Territorial space planning multi-modal data alignment method

Abstract

The invention provides an alignment method for territorial space planning multi-modal data, and belongs to the technical field of big data processing.The alignment method comprises the steps that a quadtree structure is adopted to segment a remote sensing image, a three-dimensional index is established, a graph segmentation algorithm is utilized to divide vector data space subgraphs, a space-time weight matrix fusing space weight, time weight and credibility weight is constructed, and the three-dimensional index is established; inputting the preprocessed multi-modal data into a multi-modal semantic fusion model comprising a residual neural network, a bidirectional encoder and a graph convolutional network to extract a unified semantic representation vector, identifying conflict data by calculating semantic similarity, constructing a rule priority directed graph, and calculating a rule importance score by adopting a random walk algorithm; and executing the conflict resolution rule according to the priority and outputting the aligned data set. The technical problem that efficient alignment of the multi-source heterogeneous territorial space planning data on the semantic level is difficult to realize is solved.

Description

Technical Field

[0001] This invention belongs to the field of big data processing technology, and more specifically, relates to a method for aligning multimodal data of land spatial planning. Background Technology

[0002] Territorial spatial planning involves the comprehensive processing of multimodal data, including remote sensing imagery, policy texts, and vector maps. Traditional techniques integrate data by combining manual interpretation with database association, mapping spatial data from different sources to a standard classification framework using a unified coding system, and relying on expert experience to identify data conflicts. However, in current territorial spatial planning management, due to the diverse data sources and significant differences in their representation, traditional rule-based matching and manual review methods struggle to handle massive amounts of multimodal data, resulting in inaccurate semantic understanding and inefficient conflict identification. In other words, existing technologies face the challenge of achieving efficient semantic alignment of multi-source, heterogeneous territorial spatial planning data. Summary of the Invention

[0003] In view of this, the present invention provides an alignment method for multimodal land spatial planning data, which can solve the technical problem in the prior art that it is difficult to achieve efficient alignment of multi-source heterogeneous land spatial planning data at the semantic level.

[0004] This invention is implemented as follows: It provides a method for aligning multimodal data in land spatial planning, including collecting geospatial data, policy text data, and socioeconomic data, preprocessing them, segmenting remote sensing images into image tiles according to a quadtree structure, establishing a three-dimensional index structure, and storing them in a distributed object storage system, using a graph segmentation algorithm to divide vector data into multiple spatial subgraphs and establishing a boundary node feature cache pool, constructing a spatiotemporal weight matrix to calculate the spatial weight value, temporal weight value, and confidence weight value of each plot, inputting remote sensing images, policy text data, and vector data into a multimodal semantic fusion model to extract a unified semantic representation vector, calculating semantic similarity values ​​based on the unified semantic representation vector and determining semantic conflicts, extracting the rule weight values ​​corresponding to conflicting data to construct a rule priority directed graph, using a random walk algorithm to calculate the importance score value of rule nodes, executing conflict resolution rules according to priority, and outputting an aligned dataset.

[0005] The preprocessing includes standardizing geospatial data to the CGCS2000 coordinate system, semantically standardizing policy text data to GB / T 21010-2017 land category coding, and aligning socioeconomic data to a unified time period.

[0006] The coordinate system unification process employs a seven-parameter transformation model to convert geospatial data from different sources from the original coordinate system to the National 2000 coordinate system.

[0007] The image tiles are 256×256 pixels in size, and the three-dimensional index structure includes a spatial dimension index encoded with GeoHash, a time dimension index sorted by timestamps, and a resolution dimension index with a pyramid hierarchy.

[0008] The distributed object storage system adopts a Ceph storage architecture, storing frequently accessed image tiles in the solid-state drive layer and infrequently accessed image tiles in the hard disk array layer.

[0009] The graph partitioning algorithm uses the METIS algorithm to achieve a balanced partition of the graph by minimizing the number of edges between spatial subgraphs, so that the number of nodes in each spatial subgraph is approximately equal and the number of edges across subgraphs is minimized.

[0010] The spatial weight value is calculated using a Gaussian kernel function, the time weight value is calculated using an exponential decay function, and the credibility weight value is determined based on the authority of the data source.

[0011] The spatiotemporal weight matrix is ​​decomposed into provincial block matrices according to the provincial administrative divisions using a block sparse matrix decomposition strategy, and the non-zero weight values ​​are stored in the provincial block matrices using a compressed sparse row format.

[0012] The multimodal semantic fusion model includes a remote sensing image feature extraction branch using a residual neural network structure, a policy text feature extraction branch using a bidirectional encoder structure, and a vector data feature extraction branch using a graph convolutional network structure.

[0013] The multimodal semantic fusion model includes a cross-modal attention layer that dynamically adjusts the attention weight coefficients between different modal feature vectors based on the spatial weight values ​​in the spatiotemporal weight matrix and the credibility weight values ​​of the data source.

[0014] The training of the multimodal semantic fusion model employs a contrastive learning loss function to narrow the distance between the unified semantic representation vectors of positive sample pairs and widen the distance between the unified semantic representation vectors of negative sample pairs.

[0015] The pre-trained weights of the bidirectional encoder structure are derived from a Chinese pre-trained language model, which includes the Wudao model, the Wenxin model, or the Tongyi model.

[0016] The semantic similarity value is calculated using cosine similarity measurement, and the similarity threshold is determined based on the statistical distribution of historical conflict data as the semantic similarity value corresponding to the 10th percentile.

[0017] The rule weight values ​​include a constraint strength weight value determined according to the rule's legal hierarchy and a timeliness weight value calculated according to the rule's effective time.

[0018] The rule priority directed graph uses conflicting rules as nodes and the priority relationships between rules as directed edges. The random walk algorithm iterates to bring the importance scores of each rule node to a stable value.

[0019] All data in the aligned dataset maintain semantic consistency and do not conflict with policies or space. The aligned dataset is used for the delineation of the three zones and three lines and land use control decisions in territorial spatial planning.

[0020] This invention solves the technical problem of low semantic alignment efficiency for multi-source heterogeneous data by constructing a multimodal semantic fusion model to extract a unified semantic representation vector and combining a spatiotemporal weight matrix and a directed graph of rule priorities to achieve intelligent conflict resolution. Addressing the limitation of traditional manual interpretation in handling massive amounts of multimodal data, this invention employs residual neural networks, bidirectional encoders, and graph convolutional networks to extract deep semantic features from image, text, and vector data respectively. A cross-modal attention mechanism is used to fuse multi-source features to form a unified semantic space, enabling comparison of different modalities within the same semantic framework, overcoming the limitations of insufficient semantic understanding in traditional rule matching methods. This invention automatically calculates rule priorities and executes intelligent conflict resolution using a random walk algorithm, replacing the manual review process that relies on expert experience, thus improving the efficiency and accuracy of conflict identification and handling. In summary, this invention solves the technical problem mentioned in the background art of achieving efficient alignment of multi-source heterogeneous land spatial planning data at the semantic level. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the quadtree segmentation and three-dimensional indexing structure of remote sensing images.

[0023] Figure 3 Flowchart for constructing the spatiotemporal weight matrix and block sparse decomposition.

