Method and device for dynamic estimation of carbon stock of an ecosystem

Abstract

The application provides a kind of ecosystem carbon physical quantity dynamic estimation method and device, comprising: obtaining the first modal environment perception data of target ecological entity and second modal spatial distribution data;The spatiotemporal feature extraction operation is executed to the first modal environment perception data, and the dynamic evolution characteristic vector is obtained;The boundary analysis operation is executed to the second modal spatial distribution data, and the static boundary topology matrix is obtained;Dynamic evolution characteristic vector is input into feature projection layer to execute dimension reduction processing, and generate feature projection sequence;Space mapping mechanism is constructed based on static boundary topology matrix;According to the space mapping mechanism, the space weight distribution operation is executed to feature projection sequence, and the weighted feature sequence is obtained;The multi-scale aggregation operation is executed to weighted feature sequence, and the fusion evaluation result of target ecological entity is generated.The application improves the higher objectivity and spatial feature adaptability of evaluation result.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more specifically, to a method and apparatus for dynamic estimation of ecosystem carbon content. Background Technology

[0002] With the continuous advancement of standardized accounting for the gross ecological product (GEP), the key to revitalizing dormant ecological assets through digital platforms and transforming them into data hubs with clear economic evaluation significance has become a crucial approach to addressing local fiscal dilemmas and promoting the transformation of relevant entities. Against this backdrop, it is necessary to conduct scientific and objective technical evaluations and data mapping of the natural rights and ecological function values ​​of ecological entities to support the digital realization mechanism of ecological product value and the construction of digital platforms.

[0003] Existing ecological entity evaluation and processing schemes typically employ a single-dimensional data overlay statistical processing architecture. This scheme first collects basic indicator signals of the ecological entity using sensors at fixed sites and directly aggregates them into an environmental index. Then, it extracts the boundary contour area of ​​the ecological entity based on a geographic information system and substitutes it into a pre-defined fixed conversion formula to calculate a static equity value. Finally, it performs a simple linear weighted summation of the environmental index and the static equity value to output the overall evaluation score of the ecological entity.

[0004] However, this processing method has obvious technical flaws. By directly calculating and linearly superimposing environmental signals and geographical areas independently, it severs the deep mapping relationship between the dynamic evolution of ecological entities in the continuous time dimension and their distribution topology in the spatial boundary dimension. This results in an inability to effectively reduce the dimensionality of high-dimensional and complex evolutionary features, and an inability to perform differentiated weight allocation on feature sequences in the spatial dimension. It also has weak multi-scale aggregation processing capabilities for feature sequences, making it difficult for the final output evaluation results to truly and accurately reflect the complex dynamic fusion features of ecological entities, leading to low objectivity in the final data evaluation results. Summary of the Invention

[0005] This application provides a method and apparatus for dynamic estimation of ecosystem carbon content, in order to at least alleviate the above-mentioned technical problems.

[0006] A method for dynamic estimation of ecosystem carbon biomass includes the following steps: Acquire first modal environmental perception data and second modal spatial distribution data of the target ecological entity. The first modal environmental perception data is used to characterize the environmental perception state of the target ecological entity in the continuous time dimension, and the second modal spatial distribution data is used to characterize the spatial distribution state of the target ecological entity in the spatial boundary dimension. Spatiotemporal feature extraction is performed on the first modal environmental perception data to obtain a dynamic evolution feature vector that characterizes the ecological function value of the target ecological entity. Perform boundary resolution operation on the second modality spatial distribution data to obtain a static boundary topology matrix that characterizes the natural rights of the target ecological entity; The dynamically evolved feature vector is input into a preset feature projection layer to perform dimensionality reduction processing, generating a feature projection sequence corresponding to the dynamically evolved feature vector; A spatial mapping mechanism corresponding to the static boundary topology matrix is ​​constructed based on the static boundary topology matrix. According to the spatial mapping mechanism, a spatial weight allocation operation is performed on the feature projection sequence to obtain a weighted feature sequence formed by the spatial mapping mechanism after applying it to the feature projection sequence; A multi-scale aggregation operation is performed on the weighted feature sequence to generate a fusion evaluation result of the target ecological entity. The fusion evaluation result is used to characterize the fusion evaluation result formed by the dynamic evolution feature vector and the static boundary topology matrix after being processed step by step by the feature projection sequence, the spatial mapping mechanism, and the weighted feature sequence.

[0007] Optionally, before the step of acquiring the first modal environmental perception data and the second modal spatial distribution data of the target ecological entity, the method further includes: Acquire initial raw acquisition signals uploaded by multiple sensing nodes deployed in the target ecological entity, wherein the initial raw acquisition signals carry sensing node identifiers and acquisition time markers corresponding to the multiple sensing nodes; Anomaly removal processing is performed on the initial raw acquisition signal to obtain a cleaned signal corresponding to the initial raw acquisition signal; A resampling operation is performed on the cleaned signal to generate aligned time series data, wherein the aligned time series data retains the correspondence between the sensing node identifier and the acquisition time marker; The aligned time-series data is used as the first modal environment perception data so that the first modal environment perception data can inherit the environment perception state in the initial raw acquisition signals uploaded by the multiple sensing nodes.

[0008] Optionally, the step of performing spatiotemporal feature extraction on the first modality of environmental perception data to obtain a dynamic evolution feature vector characterizing the ecological function value of the target ecological entity includes: The first modal environment perception data is divided into multiple consecutive time window data, and the multiple consecutive time window data are arranged in chronological order according to the first modal environment perception data. Perform a frequency domain transformation operation on each time window data to extract the frequency domain energy spectrum feature sequence corresponding to each time window data; The frequency domain energy spectrum feature sequences corresponding to multiple time window data are spliced ​​together in chronological order to construct the dynamic evolution feature vector corresponding to the first modal environment perception data.

[0009] Optionally, the step of performing boundary resolution on the second modal spatial distribution data to obtain a static boundary topology matrix characterizing the natural rights of the target ecological entity includes: The latitude and longitude coordinates of multiple key nodes are extracted from the second modality spatial distribution data. The multiple key nodes are used to characterize the boundary nodes of the target ecological entity in the spatial boundary dimension. Based on the latitude and longitude coordinate information, a connectivity analysis operation is performed on the multiple key nodes to determine the spatial adjacency relationship between the multiple key nodes; An undirected graph model is constructed based on the latitude and longitude coordinates and the spatial adjacency relationship. The undirected graph model is used to represent the multiple key nodes and the spatial adjacency relationship between the multiple key nodes. The adjacency matrix of the undirected graph model is extracted as the static boundary topology matrix, so that the static boundary topology matrix can carry the latitude and longitude coordinate information and spatial adjacency relationship in the second modality spatial distribution data.

[0010] Optionally, the step of inputting the dynamically evolved feature vector into a preset feature projection layer to perform dimensionality reduction processing and generate a feature projection sequence corresponding to the dynamically evolved feature vector includes: Obtain the preset weight matrix, preset bias vector, and nonlinear activation component of the feature projection layer; The dynamic evolution feature vector is multiplied by the preset weight matrix to obtain an intermediate feature vector; The preset bias vector and the intermediate feature vector are added together to obtain the biased vector. The nonlinear activation component is used to perform a nonlinear transformation operation on the biased vector to obtain the feature projection sequence. The feature projection sequence is the dimensionality reduction representation of the dynamically evolved feature vector after processing by the preset weight matrix, the preset bias vector, and the nonlinear activation component.

[0011] Optionally, the step of constructing a spatial mapping mechanism corresponding to the static boundary topology matrix based on the static boundary topology matrix includes: Perform eigenvalue decomposition on the static boundary topology matrix to extract the principal eigenvector corresponding to the largest eigenvalue in the static boundary topology matrix; The main feature vector is normalized to obtain a normalized feature vector; The normalized feature vector is used as the attention weight vector, and the weight mapping rule containing the attention weight vector is used as the spatial mapping mechanism, so that the spatial mapping mechanism can take over the spatial boundary weight relationship corresponding to the static boundary topology matrix.

[0012] Optionally, the step of performing a spatial weight allocation operation on the feature projection sequence according to the spatial mapping mechanism to obtain a weighted feature sequence formed by the spatial mapping mechanism after applying it to the feature projection sequence includes: Based on the attention weight vector in the spatial mapping mechanism, determine the attention weight value corresponding to each feature element in the feature projection sequence; Each feature element in the feature projection sequence is multiplied element-wise with its corresponding attention weight value to generate a mapped feature sequence. Obtain the preset smoothing filter component; The pre-defined smoothing filter component is used to perform smoothing processing on the mapped feature sequence. The processed mapped feature sequence is then used as the weighted feature sequence, so that the weighted feature sequence simultaneously incorporates the spatial weight allocation results corresponding to the feature projection sequence and the attention weight vector.

[0013] Optionally, the step of performing multi-scale aggregation on the weighted feature sequence to generate the fusion evaluation result of the target ecological entity includes: Obtain a first sliding window of a first size and a second sliding window of a second size, wherein the second size is larger than the first size; The first sliding window is used to perform local feature extraction on the weighted feature sequence to obtain a first-scale feature sequence. The second sliding window is used to perform local feature extraction on the weighted feature sequence to obtain a second-scale feature sequence. Perform a dimension concatenation operation on the first-scale feature sequence and the second-scale feature sequence to obtain the full-scale feature tensor; A numerical mapping operation is performed on the full-scale feature tensor to map the full-scale feature tensor into the fusion evaluation result in a single numerical form, so that the fusion evaluation result can inherit the multi-scale feature aggregation result formed by the weighted feature sequence under the first sliding window and the second sliding window.

[0014] Optionally, after the step of generating the fusion evaluation result of the target ecological entity, the method further includes: A preset benchmark reference threshold is obtained, which is used to perform a numerical comparison with the fusion evaluation result; The fusion evaluation result is compared with the preset benchmark reference threshold to obtain the comparison result between the fusion evaluation result and the preset benchmark reference threshold; If the comparison result indicates that the fusion evaluation result is lower than the preset benchmark reference threshold, then an environmental scheduling instruction sequence is generated, and the environmental scheduling instruction sequence corresponds to the comparison result where the fusion evaluation result is lower than the preset benchmark reference threshold; The environmental scheduling instruction sequence is sent to the execution terminal that is communicatively connected to the target ecological entity, so as to control the execution terminal to trigger the entity environment adjustment operation according to the environmental scheduling instruction sequence. The entity environment adjustment operation is the entity environment adjustment operation performed by the execution terminal after responding to the environmental scheduling instruction sequence.

