Ecological restoration effect intelligent evaluation analysis method, system, device and medium

Abstract

The application relates to an ecological restoration effect intelligent evaluation and analysis method, system, device and medium. The method comprises the following steps: inputting each space-time coordinate in a standardized sparse observation expression set into a space-time manifold mapping network to obtain a manifold coordinate; inputting the manifold coordinate into an ecological state decoder to obtain a space-time continuous ecological state field; under the time causal loop constraint, analyzing and processing the causal relationship between the processing variable field, the outcome variable and the covariate field, obtaining a causal diagram through a hybrid algorithm based on constraints and scores; based on the causal diagram, inputting the processing variable field, the outcome variable and the covariate field into a multi-modal structural equation model, and performing counterfactual inference by applying intervention and fixing the invariable exogenous noise in the multi-modal structural equation model, and quantitatively obtaining the causal effect value of each ecological restoration measure on the ecological effect. The method can deconstruct the independent contribution of the restoration measure to the response of the ecological system.

Description

Technical Field

[0001] This invention belongs to the field of ecological restoration assessment technology, and in particular relates to an intelligent assessment and analysis method, system, equipment and medium for ecological restoration effectiveness. Background Technology

[0002] With the development of technologies in the fields of ecological monitoring and ecosystem restoration, multi-source ecological monitoring technologies have emerged. By acquiring ecological data with different dimensions and spatiotemporal characteristics, these technologies provide multi-dimensional data support for evaluating the effectiveness of ecosystem restoration.

[0003] In traditional technologies, processing multi-source ecological monitoring data mainly targets the spatiotemporal differences of different monitoring data. Spatiotemporal interpolation or resampling is used to forcibly align various types of data onto a unified spatiotemporal grid. At the level of comprehensive data analysis, comprehensive index method or Pearson correlation analysis is mainly used to quantify the statistical correlation strength between different ecological indicators.

[0004] However, the smoothing errors introduced by the interpolation algorithms described above can mask the true local heterogeneity, and the information loss caused by resampling can distort key temporal phase relationships, thereby destroying the inherent structural consistency between data. In addition, the analysis of the effectiveness of ecosystem restoration has long been stuck at the correlation description stage, unable to isolate the interference of external confounding factors such as interannual climate fluctuations and regional policy changes, and unable to clarify the complex chain response pathways within the ecosystem. It is difficult to reveal the causal relationship between restoration measures and ecological responses, resulting in the inability to determine the true effectiveness of restoration measures and to provide reliable guidance for restoration practices. Summary of the Invention

[0005] Therefore, it is necessary to provide an intelligent assessment and analysis method, system, equipment, and medium for ecological restoration effectiveness that can quantitatively deconstruct and accurately assess the causal relationship between specific restoration measures and ecosystem responses, addressing the aforementioned technical problems.

[0006] Firstly, this application provides an intelligent evaluation and analysis method for the effectiveness of ecological restoration, including:

[0007] Acquire multi-source heterogeneous ecological monitoring data covering the area to be evaluated, and discretize the multi-source heterogeneous ecological monitoring data according to the data source type to obtain a standardized sparse observation expression set; the standardized sparse observation expression set includes spatiotemporal coordinates and the observation value vectors corresponding to the spatiotemporal coordinates;

[0008] The spatiotemporal coordinates of each data point in the standardized sparse observation representation set are input into the spatiotemporal manifold mapping network and forward propagated to obtain the manifold coordinates of each data point in the observation vector in a unified high-dimensional manifold space. The spatiotemporal manifold mapping network includes a spatial coding module, a temporal coding module, and a gated fusion layer.

[0009] The manifold coordinates are input into the ecological state decoder to obtain a multidimensional ecological state vector, and a spatiotemporally continuous ecological state field is constructed based on the multidimensional ecological state vector.

[0010] A processing variable field was constructed based on the implementation records of ecological restoration measures, and outcome variables and covariates were extracted from the spatiotemporal continuous ecological state field.

[0011] Using a hybrid algorithm based on constraints and fractions, under the constraint of time-causal acyclicity, the causal relationships between the variable field, outcome variable and covariate are analyzed and processed to obtain the causal graph;

[0012] Based on the causal graph, the processing variable field, outcome variable, and covariate field are input into the multimodal structural equation model. By applying interventions and fixing exogenous noise invariantly in the multimodal structural equation model, counterfactual inferences are performed to quantify the causal effect values ​​of various ecological restoration measures on ecological effectiveness. The multimodal structural equation model is composed of multiple deep network sub-modules connected according to the parent-child dependency relationship in the causal graph.

[0013] Secondly, this application also provides an intelligent evaluation and analysis system for the effectiveness of ecological restoration, including:

[0014] The data acquisition module is used to acquire multi-source heterogeneous ecological monitoring data covering the area to be evaluated, and to discretize the multi-source heterogeneous ecological monitoring data according to the data source type to obtain a standardized sparse observation expression set; the standardized sparse observation expression set includes spatiotemporal coordinates and the observation value vectors corresponding to the spatiotemporal coordinates;

[0015] The manifold coding module is used to input the spatiotemporal coordinates of each data point in the standardized sparse observation representation set into the spatiotemporal manifold mapping network for forward propagation, and obtain the manifold coordinates of each data point in the observation vector in the unified high-dimensional manifold space; the spatiotemporal manifold mapping network includes a spatial coding module, a temporal coding module and a gated fusion layer;

[0016] The ecological state module is used to input manifold coordinates into the ecological state decoder to obtain a multidimensional ecological state vector, and to construct a spatiotemporally continuous ecological state field based on the multidimensional ecological state vector.

[0017] The restoration module is used to construct a processing variable field based on the implementation records of ecological restoration measures, and to extract outcome variables and covariates from the spatiotemporal continuous ecological state field.

[0018] The causal analysis module is used to analyze and process the pairwise causal relationships between variable fields, outcome variables, and covariates under time-causal acyclic constraints using a hybrid algorithm based on constraints and scores, and to obtain a causal graph.

[0019] The effectiveness attribution assessment module is used to input the treatment variable field, outcome variable, and covariate field into the multimodal structural equation model based on the causal graph. By applying interventions and fixing exogenous noise invariantly in the multimodal structural equation model, counterfactual inference is performed to quantify the causal effect value of each ecological restoration measure on the ecological effectiveness. The multimodal structural equation model is composed of multiple deep network sub-modules connected according to the parent-child dependency relationship in the causal graph.

[0020] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned intelligent evaluation and analysis methods for ecological restoration effectiveness.

[0021] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above-mentioned intelligent evaluation and analysis methods for ecological restoration effectiveness.

[0022] The aforementioned intelligent assessment and analysis methods, systems, equipment, and media for ecological restoration effectiveness discretize multi-source heterogeneous ecological monitoring data according to data source type, obtaining a standardized sparse observation representation set composed of spatiotemporal coordinates and observation value vectors. Through a spatiotemporal manifold mapping network including spatial coding modules, temporal coding modules, and a gated fusion layer, each spatiotemporal coordinate is mapped to a unified high-dimensional manifold space to generate manifold coordinates, achieving intrinsic structural alignment of multi-source heterogeneous data on a continuous manifold. The manifold coordinates are then input into an ecological state decoder to obtain a multidimensional ecological state vector, and a high-fidelity lookup function is constructed for any continuous spatiotemporal coordinates. This method queries a continuous spatiotemporal ecological state field, transforming sparse and asynchronous raw observations into a complete and physically consistent continuous ecological state expression. Based on records of ecological restoration measures, a processing variable field is constructed. Outcome variables and covariates are extracted from the ecological state field. A hybrid algorithm based on constraints and fractions automatically discovers the causal graph structure between variables under temporal causal acyclic constraints. Using the causal graph as a topological constraint, a multimodal structural equation model is constructed. Counterfactual inference is performed by applying interventions and fixing exogenous noise, thus removing the influence of confounding factors and quantifying the independent causal effects of each restoration measure on ecological outcomes. Through a progressive processing approach—from data manifold alignment to causal graph discovery and then to counterfactual inference—a paradigm shift from traditional correlation analysis to causal quantification is achieved. This allows for the accurate deconstruction of the independent contribution of each restoration measure to the ecosystem response, excluding interference from external factors such as climate fluctuations and regional policies. This provides a causal-level scientific decision-making basis for precise policy implementation and adaptive management of ecological restoration. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the intelligent evaluation and analysis method for ecological restoration effectiveness of the present invention.

[0025] Figure 2 This is a structural diagram of the intelligent evaluation and analysis system for ecological restoration effectiveness of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] In one embodiment, such as Figure 1 As shown, an intelligent evaluation and analysis method for ecological restoration effectiveness is provided. This embodiment illustrates the application of this method to an intelligent evaluation and analysis terminal for ecological restoration effectiveness. It is understood that this method can also be applied to an intelligent evaluation and analysis server for ecological restoration effectiveness, and further to an intelligent evaluation and analysis system for ecological restoration effectiveness that includes both an intelligent evaluation and analysis terminal and an intelligent evaluation and analysis server, and is implemented through the interaction between the two. In this embodiment, the method includes the following steps:

[0028] S101. Obtain multi-source heterogeneous ecological monitoring data covering the area to be evaluated, and discretize the multi-source heterogeneous ecological monitoring data according to the data source type to obtain a standardized sparse observation expression set; the standardized sparse observation expression set includes spatiotemporal coordinates and the observation value vectors corresponding to the spatiotemporal coordinates.

