A method and system for digital evolution of an ecosystem

Abstract

The application provides a method and system for digital evolution of an ecosystem, which comprises: obtaining information about animals and plants, environmental quality assessment and ecosystem service function assessment results in an ecological protection zone according to biological data and environmental data collected in the ecological protection zone; generating a comprehensive feature matrix based on features extracted from the information about animals and plants, the environmental quality assessment and the ecosystem service function assessment results, and performing deep learning on the comprehensive feature matrix to construct a digital twin engine; constructing a digital twin large model based on the digital twin engine, static base information of a large model and a prediction model, the prediction model supporting natural evolution prediction and scenario deduction; in response to real-time data input into the digital twin large model, processing the real-time data and target data including ecosystem historical data, and displaying ecosystem panoramic dynamic information including natural evolution prediction results and / or scenario deduction results on a visual interface.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ecology, and in particular to a digital evolution method and system of an ecological system. BACKGROUND

[0002] With the improvement of ecological protection awareness, monitoring the ecological system has become the key to maintaining biodiversity and ecological balance. Real-time monitoring of the ecological system can provide early warning of environmental degradation, assessment of ecological service functions, and provide a basis for ecological restoration and sustainable management.

[0003] With the rapid development of digital technology, digital twin technology has gradually emerged and been applied in many fields. A digital twin is a virtual digital model corresponding to a physical entity, which is constructed by integrating physical models, sensor data, historical data and other multi-source information, and can reflect the state of the physical entity in real time and predict its future behavior. In the field of ecology, although there have been related technologies trying to build digital twins of ecological systems to monitor the ecological system, there are still many challenges.

[0004] 1. The data collection is single, mostly focusing on the collection of environmental data, ignoring the importance of biological data, resulting in an incomplete assessment of the ecological system function; 2. Due to the wide range of ecological systems and complex environment, there are problems such as difficulty in data collection and poor real-time performance; 3. At the data fusion level, different modal data often have different data structures and time resolutions, and simple data superposition cannot fully utilize the advantages of multi-source data, making it difficult to obtain comprehensive data that accurately reflect the state of the ecological system; 4. The comprehensive data obtained by simple data fusion cannot reflect the true state of the ecological system, resulting in distorted prediction input and ineffective prediction results; 5. Data collaboration in the fields of ecology, geographic information, computer science and other fields is difficult, and it is difficult to form a sustainable digital twin ecological circle.

[0005] Therefore, there are still many problems to be solved in applying digital twin technology to the ecological field to reflect the state of the ecological system and predict the state of the ecological system, and a feasible solution is urgently needed. SUMMARY

[0006] The present application aims to provide a digital evolution method and system of an ecological system to solve the problems existing in the prior art. The technical problems to be solved by the present application are solved by the following technical solutions.

[0007] In a first aspect, the present application provides a digital evolution method of an ecological system, comprising:

[0008] According to the biological data and environmental data collected in the ecological protection area, the information of animals and plants, the environmental quality assessment and the ecological system service function assessment results in the ecological protection area are obtained;

[0009] generate a comprehensive feature matrix based on the features extracted from the animal and plant information, the environment quality assessment situation, and the ecosystem service function assessment result, and perform deep learning on the comprehensive feature matrix to construct a digital twin engine, wherein the comprehensive feature matrix is composed of fusion feature vectors of multiple spatio-temporal units arranged in a spatio-temporal sequence, and a fusion feature vector of a single spatio-temporal unit is formed by multi-modal feature splicing of the animal and plant information, the environment quality assessment situation, and the ecosystem service function assessment result corresponding to the current spatio-temporal unit;

[0010] construct a digital twin large model based on the digital twin engine, large model static base information, and a prediction model, wherein the prediction model supports natural evolution prediction and scenario deduction under a set scenario;

[0011] In response to inputting real-time data of the digital twin large model, process the real-time data and target data, and display ecosystem panoramic dynamic information on a visual interface, wherein the target data includes ecosystem historical data, the visual interface is a virtual reality (VR) interface or an augmented reality (AR) interface, and the ecosystem panoramic dynamic information includes at least one of a natural evolution prediction result and a scenario deduction result.

[0012] In a second aspect, an embodiment of the present application provides a digital evolution system of an ecosystem, including:

[0013] An acquisition module is configured to acquire animal and plant information, environment quality assessment situation, and ecosystem service function assessment result in an ecological protection area based on biological data and environmental data collected in the ecological protection area.

[0014] A generation and learning module is configured to generate a comprehensive feature matrix based on features extracted from the animal and plant information, the environment quality assessment situation, and the ecosystem service function assessment result, and perform deep learning on the comprehensive feature matrix to construct a digital twin engine, wherein the comprehensive feature matrix is composed of fusion feature vectors of multiple spatio-temporal units arranged in a spatio-temporal sequence, and a fusion feature vector of a single spatio-temporal unit is formed by multi-modal feature splicing of the animal and plant information, the environment quality assessment situation, and the ecosystem service function assessment result corresponding to the current spatio-temporal unit.

[0015] A construction module is configured to construct a digital twin large model based on the digital twin engine, large model static base information, and a prediction model, wherein the prediction model supports natural evolution prediction and scenario deduction under a set scenario.

[0016] The processing display module is configured to process the real-time data and target data in response to inputting the real-time data of the digital twin large model, and display ecosystem panoramic dynamic information in a visual interface, the target data including ecosystem historical data, the visual interface being a virtual reality (VR) interface or an augmented reality (AR) interface, and the ecosystem panoramic dynamic information including at least one of a natural evolution prediction result and a scenario deduction result.

[0017] The technical scheme provided by the embodiments of the present application can analyze biological data and environmental data to obtain information of animals and plants in the ecological protection zone, environmental quality evaluation conditions, and ecosystem service function evaluation results, extract features from the information of animals and plants, the environmental quality evaluation conditions, and the ecosystem service function evaluation results to generate a comprehensive feature matrix, perform deep learning on the comprehensive feature matrix to construct a digital twin engine, construct a digital twin large model based on the digital twin engine, static base information of the large model, and a prediction model supporting natural evolution prediction and scenario deduction under a set scenario, process target data and real-time data input into the digital twin large model, and display ecosystem panoramic dynamic information including a natural evolution prediction result and / or a scenario deduction result in a visual interface. The digital twin large model can predict the future natural evolution state of the ecosystem and / or deduce the state of the ecosystem under different set conditions, so that the digital twin large model can more comprehensively and accurately reflect the change trend of the ecosystem, and provide strong support for ecological protection and management. The content provided by the digital twin large model can be visually displayed through VR technology or AR technology, and an immersive user experience can be provided. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A digital evolution method of an ecosystem provided by an embodiment of the present application is shown;

[0019] Figure 2 A simplified flowchart of processing real-time data and target data by a digital twin large model provided by an embodiment of the present application is shown;

[0020] Figure 3 An overall implementation flowchart of a digital evolution method of an ecosystem provided by an embodiment of the present application is shown;

[0021] Figure 4 An interaction architecture diagram provided by an embodiment of the present application is shown;

[0022] Figure 5 A digital evolution system of an ecosystem provided by an embodiment of the present application is shown. DETAILED DESCRIPTION

[0023] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0024] The embodiment of the present application provides a digital evolution method of an ecological system, as shown in the figure, comprising the following steps: Figure 1

[0025] Step 101, obtaining the information of animals and plants, the environmental quality assessment situation and the evaluation result of the ecosystem service function in the ecological protection zone according to the biological data and environmental data collected in the ecological protection zone.

[0026] After selecting the ecological protection zone, the embodiment of the present application uses the collection equipment (such as unmanned equipment equipped with various sensors) to collect biological data and environmental data in the ecological protection zone within one or more specific time periods. After obtaining the biological data and environmental data through data collection, the animals and plants in the ecological protection zone are identified based on the biological data, the environmental quality in the ecological protection zone is evaluated based on the environmental data, to obtain the information of animals and plants and the environmental quality assessment situation in the ecological protection zone, and the evaluation result of the ecosystem service function is obtained by comprehensively considering the biological data and the environmental data.

[0027] Through the identification of animals and plants in the ecological protection zone, the information of the species, quantity and distribution of animals and plants in the ecological protection zone can be understood; through the environmental quality assessment of the ecological protection zone, it is beneficial to timely find the environmental problems in the ecological protection zone, and provide a scientific basis for ecological protection and management; through the evaluation result of the ecosystem service function, the complex interaction between organisms and environment can be revealed, and the overall condition of the ecological system can be more comprehensively understood.

[0028] Step 102, generating a comprehensive feature matrix based on the features extracted from the information of animals and plants, the environmental quality assessment situation and the evaluation result of the ecosystem service function, and constructing a digital twin engine through deep learning of the comprehensive feature matrix, wherein the comprehensive feature matrix is composed of fusion feature vectors of multiple space-time units arranged in space-time sequence, and the fusion feature vector of a single space-time unit is formed by multi-modal feature splicing of the information of animals and plants, the environmental quality assessment situation and the evaluation result of the ecosystem service function corresponding to the current space-time unit.

[0029] ​In the case of obtaining the information of animals and plants, the environmental quality assessment situation and the ecosystem service function assessment result in the ecological protection zone by analyzing biological data and environmental data, feature extraction is performed in the animal and plant information, the environmental quality assessment situation and the ecosystem service function assessment result, and a comprehensive feature matrix is generated after the extracted biological layer features, environmental layer features and service layer features are aligned in time and space. On the basis of the comprehensive feature matrix, deep learning is performed on the comprehensive feature matrix to construct a digital twin engine.

[0030] The digital twin engine constructed by deep learning includes a dynamic calculation layer and a rule engine layer, which are respectively responsible for real-time calculation and logical decision, and constitute the core function of the digital twin engine; and the dynamic calculation layer is located at the bottom and is responsible for basic calculation, and the rule engine layer is located at the upper layer and is responsible for logical judgment and decision support.

[0031] In the process of constructing the digital twin engine by deep learning on the comprehensive feature matrix, the comprehensive feature matrix is taken as the input, the dynamic calculation layer is formed by deep learning, the rule engine layer is integrated with the dynamic calculation layer as the core, and the digital twin engine is encapsulated. The constructed digital twin engine has the capabilities of real-time data access, state calculation, rule judgment and response triggering, is the core driving module of the digital twin large model, supports the dynamic operation and intelligent decision of the large model.

[0032] Step 103, constructing a digital twin large model based on the digital twin engine, the large model static base information and the prediction model, the prediction model supports natural evolution prediction and scenario deduction under the set scenario.

[0033] The digital twin engine constructed by deep learning is the core calculation and decision module of the digital twin large model, and is responsible for real-time data processing, state calculation, rule response and other dynamic functions. The large model static base information provides the spatial structure and basic attributes of the large model, including information such as terrain, vegetation type, soil texture, etc. These information provides a spatial reference framework for the digital twin engine, so that the calculation results can be accurately positioned and displayed in geographical space. The prediction model is a model that supports natural evolution prediction and scenario deduction under the set scenario, that is, the prediction model can not only predict the natural evolution trend of the ecosystem, but also simulate the response of the ecosystem under different environments or policy interventions by setting different input conditions (scenario parameters), so as to realize scenario deduction; by predicting the future state of the ecosystem under natural evolution, and / or, scenario deduction under the set conditions, it can provide a strong basis for intelligent decision and sustainable management.

[0034] The digital twin engine is responsible for calculation and response; the large model static base information is a skeleton that can provide spatial structure and basic attributes; the prediction model is used to predict the future state of the ecosystem and deduce the state changes of the ecosystem under different environmental or intervention conditions; the combination of the digital twin engine, the large model static base information, and the prediction model is used to construct the digital twin large model.

[0035] It should be noted that the prediction model simulates the response of the ecosystem under different environments or policy interventions by setting different input conditions when performing scenario deduction, realizes scenario deduction, and belongs to a prediction form. Scenario deduction is essentially a hypothesis-deduction, and the core use is to predict the future (such as how the ecosystem will evolve given a certain scenario), but it can also be used for historical scenario review (such as reproducing the historical state using real scenario parameters (such as extreme drought) in a certain year in the past to verify the reliability of the digital twin large model) and sensitivity analysis (such as quickly testing the immediate impact of different parameter combinations on the ecosystem at the current time point). The embodiment of the present application takes the future state deduction as an example to introduce scenario deduction.

[0036] In step 104, in response to inputting real-time data of the digital twin large model, the real-time data and target data are processed, and the panoramic dynamic information of the ecosystem is displayed on a visual interface, the target data includes historical data of the ecosystem, the visual interface is a virtual reality (VR) interface or an augmented reality (AR) interface, and the panoramic dynamic information of the ecosystem includes at least one of natural evolution prediction results and scenario deduction results.

