Establishment method of ecological resource intelligent credible data element model

Abstract

The invention provides a method for establishing an ecological resource intelligent credible data element model. The method comprises the following steps: acquiring natural ecological basic data; obtaining natural ecology derivative data, wherein the natural ecology derivative data comprises ecological tourism description metadata and leisure health care description metadata; performing controlled feature mining on the natural ecology basic data and the natural ecology derivative data to generate ecological basic data elements and ecological derivative data elements; performing key feature extraction on the ecological basic data elements and the ecological derivative data elements to obtain basic ecological key features and derivative ecological key features; and obtaining green financial data elements, and establishing association relationships between the green financial data elements and the basic ecological key features and between the green financial data elements and the derivative ecological key features to generate an ecological resource credible data element model. According to the technical scheme, the problem that effective association between green financial data elements and ecological resource data is lacked is at least solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent processing technologies, and particularly to a method for establishing an intelligent and trustworthy data element model for ecological resources. Background Art

[0002] In the current era of parallel digitalization and sustainable development, the effective management and rational utilization of ecological resources have become increasingly crucial. Ecological resources are not only the foundation for maintaining the earth's ecological balance but also an important support for promoting the green development of the economic society. The natural ecosystem encompasses diverse components such as forests, rivers, lakes, grasslands, and arable lands, and the rich data it contains is of inestimable value for scientific decision-making, ecological protection, and the development of green industries. At the same time, with the increasing attention of people to the ecological environment, ecological derivative fields such as eco-tourism and leisure and health care have flourished, and the related data has also increased day by day.

[0003] However, there are many problems in the field of ecological resource data management currently. On the one hand, for the basic data of natural ecology, such as the descriptive metadata of forests, rivers, etc., they are often scattered in different departments and systems, lacking effective integration and unified management, resulting in poor data consistency and availability. Moreover, when analyzing these basic data, it is difficult to comprehensively and accurately mine their controlled characteristics, and it is impossible to provide strong data element support for the scientific assessment of ecological resources. On the other hand, for the natural ecological derivative data such as eco-tourism and leisure and health care, they also face the problems of data fragmentation and uneven quality, lacking effective technical means in key feature extraction, and it is difficult to accurately extract the derivative ecological key features that are of important guiding significance for the comprehensive utilization of ecological resources and industrial development. In addition, as an important force in promoting ecological protection and sustainable development, the data elements of green finance lack an effective association mechanism with ecological resource data, and it is impossible to give full play to the supporting role of green finance in the development and protection of ecological resources.

[0004] In summary, the existing ecological resource data management technologies cannot meet the current requirements for intelligent and trustworthy management of ecological resources and the coordinated development of green industries, and there is an urgent need for an innovative method to solve these problems. Summary of the Invention

[0005] To solve the above technical problems, the present application provides a method for establishing an intelligent and trustworthy data element model for ecological resources to at least solve or alleviate the problems existing in the above prior art.

[0006] To achieve the above object, according to one aspect of the present application, there is provided a method for establishing an intelligent and trustworthy data element model for ecological resources, which includes:

[0007] Obtain natural ecological basic data, where the natural ecological basic data includes at least one of forest description metadata, river description metadata, lake description metadata, grassland description metadata, and cultivated land description metadata;

[0008] Obtain natural ecological derivative data, where the natural ecological derivative data includes ecological tourism description metadata and leisure and health care description metadata;

[0009] Perform controlled feature mining on the natural ecological basic data and the natural ecological derivative data respectively to generate ecological basic data elements and ecological derivative data elements;

[0010] Extract key features from the ecological basic data elements and the ecological derivative data elements respectively to obtain basic ecological key features and derivative ecological key features;

[0011] Obtain green financial data elements and establish an association relationship between them and the basic ecological key features and the derivative ecological key features to generate an ecological resource credible data element model.

[0012] The technical solutions in this application have at least the following technical advantages:

[0013] (1) By obtaining natural ecological basic data, this solution covers at least one of the description metadata of forests, rivers, lakes, grasslands, cultivated lands, etc. This operation breaks the current situation where data is scattered in different departments and systems, realizes the centralized acquisition of basic data, and lays a foundation for subsequent integration and unified management. Performing controlled feature mining on the obtained natural ecological basic data can comprehensively and accurately extract ecological basic data elements from these basic data, thereby providing strong data element support for the scientific evaluation of ecological resources and solving the problem of difficult effective mining of controlled features in the past.

[0014] (2) This application obtains natural ecological derivative data, including ecological tourism description metadata and leisure and health care description metadata, and collects the originally fragmented ecological derivative data. Extracting key features from these natural ecological derivative data can accurately extract derivative ecological key features, provide important guidance for the comprehensive utilization of ecological resources and industrial development, and make up for the lack of effective key feature extraction technical means in this field in the past.

[0015] (3) This solution obtains green financial data elements and establishes an association relationship between them and the basic ecological key features and the derivative ecological key features. Through the establishment of this association mechanism, the supporting role of green finance in the development and protection of ecological resources can be fully exerted, solving the problem of the lack of effective association between green financial data elements and ecological resource data, and meeting the needs of intelligent and credible management of ecological resources and the coordinated development of green industries. Description of the Drawings

[0016] Figure 1 This is a schematic flow chart of a method for establishing an intelligent and trustworthy data element model of ecological resources in an embodiment of the present application. Specific implementation manner

[0017] Figure 1 This is a schematic flow chart of a method for establishing an intelligent and trustworthy data element model of ecological resources in an embodiment of the present application. As Figure 1 shown, it includes:

[0018] Obtain natural ecological basic data, where the natural ecological basic data includes at least one of forest description metadata, river description metadata, lake description metadata, grassland description metadata, and cultivated land description metadata;

[0019] Obtain natural ecological derivative data, where the natural ecological derivative data includes ecological tourism description metadata and leisure and health care description metadata;

[0020] Respectively perform controlled feature mining on the natural ecological basic data and the natural ecological derivative data to generate ecological basic data elements and ecological derivative data elements;

[0021] Respectively perform key feature extraction on the ecological basic data elements and the ecological derivative data elements to obtain basic ecological key features and derivative ecological key features;

[0022] Obtain green financial data elements, and establish an association relationship between them and the basic ecological key features and the derivative ecological key features to generate an ecological resource trustworthy data element model.

[0023] Optionally, the obtaining of the natural ecological basic data includes:

[0024] Based on the constructed multi-source data collection rules, parse the carriers carrying different types of natural ecological basic data to obtain the natural ecological basic data therefrom.

[0025] Preferably, in a specific application scenario, the specific implementation process of the above solution is as follows:

[0026] Preferably, in the natural ecological basic data acquisition link, there are n natural ecological basic data carriers, denoted as the set V = {v1, v2,..., v n}. The natural ecological basic data carried by each carrier v i (1 ≤ i ≤ n) is a high-dimensional vector where m i represents the dimension of the i-th carrier data, reflecting the data complexity. The constructed multi-source data collection rule is a non-linear mapping function f c : It is based on an improved Autoencoder structure. The Autoencoder consists of an encoder e and a decoder d. The encoder e maps the input data D bi to a low-dimensional representation z i = e(D bi ), and the decoder d then reconstructs z i into by minimizing the reconstruction error (where λ is the regularization parameter that controls the sparsity of the low-dimensional representation) to train the Autoencoder. The trained encoder e is the non-linear mapping function f in the multi-source data acquisition rule c , which analyzes the carrier data to obtain the processed data D bi ' = f c (D bi ). The finally obtained natural ecological basic data D b is: where ω i is the weight coefficient, determined by the Analytic Hierarchy Process (AHP), which reflects the importance of different carrier data. In the AHP method, a judgment matrix A = (a ij ) is constructed, where a ij represents the comparison of the importance of carrier i relative to carrier j. The maximum eigenvalue λ max of the judgment matrix and its corresponding eigenvector are calculated by the eigenvalue method, and the eigenvector is normalized to obtain the weight coefficient ω i .

[0027] The actual meanings of the various parameters in the above formula are detailed as follows: n: the number of types of natural ecological basic data carriers. v i : the i-th type of natural ecological basic data carrier. D bi : the high-dimensional vector of the natural ecological basic data carried by the i-th carrier, whose dimension m i varies according to the carrier type, reflecting the data richness. m i′ : the dimension of the data after being processed by the multi-source data acquisition rule, which may be different from m i , determined by the Autoencoder structure. f c : the non-linear mapping function trained based on the Autoencoder, used to analyze the carrier data, remove noise and redundant information, and extract key features. e: the encoder of the Autoencoder, which maps high-dimensional data to a low-dimensional space. d: the decoder of the Autoencoder, which reconstructs the original data from the low-dimensional representation. z i : the low-dimensional representation of the i-th carrier data in the Autoencoder. L: the training loss function of the Autoencoder, consisting of the reconstruction error and the regularization term, ensuring the reconstruction accuracy and the sparsity of the low-dimensional representation. λ: the regularization parameter, which adjusts the balance between the reconstruction error and the sparsity of the low-dimensional representation. ω i: The weight coefficient determined by the AHP method reflects the contribution degree of the data of the i-th carrier in the final natural ecological basic data. A: The judgment matrix in AHP, which is used to compare the relative importance of different carrier data. a ij : The element in the judgment matrix A, which represents the importance comparison of carrier i relative to carrier j. λ max : The largest eigenvalue of the judgment matrix A.

[0028] Preferably, in the link of obtaining natural ecological derivative data, let the natural ecological derivative description text be the sequence T = [t1, t2, …, t N , where t j (1 ≤ j ≤ N) is the j-th lexical unit in the text, and N is the text length. First, perform structuring processing. This application uses the model g s based on the conditional random field (CRF). CRF is defined on the undirected graph G = (V, E), where the vertex V corresponds to the lexical units in the text, and the edge E represents the relationship between the words. Let x = [x1, x2, …, x N be the sequence of feature vectors of the input text, and y = [y1, y2, …, y N be the output structured label sequence. The joint probability distribution of CRF is: Among them, Z(x) is the normalization factor, λ k is the weight parameter of the feature function f k , which is obtained by maximum likelihood estimation training. After being processed by the CRF model, the structured text S = g s (T) is obtained.

