A simplified method and system for assessing the resilience of aquatic ecosystems

Abstract

This invention discloses a method and system for simplifying indicators for assessing the resilience of aquatic ecosystems. The method generates a multimodal retrieval strategy and obtains diverse text by parsing the assessment scenario instructions using thought chain analysis. It then extracts indicator relationships from the text, analyzes the coupling relationships between indicators, constructs a knowledge graph of aquatic ecological indicators, and forms a pool of all candidate indicators. Semantic and correlation analyses are used to calculate the correlation between indicators, eliminating highly correlated redundant indicators to form a preliminary simplified indicator set. The preliminary simplified indicator set and selection criteria are output. After verification, the final simplified indicators are output. This invention solves the problems of complex indicator systems, high collinearity and redundancy, and strong subjectivity in traditional aquatic ecological assessments by integrating the logical constraints of knowledge graphs with the semantic understanding capabilities of large-scale models. It achieves scientific dimensionality reduction and dynamic optimization of assessment indicators, possessing simplicity, interpretability, and engineering application value.

Description

Technical Field

[0001] This invention relates to the field of aquatic ecological environment assessment technology, and in particular to a simplified method and system for assessing the resilience of aquatic ecological environments. Background Technology

[0002] Aquatic ecological environment resilience refers to the ability of aquatic ecosystems to maintain structural and functional stability (adaptive), rapidly recover from damage (self-repair), and adapt to new environments (adaptive) when subjected to external disturbances (such as pollution discharge, hydrological fluctuations, and climate change). With the development of environmental monitoring technology, the number of indicators available for assessing aquatic ecological environments has exploded, covering multiple dimensions including physical, chemical, biological, hydrological, and habitat aspects. In practical applications, selecting the most representative, independent, and cost-effective key indicators from a vast pool of candidate indicators has become a challenge in this field.

[0003] Currently, existing water environment assessment technologies mainly suffer from the following development status and limitations: For example, Chinese patent CN119106818A discloses an assessment method based on water ecological environment monitoring. By acquiring datasets such as topography, hydrology, and pollution sources, it uses ArcGIS to perform refined regional division of rivers and deploys monitoring equipment accordingly. Finally, it uses a random forest model to assess water quality. However, this technical solution has high indicator redundancy, lacks mechanistic constraints, and lacks a simplification mechanism. Chinese patent CN120409450A discloses a literature extraction method, system, equipment, and storage medium based on a large language model. It uses a large language model and PICO templates to automatically extract information such as research design, participant characteristics, and intervention measures from medical literature and generate structured tables. However, this method is mainly applicable to the medical field and lacks knowledge graph reasoning ability and mathematical statistical verification in the field of water ecological environment. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a simplified method and system for assessing the resilience of aquatic ecosystems, and to construct an index set and formula that is simple, interpretable, and has engineering application value.

[0005] Technical solution: The simplified index method for assessing the resilience of aquatic ecosystems described in this invention includes the following steps:

[0006] S1. Receive the user's assessment scenario instruction, use a large language model to perform mind chain parsing on the assessment scenario instruction to generate a multimodal retrieval strategy, and obtain water ecology-related text from the academic database according to the multimodal retrieval strategy.

[0007] S2. Extract the indicator relationships from the acquired text, and establish topological connections based on the causal, inclusion, collinear and power-law relationships between indicators to construct a water ecological indicator knowledge graph and form a full candidate indicator pool.

[0008] S3. Based on the water ecological index knowledge graph, calculate the correlation between the indicators through semantic analysis and correlation analysis, remove redundant indicators with high correlation, and form a primary simplified index set.

[0009] S4. Output the primary simplified index set and selection criteria; calculate the ecological resilience index for the same assessment scenario using both the full set of candidate indicators and the primary simplified indicators, and perform deviation analysis to verify the effectiveness of the primary simplified index set. If the verification is effective, output the final simplified index; otherwise, return to step S3 to increase the number of retained indicators until the verification is effective and output the final simplified index.

