Data fusion methods and data fusion systems

Abstract

This application provides a data fusion method and a data fusion system. The method includes: acquiring relevant data for the region to be evaluated; classifying the relevant data to obtain classified relevant data; performing a first fusion on relevant data of the same type to obtain multiple similar fusion features; performing a second fusion on all similar fusion features to obtain a comprehensive fusion feature; and performing a resource assessment on at least the region to be evaluated based on the comprehensive fusion feature. This solution solves the problem in existing technologies where resource assessment of a region is performed using a simple weighted average. This simple data weighted average fusion method suffers from severe information entropy loss, resulting in poor accuracy of the data fusion results and consequently, poor accuracy in automated resource assessment of a region.

Description

Technical Field

[0001] This application relates to the field of data fusion technology, and more specifically, to a data fusion method and a data fusion system. Background Technology

[0002] In modern regional management and spatial planning, it is often necessary to conduct resource assessments of a region. There are some schemes for regional resource assessment, but these schemes only use single data for assessment, or use multiple data to conduct assessments through simple weighted averages. This simple data weighted average fusion method suffers from severe information entropy loss and poor accuracy of data fusion results, leading to poor accuracy in automated resource assessment of a region. Summary of the Invention

[0003] The main objective of this application is to provide a data fusion method and a data fusion system to at least solve the problem that in the prior art, resource assessment of a certain area is carried out by a simple weighted average. This simple data weighted average fusion method results in serious loss of information entropy and poor accuracy of the data fusion results, leading to poor accuracy in automated resource assessment of a certain area.

[0004] To achieve the above objectives, according to one aspect of this application, a data fusion method is provided, comprising: acquiring relevant data of a region to be evaluated, wherein the relevant data includes at least remote sensing image data, pedestrian flow data, charging equipment data, vehicle flow data, air quality data, and news data; classifying the relevant data to obtain classified relevant data, wherein the types of the classified relevant data include spatial geographic type, structured type, unstructured type, and time-series type, wherein the spatial geographic type is the relevant data related to the spatial location of the region to be evaluated, the structured type is the relevant data with data structure, the unstructured type is the relevant data without data structure, and the time-series type is the relevant data related to time; performing a first fusion on the relevant data of the same type to obtain multiple similar fusion features, wherein the first fusion is feature fusion using machine learning, and the number of similar fusion features is the same as the number of types of relevant data; performing a second fusion on all the similar fusion features to obtain a comprehensive fusion feature, wherein the second fusion is feature fusion using machine learning; and performing a resource assessment on at least the region to be evaluated based on the comprehensive fusion feature.

[0005] According to another aspect of this application, a data fusion system is provided, comprising: one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including methods for performing any of the data fusion methods described above.

[0006] By applying the technical solution of this application, machine learning can automatically learn the nonlinear mapping relationship within the data, extract the most representative core feature vector from the original high-dimensional noisy data and perform a first fusion (i.e., "same-type fusion feature"), and then perform a second fusion on the "same-type fusion feature" through machine learning. This allows the learning of the correlation and interaction between feature vectors of different modalities. Through a two-level architecture of "intra-modal machine learning feature fusion (first fusion)" and "cross-modal machine learning global fusion (second fusion)," the traditional shallow weighted average is replaced, which can retain the information entropy of the original data to the greatest extent, thereby improving the accuracy of data fusion and the accuracy of automated resource assessment for a certain region. Attached Figure Description

[0007] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0008] Figure 1 A hardware structure block diagram of a mobile terminal providing an execution data fusion method according to an embodiment of this application is shown;

[0009] Figure 2 A flowchart illustrating a data fusion method according to an embodiment of this application is shown;

[0010] Figure 3 A structural block diagram of a data fusion apparatus provided according to an embodiment of this application is shown.

[0011] The above figures include the following reference numerals:

[0012] 102. Processor; 104. Memory; 106. Transmission device; 108. Input / output device. Detailed Implementation

[0013] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0015] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0016] For ease of description, the following explains some of the nouns or terms used in the embodiments of this application:

[0017] Multi-source heterogeneous data refers to data from different data sources with different data structures, modalities, spatiotemporal references, and semantic dimensions. In this solution, it specifically refers to multimodal data such as regionally related spatial geographic data, structured statistical data, unstructured text data, and time-series behavioral data.

[0018] Coordinate system: Geodetic coordinate system, unified geocentric coordinate system, used in this scheme to unify the coordinate benchmark of all spatial geographic data and solve the problem of spatial misalignment.

[0019] Intramodal fusion: Feature-level fusion of multi-source data within the same data modality (such as spatial raster data or structured data) to fully extract the core effective features of the modality and eliminate data redundancy within the modality.

[0020] Cross-modal fusion: Globally correlate and fuse the fusion features of different data modalities, explore the potential correlations between features of different modalities, and form a unified global fusion feature map.

[0021] Graph Neural Networks (GNNs): Models for deep learning based on graph structure data. In this scheme, the fusion features of different modalities are used as graph nodes to learn the association weights between nodes, thereby achieving cross-modal heterogeneous data fusion.

[0022] Spatiotemporal attention mechanism: An improved version of the attention mechanism, which can simultaneously capture the spatial location correlation and temporal series correlation of data. In this scheme, it is used to dynamically allocate fusion weights for different modalities and spatiotemporal dimensions of features to improve fusion accuracy.

[0023] Meta-learning, also known as learning-to-learn, is a machine learning method that enables models to quickly adapt to small datasets and avoid overfitting. In this solution, it is used to address the problem of insufficient regional data samples.

[0024] Spatial constraints: The defined control boundaries are compliance constraints that must be followed in resource assessment.

[0025] Suitability assessment: Based on multiple dimensions such as resources, ecological environment, supporting facilities, and demand, the suitability of a specific area for development is evaluated in a graded manner, which is the core link in the planning assessment.

[0026] Stacked Autoencoder (SAE): An unsupervised deep learning model composed of multiple stacked autoencoders. In this scheme, it is used for feature extraction and fusion of structured statistical data to achieve dimensionality reduction of high-dimensional features and effective information extraction.

[0027] Current resource assessment and planning suffers from several key industry pain points: 1) Severe data silos: Data is distributed across multiple departments, with inconsistent data formats, standards, and spatiotemporal benchmarks, making effective integration impossible. Planning relies heavily on expert experience and lacks scientific rigor. 2) Poor plan implementation: Many plans fail to adequately consider actual constraints, resulting in inconsistencies with those constraints. 3) Low level of digitalization: Data samples are limited and labels are incomplete. Existing digitalization methods are not suitable, and planning cannot achieve dynamic optimization throughout its entire lifecycle.

[0028] Data-driven approaches have become an inevitable trend, and the fusion of multi-source heterogeneous data is the core foundation for achieving data-driven planning. Existing technologies cannot meet the business needs of the entire resource assessment process, and there are obvious technological gaps.

[0029] The development of multi-source data fusion technology has progressed from early shallow fusion methods such as weighted averaging, DS evidence theory, and Bayesian estimation to feature-level deep fusion methods based on deep learning. This has significantly improved fusion accuracy and adaptability, and has been widely applied in smart cities, remote sensing interpretation, and recommendation systems. However, existing applications of multi-source data fusion in resource assessment are mostly concentrated in city-level passenger flow prediction and operational optimization. Fusion methods specifically for tourism planning (i.e., resource assessment in this solution) are extremely rare, and those that exist are simply transfers of general fusion methods, exhibiting three core technical shortcomings: 1) The fusion architecture is not designed to adapt to the multimodal data characteristics of tourism planning, resulting in significant loss of spatial, temporal, and semantic features and insufficient fusion accuracy; 2) It is not deeply integrated with the entire tourism planning process, and the fusion results can only serve a single resource evaluation stage, failing to support decision-making throughout the entire planning process and exhibiting poor practicality; 3) Without embedded constraints, the fusion results are risky and cannot be directly used for resource assessment.

[0030] The core challenge that existing technologies have not yet solved is how to achieve high-precision, high-practicability, and high-compliance integration of multi-source heterogeneous data in scenarios with small samples, multiple modalities, and strong compliance constraints, so as to provide implementable data support for the entire tourism planning process.

[0031] In the current field of resource assessment, the closest existing technology to this solution is the commonly used industry-standard resource evaluation technology based on multi-source data weighted fusion. This technology involves multi-source data collection and basic standardized preprocessing, combined with subjective weight assignment, and employs a single-level, numerical-level weighted summation shallow fusion method to achieve simple integration of multi-source data. Ultimately, it serves only the single stage of resource evaluation. The detailed architecture is as follows:

[0032] 1. Multi-source data collection: Collecting various types of data related to resources, including basic geographic data, ecological environment data, and network data;

[0033] 2. Data preprocessing: The collected data is filled with missing values, outliers are removed, and the data is standardized to unify the data format and units;

[0034] 3. Construction of evaluation index system: Establish a multi-level index system for resource evaluation, generally divided into target layer, criterion layer, and indicator layer;

[0035] 4. Multi-source data fusion and comprehensive scoring: Subjective weighting methods such as the Analytic Hierarchy Process (AHP) and the Delphi method are used to determine the weights of each indicator. The shallow fusion of multi-source data is completed based on the weighted summation method to obtain the comprehensive resource score of each evaluation unit.

[0036] 5. Results Output: Based on the integrated scoring results, the resources are graded and evaluated, and the resource evaluation results are output.

[0037] Therefore, the existing solution has the following drawbacks:

[0038] Disadvantage 1: The fusion architecture is too simple and cannot adapt to the characteristics of multimodal heterogeneous data, resulting in insufficient fusion accuracy and loss of core features.

[0039] Because existing general technical solutions employ a single-level, weighted summation-based shallow fusion approach, they fail to design adaptive feature extraction and fusion models tailored to the structural characteristics of different modalities of data. For spatial geographic raster / vector data, they cannot effectively extract core spatial features such as spatial topology, location relationships, and terrain; for unstructured text data, they cannot effectively extract core semantic features such as semantics, sentiment, and user preferences; and for time-series passenger flow and traffic data, they cannot effectively extract core time-series features such as time-series trends and periodic fluctuations. This results in a significant loss of core effective features across different modalities. The shallow fusion achieved solely through weighted summation fails to uncover potential correlations between features, leading to low accuracy in the fusion results. The fusion cannot reflect the full-dimensional feature information required for planning, ultimately failing to support refined end-to-end decision-making and only enabling simple resource scoring.

[0040] Disadvantage 2: It is not deeply integrated with the entire business process, resulting in poor usability and insufficient business adaptability of the integration results.

[0041] Because existing general technical solutions only design the integration method for the single stage of resource evaluation, they do not design corresponding feature systems and integration logic for the entire process of resource survey, suitability evaluation, business layout, passenger flow forecasting, and benefit assessment. The integration process only revolves around the indicator system of resource evaluation, without incorporating the traffic accessibility and passenger flow carrying capacity required for business layout, the market demand and holiday fluctuations required for passenger flow forecasting, and the core business characteristics required for benefit assessment. As a result, the integration results can only serve the single stage of resource evaluation and cannot be adapted to other core stages of the entire planning process. Data processing and integration need to be repeated for different stages, resulting in extremely low work efficiency and severely insufficient business applicability of the integration results.

[0042] Disadvantage 3: Without embedded constraints, the fusion results pose compliance risks and cannot be directly implemented.

[0043] Because existing general technical solutions do not embed constraints as the core fusion logic into the entire process, but only include ecological and environmental data as ordinary evaluation indicators in the weighted fusion, they cannot perform full-process compliance verification of the fusion results, which easily leads to situations where high-scoring areas do not meet the constraints. Since the constraints are only used as ordinary indicators, they cannot achieve compliance control with a veto power. The fusion results cannot meet the constraints, resulting in conflicts between the fusion results and the constraints, which poses serious compliance risks. They cannot be directly used for resource assessment and have extremely poor implementation.

[0044] Disadvantage 4: It is not adapted for small sample data scenarios, resulting in poor model generalization ability and easy overfitting.

[0045] Because existing general technical solutions use traditional weighted fusion methods without introducing optimization mechanisms for small sample adaptation, and most regions generally have limited data samples, incomplete labels, and missing data dimensions, when data samples are insufficient, it is impossible to learn effective feature relationships from limited samples, and the weight assignment is highly subjective. If data from other regions is introduced for model training, problems such as regional characteristic mismatch and model overfitting are likely to occur, resulting in the fusion method being usable only in regions with complete data, unable to adapt to most actual situations, and having a very narrow scope of application.

[0046] Disadvantage 5: No iterative optimization mechanism for the fusion results has been established, making it unable to adapt to the dynamic changes in regional development.

[0047] Because the existing general technical solutions use a static, one-off integration logic, they do not establish an iterative optimization mechanism for integration accuracy based on planning implementation feedback and dynamic changes. Market demand, passenger flow, environment, and industrial development are all dynamic changes, and static integration results cannot reflect these dynamic changes, resulting in poor timeliness of integration results. As time goes by, the accuracy declines rapidly and cannot continuously support the dynamic adjustment and operational optimization of resource assessment planning.

[0048] As described in the background section, existing technologies assess resources in a region using a simple weighted average. This simple data weighted average fusion method suffers from severe information entropy loss and poor accuracy of the data fusion results, leading to poor accuracy in automated resource assessment of a region. To address the above problems, embodiments of this application provide a data fusion method and a data fusion system.

