Carbon sink accounting and trend prediction system based on multi-modal data fusion

Abstract

This invention discloses a carbon sink accounting and trend prediction system based on multimodal data fusion, belonging to the field of carbon sink accounting technology. By fusing multi-source data such as lidar, multispectral remote sensing, soil eDNA, and acoustic sensor networks, it achieves accurate inversion of carbon storage at multiple scales from the single tree level to the ecosystem level. Among them, point cloud data combined with a single tree segmentation model can automatically complete the accurate measurement of single tree location, diameter at breast height, tree height, and crown width. Combined with tree species information identified by remote sensing and localized biomass expansion factor, it ensures the regional adaptability and scientific nature of carbon storage calculation, and produces a high-precision spatial distribution map, significantly improving the accuracy and spatial refinement of carbon sink accounting.

Description

Technical Field

[0001] This invention relates to the field of carbon sequestration technology, specifically a carbon sequestration and trend prediction system based on multimodal data fusion. Background Technology

[0002] A carbon sequestration and trend prediction system based on multimodal data fusion achieves high-precision carbon sequestration and future dynamic prediction by integrating radar, remote sensing, acoustic, and soil eDNA data. It is primarily suitable for precise monitoring and verification of high-standard forestry carbon sequestration projects and refined management of national-level ecological protection and restoration projects. Patent application number 202411254364.2 discloses "A method and application for assessing the community development and carbon sequestration dynamics of mangroves after northward migration," the method comprising: identifying land cover distribution through remote sensing image datasets as initial input data for the model; and using seasonal changes in carbon community and carbon sequestration capacity of newly planted mangroves as the basis for assessment. The study involved calibrating and predicting the changing trends of mangrove ecosystems; constructing a dynamic assessment model for mangrove communities and carbon sequestration; correcting model parameters; and using the model to calculate the area of ​​newly planted mangroves, primary productivity, ecosystem respiration, and net CO2 exchange flux to obtain a dynamic assessment of community development and carbon sequestration. The assessment method based on the validation model helps to understand and predict the health status and carbon sequestration function of coastal wetland mangrove systems. It enables a comprehensive assessment of community development and carbon sequestration capacity after the northward migration of mangroves and the prediction of future trends, providing a scientific basis and data support for accurately assessing the protection and restoration of coastal wetlands after the northward migration of mangroves and the effective management of blue carbon resources.

[0003] The aforementioned existing technologies have solved problems such as insufficient assessment and prediction capabilities of carbon sink functions. However, in operation, relying solely on traditional sampling and low-to-medium resolution remote sensing cannot obtain accurate structural parameters at the single-tree level. Carbon storage estimation depends on a large number of empirical assumptions, and the expression of spatial heterogeneity is coarse, making it difficult to meet the accuracy requirements of carbon sink trading. Since the key biological processes driving the carbon cycle are completely ignored, only static accounting of carbon storage or simple extrapolation based on climate variables can be performed. This results in the inability to reveal the dynamic formation mechanism of carbon sinks, and the reliability of prediction drops sharply under environmental changes or disturbances. Summary of the Invention

[0004] The purpose of this invention is to provide a carbon sink accounting and trend prediction system based on multimodal data fusion to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a carbon sink accounting and trend prediction system based on multimodal data fusion, comprising: The set of partitioning units is used to obtain the set of point clouds in a specified area collected within the current time window, divide it into multiple grid units, and construct the corresponding height matrix; The growth parameter parsing unit is used to generate a set of candidate vertex coordinates based on the height matrix, which is used as a spatial constraint to determine the point cloud corresponding to each tree instance and then analyze it to obtain the growth parameters of each instance. The carbon distribution determination unit is used to calculate the corresponding carbon storage according to the growth parameters of each tree instance. After spatialization, it generates a vegetation carbon distribution grid for the region. The soil analysis unit is used to acquire the eDNA sequencing sequence and environmental factors of all soil sampling points in a specified area within the current time window, calculate the activity index corresponding to each grid location, read the remote sensing feature vector and acoustic feature vector in the area, fuse them, and output the total primary productivity of different grid locations. The growth increment prediction unit is used to calculate the competition coefficient of each tree instance, and then analyze the growth parameters, competition coefficient, future climate data, activity index, and total primary productivity spatial distribution raster of the tree instance to output the regional vegetation carbon storage prediction sequence. The prediction report output unit is used to determine the soil carbon storage prediction sequence and generate a net carbon sink increment prediction report according to the vegetation and soil carbon storage prediction sequences.

[0006] Preferably, the set partitioning unit includes a height extraction module and a height discrimination module; The height extraction module acquires a semantically classified set of point clouds collected within a specified region during the current time window, and divides the set into horizontal planes. OK Column grid cells For unit Extraction at All points within ,in express The first in Data points, Indicates the first One grid cell, Indicates the total number of grid cells. Indicates the total number of rows. Indicates the total number of columns. express Total number of internal data points ,statistics middle axis coordinates Calculate height ,in , express of axis coordinate values, express The first in Data points of The axis coordinates are traversed sequentially. Output the corresponding height ,in express Height; The height discrimination module selects ,like If the height value is invalid, then its corresponding unit is removed. If the flag bit is set to 0, If the height value is valid, the flag bit is set to 1. If the height value is abnormal, the flag bit is set to 2. This process is repeated until the flag bit of each grid cell is determined. After counting the heights of all cells with the flag bit set to 1, the average height is taken as the background value. These are the upper and lower limits of the height, respectively.

