radius of the buffer diffusion processing is 1 km;

[0108] Step S32: performing node prior distribution based on the land evolution driving matrix and correcting misclassified spots to obtain backtracking parameters of the land optimization model;

[0109] Step S33: Optimize the initial land change detection model using the backtracking parameters of the land optimization model to obtain a land spatiotemporal optimization model.

[0110] In an embodiment of the present invention, based on the set of change patches output by the initial model of land change detection, a buffer zone diffusion process is performed on the patch boundaries within a range of 1 km to simulate the spatial spread effect that land use changes may have on neighboring areas. During the diffusion process, a spatial attenuation factor is introduced to assign different influence weights to different distance positions in the buffer zone, forming a spatial influence field with the change patch as the core and decreasing distance attenuation. The spatial diffusion field is constructed as a land evolution driving matrix through a gridded weighting method. Each cell in the matrix records the comprehensive influence intensity of the surrounding change patches, thereby reflecting the potential land use evolution trend in the region. Subsequently, based on the grid value distribution of the evolution driving matrix, a prior distribution is generated for the graph nodes in the graph neural network. The node prior setting method under the Bayesian framework is used to assign a higher prior probability to areas with significant change trends, guiding the model to pay more attention to the structural adjustment of such areas. To correct for misclassified patches caused by sample bias or label noise, a patch-level error detection mechanism is introduced. By comparing patches with high-confidence patches or external measured data, patches with significantly deviated classification results are identified and relabeled based on their neighborhood attributes, thereby reducing false changes and missed classifications. Furthermore, the prior distribution data is integrated with the corrected patch classification results to construct the backtracking parameters of the land optimization model. This parameter set includes a node confidence adjustment factor, an edge structure reconstruction factor, and a change path reassignment rule. These parameters are used to adjust the propagation weights, activation function distribution, and graph connectivity in the initial model. Finally, the backtracking parameters are embedded in the original graph neural network model, and retraining and reaggregation are performed to complete the structural optimization and parameter update of the original model, outputting a spatiotemporal land optimization model. This optimization model is responsive to spatial diffusion mechanisms and has higher classification accuracy, providing a more stable data foundation for subsequent land risk analysis and evolutionary reasoning.

[0111] It is particularly important that step S31 includes:

[0112] Step S311: Repairing the set of cracks using a triangulated network based on the land evolution driving matrix to obtain a topologically purified change patch set;

[0113] Step S312: superimposing the DEM terrain slope grid based on the topological purification change patch set to obtain the land change impact intensity grid;

[0114] Step S313: performing buffer diffusion processing on the land change impact intensity grid to obtain a land evolution driving matrix, wherein the attenuation radius of the buffer diffusion processing is 1 km.

[0115] In this embodiment of the present invention, the spatial boundaries of the change patches are processed based on a preliminarily generated land evolution driving matrix. A triangulated irregular network (TIN) algorithm is then introduced to automatically connect and repair gaps between patches. This process constructs an irregular triangulated network structure to topologically reconstruct and patch the boundaries of geometrically broken or fragmented patches, eliminating spatial holes and discontinuities caused by remote sensing interpretation errors or patch cropping. This results in a "topologically cleansed change patch set" with complete boundary structure. This patch set is then used as the basis for spatial analysis and overlaid with slope raster data generated from a digital elevation model (DEM). The slope raster uses the terrain slope value of each pixel as a spatial factor reflecting topographic relief. By spatially overlaying the raster with the change patches, a spatial correspondence between patch units and slope conditions is established, thereby generating a "land change impact intensity raster." This raster not only captures the spatial location and shape of land changes but also incorporates slope, a topographic factor that significantly constrains land use transitions. Next, a spatial buffer diffusion process is performed based on the aforementioned impact intensity raster. This involves establishing a buffer diffusion zone around each change patch with a 1 km attenuation radius. Each diffusion pixel is assigned a gradually decreasing influence weight based on a distance decay function. The diffusion function can employ an inverse distance weighted (IDW) model or a Gaussian function to simulate the spatially induced effects and propagation capabilities of land change on surrounding areas. Ultimately, the diffusion results are superimposed back onto the original patch raster to form a land evolution driving matrix. This matrix, structurally a spatially continuous raster dataset, numerically reflects the driving intensity of land change at different locations and serves as an input feature for subsequent model learning. Through these steps, a complete data generation chain is constructed, from patch structure purification, terrain factor weighting, to spatial diffusion modeling, enabling quantitative modeling and dynamic expression of the driving mechanisms of land change in spatial dimensions.

