A method and device for expressing semantics of earth science data based on meta-representation rules

Abstract

The application discloses a kind of based on meta representation rule geoscience data semantic expression method and device, belong to geoscience data processing and artificial intelligence technical field.The method includes: the ontology characteristic information of geoscience data is preprocessed and based on knowledge mapping semantic extension;Utilize large language model to enhance the function semantics of the expanded semantics, simulate multi-field expert mechanism to improve data comprehensibility;Ontology semantics and enhanced semantics are fused with multilevel feature coding and vectorization, and a unified semantic expression vector is constructed.Based on meta representation rule, combined with user demand semantics, an interpretable and computable expression vector is generated, and adaptive optimization and personalized adjustment are realized through a rule learning module.The application realizes the leap of geoscience data from "descriptive semantics" to "functional semantics", significantly improves the discoverability, reusability and interdisciplinary application ability of data, and provides support for scientific data intelligent service.

Description

Technical Field

[0001] This invention belongs to the field of Earth science data processing and artificial intelligence technology, specifically relating to a method and apparatus for semantic representation of Earth science data based on meta-representation rules. Background Technology

[0002] Earth science data is characterized by its multi-source nature, heterogeneity, massive volume, high dimensionality, and strong spatiotemporal dynamics, encompassing multiple disciplines such as remote sensing, meteorology, oceanography, ecology, and geology. This data is typically described in structured or semi-structured metadata, containing various information elements such as text descriptions, category labels, and spatiotemporal ranges. However, in practical applications, these valuable data resources often form "data silos," lacking effective semantic connections between them. Furthermore, traditional data catalogs and representation methods struggle to efficiently handle complex semantic requests to meet the demands of large-scale models and intelligent computing, severely limiting data discoverability, reusability, and interdisciplinary decision support capabilities.

[0003] Traditional data representation methods primarily focus on structure and readability, with typical approaches including columnar feature tables, feature vector and vector database representations, and graph structures. While columnar feature tables can accommodate multi-category features, they suffer from drawbacks such as complex semantic mapping relationships, significant metadata redundancy and consistency issues, and high difficulty in constructing collaborative association networks, leading to low computational read / write efficiency. Feature vector and vector database representations rely on objective mapping of data feature information, offering advantages in efficiently handling high-dimensional unstructured features and high compatibility with large models. However, they suffer from poor interpretability, lack of business semantic support, and high sensitivity to vector generation quality. Graph structures preserve semantic association information and are compatible with multi-source heterogeneous features, but face challenges such as poor interpretability when semantics are ambiguous, weak query generality, complex toolchains, and insufficient compatibility with traditional computing frameworks.

[0004] Computation-oriented big data feature representation methods, such as elastic distributed feature datasets, tensor-based feature representations, feature cubes, and distributed sparse features, are based on data structures (i.e., arrays or tensors) that meet the computational needs of computers. These representation methods do not carry semantic information themselves; the preservation and restoration of the original semantics rely entirely on metadata's complete record of the array structure, source, and transformation logic. This results in a complex system structure and a passive computation process, making it difficult to achieve data-driven intelligent computing.

[0005] In summary, existing methods for representing Earth science data have significant shortcomings in terms of semantic mapping, interpretability, computational efficiency, and business adaptability. There is an urgent need for a unified representation method that can integrate semantic understanding and efficient computation to support the intelligent application of Earth science data in the era of large-scale models. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention provides a method and apparatus for semantic representation of Earth science data based on meta-representation rules. This method can establish a unified expression of the ontological features (text, category, spatiotemporal dimension, etc.) and semantically enhanced functional semantic features of Earth science data. By establishing a proactive semantic expression method driven by expression vectors and oriented towards user use and large models through meta-representation rules, the discoverability, reusability, and scientific value of Earth science data are improved.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A semantic representation method for Earth science data based on meta-representation rules, the method comprising:

[0009] Step 1: Load and verify the feature information metadata of the Earth science data, process the feature information and separate the feature semantics, and perform structured semantic extension on the separated ontology semantics based on domain knowledge mapping to obtain extended ontology semantics;

[0010] Step 2: Use a large language model to simulate expert dialogues in multiple rounds on the extended ontology semantics to achieve functional semantic enhancement and obtain functional semantics;

[0011] Step 3: Perform text normalization and deduplication on the extended ontology semantics and functional semantics, embed and encode them using a language model, and numerically encode the non-textual semantic features. Then, fuse and concatenate the multi-source vectors into a unified semantic expression vector.

[0012] Step 4: Construct a rule vector based on the meta-representation rules, parse the semantics of user needs, and adaptively adjust and weight the rule parameters of each dimension in the rule vector in combination with the unified semantic expression vector to generate a rule expression vector that is adapted to the semantics of user needs.

[0013] Furthermore, the feature information processing in step 1 includes text segmentation, category tag extraction, and spatiotemporal information parsing: the text segmentation breaks down unstructured text fields into semantic units and counts their frequencies; the category tag extraction extracts all categories from the information and assigns weights according to priority; the spatiotemporal information parsing performs unit standardization transformation and content parsing on the spatiotemporal fields, uniformly expressing them in numerical form; and the feature semantic separation divides the data information into three types of semantic features: text features, category features, and spatiotemporal features.

[0014] Furthermore, the ontology semantic expansion based on domain knowledge mapping in step 1 includes semantic enrichment, structural normalization, and knowledge association: by extracting numerical values ​​and units, the spatial resolution field is expanded into structured information containing numerical values, units, type levels, and application applicability; by calculating the time span, the time range field is expanded into structured information containing start and end times, duration in days, and time type; and by word segmentation and classification, the keyword field is expanded into structured information including subject terms, technical terms, related keywords, and subject classification links.

[0015] Furthermore, step 2, which utilizes a large language model to conduct multi-round simulated expert dialogues, includes: configuring multiple expert roles across various professional fields and designing prompts to conduct multi-round dialogues from the dimensions of scientific value, application scenarios, technical characteristics, data quality, relevance, traceability, and usage guidelines; extracting structured information from the dialogue results to obtain functional semantic enhancement results in the dimensions of core concepts, relationships, application classifications, quality indicators, and technical terms; and quantitatively evaluating the quality of functional semantics through consistency checks, professional terminology verification, and usability assessment.

[0016] Furthermore, step 3, which involves text normalization and deduplication of extended ontology semantics and functional semantics, includes: forming coherent sentences from extended ontology semantics through structural reorganization and standardizing terminology; forming comprehensive descriptions from enhanced semantics of the large language model through temporal summarization and cross-expert information fusion; and clustering semantic texts through similarity calculation and retaining the central sentence to achieve information deduplication.