[0007] Preferably, the set partitioning unit further includes an interpolation height analysis module and a height correction module; The interpolation height analysis module iterates through each cell whose flag bit is not 1. Search unit neighborhood unit set ,like If there is a neighboring cell with a flag bit set to 1, then the value is calculated based on the neighboring cell. interpolation height ,in , Represents a set The neighboring unit with the marker bit set to 1 in the middle, express and The reciprocal of the distance between them express height, Indicates the first A cell whose flag bit is not 1, This indicates the total number of cells whose flag bit is not 1. Indicates the flag bit, Indicates the serial number, if If there is no neighboring cell with a flag of 1, then the interpolation height is... Assign the background value, and iterate through them sequentially. Output the corresponding interpolation height ,in express The interpolation height; The height correction module uses the cell heights with a flag of 1 and the interpolated cell heights with flags of 0 and 2, arranged in cell order, to generate the interpolated height sequence. Select unit ,Sure of Calculate the median of the height values ​​of all cells within the neighborhood window. ,like ,but The interference point is replaced with Make corrections if necessary, otherwise do nothing; repeat the process until all cells are selected. To control the threshold, the height sequence is filled into the height matrix of the point cloud set in row-major order. middle.

[0008] Preferably, the growth parameter analysis unit includes a matrix creation module, a position setting module, a vertex aggregation module, a key point determination module, a tree height calculation module, and a diameter at breast height (DBH) analysis module; The matrix creation module for the height matrix Perform a two-dimensional Gaussian convolution filter to generate a smoothed height matrix. and create a Binary matrices of the same size As a potential vertex labeling matrix; The location setting module iterates through... The height value of each position within the current location and all positions in its corresponding neighboring window is determined. If the height value of the current position is greater than the height values ​​of all positions in its neighboring window and the lower height limit, then... The corresponding position is set to 1, otherwise it is 0; The vertex aggregation module for After filtering, all positions with a value of 1 in the statistical matrix are counted, and the coordinates of the center point of the corresponding cell are calculated. All coordinates are then summarized to form a set of candidate vertex coordinates. The key point determination module calls the pre-trained single tree segmentation model, inputs the point cloud set into the model, predicts the tree instance to which each point belongs through internal calculation, generates a set of initial key points, uses the candidate vertex coordinate set as spatial constraints to optimize the model results, and outputs a set of point cloud instances. The tree height calculation module reads the first point cloud instance from the point cloud instance set. Tree example ,Sure The horizontal projection centroid is used as the center to define the corresponding search area. Within this area, data points with the classification label "ground" are selected from the point cloud set and stored. The corresponding set of ground points In China, utilizing and Calculate tree height ,in , Representing data points of axis coordinate values, Representing data points of axis coordinate values, Indicates the serial number; The chest diameter analysis module from Points at a height of 1.3 meters above the ground are selected and projected onto a horizontal plane to obtain a two-dimensional point set. This two-dimensional point set is then fitted, and the optimal circle diameter generated by the fitting is used as the... Breast diameter, All points projected onto the horizontal plane will The horizontal projection centroid coordinates are taken as the center position of the crown. The distance values ​​from all projection points to the center position are analyzed, and the maximum distance value is taken as the crown radius.

[0009] Preferably, the carbon distribution determination unit includes a carbon storage analysis module, a benchmark value analysis module, and a carbon distribution generation module; The carbon storage analysis module is based on tree examples. The corresponding varietal growth coefficient can be obtained by querying the horizontal projection centroid in the tree information database. ,according to and chest diameter Tree height Calculate the material coefficient ,in , Indicates the sequence number, reads the current carbon conversion coefficient. ,according to and Analysis of carbon reserves ,in ; The benchmark analysis module repeats the operation until the carbon storage of each tree instance is calculated, and then the carbon storage of the vegetation in the current time window is accumulated to obtain the total carbon storage of the vegetation in the region. After setting the spatial attributes of the raster, the carbon distribution generation module statistically analyzes the crown center position, crown radius, and carbon storage of each tree instance. It then spatializes the data using kernel density estimation, calculates the bandwidth based on the median crown radius, and controls the smooth diffusion of carbon storage from the crown center of the tree instance outward based on the bandwidth. The module calculates the carbon density and total storage of each raster cell within the region and clips it using a region boundary mask, thereby generating a spatially continuous vegetation carbon distribution raster.