[0116] Preferably, step S32 includes the following steps:

[0117] Step S321: performing inverse distance weight analysis on the land evolution driving matrix to obtain land evolution promotion data;

[0118] Step S322: Obtain a land semantic association graph; perform node prior distribution based on land evolution promotion data, and use the land semantic association graph to define an adjacency matrix to obtain a land evolution adjacency matrix, wherein the weight of the spatial distance defined by the adjacency matrix is ​​less than 500m and the type is convertible is defined as 1;

[0119] Step S323: Obtain the drone inspection land map; use the drone inspection land map to conduct on-site verification of the land evolution adjacency matrix, and correct the misclassified spots to obtain the land optimization model backtracking parameters.

[0120] In an embodiment of the present invention, a land evolution driving matrix is ​​used as input, which reflects the intensity of the influence of the change patches on the evolution of the surrounding areas within a certain spatial range. By applying the Inverse Distance Weighting (IDW) algorithm, the weighted value of each spatial cell affected by all the change patches is calculated. In the specific calculation, the inverse of the distance is used as the attenuation coefficient, so that the area closer to the change center is given a greater weight, thereby obtaining a set of continuous land evolution promotion data. This data represents the potential land use conversion tendency within the region in the form of a grid, and is an important variable reflecting the dynamics of spatial diffusion. Next, a land semantic association map is obtained, and the evolution promotion data is mapped to the map nodes to construct the prior distribution of the nodes, that is, the evolution tendency value of each node is converted into the probability of it being a potential change node through normalization mapping. Based on this, combining the node semantic information and spatial relationship constraints in the land semantic association graph, an adjacency matrix of the graph structure is defined. If the spatial distance between two nodes is less than 500 meters and their land use types are convertible, the connecting edge is assigned a weight of 1; otherwise, it is 0. This constructs a land evolution adjacency matrix that simultaneously constrains spatial constraints and semantic transformations. To ensure the rationality of the adjacency matrix and the authenticity of the data representation, land patch data acquired through drone inspections is further introduced as a basis for on-site verification. This data has high spatial resolution and real-time characteristics, and can be used to verify the actual spatial proximity and attribute consistency between map nodes. By comparing and analyzing the patrol patches with the edge structure expressed in the adjacency matrix, misclassified patches caused by remote sensing misclassification, label drift, or spatial inconsistency are identified. Based on the verification results, node or edge label correction and structural modification are performed. Finally, the above correction information is organized into the backtracking parameters of the land optimization model, including the node prior distribution adjustment value, the edge connection structure update rule and the label consistency enhancement coefficient, which provides reliable parameter input for the subsequent optimization of the graph neural network model and realizes the land evolution structure modeling with more spatial logic and semantic consistency.

[0121] Preferably, step S33 includes the following steps:

[0122] Step S331: Structured analysis of key fields of land optimization model backtracking parameters is performed to obtain land change deconstruction data;

[0123] Step S332: The data quality weight of the learning aggregator is adjusted according to the land change deconstruction data, and a model backtracking parameter set is output;

[0124] Step S333: Double-channel backpropagation optimization is performed on the land change detection initial model according to the model backtracking parameter set, and encoder freezing, fine-tuning of the classification head parameters, and iterative model optimization for 50 times are performed to obtain a land space-time optimization model.