[0029] Preferably, perform semantic analysis on the structured text S. This application adopts the method based on deep semantic embedding (DSE). Construct a deep neural network model g sa , with the structured text S as the input, and map it to a high-dimensional semantic space through multiple hidden layers to obtain the natural ecological derivative feature element F = g sa (S). This model is trained by minimizing the contrast loss function (Textranslation failed) (where M is the number of training samples, sim is the cosine similarity function, and [·] + represents taking the positive value), so that texts with similar semantics are closer in the high-dimensional space, and texts with different semantics are farther apart. Based on the natural ecological derivative feature element F, generate the natural ecological derivative data D d using a variant of the generative adversarial network (GAN) - the conditional generative adversarial network (cGAN). The generator G inputs the noise vector z and the natural ecological derivative feature element F to generate the data D d = G(z, F); the discriminator D inputs the real natural ecological derivative data and the generated data Dd , determine its authenticity. cGAN minimizes the objective function through adversarial training:

[0030] The creative physical meanings of the various parameters in the above formula are described as follows: T: The text sequence of natural ecological derivative descriptions. t j : The j-th lexical unit in the text. N: The text length. g s : The structured processing model based on conditional random fields. G: The undirected graph in CRF, used to describe the relationship between text words. x: The sequence of feature vectors of the input text, including information such as the word itself, part of speech, context, etc. y: The output structured label sequence, such as sentence component labels, etc. Z(x): The normalization factor of CRF, ensuring the legality of the probability distribution. λ k : The weight parameter of the feature function f k in CRF, and the influence degree of different features on the structured label is learned through training. f k : The feature function of CRF, describing the relationship between the label and the features of the input text. S: The text after structured processing. g sa : The semantic analysis model based on deep semantic embedding. F: The natural ecological derivative feature element, which is the representation of the structured text in the high-dimensional semantic space. L dse : The contrastive loss function of the deep semantic embedding model, used to optimize the model to better capture semantic information. M: The number of training samples. sim: The cosine similarity function, measuring the similarity between two feature vectors in the semantic space. G: The generator of cGAN, generating natural ecological derivative data according to the noise vector and the natural ecological derivative feature element. z: The noise vector, introducing randomness for the generated data. D: The discriminator of cGAN, judging the authenticity of the generated data. Real natural ecological derivative data. L cgan : The objective function of cGAN, improving the quality of the generated data through the adversarial training of the generator and the discriminator.

[0031] Preferably, in the controlled feature mining link, the dataset composed of the natural ecological basic data D b and the natural ecological derivative data D d is D = {D b , D d}. The set of business rules and feature mining constraint conditions set is represented as a constraint function set where c l : is a real-valued function, for example c lIt can be a constraint on the statistical characteristics of data (such as the mean and variance range of data) or a constraint on semantic relationships (such as the logical relationships between specific ecological indicators). Building a controlled feature mining model is a hybrid model h based on variational autoencoder-reinforcement learning (VAE-RL). In the variational autoencoder part, for the input data x ∈ D, the encoder E maps it to the latent space z = E(x), where (μ(x) and σ 2 (x) are the mean and variance respectively), and the decoder D generates reconstructed data from the latent space by minimizing the variational lower bound loss function (where D KL is the Kullback-Leibler divergence, measuring the difference between two probability distributions) to train the variational autoencoder. In the reinforcement learning part, an agent A is defined, whose state space S is the latent space z, and the action space A is the transformation operation on the latent space (such as scaling, translation, etc.). The agent selects an action a ∈ A according to the current state s ∈ S and obtains a reward r. The reward function R is designed to be related to the constraint conditions, that is (where α l is the weight coefficient, indicating the importance of different constraint conditions). The agent learns the optimal policy π(s) by maximizing the long-term cumulative reward (where γ is the discount factor, determining the importance of future rewards). Through the VAE-RL hybrid model h, controlled feature mining is performed on the natural ecological basic data and natural ecological derivative data to generate ecological basic data elements E b and ecological derivative data elements E d :

[0032] E b = h(D b ), E d = h(D d ).

[0033] In the above formulas, the creative physical meanings of each parameter are explained as follows: D: The dataset composed of natural ecological basic data and natural ecological derivative data. A constraint set composed of multiple constraint functions c l to ensure that feature mining conforms to business rules and specific constraints. h: A controlled feature mining model based on variational autoencoder-reinforcement learning. E: The encoder of the variational autoencoder, mapping the input data to the latent space. D: The decoder of the variational autoencoder, reconstructing data from the latent space. z: The latent space variable, following a normal distribution μ(x): The mean of the input data x in the latent space. σ 2 (x): The variance of the input data x in the latent space. L vae: The variational lower bound loss function of the variational autoencoder, which balances the reconstruction accuracy and the rationality of the latent space distribution. D KL : Kullback-Leibler divergence, used to measure the difference between the latent space distribution and the standard normal distribution. A: Reinforcement learning agent. S: The state space of the agent, which is the latent space z. A: The action space of the agent, including transformation operations on the latent space. r: The reward obtained after the agent executes an action. R: The reward function, related to the constraint conditions, guiding the agent to learn by punishing actions that violate the constraints. α l : Constraint condition c l : The weight coefficient in the reward function, reflecting the importance of different constraints. G t : The long-term cumulative reward of the agent starting from time step t. γ: Discount factor, determining the influence degree of future rewards on the current decision. π(s): The optimal policy of the agent, selecting the optimal action according to the current state. E b : The generated ecological basic data elements. E d : The generated ecological derivative data elements.

[0034] Preferably, in the key feature extraction stage, the ecological basic data elements E b and the ecological derivative data elements E d constitute an element set E = {E b , E d}. Set the key feature extraction direction as a high-dimensional vector field where x ∈ E, has a dimension of d θ , and its value changes with the data point x to adapt to the feature extraction requirements of different data regions. Perform multi-scale transformation on E, using a combined method based on multi-resolution analysis (MRA) and discrete wavelet transform (DWT). Let be the scale parameter. For the data sample x ∈ E, the approximation coefficient A s (x) and the detail coefficient D s (x) are obtained through DWT at scale s. Preferably, in order to combine the key feature extraction direction, weight the coefficients to obtain the weighted approximation coefficient and the weighted detail coefficient where, and are the weight vectors related to the approximation coefficient and the detail coefficient respectively extracted from the vector field . After weighting, the feature information of the ecological basic data elements and the ecological derivative data elements at different scales s is obtained

[0035] Preferably, for the feature information at different scales perform fusion analysis. This application uses a model f based on the graph neural network (GNN) f . Construct a graph structure G = (V, E), where the vertices V are the feature information at different scales (i.e., and ), and the edges E represent the relationships between features (such as spatial adjacency relationships, semantic association relationships, etc.). Through the propagation and learning process of the GNN, the basic ecological key features K b and the derived ecological key features K d are obtained: Among them, and are the feature information of the ecological basic data elements and ecological derived data elements at scale s, respectively.

[0036] In the above formula, the creative physical meanings of each parameter are described as follows: E: The set of elements composed of ecological basic data elements and ecological derived data elements, which integrates the data after controlled feature mining and provides a rich data basis for key feature extraction. Different types of data elements play different roles in subsequent analysis and jointly contribute to extracting representative key features. The high-dimensional vector field of the key feature extraction direction, whose dimension d θ determines the complexity and diversity of feature extraction. This vector field changes dynamically according to different data points x and can sensitively capture the unique features of data in different regions. For example, in ecological basic data elements, for different attributes such as forest area and river flow, the vector field can adjust the direction weights to highlight the features related to the stability of the ecosystem; in ecological derived data elements, for data such as the flow of ecological tourism tourists and the satisfaction of leisure and health care services, the vector field can focus on the key factors affecting industrial development. s: Scale parameter, and the value set is Different scales s represent different resolution observation perspectives on data. Smaller scales focus on the local details of data, such as the distribution details of specific tree species in a forest; larger scales focus on the overall trend of data, such as the change trend of forest coverage area in a large geographical area. Through multi-scale transformation, the feature information of data at different levels can be comprehensively obtained. A s (x): The approximate coefficient of the data sample x obtained by discrete wavelet transform (DWT) at scale s. It retains the low-frequency components of the data at this scale and reflects the overall trend and main features of the data. For example, when analyzing river descriptive metadata, the approximate coefficient can reflect the average change trend of river flow over a period of time. D s(x): The detail coefficients of the data sample x obtained at scale s through DWT. It contains the high-frequency components of the data at this scale, revealing the local variations and detailed information of the data. For example, when analyzing grassland descriptive metadata, the detail coefficients can highlight the local variations of grassland vegetation in specific seasons or regions, such as sudden increases or decreases in vegetation in certain areas. and are the weight vectors related to the approximation coefficients and detail coefficients respectively extracted from the vector field . They perform weighted adjustments on the approximation coefficients and detail coefficients according to the key feature extraction directions. For example, if the key feature extraction direction focuses on highlighting the long-term stable features of the ecosystem, then will assign higher weights to the part of the approximation coefficients that reflects the long-term trend; if focusing on the short-term fluctuations of the ecosystem, will emphasize the information in the detail coefficients that reflects short-term changes. The set of characteristic information of ecological basic data elements and ecological derivative data elements at different scales s, which consists of the weighted approximation coefficients and the weighted detail coefficients . This set comprehensively summarizes the characteristic representations of the data under different scales and key feature extraction directions, providing rich and diverse information for subsequent fusion analysis. G: The graph structure constructed by the graph neural network (GNN) model f f , where the vertices V are the characteristic information at different scales (i.e., and ), and the edges E represent the relationships between the features. Through this graph structure, the GNN can capture the complex associations between features, such as spatial adjacency relationships (the feature associations of ecologically adjacent regions in the geographical space) and semantic association relationships (such as the feature associations corresponding to the similarities and differences in semantics between eco-tourism and leisure and health care). K b and K d : They are the basic ecological key features and derivative ecological key features respectively. They are the results obtained after the GNN performs fusion analysis on the characteristic information at different scales, representing the features of ecological basic data elements and ecological derivative data elements that are of key significance for the assessment, development, and protection of ecological resources. These key features can provide the core basis for the subsequent correlation analysis of green finance data elements and the construction of a credible data element model for ecological resources.

