Agricultural carbon asset value evaluation method based on multi-modal fusion

Abstract

The application discloses a kind of agricultural carbon asset value evaluation methods based on multi-modal fusion, it is related to the technical field of agricultural resources and environmental economics, including the following steps: S1, generate effective reconstruction image data;S2, output multi-modal primary feature vector sequence;S3, by improving VGAE model, based on dynamic adjacency matrix is encoded to obtain latent variable representation, introduce soft nuclear potential induced distribution calibration mechanism and combine topological consistency evolution trajectory to carry out node level adaptive optimization and cleaning;S4, output agricultural carbon sink potential depth feature;S5, generate standardized agricultural physical carbon storage numerical value;S6, output agricultural carbon asset real-time unit value;S7, generate agricultural carbon asset value evaluation report.The application overcomes the limitation that traditional method data fusion is difficult, signal-to-noise ratio is low and space-time dynamic correlation is ignored, provides efficient solution for fine evaluation of agricultural carbon assets.

Description

Technical Field

[0001] This invention relates to the field of agricultural resource and environmental economics, and in particular to a method for assessing the value of agricultural carbon assets based on multimodal fusion. Background Technology

[0002] With the increasing demand for refined agricultural carbon sequestration monitoring, traditional remote sensing technologies face severe data quality challenges when dealing with missing optical data and non-stationary environmental disturbances in cloudy and rainy areas. Existing inversion estimation models, such as process-based ecological simulation methods, while utilizing biophysical mechanisms to construct a carbon flux calculation framework, primarily rely on single-phase vegetation index statistics for parameter substitution and empirical fitting. This purely statistical approach ignores the complex spatiotemporal topology (such as the dynamic evolution of land cover and the complementary correlations between multiple sensor sources) and deep physical radiative transfer mechanisms implicit in the original multimodal data. This leads to semantic distortion in the high-dimensional feature space when calculating carbon sequestration potential parameters, thus limiting the accuracy of agricultural carbon asset assessment. Furthermore, classical estimation methods often struggle to fully utilize the dynamic constraints in environmental monitoring data to adaptively adjust estimation weights, resulting in significantly biased assessment results when dealing with extreme weather or abrupt changes in soil environment, increasing the decision-making risk in agricultural carbon asset management.

[0003] Therefore, how to provide a multimodal fusion-based method for assessing the value of agricultural carbon assets is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a multimodal fusion-based method for assessing the value of agricultural carbon assets. Through a multimodal feature evolution and fusion process based on an improved VGAE model, multi-source satellite imagery and ground sensor data are acquired. Effective reconstructed image data is generated through cloud detection and temporal reconstruction. Multi-dimensional deconstruction is performed on the effective reconstructed image data to output a sequence of primary multimodal feature vectors, and a dynamic topology graph structure is constructed and input into the improved VGAE model. In the improved VGAE model, latent variable representations are obtained based on a dynamic adjacency matrix through variational encoding. A distribution calibration mechanism induced by soft-core potential energy is introduced, and the attraction and repulsion exponents of the feature distances between nodes are dynamically mapped using a multilayer perceptron. The resultant force on each node is calculated comprehensively, and the node position is updated to energy equilibrium along the negative gradient direction. Probabilistic decoding and topology error calibration are performed on the physically optimized latent variable representations, outputting a consistent calibrated feature sequence. Physical parameter calculation and environmental adaptive weighted fusion are performed on the consistent calibrated feature sequence to output deep features of agricultural carbon sink potential. This invention analyzes the deep characteristics of agricultural carbon sequestration potential, extracts net primary productivity and soil organic carbon density through semantic mapping, and incorporates environmental and soil parameters for aggregation and correction to generate standardized agricultural physical carbon storage values. Using these agricultural physical carbon storage values ​​as the underlying asset, multi-dimensional value influencing factors are integrated, and a nonlinear dynamic weighted calculation is performed to output the real-time unit value of agricultural carbon assets. A comprehensive statistical analysis is conducted on the real-time unit value of agricultural carbon assets and physical carbon storage values, and the contribution of multi-source data is analyzed based on a dynamic gating mechanism to generate an agricultural carbon asset valuation report. This invention overcomes the limitations of traditional methods, such as difficulty in data fusion, low signal-to-noise ratio, ambiguous semantic mapping, and neglect of spatiotemporal dynamic correlations, providing an efficient solution for the refined assessment of agricultural carbon assets.

[0005] According to an embodiment of the present invention, an agricultural carbon asset valuation method based on multimodal fusion specifically includes: S1. Acquire multi-source satellite imagery and ground sensor data, perform cloud detection on satellite imagery and repair optical pixels based on temporal reconstruction, adaptively call radar imagery reconstruction based on the validity of optical data, and generate effective reconstructed image data through fusion processing; S2. Perform multi-dimensional deconstruction on the effectively reconstructed image data and ground sensor data, extract image spatial features and temporal dynamic features, and output a multimodal primary feature vector sequence after feature mapping and alignment; S3. Construct a dynamic topological graph structure for the multimodal primary feature vector sequence, input the improved VGAE model, obtain the latent variable representation based on the dynamic adjacency matrix through variational encoding, introduce a distribution calibration mechanism induced by soft kernel potential energy, and perform node-level adaptive optimization and cleaning in combination with the topological consistency evolution trajectory, and output the feature sequence after consistency calibration. S4. Perform physical parameter calculation and signal-to-noise ratio evaluation on the feature sequences after consistency calibration. After environmental adaptive weighted fusion to weaken the influence weight of low signal-to-noise ratio modes, output the deep features of agricultural carbon sink potential. S5. Analyze the deep characteristics of agricultural carbon sequestration potential, extract net primary productivity and soil organic carbon density through semantic mapping, introduce environmental and soil parameters as physical correction factors for aggregation correction, and generate standardized agricultural physical carbon storage values. S6. Based on agricultural physical carbon storage values ​​as the underlying asset, and integrating multi-dimensional value influencing factors, the real-time unit value of agricultural carbon assets is output through non-linear dynamic weighted calculation. S7. Perform comprehensive statistical analysis on the real-time unit value and physical carbon storage of agricultural carbon assets, analyze the contribution of multi-source data based on the dynamic gating mechanism, output the all-weather assessment status and efficiency improvement decisions, and generate an agricultural carbon asset value assessment report.

[0006] Optionally, S1 specifically includes: S11. Perform cloud detection on multi-source satellite imagery, identify the location of damaged pixels based on the cloud detection results, and perform repair on damaged pixels through temporal reconstruction to generate an optical pixel repair sequence. S12. Calculate the statistical distribution of pixel confidence in the optical pixel restoration sequence to determine the confidence threshold, and mark the regions in the optical pixel restoration sequence that are below the confidence threshold as low-confidence regions. S13. For low-confidence areas, perform reconstruction processing on radar images in multi-source satellite imagery to generate radar reconstructed images covering the low-confidence areas. S14. Perform spatiotemporal registration between the optical pixel restoration sequence and the radar reconstructed image. Calculate the pixel-level fusion weight matrix based on the confidence threshold of the optical pixel restoration sequence and the confidence level of the radar reconstructed image, and perform pixel-level weighted fusion to output effective reconstructed image data.