[0029] As an illustration, the intelligent assessment and analysis terminal for ecological restoration effectiveness can collect multi-source heterogeneous ecological monitoring data covering the entire area to be assessed. This multi-source heterogeneous ecological monitoring data can include monitoring data collected by ground sensor networks deployed based on the Internet of Things (IoT), remote sensing image data acquired by imaging equipment on aerospace platforms, and data from artificial plot surveys conducted in the field. Among these, IoT ground sensor network data can cover key ecological indicators such as soil temperature, soil moisture, and soil conductivity; aerospace remote sensing image data can include image reflectance data after radiometric calibration and atmospheric correction, as well as characteristic data such as the normalized vegetation index derived from these data; and artificial plot survey data can include core ecological observation indicators such as the number of plant and animal populations, species diversity, and vegetation biomass. The intelligent assessment and analysis terminal for ecological restoration effectiveness can discretize and standardize various multi-source heterogeneous ecological monitoring data according to the data source's collection characteristics and storage format. Each type of raw monitoring data is decomposed into two main components: spatiotemporal coordinates and corresponding observation value vectors. This constructs a standardized sparse observation representation set. Each data unit within the set is strictly matched with unique geographic location coordinates and timestamps forming spatiotemporal coordinates. The observation value vector fully carries all ecological monitoring indicator values ​​collected at the corresponding spatiotemporal location. This achieves unified and standardized representation of heterogeneous data with different spatiotemporal resolutions and physical attributes, avoiding the scale fragmentation problem caused by traditional data splicing methods. For example, during the discretization and standardization process, the intelligent assessment and analysis terminal for ecological restoration effectiveness uniformly sets the spatial grid resolution to a standard 30-meter raster scale, and the time resampling benchmark uses the natural day as the smallest time unit. All multi-source heterogeneous data are aligned according to this unified spatiotemporal benchmark.

[0030] S102. Input the spatiotemporal coordinates of each data point in the standardized sparse observation representation set into the spatiotemporal manifold mapping network and perform forward propagation to obtain the manifold coordinates of each data point in the observation vector in the unified high-dimensional manifold space. The spatiotemporal manifold mapping network includes a spatial coding module, a temporal coding module, and a gated fusion layer.

[0031] Furthermore, the intelligent assessment and analysis terminal for ecological restoration effectiveness can input all spatiotemporal coordinates from the standardized sparse observation set one by one into a pre-constructed spatiotemporal manifold mapping network for forward propagation. The spatiotemporal manifold mapping network is composed of a spatial encoding module, a temporal encoding module, and a gated fusion layer, which are coupled in an orderly manner. The network can achieve unified normalization processing of spatiotemporal scales of multi-source heterogeneous data and autonomously learn the complex spatiotemporal correlation structure and anisotropic change characteristics hidden within the ecosystem. The spatial encoding module can perform high-dimensional feature mapping on geospatial dimensional information, the temporal encoding module can embed and encode temporal periodic features, and the gated fusion layer can rely on its own activation and regulation logic to dynamically adjust the fusion ratio of spatial and temporal features according to the spatiotemporal evolution law of ecological processes, adapting to the heterogeneous laws of ecological state evolution under different seasons and terrains. The intelligent assessment and analysis terminal for ecological restoration effectiveness utilizes the layer-by-layer feature transformation and encoding mapping of the spatiotemporal manifold mapping network to map each original spatiotemporal coordinate to the interior of a high-dimensional manifold feature space, generating manifold coordinates specific to each monitoring data point. These manifold coordinates, as an abstract low-dimensional embedded representation, can accurately locate the position nodes of each observation sample on the overall ecological state manifold, achieving distortion-free alignment mapping of monitoring data from different dimensions to a unified manifold space.

[0032] S103. Input the manifold coordinates into the ecological state decoder to obtain a multidimensional ecological state vector, and construct a spatiotemporally continuous ecological state field based on the multidimensional ecological state vector.

[0033] Optionally, the intelligent assessment and analysis terminal for ecological restoration effectiveness can uniformly input all obtained manifold coordinates into a fully trained ecological state decoder. The ecological state decoder, built using a multi-task learning architecture, can parallelly perform fitting and prediction of multiple ecological state variables based on shared manifold coordinate features, thereby outputting a multi-dimensional ecological state vector including indicators such as soil moisture content, surface temperature, vegetation canopy height, and normalized difference in vegetation index. Based on the continuously distributed multi-dimensional ecological state vectors across the entire spatiotemporal range, the intelligent assessment and analysis terminal for ecological restoration effectiveness can integrate and reconstruct a complete spatiotemporally continuous ecological state field. This spatiotemporally continuous ecological state field possesses the technical attributes of a digital twin ground state field, overcoming the spatiotemporal gaps and scale fragmentation defects of traditional discrete monitoring data, and exhibiting characteristics of continuous spatiotemporal coverage across the entire range, and self-consistency between internal physical logic and ecological evolution laws. The intelligent assessment and analysis terminal for ecological restoration effectiveness can rely on the constructed spatiotemporal continuous ecological state field to respond to query requests for any specified continuous spatiotemporal coordinates and output a real-time snapshot of the complete multidimensional ecological state vector corresponding to that spatiotemporal location, filling the data gaps in the original monitoring data under scenarios such as cloud obstruction, sampling gaps, and asynchronous time.

[0034] S104. Based on the implementation records of ecological restoration measures, a processing variable field is constructed, and the outcome variable and covariate are extracted from the spatiotemporal continuous ecological state field.

[0035] Specifically, the intelligent assessment and analysis terminal for ecological restoration effectiveness can retrieve archived materials such as implementation logs, engineering design drawings, and construction ledgers of ecological restoration projects in the area to be assessed. It digitizes and spatializes various textual and pictorial information on restoration measures, quantifying the spatial scope, time intervals, type, and intensity parameters of each restoration measure. This constructs a continuously distributed processing variable field in a spatiotemporal dimension, characterizing the intensity distribution of different restoration measures, such as tree planting density and soil conditioner application rate, within the spatiotemporal domain. Simultaneously, the terminal can extract key ecosystem response indicators as outcome variables from the constructed spatiotemporally continuous ecological state field. These outcome variables can include core indicators representing ecological restoration effectiveness, such as regional plant and animal species diversity indices, biological population numbers, and vegetation succession levels. It can also simultaneously extract covariate datasets, encompassing the initial ecological baseline state before restoration project implementation, atmospheric environmental factors such as precipitation and temperature monitored by meteorological stations, and external interference factors such as regional land use policies and socio-economic statistics. This comprehensively classifies the intervention variables, effectiveness variables, and confounding variables required for causal inference.

[0036] S105. Using a hybrid algorithm based on constraints and fractions, under the constraint of time-causal acyclicity, the causal relationships between the variable field, outcome variable and covariate are analyzed and processed to obtain the causal graph.

[0037] For example, the intelligent assessment and analysis terminal for ecological restoration effectiveness can employ a hybrid algorithm combining constraints and scores. This algorithm can utilize a PC algorithm variant optimized and improved to suit spatiotemporal panel data. Under the premise of strictly adhering to the temporal causal acyclic constraint, it systematically analyzes and infers the pairwise relationships between processing variable fields, outcome variables, and covariates. The temporal causal acyclic constraint sets the natural chronological order as a rigid rule for constructing causal associations, strictly preventing the occurrence of reverse causal arrows pointing from result events to cause events. This effectively reduces the algorithm's search space and avoids the problem of causal inversion at the statistical level. The intelligent assessment and analysis terminal for ecological restoration effectiveness identifies the direct association skeleton structure between variables by conducting conditional independence tests on each group of variables. Combined with the causal structure smoothing prior constraints of spatially adjacent sample points, it ensures the consistency of causal logic in spatially contiguous areas, corrects and improves the logical pointing relationships between variables, and finally generates a complete causal graph including processing variable nodes, outcome variable nodes, and covariate nodes, clearly defining the mixed associations, mediating transmissions, and independent influence relationships among the variables.

[0038] S106. Based on the causal graph, the processing variable field, outcome variable and covariate field are input into the multimodal structural equation model. By applying intervention and fixing the exogenous noise in the multimodal structural equation model, counterfactual inference is performed to quantify the causal effect value of each ecological restoration measure on the ecological effect. The multimodal structural equation model is composed of multiple deep network sub-modules connected according to the parent-child dependency relationship in the causal graph.

[0039] Preferably, the intelligent assessment and analysis terminal for ecological restoration effectiveness can construct a multimodal structural equation model (SEM) based on the parent-child dependency topology of nodes presented in the causal graph. This SEM model is composed of multiple deep network sub-modules connected in an orderly manner. Each deep network sub-module corresponds one-to-one with each endogenous variable node in the causal graph, strictly replicating the nonlinear correlation mapping logic between variables. The intelligent assessment and analysis terminal for ecological restoration effectiveness synchronously inputs the constructed treatment variable field, outcome variable, and covariate field into the multimodal SEM model. By solving the prediction bias of each endogenous variable node under the model's observation path, it back-calculates the posterior estimate of the exogenous noise term matching each node. Based on this, the intelligent assessment and analysis terminal for ecological restoration effectiveness can apply targeted intervention operations to the multimodal SEM model, fixing the treatment variable node corresponding to the target ecological restoration measure to a preset intervention value. At the same time, it cuts off all correlation arrows pointing to the treatment variable node in the causal graph, while keeping the exogenous noise term values ​​corresponding to the remaining nodes constant, and re-drives the model's forward propagation operation to obtain the counterfactual outcome value under the intervention conditions. The intelligent assessment and analysis terminal for ecological restoration effectiveness calculates individual causal effects by comparing observed outcome values ​​with counterfactual outcome values ​​under the same spatiotemporal coordinates. It then calculates the mathematical expectation of individual causal effects for the entire spatiotemporal sample, removes confounding factors such as climate fluctuations and social policies, and finally accurately quantifies the independent causal effect value of each ecological restoration measure's contribution to ecological effectiveness.

