Method for Dynamically Predicting Carbon Sink in Ecological Restoration Area Based on Temporal Remote Sensing

Abstract

The present invention relates to the technical field of carbon sink prediction, and specifically to a method for dynamic prediction of carbon sinks in ecological restoration areas based on time-series remote sensing. The method comprises the following steps: obtaining a multi-source time-series dataset of a target area and extracting the spatial characteristics and time-series fluctuation patterns of the multi-source time-series dataset; dividing the ecological restoration area within the target area into multiple dynamic aggregation units based on the spatial characteristics and time-series fluctuation patterns, and merging dynamic aggregation units with similar characteristics; constructing a differentiated carbon sink accounting model based on the ecosystem type of each merged dynamic aggregation unit, extracting multi-scale time-series characteristics, and screening key influencing factors; constructing a carbon sink dynamic prediction model based on the multi-scale time-series characteristics and the screened key influencing factors, and outputting the carbon sink dynamic prediction results; and solving a restoration decision plan based on the carbon sink dynamic prediction results, with the goal of maximizing carbon sink increments. The method aims to achieve dynamic prediction of carbon sinks in large-scale ecological restoration areas and improve carbon sink prediction accuracy and computational efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon sink prediction, and in particular to a method for dynamic prediction of carbon sinks in ecological restoration areas based on time series remote sensing. Background Art

[0002] Ecological restoration is a key technical approach to consolidating and enhancing the carbon sequestration function of ecosystems, increasing carbon storage in vegetation and soil. Carbon sequestration in ecologically restored areas is the ability to absorb and store greenhouse gases such as carbon dioxide. Dynamic prediction of carbon sequestration is a crucial basis for evaluating the effectiveness of restoration projects and optimizing carbon sequestration management strategies.

[0003] Currently, carbon sink prediction technology primarily relies on empirical models or statistical regression methods based on ground-based observations, which can achieve a certain degree of accuracy in carbon storage estimates for small-scale, single-ecosystem scenarios. This approach has certain limitations when addressing large-scale, heterogeneous ecological restoration areas. Due to the differences in ecosystem types and the spatial heterogeneity of different restoration measures, traditional methods are typically only able to focus on single-scenario modeling and lack the ability to characterize carbon sink processes in multiple ecosystem types, resulting in reduced accuracy in cross-regional carbon sink predictions. Furthermore, when large-scale data is directly input into the model, redundant computation increases, and traditional time series models struggle to efficiently handle multi-scale features, resulting in inefficient carbon sink predictions. Summary of the Invention

[0004] In order to achieve dynamic prediction of carbon sinks in large-scale ecological restoration areas and improve the accuracy and computational efficiency of carbon sink prediction, the present invention provides a method for dynamic prediction of carbon sinks in ecological restoration areas based on time series remote sensing. The technical solutions adopted are as follows:

[0005] The technical solution of the present invention provides a method for dynamic prediction of carbon sinks in ecological restoration areas based on time series remote sensing, the method comprising:

[0006] Obtain a multi-source time series dataset of the target area and extract the spatial features and time series fluctuation patterns of the multi-source time series dataset;

[0007] Based on the spatial characteristics and temporal fluctuation patterns, the ecological restoration area within the target area is divided into multiple dynamic aggregation units, and dynamic aggregation units with similar characteristics are merged;

[0008] Based on the ecosystem types of each dynamic aggregation unit after merging, a differentiated carbon sink accounting model is constructed to extract multi-scale temporal characteristics and screen key influencing factors;

[0009] Based on the multi-scale time series characteristics and the screened key influencing factors, a carbon sink dynamic prediction model is constructed and a carbon sink dynamic prediction result is output;

[0010] With the goal of maximizing carbon sink increment, a restoration decision plan is solved based on the carbon sink dynamic prediction results.

[0011] Furthermore, a multi-source time series dataset of the target area is obtained, and the spatial features and time series fluctuation patterns of the multi-source time series dataset are extracted, including:

[0012] Obtain a multi-source time series dataset of the target area, wherein the multi-source time series dataset includes remote sensing data, ground data, meteorological data, and the intensity of restoration measures;

[0013] Performing spatiotemporal alignment and standardization processing on the multi-source time series dataset;

[0014] Extract spatial features through weighted fusion and principal component analysis;

[0015] The time series fluctuation pattern is extracted through frequency domain transformation and time domain regression.

[0016] Furthermore, extracting the time series fluctuation pattern specifically includes:

[0017] The standardized multi-source time series data are transformed into the frequency domain, the annual and seasonal main components are identified by fast Fourier transform, and the multi-scale time series features are decomposed by discrete wavelet transform.

[0018] Perform time domain analysis on the same time series data and fit the long-term trend term through linear regression;

[0019] The principal components of the annual cycle and seasonal cycle are superimposed on the trend term, and the composite time series fluctuation characteristics are generated through weighted summation.

[0020] Furthermore, based on the spatial characteristics and temporal fluctuation patterns, the ecological restoration area within the target area is divided into multiple dynamic aggregation units, and dynamic aggregation units with similar characteristics are merged, including:

[0021] Constructing an ecological correlation matrix based on the similarity of spatial characteristics;

[0022] The spectral clustering algorithm is used to divide the initial dynamic aggregation units;

[0023] Adjacent cells with consistent feature distributions are merged through spatial proximity constraints and temporal synchronization verification.

[0024] Furthermore, based on the ecosystem types of the merged dynamic aggregation units, a differentiated carbon sink accounting model was constructed to extract multi-scale temporal characteristics and screen key influencing factors, including:

[0025] Differentiated carbon sequestration models are selected based on ecosystem types: the light energy utilization model is used in forest restoration areas, the flux coupling model is used in wetland degradation areas, and the vegetation and soil carbon coupling model is used in desertification restoration areas;

[0026] Extract annual, seasonal and monthly time series features through wavelet multi-scale decomposition;

[0027] Key influencing factors were screened based on gradient sensitivity analysis, and redundant factors with contributions below the threshold were eliminated.

[0028] Furthermore, based on the multi-scale time series characteristics and the screened key influencing factors, a carbon sink dynamic prediction model is constructed and the carbon sink dynamic prediction results are output, including:

[0029] Construct a spatiotemporal graph convolutional network to integrate the carbon flux correlation and temporal evolution of spatial neighborhood units;

[0030] Quantify the hysteresis effect of ecological restoration measures and correct the basic prediction value through exponential decay model;

[0031] Output carbon sink dynamic prediction results including spatial synergy gains.

[0032] Furthermore, the construction of the spatiotemporal graph convolutional network includes:

[0033] Dynamically balance the output weights of spatial graph convolution and time series models through the gating matrix;

[0034] A spatial graph structure is constructed based on the carbon flux correlation of spatial neighborhood units, and the edge weights are determined by the covariance of key influencing factors;

[0035] The Sigmoid function is used to generate the gating value to control the fusion ratio of spatial dependency and temporal dependency.

[0036] Furthermore, quantifying the hysteresis effects of ecological restoration measures includes:

[0037] Define the effective delay time of the repair measures and build a time offset sequence;

[0038] The partial derivatives of historical measures are calculated within a preset time window, and the weights are generated by normalizing the ratio of the sum of the absolute values of the partial derivatives.