[0024] Figure 4 This diagram illustrates the structure and feature extraction process of a multimodal semantic fusion model.

[0025] Figure 5 This is a directed graph of rule priorities and a flowchart of conflict resolution decision-making. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shown is a flowchart of a method for aligning multimodal data in land spatial planning provided by this invention. This method includes the following steps:

[0028] S1. Collect geospatial data, policy text data, and socio-economic data. Perform coordinate system unification processing on geospatial data to the CGCS2000 coordinate system, perform semantic standardization processing on policy text data to the GB / T 21010-2017 land category coding, and perform time-series alignment processing on socio-economic data to a unified time period.

[0029] S2. The remote sensing images in the geospatial data after coordinate system unification in step S1 are divided into 256×256 pixel image tiles according to the QuadTree structure. A three-dimensional index structure of spatial range-timestamp-resolution is established. The image tiles are stored in a distributed object storage system and stored in layers according to access frequency.

[0030] S3. Use a graph segmentation algorithm to divide the vector data in the geospatial data after coordinate system unification in step S1 into multiple spatial sub-graphs, so as to minimize the number of boundary nodes between spatial sub-graphs, and establish a boundary node feature cache pool for each spatial sub-graph.

[0031] S4. Construct a spatiotemporal weight matrix. Calculate the spatial weight value, temporal weight value and confidence weight value for each plot in the geospatial data after coordinate system unification in step S1. Use the block sparse matrix decomposition strategy to decompose the global spatiotemporal weight matrix into provincial block matrices. Store the non-zero weight values ​​of the provincial block matrices in a compressed sparse row format.

[0032] S5. Input the remote sensing images stored in step S2, the policy text data after semantic standardization in step S1, and the vector data divided in step S3 into the multimodal semantic fusion model, extract the multimodal feature vectors, and output the unified semantic representation vector.

[0033] S6. Calculate the semantic similarity value between different data sources based on the unified semantic representation vector output in step S5. When the semantic similarity value is lower than the similarity threshold, it is determined that there is a semantic conflict. Extract the rule weight value corresponding to the conflicting data.

[0034] S7. Construct a rule priority directed graph from the rule weight values ​​extracted in step S6. Use a random walk algorithm to calculate the importance score of each rule node in the rule priority directed graph. Sort the conflicting rules according to the importance score and execute the rule with the highest priority to complete the conflict resolution. Output the aligned dataset after resolution.

[0035] Among them, coordinate system unification is the process of converting geospatial data from different sources from the original coordinate system to the National 2000 coordinate system. The conversion process adopts a seven-parameter conversion model.

[0036] Semantic standardization is the process of mapping land use terms in policy text data to the national standard land classification coding system. The mapping process is based on terminology dictionaries and ontology mapping rules.

[0037] The time-series alignment process is the process of unifying socio-economic data with different time granularities to the same time period. For monthly data and annual data, linear interpolation and cumulative methods are used for conversion, respectively.

[0038] The QuadTree structure is a quadtree spatial index structure that recursively divides the two-dimensional space into four quadrants. Each quadrant corresponds to a coding value, and the coding value of an image tile is determined by the level and position of the quadrant in which the image tile is located.

[0039] The three-dimensional index structure includes a spatial dimension index, a time dimension index, and a resolution dimension index. The spatial dimension index uses GeoHash encoding, the time dimension index uses timestamp sorting, and the resolution dimension index uses a pyramid hierarchy.

[0040] The distributed object storage system adopts the Ceph storage architecture, which stores image tiles as objects. Image tiles with high access frequency are stored in the solid-state drive layer, while image tiles with low access frequency are stored in the mechanical hard disk array layer.

[0041] Among them, the graph partitioning algorithm adopts the METIS algorithm, which achieves balanced graph partitioning by minimizing the number of edges between spatial subgraphs, so that the number of nodes in each spatial subgraph is approximately equal and the number of edges across subgraphs is minimized.

[0042] The boundary node feature cache pool stores the feature vectors of nodes located at the boundaries of the spatial subgraph. When performing graph neural network calculations, only the feature vectors of the boundary nodes need to be synchronized, without synchronizing all the feature vectors of the nodes.

[0043] The spatial weight value is calculated using a Gaussian kernel function, and the formula is as follows: the spatial weight value is equal to an exponential function, the numerator of which is the square of the negative Euclidean distance between the plots divided by the square of the reference distance, and the denominator of which is 1. The unit of the Euclidean distance is m, the reference distance is 1000m, and the spatial weight value is dimensionless.

[0044] The time weight value is calculated using an exponential decay function, as expressed in the following formula: the time weight value is equal to the exponential function, where the numerator of the exponent is the negative time interval divided by the reference time period, and the denominator of the exponent is 1. The unit of the time interval is days, the reference time period is 365 days, and the time weight value is dimensionless.

[0045] Optionally, the credibility weight value is determined based on the authority of the data source. The credibility weight value of authoritative data released by official sources is set to 1.0, the credibility weight value of data released by third-party institutions is set to 0.8, and the credibility weight value of data collected by the public is set to 0.5.

[0046] Optionally, the block sparse matrix decomposition strategy divides the nationwide spatiotemporal weight matrix into 34 provincial block matrices according to provincial administrative divisions. Each provincial block matrix stores only the weight values ​​between blocks within the province and the weight values ​​between blocks at provincial boundaries. If only considering the province, it is decomposed into a corresponding number of blocks according to the administrative regions within the province; the same division is applied to the city or district.

[0047] The compressed sparse row format stores the non-zero weight values, column index values ​​of the non-zero weight values, and row pointer values ​​of the starting position of each row in row-major order. The storage space complexity is linearly related to the number of non-zero weight values.

[0048] The multimodal semantic fusion model is structured as follows: the input layer receives remote sensing images, policy text data, and vector data; the remote sensing image feature extraction branch uses a residual neural network structure to extract spectral feature vectors and spatial texture feature vectors; the policy text feature extraction branch uses a bidirectional encoder structure to extract semantic feature vectors and contextual feature vectors; the vector data feature extraction branch uses a graph convolutional network structure to extract topological feature vectors and attribute feature vectors; the cross-modal attention layer calculates the attention weight coefficients between different modal feature vectors, and the attention weight coefficients are dynamically adjusted according to the spatial weight values ​​in the spatiotemporal weight matrix and the credibility weight values ​​of the data source; the feature fusion layer concatenates the weighted multimodal feature vectors and maps them to a unified semantic representation vector through a fully connected layer; the output layer outputs a 512-dimensional unified semantic representation vector.

[0049] The steps for establishing the training dataset for the multimodal semantic fusion model include: collecting national territorial spatial planning data from 2015 to 2024 as the original dataset; extracting triplet samples containing remote sensing images, policy text data, and vector data from the original dataset, with a total of 500,000 triplet samples; labeling each triplet sample with a semantic consistency label, with semantically consistent triplet samples labeled as positive samples and semantically conflicting triplet samples labeled as negative samples; and dividing the labeled triplet samples into training set, validation set, and test set in a ratio of 8:1:1.