[0040] In the aforementioned intelligent evaluation and analysis method for ecological restoration effectiveness, multi-source heterogeneous ecological monitoring data covering the area to be evaluated is acquired. This data is then discretized according to data source type to obtain a standardized sparse observation representation set containing spatiotemporal coordinates and corresponding observation vectors. Each spatiotemporal coordinate is input into a spatiotemporal manifold mapping network containing a spatial coding module, a temporal coding module, and a gated fusion layer for forward propagation, obtaining the manifold coordinates of each data point in a unified high-dimensional manifold space. These manifold coordinates are then input into an ecological state decoder to obtain a multidimensional ecological state vector. A spatiotemporally continuous ecological state field is constructed, allowing querying of snapshots of multidimensional ecological state vectors corresponding to any continuous spatiotemporal coordinate. A processing variable field is constructed based on ecological restoration measure implementation records, and outcome variables and covariates are extracted from the spatiotemporally continuous ecological state field. This is then analyzed using constraint- and score-based methods. The hybrid algorithm analyzes the causal relationships between variable fields, outcome variables, and covariates under the constraint of temporal causality and acyclicity to obtain a causal graph. Based on this causal graph, the variables are input into a multimodal structural equation model composed of multiple deep network sub-modules connected according to the parent-child dependency relationship of the causal graph. By applying intervention and fixing exogenous noise, counterfactual inference is performed, and finally, the causal effect values ​​of various ecological restoration measures on ecological effectiveness are quantified. This solves the problems of inaccurate data splicing, only being able to describe correlation, and being unable to remove the interference of confounding factors in traditional methods. It improves the fidelity of multi-source data fusion and the scientificity, accuracy, and reliability of ecological restoration effectiveness assessment. It realizes the transformation from data splicing to manifold reconstruction and from correlation description to causal quantification, and provides scientific support for precise ecological restoration policy implementation.

[0041] In one embodiment, the spatiotemporal coordinates include a first type of spatiotemporal coordinates, a second type of spatiotemporal coordinates, and a third type of spatiotemporal coordinates; the first type of spatiotemporal coordinates are obtained by extracting the sampling timestamp and sensor spatial location from the time-series data collected by IoT sensors; the second type of spatiotemporal coordinate data are obtained by extracting the effective pixel center geographic coordinates and image acquisition timestamp from satellite remote sensing images after radiometric calibration and atmospheric correction; and the third type of spatiotemporal coordinates are obtained by extracting the quadrat center geographic coordinates and survey timestamp from artificial sample plot survey data.

[0042] Specifically, the intelligent assessment and analysis terminal for ecological restoration effectiveness can classify all spatiotemporal coordinates within the standardized sparse observation expression set into three categories: Category 1, Category 2, and Category 3 spatiotemporal coordinates. For time-series monitoring data continuously collected by IoT sensors over a long period, the terminal extracts the spatial location coordinates of each sensor's fixed deployment and the timestamp corresponding to each data sampling. Based on these two types of information, it constructs the Category 1 spatiotemporal coordinates, using the corresponding physical observations as the Category 1 observation value vector, fully adapting to the high temporal resolution monitoring data attributes of IoT single-point local areas. The terminal first performs radiometric calibration and atmospheric correction on the raw remote sensing images acquired by satellite, eliminating data distortion caused by radiometric errors and atmospheric scattering and absorption. Then, it extracts the geographic coordinates of the pixel center and the satellite imaging acquisition timestamp for each effective pixel in the corrected image, constructing the Category 2 spatiotemporal coordinates. The multi-band reflectance of the corresponding pixel and its derived vegetation index are used as the Category 2 observation value vector, matching the data characteristics of the wide-area spatial coverage of the remote sensing images. The intelligent assessment and analysis terminal for ecological restoration effectiveness extracts the central geographic coordinates of each survey plot and the survey timestamps corresponding to the on-site species sampling and biomass statistics based on the survey data of artificial sample plots deployed in the field. It combines these coordinates to form a third type of spatiotemporal coordinates and uses the corresponding species abundance, biomass and diversity indices as the third type of observation value vector, which is consistent with the data distribution characteristics of the discrete sampling of artificial sample plots.

[0043] The spatiotemporal coordinates of each data point in the standardized sparse observation representation set are input into a spatiotemporal manifold mapping network and forward propagated to obtain the manifold coordinates of each data point in the observation vector in a unified high-dimensional manifold space, including:

[0044] S21. Spatial augmentation of geographic coordinates in each spatiotemporal coordinate is performed using a digital elevation model covering the area to be evaluated to obtain augmented spatial coordinates. Based on the spatial coding module, the augmented spatial coordinates are mapped using a multi-scale learnable Gaussian radial basis function to obtain spatial embedding.

[0045] In a demonstrative sense, the intelligent assessment and analysis terminal for ecological restoration effectiveness can retrieve digital elevation model (DEM) data covering the entire area to be assessed. From the DEM, it calculates and generates terrain-derived feature parameters such as slope, aspect, and topographic humidity index. These topographic feature parameters are then dimensionally spliced ​​and fused with the base geographic coordinates inherent in each spatiotemporal coordinate system, achieving spatial enhancement of the original geographic coordinates and thus obtaining enhanced spatial coordinates. The terminal can also invoke the spatial encoding module built into the spatiotemporal manifold mapping network, employing a multi-scale learnable Gaussian radial basis function as the core mapping operator. This allows for high-dimensional feature mapping operations on the enhanced spatial coordinates. The multi-scale Gaussian radial basis function can adaptively fit the spatial gradation patterns of ecological states under different terrain undulations and geomorphic units, fully incorporating the natural constraints of terrain on the ecological evolution process. Finally, the spatial encoding module maps and generates spatial embeddings with rich spatial semantic information. For example, the multi-scale learnable Gaussian radial basis function has a fixed three-level spatial scale hierarchy, with 16 learnable basis functions configured at each level. The centers of the learnable basis functions are uniformly initialized according to the geographic coordinates of the region to be evaluated, and the radial width is adaptively and iteratively optimized by backpropagation of the network. This can accurately adapt to the ecological spatial gradient characteristics of different landform units such as plains, mountains, and hills.

[0046] S22. Based on the time coding module, the timestamps in each spatiotemporal coordinate are encoded and mapped by a sine and cosine coding function based on adaptive learning frequency parameters to obtain the time embedding.

[0047] Furthermore, the intelligent assessment and analysis terminal for ecological restoration effectiveness can activate the time encoding module configured by the spatiotemporal manifold mapping network. It then uses a sine / cosine encoding function based on adaptively learned frequency parameters to encode and map the timestamps carried by each spatiotemporal coordinate. The corresponding encoding formula can be... In the formula, The original timestamp to be encoded in the spatiotemporal coordinates. The frequency parameters are adaptively trained from a multi-source monitoring dataset covering the entire region. The frequency parameters are arranged on a log-linear scale, which can simultaneously capture multi-scale ecological time-series periodic patterns such as diurnal variation, seasonal variation, and interannual variation. This represents the total number of dimensions set in the time encoding. The intelligent assessment and analysis terminal for ecological restoration effectiveness relies on this encoding function to transform linear timestamps into high-dimensional time-series vectors with periodic characteristics, generating time embeddings that can characterize the temporal evolution patterns of the ecosystem. Optional, the number of time encoding dimensions. The fixed value is 16, frequency parameter The logarithmic linear distribution range is set to 0.001–365, covering the entire time series of days, months, seasons, and years, and the initial parameters are generated strictly following the logarithmic equal interval rule.

[0048] S23. By using a gated fusion layer to perform feature interaction between spatial embedding and temporal embedding, the manifold coordinates corresponding to the spatiotemporal coordinates are obtained.

[0049] Optionally, the intelligent assessment and analysis terminal for ecological restoration effectiveness can utilize a gated fusion layer in a spatiotemporal manifold mapping network to achieve cross-dimensional feature interaction and information fusion between spatial and temporal embeddings. This gated fusion layer can employ a fully connected network structure configured with a sigmoid activation function, dynamically and adaptively adjusting the fusion weights of spatial and temporal embedding features based on the evolutionary response characteristics of the ecosystem at different times. This accurately reflects the differences in the sensitivity of vegetation growth and soil hydrological changes to spatial location under different climatic conditions such as rainy and dry seasons. After achieving deep coupling of spatiotemporal features through the gated fusion layer, the intelligent assessment and analysis terminal integrates the separated spatial and temporal features into a unified abstract high-dimensional representation, solving for the unique manifold coordinates corresponding to each spatiotemporal coordinate in the unified high-dimensional manifold space.

[0050] In one embodiment, the ecological state decoder and the spatiotemporal manifold mapping network are obtained through offline joint training; the offline joint training corresponds to the following steps:

[0051] S31. Construct an ecological state decoder based on a multi-task learning architecture; the multi-task learning architecture includes multiple independent task heads.

[0052] As an illustration, the intelligent assessment and analysis terminal for ecological restoration effectiveness can employ a deep neural network architecture to construct an ecological state decoder based on a multi-task learning architecture. This multi-task learning architecture includes a shared low-level feature extraction module and independently configured task output heads. The low-level feature extraction module receives manifold coordinates output by the spatiotemporal manifold mapping network and performs deep feature mining. Multiple independent task heads correspond to different ecological state variable prediction tasks. Each independent task head can adopt a fully connected network or convolutional network structure, and each task head shares the low-level manifold coordinate features. The multi-task learning architecture can effectively ensure the physical and logical consistency between different ecological state variables, avoiding situations where the prediction results contradict the laws of ecological evolution. At the same time, the prediction accuracy of individual ecological state variables can be improved through multi-task collaborative training. For example, the multi-task learning architecture's underlying feature extraction module is configured with a 3-layer fully connected network, with 256, 128, and 64 neurons in each layer, respectively. The ReLU activation function is used uniformly, and layer normalization is configured. Each independent task head is configured with a 2-layer fully connected network, with 64 and 32 neurons. The output layer uses a linear activation function. The parameters of each task head are updated independently, sharing only the underlying manifold features, thus balancing multi-task collaboration and variable prediction independence.