[0037] After constructing the digital twin large model, real-time data is input to the digital twin large model, and the digital twin large model processes the input real-time data and target data, the real-time data includes real-time plant and animal information, real-time environmental quality evaluation, and real-time ecosystem service function evaluation results input to the digital twin large model through a data interface (such as an Internet of Things interface), and the target data includes historical data of the ecosystem, which can provide long-term trends and background information of the ecosystem to assist the digital twin large model in more accurate analysis. By processing the real-time data and the target data by the digital twin large model, the panoramic dynamic information of the ecosystem can be output and displayed on a visual interface.

[0038] The visual interface in this embodiment is a VR (Virtual Reality) interface or an AR (Augmented Reality) interface, which uses VR technology or AR technology for visual display, converts abstract data into intuitive three-dimensional images or augmented reality images, so that users can intuitively understand the panoramic dynamic information of the ecosystem through the VR interface or the AR interface, and ensure the visual experience of the users.

[0039] The dynamic information of the ecosystem panorama displayed on the visual interface includes at least one of the natural evolution prediction result and the scenario deduction result. By analyzing real-time data and historical data of the ecosystem through the digital twin large model, the future state of the ecosystem under the natural evolution scenario can be predicted, potential ecological risks can be identified in advance, preventive measures can be taken to reduce the vulnerability of the ecosystem, and resource allocation can be planned in advance to optimize ecological protection and restoration measures and improve resource utilization efficiency. Through scenario deduction by the digital twin large model with the aid of real-time data and historical data of the ecosystem, the state of the ecosystem under different conditions can be fed back, providing strong support for ecological protection decision-making and resource optimization.

[0040] After the biological data and environmental data are analyzed to obtain the information of animals and plants in the ecological protection area, the environmental quality assessment situation, and the ecological system service function evaluation result, features are extracted from the information of animals and plants, the environmental quality assessment situation, and the ecological system service function evaluation result to generate a comprehensive feature matrix. Deep learning is performed on the comprehensive feature matrix to construct a digital twin engine. Based on the digital twin engine, a large model static base information, and a prediction model supporting natural evolution prediction and scenario deduction under a set scenario, a digital twin large model is constructed. The digital twin large model processes target data and real-time data input into the digital twin large model, and displays the dynamic information of the ecosystem panorama including the natural evolution prediction result and / or the scenario deduction result on a visual interface. The future natural evolution state of the ecosystem can be predicted and / or the state of the ecosystem under different set conditions can be deduced, so that the digital twin large model more comprehensively and accurately reflects the changing trend of the ecosystem, providing strong support for ecological protection and management. The content provided by the digital twin large model can be visually displayed through VR technology or AR technology, providing an immersive user experience.

[0041] The process of obtaining the information of animals and plants in the ecological protection area, the environmental quality assessment situation, and the ecological system service function evaluation result is introduced as follows. When obtaining the information of animals and plants, the environmental quality assessment situation, and the ecological system service function evaluation result based on biological data and environmental data, the following steps are included:

[0042] Based on biological data, a biological feature vector is generated, and based on environmental data, an environmental feature vector is generated. The biological data at least includes image data, audio data, and positioning data of animals and plants, and the environmental data at least includes meteorological data and physical environmental data.

[0043] Based on a biological large model, species recognition is performed on the biological feature vector, and based on an environmental large model, environmental analysis is performed on the environmental feature vector to obtain the information of animals and plants in the ecological protection area and the environmental quality assessment situation.

[0044] The system processes information on flora and fauna, environmental quality assessment, biological feature vectors, environmental feature vectors, and comprehensive feature vectors to obtain the assessment results of the ecosystem service functions of the ecological protection area. The comprehensive feature vector is determined based on the fusion of biological feature vectors, environmental feature vectors, and dynamically adjusted biological and environmental weights.

[0045] After acquiring biological and environmental data through data collection, biological feature vectors indicating the biological characteristics within the ecological reserve are generated based on the biological data, and environmental feature vectors indicating the environmental characteristics within the ecological reserve are generated based on the environmental data. The collected biological data includes at least image data, audio data, and location data of plants and animals; the collected environmental data includes at least meteorological data and physical environmental data. For example, environmental data includes, but is not limited to, temperature, humidity, soil conditions, water quality, light intensity, air pressure, wind speed, and seasonal factors. By collecting environmental data, background information on the survival and activity of plants and animals within the ecological reserve can be provided, allowing for a more comprehensive description of the regional characteristics of the ecological reserve in conjunction with the biological data.

[0046] After generating biofeature vectors based on biological data, species identification is performed on these vectors using a pre-constructed large-scale biological model to obtain information on flora and fauna. The large-scale biological model is a species identification model; it is used to infer and analyze the biofeature vectors, obtaining the flora and fauna information within the ecological reserve, thus enabling effective species identification. After generating environmental feature vectors based on environmental data, environmental quality assessment is performed on these vectors using a pre-constructed large-scale environmental model. This large-scale environmental model is an environmental assessment model; it is used to analyze the environmental feature vectors, obtaining the environmental quality assessment results within the ecological reserve. The environmental quality assessment results output by the large-scale environmental model include, for example, an environmental quality score, which is determined based on a comprehensive evaluation of environmental factors such as water quality, soil, and meteorological conditions within the ecological reserve.

[0047] Based on biometric vectors, flora and fauna within ecological reserves can be identified; based on environmental characteristic vectors, environmental quality within ecological reserves can be assessed. Identifying flora and fauna within ecological reserves allows us to understand their species, quantity, and distribution. Assessing environmental quality within ecological reserves helps to promptly identify environmental problems, providing a scientific basis for ecological protection and management.

[0048] After obtaining the information of animals and plants in the ecological protection zone and the environmental quality assessment situation in the ecological protection zone, the biological feature vector, the environmental feature vector, the comprehensive feature vector, the information of animals and plants in the ecological protection zone and the environmental quality assessment situation are processed by using the evaluation large model to evaluate the ecosystem service function of the ecological protection zone, and an evaluation result of the ecosystem service function is obtained.

[0049] The comprehensive feature vector is determined based on fusion of the biological feature vector and the environmental feature vector. The biological feature vector reflects information such as biodiversity and species richness, and the environmental feature vector covers multiple environmental factors. By integrating the biological feature vector and the environmental feature vector to form the comprehensive feature vector, the ecological condition of the ecological protection zone can be more comprehensively evaluated.

[0050] Optionally, when the information of animals and plants, the environmental quality assessment situation, the biological feature vector, the environmental feature vector and the comprehensive feature vector are processed to obtain the evaluation result of the ecosystem service function of the ecological protection zone, the method comprises the following steps:

[0051] The biological feature vector and the environmental feature vector are combined according to the corresponding weights to splice vectors to determine the comprehensive feature vector.

[0052] Based on the optimized evaluation large model, the biological feature vector, the environmental feature vector, the comprehensive feature vector, the information of animals and plants and the environmental quality assessment situation are comprehensively analyzed to generate an evaluation result of the ecosystem service function including an evaluation result of the ecosystem species diversity and an evaluation result of the ecosystem environment.

[0053] The comprehensive feature vector is determined based on fusion of the biological feature vector, the environmental feature vector and dynamically adjusted biological weights and environmental weights. By integrating the biological feature vector and the environmental feature vector to form the comprehensive feature vector, the health condition of the ecological protection zone can be more comprehensively evaluated. Moreover, the formation of the comprehensive feature vector is not only based on the biological feature vector and the environmental feature vector, but also takes into account the dynamically adjusted biological weights and environmental weights. The biological weights and the environmental weights are dynamically adjusted based on the contribution of biological data and environmental data to the evaluation of the ecosystem service function. This dynamic adjustment mechanism can flexibly adjust the weights according to the characteristics and protection targets of different ecological regions, so as to obtain the comprehensive feature vector in a targeted manner. Compared with the prior art in which it is difficult to integrate multi-source heterogeneous data (such as biological data and environmental data) and the weights are fixed and cannot be dynamically adjusted, in the embodiments of the present application, biological data (such as images, audio and positioning) and environmental data (such as weather, soil and water quality) are integrated to generate the biological feature vector and the environmental feature vector, and the comprehensive feature vector is generated based on dynamically adjusted weights, thereby solving the problem of data integration, and using the dynamic weight adjustment mechanism can flexibly adjust the weights according to the real-time contribution of the data to the evaluation of the ecosystem service function, so as to more accurately reflect the overall condition of the ecosystem.

[0054] Exemplarily, a specific example of dynamic weight adjustment based comprehensive feature vector generation is given here:

[0055] 1. Input data and initial feature vectors

[0056] Biological feature vector (dimension=4): B = [0.2, 0.95, 0.15, 0.5] (representing Siberian tiger quantity, confidence, elk quantity, activity range); Environmental feature vector (dimension=4): E = [0.6, 0.65, 0.68, 0.67] (representing temperature, humidity, soil pH, snow depth).

[0057] 2. Dynamic weight adjustment rules

[0058] Default weights (ecological balance scenario): biological weight = 0.5, environmental weight = 0.5;

[0059] Dynamic adjustment conditions: if the quantity of key species (such as Siberian tigers) decreases by more than 10%, then biological weight = 0.7, environmental weight = 0.3; if environmental parameters (such as soil pH) exceed the threshold (<6.5 or >7.5), then environmental weight = 0.8, biological weight = 0.2.

[0060] Example scenario: the current Siberian tiger quantity has decreased by 15% compared to last month (triggering the key species protection rule), and the weight adjustment is: biological weight = 0.7, environmental weight = 0.3.

[0061] 3. Weighted fusion calculation

[0062] Weighted biological feature vector: B_weighted = biological weight * B = 0.7 * [0.2, 0.95, 0.15, 0.5] = [0.14, 0.665, 0.105, 0.35]; Weighted environmental feature vector: E_weighted = environmental weight * E = 0.3 * [0.6, 0.65, 0.68, 0.67] = [0.18, 0.195, 0.204, 0.201].

[0063] Comprehensive feature vector (after concatenation): F = concat(B_weighted, E_weighted) = [0.14, 0.665, 0.105, 0.35, 0.18, 0.195, 0.204, 0.201].

[0064] The evaluation large model is a large model for evaluating ecosystem service functions. The large model is used to intelligently analyze biological feature vectors, environmental feature vectors, comprehensive feature vectors, information of animals and plants in the ecological protection zone, and environmental quality evaluation conditions, so as to reveal the complex interaction between the biological and environmental, and more comprehensively understand the overall condition of the ecological system, and then more comprehensively, more accurately and more deeply evaluate the ecosystem service functions. The evaluation large model can be optimized from the aspects of data quality, model architecture, training strategy, evaluation index and the like, so as to improve the accuracy and reliability of the evaluation large model for evaluating the ecosystem service functions; the evaluation large model can also be pre-trained or migrated to learn on a large amount of related data, so as to improve the performance and generalization ability of the evaluation large model.

[0065] The evaluated ecosystem service functions in the embodiments of the present application include evaluation results of ecosystem species diversity and evaluation results of ecosystem environment, for example, including biodiversity, water purification, climate regulation, soil fertility maintenance and the like. Through comprehensive and deep evaluation of the ecosystem service functions, more scientific decision support can be provided for ecological protection and resource management, so that effective protection strategies can be developed.

[0066] It should be noted that the biological feature vectors and the environmental feature vectors respectively contain biological information and environmental information in the ecological protection zone. The two vectors provide basic data, and more specific biological information (mainly referring to animals and plants) and environmental quality evaluation conditions can be extracted by analyzing these data. The animal and plant information obtained based on the analysis of the biological feature vectors belongs to higher level information, for example, species diversity, distribution of key species and the like; the environmental quality evaluation conditions obtained based on the analysis of the environmental feature vectors also belong to higher level information, for example, water quality compliance, soil pollution degree and the like. Although the animal and plant information and the environmental quality evaluation conditions provide important information, they are results obtained after preliminary analysis, and some details and potential relationships in the original data can be lost. When considering the animal and plant information and the environmental quality evaluation conditions in the ecological protection zone, analyzing the biological feature vectors, the environmental feature vectors and the comprehensive feature vectors can dig out more potential information and mutual relationships.

[0067] By comprehensively analyzing the biological feature vectors, the environmental feature vectors, the comprehensive feature vectors, the animal and plant information in the ecological protection zone, and the environmental quality evaluation conditions, the complex interaction between the biological and environmental can be revealed, and the overall condition of the ecological system can be more comprehensively understood, so as to more comprehensively, more accurately and more deeply evaluate the ecosystem service functions, and obtain the evaluation results of the ecosystem service functions.