[0010] Preferably, the soil analysis unit includes an abundance calculation module, an activity index calculation module, a feature determination module, and a feature splicing module; The abundance calculation module obtains the eDNA sequencing sequences and environmental factors of all soil sampling points in the region within the current time window, for each individual sampling point. Count the number of sequences belonging to different key functional groups. And calculate the relative sequence abundance of each functional group. ,in , Indicates the first The number of sequences in each key functional group. Indicates the number of key functional groups. Indicates the first The number of sequences in each key functional group. Indicates the serial number. Indicates the first The relative sequence abundance of the key functional groups; After determining the relative sequence abundance of functional groups corresponding to each sampling point, the activity index calculation module constructs an abundance matrix. The abundance matrix and environmental factors are normalized, and then fused to generate the activity index for each sampling point. This index is then spatialized, and a raster of the spatial distribution of the activity index within the region is output. Specifically, the activity index is:

[0011] in, Indicates the first Normalized relative abundance of key functional groups Indicates the first Normalized values ​​of environmental factors Indicates at the sampling point The activity index at the site, Indicates the first The weighting coefficients of each key functional group Indicates the first The weighting coefficients of each environmental factor. Indicates the serial number. This represents the total number of environmental factors; The feature determination module acquires standardized audio data from all sound sampling points in the region within the current time window, cuts continuous audio into segments of fixed duration, generates a Mel spectrogram for each segment, analyzes it, determines the high-dimensional feature vectors corresponding to different spectrograms, performs time averaging on all segment feature vectors at each sampling point to obtain the corresponding acoustic feature vector, and generates an acoustic feature spatial distribution grid within the region by spatial interpolating the feature vectors of all sampling points. The feature stitching module reads the remote sensing feature vectors within the area, and stitches the acoustic feature vectors with the remote sensing feature vectors at each spatial grid location to form a fused feature vector. It then analyzes the fused feature vector and outputs the total primary productivity at different spatial grid locations.

[0012] Preferably, the growth increment prediction unit includes a competing tree selection module, a competition coefficient analysis module, a competition coefficient output module, and an increment calculation module; The competing wood selection module for the first Tree example ,Sure Horizontal projection centroid ,by Centered on a target area, define a corresponding search region, and select all other tree instances whose horizontal projection centroids fall within this region, and use them as the target area. Potential competing wood; The competition coefficient analysis module is selected Potential competing wood ,calculate and Distance between ,in , Indicating potential competing wood The horizontal projection centroid, Indicates the serial number, according to and Breast diameter Analysis of potential competing wood Corresponding competition coefficient ,in ; After determining the competition coefficients of all potential competing trees, the competition coefficient output module sums them up to obtain the result. Competition coefficient Iterate through each tree instance and calculate the competition coefficient for each tree instance; The incremental calculation module uses the spatial distribution grid of total primary productivity within the current time window as a baseline constraint to optimize the internal physiological parameters in the pre-trained growth response model. It inputs the spatial distribution grid of tree instance diameter at breast height (DBH), tree height, competition coefficient, future climate data, and activity index into the growth response model for analysis, outputs the DBH and height increments of tree instance, and calculates the predicted sequence of regional vegetation carbon storage in the future time period based on the DBH and height increments.

[0013] Preferably, the prediction report output unit includes a prediction sequence analysis module and a report generation module; The prediction sequence analysis module uses historical soil carbon storage data in the region and performs regression analysis with vegetation carbon storage data and climate data in the same region during the same period to construct an empirical transfer function with vegetation dynamics and climate conditions as driving factors and soil carbon storage as the response variable. By coupling the vegetation carbon storage prediction sequence and climate data in the future time period through this function, the soil carbon storage prediction sequence in the future time period is estimated. The report generation module generates an ecosystem net carbon sink increment forecast report and a future carbon sink spatial hotspot evolution atlas based on the vegetation and soil carbon storage prediction sequence for the future time period.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves accurate inversion of carbon storage at multiple scales, from single tree level to ecosystem level, by integrating multi-source data from lidar, multispectral remote sensing, soil eDNA, and acoustic sensor networks. Point cloud data combined with a single tree segmentation model can automatically complete the accurate measurement of single tree location, diameter at breast height, tree height, and crown width. Combined with tree species information identified by remote sensing and localized biomass expansion factor, it ensures the regional adaptability and scientific nature of carbon storage calculation, and produces a high-precision spatial distribution map, which significantly improves the accuracy and spatial refinement of carbon sink accounting. 2. This invention innovatively incorporates underground biological processes and above-ground ecosystem service functions into the carbon sink assessment system by extracting soil activity index and acoustic ecological features. The activity index integrates key group information such as earthworms and nematodes from eDNA sequencing with soil environmental data, enabling spatial and quantitative expression of biological factors driving soil carbon sequestration and decomposition. The fusion of acoustic features and remote sensing data transforms the activity sounds of functional groups into auxiliary parameters that characterize the primary productivity of the ecosystem. This allows carbon sink assessment to evolve from a single result calculation to a comprehensive understanding of key processes and driving forces in the carbon cycle, providing an innovative analytical dimension for a deeper understanding of the carbon sequestration function of ecosystems. Attached Figure Description

[0015] Figure 1 A schematic diagram of the overall system flow is provided for embodiments of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1This invention provides a technical solution: a carbon sink accounting and trend prediction system based on multimodal data fusion, comprising: The set of partitioning units is used to obtain the set of point clouds in a specified area collected within the current time window, divide it into multiple grid units, and construct the corresponding height matrix; The growth parameter parsing unit is used to generate a set of candidate vertex coordinates based on the height matrix, which is used as a spatial constraint to determine the point cloud corresponding to each tree instance. After that point cloud is determined, the growth parameters of each instance are analyzed, including tree height, diameter at breast height, crown center, and crown radius. The carbon distribution determination unit is used to calculate the corresponding carbon storage according to the growth parameters of each tree instance. After spatialization, it generates a vegetation carbon distribution grid for the region. The soil analysis unit is used to acquire the eDNA sequencing sequence and environmental factors of all soil sampling points in a specified area within the current time window, calculate the activity index corresponding to each grid location, read the remote sensing feature vector and acoustic feature vector in the area, fuse them, and output the total primary productivity of different grid locations. The growth increment prediction unit is used to calculate the competition coefficient of each tree instance, and then analyze the growth parameters, competition coefficient, future climate data, activity index, and total primary productivity spatial distribution raster of the tree instance to output the regional vegetation carbon storage prediction sequence. The prediction report output unit is used to determine the soil carbon storage prediction sequence and generate a net carbon sink increment prediction report according to the vegetation and soil carbon storage prediction sequences.