[0049] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0050] The methods and embodiments provided in this application can be executed on a mobile terminal, computer terminal, or similar computing device. Taking running on a mobile terminal as an example, Figure 1 This is a hardware structure block diagram of a mobile terminal for a data fusion method according to an embodiment of the present invention. Figure 1 As shown, a mobile terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those skilled in the art will understand that... Figure 1The structure shown is for illustrative purposes only and does not limit the structure of the mobile terminal described above. For example, the mobile terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0051] The memory 104 can be used to store computer programs, such as application software programs and modules, like the computer program corresponding to the data fusion method in this embodiment of the invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, thereby implementing the above-described method. The memory 104 may include high-speed random access memory and non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the mobile terminal via a network. Examples of the aforementioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The transmission device 106 is used to receive or send data via a network. Specific examples of the aforementioned networks may include wireless networks provided by the mobile terminal's communication provider. In one example, the transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to communicate with the Internet. In one example, the transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0052] This embodiment provides a method for fusing data running on a mobile terminal, computer terminal, or similar computing device. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0053] Figure 2 This is a flowchart illustrating a data fusion method according to an embodiment of this application. Figure 2 As shown, the method includes the following steps:

[0054] Step S201: Obtain relevant data for the area to be evaluated, including at least remote sensing image data, pedestrian flow data, charging equipment data, vehicle flow data, air quality data, and news data.

[0055] Specifically, the area to be assessed can be any area requiring spatial resource assessment or planning, and the aforementioned multi-source data can be obtained through data interfaces, database queries, or file uploads. Among these, remote sensing image data provides spatial information such as land cover and topography; pedestrian and vehicle traffic data reflects the region's passenger and traffic dynamics; charging equipment data indicates the status of infrastructure; air quality data assesses the quality of the ecological environment; and news data includes textual information such as trending topics and publicly available forum data.

[0056] Step S202: Classify the above-mentioned relevant data to obtain classified relevant data. The types of the classified relevant data include spatial geographic type, structured type, unstructured type and time series type. The spatial geographic type is the relevant data related to the spatial location of the area to be evaluated. The structured type is the relevant data with data structure. The unstructured type is the relevant data without data structure. The time series type is the relevant data related to time.

[0057] Specifically, the acquired raw data is automatically categorized using a pre-defined classification rule engine. The specific classification logic is as follows: remote sensing image data (such as GeoTIFF format, containing latitude and longitude coordinates and pixel matrices) is categorized as spatial geographic type; charging device data (such as tabular data containing device ID, latitude and longitude, and status) and air quality data (such as structured records containing station ID, values, and time) are categorized as structured type (if it contains time series data, it can also be categorized as time series; here, the categorization is based on the main features, or flexibly adjusted according to subsequent processing needs. In this embodiment, data with strong static statistical attributes is categorized as structured, and data with strong dynamic changes is categorized as time series); news data (such as TXT and HTML text) is categorized as unstructured type; and pedestrian flow data and vehicle flow data (such as sequences containing timestamps and corresponding values) are categorized as time series type. By detecting the data format, field type, and metadata information, the raw data is divided into these four categories so that specific machine learning models can be used to process different types.

[0058] Step S203: Perform a first fusion on the above-mentioned related data of the same type to obtain multiple similar fusion features. The first fusion is a feature fusion using machine learning. The number of similar fusion features is the same as the number of types of related data.

[0059] Specifically, for spatial geographic data (remote sensing imagery), the data is input into an improved U-Net model. This model extracts multi-scale spatial features through an encoder, restores the resolution through a decoder, and introduces a spatial attention mechanism to highlight key areas (such as buildings and vegetation), outputting a spatially fused feature vector.

[0060] For structured data (charging devices, air quality statistics), the data is input into an improved stacked autoencoder (SAE). The SAE performs nonlinear dimensionality reduction and reconstruction through a multi-layer fully connected network, extracting compact feature representations of the high-dimensional data and outputting a structured fused feature vector.

[0061] For unstructured data (news data), the input is fed into the improved BERT-wwm model. First, the text is cleaned and segmented, then the contextual semantics are extracted by the BERT encoder, and finally, the topic attention module performs weighted aggregation to output a text semantic fusion feature vector.

[0062] For time-series data (pedestrian flow, vehicle flow), the input is fed into an improved Transformer encoder. Learnable positional encoding and multi-head self-attention mechanisms are used to capture long-term dependencies and periodic patterns in the time series, outputting a time-series fusion feature vector.

[0063] Through the four independent machine learning models described above, four types of fused features of the same kind—spatial, structured, textual, and temporal—are obtained respectively. Each feature vector represents a core information summary of the data in that modality. The first fusion is performed using a machine learning model dedicated to each modality, which avoids the neglect of deep structural information within the data by traditional weighted averaging. This fully preserves the unique semantic and structural features of each modality, significantly reduces information entropy loss, and improves the fidelity of feature representation.

[0064] Step S204: Perform a second fusion on all the above-mentioned similar fusion features to obtain a comprehensive fusion feature, wherein the second fusion is a feature fusion performed using machine learning.

[0065] Specifically, the spatial fusion features, structured fusion features, textual semantic fusion features, and temporal fusion features obtained above are used as input nodes to construct a heterogeneous graph structure. Each modality's feature is treated as a node, and nodes are connected by edges established through the Pearson correlation coefficient or predefined association rules between features. This is then fed into an improved graph neural network (GNN) model. This GNN model embeds a spatiotemporal attention mechanism during message passing, dynamically calculating the importance weights between nodes of different modalities, while utilizing residual connections to alleviate gradient vanishing and oversmoothing problems. After multiple layers of graph convolution operations, the model aggregates the information from all nodes, outputting a comprehensive fusion feature vector containing global association information across all modalities. This process achieves deep cross-modal interaction and fusion, uncovering potential nonlinear associations between different data sources.

[0066] Step S205: Based on the above-mentioned comprehensive integration characteristics, at least the above-mentioned areas to be evaluated shall be subject to resource assessment.

[0067] Specifically, the integrated feature vector obtained above is input into a preset resource assessment model (such as a fully connected neural network or regression model). This assessment model, trained with historical data, can map the resource level, suitability score, or potential index of the area to be assessed based on the integrated features.

[0068] First, a resource assessment model is constructed, which is a pre-trained supervised learning model (such as a multilayer perceptron or support vector machine). Second, the aforementioned integrated features are input as input vectors into the resource assessment model. Third, the resource assessment model maps the high-dimensional integrated features into low-dimensional resource assessment indicators through its internal nonlinear mapping layer. These resource assessment indicators include resource endowment scores, ecological suitability scores, and development potential scores. Finally, based on a preset scoring threshold, the resource assessment indicators are converted into resource levels.

[0069] In this embodiment, machine learning can automatically learn the nonlinear mapping relationships within the data, extract the most representative core feature vectors from the original high-dimensional noisy data and perform a first fusion (i.e., "same-type fusion features"). By performing a second fusion on the "same-type fusion features" through machine learning, the correlation and interaction between feature vectors of different modalities can be learned. Through the two-level architecture of "intra-modal machine learning feature fusion (first fusion)" and "cross-modal machine learning global fusion (second fusion)," the traditional shallow weighted average is replaced, which can preserve the information entropy of the original data to the greatest extent, thereby improving the accuracy of data fusion and the accuracy of automated resource assessment of a certain area.

[0070] Specifically, the overall technical solution consists of six core steps, with business rules and constraints embedded throughout the entire process. The following is a detailed explanation of each step: S1, Multi-source heterogeneous data collection and standardized preprocessing; S2, Feature hierarchical screening and weight assignment based on the entire planning process; S3, Deep fusion of two-level multi-source heterogeneous data; S4, Full-process compliance verification and correction of the fusion results; S5, Business-adaptive output of the fusion results; S6, Iterative optimization of fusion accuracy.

[0071] In the specific implementation process, the aforementioned relevant data is classified to obtain the classified relevant data, which can be achieved through at least one of the following steps: If the aforementioned relevant data meets the first condition, the type of the aforementioned relevant data is determined to be the aforementioned spatial geographic type, resulting in the classified relevant data. The first condition is that the aforementioned relevant data includes one or more of longitude, latitude, and coordinates. If the aforementioned relevant data meets the second condition, the type of the aforementioned relevant data is determined to be the aforementioned structured type, resulting in the classified relevant data. The second condition is that the aforementioned relevant data includes one or more of tables, key-value pairs, numeric fields, enumerated fields, and boolean fields. If the aforementioned relevant data meets the third condition, the type of the aforementioned relevant data is determined to be the aforementioned unstructured type, resulting in the classified relevant data. The third condition is that the aforementioned relevant data includes one or more of text, JSON fields, and XML fields. If the aforementioned relevant data meets the fourth condition, the type of the aforementioned relevant data is determined to be the aforementioned time-series type, resulting in the classified relevant data. The fourth condition is that the aforementioned relevant data includes one or more of timestamps, timelines, and time-type fields.

[0072] This solution achieves accurate identification and organization of multi-source heterogeneous data through an automated classification mechanism based on the inherent attributes of the data (coordinates, structure, content type, and time characteristics). It solves the problems of low efficiency and inconsistent standards in manual classification in existing technologies, which lead to chaotic data preprocessing. It provides accurate data input for subsequent feature extraction using dedicated machine learning models for different modal characteristics, thereby improving the automation level and processing accuracy of the overall data fusion process.

[0073] In this embodiment, when the metadata of the aforementioned related data contains longitude, latitude, or coordinate information, it is determined that the first condition is met. For example, remote sensing image data usually contains a GeoTIFF header file, which includes projection parameters and a geographic reference matrix, from which latitude and longitude coordinates can be extracted; if charging device data comes from a GIS database, its attribute table usually contains "longitude" and "latitude" fields; if news data contains location tags, coordinate information can also be extracted. By parsing the header information of the data file or the database, it is detected whether there is a geospatial reference system identifier or specific latitude and longitude values. If so, the aforementioned related data is classified as a spatial geographic type.

[0074] When the aforementioned data is presented in a two-dimensional table structure (such as CSV, Excel, or SQL database tables) or a key-value pair structure (such as simple attributes in a JSON object), and its field types are primarily numeric, enumerated, or boolean (such as whether it is powered on or open), it is determined to meet the second condition. For example, if air quality data is stored in a table format, containing columns such as "Site ID," "PM2.5 value," and "Monitoring time," and the numeric columns are numeric, then it belongs to a structured type. By reading the data definition, the data type and structural characteristics of the fields are determined. If they conform to the aforementioned structured characteristics, the aforementioned data is classified as a structured type.

[0075] The third condition is met when the aforementioned data is free text (such as news content in a TXT document or social media comments) or contains semi-structured data with a non-fixed structure (such as complex nested JSON objects or XML markup language). For example, news data is usually stored in HTML or TXT format, containing a large number of natural language characters, and cannot be directly mapped to a simple numerical table; although user comments on a platform may be stored in JSON format, their "comment content" field is long text with an inconsistent structure, making it difficult to extract a uniform numerical field. By attempting to parse the structure of the data, if it is found that the main content is natural language or a complex nested structure without a fixed numerical pattern, the aforementioned data is classified as unstructured.

[0076] When the aforementioned data possesses clear time-series attributes, including timestamps, timeline indexes, or time-type fields, and the data values ​​change over time, it is determined that it meets the fourth condition. For example, pedestrian traffic data is typically recorded as a sequence of "date-hour-number of people," while vehicle traffic data is recorded as a sequence of "time point-traffic value." By checking whether the data contains an ordered time index field (such as "2023-10-01 08:00:00") and verifying whether the data exhibits a trend of change over time, if the conditions are met, the aforementioned data is classified as a time-series type.

[0077] Specifically, for step S1, multi-source heterogeneous data acquisition and standardized preprocessing, the core objective of this step is to address the issues of spatiotemporal misalignment, inconsistent formats, and noise interference in multi-source data, providing a standardized, high-quality dataset for subsequent fusion. Multi-source heterogeneous data acquisition covers the entire process, collecting data from six core data sources, clarifying the data modality, source, and business application, as shown in Table 1.

[0078] Table 1

[0079]

[0080] Specifically, for step S2, which involves feature-based hierarchical screening and weight assignment throughout the planning process, the core objective is to embed business rules and constraints into the feature system, addressing the disconnect between existing technology and business needs, and constructing a business-adaptive feature set. This mainly includes the following:

[0081] 1. Construct a three-level feature system

[0082] For the five core stages of resource assessment and planning, a three-tiered feature system is constructed: 1) Basic Feature Layer: All pre-processed original features form the basis of the feature system; 2) Business Feature Layer: For each of the five planning stages, specific business features are extracted, such as resource features, constraint features, and transportation accessibility features in the suitability assessment stage; and passenger flow carrying capacity features, supporting facility features, and demand features in the business layout stage; 3) Decision Feature Layer: Comprehensive features oriented towards the final planning decision, including cross-stage integration of multiple business / basic features and comprehensive benefit features.

[0083] The three-tier feature system clearly defines the boundaries and differences of each level from five core dimensions: data source, processing depth, business relevance, reuse scope, and decision value, as shown in Table 2:

[0084] Table 2

[0085]

[0086] The classification execution logic is as follows: first, all preprocessed data are collected as basic features; then, for each planning stage, the basic features are combined into specific business features; finally, cross-stage business features are extracted and integrated to generate decision features.

[0087] 2. Constraint Feature Marking

[0088] The features of predefined prohibited development areas are marked as veto-type constraint features. In subsequent fusion and evaluation, units with a feature value of 0 (located in prohibited development areas) are directly excluded from the scope.

[0089] 3. Calculate feature weights using the combined weighting method.

[0090] A combined weighting method using the Analytic Hierarchy Process (AHP) and entropy weighting is adopted to balance subjective business experience with objective data patterns and avoid the subjective bias of single weighting.