[0068] Exemplarily, a specific example of generating an ecosystem service function evaluation result is given here:

[0069] Based on the comprehensive analysis of biological and environmental data by the evaluation large model, the quantitative indicators of the service function mode are generated:

[0070] (1) Carbon sink capacity (t / ha):

[0071] Input layer: Vegetation coverage data: Broadleaf forest proportion (45%), coniferous forest proportion (30%); Meteorological data: Monthly average temperature (5.2℃), historical baseline temperature (6.0℃);

[0072] Processing flow:

[0073] Vegetation carbon sink potential calculation:

[0074] Broadleaf forest carbon sink coefficient α = 0.8, coniferous forest β = 0.6;

[0075] Basic carbon sink value = (0.45×α + 0.3×β) = 0.54;

[0076] Temperature adaptability correction: Sigmoid response function is adopted (needs to meet the maximum carbon sink efficiency (should be 1.0) when the optimum temperature is met):

[0077]

[0078] Where, T: represents the current environmental temperature (5.2℃ in the embodiment). T0: represents the optimum temperature of the species (or vegetation) (set to 6.0℃ in the embodiment).

[0079] Ecological significance: When the environmental temperature is equal to T0, the biological function (such as carbon sink efficiency) reaches the best state.

[0080] k: represents the sensitivity coefficient (k = 0.5), controls the steepness of the curve, and reflects the sensitivity of the species to temperature deviation.

[0081] Temperature correction factor = ≈ 0.8.

[0082] Final carbon sink capacity:

[0083] Normalized value = Basic carbon sink value × Temperature correction factor = 0.54 × 0.8 ≈ 0.43.

[0084] (2) Water conservation capacity (%)

[0085] Input layer: Hydrological data: Monthly precipitation (120mm); Soil data: pH value (7.8), snow depth (15cm); Historical baseline: Annual average precipitation (100mm);

[0086] Processing flow:

[0087] Basic water conservation capacity calculation: Precipitation factor = min(current precipitation / historical baseline, 1.2) = 1.2;

[0088] Soil water retention correction:

[0089] pH deviation penalty: when pH When the range is [6.5, 7.5], the penalty coefficient γ = 0.9;

[0090] Snow cover compensation: δ = h * (snow thickness / reference thickness)

[0091] Snow depth: Measured value (15cm);

[0092] Baseline thickness: Take the historical average snow cover thickness of the region in winter (e.g., 22.5 cm, which needs to be adjusted according to actual data);

[0093] Compensation coefficient h: ranges from 0.1 to 0.3 (calibration required), representing the contribution intensity of unit snow thickness to water conservation. In this embodiment, the value is 0.3.

[0094] δ=0.3×(15cm / 22.5cm) = 0.2;

[0095] Soil correction factor = γ + δ = 1.1;

[0096] Comprehensive calculation:

[0097] Original value = Precipitation factor × Soil correction factor = 1.2 × 1.1 = 1.32;

[0098] Normalized value = tanh(original value) ≈ 0.72.

[0099] (3) Biodiversity index

[0100] Input layer: Species data: Siberian tiger index (0.58), red deer index (0.15); Vegetation data: broad-leaved forest (45%), shrubs (25%).

[0101] Processing flow:

[0102] Species diversity calculation:

[0103] Key species weights: Siberian tiger w1=0.6, red deer w2=0.4;

[0104] Animal diversity = w1 × Siberian tiger index + w2 × red deer index = 0.6 × 0.58 + 0.4 × 0.15 ≈ 0.41;

[0105] Vegetation diversity calculation:

[0106] Shannon index calculation: H' = -∑(p i ×ln(p i ))

[0107] where p i represents the relative proportion of the i-th species (or vegetation type) in the community (i.e., the ratio of the number of individuals or coverage area of the species to the total). For example: broad-leaved forest accounts for 45% → p1= 0.45; coniferous forest accounts for 30% → p2= 0.3; shrubs account for 25% → p3= 0.25.

[0108] ln(p i ): represents the natural logarithm (base e) of p i .

[0109] If p i = 0 (a species does not exist), it is agreed that p i ×ln(p i ) = 0 (to avoid mathematical undefined).

[0110] ∑ (accumulation): represents the sum of p i ×ln(p i ) values for all species (or vegetation types).

[0111] Vegetation composition ratio: broad-leaved forest 0.45, coniferous forest 0.3, shrubs 0.25.

[0112] H' ≈ 1.03, after normalization 0.68.

[0113] Comprehensive index:

[0114] Geometric mean: sqrt(animal diversity × vegetation diversity) = sqrt(0.41 × 0.68) ≈ 0.53;

[0115] After logistic regression calibration, output 0.61;

[0116] Get service function feature vector: S = [0.65, 0.72, 0.61].

[0117] In the implementation of the above obtaining the information of the animals and plants, the environmental quality assessment situation and the ecosystem service function evaluation result, the biological characteristic vector is analyzed based on the biological large model to identify the information of the animals and plants in the ecological protection zone, the environmental characteristic vector is analyzed based on the environmental large model to evaluate the environmental quality of the ecological protection zone, the animals and plants in the ecological protection zone are identified, and information such as the types, quantities and distribution of the animals and plants in the ecological protection zone can be understood; the environmental quality of the ecological protection zone is evaluated, and the environmental problems in the ecological protection zone can be found in time, thereby providing a scientific basis for ecological protection and management.

[0118] By integrating the multi-dimensional data and dynamically adjusting the weights, the comprehensive and dynamic evaluation of the ecosystem service function is realized. Compared with the static evaluation focusing on a single or a few service functions, the comprehensive and dynamic evaluation of the ecosystem service function is realized by multi-modal data fusion and dynamic analysis, and the change trend of the ecosystem can be more truly reflected.

[0119] The scheme of constructing the digital twin engine is introduced as follows. When the comprehensive feature matrix is generated based on the features extracted from the information of the animals and plants, the environmental quality assessment situation and the ecosystem service function evaluation result, and the digital twin engine is constructed by deep learning on the comprehensive feature matrix, the following steps are included:

[0120] After the multi-source data is sequentially preprocessed and aligned, the features are extracted from the multi-source data for fusion to generate a fusion feature vector corresponding to the same time and space dimension, and a comprehensive feature matrix is constructed based on a plurality of fusion feature vectors. The multi-source data includes the information of the animals and plants, the environmental quality assessment situation and the ecosystem service function evaluation result.

[0121] The comprehensive feature matrix is subjected to deep learning to generate a dynamic kernel including a weight matrix, a response function and a threshold rule. The weight matrix is used to quantify the action relationship between the species and the environment and the service function, the response function is used to describe the response relationship between the ecosystem variables, and the threshold rule is used to define the critical state of the ecosystem. The dynamic kernel is used to calculate the state of the ecosystem, evaluate the change trend of the state of the ecosystem and identify the critical state.

[0122] The dynamic kernel and the rule engine are packaged to construct the digital twin engine, and the rule engine makes a decision response based on the output result of the dynamic kernel.

[0123] After obtaining the plant and animal information, the environmental quality assessment situation and the ecosystem service function assessment result through the above processing, multi-source data is constructed based on the plant and animal information, the environmental quality assessment situation and the ecosystem service function assessment result. The multi-source data is sequentially subjected to data preprocessing, data alignment, feature extraction and feature fusion to generate a fusion feature vector corresponding to the same time and space dimension, and a comprehensive feature matrix is constructed based on a plurality of fusion feature vectors. That is, the comprehensive feature matrix is a structured data representation generated by preprocessing, spatio-temporal alignment, feature extraction and feature fusion of multi-source heterogeneous data, and is the core input for constructing a digital twin big model. The comprehensive feature matrix can provide a unified and standardized data basis for subsequent operations.

[0124] The process of generating the comprehensive feature matrix mainly involves four steps: 1. data preprocessing and data format unification; 2. spatio-temporal alignment and gridding; 3. feature extraction and vectorization; 4. feature fusion and matrix generation.

[0125] In the data preprocessing and data format unification step, the provided data (including plant and animal information, environmental quality assessment situation, and ecosystem service function assessment result) is subjected to preprocessing operations such as denoising, outlier removal, and missing value filling, and the data format is unified (such as gridding and time series alignment). The data is processed based on dimensionless normalization or standardization.

[0126] In the spatio-temporal alignment and gridding step, the data is aligned in time and space based on a unified spatio-temporal reference to determine the standard data. For example, a unified spatial grid (such as 10m x 10m) is divided in the spatial layer, and a unified time resolution (such as 1 hour, 10 hours, etc.) is determined in the time layer. The alignment method used is, for example, using spatial interpolation to map species observation point data to the grid, and using time interpolation to align data of different sampling frequencies to a unified timestamp.

[0127] In the feature extraction and vectorization step, multi-modal features (biological features, environmental features, and service function features) are extracted from the standard data to obtain fixed-dimension feature vectors, and all feature vectors are organized by grid-time units. Biological features include image features, audio features, and positioning features, environmental features include time series statistical features extracted based on meteorological data and physical environment data, and service function features are features extracted from service function data.

[0128] In the feature fusion and matrix generation step, the feature vectors of different modalities are spliced, and the importance of the features can also be considered, and different weights are assigned to the feature vectors of different modalities. After feature fusion, each grid-time unit corresponds to a fusion feature vector, and all fusion feature vectors constitute a comprehensive feature matrix, which serves as the input for subsequent modeling.

[0129] In the implementation process of generating the comprehensive feature matrix, consistency check can be used to ensure the spatial consistency of different data sources within the same grid; integrity check can be used to identify whether there are null values or abnormal values; and logical check can be used to verify the ecological rationality between species distribution and environmental conditions (e.g., wetland species should not appear in arid areas).

[0130] By aligning species, environment, and service data to a unified grid and time axis, scale differences are eliminated, ensuring the comparability of variables in the same dimension; and the information of animals and plants complements the biological level data, the environment data describes the background information of species, and the service function provides the ecological output indicators, which complement each other to improve the information quantity; the matrix data retains effective features, improving the data signal-to-noise ratio.

[0131] Exemplarily, an embodiment of generating a comprehensive feature matrix is given here:

[0132] 1. Assume the input data (single spatio-temporal unit)

[0133] Spatio-temporal unit: Grid A of a certain ecological protection area, time: 2023-10-01 10:00.

[0134] The input data includes the following contents:

[0135] Animal and plant information (biological features): Species 1 (Northeast Tiger): Quantity = 2, Confidence = 0.95, Activity Range = 5 km²; Species 2 (Elk): Quantity = 15, Confidence = 0.90.

[0136] Environmental quality assessment (environmental features): Temperature = 18°C, Humidity = 65%, Soil pH = 6.8, Snow Depth = 20 cm.

[0137] Ecosystem service function evaluation results (service function features): Carbon sink capacity = 2.4 t / ha, Water conservation capacity = 85%, Biodiversity index = 0.78.

[0138] 2. Generate a fusion feature vector for a single spatio-temporal unit

[0139] (1) Data preprocessing

[0140] Numerical normalization (scaled to [0, 1]):

[0141] Northeast Tiger Quantity: 2→ Normalized 0.2 (assuming the maximum number of tigers in this area = 10); Temperature: 18°C→ 0.6 (assuming the temperature range [-10, 30] °C); Carbon sink capacity: 2.4 t / ha→ 0.8 (assuming the maximum carbon sink capacity = 3 t / ha).

[0142] (2) Multi-modal feature extraction

[0143] Biological feature vector (dimension = 4): [0.2, 0.95, 0.15, 0.5] (Note: 0.15 = normalized value of the number of elks, 0.5 = normalized value of the activity range); Environmental feature vector (dimension = 4): [0.6, 0.65, 0.68, 0.67] (Note: 0.65 = normalized value of humidity, 0.68 = normalized value of soil pH, 0.67 = normalized value of snow depth); Feature vector of service function mode (dimension = 3): [0.8, 0.85, 0.78] (Note: 0.85 = normalized value of water conservation capacity).

[0144] It should be noted that in this embodiment, dynamic adjustment of the weights of each mode is still preferably considered.

[0145] The weights of the three types of vectors need to satisfy the normalization condition: biological weight + environmental weight + service weight = 1; the following gives an example of dynamic weight calculation: basic weight (default scenario): biological weight = 0.4, environmental weight = 0.4, service weight = 0.2.

[0146] Exemplary adjustment rules: if the species diversity index decreases by more than a threshold value (such as 10%), then: biological weight = 0.6, environmental weight = 0.3, service weight = 0.1.

[0147] If the water quality pH exceeds the safe range (6.5-7.5), then: environmental weight = 0.7, biological weight = 0.2, service weight = 0.1.