[0037] Preferably, in the link of establishing association relationships and generating models, the green finance data element is where p represents the dimension of the green finance data, reflecting the complexity and the number of features of the green finance data. Calculate the green finance data element with the basic ecological key feature K b and the derivative ecological key feature K dRegarding the correlation between them, this application adopts a complex similarity metric function extended based on Mutual Information Mutual Information is used to measure the degree of dependence between two random variables. The extended function considers the high-order dependence relationships among multiple variables. Let X = [K b , K d , D g be a joint variable set containing basic ecological key features, derivative ecological key features, and green financial data elements. By calculating the mutual information matrix M mi among the variables in the joint variable set, and through a series of matrix transformations (such as singular value decomposition, etc.), the correlation matrix M r is obtained: M mi (i, j) = I(X i ; X j ), M r = Transform(M mi ), where I(X i ; X j ) represents the mutual information between variables X i and X j , and Transform represents a series of complex matrix transformation operations, including but not limited to matrix normalization, adjustment of singular values after singular value decomposition, and recombination of matrices, etc., to highlight the correlation characteristics between variables.

[0038] To calculate the causality between green financial data elements and basic ecological key features, derivative ecological key features, this application adopts a complex inference algorithm based on the Structural Causal Model (SCM) combined with the Bayesian Network (Bayesian Network) Construct a Bayesian network B = (G, Θ), where G is a directed acyclic graph, and the nodes represent variables (i.e., K b , K d , and D g ), and the edges represent the causal relationships between variables; Θ is a set of conditional probability distribution parameters. Through learning a large amount of data, the maximum a posteriori estimation (MAP) method is used to determine the parameter Θ, and according to the inference rules of the Bayesian network, the causal effects between variables are calculated to obtain the causal relationship graph Based on the correlation matrix M r and the causal relationship graph Construct an ecological resource trustworthy data element model Adopt a method that integrates deep learning based on a Knowledge Graph. The correlation matrix and the causal relationship graph are used as part of the Knowledge Graph, where nodes represent different data elements (green finance data elements, basic ecological key features, derivative ecological key features), and edges represent the correlation and causal relationships between them. Through a deep learning model based on the Graph Convolutional Network (GCN) Learn and reason about the Knowledge Graph to generate a credible data element model for ecological resources

[0039] The creative physical meanings of the various parameters in the above formula are explained as follows: D g : Green finance data elements, whose dimension p reflects the richness and complexity of data in the green finance field, covering various financial indicators such as green investment amount and green credit interest rate. The similarity metric function based on the extension of mutual information can more accurately capture the complex correlation between green finance data elements and ecological key features by considering high-order dependencies. X: A joint variable set containing basic ecological key features, derivative ecological key features, and green finance data elements, used to calculate mutual information. M mi : Mutual information matrix, which records the mutual information values between variables in the joint variable set and reflects the degree of dependence between variables. Transform: A series of matrix transformation operations used to transform the mutual information matrix into a correlation matrix, highlighting the correlation features between variables. M r : Correlation matrix, which shows the degree of correlation between green finance data elements and basic ecological key features, derivative ecological key features, providing important association information for subsequent model construction. The causality inference algorithm based on the Structural Causal Model combined with the Bayesian Network can deeply analyze the causal relationships between variables. B: The constructed Bayesian Network, which consists of a directed acyclic graph G and a set of conditional probability distribution parameters Θ, used to describe the causal structure and probability relationships between variables. G: The directed acyclic graph in the Bayesian Network, where nodes represent variables and edges represent causal relationships, and its structure is determined through data learning. Θ: The set of conditional probability distribution parameters of the Bayesian Network, which is learned and determined based on the data through the maximum a posteriori estimation method. Causal relationship graph, which intuitively shows the causal relationships between green finance data elements and basic ecological key features, derivative ecological key features, providing causal logic support for the credible data element model of ecological resources. A deep learning model based on the Graph Convolutional Network is used to learn and reason about the Knowledge Graph (including the correlation matrix and the causal relationship graph) to generate a credible data element model for ecological resources. The finally generated credible data element model for ecological resources integrates the correlation and causal relationship between green finance data elements and key ecological features, and can provide comprehensive and credible data support for the scientific management and evaluation of ecological resources and the development of green industries.

[0040] Therefore, based on the above solution, this embodiment has at least the following technical advantages:

[0041] 1. By constructing a multi-source data acquisition rule based on an improved autoencoder, the natural ecological basic data of different carriers can be effectively parsed and fused. Traditional methods are difficult to handle the complex structure of multi-source data, while this method uses the autoencoder to reduce the dimension and reconstruct high-dimensional tensor data, removing noise and redundant information, enabling data of different carriers to be processed within a unified framework. For example, when integrating multi-source data such as forests and rivers, key features can be accurately extracted, avoiding data loss, improving data quality, and providing a solid foundation for subsequent analysis. The analytic hierarchy process (AHP) is used to determine the weights of data of different carriers. Compared with subjective judgment, it can more scientifically reflect the importance of each carrier's data to the overall natural ecological basic data. In ecological resource management, the contributions of data of different carriers to ecological assessment are different. For example, when assessing the ecological stability of a region, forest cover data may be more influential than the local data of a small wetland. The weights determined by AHP can reasonably reflect this difference, making the integrated data more in line with the actual ecological significance.

[0042] 2. The structured processing model based on conditional random fields (CRF) and the semantic analysis method of deep semantic embedding can effectively process the natural ecological derivative description text. Traditional text processing methods are difficult to deeply mine the semantic relationships and potential structures in the text, while CRF can use the relationships between text words for structured annotation, and the DSE model makes similar semantic texts closer in the high-dimensional space by minimizing the contrast loss function, accurately extracting the natural ecological derivative feature elements. When analyzing texts related to ecological tourism and leisure and health care, key information such as tourist experiences and service quality can be accurately extracted, providing high-quality features for the generation of ecological derivative data. The conditional generative adversarial network (cGAN) is used to generate natural ecological derivative data. Compared with traditional generation methods, cGAN can use natural ecological derivative feature elements as conditions to generate more diverse and realistic scenario-compliant data. When generating data for predicting the tourist flow of ecological tourism, it can combine conditions such as tourist resource characteristics and seasonal factors to generate more real and reliable data, providing strong support for ecological tourism planning and management.

[0043] 3. The hybrid model based on variational autoencoder - reinforcement learning can conduct feature mining while meeting complex business rules and feature mining constraints. Traditional feature mining methods are difficult to balance constraints and feature optimization, while this model adjusts feature representations guided by constraints through the exploration of the reinforcement learning agent in the latent space. In ecological resource data mining, such as when mining ecological basic data elements, it ensures that data features conform to the logical relationships and statistical constraints of ecological indicators, while optimizing features to better reflect the characteristics of the ecosystem. The variational autoencoder maps data to the latent space, providing a new perspective for feature mining. The operations of the agent in the latent space can discover feature combinations and variation patterns that are difficult to detect by traditional methods. For example, when exploring ecological derivative data elements, it can mine the potential collaborative development features between eco - tourism and leisure and health care, providing a basis for the innovative development of ecological industries.

[0044] 4. The combined method based on multi - resolution analysis and discrete wavelet transform, combined with the key feature extraction direction vector field, can comprehensively obtain data features from different scales. Traditional feature extraction methods often only focus on features at a single scale or in a fixed direction, while this method can dynamically adjust the feature extraction direction according to data points, capturing both the overall trend and local details of the data at different scales. When analyzing forest ecological data, it can not only grasp the change trend of the overall forest coverage area but also pay attention to the detailed changes in the tree species distribution in specific areas, providing more comprehensive information for ecological resource assessment. Using graph neural networks to conduct fusion analysis on multi - scale feature information can effectively capture the complex correlation relationships between features. Traditional methods are difficult to handle high - dimensional and complex feature correlations, while GNN, by constructing a graph structure with different - scale features as nodes and feature relationships as edges, can deeply mine the spatial adjacency and semantic associations between features. When analyzing the correlation between ecological resources and green finance data, it can discover the potential influence paths between ecological key features and green finance indicators, providing a decision - making basis for green finance to support the development of ecological resources.

[0045] 5. The inference algorithm combining the similarity metric function based on mutual information expansion and the structural causal model with the Bayesian network can accurately analyze the correlation and causality between green finance data elements and ecological key features. Traditional analysis methods are difficult to accurately measure the high-order dependencies and causal relationships among multiple variables, while this method determines the correlation by calculating the mutual information matrix and performing complex transformations, and determines the causal relationship by using Bayesian network learning and reasoning. When studying the role of green finance in ecological resource protection, it can clarify the causal link between green investment and the improvement of ecosystem stability, as well as the correlation between green credit interest rates and the development of ecotourism, providing a scientific basis for policy formulation. The method of generating a credible data element model for ecological resources based on knowledge graph fusion and deep learning can integrate the correlation and causality to generate a comprehensive and credible data model. Traditional models are difficult to comprehensively consider various complex relationships, while this method integrates the correlation matrix and the causal relationship graph into the knowledge graph and learns and reasons through graph convolutional networks, enabling the model to reflect the internal logic of ecological resource data. In ecological resource management, this model can provide unified and accurate data support for ecological resource assessment, development, and protection, improving the scientificity and effectiveness of ecological resource management.

[0046] Optionally, the obtaining of the natural ecological derivative data includes:

[0047] Performing structured processing on the natural ecological derivative description text to obtain structured text;

[0048] Performing semantic analysis on the structured text to extract natural ecological derivative feature elements;

[0049] Based on the natural ecological derivative feature elements, natural ecological derivative data.

[0050] Preferably, in one embodiment, in the structured processing link of the natural ecological derivative description text, the natural ecological derivative description text is T = [t1, t2,..., t N , where t j represents the j-th lexical unit in the text, and N is the text length. Specifically, it includes the following steps:

[0051] (1) Lexical embedding and position encoding step

[0052] To convert the text into a vector representation, first perform an embedding operation on each lexical item t j . Adopt the relative position encoding embedding method based on Transformer. Let the lexical embedding matrix be where d vocab is the size of the vocabulary, and d emb is the embedding dimension. The embedding vector e j of the lexical item t j = V[t j , where V[tj represents extracting the row vector corresponding to the vocabulary t from the vocabulary embedding matrix V j . At the same time, in order to capture the position information of the vocabulary in the text, a position encoding vector p j is introduced. The position encoding is generated by a sine-cosine function: where j is the position index of the vocabulary, k is the dimension index of the embedding vector, and p j,k is the k-th element of the position encoding vector p j . The final vocabulary feature vector x j = e j + p j .