[0091] (1) Calculation of objective weights using the entropy weight method

[0092] For the j-th feature, calculate its information entropy. :

[0093]

[0094] in, The proportion of the j-th feature in the i-th evaluation unit. The number of evaluation units.

[0095] Calculate the objective weight of the j-th feature :

[0096]

[0097] in, The number of features.

[0098] (2) Calculation of subjective weights using the AHP method

[0099] A judgment matrix is ​​constructed based on scores from planning experts. Calculate the largest eigenvalue The corresponding feature vectors, after normalization, yield the subjective weights. And it passed the consistency test (CR<0.1). The consistency test formula is:

[0100]

[0101] Where CI is the consistency index, RI is the average random consistency index, and CR is the consistency ratio.

[0102] (3) Calculation of combined weights

[0103]

[0104] in, As the weighting coefficient, this scheme takes... It takes into account both subjective business experience and objective data patterns.

[0105] 4. Redundant feature removal

[0106] Pearson correlation coefficient is used to remove highly redundant features with correlation coefficients greater than 0.9, thereby reducing feature dimensionality and improving fusion efficiency.

[0107] In some embodiments, the aforementioned similar fusion features include a first fusion feature, a second fusion feature, a third fusion feature, and a fourth fusion feature. Performing a first fusion on the aforementioned related data of the same type yields multiple similar fusion features. Specifically, this can be achieved through at least one of the following steps: First, a first training set is formed by combining historical related data of the aforementioned spatial geographic type and corresponding first labels; a first data fusion model is trained using the first training set to obtain a spatial geographic data fusion model; wherein the first label is a historical first fusion feature corresponding to the historical related data of the aforementioned spatial geographic type in the first training set; the related data of the aforementioned spatial geographic type is input into the spatial geographic data fusion model to obtain the first fusion feature corresponding to the related data of the aforementioned spatial geographic type; Second, a second training set is formed by combining historical related data of the aforementioned structured type and corresponding second labels; a second data fusion model is trained using the second training set to obtain a structured data fusion model; wherein the second label is a historical second fusion feature corresponding to the historical related data of the aforementioned structured type in the second training set; the structured data is then combined with the historical related data of the aforementioned structured type in the second training set to obtain a fourth fusion feature. The aforementioned related data of the above type are input into the aforementioned structured data fusion model to obtain the aforementioned second fusion feature corresponding to the aforementioned related data of the aforementioned structured type; the aforementioned unstructured historical related data and the corresponding third label are combined to form a third training set, and the aforementioned third training set is used to train the third data fusion model to obtain an unstructured data fusion model, wherein the aforementioned third label is the historical third fusion feature corresponding to the aforementioned unstructured historical related data of the aforementioned type in the aforementioned third training set, and the aforementioned unstructured related data are input into the aforementioned unstructured data fusion model to obtain the aforementioned third fusion feature corresponding to the aforementioned unstructured related data; the aforementioned time-series historical related data and the corresponding fourth label are combined to form a fourth training set, and the aforementioned fourth training set is used to train the fourth data fusion model to obtain a time-series data fusion model, wherein the aforementioned fourth label is the historical fourth fusion feature corresponding to the aforementioned time-series historical related data of the aforementioned time-series type in the aforementioned fourth training set, and the aforementioned time-series related data are input into the aforementioned time-series data fusion model to obtain the aforementioned fourth fusion feature corresponding to the aforementioned time-series related data.

[0108] This approach constructs a dedicated machine learning model for each data modality and performs supervised training using historical data and its corresponding labels, achieving deep feature extraction and fusion within each modality. This modality-specific modeling avoids feature interference and information loss that general models experience when processing heterogeneous data. The extracted first, second, third, and fourth fusion features accurately represent the core information of spatial, structural, semantic, and temporal dimensions, respectively, providing high-quality, high-fidelity feature inputs for subsequent cross-modal fusion and significantly improving the accuracy and robustness of the overall data fusion architecture.

[0109] Specifically, the S3 step involves two-level deep fusion of multi-source heterogeneous data. The core objective of this step is to solve the problems of feature loss and insufficient fusion accuracy caused by single-level fusion in existing technologies. It adopts a two-level architecture of intra-modal deep feature fusion and cross-modal global correlation fusion, and embeds a meta-learning few-sample optimization module to adapt to few-sample scenarios.

[0110] The first layer is intramodal deep feature fusion (modal adaptation fusion). For the data characteristics of the four core modalities, a dedicated feature extraction and fusion model is designed to fully preserve the core features of each modality and avoid feature loss of the general model.

[0111] The second layer: cross-modal global association fusion (an improved GNN model with spatiotemporal attention mechanism), which fuses the four types of modal features output from the first layer (spatial... , Structured ,text Time sequence Cross-modal global correlation fusion is performed to explore the potential correlation between features of different modalities and output a unified global fusion feature set.

[0112] In the specific implementation process, before training the first data fusion model using the first training set to obtain the spatial geographic data fusion model, the above method further includes the following steps: obtaining a first model, wherein the first model is a U-Net model; adding a residual module after the encoding layer of the first model, so that the output of the encoding layer is used as the input of the residual module; adding a spatial attention module after the decoding layer of the first model, so that the output of the decoding layer is used as the input of the spatial attention module, thereby obtaining the first data fusion model.

[0113] This approach addresses the vanishing gradient problem during deep U-Net model training by introducing a residual module, thereby improving the model's learning ability and convergence speed. Furthermore, the introduction of a spatial attention module enables the model to dynamically focus on spatial regions in remote sensing imagery crucial for resource assessment (such as building clusters, water bodies, and specific vegetation cover areas), suppressing background noise interference. These improvements result in the extracted first fusion features possessing stronger spatial semantic representation and noise resistance, significantly enhancing the accuracy of spatial geographic data fusion.

[0114] Specifically, the spatial grid mode is an improved U-Net model, which will be described in detail below.

[0115] (1) Improvement content

[0116] Two improvements were made to the standard U-Net:

[0117] Adding a residual module to the encoding layer: Adding a residual connection (skip connection) after each encoder convolutional block, which adds the input and output of the convolutional block, alleviates gradient vanishing in deep networks, and accelerates convergence.

[0118] A spatial attention module is added to the decoding layer: After upsampling at each decoder, a spatial attention module is inserted. This module generates a spatial weight map (with the same size as the feature map) through 1×1 convolution and sigmoid activation, and weights different spatial locations of the feature map to highlight core spatial features related to resource distribution (such as terrain undulation, water body boundaries, and transportation accessibility areas).

[0119] (2) The improved model hierarchy (from input to output) is shown in Table 3.

[0120] Table 3

[0121]

[0122] (3) Improved layer input and output

[0123] Residual module (located in the encoder)

[0124] Input: Feature map from the previous layer ,size (Height × Width × Number of Channels);

[0125] Output: Feature map after two convolutional layers + BN + ReLU ,size (Downsampled) and added to the downsampled version of the input (adjusting the channels via a 1×1 convolution);

[0126] Spatial attention module (located in the decoder)

[0127] Input: Feature map upsampled by the decoder and concatenated with skip connections ,size ;

[0128] Internal operation: 1×1 convolution will Dimensionality reduced to 1 channel, Sigmoid activation generates spatial weight map Original feature map Element-wise multiplication ;

[0129] Output: Weighted feature map, with the same size.

[0130] (4) Fusion process

[0131] The DEM elevation, remote sensing image bands, constraints, and other multi-source raster data are stitched together according to the channel dimension into a multi-channel input tensor (shape). Input to the improved U-Net: The encoder extracts multi-scale spatial features layer by layer, and residual connections preserve shallow details; the decoder upsamples to restore resolution and dynamically highlights important spatial locations through spatial attention. The final output is a single-channel (or multi-channel) spatial fusion feature map. ,size ( Each pixel value represents the spatial composite feature of that location.

[0132] In some embodiments, before training the second data fusion model using the second training set to obtain the structured data fusion model, the method further includes the following steps: obtaining a second model, wherein the second model is an SAE model; adding a normalization layer after the encoding layer of the second model, such that the output of the encoding layer is used as the input of the normalization layer; adding a regularization layer after the normalization layer of the second model, such that the output of the normalization layer is used as the input of the regularization layer, to obtain the second data fusion model.

[0133] This scheme addresses the common internal covariate shift problem during deep neural network training by introducing a normalization layer, accelerating model convergence and improving training stability. Furthermore, the introduction of a regularization layer effectively reduces the model's dependence on training data, preventing overfitting with limited samples. This improvement enables the extracted second fusion features to maintain the integrity of the data's structured information while possessing stronger generalization capabilities.

[0134] Specifically, the structured mode is a stacked autoencoder (SAE), which will be described in detail below.

[0135] (1) Improvement content

[0136] Two improvements were made to the standard stacked autoencoder:

[0137] A batch normalization layer is added after each autoencoder layer: to accelerate training convergence and prevent internal covariate shift.

[0138] Adding a Dropout layer after batch normalization (dropout rate 0.2) improves the model's generalization ability and avoids overfitting (especially suitable for small sample scenarios).

[0139] (2) The improved model hierarchy (including encoder + decoder during training, and only encoder during inference) is shown in Table 4.

[0140] Table 4

[0141]

[0142] (3) Improved layer input and output

[0143] Batch Normalization Layer

[0144] Input: Feature vector output by the encoder (N is the number of samples, D is the feature dimension);

[0145] Output: Normalized vector The shape remains unchanged;

[0146] This represents the batch average. Indicates batch variance. Represents the smoothing constant. This represents the scaling parameter. This represents the offset parameter.

[0147] Dropout layer

[0148] Input: Feature vector

[0149] Output: Randomly set 20% of neurons to 0, and scale the rest. The shape remains unchanged.

[0150] (4) Fusion process

[0151] Align the data from multiple structured tables by evaluation unit, and concatenate each unit into a high-dimensional vector. The improved SAE is input, and the dimensionality is reduced layer by layer through a three-layer encoder. After each layer, batch normalization and Dropout are applied to improve stability, finally yielding a 32-dimensional compressed feature vector. .

[0152] In the specific implementation process, before training the third data fusion model using the third training set to obtain the unstructured data fusion model, the above method also includes the following steps: obtaining the third model, wherein the third model is the BERT-wwm model; adding an attention module before the output layer of the third model, so that the output of the attention module is used as the input of the output layer, thereby obtaining the third data fusion model.

[0153] This solution addresses the issues of topic dispersion and noise interference with key information that may arise in the standard BERT model when processing long texts or complex semantics by introducing a topic attention module. BERT-wwm provides rich contextual semantic representations, while the topic attention mechanism guides the model to focus on strongly relevant semantic topics, filtering out irrelevant background information. This improvement makes the third fusion feature extracted by the model more semantically focused and business-relevant, thereby significantly improving the information extraction accuracy of unstructured text data in resource assessment.

[0154] Specifically, for text semantic modalities, we will improve the BERT-wwm model, which will be introduced in detail below.

[0155] (1) Improvement content

[0156] Based on the standard BERT-wwm model, a topic attention module is added after the output layer:

[0157] First, LDA (Latent Dirichlet Allocation) is used to extract from all text. There are 10 topics, each represented as... (Same as BERT output dimension).

[0158] For the feature vector of each token output by BERT (Sequence length L), calculate its similarity with each topic, and sum the weighted values ​​to obtain the topic-aware text representation.

[0159] (2) The improved model hierarchy is shown in Table 5.

[0160] Table 5

[0161]

[0162] (3) Improved layer input and output

[0163] Topic Attention Layer

[0164] Input 1: Token feature matrix output by the BERT encoder ;

[0165] Input 2: Predefined topic matrix (Fixed parameters, not used in training);

[0166] Internal calculation: Calculate the similarity between each token and each topic: For each token, take the topic weight with the highest similarity (or use softmax weighting), and sum them up: ,in ;

[0167] Output: Topic-aware vector ;

[0168] Output fully connected layer

[0169] enter: ;

[0170] Output: ;

[0171] (4) Fusion process

[0172] Multiple texts related to the region are collected. Each text is independently input into an improved BERT-WWM model: first, BERT-WWM extracts context-related token feature sequences; then, a topic attention layer calculates the association between the text and 20 topics, outputting a 768-dimensional topic-aware vector; finally, the dimensionality is reduced to 256 dimensions through a fully connected layer. The vectors of all texts are averaged according to evaluation units (e.g., by region) to obtain the final text fusion feature. .

[0173] In some embodiments, before training the fourth data fusion model using the fourth training set to obtain the time-series data fusion model, the method further includes the following steps: obtaining a fourth model, wherein the fourth model is a Transformer model; adding an encoding layer after the input layer of the fourth model, such that the output of the input layer is used as the input of the encoding layer; adding a multi-head self-attention module after the encoding layer of the fourth model, such that the output of the encoding layer is used as the input of the multi-head self-attention module, to obtain the fourth data fusion model.

[0174] This scheme addresses the issues of low dimensionality and insufficient semantic information in the original time-series data by introducing a linear encoding layer, thereby enhancing the model's ability to nonlinearly map time-series features. Furthermore, the introduction of a multi-head self-attention module, particularly combined with a time interval bias mechanism, enables the model to simultaneously capture long-term dependencies and short-term fluctuations in the time-series data and dynamically allocate importance weights at different time steps. This improvement allows the fourth fusion feature extracted by the model to more accurately reflect the periodic changes and trends in the time-series data, thus significantly improving the accuracy and predictive power of time-series data fusion.

[0175] Specifically, the timing mode is an improved Transformer encoder, which will be described in detail below.