[0148] In a scenario where key species protection is prioritized, the input data is: the monthly comparison of the number of Siberian tigers decreases by 15% (triggering the protection rule); the environmental parameters are normal (pH = 6.8, temperature = 18°C); the carbon sink amount is stable, and the weight adjustment is: biological weight = 0.6, environmental weight = 0.3, service weight = 0.1.

[0149] Weighted feature vector:

[0150] Biological feature vector B = [0.2, 0.95, 0.15, 0.5] → 0.6 * B = [0.12, 0.57, 0.09, 0.3];

[0151] Environmental feature vector E = [0.6, 0.65, 0.68, 0.67] → 0.3 * E = [0.18, 0.195, 0.204, 0.201];

[0152] Service function vector S = [0.8, 0.85, 0.78] → 0.1*S = [0.08, 0.085, 0.078].

[0153] 3. Feature splicing

[0154] The vectors from the three modalities are concatenated into a fused feature vector (dimension=11):

[0155] The fused feature vector is: F = concat(0.6B, 0.3E, 0.1S) = [0.12, 0.57, 0.09, 0.3, 0.18, 0.195, 0.204, 0.201, 0.08, 0.085, 0.078].

[0156] 4. Construct a comprehensive feature matrix

[0157] Repeat the above process to generate fused feature vectors from multiple spatiotemporal units, and arrange them in spatiotemporal order:

[0158]

[0159] The final generated comprehensive feature matrix has the following form, for example:

[0160] [0.12, 0.57, 0.09, 0.30, 0.18, 0.195, 0.204, 0.201, 0.08, 0.085, 0.078], Grid A 10:00;

[0161] [0.18, 0.552, 0.108, 0.30, 0.186, 0.189, 0.210, 0.195, 0.079, 0.083, 0.077], Grid B 11:00;

[0162] [0.00, 0.00, 0.00, 0.00, 0.165, 0.210, 0.195, 0.180, 0.07, 0.075, 0.065], Grid C 10:00.

[0163] After obtaining the comprehensive feature matrix, a deep learning model (such as a neural network) is used to perform deep learning on the comprehensive feature matrix, outputting a weight matrix, response function, and threshold rules, which together constitute a dynamic kernel, providing logical support for subsequent calculations. Then, using the dynamic kernel as the core of computation, a rule engine is integrated and encapsulated as a digital twin engine. The digital twin engine has real-time data access, state calculation, rule judgment, and response triggering capabilities, and is the core driving module of the large digital twin model, supporting the dynamic operation and intelligent decision-making of the large model.

[0164] wherein the species, environment, and service function are in the form of variables in the deep learning model, for example, the species variables include species richness, diversity index, presence of key species, the environment variables include water quality pH, soil organic matter, temperature, precipitation, and the service function variables include carbon sink capacity, water conservation capacity, pollination service intensity, etc. The weight matrix includes the coupling weights between species-environment, species-service function, environment-service function, and species and environment-service function. The response function represents how one variable (such as an environmental factor) responds when another variable (such as species richness or service function) changes, and is used to describe the response relationship between variables in the ecosystem; it defines the influence mode and strength between variables, which can be linear, nonlinear, or conditional relationship. For example, the response function can represent the improvement of water quality on species richness or the enhancement of carbon sink capacity with the increase of vegetation coverage.

[0165] In the deep learning of the comprehensive feature matrix, the deep learning model is inputted with the comprehensive feature matrix, the coupling relationship between species, environment, and service function is learned through model training, and the weight matrix between species, environment, and service function is outputted; a nonlinear activation layer is embedded in the deep learning model to learn the response relationship between variables and obtain the response function; a loss function constraint is introduced in the model training process to automatically learn the key ecological threshold and obtain the threshold rule.

[0166] The deep learning model is used to learn the coupling relationship between species, environment, and service function from the comprehensive feature matrix; the nonlinear activation layer is used to introduce nonlinear transformation, so that the model can fit more complex function relationships; the nonlinear activation layer is embedded in the model to enhance the modeling ability of the model for the nonlinear response relationship between variables in the ecosystem, fit the complex function relationship between species, environment, and service function variables in the ecosystem, and thus determine the response function; by introducing the loss function constraint, the threshold information can be learned, and thus the threshold rule is obtained.

[0167] Learning the coupling relationship between species, environment, and service function is essentially learning the coupling relationship between variables, and learning the coupling relationship between species, environment, and service function is essentially modeling the interaction and response mechanism between variables through deep learning. The deep learning model automatically extracts the coupling weights, response function, and threshold rule between variables by learning the correlation, causality, or response mode between them.

[0168] The coupling weights of species and environment, the coupling weights of species and service function, the coupling weights of environment and service function, and the coupling weights of multi-element synergy are obtained by deep learning on the comprehensive feature matrix to form a weight matrix. The coupling weights of multi-element synergy can be understood as the joint influence or coupling weights of species and environment on ecosystem service function. It not only considers the influence of species on service function or the influence of environment on service function, but also comprehensively considers the interaction between species and environment and the synergistic effect of them on service function (such as carbon sink, water conservation, biodiversity maintenance, etc.). As an example, the species-environment-service function synergy weight indicates how the change of species responds to the service function under specific environmental conditions, or indicates how the environmental change affects the service function under specific species composition. For example, under the joint action of increased precipitation (environmental factor) and improved vegetation coverage (species factor), the water conservation capacity is significantly improved, and the coupling weights of multi-element synergy represent the positive influence strength of the joint action of the two on water conservation. For another example, under the joint action of soil acidification (environmental factor) and reduction of key species (species factor), the soil carbon sink capacity is significantly reduced, and the coupling weights of multi-element synergy represent the negative influence strength of the joint action of the two on carbon sink.

[0169] The dynamic kernel composed of the weight matrix, the response function and the threshold rule is responsible for real-time calculation of the interaction and state change between each element (species, environment, service function) of the ecosystem, analysis of the state change trend of the ecosystem, identification of critical situations, and output of the current system state, the analyzed state change trend, and the identification of whether the critical condition is reached. The rule engine performs logical judgment and response actions based on the output of the dynamic kernel to realize intelligent feedback and decision support.

[0170] Specifically, for the dynamic kernel, it is mainly used for calculating the real-time state of the current ecosystem, analyzing the state change trend and identifying the critical state. For example, in calculating the real-time state of the ecosystem, the state of species (species richness, diversity, distribution, etc.), the state of environment (water quality, soil, weather, etc. environmental parameters), and the state of service function (carbon sink capacity, water conservation capacity, biodiversity maintenance capacity, etc.) are evaluated. In analyzing the state change trend, the latest data, historical data of the ecosystem and the response function are used to calculate the change direction and amplitude of the current state relative to the previous time; for example, the upward or downward trend of species richness, the improvement or deterioration trend of water quality, and the strengthening or weakening trend of service capacity are evaluated. In identifying the critical state, it is judged whether the ecological threshold is reached (such as whether the pH is less than 6.5 or whether the species richness is decreased by more than 20%). The dynamic kernel can provide input content for the rule engine, such as transmitting the state value, state change trend and critical judgment result of the current ecosystem to the rule engine, and the rule engine processes accordingly based on the input.

[0171] It should be noted that the dynamic kernel mainly deals with real-time data, but the ecosystem historical data as background information enhances the accuracy and reliability of the analysis, which is used to assist in a more comprehensive and accurate understanding of the changes in the ecosystem.

[0172] For the rule engine, it is mainly responsible for executing the following tasks based on the current state, state change trend, and critical identification results output by the dynamic kernel: 1. Logical judgment and decision triggering; 2. Response action execution. When logical judgment and decision triggering, it judges whether the preset business rules or response conditions are met, such as triggering acidification warning when pH is less than 6.5, triggering ecological degradation alarm when species richness decreases by more than 20%, and triggering service adjustment suggestion when carbon sink valuation decreases by more than 10%. When executing response actions, based on the judgment results, it automatically executes matching actions, such as issuing warning information and recommending intervention measures (such as planting water purification plants, limiting pollution, etc.). The rule engine can also support feeding back the judgment results and response actions to the visualization interface for relevant personnel to view risks and intervention suggestions, providing auxiliary information for ecological protection and resource regulation decision-making. It should be noted that the rule engine not only considers the current state, but also combines the state change trend for logical judgment, thereby realizing more intelligent and forward-looking ecological response and decision support.

[0173] That is, the dynamic kernel calculates the current state of the ecosystem, analyzes the state change trend and critical identification, and provides the basis for real-time monitoring and response; the rule engine, based on the current state, state change trend, and critical identification results output by the dynamic kernel, combines the preset business rules, executes logical judgment and triggers corresponding response actions, realizes intelligent feedback and decision support of the ecosystem. The rule engine, as the key module to realize intelligent feedback and decision support, forms a "judgment + response" closed-loop mechanism with the dynamic kernel, and builds a digital twin engine as the core driving module of the digital twin large model.

[0174] In the above implementation scheme of building a digital twin engine, by deep learning on the comprehensive feature matrix, a dynamic kernel for calculating the state of the ecosystem, evaluating the state change trend, and identifying the critical state is generated, and the dynamic kernel and the rule engine are integrated to build a digital twin engine, which can provide guarantee for the generation of the subsequent digital twin large model.

[0175] The implementation process of generating a prediction model, building a digital twin large model based on a digital twin engine and a prediction model is introduced as follows. When generating a prediction model, the following steps are included:

[0176] Determine a data sample set based on the labeled scenario category of the ecosystem historical sample and the label data indicating the real observation value of the target moment, the data sample set including a training set, a validation set and a test set, the label data belonging to the model training target, and the target moment being a future moment relative to a historical moment;

[0177] Based on the training set, perform model training on a machine learning model applicable to the time series prediction of the ecosystem, update the model parameters through an optimization algorithm, and evaluate the model performance based on the validation set;

[0178] Evaluate the prediction accuracy and scenario response consistency of the model based on the test set, and optimize the model structure and hyperparameters based on the evaluation index to determine the prediction model.

[0179] The prediction model trained in this embodiment has the functions of predicting the future state of the ecosystem and deducing the state of the ecosystem, which do not conflict with each other and belong to two core functions of the prediction model, and cooperatively constitute the intelligent analysis and decision-making capability of the digital twin large model. When predicting the future state of the ecosystem, the natural evolution state of the ecosystem at a specific future time is predicted based on historical and current data, providing a reference for future trends; when deducing, the deduced results are given based on the provided data combined with the set scenario.

[0180] In the prediction model training phase, the scenario category, real observation value and ecosystem historical sample need to be involved in the process. The training target of the prediction model is to predict the future state and have the response ability to different scenario inputs. The ecosystem historical sample is used to describe the state of the ecosystem (record the historical conditions of species, environment and service function), record the evolution process (record the changes of the ecosystem over time), reflect the relationship between variables (reflect the interaction between species, environment and service function), and identify abnormal events (capture sudden impacts such as pollution, disasters and invasions). As the basic data for model training and verification, it helps the model to learn the operation rules of the ecosystem and the relationship between variables; by labeling the scenario category of the ecosystem historical sample, the model can learn what changes occur under what scenario when learning the historical rules, thereby having the response ability to different scenarios. The label data indicating the real observation value is a necessary condition for training the prediction model, which is the target output of the model training, so that the model learns the mapping relationship between the input and the output, and then realizes the accurate prediction of the future state. In the deduction mode, label data is not needed because the goal of deduction is to simulate the response under a specific scenario, rather than to learn the mapping relationship between the input and the output.

[0181] The process of model training includes the following stages: data preparation (collecting historical samples, labeled data, scenario categories, dividing data sets), feature engineering (extracting time series features, constructing spatio-temporal features, scenario encoding, standardization processing), model construction (multi-input structure, model selection, introduction of enhancement mechanism), model training (supervised learning, multi-task learning, performance evaluation), model evaluation and optimization (performance evaluation, optimization method).

[0182] In the data preparation stage, historical samples of the ecosystem (such as species status, environmental status, service function status) are collected as input features; real observations at a certain time in the future are collected as label data; each sample is labeled with a scenario category (such as climate scenario, intervention measure type), and the source of the scenario category is specified; the training set, validation set, and test set are divided to ensure uniform distribution of scenario categories in each set.

[0183] In the feature engineering stage, time series features (sliding window mean, trend, seasonality) are extracted, and the sliding window size and seasonality period are specified; spatio-temporal features (neighborhood mean, spatial lag) are constructed, and the definition of spatial neighborhood is specified; scenario categories are encoded into numerical values or embedding vectors; features are standardized to improve training efficiency.