[0053] S32. Using the manifold coordinates output by the spatiotemporal manifold mapping network as shared input, soil moisture content, surface temperature, vegetation canopy height and normalized vegetation index are predicted by each independent task head to obtain a multidimensional ecological state vector.

[0054] Furthermore, the intelligent assessment and analysis terminal for ecological restoration effectiveness can use all manifold coordinates output by the spatiotemporal manifold mapping network as shared input to the ecological state decoder. The manifold coordinates are input to the underlying feature extraction module for feature transformation and fusion to obtain shared high-dimensional features that can characterize the ecological state of the entire region. These shared high-dimensional features are then input into each independent task head. Each independent task head completes the fitting and prediction of specific ecological state variables according to its corresponding prediction target. One independent task head can be used specifically to predict soil moisture content, another can be used specifically to predict surface temperature, and the other two can be used to predict vegetation canopy height and normalized vegetation index, respectively. Through the parallel prediction of each task head, a multi-dimensional ecological state vector containing soil moisture content, surface temperature, vegetation canopy height, and normalized vegetation index is finally output, realizing the synchronous prediction of multi-dimensional ecological state.

[0055] S33. Based on the training of randomly sampled points in the entire spatiotemporal domain without observations, the derivative of the multidimensional ecological state vector with respect to time is calculated by automatic differentiation, and the residual between the derivative and the net flux is calculated to obtain the first physical loss term; the net flux is calculated through the ecosystem process equation.

[0056] Optionally, the intelligent assessment and analysis terminal for ecological restoration effectiveness can construct a training set containing multi-source heterogeneous ecological monitoring data during the offline training phase. During training, a large number of configuration points are randomly sampled from the entire spatiotemporal domain of the training set where no observations are found. These configuration points are evenly distributed across the entire spatiotemporal range of the area to be evaluated, and are used to apply physical process constraints. The intelligent assessment and analysis terminal for ecological restoration effectiveness can use automatic differentiation techniques to calculate the derivative of each ecological state variable with respect to time in the multidimensional ecological state vector. Simultaneously, it calculates the corresponding net flux through the ecosystem process equation. For example, the ecosystem process equation can be a simplified one-dimensional soil moisture balance equation, the formula of which is... In the formula, Represents soil moisture content in a multidimensional ecological state vector. Represents time, It is the derivative of soil moisture content with respect to time. These are known meteorological driving variables, specifically rainfall data obtained from meteorological station monitoring. It is the evaporation rate. It is surface runoff. This refers to the amount of deep leakage, among which, , , Both can be expressed as soil moisture content. Differentiable functions of the normalized vegetation index. The intelligent assessment and analysis terminal for ecological restoration effectiveness calculates the residual between the derivative obtained through automatic differentiation and the net flux calculated through this equation. The mean square error of this residual is used as the first physical loss term to constrain the ecological state prediction results to conform to the physical laws of soil moisture balance. Optionally, random sampling points are uniformly sampled in the entire time and space domain at a density of 50 points per square kilometer, with a fixed sampling of 2048 unobserved points in a single training batch; evapotranspiration. Surface runoff Deep leakage The differentiable functions are all fitted with quadratic polynomials. A continuous differentiable expression is constructed with soil moisture content and normalized vegetation index as independent variables. After the physical loss term is calculated, a basic residual tolerance threshold of 0.01 is set. If the threshold is exceeded, the weight of the physical constraint is increased.

[0057] S34. Based on the cross-variable empirical constraint relationship, calculate the relationship between the ecological state vectors of each dimension in the multidimensional ecological state vector and the deviation between the two dimensions and the known general function form to obtain the second physical loss term.

[0058] Specifically, the intelligent assessment and analysis terminal for ecological restoration effectiveness can pre-determine the cross-variable empirical constraints between ecological state variables in each dimension of the multidimensional ecological state vector. These cross-variable empirical constraints are empirical laws with clear physical meaning, derived from long-term ecological observations, such as the empirical correlation between the normalized vegetation index (NWRI) and the leaf area index (LAI). The terminal can preset the known general function form corresponding to these empirical constraints, substitute the predicted values ​​of each relevant dimension in the multidimensional ecological state vector into this general function form, calculate the deviation between the predicted variable relationships and the known general function form, and use the mean square error of this deviation as the second physical loss term. This ensures that the relationships between the indicators in the multidimensional ecological state vector conform to the empirical laws of ecological evolution, further improving the physical rationality of the prediction results. For example, the NWRI and LAI use a classical logarithmic fitting general function as the constraint benchmark. The terminal calculates the absolute deviation between the predicted variable relationships and this general function, setting 0.02 as the upper limit of deviation tolerance. If the deviation exceeds the upper limit, the penalty of the second physical loss term is increased to ensure that the variable correlation conforms to objective ecological laws.

[0059] S35. Calculate the data fitting loss between the multidimensional ecological state vector and the corresponding observation vector, which are concentrated on the spatiotemporal coordinates with observations, and then sum the data fitting loss, the first physical loss term and the second physical loss term by weight to obtain the total loss function.

[0060] For example, the intelligent assessment and analysis terminal for ecological restoration effectiveness can extract all spatiotemporal coordinates with observed values ​​from the training set. It then compares the predicted values ​​of the multidimensional ecological state vectors corresponding to these spatiotemporal coordinates with the actual observed value vectors in the training set dimension by dimension. The data fitting loss is obtained by calculating the mean square error between the two. The data fitting loss reflects the degree of deviation between the predicted values ​​and the actual observed values; the smaller the deviation, the better the data fitting effect. The intelligent assessment and analysis terminal for ecological restoration effectiveness can set multiple balancing hyperparameters to weight the data fitting loss, the first physical loss term, and the second physical loss term, respectively. The weighted losses are then summed to obtain the total loss function, the expression of which is: In the formula, Represents the data fitting loss. This represents the sum of the first physical loss term and the second physical loss term. It is a regularization term for manifold coordinates, used to encourage the smoothness and compactness of high-dimensional manifold spaces and avoid overfitting in the model. and All are balanced hyperparameters. The intelligent evaluation and analysis terminal for ecological restoration effectiveness can adaptively adjust their values ​​based on training results to ensure a balance between data fitting accuracy and physical constraints. Optional, balanced hyperparameters... The initial setting is 0.5, and the adjustment range is limited to 0.1 to 1.0. The initial value is set to 0.3, and the adjustment range is limited to 0.05–0.8; regularization term Using L2 regularization, weight decay constraints are applied to all feature parameters of the manifold coordinates to force the high-dimensional manifold space to maintain a smooth and compact distribution, thus avoiding model overfitting.

[0061] S36. Based on the total loss function, the parameters of the spatiotemporal manifold mapping network and the ecological state decoder are updated synchronously through gradient backpropagation to obtain the spatiotemporal manifold mapping network and the ecological state decoder.

[0062] Optionally, the intelligent evaluation and analysis terminal for ecological restoration effectiveness can employ the Adam optimizer as the parameter update optimizer. This optimizer can adaptively adjust the learning rate, improving training convergence speed and parameter optimization accuracy. The intelligent evaluation and analysis terminal for ecological restoration effectiveness uses the backpropagation algorithm to calculate the gradient of the total loss function with respect to all learnable parameters in the spatiotemporal manifold mapping network and the ecological state decoder. Based on the gradient direction and magnitude, the parameters of both networks are updated synchronously. During the update process, both observed sampling points and unobserved configuration points in the training set are considered, ensuring both the good fit between the predicted results and the actual observation data, and that the predicted results conform to physical laws and empirical constraints. The intelligent evaluation and analysis terminal for ecological restoration effectiveness can repeat the above training process until the value of the total loss function converges to a preset threshold. At this point, training stops, resulting in a fully trained spatiotemporal manifold mapping network and ecological state decoder, capable of stably outputting high-fidelity manifold coordinates and multidimensional ecological state vectors. For example, when the intelligent evaluation and analysis terminal for ecological restoration effectiveness uses the Adam optimizer, the initial learning rate is fixed at 0.001, the training batch size is set to 256, and the weight decay coefficient is set to 1e-5. The model convergence condition is set as follows: the total loss function fluctuates less than 1e-4 for 50 consecutive iterations, and the gradient magnitude of the network parameter update is less than 1e-3. If both conditions are met, the training is judged to have converged and the parameter update is stopped.

[0063] In one embodiment, a hybrid algorithm based on constraints and scores is used to analyze the pairwise causal relationships between variable fields, outcome variables, and covariates under time-causal acyclic constraints, resulting in a causal graph, including:

[0064] S41. For any two variables among the treatment variable field, outcome variable, and covariate, calculate the empirical conditional mutual information partial derivative correlation statistic of the two variables in the regenerating kernel Hilbert space under different condition sets by using the kernel-based conditional independence test, and obtain the skeleton connection between the variables.

[0065] In a schematic manner, the intelligent assessment and analysis terminal for ecological restoration effectiveness can incorporate all processing variables, outcome variables, and covariates into a unified variable node set. It then sequentially selects any two sets of variables within this set as the detection objects and employs a kernel-based conditional independence test to quantify the correlation. The terminal can use the normalized form of the Hilbert-Schmidt independence criterion as the testing benchmark, mapping the two sets of variables to be tested onto the regeneration kernel Hilbert space. Given different levels of condition sets, it solves for the correlation statistics of the empirical conditional mutual information partial derivatives of the two sets of variables, using the magnitude of these statistics to measure the degree of conditional independence between variables. The terminal can systematically traverse the conditional independence test process for all pairwise combinations of variables, starting from an empty set and gradually increasing the size of the condition set. Based on the test results, it selects directly correlated variable pairs, builds a partial ancestor graph structure layer by layer, and uses this structure to identify stable skeletal connections between all variables, accurately classifying variable combinations with and without direct correlation. For example, the conditional independence test uniformly uses the Gaussian kernel as the kernel function, and traverses and searches according to the rule of gradually increasing from the empty set to the 5th order condition set. The significance level of the conditional independence test is set to 0.05, and the Hilbert-Schmidt independence criterion is used for normalization calculation. When the statistic is greater than the significance threshold, it is determined that there is a skeleton connection relationship between the variables.

