Offshore ecological environment health assessment method based on inference knowledge graph

Abstract

The invention discloses an inference knowledge graph-based offshore ecological environment health assessment method, which comprises the following steps of: acquiring water body parameters in real time, acquiring large-range monitoring information through satellite remote sensing data, and accessing literature data to construct a background knowledge base; defining a core entity type, establishing a relationship type between entities, storing the entities and the relationships through a graph database, combining the graph database with a reasoning model, and performing node classification, link prediction and graph representation learning on the entities and the relationships in the graph database; updating the state of the knowledge graph according to the detection data received in real time; and a health threshold is set in combination with historical data and regional features, health state registration and change trend analysis are generated, and an evaluation result and a reasoning process are visually displayed. The knowledge graph and ecological environment health assessment are combined, an intelligent decision support tool is provided for offshore ecological environment management, and the modeling and reasoning capabilities of the dynamic evolution process of the offshore ecological environment are improved.

Description

Technical Field

[0001] This invention relates to the field of marine ecological environment monitoring and intelligent assessment technology, and in particular to a method for assessing the health of nearshore ecological environment based on reasoning knowledge graphs. Background Technology

[0002] Under the dual pressures of human activities and global change, the ecological environment of my country's coastal waters continues to deteriorate. Pollutant discharge, coastal engineering and offshore operations, overfishing, and invasive alien species have caused numerous problems, including the degradation of marine water environments, destruction of marine habitats, depletion of marine biological resources, and frequent marine ecological disasters. Ecosystem health refers to the ability of an ecosystem to maintain a normal state over a long period, preserving the diversity, complexity, and vitality of its overall functions, ensuring stability and sustainable development, and rapidly recovering from external stresses. Currently, the main methods for assessing marine ecosystem health include the indicator species method and the indicator system method. The indicator species method is simple and easy to implement, but its indicators are singular, making it difficult to reflect complex ecosystems, and the species selection criteria are unclear, lacking universality. The indicator system method can comprehensively reflect the quality of the ecosystem and is currently the most commonly used evaluation method for marine ecosystem health both domestically and internationally. However, it involves a large amount of text and data that needs to be correlated with professional expertise, making it quite challenging.

[0003] The industry currently faces the following problems: First, marine environmental data comes from a wide range of sources, including remote sensing monitoring, sensor networks, historical documents, and expert knowledge. These data are heterogeneous in format and inconsistent in standards, making effective integration and utilization difficult. Second, traditional assessment systems are mostly based on fixed indicator systems, lacking the ability to model and reason about the dynamic evolution of the ecological environment, and thus unable to adapt to the spatiotemporal changes in the marine environment. Summary of the Invention

[0004] To overcome the aforementioned problems in the existing technology, this invention proposes a nearshore ecological environment health assessment method based on reasoning knowledge graphs.

[0005] The technical solution adopted by this invention to solve its technical problem is: a method for assessing the health of nearshore ecological environment based on reasoning knowledge graph, comprising the following steps: Step 1, Multi-source data acquisition: Real-time acquisition of water parameters, acquisition of large-scale monitoring information through satellite remote sensing data, and construction of a background knowledge base by accessing literature; Step 2, Knowledge Graph Construction: Define core entity types, establish relationship types between entities, store entities and relationships through a graph database, combine the graph database with the reasoning model, and perform node classification, link prediction, and graph representation learning on the entities and relationships in the graph database; Step 3: Update the state of the knowledge graph obtained in Step 2 based on the detection data received in real time in Step 1; Step 4, Health Status Assessment: Construct an ecological environment health assessment system that includes both the ecological environment itself and the impact of human activities. Set health thresholds by combining historical data and regional characteristics, generate a health status registration and trend analysis, and visualize the assessment results and reasoning process.

[0006] The aforementioned method for assessing the health of nearshore ecological environment based on reasoning knowledge graphs includes, in step 2, core entity types including marine area nodes, monitoring station nodes, water quality parameter nodes, biological population nodes, ecological process nodes, and time nodes.

