An ecological restoration method and system based on multi-modal data fusion

Abstract

This invention provides an ecological restoration method and system based on multimodal data fusion, relating to the field of data processing technology. The method includes: collecting multimodal ecological data of a target ecological region; extracting features from the multimodal ecological data to obtain synchronous modal feature data and text feature vectors; performing feature-level fusion on the synchronous modal feature data to determine synchronous fusion features; interpolating and enhancing the synchronous fusion features to generate enhanced fusion features; performing decision-level evidence fusion on the enhanced fusion features and text feature vectors to obtain a multimodal ecological state evidence body; identifying key ecological constraint factors and uncertainty intervals in the target ecological region; constructing an ecological vulnerability distribution map; performing dynamic Monte Carlo simulation in the ecological restoration measures space to generate a multi-branch ecological restoration scenario path set; and generating a spatiotemporal control script for the execution of ecological restoration projects based on the multi-branch ecological restoration scenario path set.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an ecological restoration method and system based on multimodal data fusion. Background Technology

[0002] With the increasing severity of ecological and environmental problems, the degradation of various ecosystems is becoming increasingly prominent globally. Issues such as vegetation degradation, water body shrinkage, soil erosion, and declining biodiversity are gradually becoming important factors restricting regional sustainable development. In order to effectively restore damaged ecosystems, various ecological restoration projects are gradually being widely applied in fields such as watershed management, mine revegetation, wetland restoration, and desertification control.

[0003] Ecosystems are inherently complex, dynamic, and characterized by multi-factor coupling. Their evolution is often influenced by a combination of factors, including climate, hydrology, soil, vegetation, and human activities. Therefore, how to utilize multi-source heterogeneous data to comprehensively analyze the state of ecosystems and formulate more scientific and rational ecological restoration strategies has become an important research direction in the field of ecological governance.

[0004] However, in the existing ecological restoration technology system, most methods still rely mainly on the analysis of single types of data, such as vegetation index assessment methods based on remote sensing images or ecological indicator analysis methods based on environmental monitoring data. When faced with complex ecosystems, these methods often fail to fully reflect the ecological change process under the coupling of multiple factors, resulting in a lack of in-depth exploration of the potential structural relationships between different modal data. At the same time, in terms of ecological restoration decision-making, traditional methods mostly select restoration measures based on expert experience or fixed rules, lacking a systematic simulation and probabilistic assessment mechanism for the evolution path of different restoration strategies, making it difficult to cope with the uncertainty and dynamic changes that are common in ecosystems. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide an ecological restoration method based on multimodal data fusion, which can solve the technical problem that traditional methods mostly select restoration measures based on expert experience or fixed rules, lack a systematic simulation and probability assessment mechanism for the evolution path of different restoration strategies, and are difficult to cope with the uncertainty and dynamic changes that are common in ecosystems.

[0006] A first aspect of this invention proposes an ecological restoration method based on multimodal data fusion, comprising:

[0007] S1: Collect multimodal ecological data of the target ecological area, including synchronous ecological data and asynchronous text data;

[0008] S2: Using a pre-trained single-modal encoder, features are extracted from synchronous ecological data and asynchronous text data respectively to obtain synchronous modal feature data and text feature vectors;

[0009] S3: Using the marginal Fisher analysis method, feature-level fusion of synchronous modal characteristic data is performed to determine the synchronous fusion characteristics;

[0010] S4: By using a data augmentation method based on shared mixing coefficients, the synchronous fusion features are interpolated and enhanced to generate enhanced fusion features;

[0011] S5: Decision-level evidence fusion is performed on the enhanced fusion features and text feature vectors to obtain a multimodal ecological state evidence body;

[0012] S6: Based on the basic probability distribution of each ecological proposition in the multimodal ecological state evidence body, identify multiple key ecological constraint factors and uncertainty intervals in the target ecological region;

[0013] S7: Construct an ecological vulnerability distribution map based on the coupling relationship between key ecological constraint factors;

[0014] S8: Combining key ecological constraint factors, uncertainty intervals, and ecological vulnerability distribution maps, dynamic Monte Carlo simulations are performed in the ecological restoration measures space to generate a multi-branch ecological restoration scenario path set containing probability weights;

[0015] S9: Generate a spatiotemporal control script for the execution of ecological restoration projects based on the multi-branch ecological restoration scenario path set.

[0016] A second aspect of this invention proposes an ecological restoration system based on multimodal data fusion, comprising: a processor and a memory;

[0017] The memory stores programs or instructions that can run on a processor, which, when executed by the processor, implement steps of the ecological restoration method based on multimodal data fusion as described in the first aspect.

[0018] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following:

[0019] In this embodiment of the invention, feature-level fusion of synchronous modal features is performed using marginal Fisher analysis, combined with data augmentation using shared mixing coefficients for interpolation enhancement. This improves the fused features' ability to characterize and robustness against complex ecological and environmental changes. By analyzing the basic probability distribution in the evidence body, key ecological constraint factors are identified and uncertainty intervals are determined, revealing the main limiting factors in the ecosystem. Furthermore, an ecological vulnerability distribution map is constructed through the coupling and propagation relationships of key ecological constraint factors, achieving a refined characterization of the spatial distribution characteristics of ecological risks. Based on this, a multi-branch ecological restoration scenario path set containing probability weights is generated through dynamic Monte Carlo simulation, and a spatiotemporal control script for ecological restoration engineering execution is generated accordingly. This enables probabilistic assessment and dynamic decision-making of ecological restoration strategies, improving the scientific rigor, adaptability, and implementation efficiency of ecological restoration schemes. Attached Figure Description

[0020] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0021] Figure 1 This is a flowchart illustrating an ecological restoration method based on multimodal data fusion provided in an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the structure of an ecological restoration system based on multimodal data fusion provided in an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] The following description, in conjunction with the accompanying drawings, details the ecological restoration method based on multimodal data fusion provided by the embodiments of the present invention through specific examples and application scenarios.

[0025] Reference manual attached Figure 1 The diagram illustrates a flowchart of an ecological restoration method based on multimodal data fusion provided by an embodiment of the present invention.

[0026] This invention provides an ecological restoration method based on multimodal data fusion, which may include the following steps:

[0027] S1: Collect multimodal ecological data of the target ecological area, including synchronous ecological data and asynchronous text data.

[0028] The target ecological area refers to a specific geographical area where ecological status assessment and ecological restoration planning are required. This area can be a watershed, wetland, mining area, forest land, grassland, or other areas with ecological governance needs. Multimodal ecological data refers to a collection of ecological and environmental data obtained from different data sources and in different information expression forms, used to reflect the state of the ecosystem from multiple dimensions. Synchronous ecological data refers to data types acquired at the same or approximately the same time scale and capable of time-aligned processing. These typically include remote sensing or visual image data, time-series monitoring data from environmental sensors, and ecological acoustic audio data. This type of data can reflect the dynamic changes of the ecosystem in real-time or near real-time. Asynchronous text data refers to ecological information textual materials not collected in a strictly time-series format, such as ecological survey reports, patrol records, expert evaluation opinions, historical governance records, and ecological monitoring explanatory documents. This type of data usually contains a large amount of semantic knowledge and empirical information, which can supplement the background knowledge and long-term change information of the ecosystem.