[0015] A dynamic estimation device for ecosystem carbon physical quantities includes: An ecological entity multimodal data acquisition module is used to acquire first modal environmental perception data and second modal spatial distribution data of a target ecological entity. The first modal environmental perception data is used to characterize the environmental perception state of the target ecological entity in the continuous time dimension, and the second modal spatial distribution data is used to characterize the spatial distribution state of the target ecological entity in the spatial boundary dimension. The dynamic evolution feature vector generation module is used to perform spatiotemporal feature extraction operations on the first modal environmental perception data to obtain a dynamic evolution feature vector that characterizes the ecological function value of the target ecological entity. The static boundary topology matrix generation module is used to perform boundary parsing operations on the second modal spatial distribution data to obtain a static boundary topology matrix that characterizes the natural rights and interests of the target ecological entity. The feature projection sequence generation module is used to obtain a pre-configured feature projection layer, and input the dynamically evolved feature vector into the pre-configured feature projection layer to perform dimensionality reduction processing, thereby generating a feature projection sequence corresponding to the dynamically evolved feature vector; A spatial mapping mechanism construction module is used to construct a spatial mapping mechanism corresponding to the static boundary topology matrix based on the static boundary topology matrix. The weighted feature sequence generation module is used to perform a spatial weight allocation operation on the feature projection sequence according to the spatial mapping mechanism to obtain a weighted feature sequence formed by the spatial mapping mechanism after applying it to the feature projection sequence. The fusion evaluation result generation module is used to perform multi-scale aggregation operation on the weighted feature sequence to generate the fusion evaluation result of the target ecological entity. The fusion evaluation result is used to characterize the fusion evaluation result formed by the dynamic evolution feature vector and the static boundary topology matrix after being processed step by step by the feature projection sequence, the spatial mapping mechanism and the weighted feature sequence.

[0016] The technical advantages of the technical solution provided in this application are: This application presents a method and apparatus for dynamic estimation of ecosystem carbon content. Addressing the shortcomings of traditional single-data overlay statistical architectures that directly calculate and linearly superimpose environmental signals and geographical areas, severing the connection between the dynamic evolution of ecological entities in the continuous time dimension and their topological structure in the spatial boundary dimension, this method extracts dynamic evolution feature vectors from first-modal environmental perception data through spatiotemporal feature extraction and obtains static boundary topological matrices from second-modal spatial distribution data through boundary resolution. This solves the feature fragmentation problem caused by isolated data dimensions in traditional schemes. Compared to the direct numerical aggregation of traditional schemes, this application transforms first-modal and second-modal data into feature vectors and topological matrices, providing a stronger computational foundation for subsequent deep mapping and correlation of data.

[0017] Based on dynamically evolving feature vectors and static boundary topology matrices, this application generates feature projection sequences by inputting the dynamically evolving feature vectors into a preset feature projection layer for dimensionality reduction. A spatial mapping mechanism is then constructed based on the static boundary topology matrix. Following this mechanism, a spatial weight allocation operation is performed on the feature projection sequences to obtain a weighted feature sequence. This effectively solves the technical deficiency of traditional schemes in being unable to reduce the dimensionality of high-dimensional evolving features and perform differentiated weight allocation in the spatial dimension. Traditional schemes mainly rely on fixed conversion formulas to calculate single values, while this application utilizes a spatial mapping mechanism directly applied to the dimensionality-reduced feature projection sequences. This allows dynamic features in the continuous time dimension to keenly perceive the weight distribution of spatial boundary features. Compared to the traditional fixed and undifferentiated calculation mode, this approach achieves a higher degree of integration in the spatial allocation of feature weights and more accurate data mapping.

[0018] Finally, by performing multi-scale aggregation operations on the weighted feature sequences to generate a fusion evaluation result for the target ecological entity, this approach addresses the technical deficiency of traditional schemes, which suffer from weak multi-scale aggregation capabilities and consequently low objectivity. Traditional schemes often only provide simple single-level summation outputs, while this application performs hierarchical multi-scale feature aggregation on the weighted feature sequences that incorporate spatial feature weights. This allows the generated fusion evaluation result to comprehensively reflect the complex interaction between dynamic evolution and static boundaries. Compared to the coarse-grained linear superposition scores output by traditional methods, the fusion evaluation result output by this application better maps the objective dynamic state of the ecological entity, significantly improving the overall reliability and objectivity of the technical evaluation. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating an application scenario of a dynamic estimation method for ecosystem carbon content according to an embodiment of this application.

[0020] Figure 2 This application provides a method for dynamically estimating the physical amount of carbon in an ecosystem.

[0021] Figure 3 This application provides an embodiment of a dynamic estimation device for the amount of carbon in an ecosystem.

[0022] Figure 4 This is an electronic device according to an embodiment of the present application. Detailed Implementation

[0023] like Figure 1 The diagram shown is a schematic representation of a dynamic estimation scenario for ecosystem carbon content according to an embodiment of this application. Figure 2 The image shows an embodiment of a dynamic estimation method for ecosystem carbon content, which includes the following steps: A multi-source heterogeneous sensing data stream of the target ecological area is acquired, and multi-dimensional spatiotemporal feature extraction processing is performed on the multi-source heterogeneous sensing data stream to obtain a multi-modal spatiotemporal feature matrix; The multimodal spatiotemporal feature matrix is ​​mapped onto a pre-constructed ecological topology hypergraph to generate a dynamic hypergraph structure with node feature states and hyperedge weights. The ecological topology hypergraph includes nodes representing single environmental elements and hyperedges representing the collaborative coupling relationships of multiple environmental elements. The node feature states are derived from the feature assignment results of the multimodal spatiotemporal feature matrix on the nodes, and the hyperedge weights are derived from the correlation characterization results of the multimodal spatiotemporal feature matrix on the collaborative coupling relationships of the multiple environmental elements. A hyperedge message passing mechanism is executed on the dynamic hypergraph structure to iteratively update the node feature state and the hyperedge weight based on the node feature state and the hyperedge weight, and output the converged collaborative topology feature sequence. The co-topological feature sequence is subjected to manifold dimensionality reduction projection processing by a pre-trained spatiotemporal manifold decoder to generate the current periodic carbon quantity estimation data corresponding to the target ecological region.

[0024] Optionally, the step of performing multi-dimensional spatiotemporal feature extraction processing on the multi-source heterogeneous sensing data stream to obtain a multimodal spatiotemporal feature matrix includes: The multi-source heterogeneous sensing data stream is spatially gridded according to a preset spatial grid granularity to obtain spatial sub-block data. Extract remote sensing spectral features, meteorological sequence features, and soil micro-sensing features from the data of each spatial sub-block; The remote sensing spectral features, meteorological sequence features, and soil microsensing features are spliced ​​together in a time sequence according to a preset time sliding window to obtain the multimodal spatiotemporal feature matrix.

[0025] Preferably, when performing multi-dimensional spatiotemporal feature extraction processing on the multi-source heterogeneous sensing data stream, the multi-source heterogeneous sensing data stream is first spatially gridded according to a preset spatial grid granularity. Here, the multi-source heterogeneous sensing data stream refers to a set of initial sensing signals uploaded in real time by multiple sensing nodes deployed in the target ecological entity, which differ in acquisition principles, data formats, time frequencies, and spatial references. Examples include pixel-level reflectance data from satellite remote sensing images, hourly observation data from ground meteorological stations, and profile sensing data from wireless sensor networks buried in the soil. The preset spatial grid granularity is a predefined spatial segmentation reference based on the spatial extension scale of the target ecological entity and the inherent spatial resolution of the aforementioned types of sensing data. For example, when the nadir resolution of the remote sensing image is 30 meters, the average control radius of the meteorological station is 200 meters, and the soil sensor deployment density is one point per hectare, the spatial grid granularity can be set to a square grid with a side length of 100 meters. During the segmentation process, the area within the geographical boundary of the target ecological entity is divided into non-overlapping spatial sub-blocks according to the granularity of the square spatial grid. Each spatial sub-block has a definite central latitude and longitude coordinate and a unique block number. The geographical location information carried in each multi-source heterogeneous sensing data stream is mapped to the spatial sub-block corresponding to its spatial location, thereby organizing the entire data stream into spatial sub-block data that corresponds one-to-one with each spatial sub-block.

[0026] Preferably, after obtaining the data for each spatial sub-block, remote sensing spectral features are extracted. For satellite remote sensing image data belonging to a certain spatial sub-block, the surface reflectance values ​​of that pixel in each spectral band are extracted, such as reflectance values ​​in the blue, green, red, near-infrared, and shortwave infrared bands. Based on these band reflectance values, at least one vegetation index is calculated, such as the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EDI). This vegetation index can sensitively reflect the proportion of photosynthetically active radiation absorbed by vegetation and has a physical correlation with the primary productivity and carbon sink intensity of the ecosystem. The remote sensing spectral features are constructed into a multi-dimensional vector, where each dimension corresponds to the numerical representation of the spatial sub-block in a certain spectral band or vegetation index. This allows the vector to characterize the growth vitality and biochemical composition state of vegetation within the spatial sub-block from the perspective of canopy spectral response, providing a characteristic basis for subsequent capture of the evolutionary field of carbon content at the optical property level.

[0027] Preferably, the data of the same spatial sub-block includes the source of meteorological sequence features. Based on long-term observation records of meteorological stations deployed near or within the target ecological entity, meteorological elements such as station-level temperature, precipitation, photosynthetically active radiation, and saturated vapor pressure difference are extrapolated to the central coordinates of the spatial sub-block using spatial interpolation algorithms, and meteorological sequence features of the spatial sub-block are formed according to a uniform time step. This meteorological sequence feature is represented as a matrix or vector set that changes over time, where the rows correspond to continuous time sampling points, the columns correspond to different meteorological variables, and the elements at the intersection of rows and columns represent the specific values ​​of a certain meteorological variable for the spatial sub-block at a certain moment. Meteorological conditions directly regulate the photosynthetic carbon sequestration process of vegetation and the soil respiration release process. Therefore, embedding meteorological sequence features into a multimodal feature matrix allows environmental driving information in the time dimension to be completely preserved in the subsequent dynamic evolution feature extraction process.

[0028] Preferably, soil micro-sensing features are also extracted from the data of each spatial sub-block. For points within a spatial sub-block equipped with soil sensing probes, time-series sensing variables such as soil temperature, soil volumetric water content, soil electrical conductivity, and soil organic carbon content retrieved using near-infrared spectroscopy are directly acquired. For spatial sub-blocks without sensing probes, representative values ​​are generated using the sensing values ​​from nearby probes and spatial autocorrelation weighting or geostatistical interpolation methods. The soil micro-sensing features also form a time-series structure, with the values ​​at each time point reflecting the physicochemical properties closely related to carbon accumulation and turnover in the soil microenvironment of that spatial sub-block. This is a key information source for describing the potential for carbon quantity changes in target ecological entities from the perspective of underground processes.