[0125] In the embodiment of the application, the structured analysis of the key fields of the land optimization model backtracking parameters means that the original backtracking parameter data is feature extracted and semantically decomposed, and the complex and multi-dimensional parameter information is converted into calculable land change deconstruction data by constructing a unified data structure. This process involves extraction, format standardization and attribute label assignment of parameter fields, so that subsequent data can be processed in a clear and consistent form. Then, based on the obtained land change deconstruction data, the data quality weight in the learning aggregator is adjusted. This step reflects the difference in the contribution of different input data to model training by dynamically updating the weight distribution, aiming to improve the information utilization efficiency in the data-driven process. When adjusting the weight, statistical analysis or weighting strategy can be used to ensure that high-quality data occupies a larger proportion in model training, thereby optimizing the generation of the parameter set. Subsequently, the double-channel backpropagation optimization is performed on the initial model for land change detection using the backtracking parameter set, that is, in the model training process, the error backpropagation is performed on the information of the two parallel channels at the same time to enhance the model's ability to capture spatio-temporal features. This step includes freezing the encoder part to maintain its feature extraction stability, while fine-tuning the classification head part by limiting the parameter update range to control the effectiveness of the training process and prevent overfitting. The entire fine-tuning process is performed iteratively for about 50 times to gradually adjust the classifier weight parameters and gradually improve the model's performance in land spatio-temporal change recognition. In summary, this technical means covers structured analysis of parameter data, dynamic weight adjustment, and iterative training optimization based on a double-channel mechanism, forming a systematic and data-driven model improvement path.

[0126] Preferably, step S4 comprises the following steps:

[0127] Step S41: Output land spatio-temporal change data based on the land spatio-temporal optimization model; and construct a land supply risk heat map based on the land spatio-temporal change data;

[0128] Step S42: Evolve a holographic sand table based on the land supply risk heat map to obtain a national space evolution sand table;

[0129] Step S43: Conduct full-process traceability analysis based on the national land space evolution sandbox, and construct a full-process report to obtain a land spatiotemporal change report.

[0130] In an embodiment of the present invention, the spatiotemporal land change data output by the land spatiotemporal optimization model, which typically includes dynamic change characteristics of land use, coverage, and function in both temporal and spatial dimensions, are uniformly formatted and used as input for subsequent analysis. Subsequently, a land supply risk heat map is constructed based on this spatiotemporal change data. This process involves spatial statistics and weighting of indicators such as land change intensity, frequency, and impact range. Geographic Information System (GIS) technology is used to map risk indicators to gridded geographic units, forming a distributed risk heat distribution map. This heat map intuitively reflects the spatial differences and changing trends of land supply risks in different regions. Furthermore, based on the generated land supply risk heat map, a holographic sand table of land space evolution is constructed. This step involves multidimensional data fusion, combining the risk heat map with topography, basic geographic information, and other relevant spatiotemporal data sets. Through three-dimensional modeling and rendering technology, a holographic sand table with interactivity and dynamic evolution is formed, achieving a comprehensive multi-scale and multi-perspective display of land space changes. Finally, a full-process traceability analysis was conducted using a land space evolution sandbox. This focused on tracking the driving factors, evolution paths, and associated impacts of spatiotemporal land change. Through correlation data mining and time series analysis, the causal chain and spatial diffusion patterns of land change were analyzed, leading to the generation of a systematic and structured full-process report. This report included the spatiotemporal distribution of change data, risk level classification, and an explanation of the evolutionary process, ensuring the integrity and logical continuity of the reported data. In summary, this technical approach relies on multi-source spatiotemporal data processing, spatial statistical mapping, multidimensional dynamic visualization, and traceability analysis algorithms, forming a complete closed-loop process from data output to risk assessment, visualization, and report generation.

[0131] Preferably, step S41 includes the following steps:

[0132] Step S411: outputting land spatiotemporal change data based on the land spatiotemporal optimization model;

[0133] Step S412: Obtaining a basic farmland protection vector; using the basic farmland protection vector to overlay the spatiotemporal land change data, and calculating the area ratio of the conflicting areas, to obtain a land conflict area map. If the judgment threshold for the conflicting area ratio exceeds a preset threshold of 15%, a red alert is issued.

[0134] Step S413: Use the land conflict area map to perform red-yellow risk area analysis to obtain a land supply risk heat map.