[0029] In one possible implementation, the synchronized ecological data includes visual image data, environmental sensor time-series data, and audio data.

[0030] It should be noted that synchronous ecological data can continuously reflect the dynamic changes of ecosystems in time and space, while asynchronous text data can supplement ecological survey experience, historical governance, and expert knowledge, enabling ecological status analysis to not only rely on real-time monitoring data but also incorporate long-term accumulated ecological knowledge. This multimodal data acquisition method provides a more comprehensive and richer data foundation for subsequent feature extraction, multimodal fusion, and ecological status determination, thereby improving the accuracy and stability of ecological status identification.

[0031] S2: Using a pre-trained unimodal encoder, features are extracted from synchronous ecological data and asynchronous text data respectively to obtain synchronous modal feature data and text feature vectors.

[0032] Among them, a pre-trained single-modal encoder refers to a feature extraction model trained in advance on large-scale data for a specific data modality (such as images, time-series data, audio, or text). This model can transform raw data into representative feature representations. For example, a visual encoder is used to extract spatial structure and texture features of images, a time-series encoder is used to extract trend features in environmental sensor data, an audio encoder is used to extract eco-acoustic spectral features, and a text encoder is used to extract semantic features of text. Synchronous modal feature data refers to the structured feature representation obtained after processing by various synchronous data encoders, used to describe the spatial or temporal features of the ecological environment. Text feature vectors refer to the vectorized representation obtained after semantic modeling of text data by a text encoder, used to represent the ecological semantic knowledge contained in the text information.

[0033] Specifically, the collected multimodal ecological data is first divided into data types: visual image data, environmental sensor time-series data, audio data, and text data. Each data type undergoes preprocessing to match its respective encoder input format. This preprocessing includes at least data cleaning, time alignment, scale unification, missing data completion, and standardization. Then, for each modality, its corresponding pre-trained single-modal encoder is invoked. Visual image data is input into the pre-trained visual encoder to extract spatial texture and target distribution features; environmental sensor time-series data is input into the pre-trained temporal encoder to extract temporal variation features; audio data is input into the pre-trained audio encoder to extract spectral and acoustic event features; and text data is input into the pre-trained text encoder to extract semantic expression features. Next, each pre-trained single-modal encoder outputs a high-dimensional feature representation for its corresponding modality. These high-dimensional feature representations are then organized into multiple independent feature data sets. Each feature data set corresponds to a feature expression result of a specific modality within the current target ecological region, thus completing the feature extraction of the multimodal ecological data.

[0034] It should be noted that by utilizing a pre-trained unimodal encoder to extract targeted features from different types of ecological data, the original multi-source data can be uniformly transformed into structured high-dimensional feature representations, thereby improving the computability and fusionability of multimodal data. This approach not only reduces noise and redundant information in the original data but also enhances the sensitivity of feature representations to changes in ecological state, providing a more stable and discriminative data foundation for subsequent multimodal feature fusion and ecological state discrimination, thus improving the accuracy and reliability of the entire ecological restoration analysis process.

[0035] S3: Using the marginal Fisher analysis method, feature-level fusion is performed on the synchronous modal feature data to determine the synchronous fusion features.

[0036] Marginal Fisher analysis is a statistical analysis method used for feature dimensionality reduction and discriminative feature extraction. Its core idea is to construct a feature mapping matrix to map the original features to a new latent space, making the distances between samples of the same class as close as possible in the mapping space, while maximizing the separation between samples of different classes, thereby enhancing the discriminative power of the features. Synchronous fusion features refer to the fusion feature representation formed in a shared latent space after processing by marginal Fisher analysis. This feature simultaneously contains key information from multiple synchronous modalities.

[0037] In one possible implementation, S3 specifically includes:

[0038] S301: Construct a multimodal feature matrix based on the synchronous modal feature data.

[0039] Specifically, features from different modalities (such as vision, sensors, and audio) are organized into a feature matrix, with each row corresponding to the features of a sample.

[0040] S302: Calculate the covariance matrix between the modal features in the multimodal feature matrix.

[0041] Specifically, firstly, feature data for each of the two modalities to be analyzed are extracted from the multimodal feature matrix across all samples, and the average values ​​of these two modal features are calculated separately. Then, for each sample, the deviation of the sample from the average value of the first modal feature and the deviation of the sample from the average value of the second modal feature are calculated. Next, these two deviations are combined to obtain the degree of common variation of the sample between the two modal features. Finally, the degrees of common variation for all samples are summed and normalized according to the sample size to obtain the covariance matrix between the two modal features. Through this process, the joint fluctuation relationship of different modal features across all samples can be characterized, thus providing a foundation for subsequent correlation analysis and fusion calculations between multimodal features.

[0042] S303: Using marginal Fisher analysis, each modal feature is mapped to a shared latent space.

[0043] Specifically, the sample features are first divided into different classes (e.g., c1 and c2) according to their categories. The class centers of each class are calculated. For each sample, a linear mapping is performed to obtain the latent space representation. Then, the squared distance from the mapped sample to its class center is calculated and summed within the class as the numerator. Simultaneously, the squared distance from the unmapped sample to its class center is calculated and summed within the class as the denominator. For each class, the ratio of the unmapped intra-class scatter to the unmapped intra-class scatter is calculated and summed to obtain the final value. By minimizing Learning W significantly reduces the intra-class distance relative to the original space after mapping (making the intra-class space more compact), thus forming clearer class clusters in the shared latent space. When class clusters are compressed, the overlap between different clusters naturally decreases, and separability increases (reflecting "easier to separate classes"). This is called the criterion function. This represents a linear mapping matrix.

[0044] For example, taking the binary classification of ecological states as an example, c1 represents "severe degradation" and c2 represents "slight degradation", and the multimodal synchronization feature of each sample point is f. i In the original space, the two types of sample points may be loosely distributed and partially overlapped due to season, terrain, and noise. The calculation and minimization of these factors are crucial. The resulting W will project samples belonging to "severe degradation" into a denser point cloud, and samples belonging to "slight degradation" into another denser point cloud. The boundaries between the two point clouds are clearer, making subsequent discrimination based on "which type of center is closer / which type of evidence is stronger" more stable, thereby achieving synchronous fusion feature construction under the shared potential space.

[0045] S304: Determine the fusion-shared features based on the covariance matrix and mapping results.

[0046] S305: Normalize the fusion and sharing features to determine the synchronous fusion features.

[0047] It should be noted that, based on the traditional marginal Fisher analysis method, a structured fusion of multimodal ecological data is achieved by introducing a multimodal feature collaborative modeling mechanism and covariance structure constraints. Furthermore, the marginal Fisher analysis method maps multimodal features to a shared latent space, and a multimodal covariance constraint term is introduced when solving the mapping matrix. This ensures that the structural relationships between different modal features in the latent space are preserved, thus avoiding the modal information loss problem that may occur when traditional marginal Fisher analysis only performs dimensionality reduction on a single modality or a simple feature set.

[0048] S4: By using a data augmentation method based on shared mixing coefficients, the synchronous fusion features are interpolated and enhanced to generate enhanced fusion features.