[0029] Preferably, after extracting remote sensing spectral features, meteorological sequence features, and soil microsensor features respectively, a temporal splicing operation is performed on the three according to a preset time sliding window. The preset time sliding window refers to a time interval division strategy with a fixed time length and sliding step size. For example, the time window length is 7 consecutive days of sensing data, and the sliding step size is 1 day, that is, two adjacent windows overlap for 6 days on the time axis. For each spatial sub-block, the remote sensing spectral feature vector, meteorological sequence feature vector, and soil microsensor feature vector falling within the same time window are spliced ​​end to end in the feature dimension to form a fused feature vector representing the overall environmental state of the spatial sub-block within the time window. As the time sliding window slides forward along the time axis, a series of fused feature vectors arranged in chronological order are generated for the spatial sub-block, constituting the temporal feature fragment of the spatial sub-block. Before stitching, the remote sensing spectral features, meteorological sequence features, and soil microsensor features were normalized to map the values ​​of different physical dimensions to a unified numerical range, so as to eliminate the unwanted bias interference caused by the difference in dimensions to the subsequent dimensionality reduction projection and spatial weight allocation.

[0030] Preferably, the temporal feature fragments corresponding to each spatial sub-block are regarded as perceptual expression entities spanning both spatial and temporal dimensions. Based on the spatial adjacency relationships and block numbers of the spatial sub-blocks within the geographical boundaries of the target ecological entity, spatial dimensions are stacked and combined to form a multimodal spatiotemporal feature matrix. Structurally, this multimodal spatiotemporal feature matrix is ​​a multidimensional array. Its first dimension represents the continuous time window index, the second dimension represents the spatial sub-block index, and the third dimension represents the fused feature channels. Any element in the array corresponds to the value of a certain fused feature component within a specific time window and spatial sub-block. Thus, through spatial gridding, multimodal feature extraction, and temporal sliding window stitching, a multimodal spatiotemporal feature matrix capable of simultaneously carrying the deep mapping relationship of the target ecological entity in both temporal evolution and spatial distribution dimensions is generated from multi-source heterogeneous perceptual data streams. The multimodal spatiotemporal feature matrix, as the first modality of environmental perception data, is directly input into the subsequent spatiotemporal feature extraction operation, namely the generation step of the dynamic evolution feature vector. This enables the subsequent feature projection and spatial weight allocation to obtain basic data representations that combine temporal dynamics and spatial heterogeneity, thereby providing a reliable source of perception information to ensure that the final fusion evaluation results objectively reflect the complex dynamic fusion characteristics of the carbon physical quantity of ecological entities.

[0031] Optionally, the step of mapping the multimodal spatiotemporal feature matrix onto a pre-constructed ecological topology hypergraph to generate a dynamic hypergraph structure with node feature states and hyperedge weights includes: The target node in the ecological topology hypergraph is determined based on the feature dimension labels in the multimodal spatiotemporal feature matrix; Calculate the dynamic correlation coefficient of the target node within a preset physical time window; Determine whether the dynamic correlation coefficient reaches a preset coupling threshold; When the dynamic correlation coefficient reaches the preset coupling threshold, the hyperedge is established between the target nodes, and the initial hyperedge weight corresponding to the hyperedge is determined according to the dynamic correlation coefficient. The target node's node feature state is initialized using the multimodal spatiotemporal feature matrix, and the target node, node feature state, hyperedge, and initial hyperedge weight are written into the ecological topology hypergraph to generate the dynamic hypergraph structure.

[0032] Preferably, in this step, the target node corresponding to the pre-constructed ecological topology hypergraph needs to be determined first based on the feature dimension labels in the multimodal spatiotemporal feature matrix. The ecological topology hypergraph is a topological data structure, distinct from ordinary binary graphs, used to characterize the multi-level ecological process interactions within a target ecological entity. In this ecological topology hypergraph, each target node no longer merely represents a spatial location, but is abstracted as a logical container corresponding to a certain type of feature dimension with a clear physical meaning in the multimodal spatiotemporal feature matrix. The multimodal spatiotemporal feature matrix has been constructed as a multidimensional array in previous steps. Its third dimension is the fused feature channel. Each feature channel has a feature dimension label closely related to its source. For example, the normalized difference vegetation index calculated based on blue light band reflectance constitutes one feature channel, and its corresponding feature dimension label can be defined as the vegetation photosynthetic activity index; the photosynthetically active radiation time series obtained based on spatial interpolation constitutes another feature channel, and its corresponding feature dimension label can be defined as a canopy radiation driving factor; the feature channel based on soil volumetric water content time series data can have its corresponding feature dimension label defined as a soil moisture state factor. When determining the target node, all feature dimension labels in the multimodal spatiotemporal feature matrix are traversed. A target node uniquely corresponding to each feature dimension label is located or dynamically created in the ecological topology hypergraph, thereby ensuring that each feature dimension participating in subsequent fusion and deduction can obtain an independent representation position in the topological structure. The technical significance of this processing action lies in the fact that it transforms the multimodal spatiotemporal feature matrix from a flattened representation based on data arrays to a topological representation based on nodes as representation units, laying a structural foundation for subsequent capture of dynamic interactions across modalities and processes.

[0033] Preferably, after determining the target nodes corresponding to the ecological topology hypergraph, the dynamic correlation coefficients of these target nodes within a preset physical time window are calculated. Here, the preset physical time window is a time segment with a clear physical meaning and a fixed duration. It is consistent with or related to the preset time sliding window used to generate the multimodal spatiotemporal feature matrix. For example, a physical time window can correspond to a heartbeat-like sliding data segment spanning seven consecutive days. For any two target nodes, continuous feature value sequences spanning all spatial sub-blocks within the current physical time window are extracted from the multimodal spatiotemporal feature matrix data blocks indexed by their corresponding feature dimension labels. Subsequently, a designed dynamic correlation analyzer is used to calculate the dynamic correlation coefficient between these two feature value sequences. This dynamic correlation analyzer differs from conventional overall correlation coefficient calculations. It first uses an adaptive time delay probe operator to perform a time-step hysteresis scan of the two feature value sequences to identify the time propagation delay of one feature sequence relative to the other and aligns this propagation delay. Next, after time-delay alignment, the ratio of the product of the covariance and standard deviation of the two sequences in the sliding sub-segment is calculated, resulting in a real number between positive and negative 1, which is the dynamic correlation coefficient. The absolute value of this dynamic correlation coefficient directly quantifies the interaction strength of the ecophysical processes represented by the two target nodes under the current environmental driving force. For example, the absolute value of the dynamic correlation coefficient between the target node representing soil moisture state factors and the target node representing canopy radiation driving factors may rise to nearly 0.8 within the physical time window of the dry season, reflecting the strong coupling control of water stress on light energy utilization efficiency; while within the water-sufficient window, the absolute value of the coefficient may drop to around 0.3, reflecting the relaxation of coupling. This time-delay-aware dynamic coefficient calculation method provides a data-driven, high-resolution quantification means for keenly capturing the complex material and energy transfer relationships within the ecosystem in the time dimension.

[0034] Preferably, after calculating the dynamic correlation coefficient between any two target nodes within the preset physical time window, a judgment operation is immediately performed to determine whether the dynamic correlation coefficient reaches a preset coupling threshold. The preset coupling threshold is a pre-configured numerical limit used to define whether there is a substantial ecological process coupling relationship between nodes. This limit avoids incorporating weak, spurious correlations caused by data noise into the topology. For example, the preset coupling threshold can be set to a real number with an absolute value greater than 0.5. When performing the judgment operation, if the absolute value of the dynamic correlation coefficient is lower than or equal to the preset coupling threshold, it is determined that the two target nodes have not formed an effective ecological process coupling within the current physical time window, and the hyperedge establishment action is not performed; conversely, once the absolute value of the dynamic correlation coefficient is higher than the preset coupling threshold, the subsequent hyperedge construction process is triggered. The technical advantage of this threshold screening mechanism is that it makes the final generated dynamic hypergraph structure not a fixed fully connected network, but a sparse topology that evolves dynamically with changes in environmental driving conditions. Only connections with strong physical interactions are retained, which is highly consistent with the objective law that the carbon cycle process in the target ecological entity is non-stationary in time and non-uniform in space.

[0035] Preferably, when the dynamic correlation coefficient reaches the preset coupling threshold, a hyperedge can be established between the two target nodes that meet the condition. In a hypergraph structure, a hyperedge can connect two or more target nodes, which differs from the limitation in ordinary graph theory that an edge can only connect two vertices. Therefore, if there are three or more target nodes within the same physical time window, and the dynamic correlation coefficients between each pair of them are all higher than the preset coupling threshold, a single hyperedge spanning multiple nodes can be constructed between these target nodes to characterize a high-order functional module involving the collaborative participation of multidimensional ecological processes, such as a water and carbon flux joint regulation unit formed by the coupling of leaf photosynthesis, soil water supply, and atmospheric transpiration pull. Simultaneously with establishing the hyperedge, the initial hyperedge weight corresponding to the hyperedge needs to be determined based on the dynamic correlation coefficient. The initial hyperedge weight is a aggregate quantification of the overall coupling strength between all target nodes connected by this hyperedge. The initial hyperedge weight can be determined by calculating the arithmetic mean or root mean square of the absolute values ​​of the dynamic correlation coefficients between all pairs of target nodes connected by the hyperedge, and using this aggregated value as the initial hyperedge weight. This initial hyperedge weight is a non-negative real number, and its value objectively reflects the compactness of the interaction within the functional module, providing a differentiated basis for allocating transmission capabilities for subsequent information propagation and feature updates on the hypergraph.

[0036] Preferably, after constructing the hyperedge and determining its initial weight, the node feature state of the target node is initialized using the multimodal spatiotemporal feature matrix. The node feature state is a real-valued vector attached to each target node, representing the overall activation level of that node within the current physical time window. The specific execution method of the initialization assignment operation is as follows: for each target node, the feature data of the multimodal spatiotemporal feature matrix across all spatial sub-blocks within the current physical time window is indexed according to its feature dimension label. A spatial aggregation operation is performed on the feature data of these spatial sub-blocks, for example, calculating the mean or median of the feature channel values ​​for all spatial sub-blocks. The aggregated value or a set of values ​​is then used as the initial value of the node feature state of the target node. Through this step, the environmental perception information originally scattered across various spatial sub-blocks and carried by the multimodal spatiotemporal feature matrix is ​​summarized and loaded onto each target node of the ecological topology hypergraph, giving the topological structure a specific physical state connotation.

