Ecological asset dynamic monitoring and evaluation method

Abstract

The invention provides an ecological asset dynamic monitoring and evaluation method. The method comprises the following steps: dynamically collecting attribute features of target ecological assets based on an ecological asset dynamic monitoring model, and forming a dynamic attribute feature sequence according to the attribute features; constructing a dynamic attribute knowledge graph based on the dynamic attribute feature sequence; and based on the ecological asset evaluation model, performing dynamic asset value evaluation on the target ecological assets according to the dynamic attribute knowledge graph to generate dynamic value sequence points and adding the dynamic value sequence points to a dynamic value tree. According to the method, the attribute change of the ecological assets in time and space dimensions is comprehensively captured, the change of the ecological assets can be reflected in real time through a dynamic construction mode, dynamic asset value evaluation is performed based on the ecological asset evaluation model and according to the dynamic attribute knowledge graph, and the evaluation efficiency is improved. The dynamic change characteristics of the ecological asset value along with time, environment and various factors can be fully considered, the scientificity and practicability of ecological asset value evaluation are remarkably improved, and powerful support is provided for ecological protection and management decision.

Description

Technical Field

[0001] The present invention relates to the field of intelligent processing technologies, and particularly to a method for dynamically monitoring and evaluating ecological assets. Background Art

[0002] In the field of dynamic monitoring and evaluation of ecological assets, with the continuous improvement of the requirements for ecological protection and resource management, traditional technologies face many bottlenecks.

[0003] Currently, most ecological asset monitoring methods rely on a combination of a dispersed sensor network and manual inspections. Although the sensor network can collect basic environmental data, it lacks a systematic data fusion and analysis mechanism. The data from different types of sensors are independent of each other, making it difficult to construct an overall picture of ecological assets. Manual inspections have problems such as low efficiency, strong subjectivity, and limited coverage, and cannot achieve real-time, dynamic, and comprehensive monitoring of the attribute characteristics of ecological assets. In the data processing link, existing technologies usually use conventional data mining algorithms to analyze monitoring data. These algorithms are mostly based on the surface features and simple correlation relationships of the data, and it is difficult to mine the complex non-linear relationships and potential laws in the ecosystem, and cannot effectively cope with the high-dimensional, multi-source, and dynamic characteristics of ecological data.

[0004] In the construction of ecological asset knowledge graphs, traditional methods often adopt manual entry or automated extraction based on fixed rules. Manual entry is extremely inefficient and prone to human errors, and it is difficult to meet the processing requirements of large-scale ecological data. Automated extraction based on fixed rules depends on pre-set templates and patterns. Facing the diverse semantic expressions and complex professional terms in the ecological field, it is difficult to accurately extract entity, relationship, and attribute information, and the constructed knowledge graph lacks flexibility and dynamic update capabilities and cannot reflect the changes in ecological assets in a timely manner.

[0005] In the aspect of ecological asset value assessment, existing assessment technologies mostly adopt static assessment index systems and fixed assessment models. For example, the economic value of ecological assets is evaluated by the market value method, and the ecological service function value is estimated by the replacement cost method. However, these methods do not fully consider the dynamic evolution process of the ecosystem and ignore the characteristics of the value of ecological assets changing with time, environment, and human activities. At the same time, the weight allocation of various influencing factors in the assessment process is often based on subjective experience or simple statistics of historical data, and it is impossible to accurately quantify the complex relationship between the value of ecological assets and various factors, resulting in the lack of timeliness and accuracy of the assessment results and making it difficult to meet the actual needs of dynamic management and scientific decision-making of ecological assets. Summary of the Invention

[0006] In order to solve the above technical problems, the present application provides a method for dynamically monitoring and evaluating ecological assets to at least solve or alleviate the problems existing in the above prior art.

[0007] To achieve the above object, according to one aspect of the present application, there is provided an ecological asset dynamic monitoring and evaluation method, which includes:

[0008] An ecological asset dynamic monitoring and evaluation method, characterized by including:

[0009] Dynamically collect the attribute characteristics of the target ecological assets based on the ecological asset dynamic monitoring model and form a dynamic attribute characteristic sequence accordingly;

[0010] Based on the dynamic attribute characteristic sequence, construct a dynamic attribute knowledge graph;

[0011] Based on the ecological asset evaluation model, dynamically evaluate the value of the target ecological assets according to the dynamic attribute knowledge graph to generate dynamic value sequence points and add them to the dynamic value tree.

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

[0013] ① Traditional monitoring means are scattered and lack systematicness. By constructing an ecological asset dynamic monitoring model in the present application, it is possible to dynamically collect the attribute characteristics of the target ecological assets and form a dynamic attribute characteristic sequence. This model can comprehensively capture the attribute changes of ecological assets in the time and space dimensions. Compared with the traditional scattered and isolated monitoring methods, it realizes the dynamic and comprehensive collection of the attribute characteristics of ecological assets, effectively avoids data missing and one-sidedness, significantly improves the accuracy and integrity of monitoring, and provides a reliable data basis for subsequent evaluation.

[0014] ② Traditional knowledge graph construction relies on manual entry or extraction based on fixed rules, with low efficiency and inaccuracy. In the present application, a dynamic attribute knowledge graph is constructed based on the dynamic attribute characteristic sequence, which can automatically and accurately extract entity, relationship, and attribute information related to ecological assets. Through the dynamic construction method, it can reflect the changes of ecological assets in real time. Compared with the traditional static and passive knowledge graph construction methods, it has stronger flexibility and adaptability, and can quickly and accurately integrate ecological asset information to form a complete and dynamic knowledge graph system.

[0015] ③The existing ecological asset value assessment uses a static index system and a fixed model, which cannot reflect the dynamic changes in value. Based on the ecological asset assessment model, this application conducts dynamic asset value assessment according to the dynamic attribute knowledge graph, and can fully consider the dynamic change characteristics of ecological asset value over time, environment, and various factors. By generating dynamic value sequence points and adding them to the dynamic value tree, it intuitively displays the dynamic evolution process of ecological asset value and its causal relationship, provides a scientific, real-time, and dynamic value assessment result for ecological resource management, effectively solves the problems of lag, lack of timeliness, and inaccuracy in traditional assessment results, significantly improves the scientificity and practicality of ecological asset value assessment, and provides strong support for ecological protection and management decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flowchart of a method for dynamically monitoring and assessing ecological assets according to an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] As Figure 1 shown, a method for dynamically monitoring and assessing ecological assets provided by an embodiment of this application includes:

[0018] Dynamically collect the attribute characteristics of the target ecological asset based on the ecological asset dynamic monitoring model and form a dynamic attribute characteristic sequence accordingly;

[0019] Construct a dynamic attribute knowledge graph based on the dynamic attribute characteristic sequence;

[0020] Based on the ecological asset assessment model, conduct dynamic asset value assessment on the target ecological asset according to the dynamic attribute knowledge graph to generate dynamic value sequence points and add them to the dynamic value tree.

[0021] Preferably, in a scenario, the above technical implementation is described in a replaceable or preferred manner.

[0022] 1. Dynamically collect the attribute characteristics of the target ecological asset based on the ecological asset dynamic monitoring model and form a dynamic attribute characteristic sequence

[0023] 1.1 Construction of the ecological asset dynamic monitoring model

[0024] Let the target ecological asset be a special wetland ecosystem. At time step t (t = 1, 2,..., T), for the i-th attribute characteristic x i (t), its calculation comprehensively considers multi-scale spatio-temporal factors, autocorrelation effects, and external disturbances.

[0025]

[0026] Among them: K represents the number of categories of different influencing factors, such as including local ecological environment, regional climate patterns, different levels of human activities, etc. α ik is the global weight coefficient of the k-th type of influencing factor on the i-th type of attribute characteristic, and 0 < α ik < 1, which is determined by a method combining the analytic hierarchy process (AHP) and principal component analysis (PCA) through multiple rounds of iteration to comprehensively consider the relative importance of each factor. S is the spatial range of the ecosystem, and s represents a point in space. β ik (s) is the local weight function of the k-th type of influencing factor on the i-th type of attribute characteristic at the spatial point s. It is a spatial weight distribution based on the Gaussian kernel function, reflecting the influence differences at different spatial positions. Among them, μ ik is the spatial center position where the k-th type of influencing factor has the greatest influence on the i-th type of attribute characteristic, and σ ik is the standard deviation controlling the spatial influence range. γ ik (t - τ ik (s)) is the time lag function, and τ ik (s) is the time lag amount of the k-th type of influencing factor on the i-th type of attribute characteristic at the spatial point s, which is determined according to the autocorrelation analysis and Granger causality test of historical data. γ ik (t - τ ik (s)) adopts the exponential decay function γ ik (t - τ ik (s)) = exp(-λ ik |t - τ ik (s)|), where λ ik is the decay coefficient. f ik is a special non-linear function, which describes the relationship between the k-th type of influencing factor on the i-th type of attribute characteristic and the ecosystem state S(t), external environmental factors E(t), and human activity factors H(t) at the spatial point s. It is trained using a deep convolutional neural network (DCNN) combined with a long short-term memory network (LSTM). The inputs are S(t), E(t), H(t), and s, and the output is the influence value on x i (t). ∈ i (t) is the random error term, which follows a mixed normal distribution where L is the number of mixed components, and ω il is the weight of the l-th mixed component, used to describe the comprehensive influence of measurement errors from multiple sources and unconsidered factors.

[0027] 1.2 Formation of dynamic attribute characteristic sequences

[0028] Suppose there are a total of n types of attribute characteristics, and the attribute characteristic vector at time t is The dynamic attribute feature sequence is {X(1), X(2), …, X(T)}. To consider the correlation and dynamic change trend among attribute features, the present application further processes the sequence to obtain a weighted dynamic attribute feature sequence. Among them: W(t) is a time-varying weight matrix of n×n, and its element w pq (t) represents the influence weight of the p-th attribute feature on the q-th attribute feature at time t. w pq (t) is learned through a dynamic Bayesian network (DBN) and dynamically adjusted according to the causal relationship and time dependence relationship among attribute features.

[0029] 2. Construct a dynamic attribute knowledge graph based on the dynamic attribute feature sequence

[0030] 2.1 Entity recognition and classification

[0031] Entities are recognized from the dynamic attribute feature sequence, and a bidirectional gated recurrent unit with attention mechanism (BiGRUAttention) model is adopted. Suppose the input dynamic attribute feature sequence is For the input at time step t After passing through the BiGRU layer, the hidden state h t is obtained:

[0032] Then, the attention weight α at each time step is calculated through the attention mechanism t :

[0033] where W a , b a and v are learnable parameters. Finally, the weighted hidden state is obtained. Entity classification is performed through a fully connected layer: where y is the probability distribution of entity classification, and W c and b c are learnable parameters.

[0034] 2.2 Relationship extraction

[0035] To determine the semantic relationship between entities, a joint model based on a graph convolutional network (GCN) and a multi-head attention mechanism is adopted. Suppose the entity set is E = {e1, e2, …, e m}, the feature vector corresponding to each entity e j is f j , and an adjacency matrix A is constructed to represent the initial connection relationship between entities. After passing through the GCN layer, the updated entity feature vector f j′ is obtained:

[0036] Among them, W gcn and b gcn are learnable parameters of the GCN layer. Then, the relationship strength between entities is calculated through the multi-head attention mechanism:

[0037] Among them, Q = W q F′, K = W k F′, V = W v F′, F′ = [f 1′ , f 2′ , …, f m′ T , W q , W k and W v are learnable parameters, and d k is the dimension of the key vector. After multi-head attention, the relationship strength matrix R is obtained, and its element r jk represents the relationship strength between entities e j and e k .

[0038] 2.3 Knowledge Graph Construction

[0039] Integrate the identified entities and the extracted relationships into the knowledge graph. The knowledge graph is represented as a graph G = (V, E), where V is the set of nodes corresponding to entities; E is the set of edges corresponding to the relationships between entities. To consider the dynamics of the knowledge graph, this application introduces a time dimension to obtain a dynamic knowledge graph G(t) = (V(t), E(t)), where V(t) and E(t) change dynamically with time t. The attribute updates of nodes and edges are optimized through a reinforcement learning algorithm to adapt to the dynamic changes of the ecological asset attribute characteristics.