[0010] Furthermore, step S1, which utilizes a large language model to perform thought chain parsing on the evaluation scenario instructions to generate a multimodal retrieval strategy, includes:

[0011] The assessment scenario instructions are broken down into assessment objects, core objectives, and key stress factors;

[0012] Based on the decomposition results, a multimodal retrieval strategy is generated that includes dense vector queries and Boolean logic queries.

[0013] Furthermore, step S1 also includes time-sensitive denoising processing of the acquired water ecology-related texts: the time-sensitive denoising processing includes identifying and filtering obsolete standards and low-citation literature published earlier than a preset threshold.

[0014] Furthermore, step S1 also includes obtaining water ecology standard text from the standard library according to the multimodal retrieval strategy; step S3 also includes using the standard text in the standard library to verify whether the water ecology-related text has a standard sampling method, and discarding it if it does not.

[0015] Furthermore, in S2, the establishment of topological connections based on causal, inclusion, collinearity, and power-law relationships between indicators includes: using natural language processing methods to identify logical connectors and mathematical relationship descriptions in the text; defining unidirectional driving relationships between indicators as causal edges; defining hierarchical subordinate relationships as inclusion edges; defining statistically strong correlations as collinearity edges; and defining power-law relationships as power-law edges.

[0016] Furthermore, step S3, which calculates the correlation between indicators through semantic analysis and correlation analysis, includes:

[0017] Calculate the mathematical correlation coefficients between indicators using principal component analysis or factor analysis;

[0018] Analyze the semantic overlap between indicators using a large language model;

[0019] A comprehensive correlation matrix is ​​constructed by integrating the mathematical correlation coefficient and the semantic overlap.

[0020] Furthermore, in step S3, the correlation between indicators is calculated through semantic analysis and correlation analysis, and redundant indicators with high correlation are eliminated to form a preliminary simplified indicator set, including:

[0021] Using a large language model, a cluster of highly redundant indicators is constructed based on the comprehensive correlation matrix. In each cluster, the indicator with the highest information entropy is retained as the representative, and the remaining redundant indicators in the cluster are removed.

[0022] The large language model is used to screen and remove difficult-to-measure indicators, which are indicators whose monitoring cost, monitoring time, or data acquisition difficulty exceeds their corresponding threshold.

[0023] During the simplification process, the real-time monitoring indicator set covers the three functional dimensions of the aquatic ecosystem: self-balance, self-repair, and self-adaptation. When the number of indicators in a certain dimension is lower than its corresponding threshold, the indicator with the highest information entropy in that dimension is forcibly retained.

[0024] Furthermore, the increase in the number of retention indicators in step S4, which is a return to step S3, includes lowering the corresponding thresholds for the dimensions of aquatic ecosystem self-balancing, self-repairing, and / or adaptive functions.

[0025] Furthermore, after outputting the initial simplified index set and selection criteria in step S4, the process also includes receiving the user's annotation instructions, updating the large language model preferences based on the human feedback reinforcement learning mechanism, and outputting the corrected index set and index formula according to the annotation instructions.

[0026] The ecological resilience index of the assessment scenario is calculated using both the full pool of candidate indicators and the modified primary simplified indicators.

[0027] Furthermore, in step S4, the ecological resilience index is calculated for the same assessment scenario using both the full set of candidate indicators and the primary simplified indicators, and a bias analysis is performed to verify the effectiveness of the primary simplified indicator set, including:

[0028] Calculate the proportion of information in the full candidate indicator pool that the simplified indicator set explains to verify the indicator coverage;

[0029] Ecological resilience index was calculated for historical cases of the assessment scenario using both full candidate indicators and primary simplified indicators. The deviation between the assessment results using full candidate indicators and primary simplified indicators was calculated to verify the effectiveness of the indicators.