[0176] (1) Improvement content

[0177] Two improvements were made to the standard Transformer encoder:

[0178] Learnable temporal positional encoding: using a trainable positional embedding matrix ( For time steps, Instead of fixed sine / cosine encoding, it enables the model to adaptively learn cyclical patterns such as seasonality and holidays.

[0179] Multi-head self-attention with time interval bias: Incorporating the time interval into the attention score calculation. Related bias terms The bias value decreases with the interval: ( , (days). The improved attention formula is:

[0180] ;

[0181] (2) Improved model hierarchical relationship

[0182] Overall structure: Input projection → Position encoding → N-layer improved Transformer encoder → Global average pooling → Output fully connected layer. The internal layers of a single encoder (a total of N_layer=3 layers) are shown in Table 6.

[0183] Table 6

[0184]

[0185] The post-processing layers after the encoder are shown in Table 7.

[0186] Table 7

[0187]

[0188] (3) Improved layer input and output

[0189] Learnable positional coding layer

[0190] Input: The matrix of the original time series data after linear projection ;

[0191] Trainable parameters: ;

[0192] Output: ;

[0193] Improved multi-head self-attention layer

[0194] enter: ;

[0195] Internal computation: generation , , Calculate attention score Add the bias matrix (shape (Only related to positional difference), multiplied by softmax. ;

[0196] Output: After splicing multiple heads, a linear transformation is performed back. .

[0197] (4) Fusion process

[0198] Aligning multiple time series data to a unified daily granularity to form a multivariate time series matrix. ,in (sky), The number of variables. First, a linear projection layer is used to... Mapped to Dimension; then stack learnable positional encodings; input to an improved Transformer encoder with 3 stacked layers (each layer using multi-head self-attention with time-interval bias); the encoder output remains the same. The sequence features are obtained; finally, a 128-dimensional vector is obtained through global average pooling, and then the dimensionality is reduced to 128 dimensions through the output fully connected layer (in fact, the dimension remains unchanged and can be adjusted as needed), thus obtaining the temporal fusion features. (N is the number of evaluation units. If each unit has an independent time series, then N>1; if it is a time series of the entire region, then N=1).

[0199] In the specific implementation process, a second fusion is performed on all the above-mentioned similar fusion features to obtain comprehensive fusion features. This can be achieved through the following steps: obtaining a fifth model, wherein the fifth model is a GNN model; adding a spatiotemporal attention module after the input layer of the fifth model, so that the output of the input layer is used as the input of the spatiotemporal attention module; adding a residual connection module before the normalization layer of the fifth model, so that the output of the residual connection module is used as the input of the normalization layer, to obtain a fifth data fusion model; combining multiple above-mentioned similar fusion features into a multi-type dataset; combining the multi-type dataset and the corresponding fifth label into a fifth training set; training the fifth data fusion model using the fifth training set to obtain a comprehensive data fusion model, wherein the fifth label is the historical fifth fusion feature corresponding to the historical multi-type dataset of the fifth training set; inputting the multi-type dataset into the comprehensive data fusion model to obtain the comprehensive fusion features corresponding to the multi-type dataset.

[0200] This scheme addresses the problem of traditional GNNs' inability to effectively distinguish the importance of different modalities and spatiotemporal dimensions when processing multi-source heterogeneous data by introducing a spatiotemporal attention module. It achieves dynamic weighted fusion of cross-modal features and uncovers potential correlations between different data sources. Furthermore, by introducing residual connections, it solves the information loss problem caused by excessive feature smoothing in deep graph neural networks, preserving the original feature details of nodes (modalities). This improvement results in more expressive and robust integrated features extracted by the model, comprehensively reflecting the interaction of spatial, structural, semantic, and temporal multidimensional information of the region to be evaluated, thereby significantly improving the overall accuracy of resource assessment.

[0201] Specifically, the following provides a detailed introduction to cross-modal global correlation fusion.

[0202] 1. Improved Graph Neural Network (GNN) Model

[0203] (1) Improvement content

[0204] Based on the standard Graph Neural Network (GNN, similar to Graph Convolutional Network GCN), two key improvements are made:

[0205] Improvement ①: Embed a spatiotemporal attention mechanism in each layer of the GNN;

[0206] Traditional GCNs use fixed weights (based on adjacency matrix degree normalization) to aggregate neighbor nodes, which cannot distinguish the importance of different neighbor nodes. This scheme introduces a spatiotemporal attention mechanism in the message-passing stage of GNNs, dynamically assigning attention weights to different neighbors of each node. At the same time, it combines spatiotemporal location encoding, enabling the model to capture both the spatial location association and the temporal series association between nodes.

[0207] Improvement ②: Add residual connections between GNN layers;

[0208] Adding residual connections between the outputs and inputs of adjacent GNN layers alleviates the oversmoothing problem of deep GNNs and preserves the original feature information of the nodes.

[0209] (2) Improved model hierarchy (from input to output)

[0210] This layer receives the features of the four modal nodes from the first layer, constructs a heterogeneous graph, and then performs message passing through multiple stacked improved GNN layers, finally outputting the fused node features. The overall hierarchy is shown in Table 8.

[0211] Table 8

[0212]

[0213] (3) Improved layer input and output

[0214] ① Spatiotemporal attention mechanism (embedded inside the improved GNN layer)

[0215] In each improved GNN layer, the computation of message passing is divided into the following sub-steps, as shown in Table 9.

[0216] Table 9

[0217]

[0218] Attention weight calculation formula (combined with spatiotemporal location encoding):

[0219] in:

[0220] , Node features after linear transformation;

[0221] Spatiotemporal location encoding between node i and node j (combining spatial distance features and temporal difference features);

[0222] Feature splicing operation;

[0223] Learnable attention vectors;

[0224] Spatiotemporal location coding Generation:

[0225] Spatial component: Calculate the geographical distance (such as Euclidean distance or spherical distance) between node i and node j, and normalize it as a one-dimensional feature.

[0226] Time component: If the node represents different time segments (such as time series modal nodes with explicit timestamps), calculate the time difference (number of days), and use it as a multidimensional feature after RBF encoding.

[0227] final It is a concatenated vector of spatial and temporal features.

[0228] ② Residual connectivity + layer normalization

[0229] enter: (Output from the previous layer) and (Current improved GNN layer output);

[0230] operate: ;

[0231] Output: The node feature matrix after residual connection, with the dimensions unchanged.

[0232] (4) Details of heterogeneous graph construction

[0233] Number of nodes Where K is the number of evaluation units. For simplicity, this scheme constructs a heterogeneous graph for each evaluation unit, containing 4 nodes (corresponding to spatial, structured, textual, and temporal features, respectively). In practice, all evaluation units can share the same set of nodes (i.e., 4 nodes), or each unit can have its own graph constructed (4 nodes × number of units). This scheme uses an independent graph construction for each evaluation unit, with 4 nodes and an adjacency matrix. Calculated based on the Pearson correlation coefficient between features.

[0234] Adjacency Matrix For the current evaluation unit, calculate the Pearson correlation coefficients between each pair of the four modal features. ,like but ,otherwise At the same time, a self-loop is added to each node ( ).

[0235] Initial node feature matrix :shape ,in The dimensions of each modal feature after unified projection (in this invention) Flatten the mesh diagram and align it with other modalities, projecting all modalities to the same dimension. ).

[0236] 2. Cross-modal global correlation fusion process

[0237] Step 1: Feature Alignment and Projection

[0238] Spatial integration features For grid diagram (size) The feature vector corresponding to each grid is extracted according to the evaluation unit (e.g., grid), forming... The matrix ( Alternatively, the channel vector of each grid cell can be directly taken. The features of each grid cell are projected onto a unified dimension through a fully connected layer. .

[0239] Structured fusion features Each evaluation unit corresponds to a 32-dimensional vector, which is projected onto 128 dimensions.

[0240] Text fusion features Each evaluation unit corresponds to a 256-dimensional vector, which is projected onto a 128-dimensional vector.

[0241] Temporal fusion features Each evaluation unit corresponds to a 128-dimensional vector (if the time series data is a whole region, then each unit is the same), which is projected onto 128 dimensions.

[0242] After projection, each evaluation unit receives one The node feature matrix.

[0243] Step 2: Spatiotemporal Location Encoding

[0244] Spatial Location: Each evaluation unit has center point coordinates (latitude and longitude or projected coordinates). The spatial distance between nodes is calculated (e.g., spatial modal nodes and structured modal nodes belong to the same evaluation unit, so the distance is 0; the distance between nodes in different evaluation units is the distance between the unit centers). Within each evaluation unit, the spatial distance between the four nodes is 0, but this needs to be considered when crossing units. For simplicity, this scheme constructs the graph independently within a single evaluation unit, and the spatial location encoding is only used as a node's own attribute (not an edge attribute). Specifically: A spatial coordinate vector (latitude and longitude normalized and concatenated) is appended to each node, and then mapped through a fully connected layer to... The dimension is added to the initial feature.

[0245] Time Location: Time-series modal nodes have timestamps (e.g., prediction date). The timestamps (year-month-day) are converted to Julian days, normalized, and used as one-dimensional features. These features are then mapped and added to the corresponding node features. Other modal nodes can have their timestamps set to the current planning baseline date.

[0246] Step 3: Construct a heterogeneous graph

[0247] For each evaluation unit, construct a graph with 4 nodes. :

[0248] Each node is characterized by a 128-dimensional vector with spatiotemporal location encoding superimposed in step 2.

[0249] Edge set Based on the Pearson correlation coefficient between features, the correlation coefficient between each pair of features of the four nodes is calculated. If the absolute value is greater than 0.3, an edge is connected. Considering that all nodes belong to the same evaluation unit, there are potential connections between them. This scheme adopts a fully connected graph (i.e., each node is connected to all other nodes) to ensure full information exchange.

[0250] Step 4: Stacking and improving GNN layers for message passing

[0251] Initial node feature matrix Input the first improved GNN layer. Execute the following within each layer:

[0252] Linear transformation: ,in (Keep the dimensions unchanged).

[0253] For each node i, calculate the attention weights of its neighbor j. (Using the spatiotemporal attention mechanism, since there is no spatial distance difference between nodes, the spatial location encoding only participates in the calculation as a feature of the node itself; the time difference may only be reflected when the sequential node is compared with other nodes.)

[0254] Weighted aggregation: .

[0255] activation: .

[0256] Residual connectivity + layer normalization: .

[0257] Stacking Layers (this scheme uses 2 layers). Through multi-layer message passing, the features of each node are integrated with the information of other modal nodes, realizing cross-modal global association.

[0258] Step 5: Output global fusion features

[0259] The node feature matrix output from the last layer Flatten (or piece together) rows to get a A dimensional vector serves as the global fusion feature of this evaluation unit. Repeat the above process for all evaluation units to obtain the final cross-modal global fusion feature set. .

[0260] 3. Meta-learning few-shot optimization (independent of GNN layers, but applied to the entire cross-modal fusion model)

[0261] To address the issue of insufficient data samples, a Model-Independent Meta-Learning (MAML) algorithm is embedded during the training process of a cross-modal fusion GNN model. The specific steps are as follows:

[0262] Sample multiple tasks (each task containing a support set and a query set) from multiple data-rich regions (source domains) to train the model's initialization parameters. This allows the model to achieve good performance after fine-tuning with only a small number of labeled samples (e.g., 5 samples) in the target region (target domain).

[0263] The meta-learning module does not affect the hierarchical structure of the GNN; it only adds an outer loop during the training phase. The finely tuned model parameters are used during the inference phase.

[0264] In addition, the core objective of the full-process compliance verification and correction of the S4 step fusion results is to address the high compliance risk of existing technology fusion results and ensure that the fusion results meet the requirements and can be directly implemented.

[0265] 1. Construct a compliance verification rule base and incorporate constraint rules, the core of which include: 1) Prohibited development rule: prohibit development and construction; 2) Restricted development rule: restrict development and only allow the construction of low-impact supporting facilities; 3) Threshold control rule: must not exceed the regional threshold and the construction of supporting facilities must not exceed the boundary.

[0266] To implement the aforementioned threshold control rules, it is necessary to quantitatively calculate the development intensity and supporting infrastructure construction scale of each evaluation unit from the fusion results. The specific method is as follows:

[0267] (1) Determination of development intensity

[0268] Development intensity is a comprehensive indicator that reflects the degree of pressure to be implemented within a specific evaluation unit. This solution integrates a global feature set. The following sub-features are extracted and weighted to obtain the development intensity index. :

[0269] Construction demand characteristics: from the perspective of spatial integration Sub-features related to the suitability of construction land are extracted, and the probability of the unit being developed into construction land is predicted by a logistic regression model. .

[0270] Pressure bearing characteristics: from time-series fusion characteristics Extract the predicted daily peak tourist density (people / hectare) and total annual number of tourists .

[0271] Facility density characteristics: from the perspective of structured integration characteristics Extract the density of existing homestays, restaurants, parking lots, and other facilities. (units / km²) and the number of planned new facilities .

[0272] The formula for calculating the comprehensive development intensity index is:

[0273]

[0274] in The regional baseline density (e.g., 50 people per hectare). Weights are assigned to the maximum permitted facility density in the plan (e.g., 20 facilities / km²). Determined by AHP method (e.g., 0.4, 0.35, 0.25). Final Normalization to The higher the value, the greater the development intensity.

[0275] (2) Determination of regional carrying capacity threshold

[0276] Bearing threshold Based on pre-calculations using environmental data, the following method is adopted for each evaluation unit:

[0277] Ecological factors such as vegetation cover (NDVI), water conservation capacity, and soil erosion modulus of this unit were collected.

