Forest carbon sink potential prediction method based on artificial intelligence

Abstract

The invention discloses a forest carbon sink potential prediction method based on artificial intelligence. The method comprises the steps of data fusion, dynamic graph neural network representation, carbon flux prediction, uncertainty quantification and scene-driven simulation and forest carbon sink potential prediction. The invention relates to the technical field of ecological environment monitoring, in particular to a forest carbon sink potential prediction method based on artificial intelligence, and the method comprises the steps: constructing a height difference, species connectivity and multi-source feature similarity three-layer heterogeneous graph, and extracting space-time embedding in combination with a time decay edge weight and a node memory dynamic graph neural network; embedding and mapping the static features and the dynamic graph into monthly, seasonal and annual subspaces, and predicting the carbon flux in parallel through space and time self-attention, cross attention and a multi-scale regression head; estimating uncertainty by using Bayesian Monte Carlo sampling, and generating scenario prediction; and finally, calculating the carbon sink potential based on the average predicted value and the plot area accumulation, and comparing with a baseline to quantify the gain.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ecological environment monitoring, and in particular to a forest carbon sink potential prediction method based on artificial intelligence. BACKGROUND

[0002] With the global climate change and the proposal of carbon neutralization target, forest carbon sink as an important natural carbon absorption system has attracted widespread attention. Traditional forest carbon sink potential evaluation methods often have limited precision, lagging update and poor model universality in spatial coverage and time response. In the face of massive growth of multi-scale, multi-source and heterogeneous data, an intelligent prediction method capable of automatically and efficiently fusing multi-dimensional data such as remote sensing images, meteorological data and forest stand structure is urgently needed to improve the estimation accuracy and shorten the model iteration cycle.

[0003] With the development of artificial intelligence technology, it provides a new possibility for forest carbon sink potential prediction. Based on artificial intelligence algorithm, rapid modeling can be realized under the condition of small and medium sample. The spatial features of high-resolution remote sensing images and the forest stand topology structure are fully explored, and the dynamic change law of carbon sink is captured by combining the time series model to realize higher precision spatial and time interpolation and carbon flux inversion, which provides strong technical support for large-scale forest carbon sink potential evaluation. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides a forest carbon sink potential prediction method based on artificial intelligence, in order to solve the problems that the traditional forest carbon sink potential prediction method is difficult to depict the terrain difference, ecological corridor connectivity and forest stand characteristic similarity at the same time, and the static model cannot reflect the influence of growth evolution on the temporal and spatial variation of carbon flux, the present scheme constructs a three-layer heterogeneous graph based on height difference, species connectivity and multi-source fusion characteristic similarity, and introduces a dynamic graph neural network with time decay edge weight and node memory unit; in order to solve the problem that the traditional forest carbon flux prediction method often relies on a single static feature when capturing dynamic changes at different time scales and spatial correlation across plots, and it is difficult to consider seasonal and disturbance event patterns, the present scheme projects the static fusion features and dynamic graph embedding into three subspaces of month, season and year respectively, combines spatial self-attention dynamic identification of cross-plot influence path, time self-attention mining of duration mode, and introduces cross-branch cross-attention fusion, and finally outputs carbon flux prediction in parallel with multi-scale regression head; in order to solve the problem that the traditional forest carbon sink prediction method often cannot quantify the uncertainty of the model itself and lacks diversified future climate scenario analysis, the present scheme estimates the prediction residual distribution through Bayesian Monte Carlo sampling, and combines the Transformer encoding and conditional adjustment network of the climate scenario to generate scenario-based carbon flux prediction and uncertainty measurement; in order to solve the problem that the traditional carbon sink potential evaluation is mainly point estimation or single scenario, and lacks comprehensive comparative analysis of cumulative potential and gain rate under different scenarios, the present scheme calculates the cumulative potential of plots and the whole region under each scenario based on the multi-scale average prediction value and the area of the plot, and aggregates and compares with the baseline without scenario, and finally quantifies the carbon sink potential gain.