[0007] Optionally, S2 specifically includes: S21. Divide the effectively reconstructed image data into regular grid units, perform texture statistics and gradient direction statistics for each regular grid unit, extract texture feature vectors and gradient feature vectors, and concatenate the texture feature vectors and gradient feature vectors to generate grid-level spatial features. S22. Based on the timestamp index, the image data is effectively reconstructed, a multi-frame image sequence is constructed, the temporal brightness change statistics of the corresponding pixels in the multi-frame image sequence are calculated, and the time-dependent features are extracted based on the temporal brightness change statistics to generate the temporal dynamic features of the image. S23. Parse the data type labels of the ground sensor data. If the data type label is a continuous value, normalize it. If the data type label is a discrete event, perform one-hot encoding mapping and generate sensor encoding feature vectors based on the processing results. S24. Perform an outer product operation on the grid-level spatial features and the temporal dynamic features of the image to generate a feature interaction tensor. Perform summation and normalization operations on the feature interaction tensor to generate an image fusion feature vector that integrates spatiotemporal correlation information. S25. Calculate the mutual information statistics between the image fusion feature vector and the sensor-coded feature vector, construct a spatial transformation matrix based on the mutual information statistics, and use the spatial transformation matrix to project the sensor-coded feature vector onto the feature space of the image fusion feature vector, and output a multimodal primary feature vector sequence.

[0008] Optionally, the improved VGAE model includes a graph adjacency matrix construction layer, a variational graph convolutional coding layer, a molecular dynamics interaction simulation layer, and a topology error calibration layer: The graph adjacency matrix construction layer is used to dynamically calculate the feature distance and similarity between nodes in real time based on the multimodal primary feature vector sequence, adaptively construct and evolve the dynamic adjacency matrix, and perform spatial mapping of node features based on the topological association of the dynamic adjacency matrix to form a spatiotemporally aligned dynamic topological graph structure. The variational graph convolutional coding layer is used to perform continuous graph convolutional aggregation and update on the node features of the dynamic topological graph structure based on the dynamic adjacency matrix, so as to dynamically generate local high-order semantic features, and to solve the mean vector and log standard deviation vector of the latent space variables in real time through linear transformation mapping. The molecular dynamics interaction simulation layer is used to introduce a soft-nuclear potential-induced distribution calibration mechanism from computational biology. The specific execution process includes: By using reparameterization, a linear transformation is performed on the mean vector and the log-standard deviation vector to generate latent variable representations for the nodes. Calculate the feature distance between the latent variable representations of any two nodes, and dynamically map the attraction exponent coefficient and repulsion exponent coefficient based on the node features using a multilayer perceptron. These coefficients are then used as power-law adjustment parameters for the attraction and repulsion forces with respect to the feature distance. The attraction index coefficient is used to perform a reciprocal operation on the smoothed feature distance to generate attraction, and the repulsion index coefficient is used to perform an exponential operation on the smoothed feature distance to generate repulsion. The total resultant force on the node is calculated by combining the gravitational and repulsive forces, and the node position is iteratively updated along the negative gradient direction of the resultant force until the system reaches energy equilibrium. The physically optimized hidden variable representation is then output. The topology error calibration layer is used to perform probabilistic decoding on the physically optimized latent variable representation to dynamically reconstruct the topological association, and generate a topology consistency index based on the reconstruction difference evolution. Based on the dynamic evolution trajectory of the topology consistency index, adaptive node cleaning is performed, and finally, the consistent calibration feature sequence is output.

[0009] Optionally, S4 specifically includes: S41. Based on the characteristic sequence after uniform calibration, calculate the multiple scattering flux and absorption attenuation of light in the vegetation canopy. By minimizing the residual error between the measured radiation value and the simulated radiation value, iteratively adjust the canopy parameters until the residual error converges, and output the spectral parameter inversion data. S42. Calculate the time series variance of the spectral parameter inversion data, and use the time series variance to quantify the data dispersion to generate a local signal-to-noise ratio evaluation index. S43. Perform kernel density estimation on the local signal-to-noise ratio evaluation index, determine the signal-to-noise classification boundary based on the peak characteristics of the kernel density estimation, and map the local signal-to-noise ratio evaluation index to adaptive weight coefficients according to the signal-to-noise classification boundary. S44. Obtain environmental factor monitoring data, calculate the correlation coefficient between spectral parameter inversion data and environmental factor monitoring data, use the correlation coefficient to perform weighted correction on spectral parameter inversion data, and generate environmentally corrected feature data. S45. Perform weighted aggregation operation on the environmentally corrected feature data using adaptive weight coefficients to output the deep feature of agricultural carbon sequestration potential.

[0010] Optionally, S5 specifically includes: S51. Map the deep features of agricultural carbon sequestration potential to the semantic feature space of carbon sequestration, analyze the distribution clustering of feature vectors to extract the feature components of net primary productivity and soil organic carbon density, and generate a set of basic parameters for carbon sequestration. S52. Obtain temperature and humidity data and soil physicochemical parameters from environmental monitoring data, calculate the first correlation coefficient between the net primary productivity characteristic component and temperature and humidity data, calculate the second correlation coefficient between the soil organic carbon density characteristic component and soil physicochemical parameters, and construct an environmental constraint weight matrix based on the first and second correlation coefficients. S53. Perform dimensional standardization on the carbon sink basic parameter set, and use the environmental constraint weight matrix to perform weighted aggregation operation on the processed carbon sink basic parameter set to generate standardized agricultural physical carbon storage values.

[0011] Optionally, S6 specifically includes: S61. Obtain agricultural carbon sink market transaction data and agricultural policy support indicators, perform numerical normalization and dimension alignment on agricultural carbon sink market transaction data and agricultural policy support indicators, and generate a multi-dimensional value influence factor matrix. S62. Perform principal component analysis on the multidimensional value impact factor matrix to extract the feature contribution rate, calculate the dynamic weight coefficient of each dimension factor based on the feature contribution rate, and use the dynamic weight coefficient to perform weighted summation on the multidimensional value impact factor matrix to generate a comprehensive value correction index. S63. Perform a nonlinear mapping operation between the agricultural physical carbon storage value and the comprehensive value correction index, perform numerical validity verification on the result of the nonlinear mapping operation, and output the real-time unit value of agricultural carbon assets.

[0012] Optionally, S7 specifically includes: S71. Perform spatiotemporal alignment and aggregation operations on the real-time unit value of agricultural carbon assets and the standardized agricultural physical carbon storage values ​​to construct a multidimensional value time series dataset, calculate the statistical moment characteristics of the multidimensional value time series dataset, and generate a carbon asset statistical feature vector. S72. Perform correlation calculation and importance scoring on the statistical feature vector of carbon assets, calculate the dynamic attention weight of the feature dimension based on the importance score, construct the gated weighting coefficient using the dynamic attention weight, and perform weighted fusion operation on the statistical feature vector of carbon assets using the gated weighting coefficient to generate a deep state representation. S73. Calculate the historical data distribution characteristics of real-time monitoring data to generate a dynamic baseline. Perform deviation analysis based on the deep state characterization and the dynamic baseline. Generate differentiated efficiency-enhancing decision-making strategies based on the deviation analysis results. Encapsulate the deep state characterization, differentiated efficiency-enhancing decision-making strategies, and real-time monitoring data to generate an agricultural carbon asset value assessment report.