[0017] Furthermore, step 3, which involves numerically encoding non-textual semantic features, includes: generating dimensional vectors for subject classification, data type, and earth science data classification using a hybrid encoding method for categorical features; generating dimensional vectors for region type, time type, and spatial resolution type for spatiotemporal features based on domain knowledge using semantic encoding; and generating dimensional vectors for spatial resolution, time span, description length, data size, and update frequency for numerical features using a logarithmic normalization method.

[0018] Furthermore, step 4, which involves constructing a rule vector based on meta-representation rules, includes: predefining a fixed-dimensional rule vector consisting of six categories: similarity calculation rules, clustering analysis rules, semantic retrieval rules, interpretation rules, generation rules, and interaction rules. Each dimension corresponds to a specific representation rule or strategy. The interaction rules encode data format and structure attributes, quality and accuracy attributes, recommendation tool and method attributes, and application guidance attributes, which are used to provide data processing operation guidance, tool recommendations, and code examples to achieve semantically consistent interaction among data, model, and user.

[0019] On the other hand, the present invention provides a semantic representation device for Earth science data based on meta-representation rules, comprising:

[0020] The ontology semantic preprocessing module is used to load and verify the feature information metadata of Earth science data, process feature information and separate feature semantics, and perform structured semantic extension on the separated ontology semantics based on domain knowledge mapping to obtain extended ontology semantics.

[0021] The semantic enhancement module based on the LLM model is used to perform multi-round simulated expert dialogue on the extended ontology semantics using a large language model in order to achieve functional semantic enhancement and obtain functional semantics.

[0022] The multidimensional feature vectorization module is used to perform text normalization and deduplication on the extended ontology semantics and functional semantics, embedding and encoding them through a language model, and numerically encoding non-textual semantic features, and fusing and splicing multi-source vectors into a unified semantic expression vector.

[0023] The expression vector generation module is used to construct rule vectors based on meta-representation rules, parse user demand semantics, and adaptively adjust and weight the rule parameters of each dimension in the rule vector in combination with the unified semantic expression vector to generate a rule expression vector that is adapted to the user demand semantics.

[0024] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for semantic representation of Earth science data based on meta-representation rules.

[0025] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for semantic representation of Earth science data based on meta-representation rules.

[0026] The beneficial effects of this invention are as follows:

[0027] First, by introducing ontology semantic extension based on knowledge mapping and functional semantic enhancement mechanisms based on large language models, this invention achieves a qualitative leap in Earth science data from "descriptive semantics" to "functional semantics." Traditional methods can only describe "what" the data is, while this invention can further reveal "why the data is used, how it is applied, and where it is used," significantly improving the richness, professionalism, and comprehensibility of semantic expression, and laying a semantic foundation for interdisciplinary data discovery and knowledge mining.

[0028] Second, this invention constructs an expression vector generation framework based on meta-representation rules, encoding multiple types of rules, including similarity calculation, cluster analysis, semantic retrieval, interpretation, generation, and interaction, into learnable rule vectors, achieving deep integration of semantic features and user demand semantics. This method retains the high-dimensional computational efficiency and large-model compatibility of vector representation while embedding business semantics and interpretable decision logic, solving the technical shortcomings of traditional feature vectors such as lack of interpretability, weak universality of graph structure queries, and reliance on external metadata for tensor representation.

[0029] Third, this invention supports incremental learning and online adaptation mechanisms, enabling it to learn the optimal representation strategy end-to-end from massive amounts of labeled data and user feedback, and adjust rule vectors in real time according to personalized needs, giving the system continuous evolution capabilities. Based on this, this invention provides semantic understanding and knowledge service support with high accuracy, strong interpretability, and adaptive capabilities for diverse application scenarios such as interdisciplinary data discovery, SDG indicator-supported analysis, and computationally ready data recommendation. Attached Figure Description

[0030] Figure 1 This is a flowchart of a semantic representation method for Earth science data based on meta-representation rules according to the present invention.

[0031] Figure 2 This is a flowchart of proactive data representation based on meta-representation rules;

[0032] Figure 3 This is a schematic diagram of the functional semantic enhancement principle based on the LLM model.

[0033] Figure 4 This is a diagram illustrating the principle of rule triggering conditions based on user demand semantics.

[0034] Figure 5 A schematic diagram illustrating the principle of generating data semantic representation vectors based on rule vectors. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0036] The terminology used in the description of the embodiments of this application adopts commonly used terms in this technical field and has the same meaning as generally understood in the field to which the embodiments of this application pertain. The nouns and terms used in the embodiments of this application are explained and are subject to the following interpretations.

[0037] Earth science data is a massive collection of data from multiple sources, including satellite remote sensing, ground monitoring, social perception, and simulation models, which has strong spatiotemporal attributes and is interconnected with the Earth system.

[0038] Earth science data feature information refers to the feature information extracted based on earth science data, including fields such as name, description, keywords, type, subject classification, earth science data classification, SDG classification, spatial resolution, temporal resolution, geographical range, time range, processing method description, author, and institution.

[0039] Ontological semantics, or ontological feature information, refers to the semantic information of the data ontology contained in the feature information of Earth science data.

[0040] Ontology semantic extension refers to the professional extension of ontology semantic information based on domain knowledge mapping in terms of concepts, classification systems, and related knowledge.

[0041] Functional semantics refers to the ability to express, understand, or utilize ontology semantics through information fusion in fields such as natural language processing (NLP), knowledge graphs, and earth science data fusion, thereby expanding functional information in areas such as data tracing, usage methods, and application domains. In this application's embodiments, the data value semantics enhanced by an LLM model constructed using local knowledge base based on semantic information expanded from ontology semantics is referred to as functional semantics.

[0042] Metarepresentation rules refer to a set of rules used to guide and standardize "how to express information." They are rules about "representation methods," rather than rules that directly describe data information. They define the structure, constraints, interpretation methods, and usage specifications of data representation. In this embodiment, metarepresentation rules are used to learn rules from a unified representation vector to generate a semantic representation vector.

[0043] A rule vector is a dense numerical vector that serves as a mathematical encoding of rules. The essence of rule vector encoding is "the rule strategy for how to express information," rather than the data information itself, but rather a set of adjustable parameters learned by the system from data or user feedback. The rule vectors used in this application are 64-dimensional.

[0044] Expression vector is the semantic term for rule vector. It refers to the rule vector that is constructed by the rule learning model based on the semantics of user needs and can autonomously guide the method to analyze data. In essence, it is a 64-dimensional vector translated from rule vector. It expresses the data expression strategy for the semantics of user needs for a specific query.