[0039] Furthermore, the spatial synergy gain includes:

[0040] According to the adjacency relationship of dynamic aggregation units, the contribution weight of similar measures in adjacent units is increased;

[0041] The synergy coefficient is dynamically adjusted through iterative optimization to maximize the spatial synergy gain.

[0042] Furthermore, with the goal of maximizing carbon sink increment, a restoration decision plan is solved based on the carbon sink dynamic prediction results, including:

[0043] Construct a multi-objective optimization model with constraints including budget limits, mutual exclusivity of measures, and biodiversity protection;

[0044] The global optimization problem is grouped by ecological type through an adaptive decomposition algorithm, and restoration decision plans are solved in parallel.

[0045] The present invention has the following beneficial effects:

[0046] The method for dynamic prediction of carbon sinks in ecological restoration areas based on time-series remote sensing provided by the present invention divides large-scale and complex ecological restoration areas into relatively homogeneous small units through the division of dynamic aggregation units, thereby reducing the computational complexity and improving the computational efficiency of carbon sink prediction in large-scale areas. By utilizing a differentiated carbon sink accounting model, the carbon sink mechanism is quantified based on the characteristics of different ecosystem types, overcoming the problem that traditional models cannot accurately reflect the differences in ecosystems, thereby improving the accuracy of carbon sink prediction. Through the fusion of multi-scale time series features and the screening of key influencing factors, the carbon sink dynamic prediction model can capture the spatiotemporal characteristics and key driving factors of the dynamic changes in carbon sinks, and further optimize the carbon sink prediction effect. Ultimately, solving the restoration decision-making plan with the goal of maximizing the carbon sink increment can provide a scientific and reasonable decision-making basis for ecological restoration projects, and realize the integration from carbon sink prediction to restoration decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 A flow chart of a method for dynamic prediction of carbon sinks in ecological restoration areas based on time-series remote sensing provided by one embodiment of the present invention;

[0049] Figure 2 A schematic diagram of a differentiated carbon sink accounting model provided by one embodiment of the present invention;

[0050] Figure 3 A schematic diagram of a carbon sink dynamic prediction model provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation method, structure, features and effects of a dynamic prediction method for carbon sinks in ecological restoration areas based on time-series remote sensing proposed by the present invention. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0052] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0053] The specific scheme of the method for dynamic prediction of carbon sinks in ecological restoration areas based on time series remote sensing provided by the present invention is described in detail below with reference to the accompanying drawings.

[0054] See also Figure 1 , which shows a method flow chart of a method for dynamic prediction of carbon sinks in ecological restoration areas based on time series remote sensing provided by one embodiment of the present invention, the method comprising:

[0055] Step S100: Acquire a multi-source time series dataset for the target area and extract the spatial characteristics and time series fluctuation patterns of the multi-source time series dataset. Specifically, this embodiment takes the integrated protection and restoration project for plateau wetland grassland area A as an example, selects typical ecosystem restoration types such as desertification restoration, wetland degradation, and forest restoration areas as research objects, and conducts carbon sink prediction and decision-making research in plateau wetland grassland area A after ecological restoration. The goal is to achieve dynamic carbon sink prediction in large-scale ecological restoration areas and improve carbon sink prediction accuracy and computational efficiency.

[0056] Step S100: specifically includes:

[0057] Step S110: Acquire a multi-source time series dataset of the target area, the multi-source time series dataset including remote sensing data, ground data, meteorological data, and restoration measure intensity; specifically, remote sensing data includes: vegetation index, used to analyze the dynamic changes of vegetation growth; surface temperature, used to assist in judging the thermal and moisture conditions of the ecosystem; land cover, which can be combined with historical ecological restoration project boundary maps to extract the spatial distribution of desertified areas, wetland degradation areas, and forest restoration areas; ground data includes: soil attribute data, and several sampling points are set in typical desertified areas, wetland degradation areas, and forest restoration areas to collect parameters such as 0-30cm soil organic carbon content, pH value, and bulk density; the restoration measure intensity includes: desertified areas: grass planting and sand fixation area, vegetation recovery rate; wetland degradation areas: water replenishment and water level monitoring data; forest restoration areas: afforestation density and tree species composition;

[0058] Step S120: performing spatiotemporal alignment and standardization on the multi-source time series datasets; specifically, constructing a three-dimensional spatiotemporal data matrix, which can be expressed as:

[0059]

[0060] Where, Represents a three-dimensional spatiotemporal data matrix; Indicates the time step, that is, the time interval and sequence length of data collection; Indicates the number of spatial units, that is, the number of grid cells in the plateau wetland grassland area A at the preset resolution, which is used to divide the spatial units of the study area; Represents data dimensions, including various indicators of multi-source data such as remote sensing data, ground data, and meteorological data;

[0061] For the time series of each grid cell, Z-score standardization is used, and the spatiotemporal data matrix after standardization can be expressed as .

[0062] Step S130: Extract spatial features through weighted fusion and principal component analysis. Specifically, first, based on the historical carbon sink data of plateau wetland grassland area A, the historical carbon sink data and various ecological factors are combined into a training dataset. Machine learning models such as random forest and gradient boosting trees are selected for training. During the model training process, the contribution weight of each ecological factor to carbon sink is determined through feature importance assessment. , indicating the The weight of the multidimensional data in the carbon sink impact is used to reflect the relative importance of the factor to the carbon sink; then, principal component analysis is performed on the multidimensional time series of each spatial unit, and a few principal components that can explain the main changes in the data are extracted by calculating the covariance matrix, eigenvalues and eigenvectors. , indicating the The principal component vectors of the spatial units were obtained through principal component analysis. Finally, the weighted high-weight features were combined with the principal component analysis results to generate a comprehensive spatial feature vector to characterize the spatial heterogeneity of vegetation cover, soil carbon storage, etc. in the restoration area. The comprehensive spatial feature vector can be expressed as:

[0063]

[0064] Where, Indicates the The comprehensive spatial feature vector of the spatial units; Indicates the The first spatial unit dimensional data; this embodiment uses machine learning and principal component analysis methods to deeply explore the relationship between various ecological factors and carbon sinks in multi-source data. By determining factor weights and extracting principal components, it effectively integrates data information, highlights key spatial features, reduces data dimensions while retaining main information, and enables the generated comprehensive spatial feature vector to accurately reflect the spatial heterogeneity of the restoration area, providing high-quality spatial feature input for the carbon sink prediction model, which helps to improve the accuracy and efficiency of carbon sink prediction.

[0065] Step S140: extracting the time series fluctuation pattern through frequency domain transformation and time domain regression;

[0066] Step S140 specifically includes:

[0067] Step S141: Perform frequency domain transformation on the standardized multi-source time series data, identify the main components of the annual and seasonal cycles through fast Fourier transform, and decompose the multi-scale time series features through discrete wavelet transform. Specifically, the carbon sink time series data is converted to the frequency domain through fast Fourier transform. In the frequency domain space, by analyzing the amplitude of the data, the annual and seasonal components with significant energy are identified. These components reflect the regular fluctuations of the carbon sink with the change of seasons and years. For example, the enhanced photosynthesis during the peak season of vegetation growth leads to a periodic increase in the carbon sink in a specific season. At the same time, the discrete wavelet transform is used to perform multi-scale decomposition on the time series data, decomposing the original sequence into sub-sequences at different frequency levels, and further exploring the characteristics of carbon sink changes at different time scales, such as the impact of short-term climate anomalies on carbon sinks, or the trend of carbon sink changes brought about by the gradual implementation of long-term ecological restoration measures.