[0018] The set partitioning unit includes a height extraction module and a height discrimination module; The height extraction module acquires a semantically classified set of point clouds from a specified region collected within the current time window, and divides the set into horizontal planes. OK Column grid cells For unit Extraction at All points within ,in express The first in Data points, Indicates the first One grid cell, Indicates the total number of grid cells. Indicates the total number of rows. Indicates the total number of columns. express Total number of internal data points ,statistics middle axis coordinates Calculate height ,in , express of axis coordinate values, express The first in Data points of The axis coordinates are traversed sequentially. Output the corresponding height ,in express Height; Height discrimination module selection ,like If the height value is invalid, then its corresponding unit is removed. If the flag bit is set to 0, If the height value is valid, the flag bit is set to 1. If the height value is abnormal, the flag bit is set to 2. This process is repeated until the flag bit of each grid cell is determined. After counting the heights of all cells with the flag bit set to 1, the average height is taken as the background value. These are the upper and lower limits of the height, respectively. The set partitioning unit also includes an interpolation height analysis module and a height correction module; The interpolation height analysis module iterates through each cell whose flag bit is not 1. Search unit neighborhood unit set ,like If there is a neighboring cell with a flag bit set to 1, then the value is calculated based on the neighboring cell. interpolation height ,in , Represents a set The neighboring unit with the marker bit set to 1 in the middle, express and The reciprocal of the distance between them express height, Indicates the first A cell whose flag bit is not 1, This indicates the total number of cells whose flag bit is not 1. Indicates the flag bit, Indicates the serial number, if If there is no neighboring cell with a flag of 1, then the interpolation height is... Assign the background value, and iterate through them sequentially. Output the corresponding interpolation height ,in express The interpolation height; The height correction module uses the cell heights with a flag of 1 and the interpolated cell heights with flags of 0 and 2, arranged in cell order, to generate the interpolated height sequence. Select unit ,Sure of Calculate the median of the height values ​​of all cells within the neighborhood window. ,like ,but The interference point is replaced with Make corrections if necessary, otherwise do nothing; repeat the process until all cells are selected. To control the threshold, the height sequence is filled into the height matrix of the point cloud set in row-major order. middle; The growth parameter analysis unit includes a matrix creation module, a position setting module, a vertex aggregation module, a key point determination module, a tree height calculation module, and a diameter at breast height (DBH) analysis module. The matrix creation module for height matrices Perform a two-dimensional Gaussian convolution filter to generate a smoothed height matrix. and create a Binary matrices of the same size As a potential vertex labeling matrix; Location setting module traversal The height value of each position within the current location and all positions in its corresponding neighboring window is determined. If the height value of the current position is greater than the height values ​​of all positions in its neighboring window and the lower height limit, then... The corresponding position is set to 1, otherwise it is 0; Vertex aggregation module After filtering, all positions with a value of 1 in the statistical matrix are counted, and the coordinates of the center point of the corresponding cell are calculated. All coordinates are then summarized to form a set of candidate vertex coordinates. The key point determination module calls the pre-trained single tree segmentation model, inputs the point cloud set into the model, predicts the tree instance to which each point belongs through internal calculation, and generates a set of initial key points. It uses the candidate vertex coordinate set as a spatial constraint to optimize the model's results, thereby outputting a set of point cloud instances. The specific process of optimizing the model results using the set of candidate vertex coordinates as spatial constraints is as follows: S11. Based on the set of candidate vertex coordinates and the initial set of key points generated by the model, perform spatial nearest neighbor analysis. For each position coordinate in the candidate vertex set, check whether there is a corresponding model key point within its preset search range. If there is, establish a mapping relationship between the candidate vertex and the model key point; otherwise, record the candidate vertex as an unmatched vertex. S12. For each key point generated by the model, back-verify whether there are candidate vertices within its search range. If they exist, establish a mapping relationship between the model key point and the candidate vertex; otherwise, record the key point as an isolated key point. S13. For candidate vertices with mapping relationships, use their position coordinates to correct the key points with relationships, and perform a rigid translation of the point cloud of the same tree instance. S14. Take the unmatched candidate vertices as key points of new tree instances, count the point clouds of unassigned tree instances with the classification label of vegetation, search and gather unassigned point cloud fragments within a preset range with the candidate vertices as the center, form new trees, and assign them numbers. For isolated key points, they are considered as false detections and deleted. S15. Statistically optimize the key points and height distribution statistics of each tree instance. If there are two tree instances whose key point distance on the horizontal projection is less than the preset minimum crown width spacing and whose statistical similarity is greater than the preset similarity threshold, then merge the two into one instance. The key points of the merged instance are taken as the average position of the geometric center of each original instance, and their point clouds are integrated. S16. For a single optimized instance, if there are multiple candidate vertices within its horizontal projection range, and the local point cloud corresponding to each vertex position shows independent height peaks and obvious valley separations, it is determined that multiple trees are mistakenly merged. The peak positions are taken as the key points of the new tree instance. Based on the spatial distance and the absolute height difference, the point cloud in the original instance is redistributed to the corresponding tree instance. After verification, the original instance is officially split into multiple independent tree instances. The single-tree segmentation model comprises a point cloud feature extraction backbone network, a semantic-instance dual-branch prediction head, and a post-processing clustering module. The feature extraction network uses a multilayer perceptron and graph attention layer stack to encode the coordinates of the input point cloud layer by layer. The dual-branch prediction head outputs the semantic category (vegetation / non-vegetation) probability, tree instance feature vector, and spatial offset vector for each point. The post-processing module uses a density-based clustering algorithm to aggregate points with similar instance features and spatial proximity that are semantically vegetation into independent tree instances. The model is trained using a labeled tree instance point cloud dataset, augmented by random sampling, and optimized using an end-to-end approach with a joint loss function. This function combines the cross-entropy loss for semantic classification, the variance-covariance loss for instance feature vectors, and the smoothing L1 loss for offset vectors. Training uses the Adam optimizer with a learning rate of 0.001 and a batch size of 8. Iterative training is performed on a GPU-equipped server until convergence. This model can accurately segment tree instances directly from normalized point clouds. The tree height calculation module reads the first point cloud instance from the point cloud instance set. Tree example ,Sure The horizontal projection centroid is used as the center to define the corresponding search area. Within this area, data points with the classification label "ground" are selected from the point cloud set and stored. The corresponding set of ground points In China, utilizing and Calculate tree height ,in , Representing data points of axis coordinate values, Representing data points of axis coordinate values, Indicates the serial number; The chest diameter analysis module starts from Points at a height of 1.3 meters above the ground are selected and projected onto a horizontal plane to obtain a two-dimensional point set. This two-dimensional point set is then fitted, and the optimal circle diameter generated by the fitting is used as the... Breast diameter, All points projected onto the horizontal plane will The horizontal projection centroid coordinates are used as the center position of the crown. The distance values ​​from all projection points to the center position are analyzed, and the maximum distance value is taken as the crown radius. The carbon distribution determination unit includes a carbon storage analysis module, a baseline value analysis module, and a carbon distribution generation module; The carbon storage analysis module is based on tree examples. The corresponding varietal growth coefficient can be obtained by querying the horizontal projection centroid in the tree information database. ,according to and chest diameter Tree height Calculate the material coefficient ,in , Indicates the sequence number, reads the current carbon conversion coefficient. ,according to and Analysis of carbon reserves ,in ; According to tree examples The corresponding varietal growth coefficient can be obtained by querying the horizontal projection centroid in the tree information database. The specific process involves pre-collecting forest resource survey data, high-resolution remote sensing classification maps, and historical sample plot data within a designated area. Tree species information is then bound to their spatial distribution range (vector planes or raster pixels), and each tree species is associated with a pre-stored growth coefficient database. Together, these elements form a tree species-growth coefficient association database with spatial indexing capabilities. Once tree instances are completed... After segmenting and calculating its horizontal projection centroid, the coordinates are used as spatial query conditions to extract raster cell values ​​in the associated database, quickly retrieve the tree species name to which the coordinates belong, and retrieve the corresponding cultivar growth coefficient from the database for subsequent carbon storage calculation. The baseline analysis module repeats the operation until the carbon storage of each tree instance is calculated, and then the carbon storage of the vegetation in the region is accumulated to obtain the total carbon storage of the region within the current time window. After setting the spatial attributes of the raster in the carbon distribution generation module, the crown center position, crown radius and carbon storage of each tree instance are statistically analyzed. The spatialization is performed by kernel density estimation. The bandwidth is calculated according to the median crown radius. Based on the bandwidth, the carbon storage is controlled to smoothly diffuse from the crown center of the tree instance outward. The carbon density and total storage of each raster cell in the region are calculated and clipped with a region boundary mask to generate a spatially continuous vegetation carbon distribution raster. The soil analysis unit includes an abundance calculation module, an activity index calculation module, a feature determination module, and a feature splicing module; The abundance calculation module obtains the eDNA sequencing sequences and environmental factors of all soil sampling points in the region within the current time window. For a single sampling point... Count the number of sequences belonging to different key functional groups. And calculate the relative sequence abundance of each functional group. ,in , Indicates the first The number of sequences in each key functional group. Indicates the number of key functional groups. Indicates the first The number of sequences in each key functional group. Indicates the serial number. Indicates the first The relative sequence abundance of the key functional groups; After determining the relative sequence abundance of functional groups corresponding to each sampling point, the activity index calculation module constructs an abundance matrix. The abundance matrix and environmental factors are normalized and then fused to generate the activity index for each sampling point. This index is then spatialized, and a raster of the spatial distribution of the activity index within the region is output. Specifically, the activity index is:

[0019] in, Indicates the first Normalized relative abundance of key functional groups Indicates the first Normalized values ​​of environmental factors Indicates at the sampling point The activity index at the site, Indicates the first The weighting coefficients of each key functional group Indicates the first The weighting coefficients of each environmental factor. Indicates the serial number. This represents the total number of environmental factors; The feature determination module acquires standardized audio data from all sound sampling points in the region within the current time window, cuts continuous audio into segments of fixed duration, generates a Mel spectrogram for each segment, analyzes it, determines the high-dimensional feature vectors corresponding to different spectrograms, performs time averaging on all segment feature vectors at each sampling point to obtain the corresponding acoustic feature vector, and generates an acoustic feature spatial distribution raster within the region by spatial interpolating the feature vectors of all sampling points. The feature stitching module reads the remote sensing feature vectors within the region. At each spatial grid location, it stitches the acoustic feature vector with the remote sensing feature vector to form a fused feature vector. It then analyzes the fused feature vector and outputs the total primary productivity at different spatial grid locations. The specific process of concatenating acoustic feature vectors and remote sensing feature vectors to form a fused feature vector is as follows: For each spatial grid, the acoustic feature vector corresponding to that location is extracted, and the remote sensing feature vector corresponding to that location is also extracted. The concatenation operation is performed in the feature channel dimension, that is, the acoustic feature vector and the remote sensing feature vector are connected end to end to generate a new high-dimensional joint feature vector. This feature vector completely retains all the information of the original acoustic and remote sensing modes, and the ecological process features and surface parameter features are aligned in position and juxtaposed in dimension in the same feature space. The growth increment prediction unit includes a competing tree selection module, a competition coefficient analysis module, a competition coefficient output module, and an increment calculation module; The competitive wood selection module for the first Tree example ,Sure Horizontal projection centroid ,by Centered on a target area, define a corresponding search region, and select all other tree instances whose horizontal projection centroids fall within this region, and use them as the target area. Potential competing wood; Competition coefficient analysis module selection Potential competing wood ,calculate and Distance between ,in , Indicating potential competing wood The horizontal projection centroid, Indicates the serial number, according to and Breast diameter Analysis of potential competing wood Corresponding competition coefficient ,in ; After determining the competition coefficients of all potential competing trees, the competition coefficient output module sums them up to obtain the result. Competition coefficient Iterate through each tree instance and calculate the competition coefficient for each tree instance; The incremental calculation module uses the spatial distribution grid of total primary productivity within the current time window as a baseline constraint to optimize the internal physiological parameters in the pre-trained growth response model. It inputs the spatial distribution grid of tree instance diameter at breast height (DBH), tree height, competition coefficient, future climate data, and activity index into the growth response model for analysis, outputs the DBH and height increments of tree instance, and calculates the predicted sequence of regional vegetation carbon storage in the future time period based on the DBH and height increments. The growth response model is as follows: S21. The growth parameter time series and competition coefficient of the same batch of tree instances obtained by inversion from historical point clouds, and the historical climate data of the same period and the spatial distribution raster of activity index and total primary productivity in the region are read to construct the training dataset and the validation dataset. S22. Select the 3-PG model and define the prior probability distribution for its key physiological parameters to be optimized. Extract a set of candidate parameters from the prior distribution, substitute the candidate parameters into the model, input historical climate data, activity index and competition coefficient in the training dataset, simulate the total primary productivity of different sampling points and the diameter at breast height and height increment of each sample tree instance during the entire observation period, compare the differences between the simulated diameter at breast height and height increment, total primary productivity and actual measurements, and calculate the joint likelihood under this set of parameters. S23. According to Bayes' theorem, by combining the prior distribution and joint likelihood, update the posterior distribution of the parameters, use the validation dataset to evaluate the predictive performance of the calibrated model, take the median of the posterior distribution as the localized parameter set, and combine this set of parameters with the structure of the 3-PG model to form a localized growth response model. The forecast report output unit includes a forecast sequence analysis module and a report generation module; The prediction sequence analysis module uses historical soil carbon storage data in the region and performs regression analysis with vegetation carbon storage data and climate data in the same region during the same period to construct an empirical transfer function with vegetation dynamics and climate conditions as driving factors and soil carbon storage as the response variable. By coupling the vegetation carbon storage prediction sequence and climate data in the future time period through this function, the soil carbon storage prediction sequence in the future time period is estimated. The report generation module generates a forecast report on the net carbon sink increment of the ecosystem and an atlas of the evolution of future carbon sink spatial hotspots based on the predicted sequence of vegetation and soil carbon storage over a future time period.