[0030] Based on the same inventive concept, the simplified index system for assessing the resilience of aquatic ecological environments described in this invention includes:

[0031] The retrieval and parsing module is used to receive the user's assessment scenario instructions, use a large language model to perform thought chain parsing on the assessment scenario instructions to generate a multimodal retrieval strategy, and retrieve water ecology-related texts from the academic database according to the multimodal retrieval strategy.

[0032] The graph construction module is used to extract indicator relationships from the acquired text. Using indicators as nodes, topological connections are established based on causal, inclusion, collinearity, and power-law relationships between indicators to construct a water ecological indicator knowledge graph and form a pool of full candidate indicators.

[0033] The simplified calculation module is used to calculate the correlation between indicators based on the water ecological indicator knowledge graph, through semantic analysis and correlation analysis, and to remove redundant indicators with high correlation to form a primary simplified indicator set.

[0034] The output and verification module is used to output the initial simplified index set and the selection criteria; the ecological resilience index is calculated for the same assessment scenario using the full set of candidate indicators and the initial simplified indicators respectively, and deviation analysis is performed to verify the effectiveness of the initial simplified index set. If the verification is effective, the final simplified index is output; otherwise, the process returns to step S3 to increase the number of retained indicators until the verification is effective and the final simplified index is output.

[0035] Based on the same inventive concept, the aquatic ecological environment resilience assessment method of the present invention includes: receiving an assessment scenario instruction from a user, using simplified indicators generated by the indicator simplification method for aquatic ecological environment resilience assessment, performing a resilience assessment on the assessment scenario, and obtaining an assessment result.

[0036] Beneficial effects: Compared with the prior art, the advantages of the present invention are: (1) The present invention introduces knowledge graphs and water environment standards, eliminates redundant and difficult-to-measure indicators, reduces monitoring costs and improves the feasibility of indicators; (2) The present invention utilizes water ecological resilience mechanism constraints (self-balancing, self-repairing, self-adaptation) and data back substitution to ensure that the simplified indicator set data is concise, avoids the accidental deletion of key resilience factors, and has interpretability in ecological mechanism; (3) The present invention effectively makes up for the lack of timeliness of traditional indicators by using multimodal retrieval and thinking chain analysis of the latest standard texts and academic literature. Attached Figure Description

[0037] Figure 1 This is a flowchart of the index simplification method according to an embodiment of the present invention.

[0038] Figure 2 This is a schematic diagram of a water ecological index knowledge graph according to an embodiment of the present invention.

[0039] Figure 3 This is a simplified system architecture diagram for indicators according to an embodiment of the present invention. Detailed Implementation

[0040] This application addresses the following shortcomings in current aquatic ecological environment assessments: First, the indicator system is complex and highly redundant, with excessive physical, chemical, biological, and hydrological indicators leading to high assessment costs and diluting the weight of key resilience factors; second, it lacks both mechanistic and data constraints, neglecting the causal logic within the aquatic ecosystem, resulting in selected indicators having large data variances but unclear ecological significance; third, it has a low level of intelligence and is difficult to update dynamically, with existing indicator databases lagging behind academic research and the discovery of new pollutants, lacking the ability to automatically mine new indicators from massive amounts of unstructured literature and verify their timeliness. Based on the above analysis, to address the core problems in current aquatic ecological environment assessment, such as high redundancy of indicator systems, lack of ecological mechanism constraints in the selection process, inability to dynamically adapt to specific scenarios, and low level of intelligence, this invention proposes a novel method that integrates standard texts, academic literature, and expert knowledge. This method utilizes intelligent means to automatically acquire all indicators, perform logical correlation analysis, and simplify multi-dimensional constraints. Leveraging the semantic understanding capabilities of large language models and the logical reasoning capabilities of knowledge graphs, combined with statistical dimensionality reduction and ecological resilience dimension constraints, it constructs an indicator set and index formula that possesses simplicity, interpretability, and engineering application value. The technical solution of this invention will be further described below with reference to the accompanying drawings.

[0041] like Figure 1 As shown, the simplified index method for assessing the resilience of aquatic ecosystems includes the following steps.

