A forestry carbon sink prediction method based on big data

Abstract

The present application relates to the technical field of forestry carbon sink measurement, in particular to a forestry carbon sink prediction method based on big data, which collects multi-temporal remote sensing, meteorological, soil and forestry quadrat multi-source data, enhances the weak echo signal of understory vegetation through adaptive clutter suppression and coherent accumulation, and extracts the hierarchical vegetation structure characteristics; based on the time-temperature equivalent principle, the monthly carbon sink equivalent attenuation coefficient is calculated, and the carbon accumulation baseline rate is dynamically corrected; a multi-carbon pool coupling recursive framework is constructed to obtain the benchmark carbon sink prediction value; longitudinal federal transfer learning is used to complete generalization correction, and the contribution degree of the driving factor is quantified combined with the channel attention mechanism. The present application improves the estimation accuracy of understory carbon pool and the prediction stability under extreme climate, takes into account data privacy and result interpretability, and is suitable for forestry carbon sink measurement, certification and forest management.

Description

Technical Field

[0001] This invention relates to the field of forestry carbon sequestration technology, specifically a forestry carbon sequestration prediction method based on big data. Background Technology

[0002] Forestry carbon sequestration is the process by which forest ecosystems absorb atmospheric carbon dioxide through plant photosynthesis and fix it in vegetation and soil. It is a crucial pathway for addressing global climate change and achieving "dual carbon" goals, and is also a core type of emission reduction project in the carbon trading market. Accurate and reliable forestry carbon sequestration forecasting is the core foundation for carbon sequestration project verification, ecological compensation policy formulation, and forest management optimization. With the rapid development of remote sensing, big data, and artificial intelligence technologies, forestry carbon sequestration forecasting methods based on multi-source data are gradually replacing traditional manual quadrat surveys and becoming the mainstream technology in the industry.

[0003] In existing technologies, forestry carbon sink prediction mainly revolves around remote sensing data inversion and machine learning model construction, but there are still many technical shortcomings in practical applications.

[0004] First, the extraction accuracy of features related to forest carbon pools is insufficient. Existing remote sensing data preprocessing methods mostly focus on feature extraction of tree canopy, lacking effective enhancement methods for weak echo signals from understory shrubs, grasses, and litter. Conventional cloud removal and topographic correction often use interpolation or radiometric correction methods. Affected by topographic shading, canopy shading, and cloud scattering, the extraction error of understory vegetation structure features generally exceeds 25%. However, the shrub and grass layer, litter layer, and soil organic carbon pool account for 20%-35% of the total forest carbon sink. Feature extraction errors directly lead to systematic biases in carbon sink estimation results, making it difficult to meet the accuracy requirements of carbon sink verification.

[0005] Second, carbon sink predictions under climate stress suffer from poor stability. Existing models often employ static biomass conversion coefficients and fixed growth rates, incorporating meteorological factors as mere input features without adequately considering the lagged cumulative effects of extreme heat, drought, and other climate stresses on forest carbon sink functions. Given the frequent occurrence of extreme global climate events, forest carbon accumulation rates in anomalous climate years exhibit significant declines and lagged recovery. Static parameter models cannot quantify the long-term impact of stress duration and intensity on carbon sink functions, leading to prediction errors in extreme climate years exceeding 40% compared to normal years, and insufficient reliability and interannual comparability of the prediction results.

[0006] Third, multi-source data fusion is ineffective and has weak cross-regional generalization ability. Existing methods often use simple normalization to achieve multi-source heterogeneous data fusion, which easily loses the nonlinear coupling relationship between features and cannot fully explore the complementary value of multi-source data. At the same time, most forest areas in my country lack sufficient ground quadrat survey data, resulting in insufficient model training samples. Directly applying general models from other regions will lead to serious accuracy degradation. Traditional transfer learning methods require centralized sharing of raw data from various regions, which poses risks of privacy leakage and security of forestry resource data, making them difficult to implement.