[0045] In response to the surge in Earth science data resources, traditional big data association mining, data discovery, and knowledge graph techniques largely rely on keyword matching, normalized labels, feature vector representation, graph structure / statistical methods, and tensor representation. These methods have shortcomings in terms of semantic understanding, interpretability, generalization ability, computational efficiency, and multi-task adaptability of data feature information. This application's embodiments transform the feature information of massive Earth science data into high-dimensional vectors containing semantic features, and further generate accurate expression vectors through meta-representation rules. This is a paradigm oriented towards the integration of "semantic interpretability + computational efficiency," constructing a rule-based semantic space embedding, semantic enhancement-driven data understanding, lightweight representation, and efficient use compatible expression vectors. This effectively improves the expression and application of data feature information, providing technical support for scenarios such as interdisciplinary data discovery, SDG indicator-supported analysis, and computationally ready data recommendation.

[0046] Reference Figure 1 This application provides a semantic representation method for Earth science data based on meta-representation rules. Its core idea is "how to solve the problem of proactively expressing data semantic information according to user needs" (see [reference]). Figure 2 The main steps include ontology semantic preparation and processing, semantic enhancement based on the LLM model, unified expression of semantic feature vectorization, and expression vector generation based on meta-representation rules. Specifically:

[0047] Step 1: Ontology semantic preprocessing and knowledge-mapping-based semantic extension; Loading and verifying the feature information metadata of Earth science data, processing feature information and separating feature semantics, and performing structured semantic extension on the separated ontology semantics based on domain knowledge mapping to obtain extended ontology semantics;

[0048] The semantics of Earth science data ontology primarily derive from metadata descriptions of entity data, including fields such as product name, description, keywords, product type, subject classification, Earth science data classification, SDG classification, spatial resolution, temporal resolution, geographical extent, time range, processing method description, author, and institution. These fields can be either text or numerical. Taking the "2020 Global 30m Farmland Spatial Distribution Product" dataset ontology semantics as an example, some key fields include:

[0049] {

[0050] Product Name: "2020 Global 30m Farmland Spatial Distribution Product"

[0051] "Product Type": "Stored using a block-based method...representing arable land",

[0052] Product Overview: Based on the Earth Observations Organization's Global Agriculture Monitoring Flagship Program...the overall accuracy is 89.4%.

[0053] Subject Classification: 42020 Photogrammetry and Remote Sensing Technology

[0054] Keywords: arable land; global; remote sensing; 30m resolution

[0055] Spatial resolution: 30m

[0056] "Time resolution": "years",

[0057] Projection method: WGS84 coordinate system, latitude and longitude projection

[0058] "Geographical Scope": "The geographical scope covers 180° east longitude to 180° west longitude and 60° south latitude to 80° north latitude. High-latitude regions in both the Northern and Southern Hemispheres and the oceans are not within the monitoring scope."

[0059] "Start Time": "2020-01-01",

[0060] End Time: 2020-12-31

[0061] Data type: "raster"

[0062] Data format: "GeoTiff"

[0063] "SDGs Classification": "SDG 2.3 Agricultural Productivity and Income, SDG 2.4 Sustainable Food Production Systems"

[0064] Earth Science Data Classification: Land Surface, Farmland, Earth Observation, Satellite Remote Sensing

[0065] ...}

[0066] Data loading and inspection: By loading the feature information metadata file, missing value checks, data format consistency verification, and field rationality verification are performed.

[0067] Missing value checking fills missing values ​​in text and numeric fields of the original information with standard empty strings and 0 values ​​respectively, avoiding parsing errors caused by NaN, and also applies to the statistical information of preprocessed records. Data format consistency verification unifies the data types in the original information, such as unifying text fields to string types and time-related fields to floating-point types, and standardizes feature formats, using a unified container type when extracting features to ensure consistent feature structures across different records. Field rationality verification verifies the rationality of the spatiotemporal range, feature existence, and statistical feature validity.

[0068] Feature information processing: Perform data cleaning and standardization, including text word segmentation, category label extraction (supporting comma / slash separation), spatiotemporal information parsing (e.g., uniformly converting "0.5°" to metric units), etc.

[0069] Text word segmentation is intended to split unstructured text fields (such as "description", "keywords", etc.) into basic units for semantic analysis, to facilitate subsequent vector encoding and semantic representation. A string segmentation method is adopted for text feature extraction, and the occurrence frequency of segmented word semantics is processed to weaken the repeated use of common terms, and the occurrence frequency of different words is counted respectively, so as to improve the accuracy of semantic expression. Category label extraction is to extract all categories in the information, and the relevant labels extracted from each field are directly encoded into the vector according to the label weight calculation logic, and weights are assigned according to priority. Spatiotemporal information parsing is to identify, perform unit standardization conversion and content parsing on the spatiotemporal fields (such as "spatial resolution", "temporal resolution", "geographic range", etc.) in the original information, so as to uniformly express the information in a numerical form.

[0070] Feature semantic separation: Divide data information into three types of semantic features, namely text features (product name, description, processing method description, etc.), category features (product type, SDG classification, geoscience data classification) and spatiotemporal features (spatial / temporal resolution, geographic / temporal range).

[0071] Ontology semantic expansion based on knowledge mapping: A process of structurally enhancing ontology semantics based on a predefined domain knowledge graph and mapping rules, whose objectives are embodied in semantic enrichment, structure standardization, and knowledge association. Semantic enrichment refers to converting simple field values into a rich semantic network; structure standardization refers to unifying the expression modes of different information; knowledge association refers to establishing links between feature information and the domain knowledge system.

[0072] Taking the 2020 global 30m cropland spatial distribution product as an example, in its ontology semantics, "spatial resolution": "30m", by extracting the value and unit and classifying the resolution level, it is expanded to:

[0073] {

[0074] "resolution": "30m",

[0075] "value": 30,

[0076] "unit": "meter",

[0077] "type": "medium-high resolution",

[0078] "features": ["Land use monitoring", "Farmland identification", "Regional scale analysis"],

[0079] "Applicability": Suitable for farmland monitoring and agricultural planning.