[0013] The beneficial effects of this invention are: (1) This invention achieves continuous spatiotemporal representation of all-weather agricultural monitoring data by constructing an adaptive fusion processing flow based on the validity of optical data and radar image reconstruction. The location of damaged pixels is identified using cloud detection results. Optical pixels are repaired through temporal reconstruction, and the statistical distribution of pixel confidence is calculated. Low-confidence areas are adaptively marked, and radar image reconstruction is invoked to perform coverage. Combining the spatiotemporal registration results of the optical pixel repair sequence and the radar reconstructed image, a pixel-level fusion weight matrix is ​​calculated based on the confidence level, and a weighted fusion operation is performed. This process converts heterogeneous image data affected by meteorological conditions into spatiotemporally continuous and effective reconstructed image data, eliminating monitoring blind spots and pixel-level confidence differences caused by missing optical data, and providing a foundational data with high spatiotemporal integrity and reliability.

[0014] (2) This invention establishes a physical field-guided node topology evolution and feature depth cleaning system by adopting an improved VGAE model and a distribution calibration mechanism induced by soft core potential energy. The graph adjacency matrix construction layer calculates the feature distance and similarity between nodes based on the multimodal primary feature vector sequence, and adaptively constructs and evolves the dynamic adjacency matrix. The molecular dynamics interaction simulation layer generates latent variable representations using reparameterization, dynamically maps the attraction and repulsion exponents through a multilayer perceptron, calculates the attraction and repulsion forces, and iteratively updates the node positions along the negative gradient direction of the resultant force until the system reaches an energy equilibrium state. The topology error calibration layer performs probabilistic decoding based on the physically optimized latent variable representations, generates topology consistency indices using reconstructed differential evolution, and performs adaptive node cleaning. This system achieves adaptive distribution correction from the geometric feature space to the physical latent space through dynamic topology graph construction, molecular dynamics position optimization, and topology consistency error verification, ensuring that the feature sequence after output consistency calibration maintains the consistency of local topology and global semantics. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an agricultural carbon asset valuation method based on multimodal fusion proposed in this invention; Figure 2 This is a flowchart illustrating the working principle of the improved VGAE model, a multimodal fusion-based method for assessing the value of agricultural carbon assets proposed in this invention. Detailed Implementation

[0016] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figure 1 and Figure 2 A multimodal fusion-based method for assessing the value of agricultural carbon assets, specifically including: S1. Acquire multi-source satellite imagery and ground sensor data, perform cloud detection on satellite imagery and repair optical pixels based on temporal reconstruction, adaptively call radar imagery reconstruction based on the validity of optical data, and generate effective reconstructed image data through fusion processing; S2. Perform multi-dimensional deconstruction on the effectively reconstructed image data and ground sensor data, extract image spatial features and temporal dynamic features, and output a multimodal primary feature vector sequence after feature mapping and alignment; S3. Construct a dynamic topological graph structure for the multimodal primary feature vector sequence, input the improved VGAE model, obtain the latent variable representation based on the dynamic adjacency matrix through variational encoding, introduce a distribution calibration mechanism induced by soft kernel potential energy, and perform node-level adaptive optimization and cleaning in combination with the topological consistency evolution trajectory, and output the feature sequence after consistency calibration. S4. Perform physical parameter calculation and signal-to-noise ratio evaluation on the feature sequences after consistency calibration. After environmental adaptive weighted fusion to weaken the influence weight of low signal-to-noise ratio modes, output the deep features of agricultural carbon sink potential. S5. Analyze the deep characteristics of agricultural carbon sequestration potential, extract net primary productivity and soil organic carbon density through semantic mapping, introduce environmental and soil parameters as physical correction factors for aggregation correction, and generate standardized agricultural physical carbon storage values. S6. Based on agricultural physical carbon storage values ​​as the underlying asset, and integrating multi-dimensional value influencing factors, the real-time unit value of agricultural carbon assets is output through non-linear dynamic weighted calculation. S7. Perform comprehensive statistical analysis on the real-time unit value and physical carbon storage of agricultural carbon assets, analyze the contribution of multi-source data based on the dynamic gating mechanism, output the all-weather assessment status and efficiency improvement decisions, and generate an agricultural carbon asset value assessment report.

[0018] In this embodiment, S1 specifically includes: S11. Read the grayscale value of each pixel in the multi-source satellite image. The grayscale value of the pixel ranges from 0 to 255. If the grayscale value is greater than 240, the pixel is identified as a cloud pixel. Mark the coordinate position of the cloud pixel in the image matrix as the damaged pixel position. Extract the grayscale value of the damaged pixel position in the time dimension of the previous 10 frames of images. Calculate the weighted time decay average of the grayscale values ​​of the previous 10 frames as the repair value. Fill the repair value into the damaged pixel position of the current frame image to generate an optical pixel repair sequence.

[0019] S12. Calculate the time series standard deviation of the grayscale value of each pixel in the optical pixel restoration sequence, set the confidence threshold to 15, and the unit is grayscale level; mark the area with a standard deviation value greater than 15 in the optical pixel restoration sequence as a low confidence area, and generate a low confidence area mask matrix. In the mask matrix, the value 1 represents the low confidence area and the value 0 represents the high confidence area.

[0020] S13. Extract radar image band values ​​from multi-source satellite imagery, perform orthorectification on the radar images to eliminate geometric distortion, and extract local texture gradient features of the radar images in low-confidence regions. Use the radar texture gradient features to guide a nonlocal mean algorithm to perform texture reconstruction on low-confidence regions in the optical pixel restoration sequence to generate a preliminary restored optical texture image. Calculate the cumulative distribution function of the radar backscattering coefficient, and use this function to map the gray values ​​in the preliminary restored optical texture image to a gray range consistent with the surrounding optical background to generate a radar reconstructed image covering the low-confidence region.

[0021] S14. Align the optical pixel restoration sequence with the radar reconstructed image in the geographic coordinate system. Calculate the time series standard deviation of the optical pixel restoration sequence. Divide the standard deviation by the sum of the gray mean and 1e-6, and take the reciprocal to obtain the basic coefficient of variability. Introduce the confidence threshold from S12 as a constraint term. If the standard deviation is greater than 15, multiply the basic coefficient of variability by a penalty coefficient of 0.2 to obtain the optical variability confidence level. Calculate the local entropy value of the radar reconstructed image in the low-confidence region. Normalize the local entropy value to obtain the radar texture confidence level. Set the value 1e-6 as a smoothing factor. Add the optical variability confidence level and the radar texture confidence level, and add the smoothing factor. Use the sum as the denominator and the optical variability confidence level as the numerator. Calculate the ratio and use it as the fusion weight of the optical pixel restoration sequence. Subtract the optical fusion weight from 1 and use the difference as the fusion weight of the radar reconstructed image. Multiply the optical pixel restoration sequence by the optical fusion weight and the radar reconstructed image by the radar fusion weight. Add the two products together and output the effective reconstructed image data.

[0022] The adaptive multi-source image fusion process proposed in this step is similar to the traditional single optical image restoration process in that it is based on spatiotemporal statistical analysis and pixel reconstruction theory. That is, it fills in missing information by utilizing the correlation of image sequences in the time dimension, identifies data quality abnormal areas based on the statistical distribution characteristics of pixel values, and uses a weighted fusion strategy to integrate multi-source information to generate continuous observation image data.