[0184] In the model construction stage, a multi-input structure is adopted, such as selecting LSTM, Transformer, GNN, etc., introducing conditional generation mechanism or attention mechanism to enhance scenario response capability, and specifying the specific form of conditional input.

[0185] In the model training stage, supervised learning is performed using the training set, and the loss function uses mean square error (MSE) or mean absolute error (MAE) to measure the prediction error; multi-task learning is used to predict the state under natural evolution and scenario intervention simultaneously; performance is evaluated on the validation set to prevent overfitting.

[0186] In the model evaluation and optimization stage, the prediction accuracy and scenario response consistency are evaluated on the test set to determine the evaluation target, adjust the model structure and hyperparameters, and improve the robustness and generalization ability. Prediction accuracy is used to measure the closeness between model predicted values and real observed values, and scenario response consistency is used to measure the rationality and stability of model output results under different scenarios.

[0187] After generating the prediction model through model training, the digital twin large model is constructed based on the digital twin engine, the large model static base information, and the prediction model, including:

[0188] Load the large model static base information used to construct the three-dimensional spatial skeleton and basic attributes to the digital twin engine, complete spatial registration and logical binding;

[0189] integrating the trained prediction model set into the digital twin engine;

[0190] configuring various model parameters in the digital twin engine for initialization of the digital twin large model.

[0191] The large model static base information is used to construct the three-dimensional spatial skeleton and basic attributes of the digital twin large model, which provides the terrain structure and ecological attributes to construct the three-dimensional spatial skeleton and basic attributes of the digital twin large model, and is the spatial and attribute basis for the operation of the large model.

[0192] The three-dimensional spatial skeleton, for example, includes terrain DEM (Digital Elevation Model), spatial gridding, and spatial positioning. The terrain DEM provides the elevation, slope, slope direction, and other information of each spatial point, thereby providing continuous elevation and terrain attributes; spatial gridding is to divide the ecological region into regular grids (such as 10m x 10m), each grid corresponds to a spatial unit, and then the DEM is discretized into numbered grid units; spatial positioning is to position and superimpose all dynamic data (species distribution, environmental parameters) based on the skeleton, that is, to accurately map the real-time data of species, environment, etc. to the corresponding grid unit with grid ID.

[0193] The basic attributes, for example, include a vegetation type grid and a soil texture map, the vegetation type grid provides vegetation gridded attribute values, and the soil texture map provides soil gridded attribute values; by unifying the grid coordinates, the vegetation, soil, and dynamic data are aligned under the same spatial reference, avoiding misalignment. For the vegetation type grid, by marking specific vegetation types (forest, grassland, wetland, etc.) in each spatial grid, the coverage characteristics are described, and these vegetation types directly affect the habitat suitability, carbon sink capacity, and other ecological processes of species, thereby providing key ecological attribute inputs for the digital twin large model. The soil texture map provides soil ecological parameters by giving the soil type, organic matter content, pH value, and other attributes of each grid, and these parameters have important influences on plant growth rate, nutrient supply, pollutant adsorption, and water purification efficiency, and are the basic inputs for the digital twin large model to simulate ecological processes.

[0194] The process of initializing the digital twin large model based on the digital twin engine, the large model static base information, and the prediction model mainly includes the following steps: 1. loading the large model static base information into the digital twin engine to provide a three-dimensional spatial skeleton and basic attributes for the large model, ensuring the accuracy and authenticity of the large model in space; 2. integrating the large model static base information with the dynamic kernel and rule engine in the digital twin engine through spatial registration and data association, establishing a correspondence between space and attributes, enabling the digital twin engine to perform state calculation and response judgment based on the large model static base information; 3. integrating the prediction model into the digital twin engine as a functional module for future state prediction and scenario deduction; 4. configuring parameters in the digital twin engine such as coupling weight, response function, and threshold rule based on the characteristics of the actual ecosystem (based on historical long-term statistics such as multi-year average species richness, seasonal water quality fluctuation range, soil background value, etc.), and initializing the large model state to ensure that the large model can accurately reflect the initial conditions of the ecosystem.

[0195] The coupling weight is used to quantify the interaction strength between species and environment, service function, and is a key parameter in the large model that describes the relationship between variables; the response function is used to describe the response relationship between variables, and is a function that describes the change rule between variables in the large model; the threshold rule is used to define the critical state of the ecosystem, and is used to judge whether the warning or response condition is reached, and is the basis for logical judgment and response triggering in the large model. These parameters together constitute the dynamic kernel of the digital twin large model, and are key model parameters in the digital twin large model that describe the relationship between variables, change rules, and critical states of the ecosystem, and are the core components of the digital twin large model that accurately reflect the state and behavior of the ecosystem. At the same time, they are also the basis for the large model to realize real-time calculation, state update, trend analysis, critical judgment, and response triggering.

[0196] In the above implementation process, by pre-training a prediction model with the function of predicting future state and deduction function, the prediction model is integrated into the digital twin engine, which can build a digital twin large model that supports future state prediction and state deduction.

[0197] The process of displaying the panoramic dynamic information of the ecosystem in the visualization interface is introduced as follows. In response to the input real-time data of the digital twin large model, the real-time data and target data are processed, and the panoramic dynamic information of the ecosystem is displayed in the visualization interface, including:

[0198] Based on the data interface, the obtained real-time data is connected to the digital twin engine to drive the digital twin engine to update the model state of the digital twin large model. The real-time data includes real-time plant and animal information, real-time environmental quality assessment, and real-time ecosystem service function assessment results.

[0199] The real-time data and the target data are processed based on the updated digital twin large model, and an ecosystem operation result and an ecosystem prediction result are output, the ecosystem operation result including at least one of an ecosystem current state, an ecosystem state change trend, an ecosystem critical state, and a decision response, and the ecosystem prediction result including at least one of a natural evolution prediction result and a scenario deduction result;

[0200] Ecosystem panoramic dynamic information is generated based on the ecosystem operation result and the ecosystem prediction result provided by the digital twin large model, and is visually displayed through a VR interface or an AR interface.

[0201] The digital twin engine obtains real-time data (real-time plant and animal information, real-time environmental quality evaluation, and real-time ecosystem service function evaluation results) through a data interface (such as an Internet of Things interface), and dynamically updates the state of the digital twin large model based on the obtained real-time data, so as to keep the large model synchronized with the real ecosystem.

[0202] After updating the digital twin large model, the real-time data and the target data are processed based on the updated digital twin large model, and an ecosystem operation result and an ecosystem prediction result are output. The provided ecosystem operation result includes at least one of an ecosystem current state, an ecosystem state change trend, an ecosystem critical state, and a decision response, and the provided ecosystem prediction result includes at least one of a natural evolution prediction result and a scenario deduction result. After the digital twin large model outputs the ecosystem operation result and the ecosystem prediction result, ecosystem panoramic dynamic information is generated based on the ecosystem operation result and the ecosystem prediction result, and is visually rendered through VR technology or AR technology, and is visually displayed on a VR interface or an AR interface.

[0203] In the processing of the real-time data and the target data based on the updated digital twin large model, and the output of the ecosystem operation result and the ecosystem prediction result, the following steps are included:

[0204] The real-time data and the ecosystem historical data are processed based on the updated digital twin large model, the ecosystem current state is calculated, the ecosystem state change trend is evaluated, the ecosystem critical state is identified, the ecosystem evaluation information is obtained, and the decision response is fed back based on logical judgment.

[0205] The scenario deduction is performed based on the ecosystem evaluation information, the set scenario parameters, and the target data, and / or the natural evolution prediction is performed based on the ecosystem evaluation information and the target data, and the ecosystem prediction result is output, and the target data further includes external driving factors, time information, and space information.

[0206] During the processing, the digital twin engine performs state calculation based on real-time data, ecosystem historical data, large model static base information, using a dynamic kernel, evaluates the current state of the ecosystem, the trend of the state of the ecosystem, and identifies whether a critical state is reached; the rule engine performs logical judgment and response actions according to the output of the dynamic kernel, such as triggering an early warning, to provide decision support for ecological protection and management. The current state of the ecosystem, the trend of the state of the ecosystem, and the critical state of the ecosystem output by the dynamic kernel are provided as ecosystem evaluation information to the prediction model.

[0207] When the prediction model performs natural evolution prediction, the data provided by the dynamic kernel is one of the core inputs, and the ecosystem historical data, time information, spatial information, and external driving factors are also needed to predict the future natural evolution path of the ecosystem.

[0208] The data provided by the dynamic kernel includes the current state of the ecosystem (such as species richness, environmental quality, service function), the trend of the state of the ecosystem (such as upward, downward, fluctuation trend), and the identified critical state of the ecosystem. Ecosystem historical data is used for context modeling, including, for example, species identification history, environmental monitoring history, service function history evaluation results, which can capture long-term evolution rules and seasonal changes. Time information includes the current time point, the prediction start time, and the prediction length (such as 7 days, 30 days in the future), which is used to determine the prediction window and the length of the model input sequence. Spatial information includes, for example, grid ID, regional attributes (such as terrain, vegetation, soil), which is used for spatial coupling modeling. External driving factors include, for example, weather forecast data (such as future precipitation, temperature), human activity intensity (such as development, pollution).

[0209] The data provided by the dynamic kernel, the ecosystem historical data, the time information, the spatial information, and the external driving factors are input into the prediction model as input data, and the prediction model performs natural evolution prediction to output the ecosystem prediction result (natural evolution prediction result).

[0210] When the prediction model performs scenario deduction, the data provided by the dynamic kernel is one of the core inputs, and the scenario parameters, ecosystem historical data, time information, spatial information, and external driving factors are also needed. Scenario parameters are set by the user or preset, such as water replenishment, temperature change, vegetation restoration area, etc., which are used to define the deduction scenario. The above data is input into the prediction model, and the prediction model performs scenario deduction to output the ecosystem prediction result (scenario deduction result).

[0211] Figure 2 A simplified flowchart of the digital twin large model provided by the embodiment of the present application for processing real-time data and target data is shown.

[0212] The dynamic kernel receives real-time data accessed through the Internet of Things (real-time plant and animal information, real-time environmental quality assessment, and real-time ecosystem service function assessment results), combines historical ecosystem data, static base information of large models, coupling weights, response functions, and threshold rules, calculates the current state of the ecosystem, analyzes the trend of the state of the ecosystem, identifies whether the state of the ecosystem reaches a critical condition, and outputs content including the current state value of the ecosystem, the state change trend, and the critical identification situation.

[0213] The rule engine executes business logic judgment and response actions based on the output of the dynamic kernel. The rule engine receives the current state value of the ecosystem, the state change trend, and the critical identification situation output by the dynamic kernel, judges whether the warning condition is met based on the set response strategy, and executes the response and outputs the decision suggestion.

[0214] The prediction model receives the output content provided by the dynamic kernel, combines historical ecosystem data, time information, spatial information, and external driving factors to perform natural evolution prediction, and outputs natural evolution prediction results; and / or, the prediction model receives the output content provided by the dynamic kernel, combines scenario parameters, historical ecosystem data, time information, spatial information, and external driving factors to perform scenario deduction, and outputs scenario deduction results.

[0215] The decision response provided by the rule engine, the current state value of the ecosystem, the state change trend, and the critical identification situation output by the dynamic kernel, and the natural evolution prediction results and / or scenario deduction results provided by the prediction model are rendered in a visual interface to drive the digital twin engine to refresh the large model in real time, ensuring that the large model is synchronized with the real ecosystem. The dynamic kernel, the rule engine, and the prediction model cooperate with each other to output the panoramic dynamic information of the ecosystem based on real-time data, support real-time monitoring, intelligent response, and future deduction.

[0216] The ecosystem operation results provided by the digital twin large model cover the current state of the ecosystem, the state change trend, the critical state identification, and the response actions. The ecosystem prediction results provided by the digital twin large model include natural evolution prediction results and / or scenario deduction results, which can comprehensively reflect the operation status, response mechanism, and prediction of the ecosystem.

[0217] In a specific example of the present application, the panoramic dynamic information of the ecosystem displayed through the VR interface or the AR interface includes the current state of the ecosystem, the state change trend of the ecosystem, the critical state of the ecosystem, the decision response result, the natural evolution prediction result, and the scenario deduction result.

[0218] The current state of the ecosystem includes the state of species, the state of environmental quality, and the state of ecosystem service functions. The state of species, for example, includes the real-time identification of species names, richness, diversity index, and spatial distribution heat map within each grid. The state of environmental quality, for example, includes real-time grid values of water quality, soil, and weather. The state of ecosystem service functions includes, for example, carbon sink flux, water conservation capacity, biodiversity maintenance index, and pollutant purification efficiency in grid form. The real-time environmental quality assessment is a comprehensive evaluation based on raw data. The digital twin model converts these evaluation results into continuous physical quantities (such as pH, dissolved oxygen, and organic matter) for each grid, resulting in an environmental state. The real-time ecosystem service function assessment is a comprehensive scoring or grading evaluation of raw data. The digital twin model converts the evaluation results into continuous physical quantities or fluxes for each grid, which are service functions.