[0050] The training steps of the multimodal semantic fusion model include: initializing the weight parameters of the residual neural network structure as ImageNet pre-trained weights, initializing the weight parameters of the bidirectional encoder structure as pre-trained weights, and initializing the weight parameters of the graph convolutional network structure as random normal distributions; setting the batch size to 32, the learning rate to 0.0001, and the number of training epochs to 100; training the model using a contrastive learning loss function, which narrows the distance between the unified semantic representation vectors of positive sample pairs and widens the distance between the unified semantic representation vectors of negative sample pairs; evaluating the model performance on the validation set every 10 epochs; stopping training and saving the model weight parameters with the smallest validation set loss value when the validation set loss value does not decrease for 5 consecutive epochs.

[0051] The pre-trained weights of the bidirectional encoder structure are derived from Chinese pre-trained language models, including the Wudao model developed by the Academy of Artificial Intelligence, the Wenxin model developed by Baidu, and the Tongyi model developed by Alibaba.

[0052] The calculation process for the attention weight coefficient in the cross-modal attention layer is as follows: First, the dot product of the query vector and the key vector is calculated and divided by a scaling factor to obtain the attention score. The scaling factor is the square root of the key vector dimension. Then, the attention score is dynamically adjusted using the following formula: The adjusted attention score equals the original attention score multiplied by the normalized value of the spatial weight, divided by the normalized standard value of the spatial weight, and then multiplied by the normalized value of the credibility weight, divided by the normalized standard value of the credibility weight. The normalized standard value of the spatial weight and the normalized standard value of the credibility weight are both 1.0. The adjusted attention score is dimensionless. Finally, the adjusted attention score is softmax normalized to obtain the attention weight coefficient.

[0053] The multimodal feature vectors include the spectral feature vectors and spatial texture feature vectors output by the remote sensing image feature extraction branch in step S5, the semantic feature vectors and contextual feature vectors output by the policy text feature extraction branch, and the topological feature vectors and attribute feature vectors output by the vector data feature extraction branch.

[0054] The semantic similarity value is calculated using the cosine similarity metric, and the formula is as follows: The semantic similarity value is equal to the inner product of two unified semantic representation vectors divided by the product of the magnitudes of the two unified semantic representation vectors. The inner product value divided by the inner product reference value is the normalized inner product, and the magnitude product divided by the magnitude reference value is the normalized magnitude product. The semantic similarity value is equal to the normalized inner product divided by the normalized magnitude product. The inner product reference value is 100, the magnitude reference value is 10, the semantic similarity value ranges from -1 to +1, and the semantic similarity value is dimensionless.

[0055] The similarity threshold is determined based on the statistical distribution of historical conflict data. The semantic similarity values ​​of historical conflict data are sorted from smallest to largest, and the semantic similarity value corresponding to the 10th percentile is selected as the similarity threshold. The current similarity threshold is set to 0.65.

[0056] The rule weight values ​​include the rule's binding strength weight value and the rule's timeliness weight value. The binding strength weight value is determined according to the rule's legal level: the binding strength weight value is set to 1.0 for rules at the national law level, 0.7 for rules at the local regulation level, and 0.5 for rules at the departmental rule level. The timeliness weight value is calculated based on the rule's effective date, and the calculation formula is as follows: the timeliness weight value is equal to an exponential function, where the numerator of the exponent is the number of days since the rule took effect divided by the reference timeliness period, and the denominator of the exponent is 1. The unit of the number of days is days, and the reference timeliness period is 1825 days. The timeliness weight value is dimensionless.

[0057] In this method, the rule priority directed graph uses conflicting rules as nodes and the priority relationship between rules as directed edges. If rule A takes precedence over rule B, a directed edge from A to B is established between node A and node B.

[0058] The random walk algorithm simulates a random walk process on a rule priority directed graph. Initially, the importance score of each rule node is set to 1 divided by the total number of rule nodes. In each iteration, the rule node distributes 85% of its importance score equally to all successor nodes pointed to by the rule node and retains 15%. After 50 iterations, the importance score of each rule node converges to a stable value.

[0059] In the conflict resolution process, rules are executed sequentially from highest to lowest importance score. When a rule is executed, the data constrained by the rule is modified to meet the rule's requirements. Once a conflict is resolved after executing a rule, the execution of subsequent rules stops.

[0060] The aligned dataset includes deconstructed geospatial data, policy text data, and socioeconomic data. All data in the aligned dataset maintains semantic consistency and does not have policy-spatial conflicts. The aligned dataset is used for the delineation of the three zones and three lines and land use control decisions in territorial spatial planning.

[0061] Optionally, the present invention also provides a method for forming a multimodal data alignment system for land spatial planning by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions execute the above-described method when running in the computer.

[0062] The specific implementation methods of the above steps are described in detail below.

[0063] The specific implementation of step S1 is as follows: First, geospatial data is obtained from the land and resources management department, policy text data is obtained from the policy and regulations database, and socio-economic data is obtained from the statistical yearbook database. When performing coordinate system unification processing on the geospatial data, a seven-parameter transformation model is used to transform the spatial coordinates under different coordinate systems to the CGCS2000 coordinate system. The seven-parameter transformation model realizes the transformation between coordinate systems through three translation parameters, three rotation parameters, and one scale parameter. The transformation process ensures that the accuracy error of the spatial position is less than 0.5m. When performing semantic standardization processing on the policy text data, a terminology dictionary is established to map the land use terms in the policy text to the GB / T 21010-2017 land category coding system. The mapping process identifies the semantic equivalence, hierarchical relationship, and correlation relationship between terms through ontology mapping rules. When performing time-series alignment processing on the socio-economic data, monthly data is converted into daily data through linear interpolation, and annual data is converted into monthly data through accumulation, so that the time granularity of all data is unified to the monthly cycle. The purpose of these steps is to eliminate the heterogeneity of multi-source data in coordinate system, semantic standard, and time granularity.

[0064] The specific implementation of step S2 involves performing spatial segmentation on the remote sensing images in the geospatial data after coordinate system unification processing in step S1. A quadtree spatial index structure is used to recursively partition the two-dimensional space, dividing it into four equal quadrants at each step. The recursion depth is set to 18 levels to accommodate the 256×256 pixel image tile size. A unique encoding value is generated for each image tile, consisting of the quadtree level and quadrant position. When establishing the three-dimensional index structure, GeoHash encoding is used to encode the two-dimensional latitude and longitude coordinates into a one-dimensional string, with an encoding precision set to 8 bits to ensure... The spatial resolution reaches 38m×19m. A B+ tree index is built using timestamp sorting for the time dimension, and a multi-scale index is built using a pyramid hierarchical approach for the resolution dimension. When storing image tiles in a distributed object storage system, they are stored in layers based on access frequency statistics. The access frequency threshold is set to 10 times per month. Image tiles with access frequencies higher than the threshold are stored in the solid-state drive layer to provide millisecond-level access speeds, while image tiles with access frequencies lower than the threshold are stored in the mechanical hard disk array layer to reduce storage costs. The purpose of these steps is to establish an efficient remote sensing image storage and retrieval mechanism to support rapid access to large-scale data.