[0023] The difference lies in that this invention breaks through the limitations of traditional methods that rely solely on optical temporal interpolation, which leads to the loss of texture details and signal distortion in areas with full cloud coverage. It adds a confidence threshold determination step to accurately define low-confidence areas and replaces single optical interpolation with a radar image reconstruction mechanism. It maps the optical pixel restoration sequence and the radar reconstructed image as heterogeneous data sources, performs adaptive weight matrix operations based on pixel-level confidence differences, and finally combines the optical confidence threshold and radar confidence to output fusion weights, rather than simply using temporal mean filling.

[0024] The beneficial effects of the improvements are that, by using confidence threshold determination and radar reconstruction intervention, the hard physical features of radar penetration imaging are embedded into the optical image restoration process. This breaks the limitation of traditional methods that cannot obtain effective surface information under thick cloud cover, leading to data interruption. It achieves a precise conversion from single-modal temporal filling to multi-modal heterogeneous complementarity. This design significantly enhances the robustness of all-weather data acquisition, accurately reconstructs surface texture in low-confidence areas, and, combined with spatiotemporal registration and weighted fusion, effectively improves the integrity and spatial consistency of the effectively reconstructed image data, providing a high-quality data foundation for subsequent feature extraction.

[0025] In this embodiment, S2 specifically includes: S21. Divide the effectively reconstructed image data into regular grid units with a size of 16 pixels by 16 pixels. For each regular grid unit, calculate the variance of gray values ​​in the region as a texture feature component. Calculate the sum of the absolute values ​​of the differences of gray values ​​in the horizontal, vertical and two diagonal directions as gradient feature components. Concatenate the texture feature components and gradient feature components into a 4-dimensional vector to generate a grid-level spatial feature matrix composed of the 4-dimensional vectors of all grid units arranged in spatial order.

[0026] S22. Based on the timestamp index, the image data is effectively reconstructed, and a multi-frame image sequence containing 10 consecutive frames of images is constructed. For all pixels in each regular grid cell, the sum of the absolute values ​​of the first-order differences of the average gray value of the corresponding area in the multi-frame image sequence as a time difference is calculated as a time-series brightness change statistic. The time-series brightness change statistic is activated by a linear rectification function to generate an image time-series dynamic feature vector consistent with the number of grid cells.

[0027] S23. Parse the data type label of the ground sensor data. If the data type label is a continuous value, divide the value by the maximum measured value and multiply it by the value 1 for normalization. If the data type label is a discrete event, map it into a binary vector with a dimension of 4 and generate a sensor coded feature vector based on the processing result.

[0028] S24. Read the grid-level spatial feature matrix and the image temporal dynamic feature vector. Multiply the value corresponding to each grid in the image temporal dynamic feature vector with the four elements of the corresponding row in the grid-level spatial feature matrix to obtain the weighted grid feature matrix. Do not accumulate or compress the grid dimensions to generate an image fusion feature matrix that integrates spatiotemporal correlation information and retains spatial structure.

[0029] S25. Calculate the average value of the image fusion feature matrix in the grid dimension to obtain the global fusion feature vector. Calculate the Pearson correlation coefficient between the global fusion feature vector and the sensor-coded feature vector and take the absolute value as the mutual information statistic. Calculate the outer product matrix between the global fusion feature vector and the sensor-coded feature vector. Multiply the outer product matrix with the mutual information statistic to construct a dimension-matched off-diagonal spatial transformation matrix. Use the spatial transformation matrix to perform a linear transformation mapping on the sensor-coded feature vector, project the sensor-coded feature vector onto the feature space of the image fusion feature vector, and output a multimodal primary feature vector sequence.

[0030] In this embodiment, the improved VGAE model includes a graph adjacency matrix construction layer, a variational graph convolutional coding layer, a molecular dynamics interaction simulation layer, and a topology error calibration layer: The graph adjacency matrix construction layer reads each feature vector in the multimodal primary feature vector sequence as a graph node, calculates the Euclidean distance between all pairs of nodes as the feature distance, substitutes the feature distance into the Gaussian radial basis function to calculate the similarity, sets the Gaussian kernel width coefficient to 1.0, and calculates the exponential function value of the negative square of the feature distance divided by the coefficient 1.0 as the node similarity; sets the adjacency threshold to 0.5. If the node similarity is greater than 0.5, a connection is established between the corresponding nodes, and the similarity value is filled into the corresponding position in the matrix; otherwise, the value is filled into 0, thus constructing a dynamic adjacency matrix; the dynamic adjacency matrix is ​​used to index and aggregate the multimodal primary feature vector sequence to generate a spatiotemporally aligned dynamic topological graph structure.

[0031] The variational graph convolutional coding layer is used to construct a variational encoder network containing two graph convolutional layers. The activation function is set to a linear rectified function. The node features of the dynamic topological graph structure are input into the first graph convolutional layer, and the multiplication operation between the dynamic adjacency matrix and the node feature matrix is ​​performed. The result is input into the second graph convolutional layer, and matrix multiplication and linear rectified function activation are performed again to output local high-order semantic features. A fully connected layer is constructed with 32 neurons. The local high-order semantic features are input into the fully connected layer for linear transformation, and the mean vector and log-standard deviation vector of the latent space variables are output respectively.

[0032] The molecular dynamics interaction simulation layer is used to introduce a soft-nuclear potential-induced distribution calibration mechanism from computational biology. The specific execution process includes: Standard normal distribution noise is introduced as a random disturbance. The standard deviation is set to 1. The standard deviation vector is obtained by calculating the exponential function value of the log-standard deviation vector. The sum of the product of the mean vector and the standard deviation vector with the standard normal distribution noise is calculated to generate the latent variable representation of the node. Calculate the Euclidean distance between the latent variable representations of any two nodes as the feature distance, construct a multilayer perceptron with 2 input layer nodes, 8 hidden layer nodes, and 2 output layer nodes, use the hyperbolic tangent function as the activation function, input the feature distance into the multilayer perceptron, and output the attraction exponent coefficient and the repulsion exponent coefficient; perform the absolute value operation on the repulsion exponent coefficient to ensure that it is a non-negative number.

[0033] Set the smoothing factor to a value of 1e-5, calculate the sum of the feature distance and the smoothing factor value of 1e-5 as the smoothed distance; calculate the value of 1 divided by the smoothed distance, and multiply the result by the attraction exponent coefficient as the attraction force; calculate the exponential function value of the smoothed distance, and multiply the result by the repulsion exponent coefficient as the repulsion force. The difference between the attractive and repulsive forces is calculated to obtain the total resultant force on the node. The learning rate is set to 0.01. The latent variable representation of the node is updated by multiplying the learning rate of 0.01 by the total resultant force. This process is repeated until the total resultant force is less than 0.001. The physically optimized latent variable representation is then output.

[0034] The topology error calibration layer is used to construct the fully connected layer of the inner decoder. The activation function is set to the Sigmoid function. The physically optimized latent variable representation is input into the fully connected layer of the inner decoder, and the decoder outputs the reconstructed node probability connection matrix. The squares of the differences between the corresponding elements of the reconstructed node probability connection matrix and the dynamic adjacency matrix are calculated and summed to obtain the topology reconstruction error. The error cleaning threshold is set to 0.1. If the topology reconstruction error of a node is greater than 0.1, the node is determined to be an outlier and is removed from the feature sequence. Otherwise, the original feature value is retained, and the final output is the consistent calibrated feature sequence.