[0049] In feature augmentation, the shared mixing coefficient is a weighting parameter used to control the proportion of linear combinations between two sample features. Its value typically ranges from 0 to 1, determining the contribution ratio of each feature in generating a new feature. Data augmentation refers to methods that generate new feature samples by transforming, combining, or perturbing existing sample features, thereby expanding the data distribution range and improving the model's generalization ability. Enhanced fusion features refer to new feature representations generated through interpolation using shared mixing coefficients based on synchronously fused features. These features expand the feature space distribution while maintaining the original semantic structure.

[0050] It should be noted that, compared with traditional simple data copying or random perturbation methods, this method generates new features by weighting and combining the features of two real samples. This not only maintains the semantic continuity of the original features, but also forms a smoother and richer data distribution in the feature space, thereby improving the model's adaptability to complex ecological and environmental changes.

[0051] In one possible implementation, S4 specifically includes:

[0052] S401: Construct a batch of synchronous fusion features based on synchronous fusion characteristics.

[0053] S402: Perform pairwise sampling on the batch of synchronous fusion features to determine multiple pairs of synchronous fusion features.

[0054] Specifically, the batch of synchronous fusion features is sampled in pairs, and the batch of synchronous fusion features is divided into a first synchronous fusion feature subset and a second synchronous fusion feature subset according to the sample order. The corresponding feature vectors in the first synchronous fusion feature subset and the second synchronous fusion feature subset are paired one by one to determine several pairs of synchronous fusion features.

[0055] S403: Define the shared mixing coefficient.

[0056] Specifically, first, two shape parameters, φ1 and φ2, are set for the Beta distribution to control the distribution shape of the random numbers. Then, a random number between 0 and 1 is randomly sampled from this Beta distribution; this random number is the shared mixing coefficient φ. Since the Beta distribution is limited to the range [0,1], the generated φ can be used as a linear interpolation weighting coefficient between two feature vectors. When φ is close to 1, the generated enhanced feature is closer to the first sample feature. When φ is close to 0, the generated enhanced feature is closer to the second sample feature. When φ is close to 0.5, it indicates that the two sample features are fused in approximately equal proportions.

[0057] S404: Within the shared latent space, the shared mixing coefficients and multiple sets of synchronous fusion feature pairs are combined to perform shared interpolation on the synchronous fusion features to determine the enhanced features.

[0058] Specifically, firstly, feature vectors of two corresponding samples are selected from the synchronous fusion feature set, denoted as the first synchronous fusion feature and the second synchronous fusion feature, respectively. Then, a shared mixing coefficient φ (within the range of 0 to 1) is introduced, and the two feature vectors are weighted using a linear combination method: the first synchronous fusion feature is multiplied by the weight φ, and the second synchronous fusion feature is multiplied by the weight 1 - φ. The two results are then summed to obtain a new feature vector, i.e., the enhanced feature. Samples from the same batch are interpolated using the same mixing coefficient within the shared latent space to maintain consistency in the multimodal semantic structure.

[0059] Specifically, since the same mixing coefficients are used in the feature calculation of the same set of samples, this interpolation process is called shared interpolation. It can generate new sample features while maintaining the semantic consistency of features, so as to expand the feature space and improve the robustness of the model to changes in multimodal data.

[0060] S405: Normalize the enhanced features to generate enhanced fusion features.

[0061] It should be noted that by introducing an interpolation enhancement mechanism based on shared mixing coefficients into the shared latent space, data enhancement processing is performed on the synchronously fused features, enabling the original feature space to generate more new feature samples with reasonable semantic structures, thereby effectively alleviating the problem of insufficient number or uneven distribution of ecological data samples.

[0062] S5: Decision-level evidence fusion is performed on the enhanced fusion features and text feature vectors to obtain a multimodal ecological state evidence body.

[0063] Among them, decision-level evidence fusion refers to the fusion of judgment results from different information sources at the decision-making level, forming a unified judgment result by integrating the support levels of multiple evidence sources. Multimodal ecological state evidence refers to the evidence representation structure formed by integrating information from different modalities through evidence fusion methods. It contains the basic probability allocation values ​​corresponding to each ecological state proposition, used to describe the degree of support for different ecological states and their uncertainty.

[0064] Among them, the decision-level evidence fusion adopts the Dempster-Shafer theory.

[0065] It should be noted that, compared to methods that rely solely on single-modal data for judgment, this step utilizes both synchronous ecological observation data and textual semantic information. This allows the ecological status analysis to incorporate both real-time monitoring information and historical experience and expert knowledge, thus forming a more comprehensive basis for judgment. This approach not only improves the accuracy of ecological status identification but also preserves uncertainties in the results, providing a more reliable and interpretable basis for subsequent identification of key ecological constraints and ecological restoration decisions.

[0066] In one possible implementation, S5 specifically includes:

[0067] S501: Calculate the similarity between the enhanced fusion features and the prototype vectors of each ecological state.

[0068] Among them, the ecological state prototype vector refers to the representative vector constructed in the feature space for each preset ecological state category (such as ecological health, mild degradation or severe degradation, etc.), which is used as the reference center for different ecological states.

[0069] Specifically, firstly, a prototype vector for each preset ecological state is pre-established. This prototype vector represents the typical distribution center of that ecological state in the feature space. Then, for the enhanced fusion feature corresponding to the current target ecological region, matching calculations are performed sequentially with each prototype vector. By comparing the directional consistency and numerical similarity between the enhanced fusion feature and each prototype vector in the feature space, similarity values ​​are obtained for the enhanced fusion feature relative to each prototype vector. A higher similarity value indicates a closer resemblance between the current enhanced fusion feature and the corresponding prototype vector, suggesting that the target ecological region is more likely to be in that ecological state. Finally, a set of similarity results corresponding one-to-one with each ecological state is obtained, providing a foundation for subsequent calculations of the probability distribution of the target ecological region belonging to each ecological state.

[0070] S502: Based on similarity and using the softmax function, calculate the probability distribution of the first ecological state:

[0071]

[0072] in, This indicates that the enhanced fusion feature vector is known. Under the given conditions, the probability that sample y belongs to the j-th ecological state is given by exp(), which represents an exponential function. This represents the prototype vector of the j-th ecological state. This represents a temperature parameter used to adjust the smoothness of the probability distribution. Indicates transpose. This represents the prototype vector of the k-th ecological state, where K represents the total number of ecological state categories. This represents the enhanced fusion feature vector corresponding to the i-th sample.

[0073] Among them, the probability distribution of the first ecological state refers to the probability distribution result of the sample belonging to each ecological state category calculated based on the enhanced fusion features.

[0074] Specifically, the similarity between enhanced fusion features and prototypes of various ecological states is converted into comparable probability values. In other words, instead of directly stating "which category it most resembles," it first calculates the degree of closeness between the sample and each prototype ecological state, and then uses softmax normalization to obtain the probability distribution of the sample belonging to each ecological state. The advantages of this approach are twofold: firstly, it quantifies the degree to which a sample belongs to different ecological states; secondly, it facilitates subsequent decision-level evidence fusion with the probability results from the text modality, resulting in a more stable multimodal ecological state judgment.

[0075] S503: Calculate the probability distribution of the second ecological state based on text feature vectors.

[0076] The second ecological state probability distribution refers to the ecological state probability distribution result calculated based on the text feature vector.