[0066] S42. Based on the temporal causal acyclic constraint, causal directions are constructed according to the skeleton connection, and a smoothing prior is applied to the causal directions between spatially adjacent variable nodes to obtain the causal graph; the temporal causal acyclic constraint sets the temporal sequence as a hard constraint of the causal topology and prohibits the generation of causal directions from result to cause; the causal graph includes processing variable nodes, outcome variable nodes and covariate nodes.

[0067] Furthermore, the intelligent assessment and analysis terminal for ecological restoration effectiveness uses the acquired variable skeleton connection as its basic framework. It introduces a temporal causal acyclic constraint to initially construct causal orientation. This constraint strictly sets the natural temporal sequence as a rigid rule for causal deduction, strictly prohibiting the generation of reverse arrows pointing from results to causes in the topological structure. This reduces the computational space for causal structure search and avoids the causal inversion problem that easily occurs in statistical association from the bottom up. The terminal can also apply smoothing prior constraints to spatially adjacent variable nodes, assuming that the causal topological structure between variables within a spatially contiguous area remains consistent. Local adjustments to the causal direction are only allowed when multi-source monitoring data show significant deviations, thus ensuring the spatial continuity and stability of the overall causal structure. The terminal integrates skeleton connection relationships, temporal causal acyclic constraints, and spatial smoothing prior constraints to refine the pointing logic and dependency relationships between variable nodes, ultimately generating a complete causal graph. This graph fully includes processing variable nodes, outcome variable nodes, and covariate nodes, clearly defining parent-child dependencies, mixed associations, and mediating relationships between nodes. Optionally, the temporal causal acyclic constraint uses the natural day as the smallest granularity for judgment, and strictly prohibits subsequent time variables from pointing to previous time variables; the spatial adjacency is set with a distance threshold of 500 meters, and spatial samples within 500 meters have the same causal structure by default. At the same time, a smoothing prior penalty coefficient of 0.2 is set, and local adjustment of causal direction is only allowed when the data feature deviation exceeds the penalty threshold.

[0068] In one embodiment, based on a causal graph, the processing variable field, outcome variable, and covariate field are input into a multimodal structural equation model. Counterfactual inference is then performed by applying interventions and fixing exogenous noise in the multimodal structural equation model to quantify the causal effect values ​​of various ecological restoration measures on ecological outcomes, including:

[0069] S51. Instantiate the nonlinear functions corresponding to each endogenous variable node in the multimodal structural equation model into deep neural network sub-modules, and connect the inputs and outputs of each deep neural network sub-module according to the parent-child dependency relationship in the causal graph to obtain the causal effect estimation network.

[0070] In a schematic representation, each endogenous variable node in a multimodal structural equation model corresponds to a nonlinear function. The intelligent assessment and analysis terminal for ecological restoration effectiveness can instantiate these nonlinear functions into independent deep neural network sub-modules. Each deep neural network sub-module can adopt a fully connected network, convolutional neural network, or Transformer network structure, with the specific structure adaptively selected based on the complexity of the corresponding endogenous variable's features. The intelligent assessment and analysis terminal for ecological restoration effectiveness strictly follows the parent-child dependency relationships of nodes in the causal graph, connecting the inputs and outputs of each deep neural network sub-module accordingly. That is, the output of the deep neural network sub-module corresponding to the parent node is directly connected to the input of the deep neural network sub-module corresponding to its child node. This ensures that the entire network structure strictly replicates the topological constraints of the causal graph, avoiding connections that violate causal logic. Ultimately, a complete causal effect estimation network is constructed, which can accurately represent the nonlinear causal mapping relationships between variables. For example, the deep neural network sub-modules corresponding to each endogenous variable are uniformly set with a 3-layer fully connected structure. The middle layer uses the ReLU activation function, and the output layer uses the linear activation function. The input and output dimensions of all sub-modules strictly match the relationship between nodes in the causal graph, and arbitrary connections across dimensions are prohibited to ensure that the topology strictly replicates the causal logic.

[0071] S52. Input the processing variable field, outcome variable, and covariate field into the causal effect estimation network, calculate the predicted value of each endogenous variable node under the observation path, and calculate the error between the predicted value and the observed value vector to obtain the posterior estimate of the exogenous noise term corresponding to the endogenous variable node.

[0072] Furthermore, the intelligent assessment and analysis terminal for ecological restoration effectiveness can simultaneously input the constructed processing variable field, outcome variable, and covariate field into the causal effect estimation network, driving the network to perform forward propagation calculations according to the observation path. This means that without any intervention, the propagation path completely follows the original causal relationship, calculating the predicted values ​​for each endogenous variable node. The intelligent assessment and analysis terminal can then compare the predicted values ​​of each endogenous variable node with the corresponding observed value vector dimension by dimension. By calculating the deviation between the two, it can infer the posterior estimate of the exogenous noise term corresponding to each endogenous variable node. Here, exogenous noise refers to independent interference factors that are not affected by other endogenous variables in the model. Its posterior estimate can instantiate unobservable exogenous interference factors, ensuring the accuracy of subsequent counterfactual inferences. The intelligent assessment and analysis terminal can use the least squares method or the maximum likelihood estimation method to calculate this deviation, thereby obtaining accurate posterior estimates of the exogenous noise term. For example, the intelligent assessment and analysis terminal for ecological restoration effectiveness uses the maximum likelihood estimation method to back-calculate the posterior estimate of the exogenous noise term. With the maximization of the log-likelihood function between the observed value and the model prediction value as the optimization objective, it iteratively solves to obtain the unique corresponding exogenous noise value, ensuring the uniqueness and accuracy of the instantiation of exogenous interference factors.

[0073] S53. Intervene in the causal effect estimation network and rerun the forward propagation of the causal effect estimation network to obtain the counterfactual outcome values ​​corresponding to the spatiotemporal coordinates under intervention. The intervention operation corresponds to forcibly setting the values ​​of the treatment variable nodes corresponding to the target ecological restoration measures to the preset intervention values, cutting off all arrows pointing to the treatment variable nodes in the causal graph, and keeping the posterior estimates of the exogenous noise terms of other nodes unchanged.

[0074] Optionally, the intelligent assessment and analysis terminal for ecological restoration effectiveness can perform targeted intervention operations on the causal effect estimation network for each ecological restoration measure to be evaluated. The core of the intervention operation is to isolate the treatment variable node corresponding to the target restoration measure and eliminate the influence of other factors on the treatment variable. Specifically, the intelligent assessment and analysis terminal forcibly sets the value of the treatment variable node corresponding to the target ecological restoration measure to a preset intervention value. This intervention value can be a preset threshold for the intensity of restoration measure implementation, the implementation scope threshold, etc. At the same time, it cuts off all arrows in the causal graph pointing to the treatment variable node, completely blocking the driving effect of other variables on the treatment variable. During this process, the intelligent assessment and analysis terminal keeps the posterior estimates of the exogenous noise terms corresponding to all other nodes unchanged, ensuring that the influence of other interfering factors remains consistent except for the target treatment variable. Then, it re-runs the forward propagation operation of the causal effect estimation network to obtain the predicted value of the outcome variable corresponding to each spatiotemporal coordinate in the area to be evaluated under the intervention scenario, that is, the counterfactual outcome value of the spatiotemporal coordinate under the intervention. For example, the intervention values ​​for remediation measures are pre-defined into three intensity levels: low, medium, and high. The values ​​for each level are determined according to the ternary digit of the actual amount of remediation measures implemented in the region. During intervention, the fixed level values ​​do not fluctuate randomly. At the same time, all reverse causal arrows are strictly cut off to prevent indirect interference from other variables on the treated variable.

[0075] S54. Compare the counterfactual outcome values ​​and observation vectors at the same spatiotemporal coordinates to calculate the individual causal effect, and calculate the expected value of the individual causal effect at each spatiotemporal coordinate covering the area to be evaluated to obtain the causal effect value.

[0076] Specifically, the intelligent assessment and analysis terminal for ecological restoration effectiveness can extract counterfactual outcome values ​​and original observation vectors corresponding to all spatiotemporal coordinates within the area to be assessed. It compares the counterfactual outcome values ​​with the observation vectors at the same spatiotemporal coordinates one by one, calculating the difference between them. This difference represents the individual causal effect of the target ecological restoration measure at that spatiotemporal coordinate, reflecting the degree of independent impact of the restoration measure on ecological effectiveness at a single spatiotemporal location. The terminal can traverse all spatiotemporal coordinates covering the area to be assessed, collecting the individual causal effects corresponding to all coordinates. Then, by calculating the mathematical expectation of all individual causal effects, it obtains the average causal effect value of the target ecological restoration measure on ecological effectiveness, i.e., the final causal effect value. For example, for a tree species planting restoration measure, its causal effect value can be quantified as the percentage increase in the target bird population after excluding interference from climatic factors such as precipitation and temperature, clearly presenting the true effectiveness of the restoration measure. The intelligent assessment and analysis terminal can use the arithmetic mean or weighted average method to calculate this expectation, ensuring the objectivity and accuracy of the results. For example, the overall individual causal effect is calculated using a spatial partition weighted average method, with the ecological protection priority of each spatial grid unit as the weighting coefficient. The higher the priority, the greater the weight. The final weighted expected result is used as the standard causal effect value of the restoration measure, thereby improving the spatial rationality of the assessment results.