[0005] The technical scheme adopted by the present application is as follows: a forest carbon sink potential prediction method based on artificial intelligence, the method comprising the following steps:

[0006] Step S1: data fusion, collecting multi-dimensional data of the predicted forest area, performing feature coding, then projecting the obtained hidden vectors to the same latent space through linear projection, and finally splicing the projected features of images, point clouds and measured samples by channel to generate fusion features;

[0007] Step S2: dynamic graph neural network representation, constructing a three-layer heterogeneous space graph; then applying a dynamic graph neural network with time decay edge weight and node memory unit on the multi-layer graph, capturing the spatial relationship and growth evolution of the forest plot over time through message aggregation and state update, and finally outputting node high-dimensional embedding;

[0008] Step S3: Carbon flux prediction, projecting the static fused features and dynamic graph embedding into monthly, seasonal, and annual subspaces respectively, and performing average pooling on the time series data within the subspaces to obtain multi-scale aggregated representations; then performing spatial self-attention, temporal self-attention, and cross-branch cross-attention fusion on each scale, and using a multi-scale linear regression head to predict the carbon flux at each scale in parallel after fusing the spatial and temporal information;

[0009] Step S4: Uncertainty quantification and scenario-driven simulation, first performing Bayesian Monte Carlo sampling on the multi-scale carbon flux prediction results to simulate the residual distribution and estimate the prediction uncertainty; then constructing a sequence of future climate scenarios, mapping them into latent vectors using a Transformer encoder, and adjusting each sampling result according to the scenario using a conditional adjustment network, finally outputting the carbon flux;

[0010] Step S5: Forest carbon sink potential prediction, based on the average carbon flux prediction value, combining the area of each plot and the corresponding time length at each scale, and cumulatively calculating the carbon sink potential of each plot; then aggregating the carbon sink potential of all plots and comparing it with the baseline carbon sink potential without scenarios to obtain the regional carbon sink potential and relative gain rate.

[0011] Further, in step S1, the data fusion specifically includes the following steps:

[0012] Step S11: Data collection, collecting high-resolution remote sensing images, laser radar point cloud data, plot measured samples, plot area, historical carbon flux observation values, and future temperature and precipitation prediction data for the predicted forest area; wherein each remote sensing image contains red, green, blue, and near-infrared bands; each point cloud contains spatial coordinates and reflectivity; and the plot measured samples contain the average tree height, average breast height area, and the most numerous species of the plot;

[0013] Step S12: Feature encoding and projection, designing independent encoders for images, point clouds, and measured samples respectively to map different modalities to the same latent space;

[0014] Step S13: Feature fusion output.

[0015] Further, in step S2, the dynamic graph neural network representation specifically includes the following steps:

[0016] Step S21: Multi-level spatial graph construction, dividing the plot into nodes and constructing multi-level edges according to the terrain, ecological corridor, and forest similarity to form a three-layer heterogeneous graph;

[0017] Step S22: Dynamic graph neural network spatio-temporal representation, introducing time decay edge weights and node memory units using a dynamic graph neural network;

[0018] Step S23: high-dimensional embedding extraction of spatio-temporal evolution, after the iteration is completed, a high-dimensional embedding with spatial structure and growth evolution information is obtained.

[0019] Further, in step S3, the carbon flux prediction, the node spatio-temporal representation set of step S2 and the fusion features of step S1 are obtained, and the carbon flux prediction is performed, specifically including the following steps:

[0020] Step S31: spatio-temporal feature construction, the fusion features of each node are fused with the historical spatio-temporal representation sequence, and are projected into three subspaces of month scale, season scale and year scale respectively;

[0021] Step S32: introducing spatial self-attention, for each scale, self-attention is performed between the node set to dynamically capture the cross-plot carbon flux interaction path;

[0022] Step S33: introducing time self-attention, performing sequence self-attention on the duration representation of each node at the scale to mine seasonal and disturbance event patterns;

[0023] Step S34: cross-branch cross-attention fusion, fusing the spatial branch and the time branch through cross-attention;

[0024] Step S35: multi-scale carbon flux prediction, configuring an independent linear regression head for each scale to output carbon flux prediction.

[0025] Further, in step S4, the uncertainty quantification and scenario-driven simulation, specifically including the following steps:

[0026] Step S41: Bayesian Monte Carlo sampling, performing Monte Carlo sampling on the multi-scale carbon flux prediction set output by step S3 to estimate the uncertainty of the prediction;

[0027] Step S42: scenario set definition and coding, constructing a climate scenario, each scenario containing a climate prediction of a future time step, and mapping the scenario to a latent vector through a scenario encoder;

[0028] Step S43: uncertainty simulation and aggregation, for each Monte Carlo sampling result and each scenario, using a conditional adjustment network to generate a scenario prediction.