[0040] 3. Based on the ecological asset evaluation model, dynamically evaluate the target ecological asset according to the dynamic attribute knowledge graph to generate dynamic value sequence points and add them to the dynamic value tree

[0041] 3.1 Construction of the Ecological Asset Evaluation Model

[0042] Suppose the total value V(t) of the target ecological asset consists of values at multiple levels, and a hierarchical weighted evaluation model is adopted: Among them: L is the number of value levels, for example, divided into levels such as direct economic value, indirect ecological service value, and potential value. ω l (t) is the weight of the l-th value level at time t, which is dynamically adjusted through the dynamic analytic hierarchy process (DAHP) combined with expert opinions and real-time data. m l is the number of value components under the l-th value level. w il ​(t) is the weight of the i-th value component at the l-th value level at time t, determined by a combined weighting method based on the entropy weight method and grey relational analysis. v il (t) is the value of the i-th value component at the l-th value level at time t, and its calculation takes into account the dynamic attribute knowledge graph G(t) and the attribute feature sequence where η il (τ) is the time discount function, and the hyperbolic discount function is adopted k il is the discount coefficient, determined according to the time sensitivity of different value components. g il is a special non-linear function, trained by the Deep Deterministic Policy Gradient (DDPG) algorithm in deep reinforcement learning, with the input being G(τ) and the output being v il (τ).

[0043] 3.2 Generation of dynamic value sequence points and addition to the dynamic value tree

[0044] At each time step t, the calculated total value V(t) constitutes a dynamic value sequence point. The dynamic value tree in this application is a multi-way tree structure T = (N, E T ), where N is the set of nodes, and each node n corresponds to the value V(t) at a time step and related attribute information (such as value composition, influencing factors, etc.); E T is the set of edges, and the edges represent the chronological order in time and the value transfer relationship. Add the dynamic value sequence point V(t) at each time step and its related information to the dynamic value tree, use the tree insertion algorithm, and update the tree structure and node attributes simultaneously to reflect the dynamic changes and causal relationships of the ecological asset value.

[0045] For this reason, the above replaceable or preferred technical implementations have the following technical advantages.

[0046] 1. Traditional ecological asset monitoring mostly uses simple linear regression models or static analysis methods with a single data source. These methods often only consider a few factors and ignore the spatio-temporal complexity and dynamic changes of the ecosystem. For example, predicting vegetation growth based only on average temperature and precipitation data for a certain period without considering local environmental differences in different regions and time lag effects. In addition, traditional methods handle measurement errors and unconsidered factors relatively simply, usually assuming a single normal distribution and unable to accurately describe special error sources.

[0047] In this application, the spatial weight function β ik (s) and the time lag function γ ik (t - τ ik(s)) can comprehensively consider the impacts of different spatial positions and time delays on the attribute characteristics of ecological assets. In wetland ecosystems, local environmental factors such as hydrological conditions and soil types vary greatly in different regions. The spatial weight function can more accurately capture these differences. At the same time, there is often a time lag in the response of the ecosystem. The time lag function reflects this delay effect and improves the accuracy of monitoring. Use a special non - linear function f ik to describe the relationship between the state of the ecosystem, external environmental factors, and human activity factors. Through training with a deep convolutional neural network (DCNN) combined with a long short - term memory network (LSTM), it can learn the complex non - linear relationships between these factors. Compared with traditional linear models, it is more in line with the actual operation rules of the ecosystem. The random error term ∈ i (t) follows a mixed normal distribution, which can more accurately describe the combined effects of measurement errors from multiple sources and unconsidered factors, improving the robustness of the model.

[0048] 2. Traditional entity recognition and relation extraction methods are mostly based on rules or simple machine - learning algorithms, such as the Naive Bayes classifier and the Support Vector Machine. These methods have limited capabilities for special text and semantic understanding and are difficult to handle the rich professional terms and special semantic relationships in the ecological field. At the same time, traditional methods often ignore the time - series information and context information of the data, resulting in lower accuracy of entity recognition and relation extraction.

[0049] In this application, a bidirectional gated recurrent unit with attention mechanism (BiGRUAttention) model is used for entity recognition. The attention mechanism can automatically focus on the important parts of the input sequence. When processing the dynamic attribute feature sequence of ecological assets, it can better capture the key information at different time steps and improve the accuracy of entity recognition. For example, when identifying different biological species entities in a wetland ecosystem, it can classify them more accurately according to their dynamic characteristics such as growth cycle and environmental adaptability. A joint model based on graph convolutional network (GCN) and multi - head attention mechanism is used for relation extraction. GCN can effectively process graph - structured data, representing the entities and relationships in the ecological asset knowledge graph as a graph structure and updating the feature representations of entities through graph convolution operations. The multi - head attention mechanism can capture the relationship strength between entities from multiple perspectives, improving the accuracy and robustness of relation extraction. In a wetland ecosystem, it can more accurately identify complex relationships such as competition and symbiosis between biological species.

[0050] 3. Traditional ecological asset assessment models mostly adopt static assessment methods, such as the market value method, replacement cost method, etc. These methods often only consider the current state and direct economic value of ecological assets, ignoring the dynamic changes of the ecosystem and the indirect ecological service value. At the same time, traditional methods often use fixed weights and simple linear combinations for value calculation, unable to reflect the dynamic characteristics of ecological asset value changing over time and with the environment. In this application, through a hierarchical weighted assessment model is implemented, which divides the value of ecological assets into multiple levels, such as direct economic value, indirect ecological service value, and potential value, etc., and dynamically adjusts the weights of each level and each component through the dynamic analytic hierarchy process (DAHP) and combined weighting method. This method can more comprehensively consider the value composition of ecological assets, and dynamically adjust the weights according to different time and environmental conditions, improving the accuracy and flexibility of the assessment. The calculation of the value component v il (t) takes into account the dynamic attribute knowledge graph G(t) and the attribute feature sequence and introduces the time discount function η il (τ). The non-linear function g trained by the deep deterministic policy gradient (DDPG) algorithm in deep reinforcement learning il can learn the complex relationships between the value of ecological assets and various factors. The time discount function can reflect the discount effect of future value, which is more in line with the actual economic decision-making needs.

[0051] 4. Traditional ecological asset assessment results are usually presented in the form of static reports or simple numerical values, unable to intuitively display the dynamic changes and causal relationships of ecological asset value over time. When making decisions, it is difficult for decision-makers to obtain comprehensive information from these static results, nor can they conduct effective prediction and analysis.

[0052] In this application, since the dynamic value tree is a multi-way tree structure, each node corresponds to the value at a time step and related attribute information, and the edges represent the chronological order and value transfer relationship in time. By adding the dynamic value sequence points and their related information to the dynamic value tree, the dynamic change process and causal relationships of ecological asset value can be intuitively displayed. Decision-makers can traverse the dynamic value tree to understand the value composition, influencing factors, and value change trends at different time points, so as to make more scientific and reasonable decisions. For example, in the protection and management of wetland ecosystems, decision-makers adjust the protection strategy and resource allocation according to the information of the dynamic value tree to maximize the value of ecological assets.

[0053] Optionally, before dynamically monitoring the attribute characteristics of the target ecological asset based on the ecological asset dynamic monitoring model and forming a dynamic attribute feature sequence accordingly, it includes:

[0054] Obtain the attribute feature carrier collected by the attribute sensor, and perform feature extraction on the attribute feature carrier to obtain the attribute features of the target ecological asset.

[0055] Preferably, in a scenario, the above technical implementation is described in a replaceable or preferred manner.

[0056] 1. Obtain the attribute feature carrier collected by the attribute sensor

[0057] 1.1 Attribute sensor and data collection

[0058] In ecological asset monitoring, multiple types of attribute sensors work together to collect data. Taking the marine ecosystem as an example, the sensors involved are: water quality sensors: used to measure water quality parameters such as pH (acidity), dissolved oxygen (DO), chemical oxygen demand (COD), salinity, etc. Biological sensors: monitor biological-related indicators such as plankton quantity and chlorophyll concentration. Meteorological sensors: collect meteorological information such as sea surface temperature, wind speed, wind direction, and light. These sensors collect data at different positions and depths, and the data is recorded in the form of a time series and stored in a distributed database. Each data record contains the sensor ID, the collection time t, the collection position p = (x, y, z) (three-dimensional coordinates), and the measured value v.

[0059] 1.2 Temporal and spatial distribution of the attribute feature carrier

[0060] Let S be the set of sensors, and s ∈ S represents a sensor. For the sensor s, the data collected at time t and position p constitutes an attribute feature carrier d(s, t, p). Considering the spatio-temporal characteristics of the sensor network, there is spatio-temporal correlation in data collection. This application uses the spatio-temporal covariance function C(t1, p1, t2, p2) to describe this correlation:

[0061] where: σ 2 is the overall variance of the data. θ t and θ p are the smoothing parameters for time and space respectively, controlling the attenuation rate of spatio-temporal correlation. λ t and λ p are the scale parameters for time and space respectively, determining the range of spatio-temporal correlation.

[0062] 2. Perform feature extraction on the attribute feature carrier

[0063] 2.1 Feature extraction model

[0064] Let D be the set of attribute feature carriers. For the i-th type of attribute feature x i , its feature extraction process is expressed as:

[0065]

[0066] Where: S is the set of sensors. t' is the historical time, from 1 to the current time t. Ω is the set of positions within the monitoring area. α i (s, t′, p) is the weight coefficient of the i-th attribute feature at sensor s, time t', and position p, which is dynamically determined by spatio-temporal weighted regression (STWR) combined with the Kalman filtering algorithm. STWR takes into account the spatio-temporal non-stationarity of the data, and the Kalman filter is used to update the weights in real time. f i is a special feature transformation function that takes into account the collected data d(s, t′, p), the time difference t - t′, and the position distance ||p - p0|| (p0 is the position of the target monitoring point). Taking the extraction of biodiversity features of the marine ecosystem as an example, f i is a convolutional neural network recurrent neural network (CNNRNN) model based on deep learning. The input is the biological data collected by different sensors at different spatio-temporal points, and the output is the estimated value of biodiversity. ∈ i (t) is the random error term, which follows a mixture Gaussian distribution where K is the number of mixture components, ω ik is the weight of the k-th mixture component, used to describe errors from multiple sources, such as sensor noise, model uncertainty, etc.

[0067] 2.2 Feature Transformation Function

[0068] Taking the biodiversity feature extraction function f based on CNNRNN as an example: First, the CNN part is used to extract spatial features. Let the input biological data matrix be X i (data from sensor s at time t'), and after passing through l convolutional layers, the feature map is obtained s,t′ (from sensor s at time t'), and after passing through l convolutional layers, the feature map is obtained Where: W l is the convolutional kernel of the l-th layer, and * represents the convolution operation. b l is the bias term. ReLU is the activation function, ReLU(x) = max(0, x). Then, the feature map output by the CNN is flattened and input into the RNN part. The update formula for the hidden state of the RNN is: h s,t = tanh(W h h s,t′-1 + W x F s,t′ + b h ), where: h s,t′ is the hidden state at time t'; W h and W xThey are the weight matrices of the hidden state and the input, respectively. b h is the bias term. Finally, the estimated value of biodiversity is obtained through the fully connected layer: f i (d(s, t′, p), t - t′, ||p - p0||) = W o h s,t′ + b o , where: W o is the weight matrix of the output layer. b o is the bias term.

[0069] 3. Obtain the attribute characteristics of the target ecological assets

[0070] Through the above feature extraction process, this application obtains n attribute characteristics of the target ecological assets at time t, forming an attribute feature vector:

[0071]

[0072] The dynamic attribute feature sequence is {X(1), X(2), …, X(T)}, where T is the total number of monitoring time steps. This sequence will serve as the basic data for subsequent construction of the dynamic attribute knowledge graph and ecological asset evaluation, providing strong support for comprehensively and accurately evaluating the value and health status of ecological assets.