[0135] In an embodiment of the present invention, based on the land spatiotemporal change data output by the land spatiotemporal optimization model, these data generally cover the land use type, utilization intensity and its spatial distribution characteristics over time, and are formatted to form a standardized spatiotemporal change data set as the basic data source for subsequent analysis. Secondly, basic farmland protection vector data is obtained. This data clearly indicates the spatial boundaries and attribute information of basic farmland in vector form, which is an accurate expression of the land protection scope. Subsequently, the basic farmland protection vector and the land spatiotemporal change data are spatially superimposed. This step is completed through the spatial analysis tool in the geographic information system. The geographical locations of the two sets of data are compared and overlapped to identify geographical units that overlap or conflict with the basic farmland protection area in the spatiotemporal change of land. For the conflicting areas, an area ratio calculation is further performed, specifically calculating the ratio of the area of ​​the conflicting area to the total area of ​​the corresponding basic farmland protection unit to quantify the degree of conflict. A preset threshold is set during the calculation process. If the conflicting area ratio exceeds 15%, the area is marked with a red warning to highlight the high-risk conflict area. Finally, based on the conflict area map data, a red-yellow risk zone classification analysis was conducted. This process, based on the percentage of land area, categorized risk areas into different levels, enabling hierarchical identification of land supply risks. Ultimately, a land supply risk heat map reflecting the spatial distribution of risk was generated. The overall technical process encompassed multi-source spatiotemporal data fusion, spatial overlay analysis based on vector data, quantitative conflict assessment, and risk grading mapping, establishing a systematic, data-driven framework for land risk identification and visualization.

[0136] In this specification, a system for feeding back land use changes based on a land spatiotemporal model is provided, which is used to execute the above-mentioned method for feeding back land use changes based on a land spatiotemporal model. The system for feeding back land use changes based on a land spatiotemporal model includes:

[0137] The land multi-source perception and benchmark construction module is used to deploy multi-source land monitoring equipment, conduct collaborative data collection and standardized processing, and build a land spatiotemporal benchmark dataset;

[0138] The semantic association graph construction and change model initialization module is used to perform triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph; construct a land use knowledge graph based on the land semantic association graph; and use the land use knowledge graph to construct an initial model for land change detection.

[0139] The land change evolution driving and model optimization module is used to output a land change patch set based on the initial land change detection model; construct a land evolution driving matrix based on the land change patch set; perform node prior distribution based on the land evolution driving matrix and update parameter errors to obtain the backtracking parameters of the land optimization model; and use the backtracking parameters of the land optimization model to optimize the initial land change detection model to obtain a land spatiotemporal optimization model.

[0140] The land risk assessment and sand table evolution display module is used to output the land supply risk heat map based on the land spatiotemporal optimization model; evolve the holographic sand table according to the land supply risk heat map, and generate a land spatiotemporal change report.

[0141] The beneficial effect of the present invention lies in the simultaneous collection of raw spatiotemporal data covering multiple dimensions, including topography, vegetation, soil, and hydrology, through multi-source land monitoring equipment. Through standardized processing, a unified format and highly consistent land spatiotemporal benchmark dataset is constructed, ensuring data integrity and comparability for subsequent analysis. Subsequently, using a triplet mapping method, the land spatiotemporal benchmark data is converted into a land semantic association graph, revealing the relationship network between different land use elements. This in turn constructs a land use knowledge graph, achieving a semantic and structured representation of land information. This knowledge graph not only enriches the multi-level attribute information of land use but also provides solid semantic support for the construction of the initial model for land change detection, enhancing the model's ability to identify dynamic land changes. Based on this, the initial model outputs a set of land change patches. By encoding the spatial distribution and characteristics of the change areas, a land evolution driving matrix is ​​formed. This matrix, through a dynamic update mechanism of node prior distributions and parameter errors, generates backtracking parameters for the land optimization model, enabling spatiotemporal feedback regulation of the land change process. The backtracking parameters are used to optimize the initial model to obtain a land spatiotemporal optimization model, improving the accuracy and timeliness of land change detection. Finally, the land supply risk heat map generated by the optimization model, combined with holographic sandbox technology, provides a multi-dimensional dynamic display of land spatial evolution and generates a systematic report on land temporal and spatial changes, supporting risk warnings and decision-making. The overall process achieves precise monitoring and risk assessment of land temporal and spatial changes through data integration, semantic expression, dynamic parameter optimization, and multi-dimensional visualization.