[0029] Further, in step S5, the forest carbon sink potential prediction, specifically including the following steps:

[0030] Step S51: carbon sink potential index calculation, based on the aggregated carbon flux average prediction of step S4 and the plot area data; calculating the cumulative carbon sink potential;

[0031] Step S52: Carbon sequestration potential aggregation, aggregating the total carbon sequestration potential of all parcels, and predicting the baseline carbon sequestration potential without scenario adjustment.

[0032] The beneficial effects achieved by the present invention using the above solution are as follows:

[0033] (1) In view of the problem that traditional forest carbon sink potential prediction methods are difficult to simultaneously characterize topographic differences, ecological corridor connectivity and forest stand characteristic similarity, and that static models cannot reflect the impact of growth evolution on the spatiotemporal changes of carbon flux, this scheme constructs a three-layer heterogeneous graph based on height difference, species connectivity and multi-source fusion characteristic similarity, and introduces a dynamic graph neural network with time decay edge weights and node memory units, thereby realizing a high-dimensional embedded representation of forest plots in the spatiotemporal evolution process and effectively capturing the spatiotemporal change law of carbon flux.

[0034] (2) In view of the problem that traditional forest carbon flux prediction methods often rely on a single static feature when capturing dynamic changes at different time scales and cross-plot spatial correlations, and are difficult to take into account seasonality and disturbance event patterns, this scheme projects static fusion features and dynamic graphs into three subspaces of month, season and year respectively. It combines spatial self-attention to dynamically identify cross-plot influence paths, temporal self-attention to mine historical patterns, and introduces cross-branch cross-attention fusion. Finally, it outputs carbon flux predictions in parallel with multi-scale regression heads, realizing refined modeling and efficient prediction of the spatiotemporal evolution of forest carbon flux.

[0035] (3) In view of the problem that traditional forest carbon sink prediction methods often have difficulty in quantifying the uncertainty of the model itself and lack diverse future climate scenario analysis, this scheme estimates the prediction residual distribution by Bayesian Monte Carlo sampling and combines the Transformer coding and condition adjustment network of climate scenario to generate scenario-based carbon flux prediction and its uncertainty measurement, thus realizing the confidence interval assessment and scenario-driven simulation of future carbon sink flux.

[0036] (4) In view of the problem that traditional carbon sink potential assessment is mainly based on point estimation or single scenario and lacks comprehensive comparative analysis of cumulative potential and gain rate under different scenarios, this scheme obtains the carbon sink potential of each plot and the whole area under each scenario by accumulating the average predicted value of multiple scales and the plot area, and compares it with the baseline without scenario, and finally quantifies the carbon sink potential gain, thereby improving the pertinence and operability of decision support. Attached Figure Description

[0037] Figure 1 A schematic diagram illustrating an artificial intelligence-based method for predicting forest carbon sequestration potential provided by this invention;

[0038] Figure 2 This is a schematic diagram of step S1;

[0039] Figure 3 This is a schematic diagram of step S2;

[0040] Figure 4 This is a schematic diagram of step S3;

[0041] Figure 5 This is a schematic diagram of step S4;

[0042] Figure 6 This is a schematic diagram of step S5.

[0043] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0045] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0046] Example 1, see Figure 1 This invention provides an artificial intelligence-based method for predicting forest carbon sink potential, which includes the following steps:

[0047] Step S1: Data fusion, collecting multi-source remote sensing images, lidar point clouds, measured samples of land parcels, land parcel areas, historical carbon flux observations, and future temperature and precipitation forecasts; using convolutional networks, ... The multilayer perceptron encodes the features of three data sources: multi-source remote sensing images, lidar point clouds, and measured samples of land parcels. The resulting hidden vectors are then mapped to the same latent space through their respective linear projections. Finally, the projection features of the images, point clouds, and measured samples are stitched together by channel to generate fused features.

[0048] Step S2: Dynamic graph neural network representation. A three-layer heterogeneous spatial graph is constructed based on the average tree height difference between plots, the number of common species, and the similarity of fusion features, forming a network structure that simultaneously reflects topography, ecological corridors, and forest stand similarity. Subsequently, a dynamic graph neural network with time-decaying edge weights and node memory units is applied to the multi-layer graph. Through message aggregation and state updates, the spatial connections and growth evolution of forest plots over time are captured, and finally, high-dimensional node embeddings are output.