[0077] In one embodiment, the multimodal structural equation model is constructed through the following steps:

[0078] S61. Construct a causal training set based on the processing variable field, outcome variable, and covariate field. With the goal of minimizing the reconstruction error of the predicted and observed value vectors of each endogenous variable node, train the causal effect estimation network through the causal training set until the convergence condition is met to obtain the initial causal effect estimation network. The initial causal effect estimation network is used to characterize the nonlinear function mapping relationship between variables that conform to the topological constraints of the causal graph.

[0079] Indicatively, the intelligent assessment and analysis terminal for ecological restoration effectiveness can first integrate and process all data from the variable field, outcome variable, and covariate field to construct a causal training set for model training. During construction, the data can be stratified and sampled according to spatiotemporal dimensions, dividing the area to be evaluated into multiple sub-regions using a spatial grid and into multiple time periods according to a time series. This ensures that each sub-region and each time period has sufficient sample size, guaranteeing the spatial representativeness and temporal integrity of the causal training set and avoiding model training bias due to uneven sample distribution. The intelligent assessment and analysis terminal for ecological restoration effectiveness uses minimizing the reconstruction error between the predicted and observed value vectors of each endogenous variable node as its core training objective. The causal training set is progressively input into the causal effect estimation network, and gradient descent or its variants, such as the Adam optimizer and RMSprop optimizer, are used to iteratively update the network parameters. During training, the reconstruction error between the predicted and observed value vectors of each endogenous variable node is calculated in real time. This reconstruction error can be quantified using metrics such as mean squared error and mean absolute error. The intelligent evaluation and analysis terminal for ecological restoration effectiveness can preset convergence conditions. These conditions can be either that the reconstruction error remains below a preset threshold for multiple consecutive iterations, or that the network parameter update amplitude is less than a preset minimum value. When the convergence condition is met, training stops, resulting in an initial causal effect estimation network. This initial causal effect estimation network strictly follows the topological constraints of the causal graph, accurately representing the nonlinear function mapping relationship between variables, laying the foundation for subsequent adversarial optimization. Optionally, the causal training set is divided into training, validation, and test samples in a 7:2:1 ratio, with spatial grid division using 1-kilometer grid units and time segmentation in quarterly intervals. The network convergence condition is set to a stable initial causal effect estimation network if the validation set reconstruction error is below 0.01 for 30 consecutive iterations.

[0080] S62. Based on the causal graph, obtain a set of sufficiently adjusted variables for blocking the backdoor path of the processing. With the goal of minimizing the statistical dependence between the processing variable node and the sufficiently adjusted variables, perform adversarial training on the initial causal effect estimation network using the set of sufficiently adjusted variables to obtain a multimodal structural equation model.

[0081] Furthermore, the intelligent assessment and analysis terminal for ecological restoration effectiveness can automatically identify the set of sufficiently adjusted variables that can block all backdoor paths between the treatment variable node and the outcome variable node based on the generated causal graph using a causal structure analysis algorithm. Backdoor paths refer to confounding paths that simultaneously affect both treatment and outcome variable nodes and may lead to biased causal inference. The set of sufficiently adjusted variables is the minimum set of variables that can completely block these backdoor paths; essentially, it is a set of various confounding factors. The intelligent assessment and analysis terminal for ecological restoration effectiveness can traverse all potential paths in the causal graph and filter out the minimum variable combination that meets the blocking conditions to obtain the set of sufficiently adjusted variables. Optionally, the intelligent assessment and analysis terminal for ecological restoration effectiveness aims to minimize the statistical dependency between the treatment variable node and the sufficiently adjusted variables. It uses the set of sufficiently adjusted variables to perform adversarial training on the initial causal effect estimation network. During adversarial training, a discriminator network can be introduced to distinguish the degree of correlation between the values ​​of the treatment variable node and the sufficiently adjusted variables. The generator is the initial causal effect estimation network, and the two achieve collaborative optimization through adversarial game theory. The intelligent evaluation and analysis terminal for ecological restoration effectiveness can use indicators such as mutual information and KL divergence to measure the statistical dependency between the processing variable node and the fully adjusted variable. Through iterative training, this statistical dependency is continuously reduced, effectively eliminating the interference of confounding factors on causal inference and improving the model's causal identification ability. When the statistical dependency between the processing variable node and the fully adjusted variable decreases to a preset threshold, and the reconstruction error of the initial causal effect estimation network remains within a reasonable range, adversarial training stops, ultimately resulting in a fully trained multimodal structural equation model. This multimodal structural equation model can stably output variable mapping relationships that conform to causal logic. For example, the discriminator network in adversarial training uses a two-layer fully connected structure and a Sigmoid activation function to output the correlation coefficient. KL divergence is used as a statistical dependency quantification indicator, with a KL divergence below 0.005 set as the convergence criterion. Once the threshold is reached, adversarial training stops, completely preserving the causal relationship of the variables while eliminating confounding biases.

[0082] In one embodiment, the method further includes:

[0083] S71. Construct a global causal loss function based on the causal effect value, and backpropagate the global causal loss function through automatic differentiation to obtain the gradient norm of each spatiotemporal coordinate.

[0084] Indicatively, the intelligent assessment and analysis terminal for ecological restoration effectiveness can construct a global causal loss function based on the causal effect values ​​of various ecological restoration measures obtained from the second-stage quantification. This global causal loss function can integrate confounding bias correction terms present in the causal inference process, accurately reflecting the accuracy of the causal inference conclusions. The terminal can employ automatic differentiation technology to perform gradient backpropagation on the constructed global causal loss function. The backpropagation process penetrates the multimodal structural equation model, the ecological state decoder, and the spatiotemporal manifold mapping network, ultimately calculating the gradient norm of each spatiotemporal coordinate corresponding to the data point. This gradient norm can quantify the impact of changes in the observed value of each data point on the global causal loss function. The calculation of the gradient norm can be achieved through the automatic differentiation interface in the deep learning framework, ensuring computational accuracy and efficiency. For example, the global causal loss function is a linear combination of a causal effect fitting bias term and a confounding bias correction term, with initial weights of 0.6 and 0.4, respectively. During training, it adaptively fine-tunes with model iterations, comprehensively measuring the overall accuracy of the causal inference results.

[0085] S72. The causal contribution score of the causal inference conclusion by using the gradient norm as a spatiotemporal coordinate.

[0086] Furthermore, the intelligent assessment and analysis terminal for ecological restoration effectiveness directly uses the gradient norm as the causal contribution score of the corresponding spatiotemporal coordinate to the causal inference conclusion. The larger the gradient norm, the stronger the influence of the data point corresponding to the spatiotemporal coordinate on the final causal effect value, and the higher its contribution to the causal inference conclusion. Conversely, the smaller the gradient norm, the lower the causal contribution score of the data point, and the weaker its influence on the causal inference conclusion.

[0087] S73. Using the data source type of spatiotemporal coordinates, the interval between the corresponding timestamp and the current time, and the acquisition characteristics as input features, a confidence score is calculated through a data confidence evaluator based on a feedforward neural network. The acquisition characteristics are one of the following: sensor calibration status information corresponding to the data source type, atmospheric condition parameters during image acquisition, and investigator identification.

[0088] Specifically, the intelligent assessment and analysis terminal for ecological restoration effectiveness can construct a data credibility evaluator based on a feedforward neural network. This evaluator is a lightweight network structure that can quickly complete the credibility assessment of a single data point, avoiding excessive consumption of computing resources. The intelligent assessment and analysis terminal for ecological restoration effectiveness uses the data source type corresponding to each spatiotemporal coordinate, the interval between the timestamp of that spatiotemporal coordinate and the current time, and the acquisition characteristics of the corresponding data source as input features. These acquisition features vary depending on the data source type. For IoT sensor data, the acquisition features can be sensor calibration status information, including parameters such as the last calibration time and calibration error. For satellite remote sensing image data, the acquisition features can be atmospheric condition parameters at the time of image acquisition, including the solar zenith angle and atmospheric aerosol thickness. For artificial plot survey data, the acquisition features can be the surveyor's identification. The intelligent assessment and analysis terminal for ecological restoration effectiveness inputs these features into a data reliability evaluator. Through forward propagation computation, it outputs a scalar value between 0 and 1 as the reliability score for that spatiotemporal coordinate data point. The closer the score is to 1, the higher the collection quality and the stronger the reliability of the data point; the closer the score is to 0, the greater the noise and the weaker the reliability of the data point. For example, the data source type and investigator identification use one-hot encoding, and continuous features such as time interval, sensor calibration error, and atmospheric aerosol thickness are normalized to the 0-1 range using Min-Max. The data reliability evaluator uses a three-layer feedforward neural network with 128, 64, and 1 neurons respectively. The output layer generates a reliability score in the 0-1 range through Sigmoid activation.

[0089] S74. The causal contribution score and the confidence score are fused to obtain a double-weighted weight. The double-weighted weight is used to weight each loss value in the data fitting loss, so that the influence of training data with low confidence scores or low causal contribution scores in the training set is automatically suppressed during training.

[0090] For example, the intelligent assessment and analysis terminal for ecological restoration effectiveness fuses the causal contribution score and the confidence score to obtain a dual-weighted weight for each spatiotemporal coordinate data point. The fusion calculation is implemented using the sigmoid activation function, and its calculation formula is as follows: In the formula, Representing the The data points are weighted by a dual weighting of spatiotemporal coordinates. This is the activation function, used to map the fusion result to a range of 0 to 1, ensuring the reasonableness of the weights; and The fusion coefficient is initially set to 1 and can be adaptively adjusted through model training. It is used to balance the influence weights of the confidence score and the causal contribution score. Representing the The credibility score of each spatiotemporal coordinate data point Representing the The causal contribution score of each spatiotemporal coordinate data point. The dual weighting weight is mainly used to weight the loss value of each data point in the data fitting loss during the joint training of the ecological state decoder and the spatiotemporal manifold mapping network. The intelligent evaluation and analysis terminal for ecological restoration effectiveness can multiply the loss value of each data point with the corresponding dual weighting weight before participating in the calculation of the total loss function. This automatically suppresses the influence of training data with low confidence scores or low causal contribution scores in the training set during model training, guiding the model to prioritize fitting high-value data with high confidence and high causal contribution. This achieves end-to-end closed-loop optimization of data fusion and causal inference, further improving the evaluation accuracy and robustness of the model.