[0068] Step S142: Perform time domain analysis on the same time series data and fit the long-term trend term through linear regression. Specifically, the linear regression model is used to perform time domain analysis on the carbon sink time series, which can be expressed as:

[0069]

[0070] in express Carbon sequestration at a given moment; represents the slope of the linear regression, which is used to quantify the long-term trend of carbon sequestration; represents the intercept, reflecting the initial level of carbon sequestration at the beginning of time; Represents the error term, which is the random fluctuation that the model cannot explain; it is solved by the least squares method and , thereby determining the long-term trend line of carbon sinks;

[0071] Step S143: Superimpose the annual and seasonal cycle principal components with the trend term, and generate a composite time series fluctuation feature through weighted summation. Specifically, superimpose the annual and seasonal cycle components obtained by FFT in step S141 and the long-term trend term obtained by linear regression in step S142, and generate a composite time series fluctuation feature through weighted summation, which can be expressed as:

[0072]

[0073] Where, Indicates the The composite time series fluctuation characteristics of spatial units are used as input for the carbon sink dynamic prediction model; It represents the carbon sink periodic component corresponding to the annual cycle and seasonal cycle frequency index after fast Fourier transform, which represents the periodic variation law of carbon sink on the annual and seasonal scales; Indicates the corresponding frequency index; It represents the long-term trend of carbon sinks obtained based on linear regression, which reflects the changes of carbon sink amount over time. The continuous change trend of

[0074] In summary, this embodiment deeply explores the periodic change patterns and long-term evolution trends in carbon sink time series data by combining frequency domain transformation with time domain regression. Frequency domain transformation can accurately identify annual and seasonal periodic components, thereby grasping the periodic fluctuation characteristics of carbon sinks over time. Time domain regression fits long-term trend terms and quantifies the direction and rate of change of carbon sinks over a longer time span. The superposition and fusion of the two generate composite time series fluctuation characteristics, which not only retains the stable periodic information of carbon sink changes, but also reflects its continuous change trend. It provides a feature input containing rich time dimension information for the carbon sink dynamic prediction model, enhances the accuracy and reliability of the carbon sink prediction model in time series, helps to more accurately predict the dynamic change trend of carbon sinks in ecological restoration areas, and provides scientific and effective time dimension data support for ecological restoration decision-making.

[0075] Step S200: Based on the spatial characteristics and temporal fluctuation patterns, the ecological restoration area within the target area is divided into multiple dynamic aggregation units, and the dynamic aggregation units with similar characteristics are merged; specifically, based on the multi-source time series data set obtained in step S100 and the spatial characteristics and temporal fluctuation patterns extracted therefrom, step S200 realizes the division and merging of dynamic aggregation units in the dynamic prediction of carbon sinks in the ecological restoration area of plateau wetland grassland area A through the following process; the ecological restoration area within the target area, that is, the plateau wetland grassland area A, is divided into dynamic aggregation units. The dynamic aggregation units are homogeneous areas divided based on the similarity of temporal and spatial characteristics, which are used to reduce the complexity of carbon sink calculation in large-scale areas, thereby improving the computational efficiency of carbon sink prediction in large-scale ecological restoration areas.

[0076] Step S200 specifically includes:

[0077] Step S210: Constructing an ecological correlation matrix based on the similarity of spatial features; specifically, using the spatiotemporal data matrix standardized in step S100 , for each spatial unit, the linear correlation between ecological variables is analyzed, and the correlation coefficient of each ecological variable (such as vegetation index, soil organic carbon content, meteorological data) in the time dimension is calculated, which can be expressed as:

[0078]

[0079] Where, Indicates the and The correlation coefficient of two ecological variables is used to quantify the degree of linear correlation between two ecological variables. The closer the absolute value of is to 1, the stronger the linear correlation is; Indicates that at time step , No. Space unit, Dimensionally normalized data; Indicates the The mean of the dimensional data in the time dimension is used to measure the average level of the ecological variable in the entire time series; Indicates the The mean of the dimensional data in the time dimension; by traversing all combinations of ecological variables, a complete ecological correlation coefficient matrix is constructed, which represents the degree of linear correlation between each ecological variable; then, a correlation coefficient threshold is set, when When it is greater than or equal to the threshold, it is considered and The ecological variables are highly correlated; for highly correlated feature pairs, the contribution weights of each ecological factor to carbon sink obtained by machine learning model training in step S100 are combined, the features that contribute more to carbon sink prediction are retained, and redundant or strongly correlated features are eliminated, thereby achieving data dimensionality reduction and obtaining a data matrix after dimensionality reduction.

[0080] Step S220: Divide the initial dynamic aggregation units using the spectral clustering algorithm. Specifically, based on the comprehensive spatial feature vector and the composite time series fluctuation feature extracted in step S100, calculate the comprehensive similarity between the spatial units, which can be expressed as:

[0081]

[0082] Where, Indicates the The spatial unit and The comprehensive similarity between spatial units; Indicates the and The square of the Euclidean distance of the comprehensive spatial feature vector of each spatial unit; The bandwidth parameter representing the spatial feature similarity is used to adjust the decay rate of the spatial feature similarity; Indicates the and The square of the Euclidean distance of the composite time series fluctuation characteristics of each spatial unit; The bandwidth parameter representing the similarity of time series features is used to adjust the decay rate of the similarity of time series features. , construct a joint similarity matrix, which comprehensively considers the similarity of spatial features and temporal features; then, construct a degree matrix based on the joint similarity matrix between spatial units, calculate the normalized Laplace matrix based on the degree matrix, perform eigendecomposition on the Laplace matrix, and select the eigenvectors corresponding to multiple minimum eigenvalues to form a low-dimensional embedding space; finally, use the K-means algorithm to cluster the embedded vectors to obtain the initial dynamic aggregation unit.

[0083] Step S230: Merge adjacent units with consistent characteristic distributions using spatial proximity constraints and temporal synchrony. Specifically, with respect to spatial proximity constraints, a geographic information system (GIS) is used to determine the proximity relationships of spatial units within plateau wetland grassland region A. A spatial distance threshold is set, and only geographically adjacent units are considered as potential merge targets to avoid the generation of isolated, spatially fragmented regions. With respect to temporal synchrony, the consistency of carbon sink trends is measured by calculating the Pearson correlation coefficient of carbon sink trends between adjacent units. The KL divergence of unit characteristic distributions is also calculated to assess the similarity of characteristic distributions. When the KL divergence of adjacent units is less than 0.2 and the Pearson correlation coefficient is greater than 0.8, the two units are deemed to have consistent characteristic distributions and synchronized carbon sink trends. These units are then merged to generate a final dynamic aggregate unit with spatiotemporal consistency.