[0077] Specifically, the calculation method for the probability distribution of the second ecological state is similar to that for the probability distribution of the first ecological state.

[0078] S504: Construct visual evidence and textual evidence based on the probability distributions of the first and second ecological states.

[0079] Among them, visual evidence and textual evidence refer to the conversion of probability distributions obtained from different modalities into basic probability allocation structures in evidence theory, which are used to represent the degree to which each ecological state proposition is supported by different modalities.

[0080] Specifically, the probability distributions of the first and second ecological states are mapped to visual evidence and textual evidence, respectively.

[0081] S505: Calculate the coefficient of conflict of evidence based on visual and textual evidence.

[0082]

[0083] Where M represents the coefficient of evidence conflict, and A i Let B represent the i-th proposition supported by visual evidence. j This represents the j-th proposition supported by the textual evidence. Denotes the empty set, m v The visual evidence represents proposition A.i The basic probability assignment value, m t The textual evidence indicates the relationship between proposition B and the proposition. j The basic probability allocation value, Proposition A i and Proposition B j There is no overlap between them.

[0084] The conflict coefficient is an indicator used to measure the degree of contradiction between different sources of evidence.

[0085] Specifically, for For example, if one piece of evidence states that "the area is in a healthy state" and another piece of evidence states that "the area is in a severely degraded state," then these two states are mutually exclusive if they cannot both be true under the current discrimination system.

[0086] Specifically, this formula is used to calculate the conflict coefficient between different modalities of evidence. Each term... This represents the conflict contribution between a set of mutually exclusive propositions. The total conflict degree M is obtained by summing all mutually exclusive combinations. Its main function is to provide conflict measurement and normalization basis for subsequent Dempster-Shafer evidence fusion.

[0087] S506: By using Dempster-Shafer theory and combining it with the coefficient of evidence conflict, the visual evidence and textual evidence are fused to obtain the quality of the fused evidence.

[0088]

[0089] in, The quality of the fused evidence is the basic probability assignment value of the fused visual and textual evidence for proposition A, i.e., the degree of support of the fused evidence for ecological state A.

[0090] Among them, the Dempster-Shafer theory is a mathematical theory for dealing with uncertain information and the fusion of multiple sources of evidence. It obtains a unified evidence support result by combining and calculating multiple sources of evidence.

[0091] Specifically, evidence from different modalities or sources is integrated using a method of "common support, conflict elimination, and normalized redistribution" to obtain a more reliable comprehensive judgment. Specifically, it first identifies the portions of evidence from both types that commonly support the same proposition, multiplies these supports, and then sums them up. Then... These non-conflicting supports are normalized to eliminate the impact of conflicting evidence on the results. In this way, the fused... Instead of relying on a single modality, it integrates information from both enhanced fusion features and textual features, making it more suitable for your multimodal ecological state assessment and subsequent restoration decisions.

[0092] S507: Construct a multimodal ecological state evidence body based on the quality of fused evidence.

[0093] Among them, the multimodal ecological state evidence body refers to the overall evidence representation structure formed by integrating visual modality and text modality evidence, which includes the basic probability allocation values ​​corresponding to each ecological state proposition.

[0094] Specifically, firstly, the quality of fused evidence corresponding to each ecological proposition obtained after fusing the aforementioned decision-level evidence is summarized, and a one-to-one correspondence is established between each ecological proposition and its corresponding fused evidence quality. Then, according to a pre-defined data organization method, all ecological propositions and their corresponding fused evidence quality are structurally encapsulated, ensuring that each ecological proposition has a corresponding support representation within the structure. The fused evidence quality characterizes the degree of support for the target ecological region belonging to the corresponding ecological state proposition after integrating both enhanced fused features and text feature vectors. Finally, all ecological propositions and their corresponding fused evidence quality are combined to form a multimodal ecological state evidence body, thereby creating a comprehensive evidence expression result that simultaneously reflects the support distribution and relative magnitudes of different ecological states, providing a basis for subsequent identification of key ecological constraint factors and ecological restoration scenario simulation.

[0095] It should be noted that by simultaneously constructing visual and textual evidence bodies and utilizing Dempster-Shafer theory for evidence fusion, the judgment results from synchronous ecological observation data and textual semantic information can be integrated within a unified evidence framework. Furthermore, the evidence conflict coefficient is used to measure and normalize conflict information between different modalities, effectively reducing the impact of single-modal errors or anomalous data on the final judgment result. This approach not only improves the stability and reliability of ecological state identification results but also preserves the uncertainty information between different modal evidence. The resulting multimodal ecological state evidence body contains multi-source information support and possesses good interpretability, thus providing more reliable data for the subsequent identification of key ecological constraints and the formulation of ecological restoration strategies.

[0096] S6: Based on the basic probability distribution of each ecological proposition in the multimodal ecological state evidence body, identify multiple key ecological constraint factors and uncertainty intervals in the target ecological region.

[0097] Ecological propositions refer to hypotheses or judgments describing the state of an ecosystem, such as "ecosystem health," "mild ecological degradation," or "severe ecological degradation," used to represent different categories of ecological states. Basic probability assignment, in evidence theory, is a numerical value representing the degree to which a particular piece of evidence supports a certain proposition; it reflects the strength of support for that ecological state from current multimodal evidence. Key ecological constraints are ecological factors that play a major limiting role in changes to the ecological state during ecosystem evolution or restoration, such as insufficient soil moisture, declining vegetation cover, water body shrinkage, or the risk of land erosion. Uncertainty intervals refer to the possible value range of ecological constraint factors derived from the degree of evidence support and uncertainty, used to represent the range of variation and cognitive uncertainty of that ecological factor in the current state.

[0098] It should be noted that, compared with traditional methods relying on single indicators or empirical judgments, this step utilizes multimodal evidence fusion results for constraint factor identification. This not only comprehensively considers the degree of support for the ecological state from different data sources but also reduces the impact of errors from a single data source, thereby improving the reliability of key ecological factor identification. Simultaneously, by constructing uncertainty intervals for ecological constraint factors, the uncertainty information in the ecosystem state can be explicitly expressed. This ensures that the ecological assessment results are no longer limited to single values ​​but reflect the possible range of changes in the ecological state in the form of intervals. This provides a more realistic, stable, and risk-assessable data foundation for subsequent ecological vulnerability analysis and the formulation of ecological restoration strategies.

[0099] In one possible implementation, S6 specifically includes:

[0100] S601: Based on the basic probability distribution of each ecological proposition in the multimodal ecological state evidence body, calculate the initial evidence contribution value of different ecological propositions to each ecological constraint factor.

[0101] The initial evidence contribution value refers to the comprehensive contribution value obtained after the evidence support of each ecological proposition is applied to the corresponding ecological constraint factor according to the preset correlation weight, which is used to measure the degree to which the factor is supported by evidence.

[0102] Specifically, firstly, basic probability assignment values ​​for all ecological propositions are obtained from the multimodal ecological state evidence body, where each ecological proposition corresponds to a support level obtained from evidence fusion. Then, the association weights between each ecological proposition and each ecological constraint factor are pre-determined; these association weights characterize the influence of the q-th ecological proposition on the r-th ecological constraint factor. Next, for the r-th ecological constraint factor, the association weights corresponding to all Q ecological propositions are sequentially multiplied by the basic probability assignment values ​​to obtain the individual evidence contribution value of each ecological proposition to that ecological constraint factor. Finally, these Q individual evidence contribution values ​​are summed to obtain the initial evidence contribution value of the r-th ecological constraint factor.