[0049] Step S3: Carbon flux prediction. The static fusion features and dynamic graph embeddings are projected onto the monthly, quarterly, and annual subspaces, respectively. The time series data in the subspaces are averaged and pooled to obtain a multi-scale aggregated representation. Then, spatial self-attention, temporal self-attention, and cross-branch cross-attention fusion are performed on each scale. After fusing spatial and temporal information, a multi-scale linear regression head is used to predict carbon flux at each scale in parallel.

[0050] Step S4: Uncertainty Quantification and Scenario-Driven Simulation. First, Bayesian Monte Carlo sampling is performed on the multi-scale carbon flux prediction results to simulate the residual distribution and estimate the prediction uncertainty. Then, a future climate scenario sequence is constructed, mapped into a latent vector through a Transformer encoder, and the scenario adjustment network is combined to adjust the sampling results for each time, and finally output the carbon flux.

[0051] Step S5: Forest carbon sink potential prediction. Based on the average carbon flux prediction value, combined with the plot area and the corresponding time length at each scale, the carbon sink potential of each plot is accumulated and calculated. Then, the carbon sink potential of all plots is aggregated and compared with the baseline carbon sink potential without scenario to obtain the carbon sink potential of the whole area and the relative gain rate.

[0052] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the data fusion specifically includes the following steps:

[0053] Step S11: Data acquisition, acquiring high-resolution remote sensing images of the predicted forest area. LiDAR point cloud data Actual measured samples of land parcels Land area Historical carbon flux observations and future temperature and precipitation forecasts; among them, each remote sensing image Includes red, green, blue, and near-infrared bands; each point cloud Includes spatial coordinates and reflection intensity; measured samples of the land parcel. , This represents the average tree height of the i-th plot. This represents the average cross-sectional area at breast height of the i-th plot. This represents the species with the most occurrences in the i-th plot;

[0054] Step S12: Feature Encoding and Projection. Independent encoders are designed for the image, point cloud, and measured samples respectively, mapping different modalities to the same latent space, as shown below:

[0055] ;

[0056] in, , and Representing parameter sets respectively , and Convolution, and sensor encoder; , and These represent the encoded image, point cloud, and measured sample hidden vectors, respectively. , Indicates image projection parameters. , Indicates the point cloud projection parameters. and Indicates the measured sample projection parameters; , and This represents the unified feature vector after projection;

[0057] Step S13: Feature fusion output, represented as follows:

[0058] ;

[0059] in, Indicates vector concatenation symbol. This indicates the splicing and fusion characteristics.

[0060] Example 3, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S2, the dynamic graph neural network representation specifically includes the following steps:

[0061] Step S21: Multi-level spatial map construction. The plots are divided into nodes, and multiple edges are constructed based on the similarity of topography, ecological corridors, and forest stands to form a three-layer heterogeneous map, as shown below:

[0062] ;

[0063] in, Let V represent a land parcel node, and let V represent the set of all land parcel nodes. This represents the average tree height difference between the i-th and j-th plot nodes. Indicates the height threshold; This represents the number of species that exist in both the i-th and j-th plot nodes. Indicates the connectivity threshold. and This represents the splicing and fusion features output in step S1. express and Similarity between them This represents a three-layer heterogeneous spatial map, where h represents the layer index.

[0064] Step S22: Spatiotemporal representation using dynamic graph neural network. This involves introducing time-decaying edge weights and node memory units using a dynamic graph neural network, as shown below:

[0065] ;

[0066] in, This represents the time decay weight, where t represents the current time. This represents an exponential function with the natural constant as its base. side Last updated time, This represents the decay rate of the edge at layer h. This represents the message mapping matrix of the h-th layer. This represents the set of neighboring nodes of node i at level h. This represents the message aggregation in the k-th round. and Represents the state transition matrix. This represents the candidate state of the node in round k. This represents the hidden state of node j in the (k-1)th round. This represents the hidden state of node i in the (k-1)th round. This represents the hidden state of node i in round k. Represents a linear rectified function;

[0067] Step S23: High-dimensional embedding extraction of spatiotemporal evolution. After completing K rounds of iterations, the final hidden state of the node is a high-dimensional embedding that combines spatial structure and growth evolution information, represented as follows:

[0068] ;

[0069] in, The spatiotemporal representation vector of node i. This represents the set of all embedded nodes.