[0073] Therefore, the above replaceable or preferred technical implementations have the following technical advantages.

[0074] 1. In traditional ecological asset monitoring data collection, the spatio-temporal correlation of sensor data is usually less considered. Different sensors work independently, and simple averaging or statistical methods are mostly used in data processing, without deeply exploring the internal connections of data in time and space. For example, in marine ecological monitoring, traditional methods only analyze the water quality data collected by sensors at different locations separately, ignoring the mutual influence between data at different locations and different time points. Moreover, traditional methods handle sensor errors simply, generally assuming a single normal distribution, and unable to accurately describe special error sources.

[0075] In this application, the spatio-temporal covariance function C(t1, p1, t2, p2) is introduced to describe the spatio-temporal correlation of sensor data. In the marine ecosystem, data such as water quality and organisms at different locations affect each other over time and space. Through this function, the spatio-temporal variation law can be captured more accurately, improving the utilization efficiency of data. For example, when predicting the number of plankton in a certain area, by combining relevant data from surrounding areas at different times, the prediction results can be made more in line with the actual situation. The mixture Gaussian distribution is used to describe the random error term \(\epsilon_i(t)\), which can more accurately reflect errors from multiple sources, such as sensor noise and model uncertainty. Compared with the traditional single normal distribution assumption, the mixture Gaussian distribution better fits the special error distribution, improving the reliability of data and the accuracy of subsequent analysis.

[0076] 2. Traditional feature extraction methods are mostly based on simple linear models or empirical formulas and cannot handle special non-linear relationships. In the ecosystem, there are special interactions among various ecological indicators, and it is difficult for traditional methods to accurately describe these relationships. For example, when evaluating marine biodiversity, traditional methods only consider the simple combination of a few biological indicators and ignore the complex influence of environmental factors (such as water quality, meteorology, etc.) on biodiversity. In addition, traditional methods usually do not consider the spatio-temporal dynamic changes of data and adopt a static feature extraction method, which cannot adapt to the dynamic characteristics of the ecosystem. In this application, the feature transformation function \(f_{i}\) based on CNNRNN is used for feature extraction. The CNN part can automatically extract the spatial features of data, and the RNN part processes the time series information of data. In marine ecological monitoring, this combination can fully explore the features of biological data in space and time, and more accurately estimate ecological indicators such as biodiversity. For example, CNN identifies the spatial distribution patterns of biological communities in different regions, and RNN tracks the change trends of biological communities over time, thus improving the accuracy and reliability of feature extraction. The weight coefficient \(\alpha\) i (s, t′, p) is dynamically determined through spatio-temporal weighted regression (STWR) combined with the Kalman filter algorithm. STWR takes into account the spatio-temporal non-stationarity of data and can adjust the weights according to the data characteristics at different spatio-temporal positions; the Kalman filter updates the weights in real time to adapt to the dynamic changes of the ecosystem. In marine ecological monitoring, the ecological environments in different seasons and regions vary greatly, and the dynamic weight adjustment makes the feature extraction process more flexible and can more accurately reflect the actual situation of the ecosystem.

[0077] 3. Traditional ecological asset monitoring and assessment methods often handle data collection, feature extraction, and assessment separately, lacking systematicness and integrity. The information transfer and interaction between various links are insufficient, resulting in deviations in the final assessment results. Moreover, traditional methods are difficult to monitor the dynamic changes of ecosystems in real time and accurately predict them, unable to meet the actual needs of ecological protection and management. This application organically combines data collection, feature extraction and other links, fully considering the spatio-temporal characteristics and complex relationships of data. Spatio-temporal correlation is considered from the data collection stage, and special models and dynamic weight adjustment are used in the feature extraction stage. The whole process forms a complete system, which can more comprehensively and accurately reflect the attribute characteristics of ecological assets. For example, in the marine ecosystem, through systematic processing, the health status and value of marine ecological assets can be more accurately evaluated. Due to the use of dynamic models and real-time updated weights, the new method can track the changes of ecosystems in real time and make more accurate predictions about future development trends. In marine ecological protection and management, decision-makers adjust protection strategies in a timely manner based on real-time monitoring and prediction results, take effective measures to address ecological problems, and improve the efficiency and effectiveness of ecological protection.

[0078] Optionally, obtain the attribute feature carrier collected by the attribute sensor, and perform feature extraction on the attribute feature carrier to obtain the attribute features of the target ecological asset, including:

[0079] Perform quantum state fusion on the obtained attribute feature carrier to generate a quantum state mixed data set;

[0080] Map the quantum state mixed data set into dynamic spatio-temporal graph structure data;

[0081] Perform cross-domain mapping on the dynamic spatio-temporal graph structure data, and endow the dynamic spatio-temporal graph structure data with cross-domain semantic information based on graph attention to generate a core feature subset with evaluation value;

[0082] Input the core feature subset into the encoder to learn the probability distribution representation of the features and generate a latent feature vector;

[0083] The generator generates enhanced feature samples according to the latent feature vector, and the discriminator distinguishes between real core features and generated feature samples;

[0084] The variational autoencoder reconstructs and optimizes the real core features and generated feature samples to generate the attribute features of the target ecological asset.

[0085] Preferably, in a scenario, the above technical implementation is described in a replaceable or preferred manner.

[0086] 1. Perform quantum state fusion on the obtained attribute feature carrier to generate a quantum state mixed data set

[0087] 1.1 High-order Quantum State Fusion

[0088] Let the set of acquired attribute feature carriers be each d i is mapped to the quantum state space to obtain the density matrix ρ i . Considering high-order quantum state fusion, a high-order fusion operator is introduced The generated quantum state mixed data set ρ mix is expressed as:

[0089] where: k is the order of fusion, k ≥ 1, and the higher the order, the higher the complexity and information integration degree of fusion. is the high-order weight coefficient, satisfying and These weight coefficients can be determined by combining quantum genetic algorithms with multi-objective optimization, considering factors such as the reliability of attribute feature carriers, their importance to the overall evaluation, and the degree of quantum entanglement between them. is the high-order fusion operator. For example, when k = 2, it is defined as where is the tensor product, [·,·] is the commutator, and λ is a tuning parameter used to balance the contributions of the tensor product and the commutator.

[0090] 2. Map the quantum state mixed data set into dynamic spatio-temporal graph structure data

[0091] 2.1 Construction of dynamic spatio-temporal graph based on fractal dimension

[0092] The dynamic spatio-temporal graph G = (V, E, T), and the node set V is mapped from the elements of the quantum state mixed data set. For node v j and v k , the edge weight e jk is calculated considering the fractal dimension and multi-scale spatio-temporal correlation.

[0093] Let the multi-scale spatio-temporal correlation function of spatio-temporal points (t j , p j ) and (t k , p k ) be C ms (t j , p j , t k , p k ), which is defined as:

[0094] where: L is the number of scales, β l is the weight coefficient of the l-th scale, and β l ≥ 0. C l(t j , p j , t k , p k ) is the spatio-temporal covariance function at the l-th scale. For example, The parameters at different scales are different.

[0095] The fractal dimension D of the node features f (x j ) is calculated by methods such as the box-counting method. The edge weight e jk is:

[0096] where f(x j , x k ) is the feature correlation function, such as cosine similarity.

[0097] 3. Map the dynamic spatio-temporal graph structure data for cross-domain mapping, and endow the dynamic spatio-temporal graph structure data with cross-domain semantic information based on graph attention to generate a core feature subset with evaluation value

[0098] 3.1 Multi-modal cross-domain mapping

[0099] Let the node feature matrix of the dynamic spatio-temporal graph G=(V, E, T) be X, and the adjacency matrix of the edges be A. The multi-modal cross-domain mapping considers multiple different mapping functions φ m (m = 1, 2, …, M), and maps the node feature matrix X to multiple new feature spaces X m′ : X m′ = φ m (X) where \(\phi_m\) is a deep neural network with different structures, such as a convolutional neural network (CNN), a recurrent neural network (RNN), etc.

[0100] Then, the results of these different feature spaces are fused to obtain the final cross-domain mapping result X': where γ m is the fusion weight coefficient, which is determined by an adaptive weight learning algorithm (such as a weight update algorithm based on gradient descent).

[0101] 3.2 Multi-head graph attention mechanism combined with fuzzy logic

[0102] The multi-head graph attention mechanism is used to endow the dynamic spatio-temporal graph structure data with cross-domain semantic information. For the node v i , the attention coefficient of the h-th head is calculated as follows:

[0103] where W h and a h are the learnable parameters of the h-th head.

[0104] Introduce fuzzy logic and define the fuzzy attention coefficient where is the fuzzy membership function, for example Sim(x i ,x j ) is the node feature similarity, and β and τ are fuzzy control parameters.

[0105] The new feature h i of node v i is: where H is the number of heads and σ is the activation function. In this way, a core feature subset with evaluation value is generated

[0106] 4. Input the core feature subset into the encoder to learn the probability distribution representation of the features and generate the latent feature vector

[0107] 4.1 Deep hierarchical encoder and Gaussian mixture variational inference

[0108] The encoder adopts a deep hierarchical structure and consists of multiple encoding layers. Let the l-th layer encoder be where

[0109] The encoder learns the probability distribution representation of the features through Gaussian mixture variational inference. The encoder outputs the parameters of the mixture Gaussian distribution of the latent feature vector z, that is, the mean μ c , standard deviation σ c and weight π c (c = 1, 2,..., C): The latent feature vector z is obtained by sampling from the mixture Gaussian distribution: where ∈ c is a random vector sampled from the standard normal distribution N(0, 1).

[0110] 5. The generator generates enhanced feature samples according to the latent feature vector, and the discriminator distinguishes between real core features and generated feature samples

[0111] 5.1 Multi-scale generative adversarial network (GAN) combined with residual connections

[0112] The generator G θ (z) adopts a multi-scale structure and consists of multiple generative sub-networks Each sub-network is responsible for generating features of different scales. where δ m is the scale fusion weight coefficient, which is optimized through a reinforcement learning algorithm. At the same time, the generator introduces residual connections. For the n-th layer generative sub-network, the output is: where y0 = z.

[0113] Discriminator also adopts a multi-scale structure to discriminate features of different scales. The loss functions of the generator and the discriminator introduce an adversarial gradient penalty term L gp :

[0114]

[0115] where λ gp is the gradient penalty coefficient, is the interpolation distribution of the real feature and the generated feature.

[0116] 6. The variational autoencoder reconstructs and optimizes the real core feature and the generated feature samples to generate the target ecological asset attribute features

[0117] 6.1 The variational autoencoder combines quantum entanglement regularization

[0118] The variational autoencoder consists of an encoder and a generator The loss function of the variational autoencoder, in addition to the reconstruction error and the KL divergence, introduces a quantum entanglement regularization term L qe :

[0119] where λ qe is the quantum entanglement regularization coefficient. The quantum entanglement regularization term L qe is used to measure the degree of quantum entanglement between the latent feature vectors z and is calculated by methods such as quantum mutual information.

[0120] 1. Loss function and optimization of the variational autoencoder (VAE)

[0121] The variational autoencoder contains an encoder and a generator Its loss function L VAE is as follows:

[0122] Reconstruction error term In this application, this error term measures the difference between the feature samples reconstructed by the generator according to the latent feature z and the real core feature In ecological asset monitoring, if the core feature covers the vegetation biomass and species diversity of the terrestrial ecosystem, or the water quality parameters and biological density of the aquatic ecosystem, etc. The expected operation It is achieved by sampling from the distribution of the latent feature z output by the encoder. This error term prompts the generator to reconstruct these features as precisely as possible to narrow the gap with the real features.