[0278] Calculate the sensitivity score using the analytic hierarchy process (AHP) or principal component analysis (PCA). (Range 0~1, the higher the value, the more sensitive it is).

[0279] Based on the sensitivity level, set the load threshold:

[0280]

[0281] This threshold represents the maximum allowable development intensity index for this unit. .

[0282] During compliance verification, if The result was determined to be a "violation exceeding the threshold" and correction is required.

[0283] (3) Determination of supporting facilities construction

[0284] Supporting facilities include service facilities (guesthouses, restaurants, visitor centers, parking lots, restrooms, etc.) and infrastructure (roads, water and electricity networks, communication base stations, etc.). This solution extracts the following information from the integration results:

[0285] Types and scale of proposed supporting facilities: Obtain a list of facilities planned for construction in each evaluation unit from the output of the business layout stage. For example, the number of guesthouse rooms. Number of parking spaces Number of dining seats Toilet squatting stalls wait.

[0286] Spatial location of supporting facilities: from the perspective of spatial integration characteristics Extract the planning coordinates (or the boundary of the evaluation unit) of each proposed facility.

[0287] Supporting facilities land area: Based on the type and scale of the facilities, the required construction land area is converted using standard quotas. (For example, each parking space is 25m², and each bed in the guesthouse is 40m²).

[0288] 2. Automatic compliance verification and correction process

[0289] (1) Spatial overlay analysis and compliance attribute matching

[0290] Step 1: Prepare the spatial data layer

[0291] Global fusion feature set: Each evaluation unit (e.g., a 30m × 30m grid unit) has a unique spatial identifier (surface or point) and carries... Eigenvectors.

[0292] Compliant layer set: Each area is a vector surface layer.

[0293] Step 2: Perform spatial overlay operation

[0294] For vector evaluation units: Use the "Identifier Overlay" or "Spatial Connection" function of the GIS spatial analysis tool to perform intersection calculations between each evaluation unit and all compliant layers.

[0295] For raster evaluation units: convert the compliant layer into a raster of the same resolution and read the attribute values ​​pixel by pixel; or sample the attributes of the compliant layer based on the coordinates of the pixel center point.

[0296] Step 3: Calculate the compliance attribute fields for each evaluation unit. For each evaluation unit, calculate and record the following fields, as shown in Table 10.

[0297] Table 10

[0298]

[0299] Execution logic: For each evaluation unit, perform spatial overlay analysis with each compliance layer in sequence, and store the above field values ​​in the attribute table of that unit.

[0300] (2) Automatic unit-by-unit verification based on rule base

[0301] For each evaluation unit, the following judgments are made in sequence. Once a certain item is matched, the category is determined and no further judgments are made.

[0302] ① Absolutely prohibit verification

[0303] Judgment conditions: in_eco_redline = true or in_permanent_basic_farm = true or in_water_source_protect = true or in_heritage_core = true.

[0304] Execution logic: If any condition is met, the unit is directly determined to be "prohibited from development", the reason is recorded as "located in a prohibited development area", and the verification process ends.

[0305] ② Buffer limit development verification

[0306] Conditions for judgment: in_heritage_control = true and not in the core area.

[0307] Execution logic: Determine if it is "restricted development" because it is "located in a constraint control zone", record the constraints that need to be followed, and end the process.

[0308] ③ Constraints on neighboring areas for development verification

[0309] Judgment conditions: in_eco_redline = false and distance_to_eco ≤ 50 meters (the threshold can be configured according to regulations).

[0310] Execution logic: Determine if it is "restricted development" due to "adjacency constraint", record the buffer distance, and end the process.

[0311] ④ Constraint boundary verification

[0312] Judgment condition: in_urban_development_boundary = false.

[0313] Execution logic: Extract the list of all proposed construction types for this unit, and check whether each business type belongs to the "Low Impact Facility List". If all proposed business types belong to the list, it is determined to be "Restricted Development" because "it is located outside the constraint boundary and only low impact facilities are allowed". If there is any business type that is not on the list, it is determined to be "Prohibited Development" because "it is located outside the constraint development boundary and contains non-low impact facilities".

[0314] ⑤ Bearing threshold verification

[0315] Operation: Calculate the development intensity index of this unit. And query its carrying threshold. .

[0316] Judgment condition: If If so, it is determined as "restricted development" because "development intensity exceeds the carrying capacity threshold," and the current situation is recorded. With maximum allowable value ;like If so, it is judged as "compliant".

[0317] Execution logic: The evaluation units are executed one by one in the above order. Each unit is eventually marked as one of three states: "development prohibited", "development restricted" or "compliant", with an explanation of the reason and necessary parameters.

[0318] (3) Handling of prohibited development units

[0319] Remove all cells marked "development prohibited" from the final fused dataset used for planning decisions.

[0320] (4) Restricting the modification of development units

[0321] For units marked "Restricted Development", the classification should be corrected according to the reasons for the restriction.

[0322] ① Restrictions caused by "constraint buffer" or "constraint proximity"

[0323] Revised objective: Prohibit high-impact facilities, retain only low-impact facilities, and reduce development intensity below a lower threshold.

[0324] Execution steps: Iterate through all proposed business types for this unit, comparing them with the "High-Impact Business Type List" and the "Low-Impact Business Type List." If a proposed business type belongs to the High-Impact Business Type List, delete it directly; if it belongs to the Low-Impact Business Type List, retain it; if it cannot be clearly categorized, delete it by default (safety first). After deleting business types from the list, recalculate the development intensity index. For constrained buffers, a stricter load threshold is used. For constrained neighbor regions, the original method is still used. ,like If the number of businesses still exceeds the threshold, the size of the remaining businesses will be reduced proportionally until the threshold is met. The filtered list of businesses and their reduced sizes will then be saved, and the unit's data will be updated. The corresponding feature fields.

[0325] ② Restrictions caused by "extra-boundary development boundaries"

[0326] Revised objective: Ensure that all proposed facilities are located within the boundaries and that their overall size is kept under control.

[0327] Execution steps: For each low-impact facility planned within the unit, check its specific spatial coordinates. If the facility is located outside the constrained development boundary, try to find a buildable area within the unit. If a buildable area exists, adjust the facility location to that area. If there is no buildable area within the unit, transfer the facility to an adjacent compliant unit, save the adjusted facility location and size, and update relevant features.

[0328] ③ Restrictions due to "development intensity exceeding the carrying capacity threshold"

[0329] Revised objective: Reduce development intensity to below a threshold through multi-level cuts.

[0330] Execution steps: Initialization, obtain current development intensity and carrying threshold .like No corrections needed, proceed directly; Level 1 Reduction – Reduce the number of proposed new facilities and calculate the excess percentage. The number of facilities to be added Reduce proportionally: (Round down) Recalculate development intensity If ≤ If the first level of reduction fails, save the results and end the correction process; the second level of reduction involves lowering the density prediction value. If the density still exceeds the limit after the first level of reduction, the predicted peak daily density of tourists for that unit is reduced by adjusting the capacity coefficient in the passenger flow prediction model. The value gradually decreases from the current value, decreasing by 10% each time, and is then recalculated. This process continues until the threshold is met or the value drops to 50% of its original value. If dropping to 50% of the original value still does not meet the threshold, the current value is recorded and the process moves to the next level. The third level of reduction involves lowering the demand probability. If the demand still exceeds the limit, the suitability-related parameters in the spatial fusion characteristics of that unit are modified, and the demand probability is recalculated. Then recalculate If the threshold is met, the modified features are saved; finally, the feature is downgraded. If the feature has undergone the above three levels of reduction, Still greater than If so, the unit is downgraded from "restricted development" to "prohibited development", marked as an undevelopable area, and the corrected feature set is output. (etc.), and the updated development intensity index (should be ≤) ).

[0331] (5) Output the final fused dataset after compliance correction.

[0332] After all evaluation units have completed the above verification and correction, the final fusion dataset is generated, which includes the status field (compliant / restricted development / prohibited development) of each unit and the corresponding reasons and correction logs.

[0333] For compliant units: retain the original fusion features.

[0334] For restricted development units: preserve the revised integration characteristics (business type, facility scale, development intensity, etc.).

[0335] For prohibited development units: the feature set is retained but marked as unavailable, and the subsequent business adaptation process is skipped directly.

[0336] Specifically, regarding the business adaptation output of the S5 step fusion results, the core objective of this step is to address the issue of poor usability of existing technology fusion results. It aims to output fusion results in a targeted manner for different stages of the planning process, achieving one-time fusion and full-process reuse.

[0337] For the five core planning stages, the corresponding integrated results will be output:

[0338] 1. Resource survey stage

[0339] Input: The global fused feature set output by S3 S1: Raw resource data collected; S2: Resource-related features in the basic feature layer.

[0340] Output: Resource category tag, resource level rating, potential level.

[0341] Step 1: Resource Category Determination. This involves determining the global fusion feature set output from S3. Extract resource-related feature subsets, combine them with the original resource data collected by S1, and determine the resource category of each resource point or evaluation unit by feature similarity.

[0342] Step 2: Resource Level Evaluation. From the basic feature layer constructed by S2, indicators such as resource scarcity, preservation integrity, and accessibility are selected. The combined weighting method determined by S2 is used to calculate the comprehensive value score of each resource, and the level is determined based on the predefined scoring thresholds of the business.

[0343] Step 3: Development Potential Assessment. From the global fusion feature set of S3... The system extracts features such as distance from the main road, density of existing infrastructure, surrounding passenger flow, and suitability index, and outputs three levels of development potential: high, medium, and low.

[0344] 2. Suitability assessment stage

[0345] Input: The global fused feature set output by S3 ; Compliance status fields output by S4 (prohibited development, restricted development, compliant); Feature weights determined by S2.

[0346] Output: Overall suitability score and level (highly suitable, moderately suitable, poorly suitable, unsuitable), and reasons for additional restrictions on development units.

[0347] Step 1: Filter evaluable units. Read the compliance status field output by S4, which includes three categories: prohibited development, restricted development, and compliant. Units marked as prohibited development are directly classified into the unsuitable level, while units marked as restricted development are reserved for scoring.

[0348] Step 2: Construct an evaluation index system. This involves using the global fusion feature set output from S3. Five dimensions of features were extracted: resource endowment, carrying capacity surplus, transportation accessibility, infrastructure support, and potential, and then normalized respectively.

[0349] Step 3: Calculate the overall suitability score. The overall suitability score for each unit is obtained by combining the weights determined by the analytic hierarchy process (AHP) and the entropy weight method, using a weighted summation.

[0350] Step 4: Level Classification. Based on the preset score threshold, the units are divided into four levels: highly suitable, moderately suitable, poorly suitable, and unsuitable. The reason for the restriction is recorded in S4 after the level of the restricted unit.

[0351] 3. Business Layout Stage

[0352] Input: The global fused feature set output by S3 Time-series fusion features Business-related features in the S2 business feature layer; preprocessed spatial geographic data (roads, water and electricity nodes) in S1; compliance status and carrying capacity threshold output in S4. .

[0353] Output: Adaptability score for each business type, list of candidate units, recommended site combinations, and suggested scale.

[0354] Step 1: Define the business format preference model. For each preset business format, extract the exclusive business features related to each business format from the business feature layer of S2.

[0355] Step 2: Calculate business type suitability. This is done from the global fusion feature set of S3. Features corresponding to preferences are extracted, and the feature weights determined by S2 are used to calculate the fit score of each unit to each business format.

[0356] Step 3: Passenger flow capacity verification. This involves verifying the time-series fusion features output from S3. Extract the predicted daily peak tourist density and combine it with the carrying capacity threshold determined by S4. Determine whether the load exceeds the limit.

[0357] Step 4: Accessibility Analysis. Extract the locations of roads and water / electricity nodes from the preprocessed S1 spatial geographic data, and calculate the distance from the unit to the infrastructure.

[0358] Step 5: Spatial Combination Optimization. An optimization algorithm is used in the candidate units to satisfy business spacing constraints and total quantity control, generating recommended site combinations.

[0359] 4. Passenger flow forecasting stage

[0360] Input: Temporal fusion features output by S3 (Historical passenger flow sequence); spatiotemporal behavioral data collected by S1 (anonymized passenger flow data, OTA booking data, holiday markers); new facility capacity output from the business layout stage.

[0361] Output: Daily passenger flow total for each scenario, unit-level tourist density distribution, and composition of tourist origin areas.

[0362] Step 1: Scenario Definition. Set up three typical scenarios: off-season, peak season, and holidays. External influencing factors are referenced from the patterns in the holiday passenger flow time series data collected by S1.

[0363] Step 2: Total Prediction. Using the temporal fusion features output by S3. The historical passenger flow sequence is input into the prediction head after the improvement of the Transformer encoder, and the total daily passenger flow for each scenario is output, and the capacity is adjusted according to the newly added facilities in the business layout.

[0364] Step 3: Spatial Allocation. Based on the spatial fusion features output by S3. Based on the spatial attractiveness characteristics, a gravity model is used to allocate the total amount to each evaluation unit, and the predicted tourist density of each unit is obtained.

[0365] Step 4: Origin Analysis. Using anonymized passenger flow data and OTA booking data from the spatiotemporal behavioral data collected by S1, identify the main source regions and adjust the proportions according to scenarios.

[0366] 5. Benefit Evaluation Stage

[0367] Input: The global fused feature set output by S3 Structured fusion features Spatial integration characteristics Passenger flow forecast output; industry data collected in S1; comprehensive benefit features and weights in the S2 decision feature layer; carrying capacity threshold output in S4. .

[0368] Output: Benefit level, overall benefit score, and recommended results for scheme comparison.