[0037] Preferably, after completing all the above construction and assignment actions, the target node, the node feature state, the hyperedge, and the initial hyperedge weight are jointly written into the ecological topology hypergraph to generate the dynamic hypergraph structure. The specific writing actions include: in the data structure of the ecological topology hypergraph, creating a node object for the target node and storing its unique node identifier; assigning the initialized node feature state vector to the state attribute of the node object; creating a hyperedge object for each newly established hyperedge, recording the identifiers of all target nodes constituting the hyperedge in its member list, and assigning the calculated initial hyperedge weight to the weight attribute of the hyperedge object. The resulting dynamic hypergraph structure, whose internal topological connections and node states are driven by the multimodal spatiotemporal feature matrix of the current physical time window, is completely output as a graph data snapshot carrying spatiotemporal dynamic semantics to subsequent graph evolution processing steps. Compared to traditional approaches that treat perceived data from different physical processes as isolated dimensions and simply sum them, this generated dynamic hypergraph structure, for the first time at the data structure level, clearly, objectively, and analytically expresses the complex ecological process topology network coupled with time lag effects, higher-order interactions, and dynamic sparse connections. This provides a highly organized data foundation for subsequent dynamic feature calculations and final multi-scale aggregation evaluation on this network.

[0038] Optionally, the step of executing a hyperedge message passing mechanism on the dynamic hypergraph structure to iteratively update the node feature states and hyperedge weights based on the node feature states and the hyperedge weights, and outputting a converged cooperative topological feature sequence includes: Extract the set of associated nodes contained in each hyperedge in the dynamic hypergraph structure and the hyperedge weight corresponding to each hyperedge; Based on the hyperedge weights corresponding to each hyperedge, the node feature states in the set of associated nodes are aggregated to obtain a hyperedge state representation vector. Based on the hyperedge state representation vector, a reverse feature distribution operation is performed on the feature states of each node in the set of associated nodes to obtain the updated node feature states. Based on the matching relationship between the updated node feature state and the hyperedge state representation vector, a weight correction operation is performed on the hyperedge weights corresponding to each hyperedge to obtain the updated hyperedge weights. Determine whether the feature change rate of the updated node feature state and the weight change rate of the updated hyperedge weight are both lower than a preset convergence threshold; When both the feature change rate and the weight change rate are lower than the preset convergence threshold, the updated node feature state and the updated hyperedge weight are concatenated into the collaborative topology feature sequence according to the preset sequence concatenation rules.

[0039] Preferably, in the specific implementation of this step, the dynamic hypergraph structure is first used as input to extract the set of associated nodes contained in each hyperedge and the corresponding hyperedge weights. Here, the dynamic hypergraph structure is generated by mapping a multimodal spatiotemporal feature matrix to a pre-constructed ecological topology hypergraph. It contains multiple target nodes and multiple hyperedges, each hyperedge connecting at least two target nodes and carrying an initial hyperedge weight. The set of associated nodes refers to the aggregation of the indices or identifiers of all target nodes connected by a certain hyperedge. This set clearly defines all procedural units participating in the same high-order ecological function coupling module. For example, a hyperedge representing a water and carbon flux joint regulation unit may contain canopy radiation driving factor nodes, soil moisture state factor nodes, and atmospheric transpiration pull factor nodes in its set of associated nodes. The hyperedge weight is a value that dynamically evolves during the iterative update process, starting from the initial hyperedge weight corresponding to the hyperedge. The initial hyperedge weight is directly used in the first iteration, while subsequent iterations use the updated hyperedge weight output from the previous cycle. This extraction process breaks down the complex hypergraph structure into message passing units centered on hyperedges, providing an independent and parallel processing granularity for subsequent information aggregation and distribution along the hyperedge.

[0040] Preferably, after obtaining the set of associated nodes and their weights for each hyperedge, the node feature states in the set of associated nodes are aggregated based on the corresponding hyperedge weights to obtain the hyperedge state representation vector. The node feature states are real-valued vectors attached to each target node, initialized using a multimodal spatiotemporal feature matrix during the generation of the dynamic hypergraph structure. The aggregation process is as follows: for the current hyperedge, traverse each target node in its set of associated nodes, extract the current node feature state vector of that target node, and use the hyperedge weight as a scaling factor to perform a weighted sum of all extracted vectors. A nonlinear activation transformation can be added if necessary. After the weighted summation, the result is normalized by considering the number of nodes in the set of associated nodes to avoid drift in representation amplitude due to differences in the number of nodes. The final generated hyperedge state representation vector has the same dimension as the node feature states, and each element's value integrates the comprehensive information of all ecological process states associated with the hyperedge under the regulation of the hyperedge weights. This hyperedge state representation vector, from the perspective of functional modules, extracts the overall operational level of processes such as canopy photosynthesis, soil respiration, and water transport under synergistic effects.

[0041] Preferably, after calculating the hyperedge state representation vector of a certain hyperedge, a reverse feature distribution operation is then performed on the feature states of each node in the associated node set based on the hyperedge state representation vector to obtain the updated node feature states. The purpose of the reverse feature distribution operation is to use the overall state aggregated by the hyperedge to correct and enrich the independent representations of each component node within the hyperedge, so that each node can "perceive" the overall operational status of its functional module. Specifically, for each target node in the associated node set, its current node feature state vector is fused with the newly generated hyperedge state representation vector. The fusion method can be to design a gating regulator, which takes the concatenated node feature state vector and the hyperedge state representation vector as input, processes them through a learnable gating parameter matrix and bias vector, and outputs a gating coefficient vector of the same dimension as the node feature state through an S-shaped growth curve function. The gating coefficient vector is multiplied element-wise with the hyperedge state representation vector, and then summed and compensated with the original node feature state vector to generate the updated node feature state of the node. For target nodes spanning multiple hyperedges, the updates obtained from each associated hyperedge are averaged or superimposed to summarize feedback signals from different functional modules. Through this reverse feature distribution operation, the updated node feature state has internalized the state information at the functional module level, and compared to before the update, it can better reflect the synergy and constraints between various driving factors in the process of ecosystem carbon biomass change.

[0042] Preferably, after updating the node feature states, a weight correction operation is performed on the hyperedge weights corresponding to each hyperedge based on the matching relationship between the updated node feature states and the hyperedge state representation vectors, to obtain the updated hyperedge weights. The matching relationship is used to measure the consistency or closeness between the node feature states of nodes within the hyperedge and the hyperedge aggregate state after the update. A specific matching relationship measurement method is as follows: The hyperedge state representation vector of the hyperedge is multiplied by the updated node feature states of each target node in the associated node set. The average of the obtained inner product values ​​is calculated, and the absolute value is taken to obtain a non-negative matching score. The higher the matching score, the more coordinated the overall state aggregated by the hyperedge is with the independent states of each member node, indicating a robust internal coupling of the associated functional modules; conversely, a lower score suggests that the current composition of the functional modules may be drifting, requiring weight adjustment. The weight correction operation uses this matching score, combined with a pre-set weight adjustment step size factor, to update the hyperedge weights. For example, an exponential moving average method is used, making the updated hyperedge weight equal to the weighted combination of the current hyperedge weight and the matching score. After this operation, the updated hyperedge weights can be adaptively adjusted according to the degree of coordination of node state evolution, so that hyperedges with more compact internal interactions can enhance their information transmission capabilities in subsequent iterations, while hyperedges with weakened internal consistency are gradually suppressed, thus objectively realizing the adaptive graph evolution of the structure.

[0043] Preferably, after a complete round of node state update and hyperedge weight update is completed, iterative convergence determination is performed. First, the feature change rate of the updated node feature state relative to the node feature state before the update, and the weight change rate of the updated hyperedge weight relative to the hyperedge weight before the update are calculated. The feature change rate is calculated by, for each target node, calculating the ratio of the norm of the difference between its node feature state vector before and after the update to the norm of the vector before the update, and then taking the arithmetic mean over all target nodes. The weight change rate is calculated by, for each hyperedge, calculating the ratio of the absolute value of the difference between its hyperedge weight before and after the update to the absolute value of the original weight, and then taking the arithmetic mean over all hyperedges. A node convergence threshold for the node feature state and a weight convergence threshold for the hyperedge weight are predefined; their values ​​can be different, for example, the node convergence threshold is set to 0.01 and the weight convergence threshold is set to 0.005. When the feature change rate is lower than the node convergence threshold and the weight change rate is also lower than the weight convergence threshold, it is determined that the message passing and evolution on the dynamic hypergraph structure has reached steady-state convergence. At this time, the updated node feature state and the updated hyperedge weight are the final values ​​in the convergence state. If either of the change rates is not lower than its corresponding threshold, the updated node feature state and the updated hyperedge weight are used as the input for a new round of iteration. The step of extracting the hyperedge weight and the set of associated nodes is returned, and the next round of hyperedge message passing process is started until both types of change rates simultaneously meet the convergence condition.

[0044] Preferably, after the iteration convergence is determined, the updated node feature states and updated hyperedge weights are concatenated into a collaborative topological feature sequence according to a preset sequence concatenation rule. The preset sequence concatenation rule is as follows: First, based on the node index order of the target nodes in the dynamic hypergraph structure, the finally converged node feature state vectors of each target node are sequentially concatenated end-to-end to form a one-dimensional node feature long vector; then, based on the hyperedge numbering order of the hyperedges in the dynamic hypergraph structure, the finally converged hyperedge weight values ​​of each hyperedge are sequentially arranged to form a hyperedge weight sequence vector; finally, the node feature long vector and the hyperedge weight sequence vector are concatenated end-to-end again, and the resulting total long vector is the collaborative topological feature sequence. This collaborative topological feature sequence, in the form of fixed-dimensional vectors, condenses the topological state and node attributes of the dynamic hypergraph structure after sufficient information interaction. It contains not only the relative activity levels of various ecological processes within the target ecological entity in the functional modules, but also preserves the distribution of coupling strength between these functional modules. This provides a structured feature representation with both high-order interactive semantics and computability for subsequent input to the feature projection layer, spatial mapping mechanism, and multi-scale aggregation operation to generate the final fusion evaluation result. This ensures that the final dynamic estimation of carbon amount can objectively accept the complex collaborative topological features of the ecosystem across modes and processes.

[0045] Optionally, the step of aggregating the node feature states in the set of associated nodes based on the hyperedge weights corresponding to each hyperedge to obtain a hyperedge state representation vector includes: Based on the connection relationship of each node in the dynamic hypergraph structure participating in the hyperedge, obtain the node connectivity index in the dynamic hypergraph structure, and extract the hyperedge weight corresponding to each hyperedge; Based on the node connectivity index and the weight of each hyperedge, aggregate attention weights are assigned to each node in the set of associated nodes. The aggregated attention weights are used to perform a weighted summation operation on the node feature states in the set of associated nodes to generate the hyperedge state representation vector.