[0142] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0143] The foregoing is considered as illustrative only of the principles of the application. Numerous modifications and changes will readily occur to those skilled in the art, and it is intended to embrace all such modifications and changes that fall within the scope of the application. Accordingly, the application is not to be restricted in scope to the specific embodiments disclosed herein but is to be accorded the full scope that the principles and novel features request appropriately granted.

Claims

1. A method for feedback of land use change based on land spatiotemporal model, characterized in that: The following steps are involved: Step S1: Deploy multi-source land monitoring equipment, conduct collaborative data collection and standardization processing, and construct a land spatiotemporal benchmark dataset; Step S2: Perform triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph; construct a land use knowledge graph based on the land semantic association graph; Using land use knowledge graph to construct initial model for land change detection; Step S3: Outputting a land change patch set based on the initial land change detection model; constructing a land evolution driving matrix based on the land change patch set; performing node prior distribution based on the land evolution driving matrix, and performing parameter error update to obtain backtracking parameters of the land optimization model; optimizing the initial land change detection model using the backtracking parameters of the land optimization model to obtain a land spatiotemporal optimization model; Step S4: Outputting a land supply risk heat map based on the land spatiotemporal optimization model; Evolve the holographic sand table based on the land supply risk heat map and generate a land spatiotemporal change report.

2. The method for feeding back land use change based on a land spatiotemporal model according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: acquiring Planet satellite image data; segmenting the Planet satellite image data according to a temporal and spatial resolution of 3 m / day, and performing NDVI vegetation index extraction processing to obtain cultivated land vegetation coverage intensity series data; Step S12: deploying land vector monitoring data and collecting a land survey base map; performing topological verification on the land survey base map, and compressing redundant nodes using Douglas-Peucker with a tolerance of 0.5m to obtain a lightweight land vector base map; Step S13: Setting the locator at the information base station and performing spatial density analysis of a grid unit of 100m×100m to obtain base station spatial density data; performing kernel density estimation on the base station spatial density data to obtain population distribution grid data; Step S14: deploy soil moisture detectors and obtain soil moisture data from the soil moisture detectors; perform Kriging spatial interpolation on the soil moisture data and classify it according to a 1km grid to obtain a precipitation erosion factor; Step S15: Fill missing values ​​in the cultivated land vegetation coverage intensity series data, lightweight land vector base map, population distribution raster data, and precipitation erosion factor, and perform time regularization processing to construct a land spatiotemporal benchmark dataset.

3. The method for feeding back land use change based on a land spatiotemporal model according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: performing triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph; Step S22: constructing a land use knowledge graph based on the land semantic association graph; Step S23: Obtain historical land change patches; perform initial settings on the learning aggregator based on the land use knowledge graph, and use historical land change patches for training to obtain an initial model for land change detection.

4. The method for feeding back land use change based on a land spatiotemporal model according to claim 3, characterized in that: Step S23 includes the following steps: Step S231: Obtain historical land change maps; Step S232: Initially set the learning aggregator with an encoder depth of 4 and an initial learning rate of 0.001 for each node of the land use knowledge graph to obtain an initial model; upload the model gradient data for each node of the land use knowledge graph to obtain the quality data of each node; Step S233: Using historical land change patches to train the initial model, and performing data quality weight calculation on each node quality data to obtain land analysis aggregation weight data; in Aggregate weight data for land analysis; Step S234: constructing an initial model for land change detection based on the land analysis aggregation weight data.