[0020] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0021] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A carbon sink accounting and trend prediction system based on multimodal data fusion, characterized in that, include: The set of partitioning units is used to obtain the set of point clouds in a specified area collected within the current time window, divide it into multiple grid units, and construct the corresponding height matrix; The growth parameter parsing unit is used to generate a set of candidate vertex coordinates based on the height matrix, which is used as a spatial constraint to determine the point cloud corresponding to each tree instance and then analyze it to obtain the growth parameters of each instance. The carbon distribution determination unit is used to calculate the corresponding carbon storage according to the growth parameters of each tree instance. After spatialization, it generates a vegetation carbon distribution grid for the region. The soil analysis unit is used to acquire the eDNA sequencing sequence and environmental factors of all soil sampling points in a specified area within the current time window, calculate the activity index corresponding to each grid location, read the remote sensing feature vector and acoustic feature vector in the area, fuse them, and output the total primary productivity of different grid locations. The growth increment prediction unit is used to calculate the competition coefficient of each tree instance, and then analyze the growth parameters, competition coefficient, future climate data, activity index, and total primary productivity spatial distribution raster of the tree instance to output the regional vegetation carbon storage prediction sequence. The prediction report output unit is used to determine the soil carbon storage prediction sequence using the regional vegetation carbon storage prediction sequence, and generate a net carbon sink increment prediction report according to the vegetation and soil carbon storage prediction sequences.