[0053] (2) Structured processing model

[0054] Preferably, in one embodiment, a model based on an improved Graph Convolutional Recurrent Network (GCRN) is used for structured processing. A text graph G=(V, E) is constructed, where the vertices V={v1, v2, …, v N} correspond to the vocabulary in the text, and the edges E represent the relationships between the vocabulary. Here, the weights of the edges are determined by calculating the semantic similarity between the vocabulary, and the semantic similarity adopts an improved method based on cosine similarity: where x i and x j are the feature vectors of the vocabulary t i and t j , ‖i - j‖ is the position distance of the vocabulary in the text, and σ is a parameter controlling the distance attenuation. When s(t i , t j ) is greater than a certain threshold τ, an edge is added between the vertices v i and v j , and the weight of the edge is s(t i , t j ).

[0055] In the GCRN model, the update formula for the hidden state h j of each vertex v j is as follows: where l is the number of layers of the GCRN, and GRU is the Gated Recurrent Unit, which is used to process sequence information. After being processed by the GCRN of L layers, the final hidden state of each vocabulary is obtained These hidden states constitute the representation of the structured text

[0056] Preferably, in one embodiment, the semantic analysis of the structured text includes the following steps:

[0057] (1) Construction of semantic analysis model

[0058] A model based on the Deep Semantic Fusion Network (DSFN) is used to perform semantic analysis on the structured text S. The DSFN model consists of multiple semantic fusion modules, and each module contains a self-attention mechanism and a multi-layer perceptron (MLP). In the self-attention mechanism, for the input structured text S, the query vector Q = W Q S, the key vector K = W K S, and the value vector V = W V S are calculated, where is a learnable weight matrix, and d att is the dimension of the attention mechanism. The self-attention output A is calculated as follows: Then, the self-attention output A is input into the multi-layer perceptron for further feature extraction. Let the input of the multi-layer perceptron be A, and it is processed through two fully connected layers and an activation function (such as ReLU): F = MLP(A) = W2ReLU(W1A + b1) + b2, where W1 and W2 are the weight matrices of the fully connected layers, and b1 and b2 are the bias vectors. After being processed by M semantic fusion modules, the natural ecological derivative feature element F M is obtained.

[0059] Preferably, in one embodiment, in the link of generating natural ecological derivative data based on the natural ecological derivative feature element, a generation model based on the fusion of a generative adversarial network (GAN) and a variational autoencoder (VAE) is used to generate natural ecological derivative data. Let the natural ecological derivative feature element be F M , and it is concatenated with a random noise vector to obtain the input vector I = [F M ; z], where is a normal distribution with a mean of 0 and a covariance matrix of the identity matrix I. The specific technical process is as follows:

[0060] (1) Generator construction

[0061] The generator G is a multi-layer neural network with the input I and the output being the generated natural ecological derivative data D g . Let the weight matrices of each layer of the generator be and the bias vector be , then: where σ is the activation function (such as tanh).

[0062] (2) Discriminator construction

[0063] The discriminator D is also a multi-layer neural network with the input being the real natural ecological derivative data Dr or the generated data D g , the output is a scalar representing the probability that the data is real data. Let the weight matrices of each layer of the discriminator be and the bias vector be Then: During the training process, the model is optimized by minimizing the adversarial losses of the generator and the discriminator. The goal of the generator is to maximize the probability that the discriminator judges the generated data as real data, and the goal of the discriminator is to maximize the probability of judging real data as real data and generated data as fake data. The specific loss functions are as follows:

[0064] By alternately optimizing the generator and the discriminator, the generated natural ecological derivative data D g becomes closer and closer to the distribution of real data.

[0065] In the above formulas, the specific physical meanings of each parameter are as follows: T: The natural ecological derivative description text sequence, which is the original text data and contains description information about eco-tourism, leisure and health care, etc. t j : The j-th lexical unit in the text, which is the basic element that constitutes the text. N: The text length, which reflects the scale of the text and affects the complexity of subsequent processing. V: The lexical embedding matrix, which maps each word in the vocabulary to a emb d-dimensional vector space, and its dimension d vocab ×d emb determines the richness of the word representation and the computational complexity. d vocab : The size of the vocabulary, that is, the number of different words in the text, which affects the number of rows of the lexical embedding matrix. d emb : The embedding dimension, which determines the length of each word vector. A larger embedding dimension can capture richer lexical semantic information. p j : The position encoding vector, which is used to add position information to the words so that the model can distinguish the same words in different positions. Its dimension is the same as that of the word embedding vector, that is, d emb . G: The text graph, which is used to describe the relationships between words in the text. By constructing a graph structure, the model can perform structured processing using the semantic associations between words. V: The vertex set of the text graph, corresponding to the words in the text. E: The edge set of the text graph, and the weight of the edge is determined by the semantic similarity between words, reflecting the closeness between words. s(t i , t j ): The words t i and t jThe semantic similarity between them comprehensively considers the vector representation of words and the positional distance, and is used to determine the weights of the edges in the text graph. σ: The parameter controlling distance attenuation, which determines the degree of influence of the lexical positional distance on the similarity when calculating the semantic similarity. A larger σ makes the influence of the positional distance on the similarity smaller. τ: The edge addition threshold. When the semantic similarity between words is greater than this threshold, an edge is added in the text graph, and its value affects the sparsity and structure of the text graph. GRU: Gated Recurrent Unit, which is used to process text sequence information and update the hidden state of vertices in the GCRN. Its internal parameters (such as weight matrices and bias vectors) are obtained through training. l: The number of layers of the GCRN, which determines the depth of the model's extraction of text structure information. Increasing the number of layers can learn more complex text structure features. L: The total number of layers of the GCRN. After L-layer processing, the final representation of the structured text is obtained.

[0066] Preferably, for each parameter in the semantic analysis stage, its physical meaning is explained as follows: S: The representation of the structured text, which is the sequence of word hidden states obtained after being processed by the GCRN, containing the structure and partial semantic information of the text. W Q ,W K ,W V : The weight matrix in the self-attention mechanism, which is used to map the structured text S to the query, key, and value vector spaces. Its dimension d att ×d emb determines the computing and representation capabilities of the attention mechanism. d att : The dimension of the attention mechanism, which affects the complexity of the self-attention calculation and the feature representation ability. A larger dimension can capture richer semantic relationships, but it also increases the computational amount. A: The self-attention output, which is obtained by calculating the query, key, and value vectors, reflecting the attention distribution between words in the text. W1, W2: The weight matrices in the multi-layer perceptron, which are used to further extract features from the self-attention output. Their dimensions are determined according to the number of input and output features. b1, b2: The bias vectors in the multi-layer perceptron, which are used to adjust the output of the neural network. ReLU: The activation function, which is used to introduce non-linearity so that the model can learn more complex semantic features. M: The number of semantic fusion modules, which determines the depth and complexity of the model's extraction of semantic features. Increasing the number of modules can learn more advanced semantic representations, but it may also lead to an increase in training time and overfitting. F M : The natural ecological derivative feature element obtained after being processed by M semantic fusion modules, which is the final result of the semantic analysis of the structured text, containing key semantic features related to eco-tourism, leisure and health care, etc.

[0067] In the data generation stage, the physical meaning of each parameter is explained as follows: F M: Natural ecological derivative feature element, as part of the input of the generation model, provides semantic constraints for the generated data, making it conform to the characteristics of natural ecological derivative data. z: Random noise vector, following a normal distribution It is used to introduce randomness into the generated data, making the generated data diverse. I: Input vector of the generator, concatenated by the natural ecological derivative feature element and the random noise vector, and its dimension is the sum of the dimension of the natural ecological derivative feature element and the dimension of the noise vector. G: Generator, which is a multi-layer neural network used to generate natural ecological derivative data according to the input vector, and its internal parameters (such as weight matrix and bias vector) are obtained through training and learning. The weight matrix of each layer of the generator determines the transformation method of the generator for the input vector. The bias vector of each layer of the generator is used to adjust the output of each layer of the generator. P: Number of layers of the generator, which determines the processing complexity of the generator for the input vector and the quality of the generated data. σ: Activation function (such as tanh), used to introduce non-linearity, enabling the generator to generate a richer data distribution. D g : Generated natural ecological derivative data, which is the output result of the generator. D: Discriminator, which is a multi-layer neural network used to determine whether the input data is real data or generated data, and its internal parameters (such as weight matrix and bias vector) are obtained through training and learning. The weight matrix of each layer of the discriminator determines the feature extraction and judgment method of the discriminator for the input data. The bias vector of each layer of the discriminator is used to adjust the output of each layer of the discriminator. Q: Number of layers of the discriminator, which determines the discrimination ability of the discriminator for data. Increasing the number of layers can improve the discrimination accuracy. p: Probability value output by the discriminator, indicating the probability that the input data is real data. L G : Loss function of the generator, used to measure the difference between the data generated by the generator and the real data distribution.

[0068] Therefore, based on the above solution, in one embodiment, it has the following technical advantages:

[0069] 1. The relative position encoding embedding method using Transformer maps words to a high-dimensional vector space through a vocabulary embedding matrix, endowing each word with a rich semantic representation. The position encoding ingeniously integrates the position information of words in the text into the vector representation through unique sine and cosine functions. In ecotourism texts, for the two words "scenic area entrance" and "visitor center", not only are their semantics distinguished through vocabulary embedding, but the position encoding can also reflect their order in describing the tour route, enabling the model to accurately capture the spatial and semantic structure of the text and providing a solid foundation for subsequent analysis. The semantic similarity between words is calculated through an improved cosine similarity formula, comprehensively considering the word vector representation and the position distance, and assigning weights to the edges of the text graph. In leisure and health care texts, "forest bathing" and "meditation activities" are semantically similar and may be adjacent in the description. The high semantic similarity prompts a strong connection to be established between the corresponding vertices in the text graph. This text graph constructed based on semantic associations uses a graph convolutional recurrent network (GCRN) to update the hidden state of the vertices, effectively mining the context relationships and semantic dependency structures among the words in the text. Compared with traditional sequence models, it can better handle the complex relationships among words in long texts and improve the accuracy of structured processing.