[0007] The above-mentioned method for assessing the health of nearshore ecological environment based on reasoning knowledge graphs defines core entity types in step 2, including named entity recognition and relation extraction. The named entity recognition uses a natural language processing model to automatically extract text. The relation extraction specifically involves: for structured and semi-structured data, using manual processing and filtering, transforming the data into triples of <entity, relation, entity> or <entity, attribute, attribute value>; For unstructured data, the required entities and relationships are extracted using a natural language processing model, transforming unstructured text into structured knowledge, and then processed using methods for structured and semi-structured data.

[0008] The aforementioned method for assessing the health of nearshore ecological environment based on reasoning knowledge graphs employs a transformer architecture in its natural language processing model. This model extracts rules from keywords and short texts, determines the distance between similar words, uses relevance scores for weighting, sums all word vectors, and matches them with regularized templates from the knowledge graph's ternary components. Based on the structured representation of the scheme, it completes the association between objects and their attributes in the rule graph and the object relationship graph.

[0009] The aforementioned method for assessing the health of nearshore ecological environments based on reasoning knowledge graphs determines the distance between similar words by calculating the syntactic distance between two adjacent word vectors in a code sequence. The specific measurement function is as follows: in, and A distance function used to calculate the intermediate distance between adjacent points, code sequence. As input to the model, the function g(.) converts each input word into a corresponding vector representation; h represents the size of the hidden embedding, and n represents the length of the code sequence; the value of the measurement function ranges from [-1, 1], the closer the value is to 1, the more similar the two vectors are; the closer the value is to -1, the less similar the two vectors are.

[0010] In the aforementioned method for assessing the health of nearshore ecological environment based on reasoning knowledge graphs, the reasoning model in step 2 is the Transformer model.

[0011] The beneficial effects of this invention are as follows: Based on a knowledge graph construction method that integrates multi-source data fusion and semantic enhancement, this invention fuses multi-source data such as remote sensing imagery, sensor monitoring, historical documents, and expert reports. It employs a Transformer model to extract ecological entities (such as phytoplankton, dissolved oxygen, eutrophication zones, etc.) and their relationships. A semantic enhancement mechanism is introduced, combining Geographic Information System (GIS) data with spatiotemporal attributes to construct a dynamic knowledge graph with spatiotemporal dimensions. Based on the knowledge graph, a graph neural network reasoning model is constructed to learn the potential influence relationships between nodes. This invention innovatively combines knowledge graph technology with ecological environment health assessment, providing an intelligent decision support tool for nearshore ecological environment management. It improves the modeling and reasoning capabilities for the dynamic evolution of the nearshore ecological environment and can adapt to the spatiotemporal changes in the marine environment. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the process of this invention; Figure 2 This is a schematic diagram illustrating the construction and application of the knowledge graph of this invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0014] like Figure 1 As shown in the figure, this embodiment discloses a method for assessing the health of nearshore ecological environment based on reasoning knowledge graph, which specifically includes the following steps: Step 1, Multi-source data acquisition: Real-time acquisition of water parameters, acquisition of large-scale monitoring information through satellite remote sensing data, and construction of a background knowledge base by accessing literature.

[0015] Deploy a marine environmental monitoring sensor network to collect parameters such as water temperature, salinity, dissolved oxygen, and chlorophyll in real time. Integrate satellite remote sensing data to obtain large-scale monitoring information such as sea surface temperature, chlorophyll concentration, and suspended matter distribution. Access historical scientific expedition data, literature, laws and regulations, and expert experience to build a background knowledge base.

[0016] We used CiteSpace to perform co-occurrence analysis on keywords in relevant literature and extracted keywords to serve as keywords for the Transformer model.

[0017] Step 2, Knowledge Graph Construction: such as Figure 2As shown, core entity types are defined, relationship types between entities are established, entities and relationships are stored in a graph database, and the graph database is combined with the inference model to perform node classification, link prediction, and graph representation learning on entities and relationships in the graph database.