[0080] }

[0081] In the ontology semantics, the time features "Start Time": "2020-01-01", "End Time": "2020-12-31", are expanded by calculating the time span and classifying the time type:

[0082] {

[0083] "start": "2020-01-01",

[0084] "end": "2020-12-31",

[0085] "span_days": 366, / / 2020 is a leap year

[0086] "temporal_type": "Annual Monitoring",

[0087] "temporal_features": ["single-year data", "full-year coverage", "agricultural production year"]

[0088] }

[0089] The keywords "arable land; global; remote sensing; 30m resolution" in the ontology semantics suffer from semantic sparsity (lack of deep association), unclassification (mixed topics and technical keywords), and lack of related concept expansion. Through word segmentation and classification, semantic association expansion, and domain knowledge linking, they are expanded as follows:

[0090] {

[0091] "original_keywords": ["arable land", "global", "remote sensing", "30m resolution"],

[0092] "themes": ["arable land", "global"],

[0093] "technologies": ["Remote Sensing", "30m Resolution"],

[0094] "related_keywords": [

[0095] / / Related to arable land

[0096] Agriculture, Crops, Farmland, Land Use, Food Security, Agricultural Production, Agricultural Monitoring, Crop Monitoring, Land Management

[0097] / / Global Related

[0098] "International", "Transnational", "Multi-regional", "Worldwide", "Global Change", "International Cooperation", "Global Agriculture"

[0099] / / Remote Sensing Related

[0100] "Satellite", "Earth Observation", "Imagery", "Monitoring Technology", "Satellite Remote Sensing", "Optical Remote Sensing", "Landsat", "Sentinel-2"

[0101] / / Resolution related

[0102] [Medium resolution, Landsat scale, Regional monitoring, Spatial analysis, Geographic Information System]

[0103] "sdg_links": ["SDG2.3 Agricultural Productivity and Income", "SDG2.4 Sustainable Food Production Systems"],

[0104] "earth_data_categories": ["farmland", "vegetation", "land use / cover"]

[0105] }

[0106] Ontology semantic extension essentially involves two aspects: firstly, adding deterministic knowledge to the foundation of ontology semantics, deepening semantics in terms of concept coverage, constructing a classification system in terms of organizational structure, and forming a network in terms of knowledge association, thus providing a foundation for subsequent semantic enhancement; secondly, providing more and richer semantic features through vectorization, and improving the accuracy and consistency of vector representation by combining deterministic knowledge.

[0107] Step 2: Semantic enhancement based on LLM model: Use a large language model to simulate expert dialogue in multiple rounds on the extended ontology semantics, and enhance the functional semantics from the dimensions of data source, technical details, application scenarios and scientific value to obtain functional semantics;

[0108] like Figure 3 As shown, the core of this step is to extend static feature information into dynamic, computable functional semantic knowledge based on ontology semantic extension. It transforms the traditional feature information that describes what the data is (What) into enhanced functional semantics, that is, why the data is used, how it is used, where it is applied, and when it is represented (Why / How / Where / When), thus achieving a semantic leap in data comprehensibility.

[0109] Domain knowledge injection strategy: The LLM model selects a locally deployed discrete deepseek-v3 model, using geoscience literature, geoscience data feature metadata datasets, scientific data corpora, and general scientific knowledge to build a local knowledge base. During training, the domain knowledge accounts for 40%, 30%, 20%, and 10%, respectively.

[0110] This invention simulates a multi-expert analysis mechanism, using data features to simulate multiple expert roles engaging in multi-round simulated dialogues via prompts. Semantic enhancement is achieved by considering aspects such as data source, technical details, usage methods, application scenarios, and scientific value. The device supports the selection and addition of multi-domain expert roles and the self-design of prompts.

[0111] For example, based on the ontological semantics of the 2020 global 30m arable land spatial distribution product, five professional expert roles were configured: a senior expert in earth science and remote sensing, an expert in SDG2 indicator research, an expert in social statistics and economics, an expert in agriculture, and an expert in urban planning. Prompt words were designed from eight aspects: scientific value, application scenarios, technical characteristics, data quality, relevance, traceability, and usage guidelines.

[0112] {

[0113] "Scientific Value": What is the core value of this dataset in Earth science research?

[0114] "Application Scenarios": In which specific research or applications can it play a role?

[0115] "Technical Features": How do technical parameters such as spatiotemporal resolution affect usage?

[0116] "Data Quality": Based on the description, what are some potential data quality issues and considerations?

[0117] "Relevance": What other datasets or research areas are there potential relevances?

[0118] "Tracing the source": What data sources and methods were used in the production process of this dataset?

[0119] User Guide: What is the format of this dataset? What software tools or methods are required?

[0120] }

[0121] Through a four-round dialogue process, including "basic information confirmation and preliminary analysis", "in-depth discussion of technical details", "specific analysis of application scenarios", and "limitations and improvement suggestions", the functional semantic enhancement based on cross-domain knowledge and applications is gradually completed.

[0122] Structured information extraction involves extracting information from simulated expert dialogues. Structured dimensions are set, including core concepts, relationships, application classifications, quality indicators, and technical terms, to obtain semantically enhanced results. For example, the semantically enhanced description of "Product Name: 2020 Global 30m Arable Land Spatial Distribution Product" is: "The dataset '2020 Global 30m Arable Land Spatial Distribution Product' is an important Earth science data product. The core concepts involved include: arable land, agriculture, remote sensing, and spatial analysis. It can be applied to: food security monitoring, agricultural planning, land use management, and other fields. The data covers mid-to-low latitude landmasses globally, with a 30-meter spatial resolution and an annual temporal resolution."

[0123] The quality assessment of functional semantics is quantitatively evaluated through consistency checks, terminology verification, and usability assessment. Taking the 2020 global 30m arable land spatial distribution product as an example, its semantic richness improved by 285%, professionalism (increased terminology) by 320%, usability (clarification of application scenarios) by 250%, and the degree of structuring increased from 0 to 85%.

[0124] Step 3: Unified expression of semantic features by vectorization: The extended ontology semantics and functional semantics are normalized and deduplicated by text, embedded and encoded by language model, and non-textual semantic features are numerically encoded. Multi-source vectors are fused and concatenated into a unified semantic expression vector.

[0125] To address the extended ontology semantics and enhanced functional semantics, a multi-level feature extraction strategy is adopted, including feature encoding and feature fusion, to map multi-dimensional semantic information into a unified vector space, achieving semantic computability, comparability, and inference capability. This step involves three sub-steps: semantic text normalization and deduplication, multi-dimensional feature embedding and encoding, and unified expression through multi-source vector fusion.

[0126] Step 3.1, Semantic text normalization and deduplication, is essentially a text preprocessing process aimed at semantic quality and model input adaptation, which transforms structured data into natural language sentences to facilitate subsequent language model encoding.