[0070] 2. The self-attention mechanism in the deep semantic fusion network (DSFN) generates query, key, and value vectors through a learnable weight matrix, calculating the attention distribution among the words in the text. When analyzing texts related to ecotourism and leisure and health care, for the description of "hiking adventure in ecotourism projects", the self-attention mechanism enables the model to focus on the semantic connections between "hiking" and "adventure" and related modifying words, highlighting key semantic information, avoiding the loss of local information, comprehensively and deeply understanding the text semantics, and improving the accuracy of semantic analysis. The self-attention output is further processed by a multi-layer perceptron, through two fully connected layers and the ReLU activation function, performing non-linear transformation and feature extraction on the word semantics. In the scenario of ecological derivative data, it can extract high-level semantic features such as "the positive impact of ecotourism on physical and mental health" from the original word semantics, integrating the relevant semantic information scattered throughout the text, providing rich and accurate semantic information for generating natural ecological derivative feature elements, and helping to explore the internal connections between ecological resources and the tourism and health care industries.

[0071] 3. A generative model based on the fusion of a Generative Adversarial Network (GAN) and a Variational Autoencoder (VAE) concatenates natural ecological derivative feature elements with a random noise vector as the input of the generator. When generating ecological tourism tourist flow prediction data, the natural ecological derivative feature elements (such as semantic features like the ecological resource characteristics of scenic spots and seasonal factors) provide semantic constraints for the generated data to ensure that the data conforms to the actual scenario of ecological tourism; the random noise vector introduces randomness, making the generated tourist flow data diverse within a reasonable range and meeting the prediction requirements in different scenarios. Compared with traditional data generation methods, it can generate more practical and diverse data. The generator and discriminator are alternately optimized through adversarial loss. The generator tries to generate natural ecological derivative data that the discriminator misjudges as real, while the discriminator continuously improves its ability to distinguish true and false data. When generating leisure and healthcare service evaluation data, as the training progresses, the data generated by the generator gradually approaches the distribution of real evaluation data, and the data quality continuously improves, providing reliable data support for ecological resource development and service improvement, effectively solving the problems of insufficient authenticity and diversity of data in traditional generative models.

[0072] Optionally, the step of respectively performing controlled feature mining on the natural ecological basic data and the natural ecological derivative data to generate ecological basic data elements and ecological derivative data elements includes:

[0073] Constructing a controlled feature mining model according to the set business rules and feature mining constraint conditions;

[0074] Based on the controlled feature mining model, performing controlled feature mining on the natural ecological basic data and the natural ecological derivative data to generate ecological basic data elements and ecological derivative data elements.

[0075] Preferably, in one embodiment, constructing the controlled feature mining model includes the following technical processing procedures:

[0076] (1) Formalization of business rules and constraint conditions

[0077] Let the set of business rules be where R i is a Boolean function R i : n is the data feature dimension. In the ecological resource management scenario, if R1 represents that "the proportion of forest area should be between 0 and 1", for the input natural ecological data feature vector x = (x1, x2,..., x n ), when the proportion component of the forest area in x meets the range requirement, R1(x) = 1, otherwise R1(x) = 0.

[0078] The feature mining constraint condition is represented by the constraint function Φ(x), which is a real-valued function Φ: For example, in ecological tourism data mining, the constraint condition may be that "the difference between the number of tourists during the peak tourist season and the carrying capacity of tourism facilities should be within a reasonable range". At this time, Φ(x) can be defined as the absolute value of the difference between the two, and whether the constraint is satisfied is judged by setting a threshold.

[0079] (2) Construction of Variational Autoencoder-Reinforcement Learning Hybrid Model

[0080] Variational Autoencoder (VAE): The input natural ecological data The encoder maps it to the latent space. Let the encoder network be a Multi-Layer Perceptron (MLP), then the mean μ(x) and log variance logσ 2 (x) of the latent space are calculated as follows: μ(x) = MLP μ (x; θ μ ) where MLP μ and is a multi-layer perceptron with θ μ and as parameters.

[0081] Sample z from the latent space: z = μ(x) + σ(x) ⊙ ∈, where ⊙ represents element-wise multiplication.

[0082] The decoder is also a multi-layer perceptron that reconstructs z into

[0083] The loss function L VAE of VAE is: where is the posterior distribution, is the prior distribution, and p(x|z) is the conditional probability that the decoder generates x. It can be further expanded as: Here k is the dimension of the latent space.

[0084] (3) Reinforcement Learning (RL)

[0085] The state s of the reinforcement learning agent is defined as s = [z, Φ(z)], and the action is the adjustment of the latent variable z.

[0086] The reward function R(s, a) is: R(s, a) = αR rec (s, a) + βR con (s, a) + γR rule (s, a)

[0087] where: is the reconstruction reward, which encourages the agent to generate data that can be well reconstructed. is the constraint reward, and δ is the constraint threshold. is the business rule reward, and λ i is the rule weight.

[0088] The agent uses the policy network π(a|s; θ π ) to select an action and updates the policy network parameter θ using the Proximal Policy Optimization (PPO) algorithm π . The PPO algorithm updates θ by maximizing the following objective function π : where is the importance sampling ratio, is the advantage estimate, and ∈ is the clipping parameter.

[0089] Preferably, in a specific embodiment, feature mining based on the controlled feature mining model includes the following technical steps:

[0090] (1) Data input and preprocessing

[0091] Normalize the natural ecological basic data and the natural ecological derivative data to obtain and The normalization function is

[0092] (2) Feature mining process

[0093] For each normalized data sample (from the basic data or the derivative data):

[0094] 1. Obtain the latent variable z through the VAE encoder. 2. The agent, according to the state s = [z, Φ(z)], uses the policy network π(a|s; θ π ) to select an action a. 3. Update the latent variable z′ = z + a. 4. Check the constraint condition Φ(z′) and the business rule R i (z′), and if not satisfied, the agent reselects the action until satisfied. 5. Reconstruct z' into 6. through the VAE decoder. Perform denormalization on the reconstructed data

[0095] Preferably, in a specific embodiment, in the process of generating the ecological basic data elements and the ecological derivative data elements, after the above feature mining process, the ecological basic data elements and the ecological derivative data elements

[0096] In the above processing process, the physical meanings of each parameter are described as follows: Business rule set It contains multiple Boolean functions, which are used to define the business requirements that data in ecological resource management need to meet, such as data range, logical relationships, etc. The constraint function Φ(x): measures the deviation degree of the data feature vector x from the constraint conditions, and ensures the rationality of data in ecological data mining. θ μ 、 and are the parameters of the multi-layer perceptrons of the encoder and decoder respectively, which determine the mapping and reconstruction of data in the latent space. L VAE : Balances the data reconstruction error and the latent space distribution, enabling the model to learn an effective representation of the data. α, β, and γ: The weight coefficients of the reconstruction reward, constraint reward, and business rule reward in the reward function. Adjusting them can control the behavior of the agent. θ π : The parameters of the policy network, which are optimized by the PPO algorithm to maximize the cumulative reward. ∈: The clipping parameter of the PPO algorithm, which prevents the policy update from being too large. X base and X derived : Are the sets of natural ecological basic data and derivative data samples respectively. N and M: Are the numbers of samples of the basic data and derivative data respectively. δ: The constraint threshold, which is used to judge whether the constraint conditions are met. λ i : The weight of the business rule R i reflects the importance of different rules.

[0097] Preferably, in one embodiment, the specific technical processing process of respectively performing controlled feature mining on the natural ecological basic data and the natural ecological derivative data to generate ecological basic data elements and ecological derivative data elements has the following technical advantages:

[0098] 1. Traditional methods usually lack systematic integration of business rules and constraint conditions. In the scenario of ecological resource management, it may simply perform manual screening after data collection, or use simple conditional statements to preliminarily filter the data. For example, when processing the data of the proportion of forest area, it may only perform manual checks during the data entry stage, and it is impossible to ensure in real time and comprehensively that the data complies with the rules during the feature mining process. For complex constraint conditions, such as the dynamic relationship between the number of tourists and the carrying capacity of facilities in the ecological tourism scenario, traditional methods are difficult to effectively handle and often cannot incorporate such constraints into the core process of data mining.

[0099] This solution formalizes the business rules as a set of Boolean functions And representing the feature mining constraint conditions as a real - valued function Φ(x) can precisely define and quantify various rules and constraints. Throughout the feature mining process, processing is strictly carried out in accordance with these rules and constraints starting from data input. For example, in the mining of ecological basic data elements, for business rules such as the proportion of forest area and the range of river flow, the model can judge in real - time whether the data feature vector meets the conditions. When mining ecological tourism data, the constraint function Φ(x) can continuously monitor the difference between the number of tourists and the carrying capacity of facilities, and make dynamic adjustments according to the set thresholds to ensure that the mined data elements fully meet the actual business requirements and constraint conditions, greatly improving the accuracy and reliability of the data.

[0100] Traditional data feature extraction and dimensionality reduction methods, such as principal component analysis (PCA), can only extract linear features and cannot handle the probability distribution and generation problems of data well. In ecological resource data, the data often has complex non - linear features, and it is difficult for traditional methods to capture the internal relationships between these features. For example, when analyzing the forest ecosystem data in natural ecological basic data, there are complex non - linear associations between features such as the species diversity of forests, vegetation coverage, and soil quality, and PCA cannot effectively extract this information. At the same time, traditional methods have limited ability in generating new data samples and are difficult to meet the needs of data expansion and prediction in ecological resource management.

[0101] The VAE in this solution uses a multi - layer perceptron as the encoder and decoder, and can learn the non - linear feature representation of data. By mapping the data to the latent space, VAE can not only extract the key features of the data, but also generate new data samples through operations on the latent space. For example, when processing natural ecological derivative data, such as the prediction of tourist flow in ecological tourism, VAE can learn the distribution law of the latent space based on historical data, and then generate tourist flow prediction data under different scenarios by sampling in the latent space. The loss function L of VAE VAE balances the reconstruction error and the latent space distribution, enabling the model to learn an effective representation of the data, providing a solid foundation for subsequent feature mining and generation tasks.

[0102] Traditional feature mining methods usually rely on fixed algorithms and parameters and lack the ability to adapt to environmental changes and constraint conditions. In the scenario of ecological resource management, business rules and constraint conditions may change with factors such as time, season, and policy, and traditional methods cannot adjust the mining strategy in a timely manner. For example, during the peak and off - peak seasons of ecological tourism, the constraint conditions between the number of tourists and the carrying capacity of facilities may be different, and it is difficult for traditional methods to dynamically adjust the feature mining process according to such changes. In addition, traditional methods often cannot make full use of the feedback information between data, resulting in the mined features may not be optimal.