[0123] KL divergence term In this application, the KL divergence is used to measure the distribution of the latent feature z output by the encoder The difference from the prior distribution p(z) (usually set to the standard normal distribution N(0, I)). In the ecological asset scenario, it makes the distribution of the latent features learned by the encoder closer to the standard normal distribution, ensuring that the latent feature space has a good structure, and the features of different ecological regions or types are reasonably distributed in the latent space, which is conducive to subsequent generation and analysis.

[0124] Quantum entanglement regularization term λ qe L qe , in this application, this term measures the degree of quantum entanglement between the latent feature vectors z, and λ qe is the adjustment coefficient. In the ecosystem, there are complex correlations among different ecological features. For example, in the marine ecosystem, water quality parameters and biological indicators affect each other. This regularization term prompts the latent feature space to capture these complex correlations, making the generated feature samples more in line with the actual situation of the ecosystem.

[0125] By calculating the gradients of the loss function L VAE with respect to the encoder parameter φ and the generator parameter θ, and then using an optimization algorithm to update the parameters. Taking the Adam optimization algorithm as an example, its update formula is as follows:

[0126] For the encoder parameter φ: For the generator parameter θ:, where α is the learning rate, and are the first-order moment estimate and the second-order moment estimate of the encoder parameter φ, and are the first-order moment estimate and the second-order moment estimate of the generator parameter θ, and ∈ is a small constant to prevent the denominator from being zero.

[0127] 2. Reconstruction Optimization and Generation of Target Ecological Asset Attribute Features

[0128] During the optimization process, as the number of iterations increases, the loss function L VAE gradually decreases. The decrease of the reconstruction error term indicates the improvement of the generator's ability to reconstruct the real core features. The decrease of the KL divergence term makes the latent feature space more in line with the prior distribution. The quantum entanglement regularization term enables the latent features to capture the complex correlations between ecological features. After multiple iterations of optimization, the encoder can effectively map the real core features to the latent feature space, and the generator can generate high-quality feature samples based on the latent features. These samples are the target ecological asset attribute features.

[0129] 3. Parameter Representation of Target Ecological Asset Attribute Characteristics

[0130] 3.1 Terrestrial Ecosystems

[0131] Forest Ecosystems

[0132] Biomass: Including aboveground biomass B a (Unit: tons / hectare) and belowground biomass B u (Unit: tons / hectare). It can be estimated by combining remote sensing data with biomass models, such as B a = f(NDVI, LAI,...), where NDVI is the normalized difference vegetation index and LAI is the leaf area index. Species diversity: Represented by the Shannon-Wiener index H, where S is the total number of species, and p i is the proportion of the number of individuals of the i-th species to the total number of individuals. Carbon sink C s (Unit: tons of carbon dioxide / hectare·year): It can be estimated by measuring biomass growth and carbon content and combining ecological models, such as C s = g(B a , B u ,...).

[0133] Grassland Ecosystems

[0134] Vegetation Coverage V c : Expressed in percentage form, it can be obtained through remote sensing image analysis or on-site quadrat surveys, such as V c = h(RGB, NIR,...), where RGB is the visible light band and NIR is the near-infrared band. Forage Yield Y g (Unit: kg / hectare): Estimated by weighing the forage in the harvested quadrats in the field, and is related to factors such as vegetation coverage and soil fertility, such as Y g = k(V c , OM,...), where OM is the soil organic matter content. Soil fertility: Including soil organic matter content OM (expressed as a percentage), soil nitrogen content N (unit: g / kg), soil phosphorus content P (unit: g / kg), and soil potassium content K (unit: g / kg), which are determined by soil sampling and chemical analysis.

[0135] 3.2 Aquatic Ecosystems

[0136] Marine Ecosystems

[0137] Water Quality Parameters: Dissolved oxygen DO (unit: mg / L), chemical oxygen demand COD (unit: mg / L), pH value, etc., which are monitored in real time by water quality sensors.

[0138] Biological Parameters: Phytoplankton Density D p(Unit: number per liter), determined by microscopic observation and counting; fish resource quantity B f (Unit: ton), estimated by fishery resource survey methods, such as B f = l(D p , temp, …), where temp is the seawater temperature. Ecological service function parameter: fishery yield Y f (Unit: ton), obtained from fishery statistical data; the coastal protection function can be evaluated by indicators such as the coastal zone terrain change rate R t and so on, such as ΔT is the change amount of the coastal zone terrain within a certain time, and T is the initial terrain parameter.

[0139] Freshwater ecosystems (rivers, lakes, etc.)

[0140] Hydrological parameters: flow rate Q (unit: cubic meters per second), water level H (unit: meter), which are monitored in real time by hydrological station measuring equipment.

[0141] Water quality parameters: total phosphorus TP (unit: milligrams per liter), total nitrogen TN (unit: milligrams per liter), transparency SD (unit: meter), measured by chemical analysis and Secchi disk. Biological parameter: benthic animal diversity is represented by the Simpson diversity index D s as aquatic plant coverage V a (expressed as a percentage), obtained by remote sensing image analysis or field investigation.

[0142] By optimizing the loss function of the variational autoencoder, the real core features and the generated feature samples are reconstructed and optimized, and finally the target ecological asset attribute features are generated.

[0143] Therefore, in the field of ecological asset dynamic monitoring and evaluation, this application shows a completely different technical essence from traditional technologies from data processing, feature extraction to model optimization, bringing significant technical advantages in the ecological asset monitoring and evaluation scenario as follows:

[0144] 1. Traditional data fusion methods mostly adopt simple weighted average or rule-based fusion strategies, and lack the mining of complex relationships between data. When constructing the spatio-temporal graph structure, usually only the spatio-temporal correlation at a single scale is considered, and it is difficult to reflect the special spatio-temporal variation characteristics of the ecosystem. For example, in forest ecological monitoring, traditional methods simply summarize the sensor data in different regions, ignoring the interactions of ecological factors in different height vegetation layers and different time scales. This application adopts a high-order fusion operator and multi-order weight coefficients Optimize the weights through the quantum genetic algorithm, fully considering the quantum entanglement characteristics and complex correlations among ecological data. Taking the wetland ecosystem as an example, this method can more accurately fuse multi-source data such as water quality, biology, and meteorology, avoid information loss caused by traditional weighted averaging, and improve the depth and accuracy of data fusion. Construction of dynamic spatio-temporal graphs based on fractal dimensions: Introduce fractal dimensions and multi-scale spatio-temporal correlation function C ms Calculate the edge weights, which can capture the self-similarity and spatio-temporal dynamic changes of the ecosystem at different scales. In marine ecological monitoring, it can more meticulously depict the interaction relationships of ecological factors in different sea areas and at different times. Compared with traditional single-scale spatio-temporal graphs, it can more realistically reflect the complex structure and dynamic evolution process of the ecosystem.

[0145] 2. Traditional cross-domain mapping methods usually rely on manually designed mapping rules and lack the ability to adaptively learn the internal characteristics of data. In terms of graph attention mechanisms, traditional methods are mostly single-attention calculations and are difficult to handle the ambiguity and multi-modal information in ecological data. For example, during ecological assessment, traditional mapping methods cannot effectively convert ecological monitoring data into features that meet the assessment requirements, and have limited ability to handle the complex relationships among ecological factors.

[0146] This application uses a variety of different mapping functions φ m (such as CNN, RNN) for feature mapping, and determines the fusion weight γ through an adaptive weight learning algorithm m , which can automatically learn the feature representations of different modal data and achieve efficient conversion from ecological monitoring data to assessment features. When evaluating the value of forest ecological assets, it can simultaneously process multi-modal data such as vegetation spectra, terrain, and meteorology, improving the accuracy and applicability of feature conversion. Calculate the feature importance from different perspectives through multi-head attention, and combine the fuzzy membership function to handle the ambiguity and uncertainty of ecological data. In the wetland ecosystem, it can more accurately identify the complex relationships among different biological populations and environmental factors. Compared with traditional single-attention mechanisms, it can more comprehensively capture the semantic information of the ecosystem and provide a richer feature subset for ecological assessment.

[0147] 3. The traditional encoder generator model has a simple structure, weak learning ability for data probability distributions, and low-quality generated feature samples. In the case of variational autoencoders, traditional methods usually do not consider the special correlations of ecological data, and the optimization process lacks pertinence. For example, when generating ecological features, the features generated by traditional models do not conform to the actual ecological situation and cannot provide effective support for ecological assessment. This application uses an encoder with a deep hierarchical structure and combines Gaussian mixture variational inference to learn the probability distribution of features, which can more accurately capture the complex distribution characteristics of ecological data. When processing ecological asset features in different regions and at different times, it can generate latent feature vectors that are more in line with the actual distribution, laying a foundation for generating high-quality ecological feature samples. The generator adopts a multi-scale structure and residual connections, combined with an adversarial gradient penalty term, which improves the diversity and authenticity of the generated samples. When generating ecological asset attribute features, it can generate feature samples at different scales and closer to the real ecological situation, and has a better generation effect than traditional generative adversarial networks. A quantum entanglement regularization term L qe is introduced into the variational autoencoder loss function to prompt the model to learn the complex correlations between ecological features. In ecological asset assessment, it can make the generated attribute features better reflect the interaction relationships within the ecosystem. The optimized feature samples provide more reliable data support for the dynamic monitoring and precise assessment of ecological assets, and have stronger adaptability and accuracy in ecological scenario applications compared to traditional variational autoencoders.

[0148] Optionally, based on the dynamic attribute feature sequence, a dynamic attribute knowledge graph is constructed, including: performing entity extraction, relationship extraction, and attribute extraction on the dynamic attribute feature sequence to respectively obtain the entities of ecological assets, the semantic relationships between entities, and the attribute information of entities; preprocessing the extracted entities, and updating the semantic relationships and attribute information based on the preprocessed entities, where the preprocessing includes alignment and disambiguation processing; loading the entities, the semantic relationships between entities, and the attribute information of entities into a graph data structure and loading it into the constructed initial-state knowledge graph to form the dynamic attribute knowledge graph of the target ecological asset.

[0149] Preferably, in a scenario, the above technical implementation is described in a replaceable or preferred manner.

[0150] 1. Perform entity extraction, relationship extraction, and attribute extraction on the dynamic attribute feature sequence

[0151] 1.1 Entity extraction

[0152] Let the dynamic attribute feature sequence be S = {s1, s2, …, s n}, where s iRepresents the i-th feature in the sequence. The goal of entity extraction is to identify entities related to ecological assets from this sequence. This application uses a Conditional Random Field (CRF) model for entity extraction. The probability formula of the CRF model is: Where: x is the input dynamic attribute feature sequence, y is the corresponding entity label sequence. Z(x) is the normalization factor, λ k Is the weight of the k-th feature function f k Obtained by learning from the training data. f k (y i-1 , y i , x, i) is a feature function used to describe the local features of the label sequence y at position i. For example, when x is a dynamic attribute feature sequence about a forest ecosystem, entities such as "trees", "birds", "soil", etc. will be identified.

[0153] 1.2 Relation Extraction

[0154] The purpose of relation extraction is to determine the semantic relations between entities. This application uses a relation extraction model based on deep learning, such as a Convolutional Neural Network (CNN). Let the input entity pair be (e1, e2), and their corresponding feature vectors be And Concatenate them into a new vector

[0155] The convolution layer formula of the CNN model is: c j = ReLU(W j v + b j ), where: W j Is the weight matrix of the j-th convolutional kernel, b j Is the bias term. ReLU is the activation function, ReLU(x) = max(0, x). After processing by the convolutional layer and the pooling layer, a fixed-length feature vector h is obtained, and then classification is performed through the fully connected layer to predict the relation r between the entity pair:

[0156] Where: W h Is the weight matrix of the fully connected layer, b h Is the bias term. The softmax function is used to convert the output into a probability distribution, m is the number of relation categories. For example, in a forest ecosystem, a relation such as "trees provide habitat for birds" will be identified.

[0157] 1.3 Attribute Extraction

[0158] Attribute extraction is to extract relevant attribute information for entities. This application uses a combination of rule-based methods and machine learning methods. Let the entity be e, and its corresponding dynamic attribute feature sequence be S e For each attribute type a, this application defines a rule set R a and a machine learning model M a .