[0035] The improved VGAE model proposed in this step is similar to the traditional VGAE model in that it is based on graph neural network representation learning and probability generation theory. That is, by constructing graph structure data, the node features of non-Euclidean space are projected to the latent space for topological information aggregation. The variational inference mechanism is used to learn the posterior probability distribution of latent variables, and both use a decoder to reconstruct the graph connection relationship to verify the effectiveness of feature extraction.

[0036] The difference lies in that this invention breaks the limitation of traditional VGAE latent space sampling, which ignores the principle of maintaining the lowest physical field energy by ignoring the structure. It adds a molecular dynamics interaction simulation layer to introduce a distribution calibration mechanism induced by soft nucleus potential energy, and replaces the standard random sampling process with physical field force optimization. It maps the node feature distance to the attraction exponent coefficient and the repulsion exponent coefficient, calculates the total resultant force of attraction and repulsion and updates the node position along the negative gradient direction until the energy equilibrium state is reached. Finally, it outputs the physically optimized latent variable representation, rather than a simple reparameterized sampling result.

[0037] The beneficial effects of the improvements are that, through soft-core potential energy induction and resultant force iterative updates, the molecular dynamics energy constraint is rigidly embedded into the latent spatial evolution process, breaking the limitations of traditional methods that are prone to topological collapse or mode confusion when feature distributions are crowded or discrete. This achieves a precise conversion from pure probability distribution mapping to physical field energy balance. The design significantly enhances the ability of the feature manifold to preserve local geometry, can accurately correct the distribution of latent variables in the energy equilibrium state, and, combined with topological error calibration, effectively improves the global structural integrity and semantic recognizability of the feature sequence after consistency calibration.

[0038] In this embodiment, S4 specifically includes: S41. Read the feature sequence after consistency calibration, use the feature sequence as the canopy leaf area index, set the solar zenith angle to 30 degrees, calculate the transmittance of light passing through the canopy using the Beer-Lambert law, calculate the absorption attenuation by subtracting the transmittance from the value 1, calculate the product of the absorption attenuation and the hemispherical reflectance as the multiple scattering flux, add the absorption attenuation and the multiple scattering flux to generate the simulated radiation value; calculate the difference between the simulated radiation value and the measured radiation value as the residual error, use the stochastic gradient descent algorithm to iteratively adjust the canopy parameters with the goal of reducing the residual error, set the learning rate to 0.01, repeat the iteration until the residual error is less than 0.001, and output the spectral parameter inversion data.

[0039] S42. Calculate the variance of the spectral parameter inversion data in the time dimension as the time series variance, calculate the square root of the time series variance as the standard deviation, calculate the mean of the spectral parameter inversion data, calculate the ratio of the mean to the standard deviation, and generate a local signal-to-noise ratio evaluation index.

[0040] S43. Substitute the local signal-to-noise ratio (SNR) evaluation index into the Gaussian kernel function, set the kernel width coefficient to 0.5, calculate the distance between the local SNR evaluation index and historical sample points, and substitute it into the Gaussian kernel function to calculate the kernel density value. Summate the values ​​to obtain the kernel density estimation curve. Solve for the derivative of the kernel density estimation curve and find the point where the derivative is zero as the peak feature point. Use the peak feature point as the SNR classification boundary. If the local SNR evaluation index is greater than the SNR classification boundary, set the adaptive weight coefficient to 1; otherwise, set it to 0.5.

[0041] S44. Obtain environmental factor monitoring data, calculate the Pearson correlation coefficient between the spectral parameter inversion data and the environmental factor monitoring data, set the scaling factor to a value of 1, use the hyperbolic tangent function to perform a nonlinear transformation on the correlation coefficient, divide it by a value of 2 and add a value of 0.5 to normalize it, and obtain the environmental correction coefficient. Multiply the spectral parameter inversion data by the environmental correction coefficient to generate the environmentally corrected feature data.

[0042] S45. Read the environmentally corrected feature data and adaptive weight coefficients, calculate the product of each element in the environmentally corrected feature data and the corresponding adaptive weight coefficient, concatenate all product results in the original order, generate a weighted multidimensional feature vector, and output the deep feature of agricultural carbon sequestration potential.

[0043] In this embodiment, S5 specifically includes: S51. Read the deep feature vector of agricultural carbon sequestration potential, construct a feature mapping network containing two fully connected layers, set the activation function to a linear rectified function, input the deep feature vector of agricultural carbon sequestration potential into the feature mapping network to perform linear transformation and nonlinear activation, and output the semantic feature vector of carbon sequestration; calculate the value of each element in the semantic feature vector of carbon sequestration and sort the values ​​from largest to smallest, select the elements corresponding to the top 20% of the largest values ​​as the net primary productivity feature components, and select the elements corresponding to the top 20% to 40% of the largest values ​​as the soil organic carbon density feature components; calculate the arithmetic mean of the net primary productivity feature components and the soil organic carbon density feature components respectively, combine the two averages to generate the basic parameter set of carbon sequestration.

[0044] S52. Obtain temperature and humidity data and soil physicochemical parameters from environmental monitoring data. Read the average net primary productivity and average soil organic carbon density from the carbon sink basic parameter set. Calculate the Pearson correlation coefficient between the average net primary productivity and temperature and humidity data as the first correlation coefficient. Calculate the Pearson correlation coefficient between the average soil organic carbon density and soil physicochemical parameters as the second correlation coefficient. Construct an identity matrix with 2 rows and 2 columns. Fill the first correlation coefficient into the first row and first column, and fill the second correlation coefficient into the second row and second column to construct an environmental constraint weight matrix.

[0045] S53. Read the carbon sink basic parameter set, calculate the sum of the values ​​of all elements in the carbon sink basic parameter set as the denominator, calculate the ratio of the value of each element to the denominator as the normalized value, and generate a dimensionally standardized dataset; read the environmental constraint weight matrix, construct the dimensionally standardized dataset into a column vector with 2 rows and 1 column, calculate the product of the environmental constraint weight matrix and the column vector, perform an accumulation and summation operation on all elements in the product result, and generate a standardized agricultural physical carbon storage value.

[0046] The carbon storage correction process based on environmental physical constraints proposed in this step is similar to the traditional direct feature inversion process in that both are based on the theory of feature semantic mapping and multidimensional parameter aggregation. That is, by projecting the abstract feature vector extracted by the deep neural network onto a specific physical index space, clustering analysis algorithm is used to analyze the spatial distribution law of the feature components, and weighted summation operation is used to aggregate the multidimensional feature components into a value representing the overall level.

[0047] The difference lies in that this invention breaks through the limitation of traditional pure feature-driven black-box mapping ignoring the interaction mechanism of environmental physicochemical factors. It adds an environmental constraint weight matrix construction step to quantify the coupling strength and replaces direct numerical output with a correlation weighted aggregation mechanism. It maps environmental monitoring data into a first correlation coefficient and a second correlation coefficient to perform weight calculation. Finally, it combines the carbon sink basic parameter set and the environmental constraint weight matrix to generate a standardized agricultural physical carbon storage value, rather than a single linear regression prediction.