[0042] S1. Scene Analysis and Text Acquisition: Receives user's evaluation scene instructions, uses a large language model to perform CoT (CoT) parsing to generate a multimodal retrieval strategy, and uses a hybrid retrieval technology combining semantic vectors and keywords to obtain water ecology-related text from standard and academic libraries, and performs time-sensitive noise reduction processing.

[0043] Specifically, the assessment scenarios include, but are not limited to, resilience assessments of aquatic ecosystems such as oceans, lakes, and wetlands. The system can receive user assessment scenario commands through an interactive interface and utilize a pre-trained large language model (such as GPT-5 or a domestically developed large model) for thought chain reasoning to decompose fuzzy commands into structured retrieval elements: assessment object, stress factor, and core objective. Based on these decomposed elements, a hybrid retrieval strategy of "dense vector + Boolean keywords" is generated. The system automatically connects to standard libraries and academic literature databases. Then, based on set time thresholds, NLP technology is used to identify "repealed" and "revised" tags in the text, filtering out obsolete standards and low-citation-rate literature to construct a high-quality domain corpus.

[0044] Furthermore, this embodiment uses lake water ecological resilience assessment as an example for illustration. The user's assessment scenario command is "Lake Water Ecological Resilience Assessment." A pre-trained large language model (e.g., GPT-5 or a domestically developed large model) is used for thought chain reasoning. The retrieval elements are: assessment object (shallow lake), stress factors (e.g., nitrogen and phosphorus load, algal species count, COD, DO, etc.), and core objectives (adaptive, self-balancing, self-repair). Relevant texts are extracted from standard libraries (e.g., "Surface Water Environmental Quality Standard" GB3838-2002) and academic literature databases (e.g., Web of Science, CNKI). Based on a set time threshold (e.g., the last 10 years), expired standards and low-citation-rate literature are filtered to construct a high-quality domain corpus.

[0045] S2. Construction of a knowledge graph of water ecological indicators and generation of a full pool: Extract the relationship between indicators from the relevant texts in the acquired academic database, construct a knowledge graph of water ecological indicators, use the indicators as nodes in the graph, and establish topological connections by analyzing the causal, inclusion, collinear and power-law relationships between the indicators to form a full pool of candidate indicators.

[0046] Specifically, the method for establishing topological connections is to use natural language processing technology to identify logical connectors and mathematical relationship descriptions in the text, define the unidirectional driving relationship between indicators as causal edges, define the hierarchical subordinate relationship as inclusion edges, and define the strong correlation in statistics as collinear edges.

[0047] Furthermore, in this embodiment, entity relation extraction techniques (NER and RE) are used to extract relations from relevant text obtained from academic databases to construct an indicator knowledge graph, such as... Figure 2 As shown, the extracted physicochemical indicators (such as TN and TP), biological indicators (such as phytoplankton biomass and benthic diversity), hydrological indicators (such as water exchange cycle and lake capacity), and habitat indicators are defined as graph nodes. A topological connection of causal edges, inclusion edges, and collinear edges is established to form a pool of all candidate indicators.

[0048] Among them, causal edges are directed edges established based on the mechanism descriptions in the literature. For example, according to the description "an increase in phosphorus concentration induces the growth of cyanobacteria", a directed edge of the form <Total Phosphorus, drives, Cyanobacteria> is established. Inclusion edges are directed edges established by identifying the classification relationships between indicators. For example, <Total Nitrogen, includes, Ammonia Nitrogen>. Collinearity edges are undirected edges established based on statistical descriptions for marking indicators that often appear simultaneously and have the same trend. For example, <CODcr, strongly correlated, CODMn>. Power-law edges are directed edges established based on the power-law relationships indicated by formulas or relevant conclusions in the academic library on the basis of causal edges and inclusion edges. For example, based on the formula DI = S×E×F×D×PT, <DI, 1, S>, <DI, 1, E>, etc. can be obtained.