[0018] The Neo4j graph database is used to store entities and relationships. Core entity types are defined: sea area, monitoring station, water quality parameters, biological population, ecological process, etc. Relationship types between entities are established: spatial affiliation, temporal evolution, causal influence, correlation, etc.

[0019] Step 2.1, define the core entity types, including: Sea area node: Sea area ID, name, geographical location (polygon coordinates), area, description, etc.; Monitoring site nodes: site ID, name, latitude and longitude, sea area, deployment time, etc.; Water quality parameter nodes: parameter ID, name, unit, standard range (maximum value, minimum value), etc.; Biological population nodes: species ID, species name, Latin name, functional group, whether it is a key species, whether it is an indicator species, etc.; Ecological process nodes: process ID, name, description (e.g., eutrophication, red tide, hypoxia, etc.); Time nodes: Time ID, timestamp (year, month, day, hour, minute, second), year, month, season, etc.

[0020] The process of ontology construction is the process of defining the specifications of concepts and relationships between concepts within a certain domain.

[0021] Knowledge extraction consists of two steps: named entity recognition and relation extraction. Named entity recognition uses the Transformer model as a deep learning model, which can automatically extract features from the text. The self-attention mechanism of the Transformer architecture allows for faster training of the corpus.

[0022] For structured data, manual processing and filtering transform the data into triples of <entity, relation, entity> or <entity, attribute, attribute value>. Information extraction from structured and semi-structured data sources is relatively simple, and there is less data noise. After manual filtering, high-quality structured triple data can be obtained.

[0023] For unstructured text, knowledge extraction techniques are used to extract the implicit knowledge within the text, obtaining the attributes and values ​​of relevant entities, as well as the relationships between entities. For unstructured data, natural language processing techniques are needed to extract the required entities and relationships from large amounts of text, transforming the unstructured text into structured knowledge. In the above formula, O (Ontology) represents the ontology, C (Class / Concept) represents the set of classes or concepts, R (Relation) represents the set of relations, F (Function) represents the set of functions, and I (Instance) represents the set of instances. The knowledge in the knowledge graph is described using triples, denoted as... .in Represents a collection of entities, which contains Different entities; A set representing relations, containing Different kinds of relationships; This represents a set of triples, recording each piece of knowledge in the knowledge base. Entities are the basic elements in a knowledge graph; each entity is unique and can be represented as <Yellow Sea nearshore, coordinates, latitude and longitude parameters>.

[0024] First, an ontology model for the domain needs to be defined. An ontology editor is used to build the schema layer from top to bottom, and then data is added to the concepts. In this type of construction method, the ontology is the schema layer of the knowledge graph, which can be used to describe the upper-level conceptual schemas of the knowledge graph.

[0025] In natural language processing, the Transformer architecture is adopted, and entities, attributes, and relationships are extracted using preset rules. Based on the needs of the scenario, rule elements need to be further expanded on the object relationship graph to form a scenario-specific rule graph. Specifically, this is achieved through rule extraction using keywords and short texts, using word vectors to determine the distance between similar words, and then matching the regular expression templates with the knowledge graph's ternary component (entity 1 relationship -- entity 2). Based on the structured representation of the scenario, the association between the rule graph and the object relationship graph of objects and their attributes is completed, enabling interconnection and interoperability between different knowledge bases and between the data foundation and the knowledge platform in business applications.

[0026] The first step is to calculate the relevance score between the vectors for "nearshore," "ecology," and "health" and every other word in the sentence. This is the "attention score." It's simply measured by the dot product of two word vectors. The function used to calculate the syntactic distance between two adjacent word vectors in the code sequence is as follows: and A distance function used to calculate the intermediate distance between adjacent points, code sequence. As input to the model, the function g(.) converts each input word into a corresponding vector representation. ℎ represents the latent embedding size, and n represents the length of the code sequence. The cosine distance function ranges from -1 to 1; the closer the value is to 1, the more similar the two vectors are; the closer the value is to -1, the less similar the two vectors are. When the angle between two vectors is 0, their cosine distance is 1, indicating that they are completely identical.