[0065] The specific implementation of step S3 involves constructing a spatial topology map from the vector data in the geospatial data after coordinate system unification in step S1. The spatial topology map uses land parcels as nodes and spatial adjacency relationships between land parcels as edges. A multi-level graph segmentation algorithm is used to recursively binary segment the spatial topology map. The segmentation process minimizes the number of edge cuts between subgraphs by optimizing the objective function. The number of edge cuts refers to the number of edges connecting different subgraphs. Simultaneously, the number of nodes in each subgraph is constrained to remain balanced, with a node balance threshold set to 15%. The recursive segmentation process continues until the number of subgraphs reaches the required number of computed nodes. Boundary nodes are extracted from each spatial subgraph. Boundary nodes are nodes connected to other subgraphs by edges. A boundary node feature cache pool is established for each spatial subgraph to store the feature vectors of the boundary nodes. The feature vector dimension is set to 256 dimensions. The purpose of this step is to decompose large-scale vector data into multiple parallel-processable spatial subgraphs to reduce the communication overhead of graph neural network training.

[0066] The specific implementation of step S4 involves calculating spatial weight, temporal weight, and reliability weight values ​​for each plot in the geospatial data after coordinate system unification in step S1. The spatial weight value is calculated using a Gaussian kernel function, with the Euclidean distance between plots as the input parameter and the spatial weight value as the output parameter. The spatial weight value decreases exponentially with increasing Euclidean distance, with a reference distance parameter set to 1000m. The temporal weight value is calculated using an exponential decay function, with the time interval between data acquisition time and the current time as the input parameter and the temporal weight value as the output parameter. The temporal weight value decreases exponentially with increasing time interval, with a reference time period parameter set to 365 days. The reliability weight value is determined based on the authority level of the data source. The credibility weight value of officially released data is set to 1.0, the credibility weight value of third-party data is set to 0.8, and the credibility weight value of publicly collected data is set to 0.5. A block sparse matrix decomposition strategy is adopted to decompose the global spatiotemporal weight matrix into 34 provincial block matrices according to provincial administrative divisions. Each provincial block matrix only stores the weight values ​​between blocks within the province and the weight values ​​between blocks at the provincial boundary. The provincial block matrices are stored in a compressed sparse row format, which only stores non-zero weight values ​​and the column index and row pointer of the non-zero weight values. The storage space requirement is linearly related to the number of non-zero weight values. The purpose of this step is to quantify the spatiotemporal credibility of the data and reduce the storage space requirement of the spatiotemporal weight matrix.

[0067] The specific implementation of step S5 involves inputting the remote sensing image stored in step S2, the policy text data after semantic standardization in step S1, and the vector data divided in step S3 into a multimodal semantic fusion model. This model uses a residual neural network structure to extract spectral feature vectors and spatial texture feature vectors from the remote sensing image, a bidirectional encoder structure to extract semantic feature vectors and contextual feature vectors from the policy text data, and a graph convolutional network structure to extract topological feature vectors and attribute feature vectors from the vector data. When calculating the attention weight coefficients between different modal feature vectors in the cross-modal attention layer, the dot product of the query vector and the key vector is first calculated and divided by the square root of the key vector dimension to obtain the attention score. Then, the attention score is dynamically adjusted based on the spatial weight and confidence weight values ​​calculated in step S4. During the adjustment process, the attention score is multiplied by the spatial weight and confidence weight values. The feature fusion layer concatenates the weighted multimodal feature vectors into a high-dimensional vector and maps it to a 512-dimensional unified semantic representation vector through a fully connected layer. The purpose of this step is to achieve semantic alignment of different modal data and generate a unified semantic representation.

[0068] The specific implementation of step S6 involves calculating a semantic similarity value from the unified semantic representation vector output in step S5. The semantic similarity value is calculated using cosine similarity measurement. The input parameters are two unified semantic representation vectors, and the output parameter is the semantic similarity value. The semantic similarity value ranges from -1 to +1. The semantic similarity value is compared with a similarity threshold, which is determined based on the statistical distribution of historical conflict data. The 10th percentile of the semantic similarity values ​​of historical conflict data is selected as the similarity threshold. The current similarity threshold is set to 0.65. When the semantic similarity value is lower than the similarity threshold, it is determined that there is a semantic conflict between the corresponding data sources. Extract the rule weight values ​​corresponding to the conflicting data. The rule weight values ​​include the constraint strength weight value and the timeliness weight value of the rule. The constraint strength weight value is determined according to the legal level of the rule. The constraint strength weight value of the national law level is set to 1.0, the constraint strength weight value of the local regulation level is set to 0.7, and the constraint strength weight value of the departmental regulation level is set to 0.5. The timeliness weight value is calculated according to the effective time of the rule. The timeliness weight value decreases exponentially with the number of days since the rule took effect. The reference timeliness period parameter is set to 1825 days. The purpose of this step is to identify semantic conflicts between multi-source data and extract the rule weight information required for conflict resolution.

[0069] The specific implementation of step S7 involves constructing a directed graph of rule priorities using the rule weight values ​​extracted in step S6. This graph uses conflicting rules as nodes and the priority relationships between rules as directed edges. A random walk algorithm is used to calculate the importance score of each rule node in the graph. The random walk algorithm simulates a random walk process on the graph. Initially, the importance score of each rule node is set to the reciprocal of the total number of rule nodes. In each iteration, the rule node distributes 85% of its importance score equally among all successor nodes it points to, retaining 15% of the importance score. The number of iterations is set to 5. To ensure that the importance score converges to a stable value, conflict rules are sorted from high to low according to their importance scores. The conflict rules are then executed sequentially to resolve conflicts. When executing a rule, the data constrained by the rule is modified to meet the rule requirements. Once a conflict is resolved after executing a rule, the execution of subsequent rules stops. The resolved aligned dataset is then output. The aligned dataset includes resolved geospatial data, policy text data, and socioeconomic data. All data in the aligned dataset maintains semantic consistency and does not have policy spatial conflicts. The purpose of this step is to automatically resolve semantic conflicts between multi-source data and output a semantically consistent aligned dataset through rule priority sorting.

[0070] It should be noted that the key technical ideas of this invention include a dynamic feature weighting mechanism based on a spatiotemporal weight matrix, a parallel computing mechanism for distributed graph segmentation, and a conflict resolution mechanism based on random walks. The dynamic feature weighting mechanism based on a spatiotemporal weight matrix achieves differentiated weighting of multimodal features by quantifying the spatial proximity, temporal decay, and credibility of data. Compared to the semantic alignment deviation caused by fixed weights in traditional methods, this invention dynamically adjusts the attention weight coefficients to give higher weights to data with high credibility and strong spatiotemporal correlation, thereby improving the accuracy of semantic alignment. The parallel computing mechanism for distributed graph segmentation decomposes large-scale vector data into parallelizable spatial subgraphs by minimizing the number of edge cuts between subgraphs. Compared to the memory overflow and communication bottlenecks caused by processing the entire graph on a single machine in traditional methods, this invention uses boundary node feature caching technology to synchronize only the boundary features between subgraphs, significantly reducing cross-node communication overhead and improving training efficiency. The conflict resolution mechanism based on random walks breaks the circular dependency of rule priorities by calculating the importance score of rule nodes. Compared to the conflict resolution failure caused by manual rule sorting in traditional methods, this invention achieves automated conflict resolution through globally consistent priority sorting. The synergistic effect of the three technical approaches lies in the fact that the spatiotemporal weight matrix provides a dynamic weight foundation for multimodal feature fusion, the distributed graph segmentation provides parallel computing capabilities for large-scale data processing, and the random walk algorithm provides an automated mechanism for resolving semantic conflicts. The three work together to achieve a complete alignment process from data preprocessing to feature fusion and then to conflict resolution. Compared with the overall performance bottleneck caused by the independent optimization of each module in the existing technology, this invention significantly improves the accuracy and efficiency of multimodal data alignment in land and space planning through end-to-end collaborative optimization.