2. The carbon sink accounting and trend prediction system based on multimodal data fusion according to claim 1, characterized in that, The set partitioning unit includes a height extraction module and a height discrimination module; The height extraction module acquires a semantically classified set of point clouds collected within a specified region during the current time window, and divides the set into horizontal planes. OK Column grid cells For unit Extraction at All points within ,in express The first in Data points, Indicates the first One grid cell, Indicates the total number of grid cells. Indicates the total number of rows. Indicates the total number of columns. express Total number of internal data points ,statistics middle axis coordinates Calculate height ,in , express of axis coordinate values, express The first in Data points of The axis coordinates are traversed sequentially. Output the corresponding height ,in express Height; The height discrimination module selects ,like If the height value is invalid, then its corresponding unit is removed. If the flag bit is set to 0, If the height value is valid, the flag bit is set to 1. If the height value is abnormal, the flag bit is set to 2. This process is repeated until the flag bit of each grid cell is determined. After counting the heights of all cells with the flag bit set to 1, the average height is taken as the background value. These are the upper and lower limits of the height, respectively.

3. The carbon sink accounting and trend prediction system based on multimodal data fusion according to claim 2, characterized in that, The set partitioning unit also includes an interpolation height analysis module and a height correction module; The interpolation height analysis module iterates through each cell whose flag bit is not 1. Search unit neighborhood unit set ,like If there is a neighboring cell with a flag bit set to 1, then the value is calculated based on the neighboring cell. interpolation height ,in , Represents a set The neighboring unit with the marker bit set to 1 in the middle, express and The reciprocal of the distance between them express height, Indicates the first A cell whose flag bit is not 1, This indicates the total number of cells whose flag bit is not 1. Indicates the flag bit, Indicates the serial number, if If there is no neighboring cell with a flag of 1, then the interpolation height is... Assign the background value, and iterate through them sequentially. Output the corresponding interpolation height ,in express The interpolation height; The height correction module uses the cell heights with a flag of 1 and the interpolated cell heights with flags of 0 and 2, arranged in cell order, to generate the interpolated height sequence. Select unit ,Sure of Calculate the median of the height values ​​of all cells within the neighborhood window. ,like ,but The interference point is replaced with Make corrections if necessary, otherwise do nothing; repeat the process until all cells are selected. To control the threshold, the height sequence is filled into the height matrix of the point cloud set in row-major order. middle.

4. The carbon sink accounting and trend prediction system based on multimodal data fusion according to claim 1, characterized in that, The growth parameter analysis unit includes a matrix creation module, a position setting module, a vertex aggregation module, a key point determination module, a tree height calculation module, and a diameter at breast height (DBH) analysis module. The matrix creation module for the height matrix Perform a two-dimensional Gaussian convolution filter to generate a smoothed height matrix. and create a Binary matrices of the same size As a potential vertex labeling matrix; The location setting module iterates through... The height value of each position within the current location and all positions in its corresponding neighboring window is determined. If the height value of the current position is greater than the height values ​​of all positions in its neighboring window and the lower height limit, then... The corresponding position is set to 1, otherwise it is 0; The vertex aggregation module for After filtering, all positions with a value of 1 in the statistical matrix are counted, and the coordinates of the center point of the corresponding cell are calculated. All coordinates are then summarized to form a set of candidate vertex coordinates. The key point determination module calls the pre-trained single tree segmentation model, inputs the point cloud set into the model, predicts the tree instance to which each point belongs through internal calculation, generates a set of initial key points, uses the candidate vertex coordinate set as spatial constraints to optimize the model results, and outputs a set of point cloud instances. The tree height calculation module reads the first point cloud instance from the point cloud instance set. Tree example ,Sure The horizontal projection centroid is used as the center to define the corresponding search area. Within this area, data points with the classification label "ground" are selected from the point cloud set and stored. The corresponding set of ground points In China, utilizing and Calculate tree height ,in , Representing data points of axis coordinate values, Representing data points of axis coordinate values, Indicates the serial number; The chest diameter analysis module from Points whose height falls within a preset constraint range are selected, projected onto a horizontal plane to obtain a two-dimensional point set, and then fitted to the two-dimensional point set. The optimal circle diameter generated by the fitted circle is used as the reference point. Breast diameter, All points projected onto the horizontal plane will The horizontal projection centroid coordinates are taken as the center position of the crown. The distance values ​​from all projection points to the center position are analyzed, and the maximum distance value is taken as the crown radius.