[0070] By performing the above operations, this scheme addresses the problem that traditional methods for predicting forest carbon sink potential struggle to simultaneously characterize topographic differences, ecological corridor connectivity, and forest stand characteristic similarities, and that static models cannot reflect the impact of growth evolution on the spatiotemporal variation of carbon flux. This scheme constructs a three-layer heterogeneous graph based on height difference, species connectivity, and multi-source fusion feature similarity, and introduces a dynamic graph neural network with time-decaying edge weights and node memory units. This enables a high-dimensional embedded representation of forest plots during the spatiotemporal evolution process, effectively capturing the spatiotemporal variation patterns of carbon flux.

[0071] Example 4, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In step S3, the carbon flux prediction obtains the spatiotemporal characterization set of nodes from step S2. and the fusion features of step S1 Carbon flux prediction involves the following steps:

[0072] Step S31: Spatiotemporal feature construction, fusing the features of each node. With historical spatiotemporal representation sequence The data is then merged and projected onto three subspaces: monthly, quarterly, and annual, as shown below:

[0073] ;

[0074] in, This represents the dynamic graph embedding of node i at time t. , , and This represents a trainable projection matrix. and The fused features, where s represents the scale index corresponding to the monthly, quarterly, and yearly scales. Represents the set of time steps at scale s. and The aggregate representation of scale s;

[0075] Step S32: Introduce spatial self-attention. For each scale s, perform self-attention among the node sets to dynamically capture the cross-plot carbon flux interaction paths, as shown below:

[0076] ;

[0077] in, , and These represent the query, key, and value projection at the s-th scale, respectively. Represents the space scaling constant. This represents the maximum edge weight between node i and node j in the three-layer heterogeneous graph output by step S2. Let represent the attention weight of node i to node j at the s-th scale. This represents the spatial self-attention output of the i-th node at the s-th scale;

[0078] Step S33: Introduce temporal self-attention and represent the epoch of each node i at scale s. Sequence self-attention is performed to uncover seasonality and interference event patterns, as shown below:

[0079] ;

[0080] in, , and These represent the query, key, and value matrices for temporal attention, respectively. Represents the time scaling constant. Indicates the time decay rate. This represents the attention weights from time t to u at node i and scale s. This represents the temporal self-attention output of the i-th node at the s-th scale;

[0081] Step S34: Cross-branch attention fusion, integrating spatial branches With time branch The cross-attention fusion is represented as follows:

[0082] ;

[0083] in, , and These represent the query, key, and value projection matrices for cross-attention, respectively. Indicates the fusion hyperparameters, Represents the normalized exponential function, Indicates the fusion weight. The final fused representation at scale s;

[0084] Step S35: Multi-scale carbon flux prediction. An independent linear regression head is configured for each scale s to output the carbon flux prediction, as shown below:

[0085] ;

[0086] in, and Let represent the learnable prediction head weights and bias parameters at the s-th scale, respectively. This represents the carbon flux prediction at node i on scale s. Represents the actual observed value. This represents the predictive coefficient for the learnable s-th scale.

[0087] By performing the above operations, this scheme addresses the problem that traditional forest carbon flux prediction methods often rely on single static features and struggle to take into account seasonality and disturbance event patterns when capturing dynamic changes at different time scales and spatial correlations across plots. This scheme projects static fusion features and dynamic graphs into three subspaces: month, season, and year. It combines spatial self-attention to dynamically identify cross-plot influence paths, temporal self-attention to mine historical patterns, and introduces cross-branch cross-attention fusion. Finally, it outputs carbon flux predictions in parallel using multi-scale regression heads, achieving refined modeling and efficient prediction of the spatiotemporal evolution of forest carbon flux.

[0088] Example 5, see Figure 1 and Figure 5 This embodiment is based on the above embodiment. In step S4, the uncertainty quantification and scenario-driven simulation specifically include the following steps:

[0089] Step S41: Bayesian Monte Carlo sampling, processing the multi-scale carbon flux prediction set output from step S3. conduct Sub-Monte Carlo sampling estimates the uncertainty of the prediction, as follows:

[0090] ;

[0091] in, This represents the predicted carbon flux at the s-th scale for the n-th plot obtained from the g-th sampling. This represents the predicted carbon flux at the nth plot and the sth scale. This represents Gaussian noise in the residual estimation. This represents the standard deviation of the residual at the nth plot and the sth scale calculated in step S3 on the training set.