[0048] The beneficial effects of the improvements are that, through correlation analysis and weighted aggregation, this invention rigidly embeds environmental physical constraints into the carbon storage accounting process, breaking the limitations of traditional methods that suffer from valuation distortion due to the lack of physical correction under sudden environmental changes. It achieves a precise conversion from data-driven inversion to strong physical mechanism constraints. This design significantly enhances the resilience to the impact of environmental factor fluctuations, can accurately calibrate the basic parameters of carbon sinks in the physical constraint space, and, combined with dimensional standardization, effectively improves the physical interpretability and valuation accuracy of agricultural physical carbon storage values.

[0049] In this embodiment, S6 specifically includes: S61. Read the agricultural carbon sink market transaction data and agricultural policy support indicators, calculate the maximum and minimum values ​​of the agricultural carbon sink market transaction data and agricultural policy support indicators, and use the maximum and minimum values ​​to perform minimum-maximum normalization processing on the agricultural carbon sink market transaction data and agricultural policy support indicators; align the normalized data by timestamp, and concatenate the agricultural carbon sink market transaction data as the first column and the agricultural policy support indicators as the second column to construct a multidimensional value influence factor matrix with the number of rows equal to the length of the time series and the number of columns equal to 2.

[0050] S62. Calculate the covariance matrix of the multidimensional value influence factor matrix, solve for the eigenvalues ​​and eigenvectors of the covariance matrix, divide the eigenvalues ​​by the sum of all eigenvalues ​​to obtain the variance contribution rate as the feature contribution rate; construct a 1-row, 2-column zero vector, fill the corresponding positions of the zero vector with the feature contribution rates to generate the feature weight vector, calculate the inner product of each row of data in the multidimensional value influence factor matrix and the feature weight vector, and generate a column vector containing elements of the time series length as the comprehensive value correction index.

[0051] S63. Construct a nonlinear mapping network containing three fully connected layers, setting the activation function to the Sigmoid function. The input layer receives the agricultural physical carbon storage value and the comprehensive value correction index, performs linear transformation and nonlinear activation, and outputs an initial value coefficient with a value range of 0 to 1. Calculate the product of the initial value coefficient and the value of 200 yuan per ton. If the product result is less than 0, correct it to 0. If the product result is greater than 200 yuan per ton, correct it to 200 yuan per ton. Output the verified real-time unit value of agricultural carbon assets.

[0052] In this embodiment, S7 specifically includes: S71. Read the real-time unit value of agricultural carbon assets and the standardized agricultural physical carbon storage value. Calculate the product of the real-time unit value of agricultural carbon assets and the agricultural physical carbon storage value as the total asset valuation. Sort the total asset valuation by timestamp to construct a multi-dimensional value time series dataset with rows equal to the length of the time series. Calculate the sum of the values ​​of all data points in the multi-dimensional value time series dataset and divide it by the number of data points to obtain the mean. Calculate the square of the difference between the value of each data point in the multi-dimensional value time series dataset and the mean, sum them up, and divide it by the number of data points to obtain the variance. Calculate the fourth central moment of the multi-dimensional value time series dataset as the kurtosis. Concatenate the mean, variance, and kurtosis in order to generate a carbon asset statistical feature vector.

[0053] S72. Read the carbon asset statistical feature vector and calculate the Pearson correlation coefficient between the carbon asset statistical feature vector and the preset label vector as the importance score; construct an attention network containing two fully connected layers, set the activation function to the hyperbolic tangent function, input the importance score into the attention network, perform linear transformation and nonlinear activation, and output dynamic attention weights with the same dimension as the carbon asset statistical feature vector; calculate the Sigmoid function value of the dynamic attention weights as the gating weighting coefficients, calculate the Hadamard product of the gating weighting coefficients and the carbon asset statistical feature vectors, and generate a deep state representation.

[0054] S73. Read real-time monitoring data, calculate the mean, variance, and kurtosis of the real-time monitoring data within the past 30-day time window, and construct a historical statistical feature vector as a dynamic baseline; calculate the difference between the corresponding elements of the deep state characterization and the dynamic baseline and take the absolute value, sum all the absolute values ​​of the differences as the deviation analysis result, set the deviation threshold to a value of 10, if the deviation analysis result is greater than a value of 10, generate an efficiency-enhancing strategy to increase carbon sink input, otherwise generate an efficiency-enhancing strategy to maintain the status quo; package the deep state characterization, efficiency-enhancing strategy, and real-time monitoring data to generate an agricultural carbon asset value assessment report.

[0055] The decision generation process based on dynamic baseline deviation analysis proposed in this step is similar to the traditional static asset valuation process in that both are based on statistical analysis and multidimensional feature mapping theory. That is, a dataset is constructed by aligning the unit value of the asset with the physical reserve value in time and space, the macroscopic distribution characteristics of the data are extracted by using the statistical moment calculation method, and a weighted fusion strategy is used to integrate multidimensional features to generate a value representing the current state.

[0056] The difference lies in that this invention breaks through the limitations of traditional single numerical evaluation that ignores historical evolution trends and dynamic decision-making needs. It adds a dynamic baseline construction step to quantify the distribution characteristics of historical data, and replaces static threshold judgment with a deep state representation and baseline deviation analysis mechanism. Real-time monitoring data is mapped to a deep state representation, and the deviation vector between this representation and the dynamic baseline is calculated to implement the strategy mapping. Finally, the deviation analysis results are combined to generate a differentiated efficiency-enhancing decision strategy, rather than a single static evaluation report.

[0057] The beneficial effects of the improvements are that, through the construction of dynamic baselines and deviation analysis, this invention rigidly embeds the temporal evolution law into the decision-making process, breaking the limitations of traditional methods that lack forward-looking guidance in dealing with market and environmental fluctuations, leading to management lags. It achieves a precise transformation from static value accounting to dynamic decision support. This design significantly enhances the ability to perceive asset state drift, accurately locates efficiency-enhancing opportunities in the deviation space, and, combined with gated weighted fusion, effectively improves the timeliness and decision-making guidance value of agricultural carbon asset assessment reports.

[0058] Example 1: To verify the feasibility of this invention in agricultural carbon sink monitoring and asset management, the method of this invention was applied to the intelligent management platform for agricultural carbon assets of a provincial agricultural green development investment company (hereinafter referred to as "Company A"). In traditional agricultural carbon sink assessment systems, statistical models based on single-temporal remote sensing images or ground-based inventory methods are typically used. These methods not only struggle to obtain continuous and effective optical data under cloudy and rainy weather conditions, but also fail to accurately capture the dynamic coupling characteristics of soil and vegetation in complex environments, easily leading to distortions in carbon storage estimation and delays in value assessment. To solve these problems, Company A decided to adopt the multimodal fusion-based agricultural carbon asset valuation method proposed in this invention.

[0059] During implementation, Company A first utilized multi-source satellite remote sensing nodes and ground-based IoT weather stations deployed at its agricultural demonstration base to acquire multispectral satellite image streams and ground environmental sensor data streams covering the entire crop growth cycle. After preprocessing operations such as cloud detection, temporal reconstruction and restoration, and adaptive fusion of radar images, effectively reconstructed image data containing both spatial texture and temporal dynamic dimensions was constructed. Simultaneously, Company A's agronomic experts precisely labeled key carbon sink parameters and measured soil physicochemical properties on the collected multi-source data, serving as a benchmark for model training and value verification.