[0127] Ontology-extended text normalization: Ontology-extended semantics is based on structured information mapped by knowledge. Through structural reorganization, coherent sentences are formed, such as ["remote sensing", "earth observation", "environmental monitoring", "land use", "arable land"] being reorganized into "This data belongs to remote sensing technology in the field of geography and is applied to agricultural monitoring." Secondly, terminology standardization addresses heterogeneous expressions such as "remote sensing image" and "RS image," and "Cropland" and "arable land." The subsequent Sentence-BERT language model encoding model will control the length of text fields.

[0128] LLM-enhanced text normalization: The semantics generated from multi-round expert dialogues based on the LLM model need further refinement to produce high-quality semantic summaries. Twenty dialogue texts generated from four rounds of dialogues involving five experts were first subjected to temporal summarization (focusing first on problem understanding, then on conclusions), and then cross-expert information fusion was performed using extractive and generative methods to form a comprehensive description.

[0129] Information deduplication involves similarity calculation, clustering sentences with a similarity greater than 0.9, and retaining the central sentence.

[0130] The normalized mathematical expression is:

[0131] (1)

[0132] (2)

[0133] In the formula, This represents a deterministic text processing flow: first rephrase, then standardize, then dedup, and finally length control. Represents semantic text; This indicates extended ontology semantics; It represents functional semantics.

[0134] Step 3.2, Multidimensional Feature Embedding Coding; Multidimensional feature embedding coding mainly refers to the numerical coding of language model-based embedding coding and non-textual semantic features.

[0135] Language model-based embedding encoding is based on the Sentence-BERT multilingual sentence embedding model, which maps text semantics into a fixed-dimensional semantic vector through a pre-trained Transformer encoder.

[0136] Semantic encoding process: ① WordPiece segmentation: Assuming a sentence s is given, it is converted into a token sequence using the WordPiece segmenter:

[0137] (3)

[0138] Where x represents the final token sequence after word segmentation. This refers to the WordPiece tokenizer, which uses a sub-word segmentation algorithm to break down complete words into finer-grained sub-words, while adding special tags to output a structured token sequence. [CLS] indicates the sub-word token; [SEP] is the category marker; [CLS] is the end separator.

[0139] ② Token segmentation and encoding: Input the token sequence into a 12-layer encoder to obtain the context-aware hidden state:

[0140] (4)

[0141] In the formula, It is a lightweight MiniLM Transformer encoder; H represents the hidden state sequence of the encoder output; This represents the hidden state vector corresponding to the label, and each , d=384.

[0142] ③ Sentence Vector Pooling: The default mean pooling method is used to average the output vectors of all non-special tokens to generate sentence embeddings. :

[0143] (5)

[0144] In the formula, is the sentence embedding vector; n is the number of non-special tokens, that is, the total number of tokens participating in the pooling after excluding [CLS] and [SEP]; This represents the hidden state vector corresponding to the i-th non-special token.

[0145] After embedded encoding, a 384-dimensional ontology extended semantic vector is obtained:

[0146] (6)

[0147] In the formula, It is a 384-dimensional BERT embedded encoding function; This indicates extended ontology semantics; This represents the extended ontology semantic vector.

[0148] The functional semantic vector is obtained by encoding and then weighting and fusing the normalized semantic texts from four rounds of expert dialogue to form a 384-dimensional vector.

[0149] (7)

[0150] In the formula, This represents the semantic vector for the corresponding round; Represents the normalized semantic text for the corresponding round; Indicates the number of expert dialogue rounds; Represents a functional semantic vector; It is the weight coefficient for the r-th round, and the weight decreases with each round.

[0151] The numerical encoding of non-textual semantic features involves transforming categorical features, spatiotemporal semantic features, and numerical features into vector representations through structured processing. The categorical feature encoding employs a hybrid encoding method to classify subject categories, data types, and Earth science data into the following five-dimensional vectors:

[0152] (8)

[0153] In the formula, Represents categorical feature encoding; This represents the Level 1 subject classification code; This indicates the level 2 subject classification code; Indicates type encoding; This represents a score indicating the complexity of data formats (such as raster, vector, and tabular) based on experience. This indicates the classification and coding of Earth science data. This represents the subject attribute of the data object d; The data type attribute representing the data object d; This represents the Earth's spheres attribute of the data object d; Represents the coding function for a first-level discipline; Represents the coding function for second-level disciplines; Represents data type mapping functions; This represents a complexity scoring function; This represents the Earth's sphere encoding function.

[0154] Spatiotemporal features include region type features, time type, and spatial resolution type. Semantic encoding based on domain knowledge is employed. For example, the spatial resolution type of satellite remote sensing images is classified into low, low-medium, medium, medium-high, and high types based on scales such as 500m, 100m, 30m, 10m, and 2m; the region type is classified into global, regional, country-specific, and local types; and the time type is classified into continuous, discrete, and instantaneous types. This constructs a 3-dimensional vector, as follows:

[0155] (9)

[0156] In the formula, Represents spatiotemporal feature encoding; Indicates the region type component; Indicates a time-type component; Indicates spatial resolution type components; Indicates the region type encoding function; Indicates a time-type encoding function; Indicates the spatial resolution level encoding function; This indicates the input region type attribute, such as "Global", "Region", "Local Area", etc. Indicates the time type attribute of the input, such as "continuous" or "discrete"; This indicates the input spatial resolution level attribute, such as "low", "medium", "high", etc.

[0157] Numerical feature encoding includes spatial resolution, temporal span, description length, data size, and update frequency. A 10-dimensional vector is constructed using logarithmic normalization, as follows:

[0158] (10)

[0159] In the formula, Represents numerical feature encoding; , ,…, Represents the corresponding component of the normalized vector of numerical features; Indicates horizontal resolution; Indicates vertical resolution; Indicates time resolution; , These represent the start and end times, respectively. Indicates duration; , These represent the word count of the ontology extended semantic summary and the word count of the LLM semantic enhancement description, respectively. Indicates the size of the data; Indicates the update frequency. Represents the logarithmic normalization function; This represents the linear normalization function.

[0160] Multi-source vector fusion unified representation involves concatenating all semantic feature vectors into a unified representation. Taking the 2020 global 30m arable land spatial distribution product as an example, a 786-dimensional vector is generated, as follows:

[0161] (11)

[0162] In the formula, This represents the final unified semantic representation vector; , It is a learnable or preset attention weight matrix.

[0163] Step 4: Generate expression vector based on meta-representation rules; construct rule vectors based on meta-representation rules, parse user demand semantics, and adaptively adjust and weight-optimize the rule parameters of each dimension in the rule vector in combination with the unified semantic expression vector to generate a rule expression vector that matches the user demand semantics.