[0092] Step S42: Scenario set definition and coding, constructing climate scenario B. Each of the scenarios Includes the future Climate predictions at each time step, and through a scenario encoder The scenario is mapped to a latent vector, as follows:

[0093] ;

[0094] Where b represents the scenario index, there are a total of B; Indicates the scenario time step index. This represents the b-th scenario, the... The climate prediction vector for each scenario time step specifically includes two dimensions: temperature and precipitation. Indicates a set of parameters Transformer context encoder, This represents the latent representation vector of the b-th scenario;

[0095] Step S43: Uncertainty simulation and aggregation, processing the results of each Monte Carlo sampling. and each scenario Adjust network conditions Contextualized predictions are generated as follows:

[0096] ;

[0097] in, Indicates a set of parameters Multilayer perceptron conditional adjustment network. This represents the predicted carbon flux for the nth plot and the sth scale in scenario b, based on the g-th sampling. This represents the average predicted carbon flux for the nth plot and the sth scale under scenario b. This represents the uncertainty measure for the nth plot and the sth scale under scenario b.

[0098] By performing the above operations, this scheme addresses the problem that traditional forest carbon sink prediction methods often struggle to quantify the uncertainty of the models themselves and lack diverse future climate scenario analysis. It uses Bayesian Monte Carlo sampling to estimate the prediction residual distribution and combines it with Transformer coding and conditional adjustment networks for climate scenarios to generate scenario-based carbon flux predictions and their uncertainty measures. This enables the assessment of confidence intervals for future carbon sink fluxes and scenario-driven simulation.

[0099] Example 6, see Figure 1 and Figure 6 This embodiment is based on the above embodiment. In step S5, the prediction of forest carbon sink potential specifically includes the following steps:

[0100] Step S51: Calculation of carbon sink potential index, based on the average predicted carbon flux after aggregation in Step S4. Land area data The cumulative carbon sequestration potential of the nth plot under scenario b is calculated as follows:

[0101] ;

[0102] in, This represents the area of ​​the nth plot of land. This represents the time length corresponding to scale s, where s=1. When s=2 When s=3 ; This represents the carbon sequestration potential of the nth plot under scenario b;

[0103] Step S52: Carbon sequestration potential aggregation. Aggregate the total carbon sequestration potential of all sites, and predict the baseline carbon sequestration potential without scenario adjustment, as shown below:

[0104] ;

[0105] in, This represents the region's carbon sequestration potential under scenario b. This indicates the baseline carbon sequestration potential without scenario adjustment. This represents the relative potential gain rate of scenario b.

[0106] By performing the above operations, this solution addresses the problem that traditional carbon sequestration potential assessments often rely on point estimations or single scenarios and lack comprehensive comparative analysis of cumulative potential and gain rates under different scenarios. This solution calculates the carbon sequestration potential of each plot and the entire region under each scenario based on the cumulative calculation of multi-scale average predicted values ​​and plot area. It then aggregates and compares these results with a baseline without scenarios, ultimately quantifying the carbon sequestration potential gain and improving the relevance and operability of decision support.

[0107] 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.

[0108] 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.

[0109] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for predicting forest carbon sequestration potential based on artificial intelligence, characterized in that, The method includes the following steps: Step S1: Data fusion. Collect multi-dimensional data of the predicted forest area, perform feature encoding, and then map the obtained hidden vectors to the same latent space through their respective linear projections. Finally, stitch together the projection features of the image, point cloud and measured sample by channel to generate fused features. Step S2: Dynamic graph neural network representation to construct a three-layer heterogeneous spatial graph; then apply a dynamic graph neural network with time decay edge weights and node memory units to the multi-layer graph, capture the spatial relationship and growth evolution of forest plots over time through message aggregation and state update, and finally output high-dimensional node embeddings. Step S3: Carbon flux prediction. The static fusion features and dynamic graph embeddings are projected onto the monthly, quarterly, and annual subspaces, respectively. The time series data in the subspaces are averaged and pooled to obtain a multi-scale aggregated representation. Then, spatial self-attention, temporal self-attention, and cross-branch cross-attention fusion are performed on each scale. After fusing spatial and temporal information, a multi-scale linear regression head is used to predict carbon flux at each scale in parallel. Step S4: Uncertainty Quantification and Scenario-Driven Simulation. First, Bayesian Monte Carlo sampling is performed on the multi-scale carbon flux prediction results to simulate the residual distribution and estimate the prediction uncertainty. Then, a future climate scenario sequence is constructed, mapped into a latent vector through a Transformer encoder, and the scenario adjustment network is combined to adjust the sampling results for each time, and finally output the carbon flux. Step S5: Forest carbon sink potential prediction. Based on the average carbon flux prediction value, combined with the plot area and the corresponding time length at each scale, the carbon sink potential of each plot is accumulated and calculated. Then, the carbon sink potential of all plots is aggregated and compared with the baseline carbon sink potential without scenario to obtain the carbon sink potential of the whole area and the relative gain rate.