[0103] In this solution, the reinforcement learning agent can perceive the latent variable z and the satisfaction of the constraints in real time by defining the state as s = [z, Φ(z)]. The reward function R(s, a) comprehensively considers the reconstruction reward, constraint reward, and business rule reward, enabling the agent to not only generate data that can be well reconstructed when exploring the latent space but also ensure the satisfaction of various constraints and business rules. For example, when mining ecological basic data elements, the agent can adjust the action a according to the current latent variable and constraints to generate data features that conform to business rules such as the proportion of forest area and the range of river flow. The parameters θ of the policy network are updated through the Proximal Policy Optimization (PPO) algorithm π , enabling the agent to continuously learn and optimize the policy, improve the feature mining ability in complex environments, and thus generate more valuable ecological basic and derivative data elements.

[0104] 4. Traditional feature mining processes are often single and linear, lacking effective interaction and feedback between various links. In the scenario of ecological resource management, different types of data (such as natural ecological basic data and natural ecological derivative data) may require different processing methods, but traditional methods are difficult to comprehensively process these data. For example, when processing ecological tourism data, traditional methods may only focus on single data such as tourist flow and ignore the association with natural ecological basic data (such as the ecological environment quality of scenic spots). In addition, traditional methods are less efficient when processing large-scale data and difficult to meet real-time requirements. The controlled feature mining model constructed in this solution organically combines links such as data input and preprocessing, feature mining process, and generation of ecological data elements. Through the unified processing of natural ecological basic data and natural ecological derivative data, the potential associations between different types of data can be mined. For example, when generating ecological basic data elements and ecological derivative data elements, the model can consider forest ecosystem data and ecological tourism data simultaneously and mine key information such as the impact of ecological tourism on the forest ecosystem. At the same time, through the synergistic effect of VAE and RL, the model can efficiently perform feature mining on the premise of meeting business rules and constraints, improving the efficiency and accuracy of data processing, and providing more comprehensive and valuable data support for ecological resource management.

[0105] Optionally, the key feature extraction of the ecological basic data elements and the ecological derivative data elements respectively to obtain the basic ecological key features and the derivative ecological key features includes:

[0106] Performing multi-scale transformation on the ecological basic data elements and the ecological derivative data elements based on the set key feature extraction direction to obtain the feature information of the ecological basic data elements and the ecological derivative data elements at different scales;

[0107] Fusion analysis is performed on the characteristic information of ecological basic data elements and ecological derivative data elements at different scales to obtain the basic ecological key features and derivative ecological key features.

[0108] Preferably, in one embodiment, the set of ecological basic data elements is The set of ecological derivative data elements is where N and M are the numbers of ecological basic data elements and ecological derivative data elements respectively. The multi-scale transformation based on the set key feature extraction direction specifically includes the following steps:

[0109] (1) Definition of the key feature extraction direction

[0110] Set the key feature extraction direction as a high-dimensional vector field where x can be or The dimension of is d θ , and its value changes with the change of the data point x to adapt to the feature extraction requirements of different data regions. In the ecological resource management scenario, for ecological basic data elements, such as data on forest area, river flow, etc., The direction can be adjusted according to the functional zoning of the ecosystem (such as core protected areas, buffer zones, etc.) to highlight the key features related to ecological protection; for ecological derivative data elements, such as data on the flow of eco-tourism visitors, satisfaction with leisure and health care services, etc., The direction can be adjusted according to market demand and industrial development trends to focus on the key factors affecting the sustainable development of the ecological industry.

[0111] (2) Multi-scale transformation

[0112] A combined method based on the non-subsampled contourlet transform (NSCT) and the fractional Fourier transform (FrFT) is used for multi-scale transformation. Let be the scale parameter. For the ecological basic data element The approximate coefficient and the detail coefficient are obtained through NSCT at scale s. For this reason, since NSCT is a multi-scale and multi-directional image decomposition method, it can effectively capture the edge and texture information of the data.

[0113] In order to combine the key feature extraction direction, the NSCT coefficients are weighted to obtain the weighted approximate coefficient and the weighted detail coefficient where, and are respectively from the vector field The weight vectors related to the approximation coefficients and detail coefficients extracted from

[0114] Meanwhile, perform a fractional Fourier transform, which is defined as:

[0115] where α is the order of the fractional Fourier transform, and K α (t, u) is the kernel function, defined as:

[0116] Here, n is an integer, and δ is the Dirac function. The fractional Fourier transform can analyze data in the time-frequency domain. By adjusting the order α, characteristic information with different time-frequency characteristics can be obtained.

[0117] After the fractional Fourier transform, the coefficient is obtained at scale s. Similarly, weighted processing is performed to obtain where is the weight vector related to the fractional Fourier transform coefficients extracted from the vector field .

[0118] For the ecological derivative data elements the above NSCT and fractional Fourier transform and their weighted processing are similarly performed to obtain

[0119] After weighted processing, the characteristic information of the ecological basic data elements at different scales s the characteristic information of the ecological derivative data elements at different scales s

[0120] Preferably, in one embodiment, in the link of fusion analysis of the characteristic information at different scales, the following steps are included:

[0121] (1) Construct a fusion analysis model

[0122] Adopt a fusion analysis model based on a quantum neural network (QNN) and a graph attention network (GAT). For the characteristic information of the ecological basic data elements at different scales construct a graph structure G base =(V base , E base ), where the vertex V base is the characteristic information at different scales (i.e., ), and the edge E baseRepresents the relationship between features. Here, the edge weights are determined by calculating the quantum correlation between features, and the quantum correlation is calculated using a method based on quantum mutual information. Let the quantum states of two features x and y be ρ x and ρ y , and their quantum mutual information I Q (x; y) is defined as: I Q (x; y) = S(ρ x ) + S(ρ y ) - S(ρ xy ), where S(ρ) is the von Neumann entropy of the quantum state ρ, defined as S(ρ) = -tr(ρ log2ρ), and ρ xy is the joint quantum state of x and y. When I Q (x; y) is greater than a certain threshold τ base , an edge is added between vertices x and y, and the weight of the edge is I Q (x; y).

[0123] Input the graph structure G base into the graph attention network, and the graph attention network calculates the importance weights of each vertex through the attention mechanism. For vertex v ∈ V base , its attention weight α v is calculated as follows: where, is a learnable attention vector, W is the weight matrix, and are the feature vectors of vertex v and its adjacent vertex u respectively, is the set of adjacent vertices of vertex v, ‖ represents vector concatenation, and LeakyReLU is the leaky rectified linear unit. Then, input the output of the graph attention network into the quantum neural network. The quantum neural network consists of quantum neurons, and the state update of the quantum neurons uses quantum gate operations. Let the initial state of the quantum neuron be |ψ0>, and after a series of operations of quantum gates U1, U2, …, U K , the final state |ψ K > = U K U K-1 … U1|ψ0> is obtained. The quantum gate U k can be a rotation gate, a phase gate, etc., and its parameters are obtained through training and learning. Through the fusion analysis of the quantum neural network and the graph attention network, the basic ecological key feature K base is obtained. For the feature information of ecological derivative data elements at different scales , a graph structure G derived =(V derived , E derived) Calculate the weights of the edges (based on quantum mutual information), and through the fusion analysis of the graph attention network and the quantum neural network, obtain the key features K of the derived ecosystem derived .

[0124] In the above technical processing process, the creative physical meanings of each parameter are described as follows: E base : The set of ecological basic data elements, which contains the data elements related to the natural ecological basis obtained after controlled feature mining, such as the quantitative representations of features such as forest area and river flow. E derived : The set of ecological derived data elements, which contains the data elements related to derived fields such as eco-tourism and leisure and health care, such as the quantitative representations of data such as the tourist flow of eco-tourism and the satisfaction degree of leisure and health care services. N: The number of ecological basic data elements, which reflects the richness and diversity of ecological basic data. M: The number of ecological derived data elements, which reflects the scale and complexity of ecological derived data. The high-dimensional vector field of the key feature extraction direction, whose dimension d θ determines the complexity and flexibility of feature extraction. In the ecological resource management scenario, this vector field can dynamically adjust the focus and direction of feature extraction according to different data types and application requirements. are the weight vectors related to the NSCT approximation coefficients, detail coefficients, and fractional Fourier transform coefficients extracted from the vector field respectively. They weight the coefficients of different transforms according to the key feature extraction direction, highlighting the information related to the key features. The set of scale parameters, s is the scale parameter among them, and different scales represent the observation perspectives of different resolutions of the data. Smaller scales focus on the local details of the data, and larger scales focus on the overall trend of the data. Ecological basic data elements The approximation coefficients and detail coefficients obtained by NSCT at scale s, which reflect the low-frequency and high-frequency feature information of the data at this scale. The weighted approximation coefficients and detail coefficients are weighted by the weight vector related to the key feature extraction direction to highlight the key features. Ecological basic data elements The fractional Fourier transform result of, by adjusting the order α of the fractional Fourier transform, feature information with different time-frequency characteristics can be obtained. Ecological basic data elements The coefficients obtained by the fractional Fourier transform of at scale s. The weighted fractional Fourier transform coefficients are weighted by the weight vector related to the key feature extraction direction to highlight the key features. For the relevant parameters of ecological derived data elements (such as ) which has a similar meaning to the ecological basic data element, except that corresponding transformation and weighting processes are performed on the ecological derivative data. G base , G derived : They are the graph structures constructed for the characteristic information of the ecological basic data element and the ecological derivative data element at different scales, respectively, and describe the association between characteristics through the relationship of vertices and edges. V base , V derived : The graph structure G base and G derived 's vertex set, which contains the characteristic information at different scales. E base , E derived : The edge set of the graph structure G base and G derived . The weight of the edge is calculated by quantum mutual information, which reflects the correlation between characteristics. I Q (x; y): Quantum mutual information, which is used to measure the quantum correlation between two characteristics x and y and is obtained by calculating the von Neumann entropy. S(ρ): The von Neumann entropy of the quantum state ρ, which is used to calculate the quantum mutual information. τ base , τ derived : Threshold, which is used to determine the addition of edges in the graph structure. When the quantum mutual information between two characteristics is greater than this threshold, an edge is added. The learnable attention vector in the graph attention network, which is used to calculate the attention weight of the vertex. W: The weight matrix in the graph attention network, which is used to perform a linear transformation on the vertex feature vector. α v : The attention weight of vertex v in the graph attention network, which reflects the importance of this vertex in feature fusion. |ψ0>: The initial state of the quantum neuron in the quantum neural network. U1, U2, …, U K : Quantum gates in the quantum neural network, which process information and extract features by operating on the state of the quantum neuron. K base : The basic ecological key features obtained through fusion analysis, which contain the characteristic information of the ecological basic data element that is crucial for ecological resource assessment, management, etc. K derived : The derivative ecological key features obtained through fusion analysis, which contain the characteristic information of the ecological derivative data element that is crucial for ecological industry development, ecological service assessment, etc.