[0159] First, use the rule set R a to match S e . If the match is successful, extract the corresponding attribute value v a . If the rule match fails, use the machine learning model M a to make a prediction. For example, for the "tree" entity, the attribute types include "tree height", "diameter at breast height", etc. Through rule matching or the machine learning model, this application extracts attribute information such as "tree height: 10 meters" and "diameter at breast height: 0.5 meters" from the dynamic attribute feature sequence.

[0160] 2. Preprocess the extracted entities and update semantic relationships and attribute information based on the preprocessed entities

[0161] 2.1 Entity alignment

[0162] The purpose of entity alignment is to merge the same entity from different sources or different representations. Let the set of extracted entities be E = {e1, e2,..., e m}. This application uses a similarity-based method for entity alignment. Define the similarity function sim(e i and e j ) between entities, for example, using cosine similarity: i , j where and are the feature vectors of entities e and e i and e j respectively. If sim(e i , e j ) is greater than a certain threshold τ, then e i and e j are considered the same entity and are merged into one entity e ij .

[0163] 2.2 Entity disambiguation

[0164] Entity disambiguation is to solve the ambiguity problem of entity names. Suppose there is an entity name n that corresponds to multiple different entities e n1 , e n2 , …, e nk. This application uses context information and a knowledge base for entity disambiguation. Define a context similarity function sim ctx (e ni , C), where C is the context in which the entity name n appears. At the same time, obtain the relevant information of each entity e ni from the knowledge base, and calculate its matching degree match(e ni , C). The final disambiguation score score(e ni ) = αsim ctx (e ni , C) + (1 - α)match(e ni , C), where α is the weight coefficient. Select the entity with the highest score as the correct entity.

[0165] 2.3 Update semantic relationships and attribute information

[0166] Based on the preprocessed entities, update the semantic relationships between entities and the attribute information of entities. If two entities are merged into one entity, then merge their relationships and attribute information. If an entity is disambiguated as another entity, then update all the relationships and attribute information related to it.

[0167] 3. Load the entities, the semantic relationships between entities, and the attribute information of entities into a graph data structure and load it into the initially constructed knowledge graph to form a dynamic attribute knowledge graph of the target ecological assets

[0168] 3.1 Graph data structure representation

[0169] Represent entities as nodes of the graph, represent the semantic relationships between entities as edges of the graph, and store the attribute information of entities in the attributes of nodes and edges. Let the node set be V, the edge set be E, and each node v ∈ V has an attribute set A v , and each edge e ∈ E has an attribute set A e . The graph data structure is represented by an adjacency matrix A, where A ij represents whether there is an edge between node i and node j. If there is an edge, then A ij = 1, and store the attribute information of the edge in the corresponding position.

[0170] 3.2 Load into the initially constructed knowledge graph

[0171] Let the initial state knowledge graph be G0 = (V0, E0, A0). Load the new node, edge, and attribute information into the initial state knowledge graph to form the dynamic attribute knowledge graph G = (V, E, A) of the target ecological assets. For a new node v ∈ V - V0, add it to the node set V and add its attribute information to A. For a new edge e ∈ E - E0, add it to the edge set E and add its attribute information to A.

[0172] Therefore, starting from the specific scenario of ecological asset dynamic monitoring and evaluation and comparing with traditional technology processing, the following technical advantages are obtained:

[0173] 1. Traditional methods mostly adopt rule-based entity extraction technology, which requires a large number of rules to be manually written and is difficult to adapt to the complex and changeable language expressions and data types in the ecological field. For example, in forest ecological monitoring, for newly emerged species names or descriptions of ecological phenomena, the rules cannot accurately identify entities. Traditional relation extraction relies on manual features and shallow machine learning models, such as the Naive Bayes classifier. These models are difficult to capture the special semantic information and context relationships in ecological data, resulting in low accuracy and recall rates of relation extraction. For example, when describing the symbiotic and predatory relationships between species in an ecosystem, traditional methods are prone to misjudgment or omission. Traditional attribute extraction methods are usually simple keyword matching, ignoring the association between attribute values and entities and contexts. In ecological scenarios, the same attribute of different entities has different value ranges and meanings, and traditional methods are difficult to accurately extract and distinguish. In this application, a conditional random field (CRF) model is used to automatically learn the context information in sequence data through feature functions without the need to manually write a large number of rules. In ecological monitoring, it can better adapt to dynamic attribute feature sequences from different sources and in different formats, and accurately identify various ecological entities, such as rare species and ecological regions. In this application, a convolutional neural network (CNN) model based on deep learning can automatically extract the features of entity pairs and capture the complex semantic information in the data. In an ecosystem, it can more accurately identify various relationships between species and between species and the environment, such as "trees provide habitats for birds" and "soil nutrients supply plants", providing more comprehensive information for ecological research and protection. In this application, by combining rules and machine learning methods, the certainty of rules is used to process common attribute extraction, and machine learning models are used to handle complex situations. In ecological asset monitoring, various attribute information of entities, such as the growth cycle of species and the environmental indicators of ecological regions, can be accurately extracted, improving the accuracy and flexibility of attribute extraction.

[0174] 2. Traditional entity alignment methods are mainly based on string matching, only considering the similarity of entity names and ignoring the semantic and context information of entities. In the ecological field, the same entity may have multiple names or expressions, and traditional methods are prone to misalignment. Traditional entity disambiguation methods rely on manually constructed knowledge bases and rules, making it difficult to handle large-scale and dynamically changing ecological data. For entities with the same name but different meanings in the ecological field, such as the same species name in different regions, traditional methods cannot accurately distinguish them. In this application, the entity alignment method based on a similarity function (such as cosine similarity) considers the feature vectors of entities and can more accurately measure the semantic similarity between entities. In ecological monitoring, the same ecological entity from different sources and with different representations can be accurately merged to avoid data redundancy and errors, improving the quality of the knowledge graph. The disambiguation method combining context similarity and knowledge base matching degree can make full use of the context information where the entity appears and the relevant knowledge in the knowledge base. In the ecological scenario, it can accurately solve the ambiguity problem of entity names, ensure the uniqueness and accuracy of entities in the knowledge graph, and provide a reliable basis for subsequent analysis and decision-making.

[0175] 3. Traditional graph data structure representation methods are relatively simple, usually only focusing on the connection relationship between nodes and edges and ignoring the attribute information of nodes and edges. In ecological asset monitoring, ecological entities and relationships have rich attributes, and traditional methods cannot comprehensively store and express this information. Traditional knowledge graph loading methods are static and difficult to adapt to the dynamic changes of ecological data. In an ecosystem, entities, relationships, and attributes change with time and the environment, and traditional methods cannot update the knowledge graph in a timely manner.

[0176] In this application, entity, relationship, and attribute information are comprehensively stored in the graph data structure, and an adjacency matrix and an attribute set are used to represent the structure and attributes of the graph. In ecological monitoring, various attributes of ecological entities (such as the biological characteristics of species, the geographical information of ecological regions, etc.) and the attributes of relationships between entities (such as the strength and duration of relationships) are completely recorded, providing richer information for ecological research. In this application, new node, edge, and attribute information are dynamically loaded into the initial knowledge graph, which can timely reflect the changes in the ecosystem. In dynamic ecological asset monitoring, the knowledge graph is updated in real time, providing the latest and most accurate information for ecological decision-making and improving the efficiency and scientific nature of ecological management.

[0177] Optionally, entity extraction, relationship extraction, and attribute extraction are performed on the dynamic attribute feature sequence to respectively obtain the entities of ecological assets, the semantic relationships between entities, and the attribute information of entities, including:

[0178] Convert the dynamic attribute feature sequence into a superposition state representation sequence;

[0179] Perform multi-scale decomposition on the superposition state representation sequence to obtain a quantum feature set containing multi-scale detailed features;

[0180] Abstract each feature element in the quantum feature set into a molecular unit;

[0181] Perform semantic feature aggregation between molecular units to generate entity candidate clusters;

[0182] Perform spatio-temporal evolution pattern recognition on the entity candidate clusters to generate a pre-mined relationship result;

[0183] Based on the connectivity and path constraints of the constructed topological graph, perform directed entity extraction, relationship extraction, and attribute extraction on the entity candidate clusters and the pre-mined relationship results to respectively obtain the entities of ecological assets, the semantic relationships between entities, and the attribute information of entities.

[0184] Preferably, in a scenario, the above technical implementation is described in a replaceable or preferred manner.

[0185] 1. Convert the dynamic attribute feature sequence into a superposition state representation sequence

[0186] Let the dynamic attribute feature sequence be S = {s1, s2, …, s n}, where s i = (s i1 , s i2 , …, s im ) is an m-dimensional vector representing the values of the i-th dynamic attribute feature in different dimensions. In the ecological asset monitoring scenario, if monitoring a forest ecosystem, these dimensions correspond to indicators such as the height, diameter at breast height, crown width of trees, and the soil humidity and light intensity of the area. This application uses a mapping function based on quantum entanglement to convert it into a superposition state representation sequence Q = {q1, q2, …, q n}. For each s i , the mapping process is as follows: Where: |ψ jk > is the quantum ground state, 2^m represents the number of all quantum state combinations, because the features in each dimension have multiple state superpositions in the quantum representation. α ijk is the probability amplitude corresponding to the quantum ground state, satisfying The calculation of the probability amplitude is implemented through a special neural network which takes s i as input and outputs α ijk : The neural network here is a deep convolutional neural network (DCNN) whose structure includes multiple convolutional layers, pooling layers, and fully connected layers, and is trained with a large amount of ecological data to learn the mapping relationship from features to probability amplitudes.

[0187] 2. Perform multi-scale decomposition on the superposition state representation sequence to obtain a quantum feature set containing multi-scale detailed features

[0188] Adopt a multi-scale decomposition operator based on quantum wavelet transform Process the superposition state representation sequence Q. The multi-scale decomposition process is expressed in a recursive form: where Q (0) = Q, l represents the scale level. The specific multi-scale decomposition operator is defined as follows: where: U pr is a set of unitary transformation matrices, and the parameters of these matrices are optimized through a quantum genetic algorithm to adapt to the characteristics of ecological data. The unitary transformation matrix U pr is used to rotate and transform the quantum state, thereby extracting feature information at different scales. w pr is the corresponding weight coefficient, satisfying The determination of the weight coefficient is through an adaptive weight adjustment algorithm, which dynamically adjusts according to the importance of features at different scales.

[0189] 3. Abstract each feature element in the quantum feature set into a molecular unit

[0190] Define an abstraction function based on quantum state projection and feature fusion to abstract each feature element (l) in the quantum feature set Q into a molecular unit where: Π s is a quantum state projection operator, which is used to project the quantum state onto a specific subspace to extract feature information in different aspects. f s is a predefined feature vector, representing different semantic feature dimensions. ⊙ represents element-wise multiplication operation. β s is the fusion weight, which is learned through a neural network based on an attention mechanism, and this network dynamically assigns weights according to the importance of features.

[0191] 4. Perform semantic feature aggregation between molecular units to generate entity candidate clusters

[0192] Define a semantic similarity function based on quantum entanglement similarity and graph neural network

[0193] where: h iand h j is a molecular unit and are feature vectors encoded by a graph neural network (GNN). The input of the graph neural network is a molecular unit graph, where nodes represent molecular units and edges represent initial associations between molecular units. W is a learnable weight matrix. σ is an activation function, such as the sigmoid function. A hybrid clustering algorithm based on density and hierarchical clustering is used to cluster molecular units, generating entity candidate clusters C = {C1, C2, …, C p}. This algorithm combines the density clustering idea of the DBSCAN algorithm and the hierarchical structure construction ability of hierarchical clustering. The specific steps are as follows: First, the DBSCAN algorithm is used to perform preliminary clustering on molecular units based on semantic similarity sim, obtaining some density-connected core points and boundary point sets. Then, the hierarchical clustering algorithm is used to further merge and partition these sets to form the final entity candidate clusters.