2. The method for predicting forest carbon sink potential based on artificial intelligence according to claim 1, characterized in that: In step S2, the dynamic graph neural network representation specifically includes the following steps: Step S21: Multi-level spatial map construction, dividing the plots into nodes, constructing multi-layer edges based on topography, ecological corridors and forest stand similarity, forming a three-layer heterogeneous map; Step S22: Spatiotemporal representation of dynamic graph neural network, using dynamic graph neural network to introduce time decay edge weights and node memory units; Step S23: Spatiotemporal evolution high-dimensional embedding extraction. After completing the iteration, a high-dimensional embedding that combines spatial structure and growth evolution information is obtained.

3. The method for predicting forest carbon sink potential based on artificial intelligence according to claim 1, characterized in that: In step S3, the carbon flux prediction involves obtaining the node spatiotemporal representation set from step S2 and the fusion features from step S1, and then performing carbon flux prediction. Specifically, this includes the following steps: Step S31: Spatiotemporal feature construction, which involves fusing the fusion features of each node with the historical spatiotemporal representation sequence and projecting them onto three subspaces: monthly, quarterly, and annual scales, respectively. Step S32: Introduce spatial self-attention. For each scale, perform self-attention among the node sets to dynamically capture the paths of cross-plot carbon flux interactions. Step S33: Introduce temporal self-attention, perform sequence self-attention on the diachronic representation of each node at the scale, and mine seasonal and interference event patterns; Step S34: Cross-branch attention fusion, fusing spatial branches and temporal branches through cross-attention; Step S35: Multi-scale carbon flux prediction. Configure an independent linear regression head for each scale and output the carbon flux prediction.

4. The method for predicting forest carbon sink potential based on artificial intelligence according to claim 1, characterized in that: In step S4, the uncertainty quantification and scenario-driven simulation specifically include the following steps: Step S41: Bayesian Monte Carlo sampling, performing Monte Carlo sampling on the multi-scale carbon flux prediction set output in step S3 to estimate the uncertainty of the prediction; Step S42: Scenario set definition and encoding, constructing climate scenarios, each scenario containing climate predictions for future time steps, and mapping scenarios to latent vectors through a scenario encoder; Step S43: Uncertainty simulation and aggregation: For each Monte Carlo sampling result and each scenario, use a conditional adjustment network to generate contextualized predictions.

5. The method for predicting forest carbon sink potential based on artificial intelligence according to claim 1, characterized in that: In step S5, the prediction of forest carbon sink potential specifically includes the following steps: Step S51: Calculate carbon sink potential index, based on the average carbon flux forecast and land area data aggregated in Step S4; calculate cumulative carbon sink potential; Step S52: Carbon sequestration potential aggregation, aggregating the total carbon sequestration potential of all parcels, and predicting the baseline carbon sequestration potential without scenario adjustment.

6. The method for predicting forest carbon sink potential based on artificial intelligence according to claim 1, characterized in that: In step S1, the data fusion specifically includes the following steps: Step S11: Data Acquisition. Collect high-resolution remote sensing images, lidar point cloud data, measured plot samples, plot area, historical carbon flux observations, and future temperature and precipitation forecasts for the predicted forest area. Each remote sensing image includes red, green, blue, and near-infrared bands; each point cloud includes spatial coordinates and reflectance intensity; the measured plot samples include the plot's average tree height, average basal area at breast height, and the most abundant species. Step S12: Feature encoding and projection. Independent encoders are designed for images, point clouds and measured samples respectively, to map different modalities to the same latent space. Step S13: Feature fusion output.

Images