[0219] When the current state of the ecosystem is displayed on the visualization interface, species distribution maps, environmental quality maps, and service function state maps can be displayed. The specific forms of species distribution maps, for example, are heat maps (using color depth to represent species density or richness), point maps (using discrete points to mark the location of each observed individual or sampling point), and three-dimensional models (superimposing species data on a three-dimensional terrain or vegetation model to provide an immersive spatial perspective). These three forms can be presented individually or in combination to form a complete visualization of species spatial distribution. Environmental quality maps, for example, plot water quality (pH, dissolved oxygen), soil (organic matter, heavy metals), and weather (temperature, humidity) parameters by grid or region to create spatial distribution maps, such as a pH distribution map or a temperature heat map, to quickly identify where the water quality is acidic or where the temperature is high. Service function state maps, for example, plot the output capacity of the ecosystem (carbon sink capacity, water conservation capacity, biodiversity index, etc.) to visually display high-value carbon sink areas, water conservation hotspots, and biodiversity cold and hot spots.

[0220] Change trends include the change trends (such as rising, falling, fluctuating) of each element (species, environment, service), the change rate and amplitude (such as the daily change amount of water quality index). The state change trend can be represented by trend curve graphs, dynamic evolution graphs, and change rate graphs. For example, trend curve graphs indicate the change curves of species richness, water quality index, and service indicators over time. Dynamic evolution graphs visually present the change process of the state of the ecosystem over time in the form of animation or time axis playback. Change rate graphs display the rising and falling speed of each ecological element in unit time in a spatial distribution form.

[0221] Critical state identification includes identifying whether the ecological threshold has been reached. It can be displayed through a warning map to show which areas trigger the ecological threshold warning, through a warning list to list all current warning information (such as pH being too low and species richness decreasing), and through a critical identification to highlight the critical areas.

[0222] For response actions and decision recommendations, the large model can automatically trigger response actions, recommend intervention measures (such as planting water purification plants, limiting pollution discharge, and ecological restoration schemes), and can use scenario simulation to simulate the ecological restoration effect under different intervention measures.

[0223] When displaying the natural evolution prediction results and scenario deduction results in the visualization interface, a variety of methods can be used to intuitively and clearly present the natural evolution future prediction and scenario deduction of the ecosystem. A dynamic evolution graph can be used to display the change of the ecosystem state over time, such as a dynamic evolution graph that displays the evolution process in the form of an animation or a time axis playback, and the display content includes species distribution changes (displaying the distribution changes of species at different time points through animation), environmental quality changes (displaying the change trend of water quality, soil, meteorological parameters, etc. over time), service function changes (displaying the change of carbon sink, water conservation, biodiversity, etc. over time). A time axis interactive method can be used to display the state of the ecosystem at different time points, such as a user dragging the time axis to view the state of the ecosystem at different time points, and the display content includes real-time data (real-time monitoring data at the current time point), historical data (recorded data at past time points), and prediction data (data at future time points). A scenario comparison graph can be used to display the change of the ecosystem state under different scenarios, such as displaying the future state of the ecosystem without intervention, and the future state of the ecosystem under set intervention measures (such as water replenishment and vegetation restoration), and the difference under different scenarios can be displayed through a chart or a map.

[0224] In an optional embodiment of the present application, the method further comprises:

[0225] In response to receiving the interactive operation on the VR interface or the AR interface, updating the model state of the digital twin large model, and obtaining updated ecosystem panoramic dynamic information output by the digital twin large model;

[0226] Based on the updated ecosystem panoramic dynamic information, updating the visualization content on the VR interface or the AR interface in real time.

[0227] The ecosystem panoramic dynamic information (ecosystem operation results and ecosystem prediction results) output by the digital twin large model is visualized through VR technology or AR technology, which enables users to intuitively understand the current state, state change trend, critical situation, response decision, and ecosystem prediction of the ecosystem. The visualization interface supports interactive operation, and based on the user's interactive operation on the visualization interface, the model state of the digital twin large model can be updated, the latest ecosystem panoramic dynamic information output by the digital twin large model can be obtained, and the visualization interface can be updated.

[0228] The visualization interface can support user interaction, for example, in the following ways:

[0229] Parameter adjustment: allow users to adjust parameters such as environmental factors, species quantity, etc., and the large model gives the response. Scenario simulation: provide different scenario options, users can view the changes of the simulated ecosystem under different scenario options. Layer control: users can choose to display or hide different data layers such as species distribution, water quality status. Warning information viewing: show warning details, users can click to view specific risk areas and suggested measures. Timeline operation: users can drag the timeline to view the historical changes and trends of the ecosystem state. The above-mentioned supported interaction modes can enable users to actively participate in model operation, explore different scenarios, and view the ecosystem state at different time points (such as historical state, future natural evolution state), thereby better understanding the ecosystem state and assisting decision-making.

[0230] The visualization interface can be a VR interface or an AR interface. The VR interface supports immersive operation, such as users can wear VR headsets, enter a three-dimensional virtual ecosystem, and interact with the virtual environment through handles, gestures, eye tracking, etc. to view the ecological state of different areas. For example: in the parameter adjustment scenario, environmental parameters (such as precipitation, temperature) can be adjusted in the virtual interface; in the information query and feedback scenario, users can click on a species or area to view detailed information (such as species name, abundance, health status); in the deduction scenario, users can set different scenario parameters to view the deduction results under each scenario parameter, and support multi-scenario comparison, users can simultaneously view the ecosystem state under different scenarios and intuitively compare the differences; in the natural evolution prediction scenario, users can drag the timeline to view the natural evolution prediction results at different time points.

[0231] The AR interface supports reality augmentation superimposition, which superimposes virtual ecological information onto the real scene through AR glasses or cameras, i.e., when users wear AR glasses or use a phone camera, they can see real-time monitoring data superimposed in the real scene in front of them, such as displaying water quality indicators above a real river or superimposing a species distribution map in a forest. Users can view real-time monitoring data (such as water quality, air quality) in the real environment, and support clicking to view details, historical trends, warning information, etc. In the scenario deduction scenario, the effects of different intervention measures (such as vegetation restoration, pollution control) can be simulated, and the differences between the current situation and the simulated ecological state, and the differences between different scenarios can be compared. In the natural evolution prediction scenario, users can view the natural evolution prediction results at different time points.

[0232] The embodiment of the application supports user interaction on the visual interface while displaying the ecosystem panoramic dynamic information on the visual interface, updates the ecosystem panoramic dynamic information in response to user interaction, realizes intuitive and real-time interaction between the user and the digital twin large model, and improves the user's interaction experience based on the VR interface or the AR interface for immersive interaction.

[0233] Figure 3 The overall implementation flowchart of the digital evolution method of the ecosystem of the embodiment of the application is shown.

[0234] Step 301, based on the biological data and environmental data collected in the ecological protection area, biological feature vectors and environmental feature vectors are generated, the biological feature vectors and the environmental feature vectors are analyzed, and the plant and animal information and the environmental quality evaluation situation are obtained.

[0235] Step 302, the biological feature vectors and the environmental feature vectors are spliced according to the corresponding weights to determine the comprehensive feature vectors.

[0236] Step 303, based on the optimized evaluation large model, the biological feature vectors, the environmental feature vectors, the comprehensive feature vectors, the plant and animal information and the environmental quality evaluation situation are comprehensively analyzed to generate the ecosystem service function evaluation result.

[0237] Step 304, data preprocessing, time and space alignment of data are performed on the multi-source data to determine the standard data, and multi-modal feature extraction is performed in the standard data to construct a comprehensive feature matrix, the multi-source data including the plant and animal information, the environmental quality evaluation situation and the ecosystem service function evaluation result, wherein the comprehensive feature matrix is composed of fusion feature vectors of multiple spatio-temporal units arranged in spatio-temporal sequence, and the fusion feature vector of a single spatio-temporal unit is formed by multi-modal feature splicing of the plant and animal information, the environmental quality evaluation situation and the ecosystem service function evaluation result corresponding to the current spatio-temporal unit.

[0238] Step 305, a dynamic kernel is generated by deep learning on the comprehensive feature matrix, and a digital twin engine is constructed by encapsulating the dynamic kernel and a rule engine.

[0239] Step 306, the large model static base information is loaded into the digital twin engine, the prediction model supporting natural evolution prediction and scenario deduction is integrated into the digital twin engine, various model parameters in the digital twin engine are configured, and the digital twin large model is initialized.

[0240] Step 307, in response to inputting real-time data of the digital twin large model, the model state of the digital twin large model is updated, the real-time data and the target data are processed based on the updated digital twin large model, the ecosystem panoramic dynamic information is output, and the visual interface is used for display.

[0241] Step 308, in response to receiving the interactive operation on the visualization interface, obtaining the latest ecosystem panoramic dynamic information, updating the visualization content on the visualization interface.

[0242] The above implementation process of the present application realizes real-time twinning of the state of the ecosystem based on digital twinning technology, realizes natural evolution prediction results and scenario deduction of the ecosystem based on integrated prediction functions, uses a visualization interface for content display and supports real-time interaction, and provides a good experience for users.

[0243] In order to further introduce the scheme provided by the embodiments of the present application, a specific example is used for illustration below. The example focuses on the monitoring and management of tropical rainforest ecosystems, aiming to realize real-time monitoring, future state prediction and scenario deduction of tropical rainforest ecosystems through multi-modal data fusion, dynamic kernel construction, natural evolution prediction, scenario deduction and AR interaction, and to provide scientific decision support for park managers.

[0244] 1. Multi-modal data acquisition and processing (acquiring and processing biological data and environmental data)

[0245] When acquiring biological data and processing the data: deploy high-resolution cameras to capture tree canopy animal activities, extract feature vectors (such as species ID, number, activity frequency) through CNN models; use directional microphone arrays to collect bird calls, generate feature vectors (species ID, call frequency, voiceprint similarity) through voiceprint recognition models; wear GPS tags on iconic species (such as monkey groups) to generate spatial trajectory feature vectors (moving speed, habitat range, migration path). Align the image, audio, and positioning data in space and time to construct biological feature vectors consistent in space and time, and dynamically assign weights, such as automatically adjusting feature weights according to data confidence.

[0246] When acquiring environmental data and processing the data: distribute soil moisture Internet of Things sensors to generate grid-based feature vectors (humidity value, spatial coordinates, timestamp); collect data based on forest canopy light meters to construct feature vectors (light intensity, daily variation curve, forest canopy transmittance); integrate micro weather stations to output vectors (temperature, rainfall, wind speed). Generate environmental feature vectors based on a unified space-time reference.

[0247] 2. Multi-source data acquisition

[0248] After obtaining the biological feature vector and the environmental feature vector, the biological feature vector is analyzed to obtain the information of animals and plants in the ecological protection zone, the environmental feature vector is analyzed to obtain the environmental quality assessment situation in the ecological protection zone, and the biological feature vector and the environmental feature vector are spliced according to weights to generate a comprehensive feature vector. The biological feature vector, the environmental feature vector, the comprehensive feature vector, the information of animals and plants, and the environmental quality assessment situation are comprehensively analyzed to obtain an ecological system service function evaluation result. Thus, the information of animals and plants, the environmental quality assessment situation, and the ecological system service function evaluation result are obtained.

[0249] 3. Dynamic kernel construction

[0250] After obtaining the information of animals and plants, the environmental quality assessment situation, and the ecological system service function evaluation result, features are extracted from the above data for fusion to generate a fusion feature vector corresponding to the same time and spatial dimension, a comprehensive feature matrix is constructed based on multiple fusion feature vectors, and a dynamic kernel is constructed by deep learning of the comprehensive feature matrix. For example, a Transformer+LSTM hybrid model is used to process the comprehensive feature matrix to generate a dynamic kernel including a weight matrix, a response function, and a threshold rule. The weight matrix quantifies the coupling relationship between light intensity and parrot breeding rate, and the ecological rationality is verified through explainability analysis; the response function fits the nonlinear relationship, such as determining the relationship between breeding rate and light, humidity, and temperature; the threshold rule is used for critical state detection, such as triggering an "extreme drought" warning when humidity < 25% and consecutive high temperature > 7 days. Moreover, the dynamic kernel supports fine-tuning through online learning.