[0049] S3. Intelligent reduction of indicators based on multi-dimensional constraints: Use a large language model combined with a statistical dimensionality reduction algorithm to generate an indicator correlation matrix, and introduce relevant indicators in the standard library for verification. Automatically eliminate highly redundant indicators and difficult-to-measure indicators in the knowledge graph, and calculate and generate a primary reduced indicator set while retaining the maximum information entropy.

[0050] Specifically, the generation of the indicator correlation matrix by using a large language model combined with a statistical dimensionality reduction algorithm specifically includes: calculating the mathematical correlation coefficients between indicators using principal component analysis or factor analysis; analyzing the semantic overlap degree between indicators using a large language model; fusing the mathematical correlation coefficients and the semantic overlap degree to construct a comprehensive correlation matrix; and the elimination of highly redundant indicators refers to identifying indicator clusters with a correlation degree higher than a preset threshold in the comprehensive correlation matrix, and retaining the indicator with the highest information entropy within the cluster as a representative.

[0051] Specifically, introducing relevant indicators in the standard library for verification means that for the information extracted from the academic library, it is judged whether it exists in the standard library, that is, it is judged whether there is a standard sampling method for the parameters or indicators extracted from the academic library. If not, it is discarded.

[0052] Specifically, the determination of difficult-to-measure indicators is based on the metadata attributes of the knowledge graph and the sampling methods specified in the standard measurement methods. If the monitoring cost, monitoring time-consuming or data acquisition difficulty label of the indicator node exceeds the preset threshold, it is determined as a difficult-to-measure indicator and is eliminated.

[0053] Specifically, it also includes the water ecological resilience dimension constraint logic, which is used to detect the coverage integrity of the indicator set for the three functional dimensions of the water ecosystem's self-balance, self-repair and self-adaptation in real time during the statistical dimensionality reduction process. When the number of indicators in a certain dimension is lower than the preset threshold, the indicator with the highest information entropy in that dimension is forcibly retained.

[0054] Furthermore, in this embodiment, core indicators are selected from the full pool, specifically including the following three sub-logic: semantic-matrix dual dimensionality reduction, measurement difficulty assessment and selection, and water ecological resilience dimension constraint.

[0055] (1) Calculate the mathematical correlation coefficient between indicators using principal component analysis or factor analysis, and analyze the semantic overlap between indicators using a large language model. The correlation coefficient focuses on the collinearity of indicators, while the semantic overlap, i.e., semantic embedding similarity, focuses on whether there are different ways of describing the same indicator. The comprehensive correlation index is obtained by summing their respective matrix weights, and a comprehensive correlation matrix is ​​constructed based on the comprehensive correlation index. Identify high-redundancy indicator clusters based on the comprehensive correlation index, i.e., indicators with high comprehensive correlation index values ​​are merged into one cluster (e.g., transparency SD and water turbidity). In each cluster, retain the indicator with the highest information entropy (i.e., the most information carried) as the representative, and remove the remaining redundant terms in the cluster.

[0056] (2) Use a large language model to extract metadata about the monitoring attributes of indicators from the knowledge graph, such as monitoring cost, time consumption, and equipment cost. For example, if the indicator "phytoplankton primary productivity" is marked as "requires carbon isotope determination", find its collinear indicator "chlorophyll a" in the knowledge graph, then replace "phytoplankton primary productivity" with "chlorophyll a" and automatically remove "phytoplankton primary productivity" to improve the applicability of the project.

[0057] (3) Built-in water ecological resilience dimension constraint logic to monitor the distribution of indicators in real time during the simplification process. The indicators are automatically mapped to three functional dimensions: self-balancing (Resistance, such as nutrient tolerance threshold), self-repair (Recovery, such as submerged plant seed bank, endogenous release in sediment), and self-adaptability (such as hydrological connectivity, morphological diversity). If statistical dimensionality reduction results in the number of indicators in a certain dimension falling below the preset threshold, the system will forcibly recall the indicator with the highest information entropy or the highest degree centrality in the knowledge graph under that dimension, to prevent the loss of resilience representation due to excessive pursuit of data simplification.