[0194] 5. Identify the spatio-temporal evolution patterns of entity candidate clusters to generate pre-mined relationship results

[0195] Let the entity candidate cluster Ct represent the set of entity candidate clusters at time t. Define a spatio-temporal evolution pattern function f(Ct1, Ct) based on a spatio-temporal graph convolutional network (STGCN): h t = ST-GCN(C t-1 , C t ), where h t is the feature representation at time t. The spatio-temporal graph convolutional network combines the ideas of graph convolutional networks and temporal convolutional networks, and can simultaneously process the spatial relationships and temporal evolutions between entity candidate clusters.

[0196] Through a classifier classify h t to obtain the pre-mined relationship result R pre : The classifier is a fully connected neural network, and its output is the probability distribution of different relationship types.

[0197] 6. Based on the connectivity and path constraints of the constructed topological graph, perform directed entity extraction, relationship extraction, and attribute extraction on entity candidate clusters and pre-mined relationship results

[0198] Construct a topological graph G = (V, E), where V is the set of nodes, corresponding to entity candidate clusters; E is the set of edges, corresponding to pre-mined relationship results.

[0199] Define the connectivity function conn(u, v) as:

[0200] The path constraint function path(u, v) is implemented through a path search algorithm based on reinforcement learning. This algorithm takes the topological graph G as the environment, and the agent searches for a path from node u to node v in the graph while satisfying certain constraints such as path length and edge weight.

[0201] In the topological graph G, directed entity extraction, relationship extraction, and attribute extraction are carried out according to connectivity and path constraints. Directed entity extraction is achieved through a selector based on the graph attention mechanism as follows: where is the set of extracted entities. Relationship extraction determines the semantic relationships between entities by analyzing the attributes of edges that satisfy connectivity and path constraints. Attribute extraction uses an inference model based on knowledge graph embedding, combined with additional information of nodes and edges, to obtain the attribute information of entities. This inference model maps the information in the topological graph to the knowledge graph and uses the semantic information of the knowledge graph for attribute inference.

[0202] Therefore, the above replaceable or preferred technical implementations, centered around the integration of quantum computing and deep learning, realize the in-depth mining of ecological asset information through quantum state conversion, multi-scale decomposition, semantic aggregation, and topology graph-based intelligent extraction of ecological asset data. Its essence is to utilize the superposition state and entanglement characteristics of quantum computing, combined with the adaptive learning ability of deep learning. Compared with traditional technologies, it has the following technical advantages:

[0203] 1. Traditional methods usually store and process ecological data in conventional forms such as numerical values and texts. For example, simply record data such as the height and quantity of trees in a forest and store it in tabular form. This way ignores the potential complex associations between data, cannot effectively capture the dynamic changes and internal laws of the ecosystem, and the data representation form is single, making it difficult to reflect the multi-dimensional characteristics of ecological data. This application uses a quantum state mapping function to convert the dynamic attribute feature sequence into a superposition state representation sequence, and the probability amplitude α ijk is learned by a deep convolutional neural network, enabling each data point to exist in the form of a quantum superposition state, reflecting the properties of multiple ecological states. This representation method can not only describe multiple dimensions of ecological characteristics simultaneously but also utilize the quantum entanglement characteristics to reflect the potential connections between different ecological factors. For example, in forest ecological monitoring, it can simultaneously represent the complex associations between tree growth and factors such as soil humidity and light intensity, and can depict the state of the ecosystem more comprehensively and accurately compared with traditional methods.

[0204] 2. Traditional feature extraction mostly adopts statistical methods with fixed windows or simple filtering algorithms, such as calculating the average value, standard deviation, etc. of forest temperature over a period of time. These methods can only extract surface features at a single scale, cannot adapt to the changes of the ecosystem at different spatio-temporal scales, and are difficult to effectively analyze special ecological processes.

[0205] This application uses a multi-scale decomposition operator based on quantum wavelet transform such as where the unitary transformation matrix U pr is optimized by a quantum genetic algorithm, and the weight coefficient w pr is adaptively adjusted. This enables the system to decompose ecological data at multiple scales and extract detailed features at different spatio-temporal resolutions. For example, when monitoring a river ecosystem, it can capture the overall flow change trend of the river (large-scale feature) and analyze the details of water quality fluctuations in local river sections (small-scale feature). Compared with traditional methods, it can understand the evolution mechanism of the ecosystem more deeply and provide richer information for ecological protection and management.

[0206] 3. Traditional semantic aggregation and clustering methods rely on manually set rules or simple similarity metrics, such as clustering algorithms based on Euclidean distance. In ecological applications, due to the ambiguity of ecological concepts and the complexity of data, these methods are prone to problems such as inaccurate clustering and semantic understanding deviation, and cannot effectively identify ecological entities and the relationships between them.

[0207] This application defines a semantic similarity function based on quantum entanglement similarity and graph neural network and uses a hybrid clustering algorithm to generate entity candidate clusters. Quantum entanglement similarity can capture the deep semantic connections between data, and graph neural network can learn the context information of molecular units in the graph structure. For example, in marine ecological monitoring, for special biological communities and environmental factors, the new technology can more accurately cluster molecular units with similar ecological functions or correlations into entity candidate clusters, thus more precisely identifying different biological populations and ecological regions in the marine ecosystem. Compared with traditional clustering methods, the clustering results are more in line with the ecological reality and improve the semantic understanding ability of ecological data.

[0208] 4. Traditional relationship mining and information extraction mainly rely on manually constructed knowledge bases and rule templates, such as extracting species relationships in ecological literature through preset keywords and grammar rules. This method is inefficient, difficult to adapt to massive dynamic data, and difficult to accurately identify newly emerging ecological relationships and complex semantics.

[0209] This application identifies the spatio-temporal evolution patterns of entity candidate clusters through a spatio-temporal dynamic graph convolutional network (STGCN), and combines a path search algorithm based on reinforcement learning for directed entity extraction, relationship extraction, and attribute extraction. For example, in the ecological asset assessment of urban green spaces, it can automatically discover the ecological connections over time between different green spaces (such as relationships like species migration and ecological function complementarity), as well as the detailed attribute information of each green space (area, vegetation type, ecological service value, etc.). Compared with traditional methods, the new technology realizes automated and intelligent information extraction, not only improving the efficiency of information processing, but also discovering complex ecological relationships that are difficult to detect by traditional methods, providing more comprehensive and accurate data support for the comprehensive assessment of ecological assets.

[0210] Optionally, entities, the semantic relationships between entities, and the attribute information of entities are loaded into a graph data structure and then into the constructed initial knowledge graph to form a dynamic attribute knowledge graph of the target ecological assets, including:

[0211] Using a graph attention network, infer the semantic relationships between entities, the attribute information of entities, the ecological dependence semantic relationships, and the material transfer semantic relationships between entities;

[0212] Quantify the relationship strengths of the ecological dependence semantic relationships and the material transfer semantic relationships to encapsulate them into weighted semantic relationship triples;

[0213] Based on the semantic relationship triples, construct an R-tree hybrid index;

[0214] Load the R-tree hybrid index into the constructed initial knowledge graph to form a dynamic attribute knowledge graph of the target ecological assets.

[0215] Preferably, in a scenario, the above technical implementation is described in a replaceable or preferred manner.

[0216] 1. Using a graph attention network, infer the semantic relationships between entities, the attribute information of entities, and mine the ecological dependence semantic relationships and the material transfer semantic relationships between entities

[0217] Let the graph data structure be G=(V, E), where V={v1, v2,..., v n} is the set of entity nodes, and E={(v i , v j )} is the set of edges, representing the existing semantic relationships between entities. Each node v i has a corresponding attribute vector x i , which contains the attribute information of the entity. For example, in a forest ecosystem, for the "tree" entity, the attribute vector contains information such as tree height, diameter at breast height, and tree species.

[0218] The propagation process of one layer of the Graph Attention Network (GAT) is represented as: Where: is the feature representation of node i at the l-th layer, W (l) is the learnable weight matrix at the l-th layer, used for linear transformation of the input features. is the set of neighbor nodes of node i. is the attention coefficient, indicating the importance of node j to node i, calculated by the following formula:

[0219] Where a (l) is the learnable attention vector, || represents the vector concatenation operation, and LeakyReLU is the activation function.

[0220] To mine the ecological dependence semantic relationship and material transfer semantic relationship between entities, this application introduces an additional relationship reasoning module. Let r ij represent the relationship feature vector between node i and node j, calculated by the following formula: Where MLP is a multi-layer perceptron, and L is the total number of layers of the graph attention network. By classifying r ij , this application determines whether there is an ecological dependence semantic relationship or a material transfer semantic relationship between node i and node j.

[0221] 2. Quantify the relationship strength of the ecological dependence semantic relationship and the material transfer semantic relationship to encapsulate them into weighted semantic relationship triples

[0222] Let R eco represent the set of ecological dependence semantic relationships, and R mat represent the set of material transfer semantic relationships. For each pair of nodes (v i , v j ) with ecological dependence or material transfer relationships, this application defines a relationship strength function s to quantify the relationship strength.

[0223] For the ecological dependence semantic relationship, the relationship strength function is defined as: Where f k is the k-th feature function, used to measure the dependence degree of node i and node j in a specific ecological aspect. For example, in a forest ecosystem, f1 represents the dependence degree of "trees" and "birds" in terms of habitat provision, and w k is the corresponding weight coefficient, learned from the training data.

[0224] For the material transfer semantic relationship, the relationship strength function is defined as: Where MLP′ is another multi-layer perceptron, and w is the learnable weight vector.

[0225] Encapsulate entities, relationships, and relationship strengths into weighted semantic relationship triples (v i , r, v j , s), where r ∈ R eco ∪R mat , and s is the corresponding relationship strength.

[0226] 3. Based on the semantic relationship triples, construct an R-tree hybrid index

[0227] Let the set of semantic relationship triples be T = {(v i , r, v j , s)}. This application maps each triple to a multi-dimensional space, where each dimension corresponds to an entity attribute or relationship feature. For example, in a forest ecosystem, the dimensions include the geographical location, species type, ecological function, etc. of the entity.

[0228] The construction process of the R-tree hybrid index is divided into the following steps:

[0229] 3.1 Data preprocessing

[0230] Convert each triple (v i , r, v j , s) into a multi-dimensional vector t, which contains entity attribute and relationship feature information.

[0231] 3.2 Node partitioning

[0232] Use the partitioning algorithm of the R-tree to partition the multi-dimensional vector t into different nodes. Let the set of vectors contained in node N be {t1, t2,..., t m}, and the minimum bounding rectangle (MBR) of the node is calculated by the following formula:

[0233] MBR N = [min 1≤i≤m t i1 , max 1≤i≤m t i1 × [min 1≤i≤m t i2 , max 1≤i≤m t i2 ×... × [min 1≤i≤m t id , max 1≤i≤m t id , where d is the dimension of the multi-dimensional space.

[0234] 3.3 Index construction

[0235] Recursively partition the node into child nodes until a set termination condition is met, such as the number of vectors in the node being less than a certain threshold. Finally, construct the R-tree hybrid index.

[0236] 4. Load the R-tree hybrid index into the initially constructed knowledge graph to form a dynamic attribute knowledge graph of target ecological assets.

[0237] Let the initially constructed knowledge graph be KG0 = (V0, E0, A0), where V0 is the set of initial entities, E0 is the set of initial relationships, and A0 is the set of initial attributes.

[0238] Load each triple (v i , r, v j , s) in the R-tree hybrid index into the initially constructed knowledge graph:

[0239] If then add v i to V0 and add its attribute information to A0.

[0240] If then add v j to V0 and add its attribute information to A0.

[0241] If then add (v i , r, v j ) to E0 and use the relationship strength s as the weight of the relationship.

[0242] Finally, form a dynamic attribute knowledge graph KG = (V, E, A) of target ecological assets, where V = V0, E = E0, and A = A0.