[0369] Step 1: Benefit Assessment. Based on the passenger flow forecast results and the industry data collected in S1, the level is assessed according to the comprehensive benefit characteristics in the S2 decision-making characteristic layer.

[0370] Step Two: Benefit Evaluation. Structured fusion features output from S3. Indicators are extracted to predict the revenue growth rate, and the rating is determined by combining relevant features in the S2 business feature layer.

[0371] Step 3: Benefit Evaluation. Spatial fusion features output from S3. Environmental baseline data such as vegetation cover NDVI and hydrological and water quality were extracted and combined with the carrying capacity threshold of S4. Determine whether the benefits are acceptable.

[0372] Step 4: Comprehensive Benefit Index. Using the features defined in the S2 decision feature layer, calculate the comprehensive benefit score according to the combined weights determined in S2, and output the comprehensive benefit contribution map of each evaluation unit.

[0373] Step 5: Solution Comparison and Optimization. Repeat the above evaluation for multiple alternative solutions, and select the solution with the highest overall benefit score and acceptable benefits as the recommended solution. If the ecosystem is unacceptable, trigger the S4 correction process to readjust the development intensity.

[0374] The core objective of iteratively optimizing the fusion accuracy of step S6 is to address the issues of static and untimely fusion results in existing technologies, establish a dynamic iterative optimization mechanism, and continuously improve fusion accuracy and business adaptability.

[0375] Input: The global fused feature set output by S3 and modal fusion features ( , , , ); the fused dataset after compliance correction from S4 output; and the actual operational data collected after the planning and implementation.

[0376] Output: Optimized feature weights, updated fusion model parameters, new round of fusion dataset and planning support results.

[0377] 1. Construct a dual-dimensional fusion accuracy evaluation index system

[0378] Step 1: Define technical indicators. Use Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Goodness of Fit. The three metrics measure the deviation between the predicted values ​​(such as suitability score and passenger flow forecast) output by the fusion model and the actual observed values. The smaller the MAE and RMSE, the better. The closer to 1, the higher the precision.

[0379] Step Two: Defining Business Metrics. Three metrics are used to measure the actual application effectiveness of the integration results: planning implementation matching degree, passenger flow forecast accuracy, and compliance pass rate (the proportion of evaluation units that pass compliance review on the first attempt).

[0380] 2. Actual data collection and indicator calculation

[0381] Step 1: Collect actual operational data. Obtain actual data every six months; also obtain change data.

[0382] Step Two: Calculate Technical Indicators. Align the predicted values ​​output by the fusion model with the actual values ​​at the same evaluation unit and time point, calculate the absolute error and squared error for each sample to obtain MAE and RMSE; calculate the ratio of the total sum of squares to the residual sum of squares to obtain... .

[0383] Step 3: Calculate business metrics. Calculate the implementation matching degree by comparing the actual number of implemented projects with the number of planned recommended projects; calculate the closeness between the predicted total passenger flow and the actual total, and the correlation between the predicted spatial distribution and the actual distribution to obtain the passenger flow prediction accuracy; calculate the proportion of units that pass approval on the first attempt to obtain the compliance approval rate.

[0384] 3. Assessment and Diagnosis

[0385] Step 1: Indicator Comparison. Compare the current indicators with the historical baseline (indicators at the time of the initial model deployment) and preset thresholds to determine whether each indicator has deteriorated.

[0386] Step Two: Attribution Analysis. If technical indicators deteriorate, further analyze whether it's due to data drift (e.g., changes in customer flow patterns) or model obsolescence; if business indicators deteriorate, analyze whether it's due to unreasonable feature weights or changes in business rules (e.g., compliance constraints). Compare the differences in indicators across different regions and business types to pinpoint the problem areas.

[0387] Step 3: Trigger optimization decision. If any metric exceeds the threshold, the fine-tuning optimization process is initiated; if all metrics are within the threshold, the current model is maintained, and the dataset is updated only for the next evaluation.

[0388] 4. Feature weight fine-tuning

[0389] Step 1: Collect feedback data. Extract positive and negative samples from actual projects, label their key features, and form a new dataset.

[0390] Step 2: Recalculate the objective weights. Using the newly collected data, recalculate the objective weights of each feature according to the entropy weight method in S2. .

[0391] Step 3: Adjust subjective weights. If the planning or expert experience changes, update the AHP judgment matrix and recalculate the subjective weights. Otherwise, retain the original subjective weighting.

[0392] Step 4: Mixed weight update. Use the same combination coefficients as S2. Calculate the new weights .

[0393] Step 5: Weight Smoothing. To maintain stability, an exponential moving average is used to merge the old and new weights: ,in .

[0394] Step Six: Verify Results. Calculate the landing matching rate using the validation set from the previous quarter. If it improves, adopt the new weights; otherwise, revert to the old weights.

[0395] 5. Fine-tuning of fusion model parameters

[0396] Treatment is categorized according to the degree of degradation of technical indicators:

[0397] Step 1: Mild Degradation. When MAE increases by 5% to 10%, perform partial layer freeze fine-tuning. Freeze the first 60% of the model's layers (e.g., the first three layers of the U-Net encoder, the first 6 layers of BERT), and train only the last 40% of the layers and the output layer. Set the learning rate to one-tenth of the initial training learning rate. Use the most recent 6 months of data as the training set, and stop training when the loss on the validation set no longer decreases.

[0398] Step 2: Moderate Degradation. When MAE increases by 10% to 20%, perform full model fine-tuning. Unfreeze all layers and update all parameters using a lower learning rate. Increase the Dropout rate or add L2 regularization to prevent overfitting. The training data includes both old and new data (70% old data, 30% new data), and training continues until the loss on the validation set converges.

[0399] Step 3: Severe Degradation Handling. When the MAE increases by more than 20% or the model structure becomes unsuitable, incremental retraining is triggered. The original model parameters are retained as initialization, and the model structure is appropriately adjusted (e.g., adding a GNN layer or adjusting the number of attention heads). The model is then retrained using all data from the past two years (starting with the old model parameters). After training, the new and old models are compared on the validation set. If the new model's MAE decreases by more than 10%, it is replaced.

[0400] 6. Fine-tuning of the compliance rule base

[0401] Step 1: Collect rejection cases. Obtain the rejected evaluation units and the reasons for rejection.

[0402] Step 2: Update the rule base. If the rejection is due to outdated spatial data boundaries, obtain the latest planning data from public channels and replace the original compliance layer; if the rejection is due to unreasonable thresholds, adjust the distance thresholds according to the approval comments; if the rejection is due to disputes over the determination of business type, add or delete business types in the list of low-impact facilities.

[0403] Step 3: Rule Version Management. Record the date, basis, and execution method of each rule change to create a traceable rule base version.

[0404] Step 4: Re-verify. Re-verify the compliance of the previous quarter's fusion results using the updated rule base and calculate the new pass rate. If the pass rate increases to the preset target (e.g., ≥95%), the new rule base is officially implemented; otherwise, continue adjustments.

[0405] 7. Output the optimized fused dataset

[0406] Step 1: Full Process Recalculation. Using optimized feature weights, fusion model parameters, and compliance rule base, re-execute the entire S1 to S5 process on the latest collected raw data to generate a new round of fused dataset.

[0407] Step Two: Comparative Analysis. Compare the new dataset with the previous version and generate an optimization report, including the improvement in each indicator and a summary of the changes.

[0408] Step 3: Deploy and Update. Deploy the updated model and rule base to the production environment and push the new planning support results to the planning department.

[0409] 8. Iteration cycle management

[0410] Step 1: Regular Cycle. The above assessment and fine-tuning process is performed every six months.

[0411] Step 2: Trigger the cycle. If a change occurs, immediately initiate a temporary iterative optimization.

[0412] In some embodiments, the first fusion of the aforementioned related data of the same type is performed to obtain multiple similar fusion features. This can be achieved through the following steps: preprocessing the aforementioned related data of different types to obtain preprocessed related data, wherein the preprocessing methods include one or more of the following: unified coordinates, unified spatial resolution, geometric correction, missing value imputation, outlier removal, normalization, text cleaning, word segmentation and entity recognition, feature vectorization, unified time, and smoothing; and performing the aforementioned first fusion on the aforementioned preprocessed related data of the same type to obtain multiple similar fusion features.

[0413] This scheme effectively addresses technical challenges associated with multi-source heterogeneous data, such as inconsistent spatiotemporal benchmarks, high noise levels, numerous missing values, inconsistent dimensions, and the difficulty in directly utilizing unstructured data, by implementing differentiated preprocessing procedures tailored to different data modalities. Unified coordinates and resolution resolve spatial data misalignment issues; missing value imputation and outlier removal improve the integrity of structured data; text cleaning and vectorization enable the structured transformation of unstructured data; and time unification and smoothing enhance the stability of time-series data. These preprocessing measures provide high-quality, standardized data input for the subsequent first-stage fusion.

[0414] Specifically, the multi-source heterogeneous data acquisition and standardized preprocessing in step S1 includes modal standardized preprocessing. For the characteristics of different modal data, a dedicated preprocessing method is adopted. The core is to unify the spatiotemporal reference, eliminate noise, and unify the feature dimensions.

[0415] (1) Spatial geographic data preprocessing includes:

[0416] Coordinate unification: All spatial data are uniformly converted to the CGCS2000 national geodetic coordinate system and uniformly projected to the Gauss-Kruger projection to solve the spatial misalignment problem;

[0417] Spatial resolution unification: All raster data are resampled to a uniform resolution of 30m×30m using nearest neighbor interpolation. The core formula is:

[0418]

[0419] Detailed explanation of parameter meanings:

[0420] Output the pixel value of the resampled raster at pixel coordinates (x, y) (x and y are non-negative integers, representing the column number and row number, respectively);

[0421] : Input the pixel value of the original raster at pixel coordinates (i,j);

[0422] Spatial resolution of the raw raster data, in meters per pixel;

[0423] The goal is to achieve a unified spatial resolution; this solution sets it at 30 meters per pixel.

[0424] The floor function ensures that the mapped coordinates fall at integer pixel positions.

[0425] : The output pixel (x,y) is mapped to the column coordinates (floating-point numbers) of the pixel in the input raster;

[0426] : The output pixel (x,y) is mapped to the pixel row coordinates (floating-point numbers) in the input raster.

[0427] The formula means that for each pixel (x, y) in the output raster, first calculate its corresponding position in the input raster (floating-point number) according to the resolution ratio, then take the nearest integer pixel coordinates (nearest neighbor interpolation), and finally assign the pixel value of the input pixel to the output pixel.

[0428] Geometric correction: Based on the 1:10000 standard topographic map, the remote sensing image is geometrically corrected, and the root mean square error (RMSE) is controlled within 0.5 pixels.

[0429] (2) Structured data preprocessing includes:

[0430] Missing value imputation: K-nearest neighbor (KNN) imputation method is used, with the core formula as follows:

[0431]

[0432] in, For missing values, The nearest neighbor number (take 5). For the first The feature values ​​of the nearest neighbors.

[0433] Outlier removal: The 3σ criterion is used to remove outliers. outliers in the range The characteristic mean, The characteristic standard deviation is denoted as .

[0434] Normalization: Min-Max normalization is used to map eigenvalues ​​to the [0,1] interval, eliminating the influence of dimensions. Core formula:

[0435]

[0436] in, These are the normalized eigenvalues. These are the original eigenvalues. , These are the maximum and minimum values ​​of this feature.

[0437] (3) Preprocessing of unstructured text data includes:

[0438] Text cleaning: Remove special symbols, stop words, and meaningless characters;

[0439] Word segmentation and entity recognition: Jieba word segmentation is used, and named entity recognition (NER) is performed based on the BERT pre-trained model to extract core entities such as keywords;

[0440] Feature vectorization: The BERT-wwm pre-trained model is used to convert the text into a 768-dimensional semantic feature vector, realizing the structural transformation of text features.

[0441] (4) Time series data preprocessing includes:

[0442] Unified time granularity: All time series data are unified to daily granularity, and the time series are aligned using the sliding window summation / mean method;

[0443] Smoothing: The moving average method is used to eliminate random noise. The core formula is:

[0444]

[0445] in, Let be the smoothed value at time t. The size of the sliding window (take 7). This represents the original time series value at time i.

[0446] In summary, this paper addresses the core pain points of existing resource assessment fields, such as difficulty in integrating multi-source heterogeneous data, disconnect from business operations, insufficient compliance, and poor adaptability to small sample scenarios. It provides a multi-source data fusion method that is adaptable to the entire business process, solving the problems of low fusion accuracy, poor practicality, high compliance risks, and narrow applicability of existing technologies.

[0447] This solution adopts a closed-loop architecture that integrates preprocessing, feature hierarchical screening, two-level multimodal deep fusion, compliance verification, business adaptation output, and iterative optimization. Its core innovative logic is as follows:

[0448] 1. For resource assessment and planning, collect six major categories of multi-source heterogeneous data, including spatial geographic data, resource data, industry data, spatiotemporal behavior data, ecological environment data, and news data, and complete standardized preprocessing to unify spatiotemporal benchmarks and feature dimensions;

[0449] 2. Construct a three-level feature system: basic feature layer, business feature layer, and decision feature layer. Complete feature selection and weight assignment through a combination weighting method, and embed a rigid compliance constraint veto mechanism.

[0450] 3. Adopting a two-level fusion core architecture: First, design dedicated models for different modal data to complete the fusion of deep features within the modality. Then, complete the cross-modal global association fusion through an improved graph neural network (GNN) with a spatiotemporal attention mechanism. At the same time, embed a meta-learning module to solve the problem of overfitting with small samples.