[0251] 4. Construction and application of digital twin large model

[0252] A digital twin engine is constructed based on the dynamic kernel and the rule engine, and a digital twin large model is constructed based on the digital twin engine, the static base information of the large model, and the prediction model. The digital twin large model is used to process real-time data and ecological system historical data, and outputs ecological system panoramic dynamic information for visual display. The ecological system panoramic dynamic information includes at least one of ecological system running results and ecological system prediction results, the ecological system running results include at least one of the current state of the ecological system, the trend of the state change of the ecological system, the critical state of the ecological system, and the decision response result, and the ecological system prediction results include at least one of natural evolution prediction results and scenario deduction results.

[0253] 5. AR interaction implementation

[0254] Based on the operation of the user on the AR interface, the display content of the visual interface is updated.

[0255] The following is an example of predicting the natural evolution of a tropical rainforest ecosystem and climate scenario. Compared with existing technologies that can only display the current state and cannot predict long-term changes, this embodiment breaks through the limitations of existing technologies by predicting natural evolution and climate scenario, providing future scenario simulation that can be intervened, and providing scientific decision support for nature reserve managers.

[0256] I. Natural evolution prediction technology implementation

[0257] (1) Data preprocessing and feature engineering

[0258] The input data includes 10 years of historical time series data, biological data includes parrot population (monthly census), tree species diversity index (annual survey); environmental data includes temperature / rainfall (daily average of weather stations), soil carbon content (quarterly sampling); ecosystem service function data includes carbon sink capacity.

[0259] Preprocess the data: Align the data by 1km×1km grid for spatial standardization, unify the time resolution to monthly, construct a 15-dimensional feature tensor (format [120 months×100 grids×15 dimensions]), and embed the coupling features output by the dynamic kernel, such as the coupling weight of temperature-parrot breeding rate, as prior knowledge.

[0260] (2) Prediction model architecture

[0261] Dual-module collaborative design (LSTM time series layer and Transformer spatial layer)

[0262] LSTM time series layer: input is 15-dimensional multi-modal features (biological + environmental + service function), double-layer structure to capture seasonal periodicity and interannual trend, output is 64-dimensional time series features (preserve historical dependency relationship). Transformer spatial layer: 4 heads of attention mechanism to identify cross-grid ecological associations (such as canopy-soil layer interaction), output fused spatio-temporal features (including global spatial relationship).

[0263] Ecological constraint mechanism

[0264] Weight matrix (64×64), initial value inherits the weight provided by the dynamic kernel to force the features to conform to ecological rules (such as temperature rising leading to carbon sink capacity decreasing); feature space transformation, suppresses feature combinations that violate ecological laws (such as herbivores exceeding vegetation carrying capacity), outputs feature tensor constrained by ecology.

[0265] (3) Prediction process

[0266] Temporal processing: LSTM encodes historical impacts (e.g., lag effects of rainfall on carbon sinks), spatial correlation: attention mechanism calculates grid interactions (e.g., drought area cascading effects), ecological correction: weight matrix projects features to ecological rule space.

[0267] (4) Result verification

[0268] Uncertainty quantification: generate 95% confidence intervals; precision comparison: rainforest measured 5-year prediction error 3.2%, traditional model error >15%.

[0269] The above natural evolution prediction implementation scheme uses spatio-temporal fusion to realize LSTM capturing time dynamics + Transformer modeling spatial correlation, and based on weight matrix to force prediction to comply with ecological rules, with greatly reduced error rate compared with traditional models; and through grid data, hybrid model architecture and ecological rule embedding, effective prediction of future period ecosystem natural evolution is realized.

[0270] II. Climate scenario deduction

[0271] (1) Scenario parameter injection mechanism

[0272] Parameter types include: boundary conditions (temperature gradient field), scenario parameters (variables used to define specific scenarios); boundary conditions are the basic input conditions for large model running, providing background environment for scenario deduction, helping large model understand how climate changes without human intervention; scenario parameters include intervention variables (such as artificial irrigation areas) and other parameters, intervention variables are used to simulate the impact of human intervention measures on climate and ecology. Data fusion: inject scenario parameters (here, intervention variables) into base data through functions, modify base data temperature and add intervention markers.

[0273] (2) Deduction workflow

[0274] Forward propagation: inject scenario parameters into digital twin large model, trigger dynamic kernel recalculation, such as weight matrix adjustment, response function update; critical state detection: monitor output variables in real time, activate rule engine to generate recommendations when carbon sink is below critical value; multi-threaded deduction: parallel computing of different scenarios, comparison of key indicator differences.

[0275] (3) Deduction result visualization

[0276] Results are displayed through dynamic heat maps, red areas indicate carbon sink below critical value, blue arrows indicate recommended irrigation priority areas. Carbon sink can reflect the response of the ecosystem to climate change, so carbon sink is used as the result of climate scenario deduction.

[0277] For example, in the basin rainforest project, through deduction, it is predicted that the carbon sink will abnormally decrease in a certain year, and according to the deduction result, it is suggested to take artificial irrigation measures to restore the carbon sink. Through the implementation of these intervention measures, the carbon sink is restored to 92% of the baseline level.

[0278] The following illustrates the case of visual interaction. The park manager uses AR glasses to make decisions, the user selects a rainforest area by gesture, and inputs the instruction of "cut down 10% of the sick trees". The digital twin large model based on the dynamic kernel and the prediction model deduces and outputs the following content in real time: the risk of disease spread decreases by 20% in the short term (response function calculation), and the loss of bird diversity is 8% in the long term (weight matrix chain reaction). The AR interface superimposes the sick tree removal range (red translucent) and the bird habitat change (such as dynamic shrinkage animation). Compared with the prior art which cannot provide real-time feedback of intervention effect, the embodiment can realize the real-time closed loop of interaction-deduction-visualization.

[0279] Figure 4 An interaction architecture diagram provided by an embodiment of the application is shown.

[0280] The AR glasses support gesture recognition and spatial anchor positioning; the interaction middleware is used to convert user gestures into operation instructions; the digital twin large model runs on an edge computing node and supports second-level response; the AR rendering engine is responsible for converting the output content of the digital twin large model into a visual AR scene in real time, so that the user can intuitively understand; the dynamic kernel database stores and manages various parameters and state data required for the large model to run, such as ecological parameters, environmental variables, weight information, etc., to provide necessary data support for the digital twin large model.

[0281] The AR glasses and the interaction middleware perform low-latency bidirectional communication. For example, a forest ranger manages the rainforest through AR glasses. The forest ranger can select an area in the glasses by gesture, which corresponds to about 2.3 hectares of endangered tree species habitat, and the forest ranger uses gestures and voice to issue instructions on the AR glasses, such as cutting down 10% of the sick trees in the selected area. The interaction middleware recognizes the gesture as a polygon area in GeoJSON format (this format can accurately describe the area selected by the forest ranger in the AR glasses), and converts the voice instruction into text.

[0282] The interaction process between the interaction middleware and the digital twin large model is as follows: the interaction middleware converts gestures into a polygon in GeoJSON format, converts voice instructions into text, parses the text instructions, identifies key information such as "cut down" and "10%", and then converts these information into standardized operation codes. For example, "cut down 10% of the sick trees" can be converted into an operation type code (such as TREE_REMOVAL) and an intervention intensity parameter (such as intensity = 0.1). In order to ensure that the selected cutting area is completely located within the ecological protection area, the interaction middleware uses topological tools to calculate the spatial relationship and verify whether the polygon area selected by the user is completely located within the ecological protection area, so as to avoid unnecessary ecological damage. Once the instructions are standardized and the spatial verification is passed, the interaction middleware sends these data to the digital twin large model.

[0283] The interaction process between the digital twin large model and the dynamic kernel database is as follows: the digital twin large model requests detailed data of the current ecosystem from the dynamic kernel database, such as sick tree distribution, healthy tree status, soil conditions, etc. The dynamic kernel database returns a response result including all related ecological parameters after data arrangement based on the request. The digital twin large model can output the current state of the ecosystem based on the obtained detailed data in real time; can simulate the immediate impact of cutting down 10% of the sick trees on the ecosystem (such as increase of light, change of growth rate of healthy trees, etc.), analyze how these changes affect the state trend of the ecosystem; can identify the critical state of the ecosystem, and once a potential critical state is identified, a decision response is made; can make natural evolution prediction and scenario deduction, such as predicting how the ecosystem will naturally evolve without additional intervention, including the natural death of sick trees and the growth of new trees, and again, under a set intervention scenario (such as different proportions of sick tree cutting), deducing the possible response of the ecosystem and evaluating the effect of different management strategies.

[0284] The interaction process between the digital twin large model and the AR rendering engine is as follows: the large model sends the output content (such as ecosystem running results and ecosystem prediction results) to the AR rendering engine, including specific numerical values and data that need to be visualized, and the AR rendering engine converts the data into heat maps, animations, etc. visual effects, ready to be displayed on AR glasses. If there is an ecological risk, the AR rendering engine will also generate a warning message.

[0285] The interaction process between the AR rendering engine and the AR glasses is as follows: through wireless connection, the visualized data is quickly transmitted to the AR glasses. The forester can use gestures to interact with the visualized results on the AR glasses, view detailed information, and the AR glasses can instantly display the impact of the forester's instructions on the ecosystem to assist decision-making.

[0286] Through the above process, the forester can clearly understand the potential impact of cutting down the sick tree and make management decisions accordingly. The real-time feedback mechanism improves the efficiency and accuracy of decision-making, enhancing the management capability of the ecosystem.

[0287] The embodiment of the present application also provides a digital evolution system of an ecosystem, as shown in the accompanying drawings, comprising: Figure 5

[0288] The acquisition module 501 is configured to acquire the information of animals and plants, the environmental quality assessment situation, and the ecosystem service function assessment result in the ecological protection area according to the biological data and the environmental data collected in the ecological protection area;

[0289] The generation learning module 502 is configured to generate a comprehensive feature matrix based on the features extracted from the information of animals and plants, the environmental quality assessment situation, and the ecosystem service function assessment result, and perform deep learning on the comprehensive feature matrix to construct a digital twin engine, wherein the comprehensive feature matrix is composed of fusion feature vectors of multiple space-time units arranged in a space-time sequence, and each fusion feature vector of a space-time unit is formed by multi-modal feature splicing of the information of animals and plants, the environmental quality assessment situation, and the ecosystem service function assessment result corresponding to the current space-time unit;

[0290] The construction module 503 is configured to construct a digital twin large model based on the digital twin engine, large model static base information, and a prediction model supporting natural evolution prediction and scenario deduction under a set scenario;

[0291] The processing and display module 504 is configured to process real-time data input into the digital twin large model and target data, and display panoramic dynamic information of the ecosystem on a visual interface, wherein the target data includes historical data of the ecosystem, the visual interface is a virtual reality (VR) interface or an augmented reality (AR) interface, and the panoramic dynamic information of the ecosystem includes at least one of a natural evolution prediction result and a scenario deduction result.

[0292] Optionally, the acquisition module comprises:

[0293] The generation sub-module is configured to generate the biological feature vector based on the biological data and generate the environmental feature vector based on the environmental data, wherein the biological data at least includes image data, audio data, and positioning data of animals and plants, and the environmental data at least includes meteorological data and physical environment data;

[0294] The acquisition sub-module is configured to perform species identification on the biological feature vector based on a biological large model and perform environmental analysis on the environmental feature vector based on an environmental large model, and acquire the information of animals and plants and the environmental quality assessment situation in the ecological protection area.​

[0295] a processing acquisition submodule, configured to process the animal and plant information, the environment quality assessment condition, the biological feature vector, the environment feature vector, and a comprehensive feature vector to obtain an ecosystem service function assessment result of the ecological protection zone, the comprehensive feature vector being generated based on fusion of the biological feature vector and the environment feature vector.

[0296] Optionally, the processing acquisition submodule comprises:

[0297] a splicing determination unit configured to splice the biological feature vector and the environment feature vector based on corresponding weights to determine the comprehensive feature vector.

[0298] an analysis generation unit configured to perform comprehensive reasoning analysis on the biological feature vector, the environment feature vector, the comprehensive feature vector, the animal and plant information, and the environment quality assessment condition based on the optimized evaluation large model to generate the ecosystem service function assessment result including an ecosystem species diversity assessment result and an ecosystem environment assessment result.

[0299] Optionally, the generation learning module comprises:

[0300] a first processing submodule configured to extract features from the multi-source data for fusion to generate a fusion feature vector corresponding to the same time and space dimension after sequentially performing data preprocessing and data alignment on the multi-source data, and to construct the comprehensive feature matrix based on a plurality of fusion feature vectors, the multi-source data comprising the animal and plant information, the environment quality assessment condition, and the ecosystem service function assessment result.