[0243] Therefore, the above replaceable or preferred technical implementation constructs a dynamic attribute knowledge graph of ecological assets with graph attention network (GAT), relationship strength quantization model, and R-tree hybrid index as the core. GAT, based on the attention mechanism, mines the potential semantic relationships of ecological entities by adaptively learning the association weights between nodes; the relationship strength quantization model accurately measures the tightness of relationships such as ecological dependence and material transfer with the help of a multi-layer perceptron (MLP) and weighted calculation; the R-tree hybrid index maps semantic relationship triples to a multi-dimensional space to achieve efficient knowledge storage and retrieval.

[0244] Compared with traditional technologies, it has the following technical advantages:

[0245] 1. Traditional ecological relationship analysis relies on manually coded rules or simple association rule mining. For example, in the study of forest ecosystems, a correspondence table of "tree species - inhabiting birds" is set manually, or the Apriori algorithm is used to mine species symbiotic relationships based on historical data. Such methods require predefined relationship types, making it difficult to discover special ecological dependencies (such as allelopathy between plants) and dynamic material transfers (such as seasonal nutrient cycling), and they cannot adapt to the dynamic changes of ecosystems.

[0246] This application uses GAT and a multi - layer perceptron (MLP) to construct a relationship reasoning model, such as By performing multi - layer attention calculations on node features, it automatically captures the non - linear associations between entities. For example, in wetland ecosystems, it can identify complex dependency chains such as "reed community - benthic organisms - water purification". At the same time, by classifying relationship feature vectors in combination with ecological domain knowledge, it can discover new ecological relationships that are difficult to detect by traditional methods, providing support for the systematic study of ecosystems.

[0247] 2. Traditional relationship strength assessment mostly uses subjective scoring methods or single - indicator calculations. For example, when evaluating species competition relationships, the competition intensity coefficient is only assigned based on the proportion of species numbers, ignoring the influence of ecological factors (such as temperature, humidity) on the relationship. This approach lacks objectivity and comprehensiveness and cannot accurately reflect the dynamic changes of ecological relationships.

[0248] This application uses a quantization model that combines a weighted feature function and a softmax function, such as to comprehensively calculate the relationship strength from multiple dimensions (such as the amount of material exchange, energy flow efficiency, niche overlap). In marine ecosystems, it can accurately quantify the seasonal fluctuations in the material transfer strength between "phytoplankton - zooplankton", providing accurate data support for ecological carrying capacity assessment and resource management.

[0249] 3. Traditional ecological knowledge graphs mostly use relational databases or simple graph storage structures, such as using SQL tables to store entity relationship attribute data. When dealing with large - scale dynamic data, this approach has low query efficiency (such as retrieving cross - regional ecological chains requires traversing a large number of tables), and it is difficult to support complex spatial relationship queries (such as finding all ecological nodes with material transfer relationships within a certain basin).

[0250] This application is based on the R-tree hybrid index, which maps semantic relation triples to a multi-dimensional space to construct a Minimum Bounding Rectangle (MBR), as shown in (Tex translation failed). In the forest ecological monitoring scenario, when querying the "ecological dependence network affected by a certain pollutant at an altitude of 800 - 1200 meters", the R-tree index can quickly locate the eligible nodes and relationships, and the retrieval efficiency is increased by dozens of times. At the same time, this index supports dynamic updates to adapt to the real-time changes of ecological data and ensure the timeliness of the knowledge graph.

[0251] 4. After the traditional knowledge graph is constructed, its structure is fixed, making it difficult to integrate newly discovered ecological entities or relationships. For example, when a new species invades an ecosystem, it is necessary to manually modify the data structure and association rules, resulting in a long update cycle and prone to data inconsistency problems.

[0252] The dynamic attribute knowledge graph of this application continuously learns new data features through GAT and automatically discovers the relationships between newly added entities; the R-tree hybrid index supports incremental updates, and new semantic relation triples can be inserted into the index structure in real time. In the urban green space ecosystem, as new plant varieties are introduced in the greening project, the system can automatically identify their ecological relationships with local species and quickly update the knowledge graph to provide an immediate decision-making basis for ecological planning.

[0253] Optionally, based on the ecological asset evaluation model, dynamically evaluate the value of the target ecological asset according to the dynamic attribute knowledge graph to generate dynamic value sequence points and add them to the dynamic value tree, including:

[0254] Based on the graph embedding layer in the ecological asset evaluation model, map the entities and relationships in the dynamic attribute knowledge graph to a low-dimensional vector space to obtain the semantic association relationships between ecological asset elements;

[0255] Based on the sequence decision layer in the ecological asset evaluation model, extract the time features of the semantic association relationships between ecological asset elements;

[0256] Based on the asset evaluation layer in the ecological asset evaluation model, generate dynamic value sequence points according to the extracted time features and add them to the dynamic value tree.

[0257] Preferably, in a scenario, the above technical implementation is described in a replaceable or preferred manner.

[0258] 1. Based on the graph embedding layer in the ecological asset evaluation model, map the entities and relationships in the dynamic attribute knowledge graph to a low-dimensional vector space to obtain the semantic association relationships between ecological asset elements

[0259] Let the dynamic attribute knowledge graph be G=(V, E), where V={v1, v2, …, v n} is a set of entity nodes, and each entity node v i carries an attribute feature vector E = {(v i , r ij , v j )} is a set of relationship edges, and r ij represents the relationship type between entity v i and v j , and the relationship edge carries a weight w ij indicating the relationship strength. Taking the forest ecosystem as an example, entity v i is "Korean pine", and the attribute feature vector x i contains information such as tree age, diameter at breast height, and volume; the relationship edge (v i , r ij , v j ) represents the "habitat provision" relationship between "Korean pine" and "squirrel", and the weight w ij reflects the degree of dependence.

[0260] The graph embedding layer adopts a fusion model of graph convolutional network (GCN) and multi-head attention mechanism (MultiHeadAttention). For the l-th layer of graph convolutional operation, the feature update formula of node v i is: Among them: is the hidden feature vector of node v i at the l-th layer. Initially, is the set of neighbor nodes of node v i ; c ij is a normalization constant, usually used to balance the influence of different node degrees; W (l) is the learnable weight matrix of the l-th layer, and b (l) is the bias vector; σ is the activation function, such as the ReLU function σ(x) = max(0, x). On this basis, the multi-head attention mechanism is introduced to enhance the ability to capture relationship semantics. The k-th attention head calculates the attention coefficient of node v i and neighbor node v j as

[0261] Among them: is the learnable weight matrix of the k-th attention head; a k is the learnable attention vector of the k-th attention head; || represents the vector concatenation operation; LeakyReLU is the ReLU activation function with leakage to avoid the problem of gradient disappearance.

[0262] Aggregate the results of K attention heads to obtain node v iThe final embedded vector z i :

[0263] where W out is the output weight matrix for dimensionality reduction. Through the above operations, entities and relationships are mapped into a low-dimensional vector space, and the geometric relationships such as the distance and angle between the low-dimensional vectors of entities and relationships can reflect the semantic association relationships between ecological asset elements. For example, the distance between the embedded vectors of "Korean pine" and "Larch" in space can reflect the similarity degree of their ecological functions.

[0264] 2. Extract time features from the semantic association relationships between ecological asset elements based on the sequential decision-making layer in the ecological asset evaluation model

[0265] Let the sequence of ecological asset element embedded vectors obtained through the graph embedding layer be Z = {z1, z2, …, z T}, where T is the time step. In the wetland ecosystem monitoring scenario, Z records the semantic association states of various ecological elements (such as reed communities, fish populations, water quality indicators) in the wetland at different time points. The sequential decision-making layer adopts an architecture that combines a gated recurrent unit (GRU) and an attention mechanism. For the t-th time step, the update formula of the GRU unit is:

[0266] r t = σ(W xr z t + W hr h t-1 + b r )

[0267] z′ t = tanh(W xz z t + W hz (r t ⊙ h t-1 ) + b z )

[0268] u t = σ(W xu z t + W hu h t-1 + b u )

[0269] h t = u t ⊙ h t-1 + (1 - u t ) ⊙ z′ t

[0270] 1. Reset Gate calculation: r t = σ(Wxr z t +W hr h t-1 +b r )

[0271] Reset gate r t For determining the hidden state h at the previous moment t-1 How much information is retained in the calculation of the current candidate hidden state. When r t tends to 0, it means that most of the hidden state information at the previous moment is discarded; when r t tends to 1, it means that a large amount of hidden state information at the previous moment is retained.

[0272] During the calculation, first, multiply the input vector z t by the weight matrix W xr to obtain the contribution of the input to the reset gate; meanwhile, multiply the hidden state h at the previous moment t-1 by the weight matrix W hr to obtain the contribution of the hidden state at the previous moment to the reset gate. Then, add these two parts of contributions together and add the bias vector b r . Finally, map the result to the interval (0, 1) through the Sigmoid activation function σ to obtain the reset gate vector r t .

[0273] 2. Candidate Hidden State: z t′ = tanh(W xz z t + W hz (r t ⊙ h t-1 ) + b z )

[0274] The candidate hidden state z t′ combines the current input z t and the hidden state information at the previous moment filtered by the reset gate, and is used to generate the hidden state at the current moment.

[0275] During the calculation, first multiply the input vector z t by the weight matrix W xz to obtain the contribution of the input to the candidate hidden state; then multiply the hidden state at the previous moment processed by the reset gate r t ⊙ h t-1 (⊙ represents element-wise multiplication) by the weight matrix W hz to obtain the contribution of this part to the candidate hidden state. Add these two parts of contributions together and add the bias vector b z, finally, the result is mapped to the interval (-1, 1) through the hyperbolic tangent activation function tanh to obtain the candidate hidden state vector z t′ .

[0276] 3. Update Gate: u t = σ(W xu z t + W hu h t-1 + b u )

[0277] The update gate u t determines the contribution ratios of the previous hidden state h t-1 and the current candidate hidden state z t′ to the current hidden state h t respectively. When u t approaches 1, a large amount of information of the previous hidden state h t-1 will be retained; when u t approaches 0, a large amount of information of the current candidate hidden state z t′ will be retained.

[0278] In the calculation, similar to the calculation of the reset gate, the input vector z t is multiplied by the weight matrix W xu , the previous hidden state h t-1 is multiplied by the weight matrix W hu , the contributions of these two parts are added together and the bias vector b u is added, and finally the update gate vector u t is obtained through the Sigmoid activation function σ.

[0279] 4. Current Hidden State: h t = u t ⊙ h t-1 + (1 - u t ) ⊙ z t′

[0280] The previous hidden state h t and the candidate hidden state z t-1 are weighted and combined through the update gate u t′ to obtain the hidden state h t at the current moment.

[0281] In the calculation, the update gate vector u t is element-wise multiplied by the previous hidden state h t-1 to obtain the retained part of the previous hidden state; 1 - u t is element-wise multiplied by the candidate hidden state z t′Perform element-wise multiplication to obtain the retained part of the candidate hidden state. Finally, add these two parts together to get the hidden state vector h at the current time step. t 。

[0282] z t : The input vector at time step t. In the scenario of ecological asset dynamic monitoring and assessment, it may be a low-dimensional vector representing the semantic associations of ecological asset elements after being processed by the graph embedding layer, with a dimension of input_size. h t-1 : The hidden state vector at time step t-1, with a dimension of hidden_size, which records the information from previous time steps and is used for the calculation at the current time step. h t : The hidden state vector at time step t, with a dimension of hidden_size, which is the output of the GRU unit at the current time step and will be used for the calculation at the next time step and subsequent temporal feature extraction and analysis. W xr : The input z t to the weight matrix of the reset gate, with a dimension of hidden_size × input_size. W hr : The weight matrix from the hidden state h at the previous time step t-1 to the reset gate, with a dimension of hidden_size × hidden_size. b r : The bias vector of the reset gate, with a dimension of hidden_size. W xz : The input z t to the weight matrix of the candidate hidden state, with a dimension of hidden_size × input_size. W hz : The hidden state h at the previous time step after being processed by the reset gate r t ⊙h t-1 to the weight matrix of the candidate hidden state, with a dimension of hidden_size × hidden_size. b z : The bias vector of the candidate hidden state, with a dimension of hidden_size. W xu : The input z t to the weight matrix of the update gate, with a dimension of hidden_size × input_size. W hu : The weight matrix from the hidden state h at the previous time step t-1 to the update gate, with a dimension of hidden_size × hidden_size. b u : The bias vector of the update gate, with a dimension of hidden_size. σ: The Sigmoid activation function, which maps the input to the interval (0, 1). tanh: The hyperbolic tangent activation function, which maps the input to the interval (-1, 1).