[0125] Preferably, in one embodiment, the above specific solutions for respectively extracting the key features of the ecological basic data element and the ecological derivative data element to obtain the basic ecological key features and the derivative ecological key features have the following technical advantages:

[0126] 1. Traditional key feature extraction methods usually adopt fixed algorithms and preset feature directions, lacking adaptability to different data regions and application scenarios. For example, when analyzing ecological basic data, a feature extraction direction based on spatial distribution may be uniformly used, and the data differences in different functional zones (such as core protected areas and buffer zones) cannot be specifically processed, making it difficult to highlight the key features related to ecological protection. In the processing of ecological derivative data, such as ecological tourism data, the focus of feature extraction cannot be dynamically adjusted according to market demands and industrial development trends.

[0127] This solution sets a high-dimensional vector field as the key feature extraction direction. Its flexibility in dimension d θ and the characteristic of changing with the data point x can accurately adapt to the feature extraction requirements of different data regions. In the processing of ecological basic data, according to the functional zones of the ecosystem, for example, in the core protected area, the distribution characteristics of rare species can be highlighted, and in the buffer zone, the characteristics of the ecological transition zone can be emphasized. In terms of ecological derivative data, it can change according to market demands. For example, during the peak tourist season, the correlation characteristics between tourist flow and service quality can be focused on, and during the off-season, the characteristics of cost control and resource maintenance can be concerned, effectively focusing on the key factors affecting the sustainable development of the ecological industry.

[0128] 2. Traditional multi-scale transformation techniques mostly rely on a single transformation method, such as simple wavelet transformation, and cannot fully combine the advantages of different transformations. Moreover, in the processing of ecological data, it is difficult to take into account both spatial information such as edges and textures and time-frequency domain information of the data. For the complex ecological system structure in ecological basic data, such as the multi-level structure of forest vegetation, a single transformation is difficult to comprehensively capture the features at different scales; in ecological derivative data, such as the time distribution and spatial layout characteristics of ecological tourism activities, traditional methods cannot effectively analyze in both the time-frequency domain and the spatial domain.

[0129] This solution adopts a combination of the non-subsampled contourlet transform (NSCT) and the fractional Fourier transform (FrFT). NSCT can effectively capture the edge and texture information of the data, providing support for the spatial feature analysis of ecological basic data, such as clearly showing the boundaries of rivers and the edges of forest patches. The fractional Fourier transform analyzes the data in the time-frequency domain and can reveal the changing laws of ecological data in terms of time and frequency, such as the seasonal and periodic changes in the tourist flow of ecological tourism. At the same time, through weighted processing, different transformation coefficients are adjusted according to the key feature extraction direction, enhancing the highlighting effect of key features. For example, when analyzing forest ecological basic data, for the regions related to the key features of ecological protection, the corresponding coefficient weights are increased to make the key features more prominent at different scales.

[0130] 3. Traditional feature fusion analysis is mostly based on simple linear combinations or correlation analysis, which cannot effectively capture the complex non-linear relationships and quantum-level correlations between features. In ecological data fusion, it is difficult to construct an accurate feature relationship model. For example, when analyzing the correlations between ecological basic data elements and ecological derivative data elements, traditional methods may only consider the superficial linear correlations and ignore the complex interactions within the ecosystem, such as the non-linear impact of eco-tourism activities on the ecological basic environment.

[0131] This solution uses a fusion analysis model based on quantum neural network (QNN) and graph attention network (GAT). By calculating the quantum correlations between features through quantum mutual information to determine the edge weights of the graph structure, it can reveal the deep quantum-level correlations between features, which is impossible for traditional correlation analysis. The graph attention network uses the attention mechanism to calculate the vertex importance weights, which can highlight the role of key features in the fusion. For example, in the fusion of ecological basic data features, it can accurately identify the features that play a key role in ecological resource assessment. The quantum neural network processes the feature information through quantum gate operations, and its quantum characteristics endow the model with stronger information processing and feature extraction capabilities, which can discover feature patterns that are difficult for traditional neural networks to find, and provide more comprehensive and in-depth key feature information for ecological resource management and the development of ecological industries.

[0132] Optionally, the obtaining of green financial data elements and establishing the association relationships between them and the basic ecological key features and derivative ecological key features to generate an ecological resource credible data element model includes:

[0133] Calculating the correlations between green financial data elements and basic ecological key features, derivative ecological key features to construct a correlation matrix;

[0134] Calculating the causalities between green financial data elements and basic ecological key features, derivative ecological key features to construct a causality graph;

[0135] Based on the correlation matrix and the causality graph, constructing an ecological resource credible data element model.

[0136] Preferably, in a specific application scenario, let the set of green financial data elements be where represents the k-th green financial data element, and P is the number of green financial data elements. The set of basic ecological key features is The set of derivative ecological key features is Q and R are the numbers of basic ecological key features and derivative ecological key features respectively. A hybrid method based on the maximum information coefficient (MIC) and kernel canonical correlation analysis (KCCA) is used to calculate the correlations.

[0137] First, calculate the green finance data elements and the key features of the basic ecosystem The maximum information coefficient The maximum information coefficient measures the maximum amount of information transmission between two variables by gridifying the data. The calculation process is as follows: Let x and y be and data sample sets respectively. Divide the value ranges of x and y into m and n intervals respectively to construct an m×n grid. Let p ij be the probability that the data point (x i , y j ) falls into the grid in the i-th row and j-th column. p i · and p ·j are the marginal probabilities of the i-th row and j-th column respectively. Then the mutual information I(x; y) is:

[0138] The maximum information coefficient is defined as the maximum value of the ratio of the mutual information I(x; y) to log2min(m, n) among all possible grid partitions, that is: where N is the number of data samples, and B is a preset bandwidth parameter, usually taken as 0.6. At the same time, kernel canonical correlation analysis is used to calculate the canonical correlation coefficient between the green finance data elements and the key features of the basic ecosystem Let X and Y be and data matrices respectively. Map the data to a high-dimensional feature space through the kernel function κ to obtain Φ(X) and Φ(Y). Kernel canonical correlation analysis aims to find projection vectors a and b such that the correlation between the projected variables u = a T Φ(X) and v = b T Φ(Y) is the largest. Define the covariance matrix Then the canonical correlation coefficient ρ is the solution of the following generalized eigenvalue problem:

[0139] Finally, by integrating the maximum information coefficient and the canonical correlation coefficient, the correlation measure between the green finance data elements and the key features of the basic ecosystem is obtained where λ1 and λ2 are weight coefficients determined by cross-validation.

[0140] Similarly, calculate the correlation measure between the green finance data elements and the key features of the derivative ecosystem ​

[0141] Construct the correlation matrix M r , and its elements are as follows: where k = 1, …, P, q = 1, …, Q, r = 1, …, R.

[0142] Preferably, in a specific application scenario, a method based on structural causal models (SCMs) and Bayesian structure learning is used to calculate causality. When constructing a structural causal model, for the green finance data element D g , the basic ecological key feature K b and the derived ecological key feature K d , there is a causal relationship, which can be expressed as: D g = f g (K b , K d , ∈ g ), K b = f b (∈ b ), K d = f d (K b , ∈ d ), where f g , f b , f d are unknown causal functions, and ∈ g , ∈ b , ∈ d are independent noise variables.

[0143] Use the Bayesian structure learning algorithm to determine the causal structure. Let be a directed acyclic graph (DAG) representing the causal relationships between variables, with nodes being the variables in D g , K b , K d , and the edges representing the causal directions. The goal of Bayesian structure learning is to find an optimal such that the posterior probability is maximized, where D is the observed data. According to Bayes' theorem, the prior probability can be set based on some prior knowledge or a uniform distribution. The likelihood function is obtained by modeling the conditional probability distribution of each node. For each node X, its conditional probability distribution P(X|Pa(X)) (where Pa(X) is the set of parents of X) can be represented by a parametric model, such as a Gaussian distribution (for continuous variables) or a multinomial distribution (for discrete variables). By maximizing the posterior probability the optimal causal structure i.e., the causal relationship graph, can be obtained.

[0144] Preferably, in a specific embodiment, in the link of constructing an ecological resource credible data element model based on a correlation matrix and a causal relationship diagram, a method based on tensor decomposition and a deep belief network (DBN) is used to construct the ecological resource credible data element model, which is specifically as follows:

[0145] Regard the correlation matrix M r as a third-order tensor where one dimension corresponds to green finance data elements, one dimension corresponds to basic ecological key features, and the other dimension corresponds to derivative ecological key features. Use higher-order singular value decomposition (HOSVD) to decompose the tensor as follows:

[0146] where is the core tensor, U 1 , U 2 , U 3 are orthogonal matrices, and x i represents the tensor product along the i-th mode.

[0147] Take the decomposed core tensor and the causal relationship diagram as the input of the deep belief network (DBN). The deep belief network is stacked by multiple restricted Boltzmann machines (RBMs). Let the input of the first RBM be x 1 , the number of visible layer nodes is equal to the number of elements of the core tensor , and the number of hidden layer nodes is H1. The energy function of the RBM is: where h 1 is the hidden layer state, is the weight between the visible layer node i and the hidden layer node j, and are the biases of the visible layer and the hidden layer respectively.

[0148] Train the RBM through the contrast divergence algorithm to obtain the weight matrix W 1 and the bias vectors b 1 , c 1 . Take the hidden layer output of the first RBM as the input of the next RBM, and train multiple RBMs in turn to finally obtain the parameters of the deep belief network.

[0149] The output of the trained deep belief network is the ecological resource credible data element model This model can output the evaluation and prediction of the ecological resource state according to the input green finance data elements, basic ecological key features, and derivative ecological key features.