[0046] Preferably, in the step of aggregating the node feature states in the associated node set to generate the hyperedge state representation vector, firstly, based on the connection relationships of each node in the dynamic hypergraph structure participating in the hyperedges, the node connectivity index in the dynamic hypergraph structure is obtained, and the hyperedge weights corresponding to each hyperedge are extracted. The dynamic hypergraph structure is generated by mapping a multimodal spatiotemporal feature matrix to an ecological topology hypergraph in the preceding steps. It contains several target nodes and several hyperedges. Each hyperedge is associated with an associated node set, which records all target nodes participating in the high-order ecological function coupling module represented by the hyperedge. The node connectivity index is a quantitative parameter that characterizes the topological importance of a target node in the dynamic hypergraph structure. It is calculated as follows: for a target node, all hyperedges in the dynamic hypergraph structure are traversed, and the total number of hyperedges in which the target node appears as a member is counted. This total number of hyperedges is used as the node connectivity index of the target node. If a target node belongs to multiple hyperedges simultaneously, its node connectivity index is higher, reflecting that the target node plays a pivotal regulatory role in the multi-economy process. The hyperedge weights are real values ​​that are dynamically updated during the iteration of the hyperedge message passing mechanism. The initial hyperedge weights are used directly in the first iteration, while subsequent iterations use the updated hyperedge weights output from the previous loop. This acquisition process decouples and quantifies the topology information and edge weight information, providing a data basis for differentiated treatment of nodes with different hub levels.

[0047] Preferably, after obtaining the node connectivity index of each target node and the hyperedge weight of each hyperedge, aggregation attention weights are assigned to each node in the associated node set according to the node connectivity index and the hyperedge weights corresponding to each hyperedge. The aggregation attention weights are used to differentiate the contribution of different target nodes within the associated node set to the overall state representation of the hyperedge during the aggregation process. Their allocation logic is simultaneously controlled by two factors: first, the normalized connection importance of the target node within the current hyperedge; and second, the module-level coupling strength carried by the hyperedge weight of the current hyperedge. Specifically, for a target hyperedge and its associated node set, the node connectivity index of each target node within the associated node set is first extracted. The node connectivity indices of all nodes are then summed and normalized, i.e., the node connectivity index of each target node is divided by the sum of the node connectivity indices of the entire set, resulting in a normalized connection coefficient between zero and one. Then, the normalized connectivity coefficient is multiplied by the weight of the target hyperedge, and the resulting product is the initial attention score of the target node in the current hyperedge aggregation task. Subsequently, to ensure the aggregation process can suppress irrelevant perturbations, a sparsity threshold is designed. Initial attention scores below a preset threshold are truncated to zero, and the remaining non-zero initial attention scores are summed and normalized again to obtain the final aggregation attention weight assigned to the target node. This design allows nodes with high internal connectivity and strong overall coupling within the hyperedge to receive larger aggregation attention weights, while the contributions of nodes located at the edge with loose module coupling are attenuated, objectively creating an adaptive focus on key driving factors.

[0048] Preferably, after allocating aggregate attention weights to each target node in the associated node set, the aggregate attention weights are used to perform a weighted summation operation on the node feature states in the associated node set to generate the hyperedge state representation vector. The node feature states are real-valued vectors attached to each target node, which have been initialized using a multimodal spatiotemporal feature matrix during the generation of the dynamic hypergraph structure, and are continuously updated through reverse feature distribution in each iteration of the hyperedge message passing. In the first iteration, the initialized node feature states are used directly, while in subsequent iterations, the updated node feature states obtained from the previous iteration are used. The specific process of the weighted summation operation is as follows: For the current target hyperedge, each target node in its associated node set is traversed, and the current node feature state vector of that target node is extracted. The node feature state vector is then scaled element-wise using the aggregate attention weights allocated to that target node, and then all scaled vectors are accumulated along their corresponding element dimensions. After accumulation, the resulting sum vector is the preliminary hyperedge state representation vector. The dimension of this vector is consistent with the dimension of the node feature state. The value of each element integrates the comprehensive information of all ecological process states associated with the hyperedge under the differentiated attention allocation mechanism, and extracts the overall operational level of processes such as canopy photosynthesis, soil respiration and water transport under synergistic effect from the perspective of high-order functional modules.

[0049] Preferably, to further improve the robustness of the hyperedge state representation vector to differences in the number of nodes inside different hyperedges, an adaptive normalization operation based on the number of effective nodes is introduced after obtaining the sum vector through weighted summation. This operation first determines a normalization factor based on the number of effective nodes actually participating in aggregation in the associated node set. Effective nodes are nodes with non-zero aggregation attention weights. To avoid abnormal amplification of the representation amplitude when there are very few effective nodes inside the hyperedge, the normalization factor can be set to the positive square root of the number of effective nodes or the logarithmic function value based on the number of effective nodes. Each element of the sum vector is divided by the normalization factor to obtain the normalized hyperedge state representation vector. Optionally, the normalized hyperedge state representation vector can be further activated nonlinearly, for example, by suppressing negative responses through a linear rectified function, to enhance the sparsity and physical interpretability of the vector representation, ultimately generating the hyperedge state representation vector of the hyperedge. This hyperedge state representation vector, as the core output of the hyperedge aggregation step, not only internalizes the differentiated node contributions and module coupling strength, but also has good adaptability to changes in the hyperedge topology scale. It provides a reliable module-level perceptual information foundation for subsequently using this representation vector for reverse feature distribution to drive the update of node feature states.

[0050] Optionally, the step of performing manifold dimensionality reduction projection processing on the cooperative topological feature sequence using a pre-trained spatiotemporal manifold decoder to generate current-period carbon quantity estimation data corresponding to the target ecological region includes: Obtain the preset reference manifold space; The cooperative topological feature sequence is embedded into the reference manifold space to obtain the coordinates of high-dimensional manifold points; Perform local linear measure calculations on the coordinates of the points in the high-dimensional manifold to extract local geometric structure parameters; Based on the local geometric structure parameters, the coordinates of the points in the high-dimensional manifold are compressed to generate a low-dimensional manifold representation tensor. The low-dimensional manifold representation tensor is input into the fully connected regression layer in the pre-trained spatiotemporal manifold decoder to generate the current period carbon mass estimation data.

[0051] Preferably, in the step of generating current-period carbon quantity estimation data for the target ecological region by performing manifold dimensionality reduction projection processing on the cooperative topological feature sequence through a pre-trained spatiotemporal manifold decoder, a preset reference manifold space is first obtained. The preset reference manifold space is a low-dimensional manifold structure constructed during the pre-training phase. Its construction process specifically involves: using multimodal spatiotemporal feature matrix from multiple historical periods of the climate zone or ecological functional zone where the target ecological entity is located, along with corresponding measured carbon flux data, to optimize and train the initial manifold space so that the geometric distance between any two coordinate points in the manifold space can reflect the similarity between the carbon quantity states of the corresponding historical ecological entities. After training, the reference manifold space is saved as a data structure containing a manifold space basis vector matrix and a mean center vector. The column vectors of the manifold space basis vector matrix span the principal directions of the manifold space, while the mean center vector marks the overall offset reference of the manifold space in the feature space. Once trained, the reference manifold space remains fixed, serving as a common reference frame for projecting subsequent new periodic co-topological feature sequences onto the low-dimensional manifold. This ensures that carbon quantity estimation results for different periods are generated within the same comparable geometric framework, avoiding estimation bias caused by projection reference drift.

[0052] Preferably, after obtaining a preset reference manifold space, the cooperative topological feature sequence is embedded into the reference manifold space to obtain high-dimensional manifold point coordinates. The cooperative topological feature sequence is a one-dimensional real vector formed by performing a hyperedge message passing mechanism on a dynamic hypergraph structure and iteratively converging according to preset sequence concatenation rules, as described in the previous steps. The specific execution method of the embedding operation is as follows: using the cooperative topological feature sequence as the input feature vector, the input feature vector is linearly projected using the manifold space basis vector matrix pre-stored in the reference manifold space, that is, the matrix product of the input feature vector and the manifold space basis vector matrix is ​​calculated, and then the product result is summed with the mean center vector. The result is the corresponding coordinate point of the cooperative topological feature sequence in the reference manifold space, and this coordinate point is defined as the high-dimensional manifold point coordinate. The dimension of the high-dimensional manifold point coordinates is usually comparable to or slightly lower than the original dimension of the cooperative topological feature sequence. However, the "high-dimensional" in its name is relative to the target dimension of the subsequent dimension compression operation, meaning that the coordinate point still carries all the manifold embedding information that has not yet been compressed, and fully preserves the comprehensive semantics of the cooperative topological feature sequence regarding the coupling strength of ecological processes and the state of functional modules.

[0053] Preferably, after obtaining the coordinates of the high-dimensional manifold point, a local linear measure calculation is performed on the high-dimensional manifold point coordinates to extract local geometric structure parameters. The local geometric structure parameters are a set of quantifiable geometric properties describing the local neighborhood morphology of the high-dimensional manifold point coordinates in the reference manifold space. Their calculation relies on a local neighborhood graph determined during the pre-training phase. This local neighborhood graph records the local connectivity structure between each training sample coordinate point and its neighboring sample coordinate points in the reference manifold space. When extracting the local geometric structure parameters, firstly, the local neighborhood range of the high-dimensional manifold point coordinates in the reference manifold space is located according to the local neighborhood graph, and the nearest neighbor training sample coordinate points included within this local neighborhood range are determined. Then, a local covariance matrix is ​​constructed using these nearest neighbor training sample coordinate points, and eigenvalue decomposition is performed on the local covariance matrix. The resulting local eigenvalue sequence and local eigenvector set are used as the local geometric structure parameters. These local geometric structure parameters objectively reflect the manifold curvature direction and degree at the location of the high-dimensional manifold point. The local eigenvalue sequence indicates the information concentration in each direction, while the local eigenvector set provides the corresponding orientation. By extracting these local geometric structure parameters, the geometric constraints required for structure-preserving dimensionality compression of the high-dimensional coordinate points are obtained.

[0054] Preferably, after extracting the local geometric structure parameters, a dimension compression operation is performed on the coordinates of the high-dimensional manifold points based on these parameters to generate a low-dimensional manifold representation tensor. The core of the dimension compression operation is to identify and retain the dominant manifold direction that is most discriminative of the current period's carbon quantity state, while suppressing information redundancy directions. Specifically, the operation involves selecting several principal eigenvalues ​​and their corresponding local eigenvectors from the sequence of local eigenvalues ​​of the local geometric structure parameters, where the cumulative contribution rate reaches a preset cumulative contribution rate threshold. The preset cumulative contribution rate threshold can be set to cover 95% or higher of the total variation, thereby determining the number of principal eigenvalues ​​as the target compression dimension. Then, a compression projection matrix is ​​constructed using these selected local eigenvectors, where the number of columns in the compression projection matrix is ​​equal to the target compression dimension, and each column represents a selected local eigenvector. The coordinates of the high-dimensional manifold points are projected onto the subspace spanned by the compression projection matrix, i.e., matrix multiplication is performed between the compression projection matrix and the high-dimensional manifold point coordinates. The resulting low-dimensional coordinate vector is the core component of the low-dimensional manifold representation tensor. The low-dimensional manifold representation tensor is significantly lower in dimension than the point coordinates of the high-dimensional manifold, but it still retains the geometric principal structure in the manifold space that is closely related to the change in carbon quantity, thus achieving a balance between information compression and structure preservation.