[0283] In the dynamic monitoring and assessment of ecological assets, GRU units are used to process time series data of the semantic association relationships between ecological asset elements. For example, after obtaining the low-dimensional vector representation of ecological asset elements through the graph embedding layer, it is used as the input z of the GRU t . As time goes by, the GRU unit continuously updates the hidden state h t , so as to capture the change characteristics of the semantic associations of ecological asset elements at different time steps. These features can be used in the subsequent asset evaluation layer to conduct dynamic value assessment of ecological assets based on the extracted time features, providing strong support for the management and decision-making of ecological assets

[0284] To further capture key time features, a time attention mechanism is introduced. Calculate the attention coefficient β of time step t to other time steps tt′ :

[0285] where Score(h t ,h t′ ) is a scoring function, which can adopt the dot product form or the additive form w s ,W hs ,W ht ,b s are learnable parameters

[0286] The final time feature vector f t is obtained through weighted summation

[0287] 3. Based on the asset evaluation layer in the ecological asset evaluation model, dynamic value sequence points are obtained according to the extracted time features and added to the dynamic value tree

[0288] The asset evaluation layer adopts a value evaluation model based on reinforcement learning. Let the state space be the time feature vector f t , and the action space be value evaluation decisions (such as value growth prediction, value risk assessment, etc.). Define the value evaluation function V(f t ), and use the deep Q-network (DQN) architecture for modeling: Q(f t ,a;θ) = w T MLP(f t ,a;θ), where: a is the action; θ is the network parameter; w is the output layer weight vector; MLP(f t ,a;θ) is a multi-layer perceptron, which takes the time feature vector and the action as inputs and outputs the Q value

[0289] Update the network parameter θ through the Bellman equation: y t = r t +γmax a′Q(f t+1 , a′; θ - ) where:

[0290] r t is the immediate reward, which can be set according to the actual value change of ecological assets in ecological asset evaluation; γ is the discount factor, used to balance the immediate reward and future rewards; θ - is the target network parameter, which is periodically copied and updated from the current network θ; L(θ) is the loss function, which is optimized by the stochastic gradient descent algorithm. After obtaining the value evaluation result V(f t ), it is added to the dynamic value tree as a dynamic value sequence point. The dynamic value tree adopts a quadtree structure, and each node stores the value evaluation information within a time interval. The division of nodes is based on the time granularity and the amplitude of value change. Let the root node of the dynamic value tree be R. For the newly generated dynamic value sequence point V(f t ), recursively search for the corresponding child node according to its timestamp t. If the node capacity is full, perform a node splitting operation to ensure the efficient storage and retrieval of value sequence points.

[0291] Therefore, the above replaceable or preferred technical implementations integrate cutting-edge technologies such as graph neural networks, sequence analysis, and reinforcement learning to construct an ecological asset dynamic value evaluation model. The graph embedding layer uses graph convolutional networks and multi-head attention mechanisms to map the entities and relationships in the knowledge graph to a low-dimensional vector space, and mines the semantic associations between ecological asset elements through node feature updates and attention coefficient calculations; the sequence decision layer uses gated recurrent units and temporal attention mechanisms to extract features from the embedded vectors of time series and capture the dynamic features of ecological assets changing over time; the asset evaluation layer is based on the deep Q-network of reinforcement learning, takes time features as states, and iteratively optimizes through the value evaluation function and Bellman equation to achieve the dynamic value evaluation of ecological assets, and stores them in a dynamic value tree with a quadtree structure. Compared with traditional technologies, it has the following technical advantages in data processing, feature extraction, and value evaluation

[0292] 1. Traditional methods usually identify ecological asset relationships based on manually defined rules or simple statistical analysis. For example, in forest ecological assessment, an "inter-species symbiotic relationship table" is formulated through expert experience, or association rule mining algorithms are used to analyze the occurrence frequencies of species to determine relationships. This method relies on prior knowledge, is difficult to discover complex and implicit ecological associations, and cannot adapt to the dynamic changes of the ecosystem.

[0293] This application uses a graph convolutional network and a multi-head attention mechanism, which can automatically learn the complex relationships between nodes. In the forest ecosystem, it can uncover indirect associations such as "nutrient cycling of Korean pine and soil microorganisms", as well as potential similarities in the ecological functions of different tree species (e.g., the functional similarity is reflected by the embedding vector distance between "Korean pine" and "larch"). Compared with traditional methods, the new technology can more comprehensively and deeply analyze the semantic relationships between ecological asset elements, providing rich data for the systematic study of ecosystems.

[0294] Traditional time series analysis mostly uses methods such as moving average and autoregressive integrated moving average (ARIMA) to perform simple trend fitting and prediction on ecological data. For example, in wetland ecological monitoring, only the average value of historical water quality data is used to predict future water quality changes, without considering the interactive effects and complex dynamic processes of multiple elements in the ecosystem.

[0295] This application uses a gated recurrent unit and a temporal attention mechanism, which can adaptively learn the long-term and short-term dependencies of the time series of ecological assets. In the wetland ecosystem, it can simultaneously capture the change characteristics of multiple elements such as the growth of reed communities, the migration of fish populations, and the fluctuations of water quality indicators at different time scales, and focus on key time points through attention coefficients (such as the impact of flood periods on wetland ecology). Compared with traditional methods, the new technology can more accurately depict the dynamic evolution process of the ecosystem, improving the prediction and response capabilities to ecological changes.

[0296] Traditional ecological asset assessment relies on a static assessment index system and fixed weight allocation. For example, the carbon sequestration value and biodiversity value of the forest ecosystem are determined through the expert scoring method, and then the total value is obtained by weighted summation. This method lacks consideration of the dynamic changes of ecological assets and cannot reflect the fluctuations of ecosystem value in real time.

[0297] This application is based on the deep Q-network of reinforcement learning, taking the time characteristics of ecological assets as the state, and optimizing the value assessment strategy through continuous trial and error and reward feedback. In practical applications, the value assessment results can be dynamically adjusted according to the real-time state of ecological assets (such as ecological damage after a forest fire and ecological improvement after a wetland restoration project). At the same time, a dynamic value tree with a quadtree structure is used for storage, supporting efficient query and update of value sequence points, and being able to quickly respond to changes in the ecosystem, providing timely and accurate value references for the dynamic management and decision-making of ecological resources.

[0298] The above description is only a preferred embodiment of the present application and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover 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 the technical features (but not limited to) disclosed in the present application that have similar functions.

Claims

1. An ecological asset dynamic monitoring and evaluation method, characterized in that Including: Dynamically collect the attribute features of the target ecological assets based on the ecological asset dynamic monitoring model and form a dynamic attribute feature sequence accordingly; Construct a dynamic attribute knowledge graph based on the dynamic attribute feature sequence; Based on the ecological asset evaluation model, dynamically evaluate the value of the target ecological assets according to the dynamic attribute knowledge graph to generate dynamic value sequence points and add them to the dynamic value tree.

2. The ecological asset dynamic monitoring and evaluation method according to claim 1, characterized in that Dynamically monitor the action of dynamically collecting the attribute features of the target ecological assets based on the ecological asset dynamic monitoring model and forming a dynamic attribute feature sequence accordingly, including: Obtain the attribute feature carriers collected by the attribute sensors and extract the features of the attribute feature carriers to obtain the attribute features of the target ecological assets.

3. The ecological asset dynamic monitoring and evaluation method according to claim 1, characterized in that Obtain the attribute feature carriers collected by the attribute sensors and extract the features of the attribute feature carriers to obtain the attribute features of the target ecological assets, including: Perform quantum state fusion on the obtained attribute feature carriers to generate a quantum state mixed data set; Map the quantum state mixed data set into dynamic spatio-temporal graph structure data; Perform cross-domain mapping on the dynamic spatio-temporal graph structure data, endow cross-domain semantic information to the dynamic spatio-temporal graph structure data based on graph attention to generate a core feature subset with evaluation value; Input the core feature subset into the encoder to learn the probability distribution representation of the features and generate a latent feature vector; The generator generates enhanced feature samples according to the latent feature vector, and the discriminator distinguishes between real core features and generated feature samples; The variational autoencoder reconstructs and optimizes the real core features and generated feature samples to generate the attribute features of the target ecological assets.

4. The ecological asset dynamic monitoring and evaluation method according to claim 1, wherein Construct a dynamic attribute knowledge graph based on the dynamic attribute feature sequence, including: Perform entity extraction, relationship extraction, and attribute extraction on the dynamic attribute feature sequence to obtain the entities of the ecological assets, the semantic relationships between entities, and the attribute information of entities respectively; Preprocess the extracted entities, and update the semantic relationships and attribute information based on the preprocessed entities. The preprocessing includes alignment and disambiguation processing; Load the entities, the semantic relationships between entities, and the attribute information of entities into a graph data structure and load them into the constructed initial state knowledge graph to form the dynamic attribute knowledge graph of the target ecological assets.

5. The ecological asset dynamic monitoring and evaluation method according to claim 4, wherein, Perform entity extraction, relationship extraction, and attribute extraction on the dynamic attribute feature sequence to obtain the entities of the ecological assets, the semantic relationships between entities, and the attribute information of entities respectively, including: Convert the dynamic attribute feature sequence into a superposition state representation sequence; Perform multi-scale decomposition on the superposition state representation sequence to obtain a quantum feature set containing multi-scale detailed features; Abstract each feature element in the quantum feature set into a molecular unit; Perform semantic feature aggregation between molecular units to generate entity candidate clusters; Perform spatio-temporal evolution pattern recognition on the entity candidate clusters to generate a relationship pre-mining result; Based on the connectivity and path constraints of the constructed topological graph, perform directional entity extraction, relationship extraction, and attribute extraction on the entity candidate clusters and the pre-mined relationship results to respectively obtain the entities of ecological assets, the semantic relationships between entities, and the attribute information of entities.

6. The ecological asset dynamic monitoring and evaluation method according to claim 4, characterized in that Load the entities, the semantic relationships between entities, and the attribute information into a graph data structure and then load it into the constructed initial knowledge graph to form a dynamic attribute knowledge graph of the target ecological assets, including: Use a graph attention network to infer the ecological dependence semantic relationships and material transfer semantic relationships among entities, the semantic relationships between entities, and the attribute information of entities; Quantify the relationship strengths of the ecological dependence semantic relationships and material transfer semantic relationships to encapsulate them into weighted semantic relationship triples; Based on the semantic relationship triples, construct an R-tree hybrid index; Load the R-tree hybrid index into the constructed initial knowledge graph to form a dynamic attribute knowledge graph of the target ecological assets.

7. The ecological asset dynamic monitoring and evaluation method according to claim 1, characterized in that Based on the ecological asset evaluation model, perform dynamic asset value evaluation on the target ecological assets according to the dynamic attribute knowledge graph to generate dynamic value sequence points and add them to the dynamic value tree, including: Based on the graph embedding layer in the ecological asset evaluation model, map the entities and relationships in the dynamic attribute knowledge graph into a low-dimensional vector space to obtain the semantic association relationships among ecological asset elements; Based on the sequence decision layer in the ecological asset evaluation model, extract time features from the semantic association relationships among ecological asset elements; Based on the asset evaluation layer in the ecological asset evaluation model, generate dynamic value sequence points according to the extracted time features and add them to the dynamic value tree.

Images