[0084] This embodiment identifies redundant features and reduces dimensionality by constructing an ecological correlation matrix, thereby reducing noise and redundant information in the data and improving the efficiency of subsequent analysis. The spectral clustering algorithm combines spatial features and temporal features to divide the initial dynamic aggregation units, fully considering the temporal and spatial similarities of ecosystems, making the division results more consistent with the actual ecological process, and accurately reflecting the differences in carbon sink characteristics in different areas within the ecological restoration area. The unit merging strategy based on spatial proximity constraints and temporal synchronization avoids the problem of spatial fragmentation, ensures the consistency of ecological attributes and carbon sink change trends within the dynamic aggregation unit, and provides a more reasonable and simplified data structure for the subsequent construction of differentiated carbon sink accounting models and dynamic prediction of carbon sinks. This processing method greatly reduces the computational complexity of carbon sink prediction in large-scale and complex ecological restoration areas, significantly improves computational efficiency, and at the same time enhances the accuracy of carbon sink prediction through precise unit division and merging.

[0085] See Figure 2 As shown, step S300: Based on the ecosystem type of each merged dynamic aggregation unit, a differentiated carbon sink accounting model is constructed, multi-scale time series features are extracted, and key influencing factors are screened. Specifically, based on the multi-source data features obtained in step S100 and the dynamic aggregation units divided in step S200, step S300 constructs a differentiated carbon sink accounting model, extracts multi-scale time series features, and screens key influencing factors in the dynamic carbon sink prediction of the ecological restoration area of plateau wetland grassland region A through the following steps: Key influencing factors are ecological factors that significantly contribute to carbon sink changes in specific ecosystem types (such as forests, wetlands, and desertified areas), and are screened through gradient sensitivity analysis. For example, vegetation cover in desertified areas, wetland water level changes, and forest leaf area index are retained as key influencing factors if the absolute value of the partial derivative of these factors exceeds a threshold.

[0086] Step S300 specifically includes:

[0087] Step S310: Select differentiated carbon sink models according to the ecosystem type. The forest restoration area adopts the light energy utilization model, the wetland degradation area adopts the flux coupling model, and the desertification restoration area adopts the vegetation and soil carbon coupling model: Specifically, the spatial features and terrain data extracted in step S100 are combined with the dynamic aggregation units divided in step S200, and a random forest classifier is used to classify the ecosystem type of each dynamic aggregation unit; taking the plateau wetland grassland area A ecological restoration area as an example, it is mainly divided into desertification ecological restoration area, wetland degradation ecological restoration area and forest ecological restoration area. For different ecosystem types, an appropriate carbon sink model is selected: the vegetation and soil carbon coupling model is used in the desertification ecological restoration area. This model considers the interaction between vegetation growth and carbon sequestration and changes in soil carbon pools; the methane-carbon dioxide flux coupling model is used in the wetland degradation ecological restoration area. This model focuses on the greenhouse gas flux exchange in wetland ecosystems; the light energy utilization rate model is used in the forest ecological restoration area. That is, the carbon sink is estimated based on the absorption and conversion efficiency of vegetation to light energy. The light energy utilization rate model estimates carbon sinks through the photosynthesis efficiency of vegetation and is suitable for forest restoration areas with high vegetation coverage and significant light energy absorption. Ecosystem type adaptive modeling is used, which can be expressed as:

[0088]

[0089] Where, Represents a dynamic aggregation unit The estimated carbon sequestration value reflects the carbon sequestration amount of the unit under a specific ecosystem type; Indicates ecosystem type For the first The weights of different carbon sequestration mechanisms reflect the relative contribution of different carbon sequestration mechanisms to the total carbon sequestration in this ecosystem type, which can be determined through model training and optimization; Indicates the basic carbon sink function corresponding to different ecosystems; Represents a dynamic aggregation unit The comprehensive spatial eigenvector of ; Represents a dynamic aggregation unit Composite time series fluctuation characteristics;

[0090] Step S320: Extract the time series features of the year, season and month by wavelet multi-scale decomposition; Specifically, obtain the composite time series fluctuation feature data of each dynamic aggregation unit from step S140, and divide it into yearly, seasonal and monthly levels. ,season ,moon Split into subsequences ; For each subsequence, the discrete wavelet transform is performed using the db4 wavelet basis function, which can be expressed as:

[0091]

[0092] Where, Represents a dynamic aggregation unit In time scale The subsequence below is the result of dividing the original composite time series fluctuation characteristics according to a specific time scale; represents the approximate coefficient, which is used to retain the low-frequency signal and reflect the long-term trend and main components of carbon sink changes; It represents the detail coefficient, characterizes the high-frequency signal, and reflects the short-term fluctuations and detailed information of carbon sink changes; The wavelet function representing the db4 wavelet basis is used to analyze the details of the signal in discrete wavelet transform; Represents the scaling function of the db4 wavelet basis, which is used to analyze the approximate part of the signal in discrete wavelet transform; , represents the number of decomposition levels;

[0093] The approximate coefficients at different scales are spliced together to generate a multi-scale feature vector, which can be expressed as:

[0094]

[0095] Where, Represents a multi-scale feature vector, characterizing the carbon sink fluctuation pattern at the annual, seasonal, and monthly levels. The vector comprehensively characterizes the carbon sink fluctuation pattern at the annual, seasonal, and monthly levels, providing rich time dimension features for carbon sink prediction. The subscripts indicate the number of layers of wavelet multi-scale decomposition, 1, 2, and 3 correspond to the characteristic data obtained from the first, second, and third layers of decomposition, respectively; the superscript 1 indicates that the time scale is yearly, the superscript 4 indicates that the time scale is quarterly, and the superscript 12 indicates that the time scale is monthly;

[0096] Step S330: Screen key influencing factors based on gradient sensitivity analysis and eliminate redundant factors whose contribution is lower than the threshold; specifically, for each dynamic aggregation unit The carbon sink accounting model is used to calculate the gradient of the carbon sink prediction value to each influencing factor; for example, in the carbon sink model of the desertification ecological restoration area, the input factor Including vegetation coverage, soil moisture content, and wind and sand erosion intensity; in the carbon sink model of wetland degradation ecological restoration area, the input factor This may include water level changes, vegetation types, and temperature. First, the partial derivative of the carbon sink prediction value with respect to the input factors is calculated, which can be expressed as:

[0097]

[0098] Where, Represents a dynamic aggregation unit The predicted carbon sink value; Indicates the Input factors, namely various ecological factors that affect carbon sink changes; represents the number of carbon sink mechanisms; Indicates the carbon sink prediction value for the Input factors The partial derivative of is used to measure the impact of changes in input factors on the predicted value of carbon sinks, that is, the gradient;

[0099] Then, the dynamic threshold screening factor can be expressed as:

[0100]

[0101] Where, Display unit The key factor set includes the input factor numbers that contribute most to carbon sink prediction; It represents the threshold ratio, preferably 0.3, to retain the top 30% of the contributing factors; Represents a dynamic aggregation unit The predicted carbon sink value; Represents the maximum absolute value of the partial derivatives of all input factors, which serves as the benchmark for determining the threshold; After Represents the standard delimiter used to separate input factors from filter conditions.