[0058] S4. Reinforcement Learning and Backtracking Validation: The system presents users with a preliminary simplified indicator set and its selection criteria, receives user annotation instructions, updates the backend preference model based on human feedback reinforcement learning, and outputs a revised indicator set and index formula according to the annotation instructions. The index formula is a LaTeX format formula similar to the eutrophication index. The revised indicator set and index formula are then back-substituted into historical case data or a standard system for secondary validation of coverage and effectiveness, generating the final simplified indicator set for assessing the resilience of the aquatic ecosystem.

[0059] Specifically, the update of the background preference model based on human feedback reinforcement learning mechanism includes recording users' retention, deletion or modification of the weight of indicators as preference data, and using the reward model to fine-tune the simplification strategy of the large language model so that the model learns the selection preferences of experts.

[0060] Specifically, the secondary verification includes coverage verification and validity verification. The secondary verification is considered successful if both coverage verification and validity verification are passed.

[0061] Coverage verification refers to calculating the proportion of information explained by the simplified indicator set in the original full candidate indicator pool, including the Nash efficiency coefficient (NSE) or information loss rate. If it is greater than a preset threshold, the verification is considered successful. Validity verification refers to calculating the evaluation results using historical case data and comparing the evaluation results of the simplified indicator set with those of the full indicator set. If the deviation is less than a preset value, the verification is considered successful. Deviation indicators include the coefficient of determination R², mean absolute error (MAE), mean absolute percentage error (MAPE), mean squared error (MSE), and root mean square error (RMSE), calculated using the following formulas.

[0062]

[0063]

[0064]

[0065]

[0066]

[0067] In this embodiment, a preliminary simplified indicator set and selection reasons are output (e.g., "Retain TP, remove SRP, reason: correlation 0.95 and TP has broader coverage"). Users add or delete indicators based on specific engineering experience (e.g., "The lake sediment is heavily polluted, sediment indicators need to be added"). Human feedback reinforcement learning (RLHF) records expert modifications and updates the backend model preferences, making the next recommendation more aligned with expert intuition. Updating the backend model preferences refers to introducing a dynamic preference weighting factor in the indicator selection logic of step S3, transforming the semantic-matrix dual dimensionality reduction criterion into a comprehensive weighted evaluation mechanism of "information entropy + expert preference score". Under this mechanism, the system comprehensively considers information entropy and preference indicators when removing redundant indicators. As the RLHF phase continues to iterate, if a specific indicator is repeatedly manually retained or frequently added by users in multiple rounds of interaction, the preference weight of that indicator in the background model will be accumulated and enhanced. As a result, in the subsequent S3 reduction calculation, even if the information entropy of that indicator is slightly lower than that of other indicators in the same cluster, the system will prioritize identifying it as a representative indicator due to its extremely high preference score, making the retention tendency of that indicator greater and greater.

[0068] In this embodiment, the system integrates a historical water environment monitoring case library (containing full monitoring data from the past 5-10 years and corresponding water environment classifications) to verify the effectiveness of the simplified indicator set. The ecological resilience index of the historical time series is calculated using all candidate indicators (e.g., 120). This is then compared with the ecological resilience index of the same time series calculated using simplified indicators (e.g., 15). A deviation analysis is performed to calculate the Nash efficiency coefficient (NSE), information loss rate, and coefficient of determination R² for the two sets of evaluation results. If NSE > 0.85, information loss rate < 10%, and R² > 0.85, the simplified indicator set is deemed effective, and the final result is output. If the criteria are not met, the system automatically reverts to S3, lowers the dimensionality reduction threshold, and increases the number of retained indicators until validation is passed.

[0069] like Figure 3 As shown, the simplified index system for assessing the resilience of aquatic ecosystems according to the present invention includes:

[0070] The retrieval and parsing module is used to receive evaluation scenario instructions, use a large language model to parse the thought chain to generate a multimodal retrieval strategy, and perform hybrid retrieval and denoising processing to obtain water ecology-related text.