[0103] For example, when the three ecological propositions of "water body shrinkage", "vegetation degradation" and "soil exposure" all have a high correlation weight with "water stress factor", the initial evidence contribution value of the ecological constraint factor of this factor will be large, indicating that it is more likely to constitute the key constraint of current ecological restoration.

[0104] S602: Based on the initial evidence contribution value and the coupling relationship between various ecological constraint factors, calculate the constraint strength after multiple coupling amplifications:

[0105]

[0106] Among them, F r This represents the coupling amplification constraint strength of the r-th ecological constraint factor. Let represent the initial evidence contribution value of the r-th ecological constraint factor. Represents the coupling propagation coefficient. Let represent the coupling weight of the s-th ecological constraint factor with respect to the r-th ecological constraint factor, and R represent the total number of ecological constraint factors. This represents the initial evidence contribution value of the s-th ecological constraint factor.

[0107] The coupling relationship refers to the mutual influence or reinforcement between different ecological constraint factors. The constraint strength after coupling amplification refers to the comprehensive strength of the constraint factors after considering the linkage and propagation between factors.

[0108] Specifically, this formula does not only consider the strength of evidence for a single factor, but also takes into account the amplifying effect of the interaction between factors. For example, "insufficient soil moisture content" exacerbates "limited vegetation growth," and "limited vegetation growth" further exacerbates "risk of surface erosion." Therefore, the coupling between these factors can cause the final strength of certain constraint factors to be higher than their individual evidence values. This approach draws on the concept of coupling propagation in complex networks and fault propagation analysis, and can reflect the true characteristics of "multi-factor interconnected constraints" in ecosystems.

[0109] S603: Based on the constraint strength after each coupling amplification, determine several key ecological constraint factors.

[0110] Specifically, when identifying key ecological constraint factors, the normalized discriminant coefficients of all ecological constraint factors are first calculated, then they are sorted according to their numerical values, and the sorting results are compared with preset discrimination criteria. If the normalized discriminant coefficient of a certain ecological constraint factor is greater than or equal to a preset threshold, or ranks among the top few in the sorting, then that ecological constraint factor is identified as a key ecological constraint factor. In other words, by converting the absolute intensity of each constraint factor into a relative proportion, the factors that dominate the overall ecological constraints are screened out, thereby avoiding misjudgments caused by differences in dimensions or local numerical fluctuations, making the identification results of key ecological constraint factors more stable and comparable.

[0111] S604: Based on the evidence support and uncertainty of each key ecological constraint factor, identify the uncertainty interval:

[0112]

[0113] in, Let r represent the uncertainty interval of the r-th key ecological constraint factor. This represents the central estimate of the r-th key ecological constraint factor. Indicates the interval scaling factor. This represents the trust level of the r-th key ecological constraint factor. This represents the likelihood of the r-th key ecological constraint factor.

[0114] Specifically, the single-point estimates of key ecological constraint factors are extended into an interval representation with evidence uncertainty, thereby more realistically describing the range of states that the factor may currently be in.

[0115] It should be noted that by introducing a coupling propagation mechanism among ecological constraint factors, the initial evidence contribution values ​​of each factor are amplified through a linked calculation. This allows the model to consider not only the evidence strength of a single factor but also the true relationships of mutual influence and chain amplification among multiple factors in the ecosystem, thus better reflecting the complexity of actual ecosystems. Simultaneously, by normalizing and ranking the constraint strengths after coupling amplification and screening key ecological constraint factors, the dominant limiting factors in the current ecological environment can be effectively identified, improving the accuracy of ecological problem diagnosis. Furthermore, by constructing uncertainty intervals based on confidence and likelihood, the single estimate of ecological factors can be extended to an interval expression containing uncertainty ranges, thus more realistically reflecting the volatility and cognitive uncertainty of the ecosystem state, providing a more robust and reliable analytical foundation for subsequent ecological vulnerability analysis and ecological restoration decisions.

[0116] S7: Construct an ecological vulnerability distribution map based on the coupling relationship between key ecological constraint factors.

[0117] Coupling relationships refer to the interactions or mutual influences between different ecological constraints. For example, a change in one ecological factor may cause a change in another, thus creating a synergistic effect. An ecological vulnerability distribution map is a spatial distribution map that spatially represents and visualizes the vulnerability of an ecosystem at different locations. This map reflects the degree of ecological risk and its differences at different locations within a target ecological region.

[0118] It should be noted that, compared with traditional methods that rely on a single indicator to assess ecological status, this step can reveal the spatial superposition and propagation effects of different ecological factors, thereby improving the accuracy of ecological risk identification. Simultaneously, the ecological vulnerability distribution map can provide clear spatial guidance for the subsequent development of ecological restoration strategies, enabling restoration measures to be prioritized in areas with higher ecological vulnerability or stronger ecological constraints, thus improving the targeting and resource utilization efficiency of ecological restoration projects.

[0119] In one possible implementation, S7 specifically includes:

[0120] S701: Based on the coupling relationship between various key ecological constraint factors, construct the ecological constraint factor coupling propagation matrix.

[0121] Among them, the ecological constraint factor coupling propagation matrix refers to the matrix structure used to describe the mutual influence relationship between different key ecological constraint factors.

[0122] Specifically, the "propagation impact of all key ecological constraint factors on a certain target factor" is uniformly normalized to construct an ecological constraint factor coupling propagation matrix. Specifically, for each element in the matrix... The value represents the relative contribution of the nth key ecological constraint factor to the vulnerability propagation of the rth key ecological constraint factor. If a factor has a large coupling weight and a high normalized discriminant coefficient, its corresponding propagation coefficient will be larger, indicating that the factor has a stronger influence on the target factor during vulnerability propagation. By repeating the above calculation for all r and s, a complete ecological constraint factor coupling propagation matrix can be formed, which is used to describe the vulnerability transmission relationship between key ecological constraint factors.

[0123] S702: Combining the constraint strengths and uncertainty intervals after various coupling amplifications, calculate the initial vulnerability source strengths of each key ecological constraint factor:

[0124]

[0125] in, This represents the initial vulnerability source strength of the r-th key ecological constraint factor. This represents the coupling amplification constraint strength of the r-th key ecological constraint factor. This represents the uncertainty amplification factor. This represents the upper bound of the uncertainty interval for the r-th key ecological constraint factor. This represents the lower bound of the uncertainty interval for the r-th key ecological constraint factor. This represents the central estimate of the r-th key ecological constraint factor. This represents a small positive number that prevents the denominator from being zero.

[0126] The initial vulnerability source strength refers to the vulnerability starting value calculated based on the constraint strength and uncertainty range of key ecological constraint factors, which is used as the initial input of the vulnerability propagation model.