[0301] a learning generation submodule configured to perform deep learning on the comprehensive feature matrix to generate a dynamic kernel including a weight matrix, a response function, and a threshold rule, the weight matrix being used to quantify the action relationship between species and environment and service function, the response function being used to describe the response relationship between ecosystem variables, and the threshold rule being used to define the critical state of the ecosystem, the dynamic kernel being used to calculate the state of the ecosystem, assess the trend of the state change of the ecosystem, and identify the critical state.

[0302] a construction submodule configured to encapsulate the dynamic kernel and a rule engine to construct the digital twin engine, the rule engine being used to make a decision response based on an output result of the dynamic kernel.

[0303] Optionally, the system further comprises:

[0304] A determining module is configured to determine a data sample set based on an ecosystem history sample with a labeled scenario category and label data indicating a real observation value at a target moment, the data sample set including a training set, a validation set and a test set, the label data belonging to a model training target, and the target moment being a future moment relative to a history moment;

[0305] A first processing module is configured to perform model training on a machine learning model applicable to ecosystem time series prediction based on the training set, update model parameters through an optimization algorithm, and evaluate model performance based on the validation set;

[0306] A second processing module is configured to evaluate prediction accuracy and scenario response consistency of the model based on the test set, and optimize model structure and hyperparameters based on evaluation indexes to determine the prediction model.

[0307] Optionally, the constructing module comprises:

[0308] A loading submodule is configured to load the large model static base information for constructing a three-dimensional space skeleton and basic attributes to the digital twin engine to complete space registration and logical binding;

[0309] An integrating submodule is configured to integrate the trained prediction model into the digital twin engine;

[0310] A configuring submodule is configured to configure various model parameters in the digital twin engine to initialize the digital twin large model.

[0311] Optionally, the processing display module comprises:

[0312] A second processing submodule is configured to access the digital twin engine based on a data interface to drive the digital twin engine to update a model state of the digital twin large model, the real-time data including real-time plant and animal information, real-time environmental quality evaluation, and real-time ecosystem service function evaluation results;

[0313] A third processing submodule is configured to process the real-time data and the target data based on the updated digital twin large model to output an ecosystem operation result and an ecosystem prediction result, the ecosystem operation result including at least one of an ecosystem current state, an ecosystem state change trend, an ecosystem critical state, and a decision response, and the ecosystem prediction result including at least one of the natural evolution prediction result and the scenario deduction result;

[0314] The fourth processing sub-module is configured to generate the ecosystem panoramic dynamic information based on the ecosystem running result and the ecosystem prediction result provided by the digital twin large model, and visually display the ecosystem panoramic dynamic information through the VR interface or the AR interface.

[0315] Optionally, the third processing sub-module comprises:

[0316] The first processing unit is configured to process the real-time data and the ecosystem historical data based on the updated digital twin large model, calculate the current state of the ecosystem, evaluate the state change trend of the ecosystem, identify the critical state of the ecosystem, obtain the ecosystem evaluation information, and feed back a decision response based on logical judgment.

[0317] The second processing unit is configured to perform scenario deduction based on the ecosystem evaluation information, the set scenario parameter, and the target data, and / or perform natural evolution prediction based on the ecosystem evaluation information and the target data, and output the ecosystem prediction result. The target data further comprises external driving factors, time information, and space information.

[0318] Optionally, the system further comprises:

[0319] The update acquisition module is configured to update the model state of the digital twin large model in response to receiving an interactive operation on the VR interface or the AR interface, and acquire updated ecosystem panoramic dynamic information output by the digital twin large model.

[0320] The update module is configured to update the visual content on the VR interface or the AR interface in real time based on the updated ecosystem panoramic dynamic information.

[0321] For the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the related parts are referred to the part of the method embodiment.

[0322] It should be noted that the above detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0323] In the above detailed description, reference is made to the accompanying drawings, which form a part hereof. In the drawings, similar symbols typically identify similar components, unless context dictates otherwise. The illustrative embodiments described in the detailed description, drawings, and claims are not meant to be limiting. Other embodiments can be used, and other changes can be made, without departing from the spirit or scope of the subject matter presented herein.

[0324] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. The present application can have various changes and modifications for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for digital evolution of an ecosystem, characterized in that, The method comprises the following steps: According to the biological data and environmental data collected in the ecological protection zone, the information of animals and plants, the environmental quality evaluation and the ecosystem service function evaluation results in the ecological protection zone are obtained; wherein, the biological feature vector is generated based on the biological data, and the environmental feature vector is generated based on the environmental data; the information of animals and plants, the environmental quality evaluation, the biological feature vector, the environmental feature vector and the comprehensive feature vector are processed to obtain the ecosystem service function evaluation results of the ecological protection zone, and the comprehensive feature vector is determined based on the fusion of biological feature vector and environmental feature vector and the dynamically adjusted biological weight and environmental weight; Based on the features extracted from the information of animals and plants, the environmental quality evaluation and the ecosystem service function evaluation results, a comprehensive feature matrix is generated, and a digital twin engine is constructed through deep learning of the comprehensive feature matrix, wherein the comprehensive feature matrix is composed of fusion feature vectors of multiple space-time units arranged in space-time sequence, and a single space-time unit fusion feature vector is formed by multi-modal feature splicing of the information of animals and plants, the environmental quality evaluation and the ecosystem service function evaluation corresponding to the current space-time unit; Based on the digital twin engine, a large model static base information and a prediction model, a digital twin large model is constructed, and the prediction model supports natural evolution prediction and scenario deduction under a set scenario; In response to inputting real-time data of the digital twin large model, the real-time data and target data are processed, and the ecological system panoramic dynamic information is displayed on a visual interface, the target data includes ecological system historical data, the visual interface is a virtual reality (VR) interface or an augmented reality (AR) interface, and the ecological system panoramic dynamic information includes at least one of natural evolution prediction results and scenario deduction results.

2. The method of claim 1, wherein: The biological data at least includes image data, audio data and positioning data of animals and plants, and the environmental data at least includes meteorological data and physical environment data; Based on the biological large model, species identification is performed on the biological feature vector, and based on the environmental large model, environmental analysis is performed on the environmental feature vector to obtain the information of animals and plants and the environmental quality evaluation in the ecological protection zone.

3. The method of claim 2, wherein, The processing of the information of animals and plants, the environmental quality evaluation, the biological feature vector, the environmental feature vector and the comprehensive feature vector to obtain the ecosystem service function evaluation results of the ecological protection zone comprises: According to the biological feature vector, the environmental feature vector and the corresponding weight, the vector splicing is performed to determine the comprehensive feature vector; Based on the optimized evaluation large model, the biological feature vector, the environmental feature vector, the comprehensive feature vector, the information of animals and plants and the environmental quality evaluation are comprehensively analyzed to generate the ecosystem service function evaluation results including the ecosystem species diversity evaluation results and the ecosystem environmental evaluation results.

4. The method of claim 1, wherein, The comprehensive feature matrix is generated based on the features extracted from the animal and plant information, the environment quality assessment situation, and the ecosystem service function assessment result, and deep learning is performed on the comprehensive feature matrix to construct a digital twin engine, including: After sequentially performing data preprocessing and data alignment on the multi-source data, features are extracted from the multi-source data for fusion to generate a fusion feature vector corresponding to the same time and spatial dimension, and the comprehensive feature matrix is constructed based on multiple fusion feature vectors, the multi-source data including the animal and plant information, the environment quality assessment situation, and the ecosystem service function assessment result; Deep learning is performed on the comprehensive feature matrix to generate a dynamic kernel including a weight matrix, a response function, and a threshold rule, the weight matrix being used to quantify the action relationship between species and environment and service function, the response function being used to describe the response relationship between ecosystem variables, and the threshold rule being used to define the critical state of the ecosystem, and the dynamic kernel being used to calculate the state of the ecosystem, assess the trend of the state change of the ecosystem, and identify the critical state; The dynamic kernel and a rule engine are packaged to construct the digital twin engine, and the rule engine makes a decision response based on the output result of the dynamic kernel.

5. The method of claim 1, wherein, The method further includes: A data sample set is determined based on an ecosystem historical sample with a labeled scenario category and label data indicating a real observation value at a target time, the data sample set including a training set, a validation set, and a test set, the label data belonging to a model training target, and the target time being a future time relative to a historical time; A machine learning model suitable for ecosystem time series prediction is trained based on the training set, model parameters are updated through an optimization algorithm, and model performance is evaluated based on the validation set; The prediction accuracy and scenario response consistency of the model are evaluated based on the test set, and the model structure and hyperparameters are optimized based on evaluation indicators to determine the prediction model.

6. The method according to claim 1 or 5, characterized in that, The digital twin large model is constructed based on the digital twin engine, large model static base information, and the prediction model, including: The large model static base information used to construct a three-dimensional spatial skeleton and basic attributes is loaded into the digital twin engine to complete spatial registration and logical binding; The trained prediction model is integrated into the digital twin engine; Model parameters in the digital twin engine are configured to initialize the digital twin large model.

7. The method of claim 4, wherein, In response to inputting real-time data into the digital twin large model, the real-time data and target data are processed, and ecosystem panoramic dynamic information is displayed on a visualization interface, including: Real-time data obtained through a data interface are connected to the digital twin engine to drive the digital twin engine to update the model state of the digital twin large model, the real-time data including real-time animal and plant information, real-time environment quality assessment situation, and real-time ecosystem service function assessment result; The real-time data and the target data are processed based on the updated digital twin large model, and ecosystem operation results and ecosystem prediction results are output, the ecosystem operation results including at least one of an ecosystem current state, an ecosystem state change trend, an ecosystem critical state and a decision response, and the ecosystem prediction results including at least one of the natural evolution prediction result and the scenario deduction result; The ecosystem panoramic dynamic information is generated based on the ecosystem operation results and the ecosystem prediction results provided by the digital twin large model, and is visually displayed through the VR interface or the AR interface.

8. The method of claim 7, wherein, The real-time data and the target data are processed based on the updated digital twin large model, and ecosystem operation results and ecosystem prediction results are output, including: The real-time data and the ecosystem historical data are processed based on the updated digital twin large model, an ecosystem current state is calculated, an ecosystem state change trend is evaluated, an ecosystem critical state is identified, to obtain ecosystem evaluation information, and a decision response is fed back based on logical judgment; The scenario deduction is performed based on the ecosystem evaluation information, set scenario parameters and the target data, and / or the natural evolution prediction is performed based on the ecosystem evaluation information and the target data, to output the ecosystem prediction results, the target data further including external driving factors, time information and space information.

9. The method according to claim 7 or 8, characterized in that, The method further includes: In response to receiving an interactive operation on the VR interface or the AR interface, a model state of the digital twin large model is updated, and updated ecosystem panoramic dynamic information output by the digital twin large model is obtained; Based on the updated ecosystem panoramic dynamic information, the visualized content on the VR interface or the AR interface is updated in real time.

10. A digital evolution system of an ecosystem, characterized by, The method further includes: The acquisition module is configured to acquire biological data and environmental data collected in the ecological protection area, and acquire animal and plant information, environmental quality evaluation information and ecosystem service function evaluation results in the ecological protection area; specifically, the acquisition module is configured to generate a biological feature vector based on the biological data, and generate an environmental feature vector based on the environmental data; and process the animal and plant information, the environmental quality evaluation information, the biological feature vector, the environmental feature vector and a comprehensive feature vector to obtain the ecosystem service function evaluation results of the ecological protection area, the comprehensive feature vector being determined based on the biological feature vector and the environmental feature vector and fusion of dynamically adjusted biological weights and environmental weights. The generating learning module is configured to generate a comprehensive feature matrix based on the features extracted from the animal and plant information, the environment quality assessment situation, and the ecosystem service function assessment result, and to perform deep learning on the comprehensive feature matrix to construct a digital twin engine. The comprehensive feature matrix is composed of fused feature vectors of multiple spatio-temporal units arranged in a spatio-temporal sequence. The fused feature vector of a single spatio-temporal unit is formed by multi-modal feature splicing of the animal and plant information, the environment quality assessment situation, and the ecosystem service function assessment result corresponding to the current spatio-temporal unit. The constructing module is configured to construct a digital twin large model based on the digital twin engine, large model static base information, and a prediction model. The prediction model supports natural evolution prediction and scenario deduction under a set scenario. The processing and display module is configured to process real-time data and target data in response to input of the digital twin large model, and to display ecosystem panoramic dynamic information on a visual interface. The target data includes ecosystem historical data. The visual interface is a virtual reality (VR) interface or an augmented reality (AR) interface. The ecosystem panoramic dynamic information includes at least one of a natural evolution prediction result and a scenario deduction result.

Images