[0071] It should be noted that this invention also solves the following technical problem: the efficiency bottleneck faced by massive multimodal land spatial data in storage and retrieval. Traditional methods store remote sensing images as whole files, requiring the loading of complete image data for large-scale spatial analysis, resulting in excessive memory consumption and slow retrieval speed. This invention segments images into standard tiles using a quadtree structure and establishes a spatial-temporal-resolution three-dimensional index, achieving on-demand loading and rapid positioning of image data. It employs a distributed object storage system to store image tiles hierarchically according to access frequency, placing hot data in the solid-state drive layer to improve access speed. Addressing the low efficiency of large-scale graph computation for vector data, this invention uses a graph segmentation algorithm to divide the global vector network into multiple spatial subgraphs and establishes a boundary node feature cache pool for each subgraph. During graph neural network computation, only boundary node features need to be synchronized without transmitting all node data, significantly reducing communication overhead in distributed computing and improving the parallel processing efficiency of large-scale spatial data.

[0072] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the technical problem of semantic alignment difficulties in multi-source heterogeneous data lies in the establishment of a deep learning-driven multimodal semantic unified representation mechanism. Traditional methods perform shallow feature matching based on predefined rules, which cannot capture the deep semantic relationships of data. However, this invention automatically learns abstract semantic features from the original data through neural networks, mapping heterogeneous data to a unified high-dimensional semantic space, making data from different modalities comparable. The cross-modal attention mechanism dynamically adjusts the feature fusion weights according to spatiotemporal weights, ensuring that the semantic alignment process fully considers spatial proximity relationships and data credibility, thus improving the accuracy of semantic representation. The rule priority directed graph transforms conflict resolution into a graph theory problem, and the random walk algorithm quantifies the importance of rules by simulating the transmission relationship between rules, realizing the automation and intelligence of conflict handling. This technical solution is logical because deep learning models have powerful feature abstraction capabilities, enabling them to discover complex semantic patterns in data that are difficult to define manually, while graph algorithms provide a mathematical computational framework for rule priority ranking.

[0073] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0074] The specific implementation of step S1 involves collecting geospatial data, policy text data, and socioeconomic data. The geospatial data undergoes coordinate system unification processing to the CGCS2000 coordinate system; the policy text data undergoes semantic standardization processing to the GB / T 21010-2017 land category coding; and the socioeconomic data undergoes time-series alignment processing to a unified time period. The coordinate system unification processing employs a seven-parameter transformation model, and the transformation formula is expressed as follows:

[0075] ;

[0076] ;

[0077] ;

[0078] Where the rotation matrix The expanded form is:

[0079] ;

[0080] In the formula, , , To bypass axis, axis, The rotation angle parameter of the axis, in radians, typically takes a range of values. to radian; , , These are coordinate components in the CGCS2000 coordinate system, in meters. , , These are the coordinate components in the original coordinate system, in meters. , , This is a translation parameter, measured in meters (m), and typically ranges from -5m to +5m. The scale factor is dimensionless and typically takes values ​​in the range of 1000-1200. to ; , , These represent the matrix products of the rotation matrix and the original coordinate components, respectively, in units of m; , , These represent the scaled coordinate components, in meters (m). Semantic standardization processing maps land use terms in policy text data to the national standard land classification coding system based on a terminology dictionary and ontology mapping rules. Temporal alignment processing uses linear interpolation and accumulation methods; monthly data is converted to annual data using linear interpolation, and annual data is converted to monthly data using accumulation.

[0081] The specific implementation of step S2 involves dividing the remote sensing imagery in the geospatial data after coordinate system unification in step S1 into 256×256 pixel image tiles according to a QuadTree structure. A three-dimensional index structure of spatial range-timestamp-resolution is established, and the image tiles are stored in a distributed object storage system and hierarchically according to access frequency. The QuadTree structure recursively divides the two-dimensional space into four quadrants, and the encoded value of the image tile is determined by the level and position of the quadrant in which the image tile belongs. The three-dimensional index structure includes a spatial dimension index using GeoHash encoding, a time dimension index using timestamp sorting, and a resolution dimension index using a pyramid hierarchy. The distributed object storage system adopts a Ceph storage architecture, storing frequently accessed image tiles in a solid-state drive layer and infrequently accessed image tiles in a hard disk array layer.

[0082] The specific implementation of step S3 involves using a graph segmentation algorithm to divide the vector data in the geospatial data after coordinate system unification in step S1 into multiple spatial subgraphs, minimizing the number of boundary nodes between the spatial subgraphs, and establishing a boundary node feature cache pool for each spatial subgraph. Specifically, the graph segmentation algorithm uses the METIS algorithm to achieve a balanced graph partition by minimizing the number of edges between spatial subgraphs, ensuring that the number of nodes in each spatial subgraph is approximately equal and the number of edges across subgraphs is minimized. The boundary node feature cache pool stores the feature vectors of nodes located at the boundaries of the spatial subgraphs. During graph neural network computation, only the boundary node feature vectors need to be synchronized, without synchronizing all node feature vectors.

[0083] The specific implementation of step S4 is to construct a spatiotemporal weight matrix, and the formula for constructing the spatiotemporal weight matrix is ​​as follows:

[0084] ;

[0085] In the formula, For the plot of land With the plot of land The overall spatiotemporal weight value between them is dimensionless; For the plot of land With the plot of land Spatial weight values ​​between them, dimensionless; For the plot of land With the plot of land The time weighting values ​​between them are dimensionless. For the plot of land With the plot of land The corresponding data credibility weight value is dimensionless. and These are the index numbers of the land parcels, dimensionless. For each land parcel in the geospatial data after coordinate system unification in step S1, spatial weight values, temporal weight values, and confidence weight values ​​are calculated. A block sparse matrix decomposition strategy is used to decompose the global spatiotemporal weight matrix into provincial block matrices. Non-zero weight values ​​are stored in the provincial block matrices using a compressed sparse row format. The spatial weight values ​​are calculated using a Gaussian kernel function, as expressed in the following formula:

[0086] ;

[0087] In the formula, For the plot of land With the plot of land Spatial weight values ​​between them, dimensionless; For the plot of land With the plot of land The Euclidean distance between them, in meters; For reference distance, a value of 1000m is usually taken. The square of the Euclidean distance, in units of ; The square of the reference distance, in units of ; It is a dimensionless ratio; It is an exponential function. The time weight value is calculated using an exponential decay function, as expressed in the following formula:

[0088] ;

[0089] In the formula, For the plot of land With the plot of land The time weighting values ​​between them are dimensionless. For the plot of land With the plot of land The time interval for the corresponding data is in days; For reference time periods, a value of 365 days is typically used. This is a dimensionless ratio. The credibility weight is determined based on the authority of the data source: officially released authoritative data has a credibility weight of 1.0, data released by third-party institutions has a credibility weight of 0.8, and data collected by the public has a credibility weight of 0.5. The block sparse matrix decomposition strategy divides the nationwide spatiotemporal weight matrix into 34 provincial block matrices according to provincial administrative divisions. Each provincial block matrix stores only the weight values ​​between blocks within the province and the weight values ​​between blocks at provincial boundaries. The compressed sparse row format stores the non-zero weight values, the column index values ​​of the non-zero weight values, and the row pointer value at the beginning of each row in row-major order.

[0090] The specific implementation of step S5 involves inputting the remote sensing image stored in step S2, the policy text data after semantic standardization in step S1, and the vector data segmented in step S3 into a multimodal semantic fusion model. This model extracts multimodal feature vectors and outputs a unified semantic representation vector. The multimodal semantic fusion model's structure includes an input layer that receives remote sensing imagery, policy text data, and vector data. The remote sensing image feature extraction branch uses a residual neural network structure to extract spectral feature vectors and spatial texture feature vectors. The policy text feature extraction branch uses a bidirectional encoder structure to extract semantic feature vectors and contextual feature vectors. The vector data feature extraction branch uses a graph convolutional network structure to extract topological feature vectors and attribute feature vectors. A cross-modal attention layer calculates the attention weight coefficients between different modal feature vectors. The calculation process involves first calculating the dot product of the query vector and the key vector and then dividing it by a scaling factor to obtain the attention score. The scaling factor calculation formula is as follows:

[0091] ;

[0092] In the formula, The scaling factor is dimensionless. The key vector dimension is dimensionless and typically takes the value 512. The original attention score calculation formula is as follows:

[0093] ;

[0094] In the formula, The original attention score is dimensionless. The query vector has a dimension of 512; This is the transpose of the key vector, with a dimension of 512; The dot product of the query vector and the transpose of the key vector is dimensionless. Then, the attention score is dynamically adjusted, as expressed in the following formula:

[0095] ;

[0096] In the formula, The adjusted attention score is dimensionless. The original attention score is dimensionless. For the plot of land With the plot of land Spatial weight values ​​between them, dimensionless; This is a normalized reference value for spatial weights, typically set to 1.0, and is dimensionless. The normalized spatial weight values ​​are dimensionless. For the plot of land With the plot of land The corresponding data credibility weight value is dimensionless. This is a normalized reference value for the credibility weight, typically set to 1.0, and is dimensionless. The normalized confidence weights are dimensionless. Finally, the adjusted attention scores are softmax normalized to obtain the attention weight coefficients. The softmax normalization function is expressed as follows:

[0097] ;

[0098] In the formula, For the first Each attention weight coefficient is dimensionless. For the first The adjusted attention score is dimensionless. The total number of attention scores is dimensionless. For summation indexing, dimensionless; It is an exponential function; The sum of all adjusted attention scores is an exponential function, dimensionless. The feature fusion layer concatenates the weighted multimodal feature vectors and maps them to a unified semantic representation vector through a fully connected layer. The output layer outputs a 512-dimensional unified semantic representation vector.

[0099] The specific implementation of step S6 involves calculating the semantic similarity value between different data sources based on the unified semantic representation vector output in step S5. When the semantic similarity value is lower than the similarity threshold, a semantic conflict is determined, and the rule weight value corresponding to the conflicting data is extracted. The semantic similarity value is calculated using cosine similarity, and the formula for calculating the vector inner product is as follows:

[0100] ;

[0101] In the formula, For vectors with vector The inner product value is dimensionless; For vectors The One component, dimensionless; For vectors The One component, dimensionless; The component index is dimensionless and ranges from 1 to 512; 512 is the dimension of the unified semantic representation vector, which is also dimensionless. The formula for calculating the vector magnitude is as follows:

[0102] ;

[0103] ;

[0104] In the formula, For vectors The modulus is dimensionless; For vectors The modulus is dimensionless; For vectors The The square of each component, dimensionless; For vectors The The semantic similarity value is the square of each component, dimensionless. The formula for calculating the semantic similarity value is as follows:

[0105] ;

[0106] In the formula, This is a semantic similarity value, dimensionless, and its value ranges from -1 to +1. This is the first unified semantic representation vector; This is the second unified semantic representation vector; The dot product of two vectors is dimensionless. For vectors The modulus is dimensionless; For vectors The modulus is dimensionless; This is a reference value for the inner product, usually taken as 100, and is dimensionless. This is a reference value for the module length, usually taken as 10, and is dimensionless. The normalized inner product value is dimensionless. It is the product of the magnitudes of two vectors, and is dimensionless. The square of the reference value for the modulus length is dimensionless. The normalized modulus product value is dimensionless. The similarity threshold is determined based on the statistical distribution of historical conflict data. The semantic similarity values ​​of historical conflict data are sorted from smallest to largest, and the semantic similarity value corresponding to the 10th percentile is selected as the similarity threshold. The current similarity threshold is set to 0.65. Rule weight values ​​include the rule's binding strength weight value and the rule's timeliness weight value. The binding strength weight value is determined according to the rule's legal level: the binding strength weight value for rules at the national law level is set to 1.0, the binding strength weight value for rules at the local regulation level is set to 0.7, and the binding strength weight value for rules at the departmental regulation level is set to 0.5. The timeliness weight value is calculated based on the rule's effective date, and the calculation formula is as follows:

[0107] ;

[0108] In the formula, For rules The timeliness weight value is dimensionless; For rules The number of days since it took effect, in days; For reference purposes, the timeframe is typically set at 1825 days. This is a rule-based index, dimensionless. It is a dimensionless ratio.