5. The carbon sink accounting and trend prediction system based on multimodal data fusion according to claim 1, characterized in that, The carbon distribution determination unit includes a carbon storage analysis module, a benchmark value analysis module, and a carbon distribution generation module; The carbon storage analysis module is based on tree examples. The corresponding varietal growth coefficient can be obtained by querying the horizontal projection centroid in the tree information database. ,according to and chest diameter Tree height Calculate the material coefficient ,in , Indicates the sequence number, reads the current carbon conversion coefficient. ,according to and Analysis of carbon reserves ,in ; The benchmark analysis module repeats the operation until the carbon storage of each tree instance is calculated, and then the carbon storage of the vegetation in the current time window is accumulated to obtain the total carbon storage of the vegetation in the region. After setting the spatial attributes of the raster, the carbon distribution generation module statistically analyzes the crown center position, crown radius, and carbon storage of each tree instance. It then spatializes the data using kernel density estimation, calculates the bandwidth based on the median crown radius, and controls the smooth diffusion of carbon storage from the crown center of the tree instance outward based on the bandwidth. The module calculates the carbon density and total storage of each raster cell within the region and clips it using a region boundary mask, thereby generating a spatially continuous vegetation carbon distribution raster.

6. The carbon sink accounting and trend prediction system based on multimodal data fusion according to claim 1, characterized in that, The soil analysis unit includes an abundance calculation module, an activity index calculation module, a feature determination module, and a feature splicing module. The abundance calculation module obtains the eDNA sequencing sequences and environmental factors of all soil sampling points in the region within the current time window, for each individual sampling point. Count the number of sequences belonging to different key functional groups. And calculate the relative sequence abundance of each functional group. ,in , Indicates the first The number of sequences in each key functional group. Indicates the number of key functional groups. Indicates the first The number of sequences in each key functional group. Indicates the serial number. Indicates the first The relative sequence abundance of the key functional groups; After determining the relative sequence abundance of functional groups corresponding to each sampling point, the activity index calculation module constructs an abundance matrix. The abundance matrix and environmental factors are normalized, and then fused to generate the activity index for each sampling point. This index is then spatialized, and a raster of the spatial distribution of the activity index within the region is output. Specifically, the activity index is: in, Indicates the first Normalized relative abundance of key functional groups Indicates the first Normalized values ​​of environmental factors, Indicates at the sampling point The activity index at the site, Indicates the first The weighting coefficients of each key functional group Indicates the first The weighting coefficients of each environmental factor. Indicates the serial number. This represents the total number of environmental factors; The feature determination module acquires standardized audio data from all sound sampling points in the region within the current time window, cuts continuous audio into segments of fixed duration, generates a Mel spectrogram for each segment, analyzes it, determines the high-dimensional feature vectors corresponding to different spectrograms, performs time averaging on all segment feature vectors at each sampling point to obtain the corresponding acoustic feature vector, and generates an acoustic feature spatial distribution grid within the region by spatial interpolating the feature vectors of all sampling points. The feature stitching module reads the remote sensing feature vectors within the area, and stitches the acoustic feature vectors with the remote sensing feature vectors at each spatial grid location to form a fused feature vector. It then analyzes the fused feature vector and outputs the total primary productivity at different spatial grid locations.

7. The carbon sink accounting and trend prediction system based on multimodal data fusion according to claim 1, characterized in that, The growth increment prediction unit includes a competing tree selection module, a competition coefficient analysis module, a competition coefficient output module, and an increment calculation module; The competing wood selection module for the first Tree example ,Sure Horizontal projection centroid ,by Centered on a target area, define a corresponding search region, and select all other tree instances whose horizontal projection centroids fall within this region, and use them as the target area. Potential competing wood; The competition coefficient analysis module is selected Potential competing wood ,calculate and Distance between ,in , Indicating potential competing wood The horizontal projection centroid, Indicates the serial number, according to and Breast diameter Analysis of potential competing wood Corresponding competition coefficient ,in ; After determining the competition coefficients of all potential competing trees, the competition coefficient output module sums them up to obtain the result. Competition coefficient Iterate through each tree instance and calculate the competition coefficient for each tree instance; The incremental calculation module uses the spatial distribution grid of total primary productivity within the current time window as a baseline constraint to optimize the internal physiological parameters in the pre-trained growth response model. It inputs the spatial distribution grid of tree instance diameter at breast height (DBH), tree height, competition coefficient, future climate data, and activity index into the growth response model for analysis, outputs the DBH and height increments of tree instance, and calculates the predicted sequence of regional vegetation carbon storage in the future time period based on the DBH and height increments.

8. The carbon sink accounting and trend prediction system based on multimodal data fusion according to claim 1, characterized in that, The prediction report output unit includes a prediction sequence analysis module and a report generation module; The prediction sequence analysis module uses historical soil carbon storage data in the region and performs regression analysis with vegetation carbon storage data and climate data in the same region during the same period to construct an empirical transfer function with vegetation dynamics and climate conditions as driving factors and soil carbon storage as the response variable. By coupling the vegetation carbon storage prediction sequence and climate data in the future time period through this function, the soil carbon storage prediction sequence in the future time period is estimated. The report generation module generates an ecosystem net carbon sink increment forecast report and a future carbon sink spatial hotspot evolution atlas based on the vegetation and soil carbon storage prediction sequence for the future time period.

Images