[0127] Specifically, if a key ecological constraint factor has a high constraint strength and a wide uncertainty range, it indicates that the factor not only has a strong limiting effect but also higher state fluctuations or cognitive uncertainty. In this case, the initial vulnerability source strength will be further amplified. Conversely, if a factor has a certain constraint strength but low uncertainty, its initial vulnerability source strength is mainly determined by the basic constraint level. This formula can integrate "constraint strength" and "uncertainty risk" into the vulnerability source strength modeling process, providing a more reasonable initial input for subsequent ecological vulnerability propagation analysis.

[0128] S703: Using the initial vulnerability source strength as the initial state for propagation, iterative diffusion is performed on the ecological constraint factor coupling propagation matrix:

[0129]

[0130] in, This represents the vulnerability value of the r-th key ecological constraint factor at the (t+1)-th iteration. Indicates the diffusion coefficient. This represents the initial vulnerability source strength of the r-th key ecological constraint factor. This represents the normalized propagation coefficient of the vulnerability impact from the s-th key ecological constraint factor to the r-th key ecological constraint factor. Let R represent the vulnerability value of the s-th critical ecological constraint factor at the t-th iteration, and let R represent the total number of critical ecological constraint factors.

[0131] Iterative diffusion refers to the process of simulating the multi-factor linkage propagation by gradually propagating and accumulating the vulnerability effects of each factor through multiple rounds of calculation under the action of the coupling propagation matrix.

[0132] Specifically, starting with the initial vulnerability source strength of each key ecological constraint factor, and combining the coupling and propagation relationships between these factors, the diffusion and accumulation process of each factor's vulnerability in the network is iteratively calculated to obtain a vulnerability distribution result that better reflects the multi-factor linkage effect. Specifically, in each iteration, a portion of the vulnerability value comes from the initial vulnerability retained by the factor itself, while another portion comes from the influence transmitted from other key ecological constraint factors through the coupling propagation matrix. As the iteration progresses, the vulnerability of each key ecological constraint factor gradually integrates its own state with the network linkage effect, eventually converging to obtain a stable factor vulnerability result, which can be used to subsequently construct an ecological vulnerability distribution map.

[0133] S704: Repeat step S703 until the convergence condition is met, then output the steady-state factor vulnerability vector.

[0134] Among them, the stable factor vulnerability vector refers to the set of stable vulnerability values ​​of each key ecological constraint factor obtained after the iterative calculation converges.

[0135] Specifically, the convergence condition is that when the L2 norm of the difference between the vulnerability vector obtained in the current iteration and the vulnerability vector obtained in the previous iteration is less than or equal to a preset threshold, it indicates that the overall change between the results of the two iterations is small enough, the vulnerability values ​​of each key ecological constraint factor have basically stabilized, and no more significant fluctuations occur. At this point, it can be determined that the iteration process has reached a convergence state, and the current vulnerability vector is output as the final steady-state vulnerability result. When the L2 norm is still greater than the preset threshold, it indicates that the vulnerability propagation in the system has not yet stabilized, and the next round of iteration calculation needs to be continued.

[0136] S705: Based on the sensitive response relationship of each spatial unit in the target ecological region to each key ecological constraint factor, the vulnerability vector of the steady-state factor is mapped to the spatial unit to obtain the ecological vulnerability distribution map.

[0137] Specifically, the target ecological region is first divided into multiple spatial units, and a sensitivity response relationship to each key ecological constraint factor is pre-established for each spatial unit. This sensitivity response relationship characterizes the degree to which a spatial unit is affected by the corresponding key ecological constraint factor. Then, the steady-state factor vulnerability vector obtained through the aforementioned iterative propagation is read, and the steady-state vulnerability values ​​of each key ecological constraint factor are weighted and allocated according to the sensitivity response weight of the corresponding spatial unit, so that each spatial unit receives vulnerability contributions from each key ecological constraint factor. Next, all vulnerability contributions received by the same spatial unit are summed to obtain the comprehensive ecological vulnerability value of that spatial unit. Finally, the comprehensive ecological vulnerability values ​​of each spatial unit are spatially correlated and visualized according to their geographical location, thereby forming an ecological vulnerability distribution map reflecting the differences in vulnerability levels at different locations within the target ecological region.

[0138] It should be noted that by constructing a coupled propagation matrix of ecological constraint factors and combining iterative diffusion calculations with the initial vulnerability source strength, ecological vulnerability assessment can simultaneously consider the interactions between multiple key ecological factors, thus more realistically reflecting the propagation process of multi-factor linkage effects in the ecosystem. By mapping the vulnerability vectors of steady-state factors to various spatial units and forming an ecological vulnerability distribution map, ecological risks can be visually displayed in spatial form. This facilitates the identification of areas with high ecological vulnerability and provides a more scientific and accurate decision-making basis for the spatial layout and resource allocation of subsequent ecological restoration measures.

[0139] S8: Combining key ecological constraint factors, uncertainty intervals, and ecological vulnerability distribution maps, dynamic Monte Carlo simulations are performed in the ecological restoration measures space to generate a multi-branch ecological restoration scenario path set containing probability weights.

[0140] The ecological restoration measure space refers to the set of various restoration measures that can be selected during the ecological restoration process, such as vegetation restoration, soil and water conservation projects, wetland restoration, or water replenishment, forming a decision space. Dynamic Monte Carlo simulation refers to the simulation and analysis of the evolutionary process of an ecosystem under different combinations of restoration measures through random sampling and multiple iterative calculations, thereby generating multiple possible restoration evolution outcomes. The multi-branch ecological restoration scenario path set refers to the ecological evolution process paths corresponding to different ecological restoration measure implementation sequences formed during the simulation process.

[0141] It should be noted that by comprehensively incorporating key ecological constraints, uncertainty intervals, and ecological vulnerability distribution maps into the analysis, and conducting dynamic Monte Carlo simulations within the ecological restoration measure space, the formulation of ecological restoration plans can be transformed from a traditional single fixed strategy into a probabilistic decision-making process based on multi-scenario simulations. Compared with traditional methods that rely on experience to select restoration measures, this step not only considers the uncertainty and spatial differences of the ecosystem but also explores multiple potential restoration pathways, thereby providing a more flexible, reliable, and risk-assessment-capable decision-making basis for subsequent ecological restoration projects.

[0142] In one possible implementation, S8 specifically includes:

[0143] S801: Construct a stochastic initial state vector for an ecological restoration scenario based on an uncertainty interval.

[0144] Among them, the random initial state vector refers to the initial state combination vector obtained by random sampling based on the uncertainty range of each key ecological constraint factor at the beginning of the ecological restoration scenario simulation. It is used to describe the state distribution of each ecological constraint factor in the target ecological area at the beginning of a certain scenario path.

[0145] Specifically, firstly, for each key ecological constraint factor, its corresponding upper and lower bounds of uncertainty interval are read, and this uncertainty interval is taken as the possible value range of the key ecological constraint factor at the initial moment of the ecological restoration scenario. Then, within each spatial unit, each key ecological constraint factor is sampled within an interval according to a preset random distribution to obtain the random initial state value of each key ecological constraint factor in that spatial unit under the current scenario. Based on this, the random initial state values ​​of all key ecological constraint factors in the same spatial unit are combined according to the factor number order to form the random initial state sub-vector of the corresponding spatial unit. Furthermore, the random initial state sub-vectors of each spatial unit in the target ecological area are concatenated in spatial order to finally construct the random initial state vector for the current ecological restoration scenario.