[0102] This example constructs a differentiated carbon sink accounting model for different ecosystem types in the ecological restoration area of Region A of the plateau wetland grassland. This model fully considers the differences in the carbon sink formation mechanisms of each ecosystem and fundamentally improves the accuracy of carbon sink estimation. On the one hand, it extracts temporal characteristics at the annual, seasonal, and monthly levels through wavelet multi-scale decomposition, comprehensively capturing the changing patterns of carbon sinks at different time scales. This enables the model to capture subtle fluctuations and long-term trends in the dynamic changes of carbon sinks, enhancing its ability to depict the time series characteristics of carbon sinks. On the other hand, it screens key influencing factors based on gradient sensitivity analysis, effectively eliminating redundant factors. While reducing model complexity and improving computational efficiency, it also highlights the core factors that significantly affect carbon sinks and effectively reduces data noise interference.

[0103] See Figure 3 As shown, step S400: based on the multi-scale time series characteristics and the screened key influencing factors, a carbon sink dynamic prediction model is constructed and the carbon sink dynamic prediction result is output; specifically, based on the multi-source data characteristics extracted in step S100, the dynamic aggregation units divided in step S200, and the differentiated carbon sink accounting model constructed in step S300 and the extracted multi-scale time series characteristics and key influencing factors, step S400 constructs a carbon sink dynamic prediction model and outputs the result in the carbon sink dynamic prediction of the plateau wetland grassland area A ecological restoration area through the following steps:

[0104] Step S400 specifically includes:

[0105] Step S410: Construct a spatiotemporal graph convolutional network to integrate the carbon flux correlation and temporal evolution patterns of spatial neighborhood units. Specifically, spatial neighborhood units refer to spatially adjacent or neighboring units within the spatiotemporally consistent dynamic aggregation units delineated in step S200. Due to their geographical proximity, these units exhibit correlations in carbon flux changes (e.g., sharing similar vegetation cover, soil properties, or meteorological conditions), and their spatial dependencies need to be modeled through a graph structure. Adjacency relationships are determined using GIS. For example, in the actual scenario of ecological restoration area A in plateau wetland grassland, a specific wetland dynamic aggregation unit is considered the central unit, and surrounding dynamic aggregation units such as surrounding wetlands, adjacent grasslands, or forests may also become its spatial neighborhood units.

[0106] Step S410 specifically includes:

[0107] Step S411: Dynamically balance the output weights of the spatial graph convolution and the time series model using a gating matrix. Specifically, this embodiment introduces a gating matrix mechanism to effectively integrate the outputs of the spatial graph convolution and time series models. During training, the backpropagation algorithm dynamically adjusts the gating matrix parameters based on the model prediction error. This balances the output weights of the spatial graph convolution layer (which captures the correlation of carbon fluxes between spatial neighborhood units) and the time series model, preventing excessive dominance of either information and ensuring that the model fully utilizes both spatial and temporal dimensions.

[0108] Step S412: Construct a spatial graph structure based on the carbon flux correlations of spatially neighboring units, with edge weights determined by the covariance of key influencing factors. Specifically, based on the adjacency relationships of the dynamically aggregated units determined in step S200 and the key influencing factors screened in step S300, the carbon flux correlations between spatially neighboring units are calculated. The covariance of the key influencing factors is used as the edge weight to construct the spatial graph structure. The covariance reflects the degree of coordination between the changes in the key influencing factors of different units. The larger the covariance, the higher the edge weight, indicating a stronger correlation between the carbon fluxes of the units, thereby constructing a graph structure that reflects spatial dependencies.

[0109] Step S413: Use the Sigmoid function to generate a gating value to control the fusion ratio of spatial dependency and temporal dependency; specifically, use the Sigmoid function to generate a gating value, which maps the input to the range of 0-1. In the spatiotemporal graph convolutional network, the gating value is used to control the fusion ratio of spatial dependency information and temporal dependency information. When the gating value is close to 1, more spatial dependency information is introduced; when it is close to 0, temporal dependency information is emphasized. In this way, the degree of fusion of the two types of information can be flexibly adjusted to adapt to the changing characteristics of carbon sinks under different spatiotemporal conditions. The update formula of the spatiotemporal graph convolutional network can be expressed as:

[0110]

[0111] Where, Indicates the Dynamic aggregation units in The predicted carbon sink value at the moment; Represents the activation function, which is used to perform nonlinear transformation on the result of weighted summation; Represents the time step The corresponding weight matrix is used to adjust the weights of the results of spatial graph convolution operations at different time steps; Indicates the Units in The characteristic matrix at the moment Perform spatial graph convolution to capture the correlation between the carbon flux of the unit and the neighboring units. Contains the Dynamic aggregation units in The multi-scale temporal features and key influencing factor information of the moment; B is the bias vector, which adds an offset to the model; Represents element-wise multiplication operation; Indicates the Units in The characteristic matrix at the moment Perform long-short-term memory network calculations to learn the temporal evolution law of the carbon sink of this unit.

[0112] Step S420: Quantify the hysteresis effect of ecological restoration measures and modify the base prediction value using an exponential decay model. Specifically, the base prediction value is the carbon sink prediction value initially output by the spatiotemporal graph convolutional network, which does not consider the hysteresis effect and spatial synergy gain of ecological restoration measures. The base prediction value is calculated based solely on multi-scale temporal characteristics and key influencing factors. It is subsequently modified using the exponential decay model and synergy coefficient to obtain the final prediction value that includes hysteresis effects and spatial synergy.

[0113] Step S420 specifically includes:

[0114] Step S421: Quantify the hysteresis effect of ecological restoration measures; specifically, first define the effective delay time of the restoration measures and construct a time offset sequence; specifically, clarify the various restoration measures in the ecological restoration area of the plateau wetland grassland area A, and set a total of Class repair measures ; Based on the knowledge and historical data of ecological restoration, define the effective delay time of each restoration measure , that is, the time interval from the implementation of restoration measures to the significant impact on carbon sinks. , for each dynamic aggregation unit Construct a time offset sequence and rearrange the historical restoration measures data according to the time offset to analyze the potential impact of restoration measures on carbon sinks at different time points; then, within the preset time window, for each dynamic aggregation unit , calculate the carbon sink prediction value for restoration measures The partial derivative of is used to measure the impact of historical measures on current carbon sinks at different time points. Type 1 repair measures The contribution coefficient of the carbon sink of a dynamic aggregation unit can be expressed as:

[0115]

[0116] Where, Indicates the Type 1 repair measures The contribution coefficient of the carbon sink of each dynamic aggregation unit is normalized and used as the weight to reflect the relative contribution of this type of restoration measures to the change of carbon sink; Indicates the Class repair measures; represents the lag time, i.e., the time interval between the implementation of the restoration measures and their obvious impact on carbon sinks; Represents the predicted carbon sink value After time shift Class repair measures The partial derivative of The changes of time The impact of carbon sink prediction values at each moment; Indicates the time offset after Class repair measures, Used to iterate over all remediation action categories.