[0164] Meta-representation rule-based representation vector generation standardizes and makes interpretable the process of vectorizing semantic features by learning and encoding a series of meta-representation rules, and achieves adaptive optimization. Meta-representation rules define how to represent vectors of multi-source heterogeneous semantic features, and how to effectively utilize these vectors in subsequent applications. For example... Figure 4 As shown, the meta-representation rule architecture includes a rule vector module, a rule learning module, and rule-based expression vector generation.

[0165] A rule vector module is a dense vector of fixed dimensions, where each dimension corresponds to a specific representation rule or strategy. Rule vectors allow for the addition of new rule modules, expanding the vector dimensions. In this application, the rule vector is initialized as a fixed 64-dimensional vector, consisting of six categories: similarity calculation rules (10 dimensions), clustering analysis rules (10 dimensions), semantic retrieval rules (10 dimensions), interpretation rules (10 dimensions), generation rules (10 dimensions), and interaction rules (14 dimensions).

[0166] Similarity calculation rules define the methods and parameters for assessing the similarity between semantic expression vectors. These rules include weight allocation sub-rules for different feature types (ontology extension, semantic enhancement, numerical, classification, spatiotemporal), distance metric sub-rules, and similarity score normalization sub-rules. Cluster analysis describes how to perform clustering analysis on semantic expression vectors, including algorithm selection sub-rules, feature selection sub-rules, and cluster quantity heuristic sub-rules. Semantic retrieval rules define vector-based semantic retrieval strategies, including query expansion sub-rules, ranking strategy sub-rules, and result diversification sub-rules. Interpretation rules define how to interpret the meaning of semantic expression vectors, including feature importance sub-rules and interpretability mapping sub-rules. Generation rules define how to generate natural language descriptions or other outputs from semantic expression vectors, including text generation style sub-rules, content selection sub-rules, and degree of structuring sub-rules. Interaction rules structure and encode the attribute information of the Earth science data itself, providing user and system-oriented usage guidance. These rules cover data format and structure attributes, quality and accuracy attributes, recommendation tool and method attributes, and application guidance attributes, thereby achieving efficient and semantically consistent interaction between "data—model—user".

[0167] like Figure 5 As shown, the triggering mechanism for rule vectors is based on semantic parsing of user input, which is then transformed into query vectors. Rule vectors are then assigned through the rule learning module. For example, user semantics are parsed into information such as "arable land area," "2020," and "country x," which are then transformed into query vectors and finally assigned to similarity calculation rules and semantic retrieval rules.

[0168] Different rule vectors are triggered under different conditions, and their triggering conditions overlap based on user semantics. The triggering conditions for similarity calculation rules are mostly related to semantics such as user data querying, system recommendation calculation, relevance analysis, and cross-domain data matching. The triggering conditions for clustering rules mostly involve semantics such as data classification, knowledge organization, and user interest topics. Semantic retrieval rules are mostly for user query processing, intelligent ranking, personalized recommendation calculation, knowledge query, and data retrieval needs. The triggering conditions for interpretation rules involve user interpretability needs, such as "Why can this data be used for farmland area analysis?", and semantics such as feature importance analysis and visualization needs. Generation rules are oriented towards user needs for natural language, such as "Which field does this data involve?" or "What is the summary?", and semantics such as structured output needs and multimodal output scenarios.

[0169] The rule learning module is the core of expression vector generation based on meta-representation rules. It is responsible for automatically learning, optimizing, and applying rule vectors from user semantics and interactions, constructing a customized expression vector that meets user needs. This vector serves as a "meta-controller" to guide the semantic expression and application of Earth science data. It includes four sub-modules: multi-task planning and learning, incremental learning, online adaptation, rule vector generation, and rule learning evaluation. The multi-task planning and learning sub-module parses user semantics and configures multi-task rules, including classification, similarity, and retrieval tasks, thereby adjusting the weights of different feature types for rule vector generation. This allows for the selection of the most suitable task type for different application scenarios and possesses strong generalization capabilities. The incremental learning sub-module learns from the new data in the vector library from step 2, continuously improving the rule vectors. The online adaptation sub-module adjusts rule vectors based on user feedback. By learning explicit / implicit feedback such as user click behavior, ratings, and dwell time, it personalizes adjustments and implements rule adjustment strategies. The rule vector generation sub-module generates rule-based initialization and optimization vectors. The rule-based learning evaluation submodule evaluates multi-task planning metrics, rule quality, and incremental learning to ensure the accuracy and rationality of rule-based expression vector generation.

[0170] The rule representation vector is a mathematical encoding based on meta-representation rules. Essentially, it is a learned, weighted rule vector guided by user needs, used to guide the semantic representation of how to use Earth science data. The formula for constructing the rule representation vector is:

[0171] (12)

[0172] In the formula, The feature representation of the k-th user semantic; Represents relevant semantic vectors or contextual semantic information in the vector library; The supervision category that indicates the semantics of the user query; Representation rule learning model; This represents the task-related loss function; This represents the initial rule vector in the model; This represents the output rule representation vector.

[0173] The core purpose of rule expression vectors based on meta-representation rules is to abstract the decision logic rules in the semantic representation process (such as feature fusion, similarity calculation, retrieval strategy, result interpretation, etc.) into a set of learnable, interpretable, and optimizable 64-dimensional expression vectors. This enables the standardization, adaptation, and context-aware unified encoding of multi-source heterogeneous semantic features of Earth science data, supporting adaptive semantic retrieval for users / large models, personalized knowledge graph construction, cross-domain transfer learning, and interpretable AI decision-making.

[0174] Its significance lies not only in significantly improving the performance and generalization ability of semantic representation vectors in downstream tasks (such as retrieval, clustering, and recommendation), but also in endowing the device with an adaptive mechanism and the ability to continuously evolve—it can automatically learn the optimal representation strategy from historical data, adjust rules in real time based on user feedback, and explicitly model the attributes and usage context of the data itself through structured interaction rules. Ultimately, it constructs a semantic understanding framework that combines high precision, strong interpretability, personalized adaptability, and cross-domain transferability, providing a reliable, trustworthy, and interactive knowledge service foundation for complex application scenarios such as earth science and sustainable development.

[0175] Taking "How much arable land did my country have in 2020?" as an example, this paper simulates the user / large model's question input to reflect the triggering mechanism of rule vectors and the generation process and analysis guidance of rule expression vectors.

[0176] First, complete the user requirements analysis:

[0177] Ontological concepts: arable land, area, China;

[0178] Value: Cultivated land area;

[0179] Classification: Land type = arable land, type = statistics, level = country level;

[0180] Time and space: 2020, the entire territory of China.