[0071] The graph construction module is used to extract relationships from relevant texts, construct a water ecological indicator knowledge graph containing indicator nodes and various topological connections, and form a full pool of candidate indicators.

[0072] The simplified calculation module is used to combine a large language model with a statistical dimensionality reduction algorithm to generate a correlation matrix, remove highly redundant and difficult-to-measure indicators, and generate a primary simplified indicator set that retains the maximum information entropy.

[0073] The interactive feedback module is used to display the indicator set, receive user annotation instructions, update the backend preference model through a human feedback reinforcement learning mechanism, and output the corrected indicator set and index formula.

[0074] The verification module integrates a historical water environment monitoring case library and an assessment deviation analysis unit. It is used to backtest the corrected index set and generate the final simplified assessment index.

[0075] The aquatic ecological environment resilience assessment method of this invention includes: receiving an assessment scenario instruction from a user; using simplified indicators generated by the indicator simplification method for aquatic ecological environment resilience assessment to perform a resilience assessment on the assessment scenario; and obtaining an assessment result. This method constructs a scientifically accurate aquatic ecological environment resilience assessment indicator system with engineering application value.

Claims

1. A simplified index method for assessing the resilience of aquatic ecosystems, characterized in that, Includes the following steps: S1. Receive the user's assessment scenario instruction, use a large language model to perform mind chain parsing on the assessment scenario instruction to generate a multimodal retrieval strategy, and obtain water ecology-related text from the academic database according to the multimodal retrieval strategy. S2. Extract the indicator relationships from the acquired text, and establish topological connections based on the causal, inclusion, collinear and power-law relationships between indicators to construct a water ecological indicator knowledge graph and form a full candidate indicator pool. S3. Based on the water ecological index knowledge graph, calculate the correlation between the indicators through semantic analysis and correlation analysis, remove redundant indicators with high correlation, and form a primary simplified index set. S4. Output the primary simplified index set and selection criteria; calculate the ecological resilience index for the same assessment scenario using both the full set of candidate indicators and the primary simplified indicators, and perform deviation analysis to verify the effectiveness of the primary simplified index set. If the verification is effective, output the final simplified index; otherwise, return to step S3 to increase the number of retained indicators until the verification is effective and output the final simplified index.

2. The method for simplifying indicators for assessing the resilience of aquatic ecosystems according to claim 1, characterized in that, Step S1, which uses a large language model to parse the thought chain of the evaluation scenario instructions to generate a multimodal retrieval strategy, includes: The assessment scenario instructions are broken down into assessment objects, core objectives, and key stress factors; Generate a multimodal retrieval strategy that includes dense vector queries and Boolean logic queries based on the decomposition results; Step S1 also includes time-sensitive denoising processing of the acquired water ecology-related texts: the time-sensitive denoising processing includes identifying and filtering obsolete standards and low-citation literature published earlier than a preset threshold.

3. The method for simplifying indicators for assessing the resilience of aquatic ecosystems according to claim 1, characterized in that, Step S1 further includes obtaining water ecology standard text from the standard library according to the multimodal retrieval strategy; Step S3 further includes verifying whether the water ecology-related text has a standard sampling method using the standard text in the standard library, and discarding it if it does not.

4. The method for simplifying indicators for assessing the resilience of aquatic ecosystems according to claim 1, characterized in that, Step S2, which establishes topological connections based on causal, inclusion, collinearity, and power-law relationships among indicators, includes: Natural language processing methods are used to identify logical connectors and mathematical relationships in text. One-way driving relationships between indicators are defined as causal edges, hierarchical subordinate relationships are defined as inclusion edges, strong statistical correlations are defined as collinear edges, and power-law relationships are defined as power-law edges.

5. The method for simplifying indicators for assessing the resilience of aquatic ecosystems according to claim 1, characterized in that, Step S3, which calculates the correlation between indicators through semantic analysis and correlation analysis, includes: Calculate the mathematical correlation coefficients between indicators using principal component analysis or factor analysis; Analyze the semantic overlap between indicators using a large language model; A comprehensive correlation matrix is ​​constructed by integrating the mathematical correlation coefficient and the semantic overlap.