[0150] In the above embodiments, the creative physical meanings of each parameter are described as follows: D g: The set of green finance data elements includes data related to green finance, such as green investment amounts, green credit interest rates, etc. These data reflect the support and impact of the financial field on ecological resources. K b : The set of basic ecological key features includes the key features related to the natural ecological foundation obtained after key feature extraction, such as biodiversity indicators of forest ecosystems, key parameters of water flow and water quality of water resources, etc. These features play a key role in the stability and function of the ecosystem. K d : The set of derivative ecological key features includes the key features related to ecological derivative fields such as eco-tourism and leisure and health care, such as key factors of tourist satisfaction in eco-tourism, demand characteristics of leisure and health care services, etc. These features reflect the value and application of ecological resources in derivative industries. P: The quantity of green finance data elements, which reflects the richness and dimension of green finance data. Q: The quantity of basic ecological key features, which reflects the scale of key features after screening and extraction of ecological basic data. R: The quantity of derivative ecological key features, which demonstrates the diversity of key features in ecological derivative data. Green finance data elements And basic ecological key features The maximum information coefficient between them is used to measure the maximum amount of information transmission between two variables and is obtained through gridded data and calculation of mutual information. m, n: The number of intervals for dividing the value ranges of data samples x and y when calculating the maximum information coefficient, which affects the calculation accuracy and complexity of the maximum information coefficient. p ij , p i . p .j : When calculating mutual information, the probabilities of data points falling into the grid and marginal probabilities are the basic parameters for calculating mutual information. N: The number of data samples, which is used to calculate covariance matrices in the calculation of the maximum information coefficient and kernel canonical correlation analysis, etc. B: The bandwidth parameter in the calculation of the maximum information coefficient, which controls the density of grid division, usually taking 0.6, and affects the calculation result of the maximum information coefficient. X, Y: In kernel canonical correlation analysis, the data matrices of green finance data elements And basic ecological key features Are the input data for calculating the canonical correlation coefficient. κ: The kernel function, which is used to map data into a high-dimensional feature space to enhance the separability and correlation analysis ability of data. Common kernel functions include Gaussian kernel, polynomial kernel, etc. C xx , C yy , C xy:The covariance matrix in kernel canonical correlation analysis, which is used to calculate the canonical correlation coefficients. ρ: The canonical correlation coefficients obtained from kernel canonical correlation analysis, which measure the correlation between two variables after projection in the high-dimensional feature space. λ1, λ2: The weight coefficients when calculating the correlation measure by integrating the maximum information coefficient and the canonical correlation coefficient. They are determined by cross-validation and used to balance the contributions of the two methods to the correlation measure. Green finance data elements And the key features of the basic ecosystem The comprehensive correlation measure, which integrates the results of the maximum information coefficient and the canonical correlation coefficient. Green finance data elements And the key features of the derivative ecosystem The comprehensive correlation measure, and the calculation method is similar to that of r kb Similar. M r : The correlation matrix, which stores the correlation measure values between green finance data elements and the key features of the basic ecosystem and the key features of the derivative ecosystem, and is an important input for subsequent model construction. f g , f b , f d : The unknown causal function in the structural causal model, which describes the causal relationship between green finance data elements, the key features of the basic ecosystem and the key features of the derivative ecosystem. ∈ g , ∈ b , ∈ d : The independent noise variables in the structural causal model, which reflect the random factors not explained in the model. Directed acyclic graph, which is used to represent the causal relationship between variables. The nodes are the variables in green finance data elements, the key features of the basic ecosystem and the key features of the derivative ecosystem, and the edges represent the causal direction. In Bayesian structure learning, when given the observed data D, the posterior probability of the causal structure is the objective function for finding the optimal causal structure. Likelihood function, which is obtained by modeling the conditional probability distribution of each node and reflects the generation probability of the observed data D under the given causal structure . Causal structure The prior probability of can be set according to prior knowledge or uniform distribution, which affects the result of Bayesian structure learning. The optimal causal structure obtained through Bayesian structure learning, that is, the causal relationship graph, shows the causal relationship between green finance data elements, the key features of the basic ecosystem and the key features of the derivative ecosystem. Regarding the correlation matrix M r as a third-order tensor for higher-order singular value decomposition to extract the latent structure and features in the data. The core tensor obtained by high-order singular value decomposition retains the tensor Main information and features of U 1 , U 2 , U 3 : An orthogonal matrix in a high-order singular value decomposition, used to transform and decompose tensors. × i : The tensor product along the i-th mode is an operation in high-order singular value decomposition. 1 : The input to the first restricted Boltzmann machine (RBM) in the deep belief network, whose dimension is equal to the core tensor H1: The number of nodes in the first RBM hidden layer, which affects the learning ability and feature extraction ability of RBM. E(x 1 ,h 1 ): The energy function of RBM, which is used to define the joint probability distribution of the visible and hidden layer states in RBM.

[0151] Optionally, the above-mentioned specific technical process of obtaining green finance data elements and establishing the association between them and the basic ecological key features and the derived ecological key features to generate an ecological resource credible data element model has significant advantages over traditional technologies in the scenario of ecological resource management and green finance association analysis:

[0152] 1. Traditional methods often rely on simple Pearson correlation coefficients, which can only measure linear relationships and are powerless against complex nonlinear relationships between ecological and financial data. For example, when analyzing green investment and forest ecosystem biodiversity, it is impossible to explore the complex dependencies hidden behind the data, which may miss key information and lead to a one-sided assessment of the correlation between the two.

[0153] This solution combines the maximum information coefficient (MIC) and kernel canonical correlation analysis (KCCA). MIC calculates mutual information through gridded data, which can capture various types of associations between variables, including nonlinear relationships, and comprehensively measure the amount of information transmission between green finance and key ecological characteristics. KCCA uses kernel functions to map data to high-dimensional space and mine deep correlations in data. Combining the two, such as when evaluating the relationship between green credit interest rates and key parameters of water resource flow and water quality, it can accurately identify complex associations, providing a more comprehensive and accurate correlation basis for ecological resource management and green financial decision-making.

[0154] 2. Traditional causal analysis is mostly based on regression models, which makes it difficult to distinguish between causal and correlation relationships, and can easily lead to wrong conclusions in complex ecological financial systems. For example, the correlation between increased satisfaction of ecotourism tourists and increased green investment may be misjudged as a causal relationship, and the inherent causal logic of ecological industry and financial support cannot be accurately revealed.

[0155] This application is based on structural causal models (SCMs) and Bayesian structure learning to clarify the causal relationships between variables and determine the optimal causal structure by maximizing the posterior probability through Bayes' theorem. It can consider various noise factors and establish appropriate conditional probability distributions for different types of variables. When analyzing the causal relationships among green finance data elements, key features of the basic ecosystem, and key features of the derived ecosystem, it can clearly present causal paths such as how green investment directly or indirectly affects the development of the eco-tourism industry and the feedback of changes in the ecological foundation on green finance policies, providing a scientific basis for policy-making and resource management.

[0156] Traditional model construction methods have weak processing capabilities for high-dimensional and multi-source data and are difficult to integrate ecological foundation, derived features, and green finance data. When constructing an ecological resource assessment model, it may only consider a single data source or simple associations, resulting in poor model prediction and evaluation capabilities and being unable to adapt to the complex and ever-changing actual situation of ecological finance.

[0157] This application uses tensor decomposition and deep belief networks (DBNs). The correlation matrix is regarded as a third-order tensor for high-order singular value decomposition to extract the latent structural features of the data and retain the complex relationships between multi-dimensional data. The DBN is stacked by multiple restricted Boltzmann machines and can effectively learn the patterns of the input data. Taking the construction of a credible data element model for ecological resources as an example, it can fully integrate the information of green finance, ecological foundation, and derived key features, accurately evaluate the state of ecological resources, and predict future trends, providing strong model support for ecological resource management and the formulation of green finance investment strategies.

[0158] The above description is only a preferred embodiment of this application and an explanation of the technical principles applied. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by the specific combination of the above technical features, but also covers other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for establishing an intelligent and trustworthy data element model of ecological resources, characterized in that, Including: Obtain natural ecological basic data, where the natural ecological basic data includes at least one of forest description metadata, river description metadata, lake description metadata, grassland description metadata, and cultivated land description metadata; Obtain natural ecological derivative data, where the natural ecological derivative data includes ecological tourism description metadata and leisure and health care description metadata; Respectively conduct controlled feature mining on the natural ecological basic data and the natural ecological derivative data to generate ecological basic data elements and ecological derivative data elements; Respectively extract key features from the ecological basic data elements and the ecological derivative data elements to obtain basic ecological key features and derivative ecological key features; Obtain green financial data elements and establish an association relationship between them and the basic ecological key features and the derivative ecological key features to generate an ecological resource trustworthy data element model.

2. The method according to claim 1, wherein The obtaining of the natural ecological basic data includes: Based on the constructed multi-source data collection rules, analyze the carriers carrying different types of natural ecological basic data to obtain the natural ecological basic data therefrom.

3. The method according to claim 1, characterized in that The obtaining of the natural ecological derivative data includes: Perform structured processing on the natural ecological derivative description text to obtain a structured text; Perform semantic analysis on the structured text to extract natural ecological derivative feature elements; Based on the natural ecological derivative feature elements, obtain the natural ecological derivative data.

4. The method according to claim 1, characterized in that The respectively conducting controlled feature mining on the natural ecological basic data and the natural ecological derivative data to generate ecological basic data elements and ecological derivative data elements includes: According to the set business rules and feature mining constraint conditions, construct a controlled feature mining model; Based on the controlled feature mining model, conduct controlled feature mining on the natural ecological basic data and the natural ecological derivative data to generate ecological basic data elements and ecological derivative data elements.

5. The method according to claim 1, characterized in that, The respectively extracting key features from the ecological basic data elements and the ecological derivative data elements to obtain basic ecological key features and derivative ecological key features includes: Based on the set key feature extraction direction, perform multi-scale transformation on the ecological basic data elements and the ecological derivative data elements to obtain the feature information of the ecological basic data elements and the ecological derivative data elements at different scales; Perform fusion analysis on the feature information of the ecological basic data elements and the ecological derivative data elements at different scales to obtain the basic ecological key features and the derivative ecological key features.

6. The method according to claim 1, wherein The obtaining of the green financial data elements and establishing an association relationship between them and the basic ecological key features and the derivative ecological key features to generate an ecological resource trustworthy data element model includes: Calculate the correlation between the green financial data elements and the basic ecological key features and the derivative ecological key features to construct a correlation matrix; Calculate the causality between the green financial data elements and the basic ecological key features and the derivative ecological key features to construct a causal relationship diagram; Based on the correlation matrix and the causal relationship diagram, construct an ecological resource trustworthy data element model.

Images