[0109] The specific implementation of step S7 involves constructing a rule priority directed graph using the rule weight values ​​extracted in step S6. A random walk algorithm is then used to calculate the importance score of each rule node in the rule priority directed graph. Conflicting rules are sorted according to their importance scores, and the rule with the highest priority is executed to resolve the conflict. The resolved aligned dataset is then output. Specifically, the rule priority directed graph uses conflicting rules as nodes and the priority relationships between rules as directed edges. If rule A takes precedence over rule B, a directed edge from A to B is established between nodes A and B. The random walk algorithm simulates a random walk process on the rule priority directed graph. The initial formula for calculating the importance score of each rule node is as follows:

[0110] ;

[0111] In the formula, For the first The initial importance score of each rule node, dimensionless; The total number of rule nodes, dimensionless; This is a dimensionless index for rule nodes. In each iteration, the importance score of the rule node is evenly distributed among all successor nodes pointed to by the rule node, with 85% reserved. The iteration formula is as follows:

[0112] ;

[0113] In the formula, For the first The rule node in the th The importance score after the next iteration is dimensionless. For the first The initial importance score of each rule node, dimensionless; The number of iterations is dimensionless and ranges from 0 to 49. Pointing to a node The set of predecessor nodes; For the first The predecessor node at the _th ... The importance score after the next iteration is dimensionless. For the first The number of successor nodes pointed to by each predecessor node, dimensionless. This is an index for the predecessor node, and is dimensionless. Assign all predecessor nodes to nodes The sum of importance scores is dimensionless. After 50 iterations, the importance scores of each rule node converge to a stable value. The conflict resolution process executes rules sequentially from highest to lowest importance score. When executing a rule, the data constrained by the rule is modified to meet the rule's requirements. Once a conflict is resolved after executing a rule, the execution of subsequent rules stops. The aligned dataset includes resolved geospatial data, policy text data, and socioeconomic data. All data in the aligned dataset maintains semantic consistency and does not exhibit policy-spatial conflicts. The aligned dataset is used for the delineation of the three zones and three lines and land use control decisions in territorial spatial planning.

[0114] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0115] The technical team first collected geospatial data for the planning area, including 8,672 scenes of Gaofen-2 satellite remote sensing imagery with a resolution of 2 meters, covering five years from 2020 to 2024, with a total image data volume of 127TB. Simultaneously, vector data was collected, including land use status layers, urban development boundary layers, and ecological protection red line layers, with a total of approximately 4.58 million vector elements. Regarding policy text data, 12 national-level legal documents, 56 provincial-level local regulations, and 203 municipal and county-level departmental rules were collected, with a total text volume of approximately 8.5 million words. Socioeconomic data came from statistical yearbooks and survey data, including indicators such as population, GDP, and land area, with a total of approximately 168,000 data records, covering both monthly and annual data granularities.

[0116] The technical team performed coordinate system unification processing on the collected geospatial data, transforming the raw data from different coordinate systems such as WGS84, Beijing 54, and Xi'an 80 to the CGCS2000 coordinate system. A seven-parameter transformation model was used to calculate the coordinate transformation parameters, and the accuracy error of the transformed coordinates was controlled within 0.5 meters. Semantic standardization processing was performed on the policy text data, mapping land use terms such as construction land, agricultural land, and unused land to the GB / T 21010-2017 land category coding system, identifying and standardizing 148 land use expressions. Time-series alignment processing was performed on the socio-economic data, converting monthly data into annual data through linear interpolation, and accumulating and adjusting annual data from different statistical calibers to unify them to a natural year time period.

[0117] like Figure 2 As shown, the technical team divided the remote sensing imagery into 256×256 pixel image tiles according to the QuadTree structure, generating a total of 213 million image tiles. A three-dimensional index structure of spatial range-timestamp-resolution was established. The spatial dimension used GeoHash encoding, dividing the planned area into 12 grid levels, with each grid corresponding to a unique GeoHash code. The temporal dimension used Unix timestamp sorting to record the acquisition time of each image tile. The resolution dimension adopted a pyramid hierarchy, constructing six pyramid levels from 0 to 5. The image tiles were stored in a Ceph-based distributed object storage system. Based on access frequency statistics, image tiles accessed more than 100 times per hour were stored in the solid-state drive layer, accounting for 18% of the total number of tiles, while the remaining tiles were stored in the hard disk array layer.

[0118] The technical team used the METIS graph segmentation algorithm to divide the vector data into 31 spatial subgraphs, each corresponding to the administrative region of a prefecture-level city. The graph segmentation process constructs the graph structure based on the spatial adjacency relationships of vector features, achieving balanced partitioning by minimizing the number of edges between subgraphs. After segmentation, the number of nodes in each subgraph ranges from 120,000 to 180,000, and the proportion of edges crossing subgraphs is 7.3% of the total number of edges. A boundary node feature cache pool was established for each spatial subgraph, identifying a total of 334,000 vector features located at the subgraph boundaries. The feature vectors of these boundary nodes were stored in the cache pool, with a feature vector dimension of 128.

[0119] like Figure 3 As shown, the technical team constructed a spatiotemporal weight matrix to calculate spatial, temporal, and reliability weight values ​​for 4.58 million plots within the planning area. Spatial weight values ​​were calculated using a Gaussian kernel function, with a reference distance set at 1000 meters. Plot pairs with an Euclidean distance of less than 500 meters between plots received a spatial weight value greater than 0.78. Temporal weight values ​​were calculated using an exponential decay function, with a reference time period set at 365 days. Data pairs with a time interval of less than 180 days received a temporal weight value greater than 0.61. Reliability weight values ​​were determined based on the data source: data provided by the National Surveying and Mapping Department was set at 1.0, data provided by provincial land departments at 0.8, and city / county level survey data at 0.5. A block-sparse matrix decomposition strategy was employed to decompose the global spatiotemporal weight matrix into three provincial block matrices according to provincial administrative divisions. Each block matrix stores only the weight values ​​between plots within the province and the weight values ​​between plots at provincial boundaries, achieving a sparsity of 96.7% for the weight matrix. The provincial block matrix is ​​stored using a compressed sparse row format, reducing the storage space from 2.8TB of the original matrix to 92GB.

[0120] The technical team input preprocessed remote sensing imagery, policy text data, and vector data into a multimodal semantic fusion model. The remote sensing imagery feature extraction branch of the model employs a ResNet-50 residual neural network structure, with initial weights set to ImageNet pre-trained weights, extracting 256-dimensional spectral feature vectors and 256-dimensional spatial texture feature vectors. The policy text feature extraction branch employs a BERT bidirectional encoder structure, with initial weights set to the Wudao model pre-trained weights, extracting 768-dimensional semantic feature vectors and 768-dimensional contextual feature vectors. The vector data feature extraction branch employs a GraphSAGE graph convolutional network structure, with initial weights set to a random normal distribution, extracting 512-dimensional topological feature vectors and 512-dimensional attribute feature vectors. The attention weight coefficients between different modal feature vectors are calculated in the cross-modal attention layer. The attention scores are dynamically adjusted based on the spatial weight values ​​and confidence weight values ​​in the spatiotemporal weight matrix. The distribution of the adjusted attention weight coefficients is shown in Table 1.

[0121] Table 1. Distribution of Cross-Modal Attention Weight Coefficients

[0122]

[0123] The feature fusion layer concatenates the weighted multimodal feature vectors to form a 3584-dimensional comprehensive feature vector, which is then mapped to a 512-dimensional unified semantic representation vector through two fully connected layers. Model training employs a contrastive learning loss function, with a batch size of 32 and a learning rate of 0.0001. Training is conducted for 100 epochs on a training set containing 500,000 triplet samples. The validation set loss reaches its minimum of 0.238 in the 67th epoch and remains unchanged for five consecutive epochs, at which point training is stopped and the model weight parameters are saved.