6. The method for simplifying indicators for assessing the resilience of aquatic ecosystems according to claim 5, characterized in that, In step S3, the correlation between indicators is calculated through semantic analysis and correlation analysis. Redundant indicators with high correlation are removed to form a preliminary simplified indicator set, which includes: Using a large language model, a cluster of highly redundant indicators is constructed based on the comprehensive correlation matrix. In each cluster, the indicator with the highest information entropy is retained as the representative, and the remaining redundant indicators in the cluster are removed. The large language model is used to screen and remove difficult-to-measure indicators, which are indicators whose monitoring cost, monitoring time, or data acquisition difficulty exceeds their corresponding threshold. During the simplification process, the real-time monitoring indicator set covers the completeness of the three functional dimensions of the aquatic ecosystem: self-balance, self-repair, and self-adaptation. When the number of indicators in a certain dimension is lower than its corresponding threshold, the indicator with the highest information entropy in that dimension is forcibly retained. Step S4 returns to step S3 to increase the number of retention indicators, including lowering the corresponding thresholds for the dimensions of aquatic ecosystem self-balancing, self-repairing, and / or adaptive functions.

7. The method for simplifying indicators for assessing the resilience of aquatic ecosystems according to claim 6, characterized in that, After outputting the initial simplified index set and selection criteria in step S4, the process also includes receiving the user's annotation instructions, updating the large language model preferences based on the human feedback reinforcement learning mechanism, and outputting the corrected index set and index formula according to the annotation instructions. The ecological resilience index of the assessment scenario is calculated using both the full pool of candidate indicators and the modified primary simplified indicators.

8. The method for simplifying indicators for assessing the resilience of aquatic ecosystems according to claim 1, characterized in that, Step S4 involves calculating the ecological resilience index for the same assessment scenario using both the full set of candidate indicators and the primary simplified indicators, and performing a bias analysis to verify the effectiveness of the primary simplified indicator set. Calculate the proportion of information in the full candidate indicator pool that the simplified indicator set explains to verify the indicator coverage; Ecological resilience index was calculated for historical cases of the assessment scenario using both full candidate indicators and primary simplified indicators. The deviation between the assessment results using full candidate indicators and primary simplified indicators was calculated to verify the effectiveness of the indicators.

9. A simplified index system for assessing the resilience of aquatic ecosystems, characterized in that, include: The retrieval and parsing module is used to receive the user's assessment scenario instructions, use a large language model to perform thought chain parsing on the assessment scenario instructions to generate a multimodal retrieval strategy, and retrieve water ecology-related texts from the academic database according to the multimodal retrieval strategy. The graph construction module is used to extract indicator relationships from the acquired text. Using indicators as nodes, topological connections are established based on causal, inclusion, collinearity, and power-law relationships between indicators to construct a water ecological indicator knowledge graph and form a pool of full candidate indicators. The simplified calculation module is used to calculate the correlation between indicators based on the water ecological indicator knowledge graph, through semantic analysis and correlation analysis, and to remove redundant indicators with high correlation to form a primary simplified indicator set. The output and verification module is used to output the initial simplified index set and the selection criteria; the ecological resilience index is calculated for the same assessment scenario using the full set of candidate indicators and the initial simplified indicators respectively, and deviation analysis is performed to verify the effectiveness of the initial simplified index set. If the verification is effective, the final simplified index is output; otherwise, the process returns to step S3 to increase the number of retained indicators until the verification is effective and the final simplified index is output.

10. A method for assessing the resilience of aquatic ecological environments, characterized in that, The method includes: receiving an assessment scenario instruction from a user, using a simplified index generated by the index simplification method for aquatic ecological environment resilience assessment as described in any one of claims 1-8, performing a resilience assessment on the assessment scenario, and obtaining an assessment result.

Images