[0146] S802: Combining the random initial state vector, construct the dynamic evolution equation of constraint factor-restoration measure in the ecological restoration measure space:

[0147]

[0148] in, This represents the state value of the l-th ecological restoration scenario path at time t+1, in the n-th spatial unit, and the n-th key ecological constraint factor. This represents the state propagation adjustment coefficient. This represents the state value of the l-th ecological restoration scenario path at time t, in the n-th spatial unit, and the n-th key ecological constraint factor. This represents the state value of the s-th key ecological constraint factor in the n-th spatial unit at time t of the l-th ecological restoration scenario path. This represents the action variable indicating whether the j-th type of ecological restoration measure is implemented in the n-th spatial unit at time t for the l-th ecological restoration scenario path. Let represent the reduction efficiency coefficient of the j-th type of ecological restoration measures on the r-th key ecological constraint factor, and v represent the resource consumption attenuation coefficient. Let J represent the resource consumption intensity of the j-th type of ecological restoration measure, and J represent the total number of ecological restoration measures.

[0149] Specifically, given a specific ecological restoration scenario path, the states of key ecological constraint factors in each spatial unit are dynamically updated. Specifically, the state of a factor at the next moment is determined by three parts: one part is the continuation of the factor's state from the previous moment; another part is the synergistic effect generated by the coupling propagation of other key ecological constraint factors; and the last part is the offsetting effect of the various ecological restoration measures currently implemented on the factor. In this way, the formula can simultaneously reflect the interaction relationships between ecological constraint factors, the dynamic evolution process within spatial units, and the state changes after restoration intervention, thus providing a basis for state updates for subsequent dynamic Monte Carlo simulations and the generation of multi-branch ecological restoration scenario paths.

[0150] S803: Based on the multimodal evidence support, spatial vulnerability state, and the constraint factor-remediation dynamic evolution equation, calculate the conditional transition probability of each remediation measure:

[0151]

[0152] in, Let represent the conditional probability of selecting the _i_th type of ecological restoration measure in the nth spatial unit at time t, for the l-th ecological restoration scenario path. Indicates temperature parameter, Let represent the adaptation weight of the nth type of ecological restoration measure to the rth key ecological constraint factor. Let ln() represent the evidence gain coefficient, and ln() represent the logarithmic function. This represents the ecological proposition corresponding to the Xth type of ecological restoration measures. The basic probability allocation value, Indicates the resource penalty coefficient. This represents the adaptation weight of the h-th type of ecological restoration measure to the r-th key ecological constraint factor. This represents the ecological proposition corresponding to the h-th type of ecological restoration measure. The basic probability allocation value, This indicates the resource consumption intensity of the h-th type of ecological restoration measures.

[0153] Specifically, under a given scenario path, time step, and spatial unit, the conditional probability of selecting each ecological restoration measure is calculated by comprehensively considering the current state of key ecological constraints, the suitability of restoration measures to each constraint, the strength of support for the relevant ecological proposition from multimodal evidence, and the resource consumption required to implement the restoration measure. In other words, this formula does not simply select a fixed optimal measure, but first calculates a comprehensive score for each candidate restoration measure, and then transforms it into a probability distribution through softmax normalization. This achieves probabilistic measure selection with prior evidence and cost constraints, providing a foundation for subsequent dynamic Monte Carlo sampling to generate multi-branch ecological restoration scenario paths.

[0154] Specifically, the conditional transition probability simultaneously integrates multimodal evidence support, spatial vulnerability status, and resource constraints for remediation measures.

[0155] S804: Dynamic Monte Carlo sampling is performed based on conditional transition probabilities to determine multi-branch ecological restoration scenario paths.

[0156] Specifically, at each simulation moment and within each spatial unit, the conditional transition probabilities corresponding to each candidate ecological restoration measure are first calculated according to the aforementioned formula, forming a normalized probability distribution for all candidate ecological restoration measures. Then, using this probability distribution as the sampling basis, an ecological restoration measure is randomly selected for the spatial unit at the current moment as the actual execution action for this round of scenario evolution, and the selected restoration measure is applied to the state update process of the corresponding key ecological constraint factor. Next, the process of "probability calculation - random sampling - state update" is continuously repeated according to time steps, so that each spatial unit gradually forms a complete restoration action evolution sequence at continuous moments. Finally, the above random sampling process is repeated multiple times, and each sampling run yields a different ecological restoration scenario path, thereby forming a multi-branch ecological restoration scenario path containing multiple possible evolution results.

[0157] S805: Calculate the probability weights of each branch of the ecological restoration scenario path and generate a multi-branch ecological restoration scenario path set containing the probability weights.

[0158] Specifically, firstly, for each branch ecological restoration scenario path, the state evolution results of each spatial unit and each key ecological constraint factor, the sequence of ecological restoration measures taken, and the corresponding resource consumption are statistically analyzed throughout the entire simulation period. Based on this, a comprehensive scenario score is calculated for the path. The comprehensive scenario score reflects at least the path's reduction effect on key ecological constraint factors, the degree of improvement to the ecological vulnerability distribution map, and the resource cost required to implement the path. Then, the comprehensive scenario scores of each branch ecological restoration scenario path are normalized and converted into corresponding probability weights, so that branch ecological restoration scenario paths with higher comprehensive scenario scores correspond to larger probability weights. Finally, each branch ecological restoration scenario path and its corresponding probability weight are associated and stored, generating a multi-branch ecological restoration scenario path set containing probability weights, so that subsequent screening, ranking, or hierarchical decision-making based on probability weights can be performed on each branch ecological restoration scenario path.

[0159] It should be noted that by establishing a dynamic evolution equation for constraint factors and restoration measures, the coupling and propagation relationships between ecological constraint factors and the intervention effects of ecological restoration measures are uniformly incorporated into the dynamic model. This allows the evolution of the ecological state to simultaneously reflect the combined impact of natural system changes and human restoration activities. Furthermore, by calculating the conditional transition probabilities of each restoration measure, the selection of restoration measures not only depends on the current ecological state but also comprehensively considers multimodal evidence support and resource consumption constraints, thereby achieving a more rational probabilistic decision-making mechanism. By generating multi-branch ecological restoration scenario paths through dynamic Monte Carlo sampling and comprehensively evaluating and assigning probability weights to each path, more feasible and effective solutions can be identified from numerous possible restoration strategies. This transforms ecological restoration decision-making from the traditional single-solution selection process to a multi-scenario assessment and probability ranking process, significantly improving the scientific rigor, flexibility, and reliability of ecological restoration planning.

[0160] S9: Generate a spatiotemporal control script for the execution of ecological restoration projects based on the multi-branch ecological restoration scenario path set.

[0161] Among them, the spatiotemporal control script refers to the set of execution instructions formed by structuring the restoration measures in the ecological restoration scenario path according to the time sequence and spatial location. The script usually contains information such as restoration measures to be implemented in different spatial areas at different time stages, resource allocation plans, and stage switching conditions, which are used to guide the specific implementation process of ecological restoration projects.

[0162] In one possible implementation, the spatiotemporal control script includes a hierarchical repair action sequence, a dynamic resource deployment strategy, and stage switching trigger conditions.