[0060] Company A extracted grid-level spatial features, temporal brightness dynamic features, and sensor-encoded feature vectors from the image using a multi-dimensional deconstruction module. A multimodal primary feature vector sequence was constructed using feature interaction tensors and mutual information spatial transformation. Next, a dynamic topology graph structure was built and input into an improved VGAE model. Latent variable representations were obtained through variational graph convolutional encoding, and a soft-core potential-induced distribution calibration mechanism from computational biology was introduced. The attraction and repulsion exponential coefficients of the feature distances between nodes were dynamically mapped using a multilayer perceptron. The resultant force was calculated comprehensively, and node positions were iteratively updated along the negative gradient direction to achieve energy equilibrium, thus optimizing the physical field of the latent variable distribution. Subsequently, node-level adaptive optimization and cleaning were performed based on the topology consistency evolution trajectory, outputting a consistent calibrated feature sequence.

[0061] In the core assessment and decision-making stage, this invention performs physical parameter calculations and signal-to-noise ratio assessments on the consistent calibrated feature sequences, and outputs in-depth features of agricultural carbon sink potential through environmentally adaptive weighted fusion. The system further analyzes these features, extracting net primary productivity and soil organic carbon density via semantic mapping, and introduces environmental and soil parameters as physical correction factors to generate standardized agricultural physical carbon storage values. Subsequently, it integrates multi-dimensional value influencing factors and outputs the real-time unit value of agricultural carbon assets through nonlinear dynamic weighted calculations. Finally, based on a dynamic gating mechanism, it analyzes the contribution of multi-source data, calculates historical data distribution characteristics to generate a dynamic baseline for deviation analysis, and generates differentiated efficiency-enhancing decision-making strategies and agricultural carbon asset value assessment reports, achieving a closed-loop leap from data perception to value decision-making.

[0062] Company A's technical team discovered during implementation that, compared to traditional manual statistics and conventional remote sensing estimation methods, the method of this invention significantly improves the accuracy and timeliness of agricultural carbon asset valuation. Traditional methods struggle to handle continuous monitoring in situations where optical data is missing and are slow to respond to dynamic changes in carbon sinks under complex environments. In contrast, the method of this invention, through improved topological distribution calibration of the VGAE model, multimodal spatiotemporal feature fusion, and dynamic baseline deviation analysis, effectively achieves accurate carbon storage inversion and real-time dynamic assessment of asset value.

[0063] To further verify the actual performance of the method of the present invention, Company A conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1. Performance Comparison of Agricultural Carbon Asset Valuation Methods for Company A

[0064] As shown in Table 1, the performance of the agricultural carbon asset intelligent management platform was comprehensively improved after applying the method of this invention. The relative error of carbon storage estimation decreased from 15.2% with traditional methods to 3.5%, and the missing rate of cloudy weather data dropped significantly from 35.0% to 2.1%, significantly improving the continuity and completeness of data acquisition and laying the foundation for accurate assessment. The asset valuation time was shortened from 30 days to 1 day, representing a qualitative leap in system timeliness. The response time to abnormal conditions was reduced from 48 hours to 0.5 hours, effectively addressing sudden environmental risks. Furthermore, the matching degree of dynamic decision-making strategies increased from 75.0% to 94.5%, and the carbon asset operation rate of return increased from 5.2% to 8.9%, significantly improving economic benefits.

[0065] Through the method of this invention, Company A has successfully realized the in-depth exploration of agricultural carbon sink potential and the dynamic and transparent management of asset value. It effectively solves the problems of difficulty in integrating multi-source heterogeneous data, lagging assessment, and single decision-making, ensuring the authenticity and tradability of agricultural carbon assets, greatly improving the intelligence and digitalization level of agricultural carbon management, significantly enhancing the company's competitiveness in the carbon trading market, and providing strong technical support for the green and low-carbon development of agriculture.

[0066] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing the value of agricultural carbon assets based on multimodal fusion, characterized in that, Includes the following steps: S1. Acquire multi-source satellite imagery and ground sensor data, perform cloud detection on satellite imagery and repair optical pixels based on temporal reconstruction, adaptively call radar imagery reconstruction based on the validity of optical data, and generate effective reconstructed image data through fusion processing; S2. Perform multi-dimensional deconstruction on the effectively reconstructed image data and ground sensor data, extract image spatial features and temporal dynamic features, and output a multimodal primary feature vector sequence after feature mapping and alignment; S3. Construct a dynamic topological graph structure for the multimodal primary feature vector sequence, input the improved VGAE model, obtain the latent variable representation based on the dynamic adjacency matrix through variational encoding, introduce a distribution calibration mechanism induced by soft kernel potential energy, and perform node-level adaptive optimization and cleaning in combination with the topological consistency evolution trajectory, and output the feature sequence after consistency calibration. S4. Perform physical parameter calculation and signal-to-noise ratio evaluation on the feature sequences after consistency calibration. After environmental adaptive weighted fusion to weaken the influence weight of low signal-to-noise ratio modes, output the deep features of agricultural carbon sink potential. S5. Analyze the deep characteristics of agricultural carbon sequestration potential, extract net primary productivity and soil organic carbon density through semantic mapping, introduce environmental and soil parameters as physical correction factors for aggregation correction, and generate standardized agricultural physical carbon storage values. S6. Based on agricultural physical carbon storage values ​​as the underlying asset, and integrating multi-dimensional value influencing factors, the real-time unit value of agricultural carbon assets is output through non-linear dynamic weighted calculation. S7. Perform comprehensive statistical analysis on the real-time unit value and physical carbon storage of agricultural carbon assets, analyze the contribution of multi-source data based on the dynamic gating mechanism, output the all-weather assessment status and efficiency improvement decisions, and generate an agricultural carbon asset value assessment report.

2. The method for assessing the value of agricultural carbon assets based on multimodal fusion according to claim 1, characterized in that, S1 specifically includes: S11. Perform cloud detection on multi-source satellite imagery, identify the location of damaged pixels based on the cloud detection results, and perform repair on damaged pixels through temporal reconstruction to generate an optical pixel repair sequence. S12. Calculate the statistical distribution of pixel confidence in the optical pixel restoration sequence to determine the confidence threshold, and mark the regions in the optical pixel restoration sequence that are below the confidence threshold as low-confidence regions. S13. For low-confidence areas, perform reconstruction processing on radar images in multi-source satellite imagery to generate radar reconstructed images covering the low-confidence areas. S14. Perform spatiotemporal registration between the optical pixel restoration sequence and the radar reconstructed image. Calculate the pixel-level fusion weight matrix based on the confidence threshold of the optical pixel restoration sequence and the confidence level of the radar reconstructed image, and perform pixel-level weighted fusion to output effective reconstructed image data.