[0181] Secondly, the semantics of user needs are transformed into a 786-dimensional vector, and based on its semantic features, the query and calculation processes are triggered to establish semantic retrieval rules, similarity calculation rules, interpretation rules, and interaction rules.

[0182] {

[0183] Semantic retrieval rules:

[0184] Query expansion: Automatically expand query terms

[0185] "Farmland" → "Agricultural land, land use, arable land resources"

[0186] "Our country" → "China, People's Republic of China, nationwide"

[0187] "Area" → "Distribution area, statistical data, spatial distribution"

[0188] Sorting strategy: Prioritization is as follows: Time relevance (2020) > Spatial relevance (China) > Data quality

[0189] Results diversification: Ensure that returned results cover time, space, and category.

[0190] Similarity calculation rules: Emphasizing the ontological characteristics of "arable land", the temporal characteristics of "2020", and the spatial characteristics of "my country".

[0191] Weighting: Increase the weight of spatiotemporal features (time (0.90), space China (0.75))

[0192] Distance metric: using weighted cosine similarity, with spatiotemporal feature weights doubled.

[0193] Interaction rules: Emphasize the use of a unified vector for related data.

[0194] Format and Structure Properties: Activate format information

[0195] Quality and accuracy attributes: clearly define the coding accuracy requirements;

[0196] Tools and Methods Attributes: Recommendation Processing Tools and Methods Based on Format Information

[0197] Application guidance attribute: Activation application scenario guidance

[0198] }

[0199] Next, the rule learning module performs initial semantic retrieval in the vector library through multi-task programming learning, and returns the results in priority order. User feedback is then recorded, and the rule vectors are fine-tuned using the online adaptation submodule. Specifically, the vector library is incrementally learned by adding semantic information from Earth science data, while simultaneously learning from accumulated user feedback. Finally, rule-based representation vectors are generated.

[0200] The last 14 dimensions (50-63) of the representation vector are encoded using the interaction rule encoding:

[0201] {

[0202] Dimensions 50-53: Format and Structural Attributes

[0203] Dimension 50 (data format type): 0.55 → indicates raster data (GeoTIFF)

[0204] Dimension 51 (Dimensional Organization): 0.10 → One-dimensional / numerical data

[0205] Dimension 52 (spatial reference): 0.90 → Indicates the use of the WGS84 coordinate system.

[0206] Dimension 53 (Time Characteristics): 0.20 → Year / Single Time Point

[0207] Dimensions 54-56: Quality and Precision Attributes

[0208] Dimension 54 (spatial resolution): 0.55 → 30m / medium resolution

[0209] Dimension 55 (time resolution): 0.2 → year

[0210] Dimension 56 (Quality Level): 0.85 → Indicates officially verified data.

[0211] Dimensions 57-60: Recommended Tools and Methods

[0212] Dimension 57 (Recommended processing tool): 0.70 → Recommended to use Python / ArcGIS / QGIS

[0213] Dimension 58 (Core Processing Method): 0.55 → Spatial statistical methods are recommended.

[0214] Dimension 59 (Visualization Method): 0.75 → Thematic Map

[0215] Dimension 60 (Data Integration Method): 0.25 → Standalone Use

[0216] Dimensions 61-63: Application Guidance Attributes

[0217] Dimension 61 (Main Application Areas): 0.2 → Agricultural Applications

[0218] Dimension 62 (Typical Output Type): 0.25 → Statistical Reports

[0219] Dimension 63 (using complexity guidelines): 0.60 → Medium complexity, requires basic GIS skills.

[0220] }

[0221] External methods and tools for analyzing interaction rule vectors based on user needs semantics:

[0222] {

[0223] Data processing operation guide:

[0224] • Data format: The data is in raster / image format. It is recommended to use the GDAL library or Rasterio library to read it.

[0225] • Spatial reference: The data uses the WGS84 coordinate system and can be overlaid with other geographic data for analysis.

[0226] • Processing tools: It is recommended to use ArcGIS, QGIS, or Python's GDAL library for spatial analysis.

[0227] Specific operation example:

[0228] • For the query 'How much arable land did my country have in 2020?', please follow these steps:

[0229] Step 1: Use GDAL or Rasterio to read the TIFF data of the 2020 Global Cultivated Land Spatial Distribution Product (GCL30_2020).

[0230] Step 2: Use China's administrative boundary vector data to crop out the cultivated land data for the Chinese region.

[0231] Step 3: Count the number of cultivated land pixels in the cropped raster and calculate the total area based on the spatial resolution.

[0232] Step 4: Output the calculation results in the form of statistical reports or thematic maps.

[0233] Code example (Python):

[0234] Python

[0235] import rasterio

[0236] from rasterio.mask import mask

[0237] import geopandas as gpd

[0238] import numpy as np

[0239] # 1. Read farmland TIFF data

[0240] with rasterio.open('GCL30_2020.tif') as src:

[0241] Farmland data = src.read(1)

[0242] profile = src.profile

[0243] # 2. Read China's borders

[0244] China's border = gpd.read_file('china_boundary.shp')

[0245] #3. Crop the China region

[0246] with rasterio.open('GCL30_2020.tif') as src:

[0247] Chinese arable land, transform = mask(src, China's border.geometry, crop=True)

[0248] # 4. Statistics on cultivated land area

[0249] Pixel size = transform[0] # Spatial resolution

[0250] Number of arable land pixels = np.sum("China arable land == 1") # Assume 1 represents arable land

[0251] Total area = Number of arable land pixels Pixel size Pixel size

[0252] print(f'The area of ​​arable land in China in 2020 is: {total area: .2f} square meters')

[0253] ```

[0254] }

[0255] In summary, this application's embodiments achieve a leap from static descriptive semantics to dynamic functional semantics in Earth science data through a full-chain technical solution encompassing ontology semantic preprocessing and expansion, LLM-based semantic enhancement, unified multi-dimensional feature vectorization, and expression vector generation based on meta-representation rules. This method not only retains the high-dimensional computational efficiency and large-model compatibility of vectorized expression but also embeds business semantics and interpretable decision logic into learnable rule vectors, endowing the system with the ability to adaptively adjust expression strategies according to user needs. By deeply integrating inherent data features with user-demand semantics, this application's embodiments generate expression vectors oriented towards diverse application scenarios, providing semantic understanding and knowledge service support with high accuracy, strong interpretability, and personalized adaptability for complex applications such as interdisciplinary data discovery, SDG indicator-supported analysis, and computationally ready data recommendation. The apparatus, equipment, and storage media provided in this application's embodiments also possess the aforementioned beneficial effects.