[0124] like Figure 4 As shown, the technical team calculated semantic similarity values ​​between different data sources based on a unified semantic representation vector, using cosine similarity as the metric. The team set a similarity threshold of 0.65 and identified 8763 pairs of semantically conflicting data through semantic similarity calculations. Conflict types included inconsistencies between policy constraints and current land use, overlap between ecological protection red lines and development boundaries, and contradictions between overall land use planning and urban master planning. The team extracted rule weight values ​​corresponding to the conflicting data, including constraint strength weight values ​​and timeliness weight values. The constraint strength weight value was determined according to the legal level of the rule: 1.0 for 12 national-level legal documents, 0.7 for 56 provincial-level local regulations, and 0.5 for 203 municipal and county-level departmental rules. The timeliness weight value was calculated based on the rule's effective date: rules effective less than 3 years ago had a timeliness weight value greater than 0.72, while rules effective more than 5 years ago had a timeliness weight value less than 0.37.

[0125] The technical team constructed the conflict rules as a directed graph of rule priorities, containing 271 rule nodes and 542 directed edges. These rules cover specific control requirements such as the classification of protection intensity for different land types, development restrictions for different areas, and land use conversion conditions for different purposes. A random walk algorithm was used to calculate the importance score of each rule node. Initially, the importance score of each rule node was set to 0.00369. In each iteration, each rule node distributed 85% of its importance score equally to its successor nodes and retained 15%. After 50 iterations, the importance scores of each rule node converged to a stable value. The top 10 rules in terms of importance score are shown in Table 2.

[0126] Table 2. Ranking of Rule Importance Scores

[0127]

[0128] like Figure 5As shown, the technical team executed conflict resolution rules in descending order of importance score, modifying conflicting data to meet the requirements of the highest priority rule during execution. In processing 8763 sets of semantic conflict data, the rule prohibiting construction and development in primary protected areas resolved 2841 conflicts, primarily concerning overlaps between the core area of ​​nature reserves and construction land planning. The rule strictly controlling linear engineering projects within ecological red lines resolved 1967 conflicts, involving situations where linear infrastructure such as roads and pipelines cross ecological protection red lines. The rule requiring zero compensation for permanent basic farmland resolved 1523 conflicts, mainly addressing the conflict between basic farmland protection and urban expansion. Other rules resolved 2432 conflicts, including various rule types such as urban development boundary control, shoreline protection, wetland buffer zones, cultural heritage protection, land use conversion approval, green space ratio control, and plot ratio restrictions. After conflict resolution, an aligned dataset is output. This dataset contains geospatial data, policy text data, and socioeconomic data that are coordinate system 1, semantically standardized, and temporally consistent. All data maintain semantic consistency and are free from policy-spatial conflicts. The technical team applied the aligned dataset to the delineation of the "three zones and three lines" and land use control decisions in the territorial spatial planning of the Yangtze River Midstream Urban Agglomeration, achieving precise delineation of ecological protection red lines, permanent basic farmland protection red lines, and urban development boundaries.

[0129] This invention represents a significant technological advancement over traditional methods that combine manual interpretation with database association. Traditional methods rely on manual comparison of spatial data from different sources and matching based on predefined rules. This shallow feature matching approach fails to capture the deep semantic relationships within the data, resulting in insufficient accuracy in semantic understanding. This invention constructs a multimodal semantic fusion model, utilizing residual neural networks, bidirectional encoders, and graph convolutional networks to automatically learn abstract semantic features from raw data. This maps heterogeneous data to a unified high-dimensional semantic space, achieving semantic comparability between different modalities. Traditional methods suffer from efficiency bottlenecks when processing massive amounts of multimodal data; manual review cannot handle millions of land parcels and thousands of policy documents. This invention, however, establishes an efficient data organization and retrieval mechanism by segmenting images using a quadtree structure, dividing vector data using graph segmentation algorithms, and compressing and storing weight values ​​using a block sparse matrix. Traditional conflict resolution methods rely on expert experience to rank rules, which is highly subjective and difficult to quantify rule priorities. This invention transforms conflict resolution into a graph theory problem by using a directed graph of rule priorities and a random walk algorithm. This enables automated calculation of rule importance and intelligent decision-making in conflict handling, improving the objectivity and consistency of conflict resolution.

[0130] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0131] Table 3. Variable Explanation Table

[0132]

[0133] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for aligning multimodal data in land spatial planning, characterized in that, The process includes collecting and preprocessing geospatial data, policy text data, and socioeconomic data; segmenting remote sensing images into image tiles using a quadtree structure and storing them in a distributed object storage system using a 3D index structure; using a graph segmentation algorithm to divide vector data into multiple spatial subgraphs and establishing a boundary node feature cache pool; constructing a spatiotemporal weight matrix to calculate the spatial weight, temporal weight, and confidence weight values ​​for each plot; inputting remote sensing images, policy text data, and vector data into a multimodal semantic fusion model to extract a unified semantic representation vector; calculating semantic similarity values ​​and determining semantic conflicts based on the unified semantic representation vector; extracting the rule weight values ​​corresponding to conflicting data to construct a rule priority directed graph; using a random walk algorithm to calculate the importance score of rule nodes; executing conflict resolution rules according to priority; and outputting an aligned dataset.

2. The method according to claim 1, characterized in that, The preprocessing includes standardizing the geospatial data to the CGCS2000 coordinate system, semantically standardizing the policy text data to the GB / T 21010-2017 land category coding, and aligning the socio-economic data to a unified time period.

3. The method according to claim 2, characterized in that, The coordinate system unification process employs a seven-parameter transformation model to transform geospatial data from different sources from the original coordinate system to the National 2000 coordinate system.

4. The method according to claim 3, characterized in that, The image tile has a size of 256×256 pixels, and the three-dimensional index structure includes a spatial dimension index encoded with GeoHash, a time dimension index sorted by timestamp, and a resolution dimension index with a pyramid hierarchy.

5. The method according to claim 4, characterized in that, The distributed object storage system adopts a Ceph storage architecture, storing frequently accessed image tiles in the solid-state drive layer and infrequently accessed image tiles in the hard disk array layer.

6. The method according to claim 5, characterized in that, The graph partitioning algorithm uses the METIS algorithm to achieve a balanced partitioning of the graph by minimizing the number of edges between spatial subgraphs.

7. The method according to claim 6, characterized in that, The spatial weight value is calculated using a Gaussian kernel function, the time weight value is calculated using an exponential decay function, and the credibility weight value is determined based on the authority of the data source.

8. The method according to claim 7, characterized in that, The spatiotemporal weight matrix is ​​decomposed into provincial block matrices according to the provincial administrative divisions using a block sparse matrix decomposition strategy, and the non-zero weight values ​​are stored in the provincial block matrices using a compressed sparse row format.

9. The method according to claim 8, characterized in that, The multimodal semantic fusion model includes a remote sensing image feature extraction branch using a residual neural network structure, a policy text feature extraction branch using a bidirectional encoder structure, and a vector data feature extraction branch using a graph convolutional network structure.

10. The method according to claim 9, characterized in that, The multimodal semantic fusion model includes a cross-modal attention layer that dynamically adjusts the attention weight coefficients between different modal feature vectors based on the spatial weight values ​​in the spatiotemporal weight matrix and the credibility weight values ​​of the data source.

Images