3. The method for assessing the value of agricultural carbon assets based on multimodal fusion according to claim 1, characterized in that, S2 specifically includes: S21. Divide the effectively reconstructed image data into regular grid units, perform texture statistics and gradient direction statistics for each regular grid unit, extract texture feature vectors and gradient feature vectors, and concatenate the texture feature vectors and gradient feature vectors to generate grid-level spatial features. S22. Based on the timestamp index, the image data is effectively reconstructed, a multi-frame image sequence is constructed, the temporal brightness change statistics of the corresponding pixels in the multi-frame image sequence are calculated, and the time-dependent features are extracted based on the temporal brightness change statistics to generate the temporal dynamic features of the image. S23. Parse the data type labels of the ground sensor data. If the data type label is a continuous value, normalize it. If the data type label is a discrete event, perform one-hot encoding mapping and generate sensor encoding feature vectors based on the processing results. S24. Perform an outer product operation on the grid-level spatial features and the temporal dynamic features of the image to generate a feature interaction tensor. Perform summation and normalization operations on the feature interaction tensor to generate an image fusion feature vector that integrates spatiotemporal correlation information. S25. Calculate the mutual information statistics between the image fusion feature vector and the sensor-coded feature vector, construct a spatial transformation matrix based on the mutual information statistics, and use the spatial transformation matrix to project the sensor-coded feature vector onto the feature space of the image fusion feature vector, and output a multimodal primary feature vector sequence.

4. The method for assessing the value of agricultural carbon assets based on multimodal fusion according to claim 1, characterized in that, The improved VGAE model includes a graph adjacency matrix construction layer, a variational graph convolutional coding layer, a molecular dynamics interaction simulation layer, and a topology error calibration layer. The graph adjacency matrix construction layer is used to dynamically calculate the feature distance and similarity between nodes in real time based on the multimodal primary feature vector sequence, adaptively construct and evolve the dynamic adjacency matrix, and perform spatial mapping of node features based on the topological association of the dynamic adjacency matrix to form a spatiotemporally aligned dynamic topological graph structure. The variational graph convolutional coding layer is used to perform continuous graph convolutional aggregation and update on the node features of the dynamic topological graph structure based on the dynamic adjacency matrix, so as to dynamically generate local high-order semantic features, and to solve the mean vector and log standard deviation vector of the latent space variables in real time through linear transformation mapping. The molecular dynamics interaction simulation layer is used to introduce a soft-nuclear potential-induced distribution calibration mechanism from computational biology. The specific execution process includes: By using reparameterization, a linear transformation is performed on the mean vector and the log-standard deviation vector to generate latent variable representations for the nodes. Calculate the feature distance between the latent variable representations of any two nodes, and dynamically map the attraction exponent coefficient and repulsion exponent coefficient based on the node features using a multilayer perceptron. These coefficients are then used as power-law adjustment parameters for the attraction and repulsion forces with respect to the feature distance. The attraction index coefficient is used to perform a reciprocal operation on the smoothed feature distance to generate attraction, and the repulsion index coefficient is used to perform an exponential operation on the smoothed feature distance to generate repulsion. The total resultant force on the node is calculated by combining the gravitational and repulsive forces, and the node position is iteratively updated along the negative gradient direction of the resultant force until the system reaches energy equilibrium. The physically optimized hidden variable representation is then output. The topology error calibration layer is used to perform probabilistic decoding on the physically optimized latent variable representation to dynamically reconstruct the topological association, and generate a topology consistency index based on the reconstruction difference evolution. Based on the dynamic evolution trajectory of the topology consistency index, adaptive node cleaning is performed, and finally, the consistent calibration feature sequence is output.

5. The method for assessing the value of agricultural carbon assets based on multimodal fusion according to claim 1, characterized in that, S4 specifically includes: S41. Based on the characteristic sequence after consistency calibration, calculate the multiple scattering flux and absorption attenuation of light in the vegetation canopy. By minimizing the residual error between the measured radiation value and the simulated radiation value, iteratively adjust the canopy parameters until the residual error converges, and output the spectral parameter inversion data. S42. Calculate the time series variance of the spectral parameter inversion data, and use the time series variance to quantify the data dispersion to generate a local signal-to-noise ratio evaluation index. S43. Perform kernel density estimation on the local signal-to-noise ratio evaluation index, determine the signal-to-noise classification boundary based on the peak characteristics of the kernel density estimation, and map the local signal-to-noise ratio evaluation index to adaptive weight coefficients according to the signal-to-noise classification boundary. S44. Obtain environmental factor monitoring data, calculate the correlation coefficient between spectral parameter inversion data and environmental factor monitoring data, use the correlation coefficient to perform weighted correction on spectral parameter inversion data, and generate environmentally corrected feature data. S45. Perform weighted aggregation operation on the environmentally corrected feature data using adaptive weight coefficients to output the deep feature of agricultural carbon sequestration potential.

6. The method for assessing the value of agricultural carbon assets based on multimodal fusion according to claim 1, characterized in that, S5 specifically includes: S51. Map the deep features of agricultural carbon sink potential to the carbon sink semantic feature space, analyze the distribution clustering of feature vectors to extract the net primary productivity feature components and soil organic carbon density feature components, and generate a set of basic carbon sink parameters. S52. Obtain temperature and humidity data and soil physicochemical parameters from environmental monitoring data, calculate the first correlation coefficient between the net primary productivity characteristic component and temperature and humidity data, calculate the second correlation coefficient between the soil organic carbon density characteristic component and soil physicochemical parameters, and construct an environmental constraint weight matrix based on the first and second correlation coefficients. S53. Perform dimensional standardization on the carbon sink basic parameter set, and use the environmental constraint weight matrix to perform weighted aggregation operation on the processed carbon sink basic parameter set to generate standardized agricultural physical carbon storage values.

7. The method for assessing the value of agricultural carbon assets based on multimodal fusion according to claim 1, characterized in that, S6 specifically includes: S61. Obtain agricultural carbon sink market transaction data and agricultural policy support indicators, perform numerical normalization and dimension alignment on agricultural carbon sink market transaction data and agricultural policy support indicators, and generate a multi-dimensional value influence factor matrix. S62. Perform principal component analysis on the multidimensional value impact factor matrix to extract the feature contribution rate, calculate the dynamic weight coefficient of each dimension factor based on the feature contribution rate, and use the dynamic weight coefficient to perform weighted summation on the multidimensional value impact factor matrix to generate a comprehensive value correction index. S63. Perform a nonlinear mapping operation between the agricultural physical carbon storage value and the comprehensive value correction index, perform numerical validity verification on the result of the nonlinear mapping operation, and output the real-time unit value of agricultural carbon assets.

8. The method for assessing the value of agricultural carbon assets based on multimodal fusion according to claim 1, characterized in that, Specifically, S7 includes: S71. Perform spatiotemporal alignment and aggregation operations on the real-time unit value of agricultural carbon assets and the standardized agricultural physical carbon storage values ​​to construct a multidimensional value time series dataset, calculate the statistical moment characteristics of the multidimensional value time series dataset, and generate a carbon asset statistical feature vector. S72. Perform correlation calculation and importance scoring on the statistical feature vector of carbon assets, calculate the dynamic attention weight of the feature dimension based on the importance score, construct the gated weighting coefficient using the dynamic attention weight, and perform weighted fusion operation on the statistical feature vector of carbon assets using the gated weighting coefficient to generate a deep state representation. S73. Calculate the historical data distribution characteristics of real-time monitoring data to generate a dynamic baseline. Perform deviation analysis based on the deep state characterization and the dynamic baseline. Generate differentiated efficiency-enhancing decision-making strategies based on the deviation analysis results. Encapsulate the deep state characterization, differentiated efficiency-enhancing decision-making strategies, and real-time monitoring data to generate an agricultural carbon asset value assessment report.

Images