5. The method for feeding back land use change based on a land spatiotemporal model according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: outputting a land change patch set based on the initial land change detection model; performing buffer diffusion processing on the land change patch set to obtain a land evolution driving matrix, wherein the attenuation radius of the buffer diffusion processing is 1 km; Step S32: performing node prior distribution based on the land evolution driving matrix and correcting misclassified spots to obtain backtracking parameters of the land optimization model; Step S33: Optimize the initial land change detection model using the backtracking parameters of the land optimization model to obtain a land spatiotemporal optimization model.

6. The method for feeding back land use change based on a land spatiotemporal model according to claim 5, characterized in that: Step S32 includes the following steps: Step S321: performing inverse distance weight analysis on the land evolution driving matrix to obtain land evolution promotion data; Step S322: Obtain a land semantic association graph; perform node prior distribution based on land evolution promotion data, and use the land semantic association graph to define an adjacency matrix to obtain a land evolution adjacency matrix, wherein the weight of the spatial distance defined by the adjacency matrix is ​​less than 500m and the type is convertible is defined as 1; Step S323: Obtain the drone inspection land map; use the drone inspection land map to conduct on-site verification of the land evolution adjacency matrix, and correct the misclassified spots to obtain the land optimization model backtracking parameters.

7. The method for feeding back land use change based on a land spatiotemporal model according to claim 5, characterized in that: Step S33 includes the following steps: Step S331: performing structural analysis of key fields on the backtracking parameters of the land optimization model to obtain land change deconstruction data; Step S332: adjusting the data quality weight of the learning aggregator according to the land change deconstruction data, and outputting a model backtracking parameter set; Step S333: perform dual-channel backpropagation optimization on the initial land change detection model according to the model backtracking parameter set, freeze the encoder, fine-tune the classification head parameters, and iterate the model optimization 50 times to obtain the land spatiotemporal optimization model.

8. The method for feeding back land use change based on a land spatiotemporal model according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: outputting land spatiotemporal change data based on the land spatiotemporal optimization model; constructing a land supply risk heat map based on the land spatiotemporal change data; Step S42: Evolving the holographic sand table based on the land supply risk heat map to obtain the national land space evolution sand table; Step S43: Conduct full-process traceability analysis based on the national land space evolution sandbox, and construct a full-process report to obtain a land spatiotemporal change report.

9. The method for feeding back land use change based on a land spatiotemporal model according to claim 2, characterized in that: Step S41 includes the following steps: Step S411: outputting land spatiotemporal change data based on the land spatiotemporal optimization model; Step S412: Obtaining a basic farmland protection vector; using the basic farmland protection vector to overlay the spatiotemporal land change data, and calculating the area ratio of the conflicting areas, to obtain a land conflict area map. If the judgment threshold for the conflicting area ratio exceeds a preset threshold of 15%, a red alert is issued. Step S413: Use the land conflict area map to perform red-yellow risk area analysis to obtain a land supply risk heat map.

10. A land use change feedback system based on land spatiotemporal model, characterized in that: The method for feeding back land use changes based on a land spatiotemporal model according to claim 1 is used to implement the method. The system for feeding back land use changes based on a land spatiotemporal model comprises: The land multi-source perception and benchmark construction module is used to deploy multi-source land monitoring equipment, conduct collaborative data collection and standardized processing, and build a land spatiotemporal benchmark dataset; The semantic association graph construction and change model initialization module is used to perform triple mapping on the land spatiotemporal benchmark dataset to obtain a land semantic association graph; construct a land use knowledge graph based on the land semantic association graph; and use the land use knowledge graph to construct an initial model for land change detection. The land change evolution driving and model optimization module is used to output a land change patch set based on the initial land change detection model; construct a land evolution driving matrix based on the land change patch set; perform node prior distribution based on the land evolution driving matrix and update parameter errors to obtain the backtracking parameters of the land optimization model; and use the backtracking parameters of the land optimization model to optimize the initial land change detection model to obtain a land spatiotemporal optimization model. The land risk assessment and sand table evolution display module is used to output the land supply risk heat map based on the land spatiotemporal optimization model; evolve the holographic sand table according to the land supply risk heat map, and generate a land spatiotemporal change report.

Images