[0091] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0092] Based on the same inventive concept, this application also provides an intelligent evaluation and analysis system for ecological restoration effectiveness, used to implement the intelligent evaluation and analysis method for ecological restoration effectiveness described above. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the intelligent evaluation and analysis system for ecological restoration effectiveness provided below can be found in the limitations of the intelligent evaluation and analysis method for ecological restoration effectiveness described above, and will not be repeated here.

[0093] In one exemplary embodiment, such as Figure 2 As shown, an intelligent evaluation and analysis system 200 for ecological restoration effectiveness is provided, including:

[0094] The data acquisition module 201 is used to acquire multi-source heterogeneous ecological monitoring data covering the area to be evaluated, and to discretize the multi-source heterogeneous ecological monitoring data according to the data source type to obtain a standardized sparse observation expression set; the standardized sparse observation expression set includes spatiotemporal coordinates and the observation value vectors corresponding to the spatiotemporal coordinates;

[0095] The manifold coding module 202 is used to input the spatiotemporal coordinates of each data point in the standardized sparse observation representation set into the spatiotemporal manifold mapping network for forward propagation, so as to obtain the manifold coordinates of each data point in the observation vector in the unified high-dimensional manifold space; the spatiotemporal manifold mapping network includes a spatial coding module, a temporal coding module and a gated fusion layer;

[0096] The ecological state module 203 is used to input manifold coordinates into the ecological state decoder to obtain a multidimensional ecological state vector, and to construct a spatiotemporally continuous ecological state field based on the multidimensional ecological state vector.

[0097] Repair module 204 is used to construct a processing variable field based on the implementation records of ecological restoration measures, and to extract outcome variables and covariates from the spatiotemporal continuous ecological state field;

[0098] The causal analysis module 205 is used to analyze and process the causal relationships between variable fields, outcome variables and covariates under the constraint of time causality and acyclicity through a hybrid algorithm based on constraints and scores, and to obtain a causal graph.

[0099] The effectiveness attribution assessment module 206 is used to input the treatment variable field, outcome variable, and covariate field into the multimodal structural equation model based on the causal graph. By applying interventions and fixing exogenous noise invariant in the multimodal structural equation model, counterfactual inference is performed to quantify the causal effect value of each ecological restoration measure on the ecological effectiveness. The multimodal structural equation model is composed of multiple deep network sub-modules connected according to the parent-child dependency relationship in the causal graph.

[0100] In one embodiment, the manifold encoding module 202 is further configured to:

[0101] By using a digital elevation model covering the area to be evaluated, the geographic coordinates in each spatiotemporal coordinate are spatially augmented to obtain augmented spatial coordinates. Based on the spatial coding module, the augmented spatial coordinates are mapped through a multi-scale learnable Gaussian radial basis function to obtain spatial embedding.

[0102] Based on the time coding module, the timestamps in each spatiotemporal coordinate are encoded and mapped by a sine and cosine coding function based on adaptive learning frequency parameters to obtain the time embedding;

[0103] By using a gated fusion layer to perform feature interaction between spatial and temporal embeddings, the manifold coordinates corresponding to the spatiotemporal coordinates are obtained.

[0104] In one embodiment, the intelligent evaluation and analysis system 200 for ecological restoration effectiveness is also used for:

[0105] Construct an ecosystem state decoder based on a multi-task learning architecture; the multi-task learning architecture includes multiple independent task heads;

[0106] Using the manifold coordinates output by the spatiotemporal manifold mapping network as a shared input, soil moisture content, surface temperature, vegetation canopy height and normalized vegetation index are predicted by each independent task head to obtain a multidimensional ecological state vector;

[0107] Based on training focused on randomly sampled points in the entire spatiotemporal domain without observations, the derivative of the multidimensional ecological state vector with respect to time is calculated by automatic differentiation, and the residual between the derivative and net flux is calculated to obtain the first physical loss term; the net flux is calculated through the ecosystem process equation.

[0108] Based on cross-variable empirical constraints, the relationship between the ecological state vectors of each dimension in the multidimensional ecological state vector and the deviation between the two dimensions and the known general function form are calculated to obtain the second physical loss term.

[0109] The training focuses on the data fitting loss between the multidimensional ecological state vector on the spatiotemporal coordinates with observations and the corresponding observation vector. The data fitting loss, the first physical loss term, and the second physical loss term are weighted and summed to obtain the total loss function.

[0110] Based on the total loss function, the parameters of the spatiotemporal manifold mapping network and the ecological state decoder are updated synchronously through gradient backpropagation to obtain the spatiotemporal manifold mapping network and the ecological state decoder.

[0111] In one embodiment, the causal analysis module 205 is further configured to:

[0112] For any two variables among the treatment variable field, outcome variable, and covariate, the empirical conditional mutual information partial derivative correlation statistic of the two variables in the regenerating kernel Hilbert space is calculated by kernel-based conditional independence test to obtain the skeleton connection between variables.

[0113] Based on the temporal causal acyclic constraint, causal directions are constructed according to the skeleton connection, and a smoothing prior is applied to the causal directions between spatially adjacent variable nodes to obtain the causal graph. The temporal causal acyclic constraint sets the temporal sequence as a hard constraint of the causal topology and prohibits the generation of causal directions from result to cause. The causal graph includes processing variable nodes, outcome variable nodes, and covariate nodes.

[0114] In one embodiment, the effectiveness attribution assessment module 206 is further configured to:

[0115] The nonlinear functions corresponding to each endogenous variable node in the multimodal structural equation model are instantiated as deep neural network sub-modules, and the inputs and outputs of each deep neural network sub-module are connected according to the parent-child dependency relationship in the causal graph to obtain the causal effect estimation network.

[0116] The processing variable field, outcome variable, and covariate field are input into the causal effect estimation network. The predicted value of each endogenous variable node under the observation path is calculated, and the error between the predicted value and the observed value vector is calculated to obtain the posterior estimate of the exogenous noise term corresponding to the endogenous variable node.

[0117] Intervention operations were performed on the causal effect estimation network, and the forward propagation of the causal effect estimation network was rerun to obtain the counterfactual outcome values ​​corresponding to the spatiotemporal coordinates under intervention. The intervention operation corresponds to forcibly setting the values ​​of the treatment variable nodes corresponding to the target ecological restoration measures to the preset intervention values, cutting off all arrows pointing to the treatment variable nodes in the causal graph, and keeping the posterior estimates of the exogenous noise terms of other nodes unchanged.

[0118] By comparing the counterfactual outcome values ​​and observation vectors at the same spatiotemporal coordinates, the individual causal effect is calculated. The expected value of the individual causal effect is then calculated at each spatiotemporal coordinate covering the region to be evaluated, thus obtaining the causal effect value.

[0119] In one embodiment, the intelligent evaluation and analysis system 200 for ecological restoration effectiveness is also used for:

[0120] A causal training set is constructed based on the processing variable field, outcome variable, and covariate field. The causal effect estimation network is trained on the causal training set with the goal of minimizing the reconstruction error of the predicted and observed value vectors of each endogenous variable node until the convergence condition is met, thus obtaining the initial causal effect estimation network. The initial causal effect estimation network is used to characterize the nonlinear function mapping relationship between variables that conforms to the topological constraints of the causal graph.

[0121] Based on the causal graph, a set of sufficiently adjusted variables is obtained to block the backdoor path of the processing. With the goal of minimizing the statistical dependence between the processing variable node and the sufficiently adjusted variables, the initial causal effect estimation network is adversarially trained using the set of sufficiently adjusted variables to obtain a multimodal structural equation model.

[0122] In one embodiment, the intelligent evaluation and analysis system 200 for ecological restoration effectiveness is also used for:

[0123] A global causal loss function is constructed based on the causal effect value, and the gradient norm of each spatiotemporal coordinate is obtained by backpropagating the global causal loss function through automatic differentiation.

[0124] The gradient norm is used as a spatiotemporal coordinate to score the causal contribution of the causal inference conclusion.

[0125] Using the data source type of spatiotemporal coordinates, the interval between the corresponding timestamp and the current time, and the acquisition characteristics as input features, a data reliability evaluator based on a feedforward neural network is used to calculate the reliability score; the acquisition characteristics are one of the following: sensor calibration status information corresponding to the data source type, atmospheric condition parameters at the time of image acquisition, and investigator identification;

[0126] The causal contribution score and the confidence score are combined to obtain a double-weighted weight. The double-weighted weight is used to weight each loss value in the data fitting loss, so that the influence of training data with low confidence score or low causal contribution score in the training set is automatically suppressed during training.