[0055] Preferably, the pre-trained spatiotemporal manifold decoder includes a geometric decoupling layer and a fully connected regression layer, which are jointly trained and their parameters are fixed before being used in the inference stage. The geometric decoupling layer receives the low-dimensional manifold representation tensor and performs manifold deconvolution and spatiotemporal context fusion operations on it. The geometric decoupling layer internally includes several cascaded temporal attention modules and manifold upsampling modules. The temporal attention module uses a learnable positional encoding matrix to extract and recalibrate the temporal evolution information implicit in the low-dimensional manifold representation tensor, enabling the active ecological process features at different time segments to obtain differentiated activation intensities. The manifold upsampling module then uses a set of trainable manifold basis functions to map the recalibrated low-dimensional features back to a high-dimensional manifold coordinate space, restoring the local detailed structures temporarily discarded due to dimensionality compression. The geometric decoupling layer ultimately outputs a high-dimensional manifold reconstruction vector with restored details. This vector, while maintaining semantic alignment with the original co-topological feature sequence, is endowed with a clearer spatiotemporal evolution context and higher feature discriminancy.

[0056] Preferably, the high-dimensional manifold reconstruction vector, after detailed restoration of the output of the geometric decoupling layer, is input into the fully connected regression layer in the pre-trained spatiotemporal manifold decoder to generate the current period's carbon flux estimation data. The fully connected regression layer is a multi-layer fully connected neural network structure, where the number of neurons in the input layer is equal to the dimension of the high-dimensional manifold reconstruction vector, the number of neurons in the hidden layers decreases progressively, and the output layer has only one neuron. A linear rectified activation function and a Dropout random deactivation mechanism are configured between the hidden layers of this fully connected regression layer. During the pre-training phase, the connection weights and bias parameters of each neuron are determined by minimizing the mean squared error loss function between the output value and the measured carbon flux data. During the inference phase, the high-dimensional manifold reconstruction vector is fed into the fully connected regression layer. After multi-layer forward layer-by-layer calculation, a single real scalar value is output. This scalar value is the current periodic carbon mass estimation data of the target ecological entity within the preset physical time window. Its physical unit is a fixed carbon mass per square meter per day or a similar commensurable physical unit. This completes the end-to-end dynamic estimation of carbon mass values ​​from the cooperative topological feature sequence through manifold dimensionality reduction projection.

[0057] Optionally, before the step of performing manifold dimensionality reduction projection processing on the cooperative topological feature sequence using a pre-trained spatiotemporal manifold decoder to generate the current period carbon quantity estimation data corresponding to the target ecological region, the method further includes: Obtain historical carbon flux observation sample sequences and corresponding real carbon quantity labels; The historical carbon flux observation sample sequence is perturbed by a preset data augmentation operation to generate positive sample pairs and negative sample pairs. Construct a comparative loss metric using the positive and negative sample pairs; Based on the contrastive loss metric and the real carbon quantity label, the initial spatiotemporal manifold decoding network is subjected to gradient update iteration, and the trained initial spatiotemporal manifold decoding network is used as the pre-trained spatiotemporal manifold decoder.

[0058] Preferably, in constructing the pre-trained spatiotemporal manifold decoder, the historical carbon flux observation sample sequence and its corresponding real carbon quantity label are first acquired. The historical carbon flux observation sample sequence originates from multimodal sensing datasets related to the carbon cycle collected over multiple historical time periods within the climate zone or ecological functional zone of the target ecological entity. Each historical carbon flux observation sample sequence corresponds to a complete time window, such as a seven-day continuous sensing data record, and this sequence has been pre-constructed into a one-dimensional real vector according to the same preset sequence splicing rules as the cooperative topological feature sequence. The real carbon quantity label is the net carbon exchange data obtained through actual observation using equipment such as eddy covariance flux towers and standardized, synchronized with each historical carbon flux observation sample sequence in time. Its physical unit is a fixed carbon mass per square meter per day, serving as the target ground truth for supervised training. This acquisition action provides a paired input-output data foundation for subsequently constructing comparative learning samples and training regression capabilities.

[0059] Preferably, after acquiring historical carbon flux observation sample sequences and their corresponding real carbon quantity labels, the historical carbon flux observation sample sequences are perturbed using preset data augmentation operations to generate positive sample pairs and negative sample pairs. The preset data augmentation operations are a set of perturbation strategies designed for ecological time-series data, including but not limited to sub-operations such as temporal local masking, frequency domain amplitude perturbation, and random sensor channel deactivation. When generating positive sample pairs, an original historical carbon flux observation sample sequence is selected as an anchor sample. A lightweight temporal local masking process is applied to this anchor sample, for example, randomly truncating and zeroing the values ​​at a consecutive small time step in the sequence, thereby generating a variant sample highly similar to the anchor sample in terms of the main ecological process characteristics. This anchor sample and the variant sample constitute a positive sample pair. When generating negative sample pairs, a historical carbon flux observation sample sequence is randomly selected from a different time period than the time window to which the anchor sample belongs as a heterogeneous sample. This heterogeneous sample is combined with the anchor sample to form a negative sample pair. In this way, positive sample pairs reflect the similar states of the same ecological entity under extremely short parallax, while negative sample pairs capture the essential differences in the carbon cycle state under different time windows, providing a comparative signal with clear physical meaning for subsequently widening the state spacing in manifold space.

[0060] Preferably, a contrastive loss metric is constructed using the positive and negative sample pairs. This metric constrains the learning direction of the manifold embedding portion in the initial spatiotemporal manifold decoding network during training. Its technical objective is to compress the geometric distance between positive sample pairs in the current reference manifold space or intermediate manifold space to a relatively small value, while simultaneously widening the geometric distance between negative sample pairs to a preset boundary value. Specifically, the construction method is as follows: positive and negative sample pairs are input into the geometric decoupling layer of the initial spatiotemporal manifold decoding network for forward propagation. Before passing through the fully connected regression layer, their respective manifold representation vectors are extracted from an intermediate layer of the geometric decoupling layer or from the final output high-dimensional manifold reconstruction vector. Then, the Euclidean distance between the manifold representation vectors of the positive and negative sample pairs is calculated, and the Euclidean distance between the manifold representation vectors of the negative sample pairs is also calculated. Based on these two distances, a marginal contrastive loss function is constructed. The value of this loss function is equal to the positive sample distance minus the negative sample distance plus the preset marginal value, compared with zero, and the larger value is taken as the contrastive loss metric. This comparative loss metric provides the network with a geometric intuition for judging the similarity of ecological states from an unsupervised perspective, avoiding overfitting that may result from relying solely on sparse carbon flux truth labels.

[0061] Preferably, the initial spatiotemporal manifold decoding network is subjected to gradient update iterations based on the contrastive loss metric and the real carbon quantity label, and the trained initial spatiotemporal manifold decoding network is used as the pre-trained spatiotemporal manifold decoder. The initial spatiotemporal manifold decoding network is a network structure containing the geometric decoupling layer and the fully connected regression layer, but with parameters in a randomly initialized or pre-initialized state. In one gradient update iteration, multiple historical carbon flux observation sample sequences of a batch are first fed into the initial spatiotemporal manifold decoding network. The network gradually extracts spatiotemporal features through the geometric decoupling layer and outputs a high-dimensional manifold reconstruction vector after detailed restoration. Then, the fully connected regression layer maps this vector to a carbon quantity estimate. By calculating the mean squared error loss between the estimate and the corresponding real carbon quantity label, the regression loss term is obtained. At the same time, the contrastive loss metric is calculated using the positive and negative sample pairs constructed from this batch of data in the manner described above. The regression loss term is weighted and summed with the comparative loss metric. The weighting coefficients can be empirically set to, for example, 0.5 and 1.0, to maintain a balance between regression accuracy and manifold structure rationality. This weighted sum is the total loss function. Then, the partial derivatives of the total loss function with respect to all learnable parameters in the geometrically decoupled layer and the fully connected regression layer of the initial spatiotemporal manifold decoding network are calculated using automatic differentiation. Gradient descent algorithms, such as adaptive moment estimation optimizers, are used to iteratively update the parameters using these partial derivatives until the total loss function converges to a stable level. Upon convergence, the network parameters are fixed, and the manifold embedding space inherent in the geometrically decoupled layer is saved as the preset reference manifold space, which includes the manifold space basis vector matrix and mean center vector. The network with fixed parameters obtained at this point is the pre-trained spatiotemporal manifold decoder. It can accurately project the cooperative topological feature sequence of any new period onto the reference manifold space with carbon quantity discrimination, and ultimately output high-precision carbon quantity estimation data for the current period. This training method, which combines contrastive and regression losses, differs from the traditional black-box training that relies solely on regression loss. It enables the latent space within the decoder to possess an explicit geometric understanding of the similarities and differences in the dynamic processes of the carbon cycle, thereby demonstrating stronger generalization ability and robustness in addressing the challenges of nonlinear and spatiotemporally heterogeneous estimation of the carbon content of target ecological entities.

[0062] Optionally, after the step of generating the current period carbon physical quantity estimation data corresponding to the target ecological region, the method further includes: Receive real-time error feedback signals uploaded by edge computing nodes pre-deployed in the target ecological area. The real-time error feedback signals are used to characterize the error feedback status between the current period carbon physical quantity estimation data and the corresponding local observation results of the edge computing nodes. Analyze the local feature deviation index in the real-time error feedback signal; Generate a corresponding topology compensation mask based on the local feature deviation index; The topology compensation mask is applied to the ecological topology hypergraph to update the hyperedge weights corresponding to the hyperedges.

[0063] Preferably, after generating the current period carbon physical quantity estimation data corresponding to the target ecological area, the method further includes receiving a real-time error feedback signal uploaded by an edge computing node pre-deployed in the target ecological area. The edge computing node pre-deployed in the target ecological area refers to an embedded device installed inside or near the target ecological entity, possessing local environmental perception and lightweight computing capabilities, such as a field observation terminal equipped with a soil carbon flux measurement chamber, a micro weather station, and a wireless communication module. This edge computing node independently collects and calculates the local observed carbon physical quantity parameters for the local plot it covers within each preset physical time window, while simultaneously receiving the current period carbon physical quantity estimation data for the corresponding time window from the cloud. The edge computing node compares the local observed carbon physical quantity parameters with the current period carbon physical quantity estimation data within the same physical time window, calculates characteristics such as residual magnitude, residual sign, and residual abnormality, and encapsulates these characteristics into a real-time error feedback signal for upward transmission. The real-time error feedback signal is used to characterize the error feedback state between the current period carbon physical quantity estimation data and the corresponding local observation results of the edge computing node. The local observation results are not the overall carbon physical quantity of the entire region, but the measured value of the carbon budget process in the microenvironment where the edge computing node is located. The error feedback state reflects the degree and pattern of difference between the two in a local spatial region.