[0027] Step 2 uses relevance scores for weighted summation of all word vectors in the sentence. Word vectors closely related to "nearshore," "ecology," and "health" contribute more to the summation (including the words themselves), while irrelevant words contribute almost nothing. The resulting vectors are new representations of "nearshore," "ecology," and "health," incorporating the context.

[0028] Repeating this process for each word vector in the sentence will produce a new sequence of vectors that encode the sentence. The meaning of this expression is: "For each word in inputs(A), calculate the relevance of that word to each word in inputs(B), and then use these scores to perform a weighted sum of the words in inputs(C)." A, B, and C are not necessarily the same unstructured input sequence.

[0029] The encoder part of the Transformer takes index data as input, with input features being evaluation index data for various ecosystems, such as aquatic environment, habitat, and biological community indicators for coral reef ecosystems. The input data should be standardized to eliminate the influence of dimensions. The importance of each index is learned through a self-attention mechanism, and then a more regularized structured data is output through a fully connected layer. Through the self-attention mechanism, the model can learn the interrelationships between indicators and the degree of impact of each indicator on ecological health, thereby assigning different attention weights. The calculation of the self-attention weights is as follows: Where Q, K, and V represent the query, key, and value matrices, respectively, which are obtained from the input index data through linear transformation. The dimension of the key vector is used. Utilizing the Transformer's Self-Attention mechanism, time-series modeling is performed on ecological indicators monitored continuously over many years to capture long-term trends and abrupt events. For example, time-series inputs are constructed for indicators such as "seagrass density" and "coverage" to focus on future health status and identify the impact and recovery processes of events such as red tides and typhoons on the ecosystem. Simultaneously, the indicator system and predicted range are used as labels to continuously optimize and train the model.

[0030] The pre-defined rules for assessing marine ecological health: GB / T42631-2023 "Guidelines for Assessment of Nearshore Marine Ecological Health" provides the classification and assessment methods for nearshore marine ecological health.

[0031] If a standard-based evaluation fails to meet certain practical needs, it may be due to insufficient monitoring points or missing data. Based on this, a personalized evaluation system should be constructed. The specific evaluation indicators remain unchanged, but are standardized using measured and monitored values. Then, the weights of the indicators and the classification of evaluation levels are determined based on expert experience.

[0032] For indicators where higher values ​​are generally more beneficial to marine environmental health (such as pH and dissolved oxygen), the normalization formula is: (1) in, Let i be the normalized value of the index. The actual monitoring value of indicator i Let i be the baseline value of index i, and when the calculated When the value is greater than 1, its value is 1; For indicators where lower values ​​are generally more beneficial to marine environmental health (such as eutrophication index, pollutants, etc.), the normalization formula is: (2) in, Let j be the normalized value of index j. The actual monitoring value of indicator j Let j be the benchmark value of index j, and when the calculated When it is greater than 1, its value is 1.

[0033] The pre-defined human environment health assessment includes economic output and cultural value. Economic output is represented by the gross marine product of the coastal area (unit: RMB 10,000 / km2, normalized using formula (1)). Cultural value is represented by the specific evaluation indicators of the protected sea area (unit: %, normalized using formula (2)) and the number of marine tourists (unit: person-times / year, normalized using formula (1)). The actual monitoring values ​​in formulas (1) and (2) are... , This is the statistical value for that year.

[0034] All indicators are also normalized. Typically, a region's gross marine product increases year by year; therefore, the region's established annual target is selected as the normalization benchmark. If no such target exists, the maximum value over the past five years is used as the normalization benchmark (my country's economic and social development is usually measured in five-year periods). The normalization benchmark for cultural value indicators is also the maximum value over the past five years. The specific normalization method follows the pre-defined marine ecological health assessment method.

[0035] The acquired knowledge set in triplet form is imported into the ontology model, and its logical hierarchy is examined and corrected. Finally, the resulting knowledge graph is queried and verified.