[0163] It should be noted that, compared with the traditional method that only provides macro-level restoration strategies, this step can transform complex simulation results into structured execution instructions, enabling restoration measures to be rationally scheduled in both time and space dimensions. At the same time, it facilitates dynamic adjustments based on actual ecological changes, thereby improving the efficiency, controllability, and scientific nature of the ecological restoration project implementation process.

[0164] Reference manual attached Figure 2 The diagram shows a schematic representation of an ecological restoration system based on multimodal data fusion provided by an embodiment of the present invention.

[0165] This invention provides an ecological restoration system 20 based on multimodal data fusion, comprising: a processor 201 and a memory 202;

[0166] The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the above-mentioned ecological restoration method based on multimodal data fusion and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0167] It should be understood that the processor 201 in this embodiment of the invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0168] It should also be understood that the memory 202 in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM).

[0169] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0170] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0171] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0172] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0173] In the embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0174] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0175] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0176] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0177] This invention provides a readable storage medium that stores a program or instructions on the medium. When the program or instructions are executed by a processor, they implement the steps of the above-described ecological restoration method based on multimodal data fusion and achieve the same technical effect. To avoid repetition, this invention will not elaborate further.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the protection scope of the present invention.

Claims

1. An ecological restoration method based on multimodal data fusion, characterized in that, include: S1: Collect multimodal ecological data of the target ecological area, wherein the multimodal ecological data includes synchronous ecological data and asynchronous text data; S2: Using a pre-trained single-modal encoder, feature extraction is performed on the synchronous ecological data and the asynchronous text data respectively to obtain synchronous modal feature data and text feature vectors; S3: Using the marginal Fisher analysis method, feature-level fusion is performed on the synchronous modal feature data to determine the synchronous fusion features; S4: The synchronous fusion features are interpolated and enhanced using a data augmentation method based on shared mixing coefficients to generate enhanced fusion features; S5: Perform decision-level evidence fusion on the enhanced fusion features and the text feature vector to obtain a multimodal ecological state evidence body; S6: Based on the basic probability allocation of each ecological proposition in the multimodal ecological state evidence body, identify multiple key ecological constraint factors and uncertainty intervals in the target ecological region; S7: Based on the coupling relationship between the key ecological constraint factors, construct an ecological vulnerability distribution map; S8: Combining the key ecological constraint factors, the uncertainty interval, and the ecological vulnerability distribution map, perform dynamic Monte Carlo simulation in the ecological restoration measures space to generate a multi-branch ecological restoration scenario path set containing probability weights; S9: Generate a spatiotemporal control script for the execution of ecological restoration projects based on the multi-branch ecological restoration scenario path set.

2. The ecological restoration method based on multimodal data fusion according to claim 1, characterized in that, The synchronized ecological data includes visual image data, environmental sensor time-series data, and audio data.

3. The ecological restoration method based on multimodal data fusion according to claim 1, characterized in that, S3 specifically includes: S301: Construct a multimodal feature matrix based on the synchronous modal feature data; S302: Calculate the covariance matrix between each modality feature in the multimodal feature matrix; S303: Using the marginal Fisher analysis method, each modal feature is mapped to a shared latent space; S304: Determine the fusion shared features based on the covariance matrix and mapping results; S305: Normalize the fusion-shared features to determine the synchronous fusion features.

4. The ecological restoration method based on multimodal data fusion according to claim 3, characterized in that, S4 specifically includes: S401: Construct a batch of synchronous fusion features based on the aforementioned synchronous fusion features; S402: Perform pairwise sampling on the batch of synchronous fusion features to determine multiple pairs of synchronous fusion features; S403: Define the shared mixing coefficient; S404: Within the shared potential space, combining the shared mixing coefficients and the multiple sets of synchronous fusion feature pairs, perform shared interpolation on the synchronous fusion features to determine the enhanced features; S405: Normalize the enhanced features to generate the enhanced fusion features.

5. The ecological restoration method based on multimodal data fusion according to claim 1, characterized in that, S5 specifically includes: S501: Calculate the similarity between the enhanced fusion features and the prototype vectors of each ecological state; S502: Based on the aforementioned similarity, and in conjunction with the softmax function, calculate the probability distribution of the first ecological state; S503: Calculate the probability distribution of the second ecological state based on the text feature vector; S504: Construct visual evidence and textual evidence based on the first ecological state probability distribution and the second ecological state probability distribution; S505: Calculate the evidence conflict coefficient based on the visual evidence body and the text evidence body; S506: By using Dempster-Shafer theory and combining the evidence conflict coefficient, the visual evidence body and the textual evidence body are fused to obtain the quality of fused evidence; S507: Construct the multimodal ecological state evidence body based on the quality of the fused evidence.

6. The ecological restoration method based on multimodal data fusion according to claim 1, characterized in that, S6 specifically includes: S601: Based on the basic probability distribution of each ecological proposition in the multimodal ecological state evidence body, calculate the initial evidence contribution value of different ecological propositions to each ecological constraint factor. S602: Based on the initial evidence contribution value, and combined with the coupling relationship between the various ecological constraint factors, calculate the constraint strength after multiple coupling amplifications; S603: Determine multiple key ecological constraint factors based on the constraint strength after each coupling amplification; S604: Identify the uncertainty interval based on the evidence support and evidence uncertainty of each of the key ecological constraint factors.

7. The ecological restoration method based on multimodal data fusion according to claim 6, characterized in that, Specifically, S7 includes: S701: Based on the coupling relationship between the key ecological constraint factors, construct the ecological constraint factor coupling propagation matrix; S702: Combine the constraint strengths after each coupling amplification with the uncertainty interval to calculate the initial vulnerability source strength of each of the key ecological constraint factors; S703: Using the initial vulnerability source strength as the initial propagation state, iteratively diffuse the propagation on the ecological constraint factor coupling propagation matrix; S704: Repeat step S703 until the convergence condition is met, then output the steady-state factor vulnerability vector. S705: Based on the sensitive response relationship of each spatial unit in the target ecological region to each of the key ecological constraint factors, the vulnerability vector of the steady-state factor is mapped to the spatial unit to obtain the ecological vulnerability distribution map.

8. The ecological restoration method based on multimodal data fusion according to claim 1, characterized in that, S8 specifically includes: S801: Based on the aforementioned uncertainty interval, construct a random initial state vector for the ecological restoration scenario; S802: Combining the random initial state vector, construct a dynamic evolution equation of constraint factor-restoration measures in the space of ecological restoration measures; S803: Calculate the conditional transition probability of each remediation measure based on the multimodal evidence support, spatial vulnerability state, and the dynamic evolution equation of the constraint factor-remediation measure; S804: Perform dynamic Monte Carlo sampling based on the conditional transition probability to determine the multi-branch ecological restoration scenario path; S805: Calculate the probability weights of each of the branch ecological restoration scenario paths, and generate a multi-branch ecological restoration scenario path set containing the probability weights.

9. The ecological restoration method based on multimodal data fusion according to claim 1, characterized in that, The spatiotemporal control script includes a hierarchical repair action sequence, a dynamic resource deployment strategy, and stage switching trigger conditions.

10. An ecological restoration system based on multimodal data fusion, characterized in that, include: Processor and memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the ecological restoration method based on multimodal data fusion as described in any one of claims 1 to 9.

Images