[0117] Step S424: According to the adjacency relationship of the dynamic aggregation unit, the contribution weight of the similar measures in the adjacent units is increased; specifically, the adjacency relationship of the dynamic aggregation unit is determined according to the spatial graph structure constructed in step S412, and the synergy coefficient is set for the case where the similar repair measures are in the adjacent units. , and Indicates the adjacent unit number, Indicates the category of repair measures;

[0118] Step S425: Dynamically adjust the synergy coefficient through iterative optimization to maximize the spatial synergy gain. Specifically, an optimization algorithm such as gradient descent is used, with the objective function of maximizing the total amount of carbon sinks, and the synergy coefficient is iteratively updated. In each iteration, the gradient of the objective function with respect to the synergy coefficient is calculated, and the synergy coefficient value is adjusted according to the gradient until the objective function converges or meets the accuracy requirements. Finally, the dynamic prediction result of carbon sinks including spatial synergy gain is obtained. The dynamic prediction result of carbon sinks including spatial synergy gain can be expressed as:

[0119]

[0120] Where, Indicates the Dynamic aggregation units in The dynamic prediction value of carbon sinks including spatial synergy gain at every moment, that is, the final prediction result after considering the hysteresis effect of restoration measures and spatial synergy; Indicates the Dynamic aggregation units in The basic carbon sink prediction value without considering the hysteresis effect and spatial synergy of restoration measures is initially output by the spatiotemporal graph convolutional network; Indicates the Type 1 repair measures Contribution coefficient of carbon sink of each dynamic aggregation unit; Represents an exponential decay function, used to describe the time-varying effects of repair measures The attenuation of Reflects the decay rate; Indicates the Class repair measures in time intensity function.

[0121] Step S430: Output the dynamic prediction result of carbon sink including spatial synergy gain; specifically, the calculated dynamic prediction value of carbon sink including spatial synergy gain The final output will provide data support for carbon sequestration management and ecological restoration decision-making in the plateau wetland grassland area A.

[0122] This embodiment effectively integrates the spatial neighborhood correlation and temporal evolution law of carbon sinks in ecological restoration areas by constructing a spatiotemporal graph convolutional network, and utilizes multi-scale temporal characteristics and key influencing factors to accurately capture the changing pattern of carbon sinks in the spatiotemporal dimension, thereby improving the accuracy of carbon sink prediction. By quantifying the lag effect of ecological restoration measures and making corrections, the prediction results are made more consistent with the delayed characteristics of the impact of restoration measures on carbon sinks in actual ecological processes, making the carbon sink prediction results more suitable for display. Taking into account spatial synergy gain and iteratively optimizing the synergy coefficient reflects the clustering effect of ecological restoration, maximizes the restoration benefit, and helps to evaluate the long-term effects of restoration measures. Overall, step S400 achieves accurate prediction of carbon sink dynamics, provides a scientific basis for carbon sink management and ecological restoration decision-making in the ecological restoration area of plateau wetland grassland area A, helps to evaluate the long-term effects of restoration measures, and improves the accuracy and reliability of carbon sink predictions in large-scale ecological restoration areas.

[0123] Step S500: maximizing the carbon sink increment as the goal, solving a restoration decision plan based on the carbon sink dynamic prediction result;

[0124] Step S500 specifically includes:

[0125] Step S510: Construct a multi-objective optimization model with constraints including budget limits, mutual exclusivity of measures, and biodiversity protection. Specifically, in the ecological restoration area of plateau wetland grassland region A, the decision variables are defined based on the dynamic aggregation units divided in step S200 and the planning period. , indicating that in the dynamic aggregation unit In time Whether to implement repair measures , The value is 0 or 1, 0 means not implemented, 1 means implemented; for example , it means that the first type of restoration measures are implemented in the third dynamic aggregation unit and the second time step; based on the dynamic prediction value of carbon sink obtained in step S430 , the contribution coefficient of the repair measures calculated in step S420 , combined with the attenuation coefficient and the delay time for the repair measures to take effect, the objective function is constructed and can be expressed as:

[0126]

[0127] Where, represents the number of time steps in the planning period; Indicates the total number of dynamic aggregation units; Indicates the Dynamic aggregation units in Dynamic prediction value of carbon sink including spatial synergy gain at every moment; Indicates the Type 1 repair measures Contribution coefficient of carbon sink of each dynamic aggregation unit; represents the attenuation coefficient; express Delay in the effectiveness of such remediation measures; represents the decay of restoration measures over time; the objective function comprehensively considers the current carbon sink forecast value and the potential increase in carbon sinks in the future after the implementation of restoration measures; the function aims to maximize the carbon sink increment of all dynamic aggregation units during the entire planning period, taking into account not only the current carbon sink forecast value, but also the potential increase in carbon sinks in the future after the implementation of restoration measures, and through the exponential decay function Reflects the attenuation of the effectiveness of restoration measures over time;

[0128] Constraints include:

[0129]

[0130] Where, represents the decision variable, unit In time Whether measures are implemented ; Indicate measures In the unit The single-period cost of represents the total budget constraint; Indicates the minimum duration of the measure; Display unit biodiversity index; represents the budget constraint, which is used to ensure that the total cost of implementing repair measures for all units does not exceed the total budget during the entire planning period; Indicates mutually exclusive constraints: at the same time , each unit Only one remediation measure can be implemented to avoid conflicting measures being carried out simultaneously; Indicates the duration constraint of the measure: To ensure that the repair measures can play an effective role, the minimum duration L of the measure is set; It is an auxiliary variable to characterize whether the measures are implemented; Class repair measures in unit , from time arrive During this period, the measure must be implemented at least Second-rate; It stated that biodiversity is one of the important goals of ecological restoration; Display unit The biodiversity index is determined by monitoring species richness, ecosystem diversity and other indicators. In order to protect biodiversity, the biodiversity index of each unit is required to be no less than the minimum threshold. ;

[0131] Step S520: The global optimization problem is grouped by ecological type and solved in parallel using an adaptive decomposition algorithm. Specifically, the global optimization problem covers the entire ecological restoration area, with the goal of maximizing the carbon sequestration increment for all dynamically aggregated units over the planning period, subject to constraints such as budget, measure mutual exclusivity, and biodiversity. Specifically, the multi-objective optimization model must determine the combination of restoration measures to be implemented for each unit at different time steps. Due to its high computational complexity, global optimization requires parallel solving by ecological type. In this embodiment, all dynamically aggregated units are divided into different ecological type groups based on their ecosystem types (desertification restoration area, wetland degradation area, forest restoration area) as determined in step S200. For example, within plateau wetland grassland area A, desertified areas with similar ecological characteristics and carbon sequestration mechanisms are grouped into the desertification restoration area ecological type group. The original global optimization problem is then decomposed into subproblems for each ecological type group. During the decomposition process, similarities between units within a group are considered, such as the similarities in vegetation, soil conditions, and carbon sequestration response mechanisms within each unit within the desertification restoration area, as well as their correlation with other groups. This reduces duplicate calculations and lowers computational complexity. Finally, the subproblems corresponding to each ecological type group are solved in parallel using multi-core processors or cluster computing resources. Optimization algorithms such as genetic algorithms and particle swarm optimization can be used to solve the optimal solution for each subproblem separately, obtaining the optimal values of the restoration decision variables for each unit within each ecological type group at different times. This significantly shortens solution time and improves computational efficiency.