[0451] 4. Conduct full-process compliance verification and correction of the fusion results to ensure they meet the requirements;

[0452] 5. Targeting the five core links of resource assessment, suitability evaluation, business layout, customer flow forecasting, and benefit evaluation, provide integrated results that are tailored to business needs;

[0453] 6. Construct a precision evaluation system that combines technology and business dimensions to achieve dynamic iterative optimization of the fusion model.

[0454] In summary, the proposed solution significantly improves fusion accuracy, fully preserves the core features of multimodal heterogeneous data, and supports refined planning decisions. This is because the solution employs a two-tiered fusion architecture: intramodal adaptive fusion and cross-modal spatiotemporal attention GNN fusion. It designs dedicated feature extraction and fusion models for the characteristics of different modalities, and dynamically learns the association weights of cross-modal features through a spatiotemporal attention mechanism. For the four core modalities, it fully extracts and preserves the spatial, semantic, and temporal core features of each modality by improving U-Net, SAE, BERT, and Transformer encoders, avoiding feature loss caused by single-layer fusion in existing technologies. By improving the GNN model, it mines the potential correlations between features of different modalities, achieving deep feature-level fusion rather than the shallow numerical-level fusion of existing technologies. This solves the problems of single fusion architecture, low fusion accuracy, and loss of core features in existing general-purpose solutions, ultimately supporting refined full-process planning decisions, rather than just simple resource scoring.

[0455] The proposed solution can be deeply integrated with the entire planning process, resulting in highly practical and business-adaptable fusion results. This is because the solution constructs a three-level feature system for the five core planning stages, embedding business rules into the entire process of feature selection, fusion, and output, rather than designing for only the resource evaluation stage. The feature system covers the business needs of the entire planning process, and the fusion process fully considers the core business characteristics of each stage. The output stage provides targeted fusion results for different planning stages, achieving one-time fusion for reuse throughout the entire process. This solves the problems of existing general technical solutions not being integrated with the entire planning process, fusion results only serving a single stage, and poor practicality. Planning agencies no longer need to repeatedly process and fuse data for different stages; the fusion results can directly support all core stages of the entire planning process, demonstrating extremely high business adaptability and practicality.

[0456] This application's solution can embed constraints throughout the entire process, and the fusion result has no compliance risks and can be directly used for planning. Because this solution treats compliance constraints as a veto item, embedding them into the entire process of feature selection, fusion, and verification, it has built a dedicated compliance verification rule base and automatic correction mechanism, rather than simply treating compliance constraints as ordinary evaluation indicators. During the feature selection stage, rigid constraints prohibiting or restricting development are marked as veto features and incorporated into the fusion logic from the source. After fusion, through spatial overlay analysis and automatic verification using the rule base, the fusion result is corrected for compliance, eliminating prohibited development areas and limiting the development intensity of development areas. This solves the problems of existing general technical solutions not embedding compliance constraints, fusion results conflicting with planning, high compliance risks, and inability to be implemented. The fusion result of this solution complies with constraints, has no compliance risks, and can be directly used in the planning process, making it highly implementable.

[0457] The proposed solution can embed a meta-learning few-shot optimization module, adapting to small-sample scenarios and having a wide range of applications. Because this solution embeds the MAML meta-learning few-shot optimization module into the cross-modal fusion model, it allows the model to learn universal initialization parameters through pre-training on multi-source region data. On small-sample datasets, only minor updates are needed for rapid adaptation, solving the problems of subjective weight assignment and model overfitting caused by insufficient data samples and incomplete labels. The model can achieve high-precision fusion without requiring a large amount of sample data from the target region, adapting to real-world situations. It also addresses the problems of existing general-purpose solutions not being adapted to small-sample scenarios, only usable in regions with complete data, and having a narrow scope of application. The proposed method maintains fusion accuracy even in regions with insufficient sample size, has a wide applicable range, and exhibits extremely strong generalization ability.

[0458] The proposed solution establishes a dynamic iterative optimization mechanism, ensuring the timeliness of the fusion results and sustainably supporting dynamic adjustments to the plan. This is because the solution constructs a dual-dimensional (technology + business) fusion accuracy evaluation index system and establishes a dynamic iterative optimization mechanism based on planning implementation feedback, rather than a static, one-off fusion logic. It can continuously update the dataset and optimize model parameters based on actual data after plan implementation; the fusion results can be updated in real time, continuously reflecting dynamic changes in development and supporting dynamic adjustments to the plan and operational optimization. It solves the problems of static, poorly time-sensitive, and unadaptable fusion results of existing general technical solutions. The accuracy of the fusion results in this solution will not rapidly decline over time, and can sustainably support the entire lifecycle management of planning, including its preparation, adjustment, and operation.

[0459] The main innovations of this solution are as follows:

[0460] Architecture Innovation: For the first time, a two-level multi-source heterogeneous data fusion architecture with intra-modal adaptive fusion and cross-modal global correlation fusion for the entire planning process is proposed. This architecture solves the problems of feature loss and low fusion accuracy caused by single-level fusion in existing technologies, and realizes high-precision deep fusion of multi-modal heterogeneous data.

[0461] Business Adaptation Innovation: For the first time, the business rules of the entire process of planning resource survey - suitability evaluation - business layout - passenger flow forecast - benefit assessment are embedded into the feature selection, integration and output process. A three-level feature system is constructed, which solves the problems of existing technology being disconnected from business and poor practicality, and realizes one-time integration and full process reuse.

[0462] Compliance innovation: For the first time, compliance constraints are used as a veto item and embedded into the entire process of multi-source data fusion. A compliance verification rule base and automatic correction mechanism have been built, which solves the problem of high compliance risk and inability to implement existing technology fusion results, and ensures that the fusion results meet the planning requirements.

[0463] Innovative scenario adaptation: For the first time, a meta-learning small sample optimization module is embedded in the multi-source data fusion method for tourism planning, which solves the problems of insufficient data samples and model overfitting, and greatly improves the generalization ability and applicability of the model.

[0464] Mechanism Innovation: For the first time, a dual-dimensional evaluation index system for the integration of technology and business has been constructed, and a dynamic iterative optimization mechanism has been established. This solves the problems of static and poor timeliness of existing technology integration results and can support dynamic management of the entire planning lifecycle.

[0465] This application also provides a data fusion apparatus. It should be noted that the data fusion apparatus of this application can be used to execute the data fusion method provided in this application. This apparatus is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that performs a predetermined function. Although the apparatus described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0466] The data fusion apparatus provided in the embodiments of this application will be described below.

[0467] Figure 3 This is a structural block diagram of a data fusion apparatus according to an embodiment of this application. Figure 3 As shown, the device includes:

[0468] The first acquisition unit 10 is used to acquire relevant data of the area to be evaluated, wherein the relevant data includes at least remote sensing image data, pedestrian flow data, charging equipment data, vehicle flow data, air quality data and news data;

[0469] The classification unit 20 is used to classify the above-mentioned related data to obtain classified related data. The types of the classified related data include spatial geographic type, structured type, unstructured type and time series type. The spatial geographic type is the related data related to the spatial location of the area to be evaluated. The structured type is the related data with data structure. The unstructured type is the related data without data structure. The time series type is the related data related to time.

[0470] The first fusion unit 30 is used to perform a first fusion on the above-mentioned related data of the same type to obtain multiple similar fusion features. The first fusion is performed by using machine learning to perform feature fusion, and the number of similar fusion features is the same as the number of types of related data.

[0471] The first fusion unit 40 is used to perform a second fusion on all the above-mentioned similar fusion features to obtain a comprehensive fusion feature, wherein the second fusion is performed by using machine learning to fuse features.

[0472] The assessment unit 50 is used to conduct resource assessment on at least the aforementioned area to be assessed based on the above-mentioned integrated characteristics.

[0473] In this embodiment, machine learning can automatically learn the nonlinear mapping relationships within the data, extract the most representative core feature vectors from the original high-dimensional noisy data and perform a first fusion (i.e., "same-type fusion features"). By performing a second fusion on the "same-type fusion features" through machine learning, the correlation and interaction between feature vectors of different modalities can be learned. Through the two-level architecture of "intra-modal machine learning feature fusion (first fusion)" and "cross-modal machine learning global fusion (second fusion)," the traditional shallow weighted average is replaced, which can preserve the information entropy of the original data to the greatest extent, thereby improving the accuracy of data fusion and the accuracy of automated resource assessment of a certain area.

[0474] In the specific implementation process, the classification unit includes a first classification module, a second classification module, a third classification module, and a fourth classification module. The first classification module is used to determine the type of the relevant data as the spatial geographic type when the relevant data meets the first condition, thereby obtaining the classified relevant data. The first condition is that the relevant data includes one or more of longitude, latitude, and coordinates. The second classification module is used to determine the type of the relevant data as the structured type when the relevant data meets the second condition, thereby obtaining the classified relevant data. The second condition is that the relevant data includes tables, key-value pairs, fields of numeric type, and fields of other types. The first category is one or more of the following: enumeration type and Boolean type. The second category is used to determine the type of the above-mentioned related data as the above-mentioned unstructured type when the above-mentioned related data meets the third condition, and obtain the above-mentioned classified related data. The third condition is that the above-mentioned related data includes one or more of the following: text, JSON format, and XML format. The third category is used to determine the type of the above-mentioned related data as the above-mentioned time series type when the above-mentioned related data meets the fourth condition, and obtain the above-mentioned classified related data. The fourth condition is that the above-mentioned related data includes one or more of the following: timestamp, time axis, and time type.

[0475] This solution achieves accurate identification and organization of multi-source heterogeneous data through an automated classification mechanism based on the inherent attributes of the data (coordinates, structure, content type, and time characteristics). It solves the problems of low efficiency and inconsistent standards in manual classification in existing technologies, which lead to chaotic data preprocessing. It provides accurate data input for subsequent feature extraction using dedicated machine learning models for different modal characteristics, thereby improving the automation level and processing accuracy of the overall data fusion process.

[0476] In some embodiments, the aforementioned similar fusion features include a first fusion feature, a second fusion feature, a third fusion feature, and a fourth fusion feature. The first fusion unit includes a first fusion module, a second fusion module, a third fusion module, and a fourth fusion module. The first fusion module is used to form a first training set by combining historical related data of the aforementioned spatial geographic type and corresponding first labels, and to train a first data fusion model using the first training set to obtain a spatial geographic data fusion model. The first label is a historical first fusion feature corresponding to the historical related data of the aforementioned spatial geographic type in the first training set. The related data of the aforementioned spatial geographic type is input into the spatial geographic data fusion model to obtain the first fusion feature corresponding to the related data of the aforementioned spatial geographic type. The second fusion module is used to form a second training set by combining historical related data of the aforementioned structured type and corresponding second labels, and to train a second data fusion model using the second training set to obtain a structured data fusion model. The second label is a historical second fusion feature corresponding to the historical related data of the aforementioned structured type in the second training set. The related data of the aforementioned structured type is input into the spatial geographic data fusion model to obtain the first fusion feature corresponding to the historical related data of the aforementioned spatial geographic type. The input to the structured data fusion model yields the second fusion feature corresponding to the structured data. The third fusion module combines the unstructured historical data and corresponding third labels into a third training set, and uses this third training set to train the third data fusion model, resulting in an unstructured data fusion model. The third label represents the historical third fusion feature corresponding to the unstructured historical data in the third training set. The input of the unstructured data into the unstructured data fusion model yields the third fusion feature corresponding to the unstructured data. The fourth fusion module combines the time-series historical data and corresponding fourth labels into a fourth training set, and uses this fourth training set to train the fourth data fusion model, resulting in a time-series data fusion model. The fourth label represents the historical fourth fusion feature corresponding to the time-series historical data in the fourth training set. The input of the time-series data into the time-series data fusion model yields the fourth fusion feature corresponding to the time-series data.

[0477] This approach constructs a dedicated machine learning model for each data modality and performs supervised training using historical data and its corresponding labels, achieving deep feature extraction and fusion within each modality. This modality-specific modeling avoids feature interference and information loss that general models experience when processing heterogeneous data. The extracted first, second, third, and fourth fusion features accurately represent the core information of spatial, structural, semantic, and temporal dimensions, respectively, providing high-quality, high-fidelity feature inputs for subsequent cross-modal fusion and significantly improving the accuracy and robustness of the overall data fusion architecture.

[0478] In the specific implementation process, the above-mentioned device further includes a second acquisition unit and a first processing unit. The second acquisition unit is used to acquire a first model before training the first data fusion model using the first training set to obtain a spatial geographic data fusion model, wherein the first model is a U-Net model. The first processing unit is used to add a residual module after the encoding layer of the first model, so that the output of the encoding layer is used as the input of the residual module, and add a spatial attention module after the decoding layer of the first model, so that the output of the decoding layer is used as the input of the spatial attention module, thereby obtaining the first data fusion model.

[0479] This approach addresses the vanishing gradient problem during deep U-Net model training by introducing a residual module, thereby improving the model's learning ability and convergence speed. Furthermore, the introduction of a spatial attention module enables the model to dynamically focus on spatial regions in remote sensing imagery crucial for resource assessment (such as building clusters, water bodies, and specific vegetation cover areas), suppressing background noise interference. These improvements result in the extracted first fusion features possessing stronger spatial semantic representation and noise resistance, significantly enhancing the accuracy of spatial geographic data fusion.

[0480] In some embodiments, the above-described apparatus further includes a third acquisition unit and a second processing unit. The third acquisition unit is used to acquire a second model before training a second data fusion model using the second training set to obtain a structured data fusion model, wherein the second model is an SAE model. The second processing unit is used to add a normalization layer after the encoding layer of the second model, such that the output of the encoding layer is used as the input of the normalization layer, and to add a regularization layer after the normalization layer of the second model, such that the output of the normalization layer is used as the input of the regularization layer, thereby obtaining the second data fusion model.