[0036] Step 2.2: Define the relationship types, which include: spatial relationships (subordinate, adjacent, and separate), monitoring relationships (monitoring and being monitored, primary monitoring stations, and some secondary monitoring stations), and temporal relationships.

[0037] Step 2.3, Knowledge Storage: The acquired triples are imported into the Neo4j graph database, and a visual query function for nearshore ecological knowledge is implemented based on Neo4j's Cypher query language. This includes node types such as "Sea Area," "Water Quality Parameters," "Biological Population," and "Pollution Source." The goal of Cypher queries is to calculate a multi-dimensional health index for a specified sea area over a specific time period. This includes, for example, assessing the compliance of water quality parameters, calculating the weighted average of scores for each parameter, and comprehensively evaluating the overall health index.

[0038] (1) Basic node matching: MATCH(sa:SeaArea{area_id:"SA001"}) # Finds the specific sea area with ID "SA001". MATCH(sa)<-[:LOCATED_IN]-(ms:MonitoringStation)# Find all monitoring stations located in this sea area (2) Data flow: Marine Node (SA001) ← Monitoring Station Nodes (MS001, MS002, ...) (3) Monitoring data correlation MATCH(ms)-[measurement:HAS_MEASUREMENT]-(wqp:WaterQualityParameter)# Connects the measurement relationship between the monitoring station and the water quality parameter. MATCH(measurement)-[:RECORDED_AT]-(ts:TimeSlice)# Associates the time information of the measurement data (4) Data Stream Extension Sea area ← Monitoring station ← Measurement relationship → Water quality parameters ↓ Time slice (5) Time filtering WHERE ts.start_time>=datetime('2023-06-01') # Filter out monitoring data from June 1, 2023 onwards to ensure the evaluation is based on the latest valid data. (6) Parameter score calculation WITH sa,wqp,avg(measurement.value)asavg_value, CASE WHEN avg_value>=wqp.standard_min AND avg_value<=wqp.standard_max THEN1.0 ELSE 0.3 END as param_score #Score 1.0 (parameter value within standard range), Score 0.3 (parameter value outside standard range).

[0039] (7) Detailed parameter evaluation: collect({parameter: wqp.name,score: param_score,value: avg_value})asparameters Use the `collect()` function to aggregate multiple rows of results into an array, where each element contains: parameter: parameter name (e.g., "dissolved oxygen") score: The score for this parameter (1.0 or 0.3) value: Parameter average (actual monitored value).

[0040] (8) Health Index Calculation: avg(param_score)as health_index The average score of all parameters is used to reflect the overall health status (range of 0.3-1.0).

[0041] In this embodiment, the health index is calculated using the monitoring data in Table 1 as an example. The specific calculation process is as follows: (1) Calculation of the average value of parameters: Dissolved oxygen: (6.8 + 4.2) / 2 = 5.5 mg / L Nutrient salts: 0.8 mg / L (2) Parameter score calculation: Dissolved oxygen: 5.5 ∈ [5.0, 8.0] → score 1.0; Nutrients: 0.8 ∉ [0.1, 0.5] → Score 0.3 (3) Health index calculation: (1.0 + 0.3) / 2 = 0.65.

[0042] Table 1 Monitoring Data In addition, for areas without monitoring stations, spatial gradient interpolation can be performed. Since recent data has a higher weight, the weight of recent true values ​​(measured values) can be increased. For example, in a coral reef ecosystem, if a region experiences high water temperature events for several consecutive years, the attention weight of the "water temperature" indicator should be automatically increased.

[0043] Step 2.4, Integrating with the Inference Model: Combining the graph database with the Transformer: Export graph data (nodes and relationships) from Neo4j as input to the Transformer. Use the Transformer model for node classification, link prediction, or graph representation learning. Save the Transformer output (such as node embeddings and prediction results) back to Neo4j for subsequent queries and visualizations.