[0132] Step S530: Dynamically adjust resource allocation and give priority to restoration measures with high carbon sink gain and strong spatial synergy; specifically, according to the synergy coefficient of step S430 , dynamically adjust the priority of measures, which can be expressed as:

[0133]

[0134] Where, Display unit The domain unit collection of represents the synergistic gain weight. This formula increases the weight of restoration measures with strong carbon sink gains and spatial synergy by accumulating the synergistic effects of adjacent units. A larger synergy coefficient indicates a stronger synergistic effect across adjacent units. Based on the adjusted contribution coefficient, resource allocation (funds, manpower, materials, etc.) is dynamically adjusted based on the principles of high carbon sink gain and strong spatial synergy. Prioritize resources for restoration measures that generate greater carbon sink increments and drive synergistic gains in adjacent units. During implementation, resource allocation plans will be continuously optimized based on actual monitoring data and feedback to ensure efficient resource utilization and maximize ecological restoration effectiveness.

[0135] In summary, this embodiment, by constructing a multi-objective optimization model, comprehensively considers multiple factors such as carbon sink increment maximization, budget constraints, mutual exclusivity of measures, and biodiversity protection, providing a scientific and reasonable framework for ecological restoration decision-making, avoiding the ecological imbalance or resource waste that may be caused by single-objective optimization. The adaptive decomposition algorithm is grouped by ecological type and solved in parallel, which effectively reduces the computational complexity of large-scale complex optimization problems, improves the solution efficiency, and can quickly obtain the restoration measure decision of each unit. Dynamically adjust the resource allocation strategy, guided by carbon sink gain and spatial synergy, give priority to the implementation of high-efficiency restoration measures, ensure that limited resources are accurately allocated and efficiently utilized, and take into account biodiversity protection while achieving carbon sink growth, effectively improving the comprehensive benefits of ecological restoration. This step has formulated a practical and feasible restoration decision-making plan for the ecological restoration area of plateau wetland grassland area A, provided strong support for the scientific implementation of ecological restoration projects, and is of great significance to promoting carbon sink enhancement and sustainable development of ecosystems in large-scale ecological restoration areas.

[0136] Taking the integrated protection and restoration project for plateau wetland grassland area A as an example, typical ecosystem restoration types such as desertification restoration, wetland degradation, and forest restoration areas were selected as research objects. Carbon sink prediction and decision-making research was carried out in plateau wetland grassland area A after ecological restoration. The goal was to achieve dynamic carbon sink prediction in large-scale ecological restoration areas and improve carbon sink prediction accuracy and calculation efficiency. The following provides a specific embodiment of the present invention:

[0137] An integrated protection and restoration project is being carried out in an area of 12,000 square kilometers in Area A of the plateau wetland grassland. The area includes desertification restoration areas, accounting for 35%: affected by wind erosion, with low vegetation coverage; wetland degradation areas, accounting for 40%: hydrological conditions have deteriorated and methane emissions have increased; forest restoration areas, accounting for 25%: artificial forests have been restored, and carbon sequestration potential is significant; in order to scientifically evaluate the effects of ecological restoration, realize dynamic prediction of carbon sinks in large-scale ecological restoration areas, and improve prediction accuracy and computational efficiency, a dynamic prediction method of carbon sinks based on time-series remote sensing is specially applied.

[0138] Data sources: Remote sensing data: Sentinel-2 for vegetation indices, MODIS for surface temperature, and Sentinel-1 for surface humidity. Meteorological data cover precipitation, temperature, and radiation. Ground data primarily includes soil organic carbon content. Restoration measures include grass reseeding, wetland water replenishment, and forest tending. All data were standardized to a spatial resolution of 1 km × 1 km to construct a three-dimensional spatiotemporal data matrix. Spatiotemporal alignment and standardization were then performed to lay the foundation for subsequent analysis. The entire area was divided into 1 km × 1 km grids, totaling 12,000 cells.

[0139] Based on the spatial characteristics extracted in step S100, such as vegetation cover and soil properties, and their temporal fluctuations, an ecological correlation matrix was constructed. By analyzing the correlations among ecological variables, redundant features were removed. For example, in the soil property data, some trace element indicators with excessive correlations were removed. A spectral clustering algorithm was then used to calculate the comprehensive similarity between units. Similar units were then merged, combining spatial proximity constraints (i.e., only units with shared boundaries or within 1 km of each other) and temporal synchrony verification (a Pearson correlation coefficient for carbon sink trends greater than 0.8 and a KL divergence of characteristic distributions less than 0.2). Ultimately, 115 dynamic aggregation units were generated, including 42 for desertified areas, 38 for wetland areas, and 35 for forested areas. Experiments showed that the computational time after dynamic aggregation was significantly reduced to 4.8 hours, compared to 72 hours for the full-area model, significantly improving computational efficiency. Appropriate models were selected for different ecosystem types. A vegetation-soil carbon coupling model was used for desertified restoration areas, focusing on the synergistic effect between vegetation restoration and soil carbon sequestration. Vegetation cover, weighted 45%, wind and sand erosion intensity, weighted 30%, were key influencing factors, while soil moisture, weighted 20%, and soil organic matter, weighted 5%, were also considered. A methane-CO2 flux coupling model was used for degraded wetland areas, with key factors including water level change, weighted 50%, vegetation type, and temperature, weighted 30%. A light energy utilization efficiency model was used for forest restoration areas, with key factors including leaf area index, weighted 40%, photosynthetically active radiation, and temperature, weighted 25%. Wavelet multiscale decomposition was used to extract temporal characteristics at the annual, seasonal, and monthly levels, comprehensively capturing the dynamics of carbon sinks and providing an accurate basis for carbon sink prediction.

[0140] This example constructs a spatiotemporal graph convolutional network, integrating the correlation of carbon fluxes between spatially neighboring units with temporal evolution patterns. This approach quantifies the lagged effects of ecological restoration measures. For example, reseeding grass seeds in desertified areas begins to significantly impact carbon sinks approximately six months after implementation; and three months after wetland water replenishment measures are implemented, carbon sink trends gradually emerge. An exponential decay model is used to correct the baseline prediction value, factoring in spatial synergy gains. This output includes a dynamic carbon sink prediction result that incorporates spatial synergy gains. Comparison with actual monitoring data shows significant improvement in prediction accuracy compared to traditional models, effectively reflecting the dynamics of carbon sinks. Ultimately, the goal is to maximize carbon sink increments, while also considering a budget limit of 100 million yuan, mutual exclusivity of measures, and biodiversity conservation. The biodiversity index of each unit must be no less than 80% of the baseline value. An adaptive decomposition algorithm is employed to group and solve the problem in parallel by ecological type. The subproblem in desertified areas prioritizes vegetation restoration measures, the subproblem in wetlands focuses on hydrological regulation measures, and the subproblem in forest areas focuses on tree tending measures. Restoration measures with high carbon sink gains and strong spatial synergy are prioritized. The final restoration decision plan was determined: 40 million yuan was invested in the desertification area to implement 150 square kilometers of reseeding grass seeds and fencing for enclosure; 35 million yuan was invested in the wetland degradation area to carry out 2 million cubic meters of water replenishment projects and wetland vegetation restoration; 25 million yuan was invested in the forest restoration area to carry out 1 million seedlings replanting and forest tending, providing a scientific and reasonable decision-making basis for the ecological restoration project.