[0481] This scheme addresses the common internal covariate shift problem during deep neural network training by introducing a normalization layer, accelerating model convergence and improving training stability. Furthermore, the introduction of a regularization layer effectively reduces the model's dependence on training data, preventing overfitting with limited samples. This improvement enables the extracted second fusion features to maintain the integrity of the data's structured information while possessing stronger generalization capabilities.

[0482] In the specific implementation process, the above-mentioned device further includes a fourth acquisition unit and a third processing unit. The fourth acquisition unit is used to acquire the third model before training the third data fusion model with the third training set to obtain the unstructured data fusion model. The third model is the BERT-wwm model. The third processing unit is used to add an attention module before the output layer of the third model, so that the output of the attention module is used as the input of the output layer to obtain the third data fusion model.

[0483] This solution addresses the issues of topic dispersion and noise interference with key information that may arise in the standard BERT model when processing long texts or complex semantics by introducing a topic attention module. BERT-wwm provides rich contextual semantic representations, while the topic attention mechanism guides the model to focus on strongly relevant semantic topics, filtering out irrelevant background information. This improvement makes the third fusion feature extracted by the model more semantically focused and business-relevant, thereby significantly improving the information extraction accuracy of unstructured text data in resource assessment.

[0484] In some embodiments, the above-described apparatus further includes a fifth acquisition unit and a fourth processing unit. The fifth acquisition unit is used to acquire a fourth model before training a fourth data fusion model using the fourth training set to obtain a time-series data fusion model, wherein the fourth model is a Transformer model. The fourth processing unit adds an encoding layer after the input layer of the fourth model, such that the output of the input layer is used as the input of the encoding layer, and adds a multi-head self-attention module after the encoding layer of the fourth model, such that the output of the encoding layer is used as the input of the multi-head self-attention module, thereby obtaining the fourth data fusion model.

[0485] This scheme addresses the issues of low dimensionality and insufficient semantic information in the original time-series data by introducing a linear encoding layer, thereby enhancing the model's ability to nonlinearly map time-series features. Furthermore, the introduction of a multi-head self-attention module, particularly combined with a time interval bias mechanism, enables the model to simultaneously capture long-term dependencies and short-term fluctuations in the time-series data and dynamically allocate importance weights at different time steps. This improvement allows the fourth fusion feature extracted by the model to more accurately reflect the periodic changes and trends in the time-series data, thus significantly improving the accuracy and predictive power of time-series data fusion.

[0486] In the specific implementation process, the second fusion unit includes an acquisition module, a first processing module, a second processing module, and a third processing module. The acquisition module is used to acquire the fifth model, wherein the fifth model is a GNN model. The first processing module is used to add a spatiotemporal attention module after the input layer of the fifth model, so that the output of the input layer is used as the input of the spatiotemporal attention module, and to add a residual connection module before the normalization layer of the fifth model, so that the output of the residual connection module is used as the input of the normalization layer, thereby obtaining the fifth data fusion model. The second processing module is used to combine multiple of the above-mentioned similar fusion features into a multi-type dataset, and to combine the above-mentioned multi-type dataset and the corresponding fifth label into a fifth training set. The fifth data fusion model is trained using the above-mentioned fifth training set to obtain a comprehensive data fusion model, wherein the above-mentioned fifth label is the historical fifth fusion feature corresponding to the historical multi-type dataset of the above-mentioned fifth training set. The third processing module is used to input the above-mentioned multi-type dataset into the above-mentioned comprehensive data fusion model to obtain the above-mentioned comprehensive fusion feature corresponding to the above-mentioned multi-type dataset.

[0487] This scheme addresses the problem of traditional GNNs' inability to effectively distinguish the importance of different modalities and spatiotemporal dimensions when processing multi-source heterogeneous data by introducing a spatiotemporal attention module. It achieves dynamic weighted fusion of cross-modal features and uncovers potential correlations between different data sources. Furthermore, by introducing residual connections, it solves the information loss problem caused by excessive feature smoothing in deep graph neural networks, preserving the original feature details of nodes (modalities). This improvement results in more expressive and robust integrated features extracted by the model, comprehensively reflecting the interaction of spatial, structural, semantic, and temporal multidimensional information of the region to be evaluated, thereby significantly improving the overall accuracy of resource assessment.

[0488] In some embodiments, the first fusion unit includes a preprocessing module and a fifth fusion module. The preprocessing module is used to preprocess the aforementioned related data of different types to obtain preprocessed related data. The preprocessing methods include one or more of the following: unified coordinates, unified spatial resolution, geometric correction, missing value imputation, outlier removal, normalization, text cleaning, word segmentation and entity recognition, feature vectorization, unified time, and smoothing. The fifth fusion module is used to perform the aforementioned first fusion on the aforementioned preprocessed related data of the same type to obtain multiple aforementioned fused features of the same type.

[0489] This scheme effectively addresses technical challenges associated with multi-source heterogeneous data, such as inconsistent spatiotemporal benchmarks, high noise levels, numerous missing values, inconsistent dimensions, and the difficulty in directly utilizing unstructured data, by implementing differentiated preprocessing procedures tailored to different data modalities. Unified coordinates and resolution resolve spatial data misalignment issues; missing value imputation and outlier removal improve the integrity of structured data; text cleaning and vectorization enable the structured transformation of unstructured data; and time unification and smoothing enhance the stability of time-series data. These preprocessing measures provide high-quality, standardized data input for the subsequent first-stage fusion.

[0490] The data fusion device includes a processor and a memory. The first acquisition unit, classification unit, first fusion unit, second fusion unit, and evaluation unit are all stored as program units in the memory. The processor executes the program units stored in the memory to achieve the corresponding functions. All of the above modules are located in the same processor; alternatively, the modules may be located in different processors in any combination.

[0491] The processor contains a kernel, which retrieves the corresponding program unit from memory. One or more kernels can be configured, and their parameters can be adjusted to address the limitations of existing technologies that use simple weighted averaging for resource assessment of a specific region. However, this simple weighted averaging method suffers from significant information entropy loss, resulting in poor accuracy of the data fusion results and consequently, inaccurate automated resource assessment of a given region.

[0492] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0493] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the data fusion method.

[0494] This invention provides a processor for running a program, wherein the program executes a data fusion method during runtime.

[0495] This invention provides a data fusion system, including one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include a method for performing any of the above-described data fusion methods.

[0496] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements at least the steps of a data fusion method. The device described herein may be a server, PC, PAD, mobile phone, etc.

[0497] This application also provides a computer program product that, when executed on a data processing device, is adapted to perform a program that initializes a fusion method step having at least the following data.

[0498] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0499] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0500] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0501] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0502] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0503] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0504] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0505] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0506] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0507] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0508] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A data fusion method, characterized in that, include: Obtain relevant data for the area to be evaluated, including at least remote sensing image data, pedestrian flow data, charging equipment data, vehicle flow data, air quality data, and news data; The relevant data is classified to obtain classified relevant data. The types of classified relevant data include spatial geographic type, structured type, unstructured type and time series type. The spatial geographic type is the relevant data related to the spatial location of the area to be evaluated. The structured type is the relevant data with data structure. The unstructured type is the relevant data without data structure. The time series type is the relevant data related to time. A first fusion is performed on the related data of the same type to obtain multiple similar fusion features. The first fusion is a feature fusion performed using machine learning. The number of similar fusion features is the same as the number of types of related data. A second fusion is performed on all the aforementioned similar fusion features to obtain a comprehensive fusion feature, wherein the second fusion is a feature fusion performed using machine learning. Based on the aforementioned integrated characteristics, at least the resource assessment of the area to be assessed shall be performed.

2. The method according to claim 1, characterized in that, The relevant data is classified to obtain classified relevant data, including at least one of the following: If the relevant data meets the first condition, the type of the relevant data is determined to be the spatial geographic type, and the classified relevant data is obtained. The first condition is that the relevant data includes one or more of longitude, latitude, and coordinates. If the relevant data meets the second condition, the type of the relevant data is determined to be the structured type, and the classified relevant data is obtained. The second condition is that the relevant data includes one or more of the following: tables, key-value pairs, fields of numeric type, fields of enumeration type, and fields of Boolean type. If the relevant data meets the third condition, the type of the relevant data is determined to be the unstructured type, and the classified relevant data is obtained. The third condition is that the relevant data includes one or more of the following: text, fields in JSON format, and fields in XML format. If the relevant data meets the fourth condition, the type of the relevant data is determined to be the time series type, and the classified relevant data is obtained. The fourth condition is that the relevant data includes one or more of the following: timestamp, time axis, and field type is time.

3. The method according to claim 1, characterized in that, The similar fusion features include a first fusion feature, a second fusion feature, a third fusion feature, and a fourth fusion feature. Performing a first fusion on the related data of the same type yields multiple similar fusion features, including at least one of the following: The historical relevant data of the spatial geographic type and the corresponding first label are combined to form a first training set. The first data fusion model is trained using the first training set to obtain a spatial geographic data fusion model. The first label is the historical first fusion feature corresponding to the historical relevant data of the spatial geographic type in the first training set. The relevant data of the spatial geographic type is input into the spatial geographic data fusion model to obtain the first fusion feature corresponding to the relevant data of the spatial geographic type. The structured type of historical related data and the corresponding second label are combined to form a second training set. The second training set is used to train a second data fusion model to obtain a structured data fusion model. The second label is the historical second fusion feature corresponding to the structured type of historical related data in the second training set. The structured type of related data is input into the structured data fusion model to obtain the second fusion feature corresponding to the structured type of related data. The unstructured historical related data and the corresponding third label are combined to form a third training set. The third training set is used to train a third data fusion model to obtain an unstructured data fusion model. The third label is the historical third fusion feature corresponding to the unstructured historical related data in the third training set. The unstructured related data is input into the unstructured data fusion model to obtain the third fusion feature corresponding to the unstructured related data. The historical related data of the time series type and the corresponding fourth label are combined to form a fourth training set. The fourth data fusion model is trained using the fourth training set to obtain the time series data fusion model. The fourth label is the historical fourth fusion feature corresponding to the historical related data of the time series type in the fourth training set. The related data of the time series type is input into the time series data fusion model to obtain the fourth fusion feature corresponding to the related data of the time series type.

4. The method according to claim 3, characterized in that, Before training the first data fusion model using the first training set to obtain the spatial geographic data fusion model, the method further includes: Obtain the first model, wherein the first model is a U-Net model; A residual module is added after the encoding layer of the first model, so that the output of the encoding layer is used as the input of the residual module. A spatial attention module is added after the decoding layer of the first model, so that the output of the decoding layer is used as the input of the spatial attention module, thus obtaining the first data fusion model.

5. The method according to claim 3, characterized in that, Before training the second data fusion model using the second training set to obtain the structured data fusion model, the method further includes: Obtain the second model, wherein the second model is an SAE model; A normalization layer is added after the encoding layer of the second model, so that the output of the encoding layer is used as the input of the normalization layer. A regularization layer is added after the normalization layer of the second model, so that the output of the normalization layer is used as the input of the regularization layer, thus obtaining the second data fusion model.

6. The method according to claim 3, characterized in that, Before training the third data fusion model using the third training set to obtain the unstructured data fusion model, the method further includes: Obtain the third model, wherein the third model is the BERT-wwm model; An attention module is added before the output layer of the third model, so that the output of the attention module is used as the input of the output layer, thus obtaining the third data fusion model.

7. The method according to claim 3, characterized in that, Before training the fourth data fusion model using the fourth training set to obtain the time-series data fusion model, the method further includes: Obtain the fourth model, wherein the fourth model is the Transformer model; An encoding layer is added after the input layer of the fourth model, such that the output of the input layer is used as the input of the encoding layer. A multi-head self-attention module is added after the encoding layer of the fourth model, such that the output of the encoding layer is used as the input of the multi-head self-attention module, thus obtaining the fourth data fusion model.

8. The method according to claim 1, characterized in that, A second fusion is performed on all the aforementioned similar fusion features to obtain a comprehensive fusion feature, including: Obtain the fifth model, wherein the fifth model is a GNN model; A spatiotemporal attention module is added after the input layer of the fifth model, so that the output of the input layer is used as the input of the spatiotemporal attention module. A residual connection module is added before the normalization layer of the fifth model, so that the output of the residual connection module is used as the input of the normalization layer, thus obtaining the fifth data fusion model. Multiple similar fusion features are combined into a multi-type dataset, and the multi-type dataset and the corresponding fifth label are combined into a fifth training set. The fifth data fusion model is trained using the fifth training set to obtain a comprehensive data fusion model. The fifth label is the historical fifth fusion feature corresponding to the historical multi-type dataset of the fifth training set. The multi-type datasets are input into the comprehensive data fusion model to obtain the comprehensive fusion features corresponding to the multi-type datasets.

9. The method according to any one of claims 1 to 8, characterized in that, The first fusion of the related data of the same type yields multiple similar fusion features, including: The relevant data of different types are preprocessed to obtain preprocessed relevant data. The preprocessing methods include one or more of the following: unified coordinates, unified spatial resolution, geometric correction, missing value imputation, outlier removal, normalization, text cleaning, word segmentation and entity recognition, feature vectorization, unified time, and smoothing. The first fusion is performed on the preprocessed related data of the same type to obtain multiple similar fusion features.

10. A data fusion system, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including a method for performing data fusion according to any one of claims 1 to 9.

Images