[0064] Preferably, after receiving the real-time error feedback signal, the local feature deviation index in the real-time error feedback signal is parsed. The local feature deviation index is a set of structured error quantification data, which includes at least an error directionality identifier, an error energy value, and an affected feature channel mapping index. The error directionality identifier indicates whether the local observation result is higher or lower than the estimated value, and can be represented by positive or negative signs; the error energy value is the normalized residual amplitude, for example, the residual is divided by the mean of the two and the absolute value is taken, and then mapped to the interval between zero and one; the affected feature channel mapping index indicates which ecological physical feature dimensions show a large deviation in this local error, for example, according to the error decomposition algorithm, the total error is reconstructed back to the various feature dimensions of the multimodal spatiotemporal feature matrix, the feature dimensions with a contribution rate exceeding a preset threshold (e.g., 10%) are selected, and the feature dimension labels corresponding to these feature dimensions are recorded. The parsing action extracts and formats this set of local feature deviation indices from the signal message into data objects that can be directly used in subsequent steps, establishing a quantitative correlation from the local error phenomenon to the multimodal feature level.

[0065] Preferably, after parsing the local feature deviation index, a corresponding topology compensation mask is generated based on the local feature deviation index. The topology compensation mask is an adjustment matrix or vector acting on the weights of hyperedges in the ecological topology hypergraph. Each element corresponds to a hyperedge in the ecological topology hypergraph and carries a weight correction coefficient. The process of generating the topology compensation mask is based on a pre-constructed mapping relationship between feature channels and target nodes: each target node in the ecological topology hypergraph corresponds to a feature dimension label in the multimodal spatiotemporal feature matrix, and the affected feature channel mapping index has been determined previously through the local feature deviation index. When generating the topology compensation mask, the affected feature channel mapping index is traversed, all target nodes corresponding to these feature channels are located in the ecological topology hypergraph, and then, based on the participation relationship of these target nodes in the graph, all hyperedges connecting these affected nodes are found. For each such hyperedge, a correction coefficient is calculated according to certain rules based on the error energy value and error directionality identifier of the affected node associated with it. For example, when multiple associated nodes simultaneously exhibit positive deviations (local observations are higher than estimates), it indicates that the overall activity of the functional module represented by the hyperedge is underestimated. The correction coefficient can be set to a value greater than one (e.g., 1.2) to enhance the information transmission weight of this module in the next iteration. Conversely, if there is a negative deviation, the correction coefficient can be set to a positive value less than one (e.g., 0.8) to weaken its dominant effect. If a hyperedge connects nodes with both positive and negative deviations, the correction direction and magnitude are determined by a weighted average of the error energy values. Thus, correction coefficients are generated for all affected hyperedges, while for hyperedges not associated with any affected nodes, the correction coefficient is set to 1.0, i.e., no compensation is made. Finally, the correction coefficients of all hyperedges are arranged in order of hyperedge number to form a topology compensation mask.

[0066] Preferably, after generating the topology compensation mask, the topology compensation mask is applied to the ecological topology hypergraph to update the hyperedge weights corresponding to the hyperedges. The ecological topology hypergraph is a dynamic hypergraph structure generated in the previous stage based on the multimodal spatiotemporal feature matrix, containing target nodes, node feature states, hyperedges, and initial hyperedge weights. The hyperedge weights currently used to generate carbon quantity estimation data based on the collaborative topology feature sequence are updated hyperedge weights after convergence through message passing iterations. The operation of applying the topology compensation mask is as follows: traverse each hyperedge in the ecological topology hypergraph, extract the current hyperedge weight of the hyperedge, multiply the hyperedge weight element-wise with the correction coefficient corresponding to the hyperedge in the topology compensation mask, and use the product as the hyperedge weight after edge feedback correction, overwriting the original value. This updated hyperedge weight will directly affect the aggregation and distribution capabilities of the hyperedge in the next round of message passing iterations, thereby correcting the expression of cross-modal coupling strength in the ecological topology hypergraph. This operation enables adaptive fine-tuning from local error feedback signals to topological parameters, allowing the ecological topology supermap to linearly incorporate ground observations, dynamically correct spatial estimation biases caused by remote sensing and model inference, and improve the sensitivity and numerical consistency of subsequent carbon quantity estimation in response to local heterogeneity and spatiotemporal abrupt events.

[0067] Optionally, after the step of generating the current period carbon physical quantity estimation data corresponding to the target ecological region, the method further includes: The current period carbon quantity estimation data is converted into a standard format data packet according to a preset data encapsulation protocol. The standard format data packet is hashed to generate a checksum string; The verification digest string is concatenated with the standard format data packet to form an on-chain evidence storage block; The on-chain evidence-stored blocks are distributed to a pre-configured distributed node network to perform consensus synchronization verification processing, thereby completing the solidification and storage of the current period's carbon physical quantity estimation data.

[0068] Preferably, after generating the current-cycle carbon physical quantity estimation data corresponding to the target ecological region, the current-cycle carbon physical quantity estimation data is converted into a standard format data packet according to a preset data encapsulation protocol. The preset data encapsulation protocol is a predefined structured data organization rule that specifies the field order, encoding method, and separator identifiers of the estimation data and its associated metadata in the data packet. Specifically, the protocol requires that the numerical value of the current-cycle carbon physical quantity estimation data, the corresponding physical unit, the start and end timestamps of the preset physical time window, the spatial block number of the target ecological entity, and the version identifier of the cooperative topological feature sequence on which the estimation is based be filled into an unsigned byte sequence in a fixed order. Each field uses fixed-length or prefix-length encoding, and fields are separated by preset separators. The converted standard format data packet is a self-describing binary or text sequence with a uniform structure, enabling any node following the same protocol to unambiguously parse its content. This conversion action aggregates the hashed estimation information into a standardized payload that facilitates subsequent hash operations and network transmission.

[0069] Preferably, after obtaining the standard format data packet, a hash function is performed on the standard format data packet to generate a checksum string. The hash function employs a designed one-way hash function that accepts a standard format data packet of arbitrary length as input. Through steps such as block division, message expansion, multi-round bit operations, and modulo addition, it outputs a fixed-length binary bit sequence. This binary bit sequence is then converted into a hexadecimal printable character representation, which is the checksum string. This hash function possesses strong avalanche effect and collision resistance characteristics, ensuring that any modification to even a single bit in the standard format data packet will cause drastic and unpredictable changes to the generated checksum string. This checksum string serves as a unique data fingerprint of the standard format data packet at this moment, and its technical function is to provide objective digital evidence for subsequent data integrity verification and tamper-proof detection.

[0070] Preferably, after generating the verification digest string, the verification digest string is concatenated with the standard format data packet to form an on-chain evidence storage block. The concatenation operation is performed according to a preset block assembly rule, which defines the order and encoding format of the components within the block. For example, an on-chain evidence storage block may contain three parts: a block header, a data payload, and a block tail. The block header contains the verification digest string, the current timestamp, and a chained pointer to the verification digest string of the previous on-chain evidence storage block, forming a hash chain structure. The data payload fully embeds the original byte stream of the standard format data packet. The block tail appends an auxiliary verification field calculated jointly by the block header and the data payload. After encapsulating the verification digest string and the standard format data packet according to this assembly rule, a data block entity that can be independently identified and stored in a distributed node network is obtained. This concatenation action binds the carbon physical quantity estimation data and the digital fingerprint used to verify its integrity to the same block, establishing an integrated storage form for content and credentials.

[0071] Preferably, after the on-chain evidence storage block is formed, it is distributed to a pre-configured distributed node network for consensus synchronization verification. The pre-configured distributed node network is a peer-to-peer network of evidence storage nodes deployed in different physical locations, interconnected via secure communication links. Each evidence storage node stores the same initial network configuration information, including the network addresses and public key certificates of other evidence storage nodes, and runs the same consensus protocol stack. The distribution action employs a diffusion broadcast, sending the newly generated on-chain evidence storage block in parallel from the initiating node to all other evidence storage nodes within the network. Upon receiving the on-chain evidence storage block, each evidence storage node initiates consensus synchronization verification, which consists of two phases: block validity verification and consensus achievement. During the block validity verification phase, each notarization node independently performs a series of verification operations: recalculating the hash value of the standard format data packet and comparing it with the checksum string in the block header to verify consistency; checking whether the previous notarized block pointed to by the chain pointer exists in the local blockchain ledger to verify chain continuity; and verifying whether the data format in the block strictly follows the preset data encapsulation protocol and block assembly rules. Only blocks that pass all checks are marked as valid blocks awaiting consensus.

[0072] Preferably, after completing the block validity verification, each evidence storage node enters the consensus-building phase, deciding whether to accept the on-chain evidence storage block through a pre-deployed Byzantine fault-tolerant consensus protocol or similar mechanism. In this consensus protocol, each evidence storage node broadcasts its verification conclusion to the network and collects the verification conclusions of other evidence storage nodes to form a global view. When more than a preset proportion of evidence storage nodes unanimously agree that the on-chain evidence storage block is valid, consensus is reached, and each network node appends the block to the end of its local blockchain ledger, forming an irreversible append record. In scenarios where some replicas are not synchronized due to network latency, the consensus protocol will also trigger a block synchronization and retransmission mechanism until all normal node ledgers reach consensus.

[0073] Preferably, after all the evidence storage nodes in the distributed node network reach a consensus and write the on-chain evidence storage block into their local ledgers, the solidification and storage of the current period's carbon physical quantity estimation data is completed. Solidification and storage refers to the data forming an immutable and uniformly linearly ordered persistent record state across multiple independent nodes. Because each evidence storage node's ledger contains the same on-chain evidence storage blockchain sequence, and any modification to a stored block will result in a hash checksum string mismatch, which will be quickly detected and isolated by the replicas of most nodes in the network, this solidification and storage mechanism ensures the historical integrity and non-repudiation of the carbon physical quantity estimation data from a data structure perspective. Through this series of processes—from estimation data generation, encapsulation, fingerprinting, and block generation to distributed consensus solidification—an end-to-end closed loop connecting data processing and secure data storage is constructed, enabling the entire lifecycle evolution trajectory of carbon physical quantity estimation to be recorded tamper-proofly, which aligns with ecological monitoring application scenarios that require high objectivity in data storage.