[0141] In summary, the method for dynamic prediction of carbon sinks in ecological restoration areas based on time-series remote sensing provided by the present invention divides large-scale and complex ecological restoration areas into relatively homogeneous small units through the division of dynamic aggregation units, thereby reducing the computational complexity and improving the computational efficiency of carbon sink prediction in large-scale areas. By utilizing a differentiated carbon sink accounting model, the carbon sink mechanism is quantified based on the characteristics of different ecosystem types, overcoming the problem that traditional models cannot accurately reflect the differences in ecosystems, thereby improving the accuracy of carbon sink prediction. Through the fusion of multi-scale time series features and the screening of key influencing factors, the carbon sink dynamic prediction model can capture the spatiotemporal characteristics and key driving factors of the dynamic changes in carbon sinks, and further optimize the carbon sink prediction effect. Ultimately, solving the restoration decision-making plan with the goal of maximizing the carbon sink increment can provide a scientific and reasonable decision-making basis for ecological restoration projects, and realize the integration from carbon sink prediction to restoration decision-making.

[0142] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0143] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A method for dynamic prediction of carbon sinks in ecological restoration areas based on time series remote sensing, characterized by: The method comprises: Obtain a multi-source time series dataset of the target area and extract the spatial features and time series fluctuation patterns of the multi-source time series dataset; Based on the spatial characteristics and temporal fluctuation patterns, the ecological restoration area within the target area is divided into multiple dynamic aggregation units, and dynamic aggregation units with similar characteristics are merged, including: Constructing an ecological correlation matrix based on the similarity of spatial characteristics; The spectral clustering algorithm is used to divide the initial dynamic aggregation units, including: constructing a degree matrix based on the comprehensive similarity between spatial units. The comprehensive similarity between spatial units can be expressed as: Where, Indicates the The spatial unit and The comprehensive similarity between spatial units; Indicates the and The square of the Euclidean distance of the comprehensive spatial feature vector of each spatial unit; The bandwidth parameter representing the spatial feature similarity is used to adjust the decay rate of the spatial feature similarity; Indicates the and The square of the Euclidean distance of the composite time series fluctuation characteristics of each spatial unit; The bandwidth parameter representing the similarity of time series features is used to adjust the decay rate of the similarity of time series features; Based on the ecosystem types of each dynamic aggregation unit after merging, a differentiated carbon sink accounting model is constructed to extract multi-scale temporal characteristics and screen key influencing factors, including: Differentiated carbon sequestration models are selected based on ecosystem types: the light energy utilization model is used in forest restoration areas, the flux coupling model is used in wetland degradation areas, and the vegetation and soil carbon coupling model is used in desertification restoration areas; Extract annual, seasonal and monthly time series features through wavelet multi-scale decomposition; Screen key influencing factors based on gradient sensitivity analysis and eliminate redundant factors whose contribution is lower than the threshold; Based on the multi-scale time series characteristics and the screened key influencing factors, a carbon sink dynamic prediction model is constructed and a carbon sink dynamic prediction result is output; With the goal of maximizing carbon sink increment, the restoration decision plan is solved according to the carbon sink dynamic prediction results. The objective function can be expressed as: Where, , representing a dynamic aggregation unit In time Whether to implement repair measures ; represents the number of time steps in the planning period; Indicates the total number of dynamic aggregation units; Indicates the Dynamic aggregation units in Dynamic prediction value of carbon sink including spatial synergy gain at every moment; Indicates the Type 1 repair measures Contribution coefficient of carbon sink of each dynamic aggregation unit; represents the attenuation coefficient; express Delay in the effectiveness of such remediation measures; Represents the decay of a repair measure over time.

2. The method for dynamic prediction of carbon sinks in ecological restoration areas according to claim 1, characterized in that: Obtain a multi-source time series dataset for the target area and extract the spatial features and time series fluctuation patterns of the multi-source time series dataset, including: Obtain a multi-source time series dataset of the target area, wherein the multi-source time series dataset includes remote sensing data, ground data, meteorological data, and the intensity of restoration measures; Performing spatiotemporal alignment and standardization processing on the multi-source time series dataset; Extract spatial features through weighted fusion and principal component analysis; The time series fluctuation pattern is extracted through frequency domain transformation and time domain regression.

3. The method for dynamic prediction of carbon sinks in ecological restoration areas according to claim 2, characterized in that: Extracting the timing fluctuation pattern specifically includes: The standardized multi-source time series data are transformed into the frequency domain, the annual and seasonal main components are identified by fast Fourier transform, and the multi-scale time series features are decomposed by discrete wavelet transform. Perform time domain analysis on the same time series data and fit the long-term trend term through linear regression; The annual cycle and seasonal cycle principal components are superimposed on the trend term, and the composite time series fluctuation characteristics are generated by weighted summation. The composite time series fluctuation characteristics under the time series fluctuation mode can be expressed as: Where, Indicates the The composite time series fluctuation characteristics of spatial units are used as input for the carbon sink dynamic prediction model; The carbon sink period component, which represents the annual and seasonal frequency index after fast Fourier transform, represents the periodic variation of carbon sink on the annual and seasonal scales; represents the frequency index; express Carbon sequestration at a given moment; represents the slope of the linear regression, which is used to quantify the long-term trend of carbon sequestration.

4. The method for dynamic prediction of carbon sinks in ecological restoration areas according to claim 1, characterized in that: Based on the multi-scale time series characteristics and the screened key influencing factors, a carbon sink dynamic prediction model is constructed and the carbon sink dynamic prediction results are output, including: Construct a spatiotemporal graph convolutional network to integrate the carbon flux correlation and temporal evolution of spatial neighborhood units; Quantify the hysteresis effect of ecological restoration measures and correct the basic prediction value through exponential decay model; Output carbon sink dynamic prediction results including spatial synergy gains.

5. The method for dynamic prediction of carbon sinks in ecological restoration areas according to claim 4, characterized in that: The construction of the spatiotemporal graph convolutional network includes: Dynamically balance the output weights of spatial graph convolution and time series models through the gating matrix; A spatial graph structure is constructed based on the carbon flux correlation of spatial neighborhood units, and the edge weights are determined by the covariance of key influencing factors; The Sigmoid function is used to generate the gating value to control the fusion ratio of spatial dependency and temporal dependency.

6. The method for dynamic prediction of carbon sinks in ecological restoration areas according to claim 5, characterized in that: Quantifying the hysteresis effects of ecological restoration measures includes: Define the effective delay time of the repair measures and build a time offset sequence; The partial derivatives of historical measures are calculated within a preset time window, and the weights are generated by normalizing the ratio of the sum of the absolute values of the partial derivatives.

7. The method for dynamic prediction of carbon sinks in ecological restoration areas according to claim 6, characterized in that: The spatial synergy gain includes: According to the adjacency relationship of dynamic aggregation units, the contribution weight of similar measures in adjacent units is increased; The synergy coefficient is dynamically adjusted through iterative optimization to maximize the spatial synergy gain.

8. The method for dynamic prediction of carbon sinks in ecological restoration areas according to any one of claims 1 to 7, characterized in that: With the goal of maximizing carbon sink increments, a restoration decision plan is solved based on the carbon sink dynamic prediction results, including: Construct a multi-objective optimization model with constraints including budget limits, mutual exclusivity of measures, and biodiversity protection; The global optimization problem is grouped by ecological type through an adaptive decomposition algorithm, and restoration decision plans are solved in parallel.

Images