[0127] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0128] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0129] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0130] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent evaluation and analysis of ecological restoration effectiveness, characterized in that, The method includes: Acquire multi-source heterogeneous ecological monitoring data covering the area to be evaluated, and discretize the multi-source heterogeneous ecological monitoring data according to the data source type to obtain a standardized sparse observation expression set; the standardized sparse observation expression set includes spatiotemporal coordinates and the observation value vector corresponding to the spatiotemporal coordinates; The spatiotemporal coordinates of each data point in the standardized sparse observation representation set are input into the spatiotemporal manifold mapping network and forward propagated to obtain the manifold coordinates of each data point in the observation vector in a unified high-dimensional manifold space; the spatiotemporal manifold mapping network includes a spatial coding module, a temporal coding module, and a gated fusion layer; The manifold coordinates are input into the ecological state decoder to obtain a multidimensional ecological state vector, and a spatiotemporally continuous ecological state field is constructed based on the multidimensional ecological state vector. A processing variable field is constructed based on the implementation records of ecological restoration measures, and outcome variables and covariates are extracted from the spatiotemporal continuous ecological state field. Using a hybrid algorithm based on constraints and scores, under the constraint of time-causal acyclicity, the causal relationships between the processing variable field, the outcome variable, and the covariates are analyzed to obtain a causal graph; Based on the causal graph, the processing variable field, the outcome variable, and the covariate field are input into the multimodal structural equation model. By applying intervention and fixing exogenous noise in the multimodal structural equation model, counterfactual inference is performed to quantify the causal effect values ​​of each ecological restoration measure on the ecological effectiveness. The multimodal structural equation model is composed of multiple deep network sub-modules connected according to the parent-child dependency relationship in the causal graph.

2. The method according to claim 1, characterized in that: The spatiotemporal coordinates include a first type, a second type, and a third type. The first type of spatiotemporal coordinates are obtained by extracting the sampling timestamps and sensor spatial locations from time-series data collected by IoT sensors. The second type of spatiotemporal coordinates are obtained by extracting the geographic coordinates of effective pixel centers and image acquisition timestamps from satellite remote sensing images after radiometric calibration and atmospheric correction. The third type of spatiotemporal coordinates are obtained by extracting the geographic coordinates of quadrat centers and survey timestamps from artificial sample plot survey data. The step of inputting each of the spatiotemporal coordinates in the standardized sparse observation representation set into a spatiotemporal manifold mapping network for forward propagation to obtain the manifold coordinates of each data point in the observation vector in a unified high-dimensional manifold space includes: The geographic coordinates in each of the spatiotemporal coordinates are spatially augmented using the digital elevation model covering the area to be evaluated to obtain augmented spatial coordinates. Based on the spatial coding module, the augmented spatial coordinates are mapped using a multi-scale learnable Gaussian radial basis function to obtain spatial embedding. Based on the time encoding module, the timestamps in each of the spatiotemporal coordinates are encoded and mapped using a sine and cosine encoding function based on adaptive learning frequency parameters to obtain the time embedding; The spatial embedding and the temporal embedding are interacted through the gated fusion layer to obtain the manifold coordinates corresponding to the spatiotemporal coordinates.

3. The method according to claim 2, characterized in that: The ecological state decoder and the spatiotemporal manifold mapping network are obtained through offline joint training; the offline joint training corresponds to the following steps: Construct the ecological state decoder based on a multi-task learning architecture; the multi-task learning architecture includes multiple independent task heads; Using the manifold coordinates output by the spatiotemporal manifold mapping network as a shared input, the soil moisture content, surface temperature, vegetation canopy height and normalized vegetation index are predicted by each of the independent task heads to obtain the multidimensional ecological state vector; Based on training focused on randomly sampled points in the entire spatiotemporal domain without observations, the derivative of the multidimensional ecological state vector with respect to time is calculated by automatic differentiation, and the residual between the derivative and the net flux is calculated to obtain the first physical loss term; the net flux is calculated through the ecosystem process equation. Based on the cross-variable empirical constraint relationship, the deviation between the relationship between the ecological state vectors of each dimension in the multidimensional ecological state vector and the known general function form is calculated to obtain the second physical loss term; The data fitting loss between the multidimensional ecological state vector and the corresponding observation vector, which are focused on the spatiotemporal coordinates with observations, is calculated, and the data fitting loss, the first physical loss term, and the second physical loss term are weighted and summed to obtain the total loss function. Based on the total loss function, the parameters of the spatiotemporal manifold mapping network and the ecological state decoder are updated synchronously through gradient backpropagation to obtain the spatiotemporal manifold mapping network and the ecological state decoder.

4. The method according to claim 1, characterized in that, The method employs a hybrid algorithm based on constraints and scores to analyze the pairwise causal relationships between the processing variable field, the outcome variable, and the covariates under time-causal acyclic constraints, thereby obtaining a causal graph, including: For any two variables among the processing variable field, the outcome variable, and the covariate, the empirical conditional mutual information partial derivative correlation statistic of the two variables in the regenerating kernel Hilbert space is calculated by using a kernel-based conditional independence test to obtain the skeleton connection between the variables. Based on the temporal causal acyclic constraint, causal directions are constructed according to the skeleton connection, and a smoothing prior is applied to the causal directions between spatially adjacent variable nodes to obtain the causal graph; the temporal causal acyclic constraint sets the temporal sequence as a hard constraint of the causal topology and prohibits the generation of causal directions from result to cause; the causal graph includes processing variable nodes, outcome variable nodes, and covariate nodes.

5. The method according to claim 4, characterized in that, Based on the causal graph, the treatment variable field, the outcome variable, and the covariate field are input into a multimodal structural equation model. Counterfactual inference is then performed by applying interventions and fixing exogenous noise in the multimodal structural equation model to quantify the causal effect values ​​of various ecological restoration measures on ecological outcomes, including: The nonlinear functions corresponding to each endogenous variable node in the multimodal structural equation model are instantiated as deep neural network sub-modules, and the inputs and outputs of each deep neural network sub-module are connected according to the parent-child dependency relationship in the causal graph to obtain the causal effect estimation network. The processing variable field, the outcome variable, and the covariate field are input into the causal effect estimation network to calculate the predicted value of each endogenous variable node under the observation path, and the error between the predicted value and the observed value vector is calculated to obtain the posterior estimate of the exogenous noise term corresponding to the endogenous variable node. An intervention operation is performed on the causal effect estimation network, and the forward propagation of the causal effect estimation network is rerun to obtain the counterfactual outcome value corresponding to the spatiotemporal coordinates under the intervention. The intervention operation corresponds to forcibly setting the value of the treatment variable node corresponding to the target ecological restoration measure to the preset intervention value, cutting off all arrows pointing to the treatment variable node in the causal graph, and keeping the posterior estimate value of the exogenous noise term of other nodes unchanged. The counterfactual outcome value and the observation vector at the same spatiotemporal coordinate are compared to calculate the individual causal effect. The expected value of the individual causal effect is then calculated at each spatiotemporal coordinate covering the region to be evaluated to obtain the causal effect value.

6. The method according to claim 5, characterized in that, The multimodal structure equation model is constructed through the following steps: A causal training set is constructed based on the processing variable field, the outcome variable, and the covariate field. The causal effect estimation network is trained using the causal training set with the goal of minimizing the reconstruction error of the predicted value and the observed value vector of each endogenous variable node until the convergence condition is met, thus obtaining an initial causal effect estimation network. The initial causal effect estimation network is used to characterize the nonlinear function mapping relationship between variables that conform to the topological constraints of the causal graph. The set of adequately adjusted variables for blocking backdoor paths is obtained from the causal graph. With the goal of minimizing the statistical dependency between the processing variable nodes and the adequately adjusted variables, the initial causal effect estimation network is adversarially trained using the set of adequately adjusted variables to obtain the multimodal structural equation model.

7. The method according to claim 3, characterized in that, The method further includes: A global causal loss function is constructed based on the causal effect value, and the gradient norm of each spatiotemporal coordinate is obtained by backpropagating the global causal loss function through automatic differentiation. The gradient norm is used as the causal contribution score of the spatiotemporal coordinates to the causal inference conclusion. Using the data source type, corresponding timestamp, and current time interval and acquisition characteristics of the spatiotemporal coordinates as input features, a confidence score is calculated through a data confidence evaluator based on a feedforward neural network; the acquisition characteristics are one of the following: sensor calibration status information corresponding to the data source type, atmospheric condition parameters during image acquisition, and investigator identification; The causal contribution score and the confidence score are fused to obtain a dual-weighted weight. The dual-weighted weight is used to weight each loss value in the data fitting loss, so that the influence of training data with low confidence scores or low causal contribution scores in the training set is automatically suppressed during training.

8. An intelligent evaluation and analysis system for the effectiveness of ecological restoration, characterized in that, The system includes: The data acquisition module is used to acquire multi-source heterogeneous ecological monitoring data covering the area to be evaluated, and to discretize the multi-source heterogeneous ecological monitoring data according to the data source type to obtain a standardized sparse observation expression set; the standardized sparse observation expression set includes spatiotemporal coordinates and the observation value vector corresponding to the spatiotemporal coordinates; The manifold coding module is used to input the spatiotemporal coordinates of each data point in the standardized sparse observation representation set into the spatiotemporal manifold mapping network for forward propagation, thereby obtaining the manifold coordinates of each data point in the observation vector in a unified high-dimensional manifold space; the spatiotemporal manifold mapping network includes a spatial coding module, a temporal coding module, and a gated fusion layer; The ecological state module is used to input the manifold coordinates into the ecological state decoder to obtain a multidimensional ecological state vector, and to construct a spatiotemporally continuous ecological state field based on the multidimensional ecological state vector. The restoration module is used to construct a processing variable field based on the implementation records of ecological restoration measures, and to extract outcome variables and covariates from the spatiotemporal continuous ecological state field. The causal analysis module is used to analyze the pairwise causal relationships between the processing variable field, the outcome variable, and the covariates under the constraint of time causality and acyclicity, using a hybrid algorithm based on constraints and scores, to obtain a causal graph; The effectiveness attribution evaluation module is used to input the treatment variable field, the outcome variable, and the covariate field into a multimodal structural equation model based on the causal graph. By applying interventions and fixing exogenous noise in the multimodal structural equation model, counterfactual inference is performed to quantify the causal effect values ​​of each ecological restoration measure on the ecological effectiveness. The multimodal structural equation model is composed of multiple deep network sub-modules connected according to the parent-child dependency relationship in the causal graph.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images