[0074] like Figure 3 As shown, this is an embodiment of an application of a dynamic estimation device for ecosystem carbon content, which includes: A multimodal spatiotemporal feature matrix generation module is used to acquire multi-source heterogeneous sensing data streams of a target ecological region, and to perform multi-dimensional spatiotemporal feature extraction processing on the multi-source heterogeneous sensing data streams to obtain a multimodal spatiotemporal feature matrix. A dynamic hypergraph structure generation module is used to map the multimodal spatiotemporal feature matrix onto a pre-constructed ecological topology hypergraph to generate a dynamic hypergraph structure with node feature states and hyperedge weights. The ecological topology hypergraph includes nodes representing single environmental elements and hyperedges representing the collaborative coupling relationship of multiple environmental elements. The node feature states are derived from the feature assignment results of the multimodal spatiotemporal feature matrix on the nodes, and the hyperedge weights are derived from the correlation characterization results of the multimodal spatiotemporal feature matrix on the collaborative coupling relationship of the multiple environmental elements. The collaborative topology feature sequence generation module is used to execute a hyperedge message passing mechanism on the dynamic hypergraph structure, so as to perform iterative update operations on the node feature state and the hyperedge weight based on the node feature state and the hyperedge weight, and output the converged collaborative topology feature sequence. The current period carbon quantity estimation data generation module is used to perform manifold dimensionality reduction projection processing on the cooperative topological feature sequence through a pre-trained spatiotemporal manifold decoder to generate the current period carbon quantity estimation data corresponding to the target ecological region.

[0075] like Figure 4 As shown, an electronic device according to an embodiment of this application includes: a processor and a memory; the memory is used to store computer programs; the processor is used to execute the program stored in the memory to implement the functions of each module of the device described in the embodiment of this application, or to implement the steps of the method described.

[0076] Figures 2-4 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for dynamic estimation of ecosystem carbon content, characterized in that, Includes the following steps: A multi-source heterogeneous sensing data stream of the target ecological area is acquired, and multi-dimensional spatiotemporal feature extraction processing is performed on the multi-source heterogeneous sensing data stream to obtain a multi-modal spatiotemporal feature matrix; The multimodal spatiotemporal feature matrix is ​​mapped onto a pre-constructed ecological topology hypergraph to generate a dynamic hypergraph structure with node feature states and hyperedge weights. The ecological topology hypergraph includes nodes representing single environmental elements and hyperedges representing the collaborative coupling relationships of multiple environmental elements. The node feature states are derived from the feature assignment results of the multimodal spatiotemporal feature matrix on the nodes, and the hyperedge weights are derived from the correlation characterization results of the multimodal spatiotemporal feature matrix on the collaborative coupling relationships of the multiple environmental elements. A hyperedge message passing mechanism is executed on the dynamic hypergraph structure to iteratively update the node feature state and the hyperedge weight based on the node feature state and the hyperedge weight, and output the converged collaborative topology feature sequence. The co-topological feature sequence is subjected to manifold dimensionality reduction projection processing by a pre-trained spatiotemporal manifold decoder to generate the current periodic carbon quantity estimation data corresponding to the target ecological region.

2. The method for dynamic estimation of ecosystem carbon content as described in claim 1, characterized in that, The step of performing multi-dimensional spatiotemporal feature extraction processing on the multi-source heterogeneous sensing data stream to obtain a multimodal spatiotemporal feature matrix includes: The multi-source heterogeneous sensing data stream is spatially gridded according to a preset spatial grid granularity to obtain spatial sub-block data. Extract remote sensing spectral features, meteorological sequence features, and soil micro-sensing features from the data of each spatial sub-block; The remote sensing spectral features, meteorological sequence features, and soil microsensing features are spliced ​​together in a time sequence according to a preset time sliding window to obtain the multimodal spatiotemporal feature matrix.

3. The method for dynamic estimation of ecosystem carbon content as described in claim 1, characterized in that, The step of mapping the multimodal spatiotemporal feature matrix onto a pre-constructed ecological topology hypergraph to generate a dynamic hypergraph structure with node feature states and hyperedge weights includes: The target node in the ecological topology hypergraph is determined based on the feature dimension labels in the multimodal spatiotemporal feature matrix; Calculate the dynamic correlation coefficient of the target node within a preset physical time window; Determine whether the dynamic correlation coefficient reaches a preset coupling threshold; When the dynamic correlation coefficient reaches the preset coupling threshold, the hyperedge is established between the target nodes, and the initial hyperedge weight corresponding to the hyperedge is determined according to the dynamic correlation coefficient. The target node's node feature state is initialized using the multimodal spatiotemporal feature matrix, and the target node, node feature state, hyperedge, and initial hyperedge weight are written into the ecological topology hypergraph to generate the dynamic hypergraph structure.

4. The method for dynamic estimation of ecosystem carbon content as described in claim 1, characterized in that, The step of executing a hyperedge message passing mechanism on the dynamic hypergraph structure to iteratively update the node feature states and hyperedge weights based on the node feature states and the hyperedge weights, and outputting a converged collaborative topological feature sequence includes: Extract the set of associated nodes contained in each hyperedge in the dynamic hypergraph structure and the hyperedge weight corresponding to each hyperedge; Based on the hyperedge weights corresponding to each hyperedge, the node feature states in the set of associated nodes are aggregated to obtain a hyperedge state representation vector. Based on the hyperedge state representation vector, a reverse feature distribution operation is performed on the feature states of each node in the set of associated nodes to obtain the updated node feature states. Based on the matching relationship between the updated node feature state and the hyperedge state representation vector, a weight correction operation is performed on the hyperedge weights corresponding to each hyperedge to obtain the updated hyperedge weights. Determine whether the feature change rate of the updated node feature state and the weight change rate of the updated hyperedge weight are both lower than a preset convergence threshold; When both the feature change rate and the weight change rate are lower than the preset convergence threshold, the updated node feature state and the updated hyperedge weight are concatenated into the collaborative topology feature sequence according to the preset sequence concatenation rules.

5. The method for dynamic estimation of ecosystem carbon content as described in claim 4, characterized in that, The step of aggregating the node feature states in the set of associated nodes based on the hyperedge weights corresponding to each hyperedge to obtain the hyperedge state representation vector includes: Based on the connection relationship of each node in the dynamic hypergraph structure participating in the hyperedge, obtain the node connectivity index in the dynamic hypergraph structure, and extract the hyperedge weight corresponding to each hyperedge; Based on the node connectivity index and the weight of each hyperedge, aggregate attention weights are assigned to each node in the set of associated nodes. The aggregated attention weights are used to perform a weighted summation operation on the node feature states in the set of associated nodes to generate the hyperedge state representation vector.

6. The method for dynamic estimation of ecosystem carbon content as described in claim 1, characterized in that, The step of generating current-period carbon quantity estimation data for the target ecological region by performing manifold dimensionality reduction projection processing on the cooperative topological feature sequence through a pre-trained spatiotemporal manifold decoder includes: Obtain the preset reference manifold space; The cooperative topological feature sequence is embedded into the reference manifold space to obtain the coordinates of high-dimensional manifold points; Perform local linear measure calculations on the coordinates of the points in the high-dimensional manifold to extract local geometric structure parameters; Based on the local geometric structure parameters, the coordinates of the points in the high-dimensional manifold are compressed to generate a low-dimensional manifold representation tensor. The low-dimensional manifold representation tensor is input into the fully connected regression layer in the pre-trained spatiotemporal manifold decoder to generate the current period carbon mass estimation data.

7. The method for dynamic estimation of ecosystem carbon content as described in claim 1, characterized in that, Before the step of performing manifold dimensionality reduction projection processing on the cooperative topological feature sequence using a pre-trained spatiotemporal manifold decoder to generate the current period carbon quantity estimation data corresponding to the target ecological region, the method further includes: Obtain historical carbon flux observation sample sequences and corresponding real carbon quantity labels; The historical carbon flux observation sample sequence is perturbed by a preset data augmentation operation to generate positive and negative sample pairs. Construct a comparative loss metric using the positive sample pairs and the negative sample pairs; Based on the contrastive loss metric and the real carbon quantity label, the initial spatiotemporal manifold decoding network is subjected to gradient update iteration, and the trained initial spatiotemporal manifold decoding network is used as the pre-trained spatiotemporal manifold decoder.

8. The method for dynamic estimation of ecosystem carbon content as described in claim 1, characterized in that, Following the step of generating the current period carbon physical quantity estimation data corresponding to the target ecological region, the method further includes: Receive real-time error feedback signals uploaded by edge computing nodes pre-deployed in the target ecological area. The real-time error feedback signals are used to characterize the error feedback status between the current period carbon physical quantity estimation data and the corresponding local observation results of the edge computing nodes. Analyze the local feature deviation index in the real-time error feedback signal; Generate a corresponding topology compensation mask based on the local feature deviation index; The topology compensation mask is applied to the ecological topology hypergraph to update the hyperedge weights corresponding to the hyperedges.

9. The method for dynamic estimation of ecosystem carbon content as described in claim 1, characterized in that, Following the step of generating the current period carbon physical quantity estimation data corresponding to the target ecological region, the method further includes: The current period carbon quantity estimation data is converted into a standard format data packet according to a preset data encapsulation protocol. The standard format data packet is hashed to generate a checksum string; The verification digest string is concatenated with the standard format data packet to form an on-chain evidence storage block; The on-chain evidence-stored blocks are distributed to a pre-configured distributed node network to perform consensus synchronization verification processing, thereby completing the solidification and storage of the current period's carbon physical quantity estimation data.

10. A dynamic estimation device for ecosystem carbon content, characterized in that, include: A multimodal spatiotemporal feature matrix generation module is used to acquire multi-source heterogeneous sensing data streams of a target ecological region, and to perform multi-dimensional spatiotemporal feature extraction processing on the multi-source heterogeneous sensing data streams to obtain a multimodal spatiotemporal feature matrix. A dynamic hypergraph structure generation module is used to map the multimodal spatiotemporal feature matrix onto a pre-constructed ecological topology hypergraph to generate a dynamic hypergraph structure with node feature states and hyperedge weights. The ecological topology hypergraph includes nodes representing single environmental elements and hyperedges representing the collaborative coupling relationship of multiple environmental elements. The node feature states are derived from the feature assignment results of the multimodal spatiotemporal feature matrix on the nodes, and the hyperedge weights are derived from the correlation characterization results of the multimodal spatiotemporal feature matrix on the collaborative coupling relationship of the multiple environmental elements. The collaborative topology feature sequence generation module is used to execute a hyperedge message passing mechanism on the dynamic hypergraph structure, so as to perform iterative update operations on the node feature state and the hyperedge weight based on the node feature state and the hyperedge weight, and output the converged collaborative topology feature sequence. The current period carbon quantity estimation data generation module is used to perform manifold dimensionality reduction projection processing on the cooperative topological feature sequence through a pre-trained spatiotemporal manifold decoder to generate the current period carbon quantity estimation data corresponding to the target ecological region.

Images