[0256] On the other hand, this invention provides a semantic representation device for Earth science data based on meta-representation rules, whose various modules can implement the various steps of the aforementioned method, specifically including:

[0257] The ontology semantic preprocessing module is used to load and verify the feature information metadata of Earth science data, process feature information and separate feature semantics, and perform structured semantic extension on the separated ontology semantics based on domain knowledge mapping to obtain extended ontology semantics.

[0258] The semantic enhancement module based on the LLM model is used to perform multi-round simulated expert dialogue on the extended ontology semantics using a large language model in order to achieve functional semantic enhancement and obtain functional semantics.

[0259] The multidimensional feature vectorization module is used to perform text normalization and deduplication on the extended ontology semantics and functional semantics, embedding and encoding them through a language model, and numerically encoding non-textual semantic features, and fusing and splicing multi-source vectors into a unified semantic expression vector.

[0260] The expression vector generation module is used to construct rule vectors based on meta-representation rules, parse user demand semantics, and adaptively adjust and weight the rule parameters of each dimension in the rule vector in combination with the unified semantic expression vector to generate a rule expression vector that is adapted to the user demand semantics.

[0261] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for semantic representation of Earth science data based on meta-representation rules.

[0262] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for semantic representation of Earth science data based on meta-representation rules.

[0263] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A semantic representation method for Earth science data based on meta-representation rules, characterized in that, The method includes: Step 1: Load and verify the feature information metadata of the Earth science data, process the feature information and separate the feature semantics, and perform structured semantic extension on the separated ontology semantics based on domain knowledge mapping to obtain extended ontology semantics; Step 2: Use a large language model to simulate expert dialogues in multiple rounds on the extended ontology semantics to achieve functional semantic enhancement and obtain functional semantics; Step 3: Perform text normalization and deduplication on the extended ontology semantics and functional semantics, embed and encode them through a language model, and numerically encode the non-textual semantic features. Then, fuse and concatenate the multi-source vectors into a unified semantic expression vector. Step 4: Construct a rule vector based on the meta-representation rules, parse the semantics of user needs, and adaptively adjust and weight the rule parameters of each dimension in the rule vector in combination with the unified semantic expression vector to generate a rule expression vector that is adapted to the semantics of user needs.

2. The semantic representation method for Earth science data based on meta-representation rules according to claim 1, characterized in that, The feature information processing in step 1 includes text segmentation, category tag extraction, and spatiotemporal information parsing: the text segmentation breaks down unstructured text fields into semantic units and counts their frequencies; the category tag extraction extracts all categories from the information and assigns weights according to priority; the spatiotemporal information parsing performs unit standardization transformation and content parsing on the spatiotemporal fields, uniformly expressing them in numerical form; and the feature semantic separation divides the data information into three types of semantic features: text features, category features, and spatiotemporal features.

3. The semantic representation method for Earth science data based on meta-representation rules according to claim 1, characterized in that, The ontology semantic extension based on domain knowledge mapping in step 1 includes semantic enrichment, structural normalization, and knowledge association: by extracting numerical values ​​and units, the spatial resolution field is extended into structured information containing numerical values, units, type levels, and application applicability; The time range field is expanded into structured information including start and end times, duration in days, and time type by calculating the time span; the keyword field is expanded into structured information including subject terms, technical terms, related keywords, and subject classification links by word segmentation and classification.

4. The semantic representation method for Earth science data based on meta-representation rules according to claim 1, characterized in that, Step 2, which utilizes a large language model to conduct multi-round simulated expert dialogues, includes: configuring multiple expert roles across various professional fields and designing prompts to conduct multi-round dialogues from the dimensions of scientific value, application scenarios, technical characteristics, data quality, relevance, traceability, and usage guidelines; extracting structured information from the dialogue results to obtain functional semantic enhancement results in the dimensions of core concepts, relationships, application classifications, quality indicators, and technical terms; and quantitatively evaluating the quality of functional semantics through consistency checks, professional terminology verification, and usability assessment.

5. The semantic representation method for Earth science data based on meta-representation rules according to claim 1, characterized in that, Step 3, which involves text normalization and deduplication of extended ontology semantics and functional semantics, includes: structural reorganization of extended ontology semantics to form coherent sentences and terminology standardization; fusion of temporal summarization and cross-expert information to form a comprehensive description of enhanced semantics of the large language model; and clustering of semantic texts by similarity calculation and retention of central sentences to achieve information deduplication.

6. The semantic representation method for Earth science data based on meta-representation rules according to claim 1, characterized in that, Step 3, which involves numerically encoding non-textual semantic features, includes: generating dimensional vectors for subject classification, data type, and earth science data classification using a hybrid encoding method for categorical features; generating dimensional vectors for region type, time type, and spatial resolution type for spatiotemporal features based on domain knowledge using semantic encoding; and generating dimensional vectors for spatial resolution, time span, description length, data size, and update frequency using a logarithmic normalization method for numerical features.

7. The semantic representation method for Earth science data based on meta-representation rules according to claim 1, characterized in that, Step 4, which involves constructing a rule vector based on meta-representation rules, includes: predefining a fixed-dimensional rule vector consisting of six categories: similarity calculation rules, clustering analysis rules, semantic retrieval rules, interpretation rules, generation rules, and interaction rules. Each dimension corresponds to a specific representation rule or strategy. The interaction rules encode data format and structure attributes, quality and accuracy attributes, recommendation tool and method attributes, and application guidance attributes. These attributes are used to provide data processing operation guidance, tool recommendations, and code examples, thereby achieving semantically consistent interaction between data, model, and user.

8. A semantic representation device for Earth science data based on meta-representation rules, characterized in that, include: The ontology semantic preprocessing module is used to load and verify the feature information metadata of Earth science data, process feature information and separate feature semantics, and perform structured semantic extension on the separated ontology semantics based on domain knowledge mapping to obtain extended ontology semantics. The semantic enhancement module based on the LLM model is used to perform multi-turn simulated expert dialogue on the extended ontology semantics using a large language model in order to achieve functional semantic enhancement and obtain functional semantics. The multidimensional feature vectorization module is used to perform text normalization and deduplication on the extended ontology semantics and functional semantics, embedding and encoding them through a language model, and numerically encoding non-textual semantic features, and fusing and splicing multi-source vectors into a unified semantic expression vector. The expression vector generation module is used to construct rule vectors based on meta-representation rules, parse user demand semantics, and adaptively adjust and weight the rule parameters of each dimension in the rule vector in combination with the unified semantic expression vector to generate a rule expression vector that is adapted to the user demand semantics.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the semantic representation method for Earth science data based on meta-representation rules as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the semantic representation method for Earth science data based on meta-representation rules as described in any one of claims 1-7.

Images