[0044] The Transformer model training process involves collecting marine ecological environment monitoring data from the past five years as the training set. The Transformer model is then used to learn the spatiotemporal evolution patterns of ecological elements.

[0045] Step 3: Update the state of the knowledge graph obtained in Step 2 based on the detection data received in real time in Step 1.

[0046] Step 4, Health Status Assessment: Set health thresholds by combining historical data and regional characteristics, generate health status registration and trend analysis, and visualize the assessment results and reasoning process.

[0047] This embodiment is based on a knowledge graph construction method that integrates multi-source data fusion and semantic enhancement. It fuses data from remote sensing images, sensor monitoring, historical documents, and expert reports, and uses a Transformer model to extract ecological entities (such as phytoplankton, dissolved oxygen, and eutrophication zones) and their relationships. A semantic enhancement mechanism is introduced, combining Geographic Information System (GIS) data with spatiotemporal attributes to construct a dynamic knowledge graph with spatiotemporal dimensions. Based on this knowledge graph, a graph neural network inference model is built to learn the potential influence relationships between nodes.

[0048] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its scope and spirit, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.

Claims

1. A method for assessing the health of nearshore ecological environment based on reasoning knowledge graphs, characterized in that, Includes the following steps: Step 1, Multi-source data acquisition: Real-time acquisition of water parameters, acquisition of large-scale monitoring information through satellite remote sensing data, and construction of a background knowledge base by accessing literature; Step 2, Knowledge Graph Construction: Define core entity types, establish relationship types between entities, store entities and relationships through a graph database, combine the graph database with the reasoning model, and perform node classification, link prediction, and graph representation learning on the entities and relationships in the graph database; Step 3: Update the state of the knowledge graph obtained in Step 2 based on the detection data received in real time in Step 1; Step 4, Health Status Assessment: Construct an ecological environment health assessment system that includes both the ecological environment itself and the impact of human activities. Set health thresholds by combining historical data and regional characteristics, generate a health status registration and trend analysis, and visualize the assessment results and reasoning process.

2. The nearshore ecological environment health assessment method based on reasoning knowledge graph as described in claim 1, characterized in that, The core entity types in step 2 include marine nodes, monitoring station nodes, water quality parameter nodes, biological population nodes, ecological process nodes, and time nodes.

3. The nearshore ecological environment health assessment method based on reasoning knowledge graph according to claim 2, characterized in that, In step 2, the core entity types are defined as named entity recognition and relation extraction. The named entity recognition uses a natural language processing model to automatically extract text. The relation extraction specifically involves: for structured and semi-structured data, using manual processing and filtering, transforming the data into triples of <entity, relation, entity> or <entity, attribute, attribute value>; For unstructured data, the required entities and relationships are extracted using a natural language processing model, transforming unstructured text into structured knowledge, and then processed using methods for structured and semi-structured data.

4. The nearshore ecological environment health assessment method based on reasoning knowledge graph as described in claim 3, characterized in that, The natural language processing model adopts a transformer architecture, extracts rules from keywords and short texts, determines the distance between similar words, uses relevance scores for weighting, sums all word vectors, and matches them with regularized templates from the knowledge graph's ternary components. Based on the structured representation of the scheme, the association between objects and their attributes in the rule graph and the object relationship graph is completed.

5. The nearshore ecological environment health assessment method based on reasoning knowledge graph according to claim 4, characterized in that, Determining the distance between similar words is achieved by calculating the syntactic distance between two adjacent word vectors in the code sequence. The specific measurement function is as follows: in, and A distance function used to calculate the intermediate distance between adjacent points, code sequence. As input to the model, the function g(.) converts each input word into a corresponding vector representation; h represents the size of the hidden embedding, and n represents the length of the code sequence; the value of the measurement function ranges from [-1, 1], the closer the value is to 1, the more similar the two vectors are; the closer the value is to -1, the less similar the two vectors are.

6. The nearshore ecological environment health assessment method based on reasoning knowledge graph according to claim 1, characterized in that, The inference model in step 2 is the Transformer model.

Images