[0007] Fourth, the predictive results lack interpretability and are insufficient to support carbon sequestration verification requirements. Most existing machine learning carbon sequestration prediction models are black-box models, which can only output the final carbon sequestration value and cannot quantify the contribution of different driving factors to the carbon sequestration. The prediction process is opaque and the results are not traceable, which fails to meet the core requirements of "measurable, reportable, and verifiable" in carbon trading and limits the application of the methods in formal carbon sequestration projects.

[0008] In addition, most existing technologies focus only on carbon sink measurement in the tree layer, and do not adequately consider the coupled flow relationship between carbon pools such as litter and soil organic carbon. They often use a simple summation method after independent estimation, ignoring carbon transfer and interaction between carbon pools, which further reduces the ecological authenticity and accuracy of the prediction results. Summary of the Invention

[0009] To address the problems in existing technologies, this invention provides a forestry carbon sink prediction method based on big data. This method solves the technical problems of low accuracy in extracting features of forest carbon pools, large prediction bias under climate stress, poor generalization in small sample areas, and insufficient interpretability of results. It achieves high-precision, high-generalization, and high-interpretability forestry carbon sink prediction, meeting the practical application needs of forestry carbon sink measurement, verification, and trading.

[0010] The technical solution adopted by this invention to solve its technical problem is: a forestry carbon sink prediction method based on big data, comprising the following steps: S1. Collect multi-source forestry big data of the target area, including multi-temporal remote sensing image data, meteorological monitoring data, soil property data and forestry quadrat survey data; S2. Spatial registration and vertical layering of multi-temporal remote sensing image data are performed. Adaptive threshold clutter suppression is used to filter out non-coherent noise. Then, phase alignment and coherent accumulation are performed on multi-temporal reflection signals of the same vertical layer to enhance the weak echo signal of forest vegetation and extract multi-dimensional vegetation structure features of tree layer, shrub layer and litter layer. S3. Based on the time-temperature equivalence principle, using the equivalent aging time as an intermediate variable, the duration of high temperature and drought stress is converted into the equivalent stress amount under the baseline conditions. The monthly carbon sink equivalent decay coefficient of the forest in the target area under climate stress is calculated, and the baseline rate of carbon accumulation of vegetation growth is dynamically corrected. S4. Construct a multi-carbon pool coupling recursive calculation framework that includes tree layer, shrub and grass layer, litter layer and soil organic carbon layer. Based on the law of conservation of matter, connect each carbon pool through carbon flux such as litter, decomposition and root input. Combine vegetation structure characteristics and equivalent attenuation coefficient, recursively calculate the carbon sink of each carbon pool with a monthly step size, and couple to obtain the baseline carbon sink prediction value. S5. Employing a feature alignment mechanism based on vertical federated transfer learning, the feature distribution of the source domain and the target domain is adapted without sharing the original data. The knowledge of the source domain model is transferred to the target domain to generalize and correct the baseline carbon sink prediction value. At the same time, the carbon sink contribution of each driving factor is quantified through the channel attention mechanism, and finally the forestry carbon sink prediction results of the target area are output.

[0011] Specifically, the adaptive clutter suppression process in step S2 is as follows: a neighborhood window is selected with the target pixel as the center, the mean and standard deviation of the signal within the window are calculated, and the noise threshold coefficient is dynamically set in combination with the terrain and coverage conditions. Incoherent clutter points that exceed the threshold are replaced with the mean of the neighborhood coherent signal. Coherent accumulation is performed on the 0-2 meter height layer under the forest to superimpose multi-temporal signals and improve the signal-to-noise ratio of weak echoes.

[0012] Specifically, the calculation process of the carbon sink equivalent decay coefficient in step S3 is as follows: first, the baseline rate of carbon accumulation in different forest stands is obtained by fitting the sample plot data based on historical years without extreme climate. Then, the equivalent aging time under high temperature stress is calculated by using the time-temperature equivalent equation in the form of Arrhenius. The drought stress effect is corrected by combining the relative humidity of the soil, and finally the monthly equivalent decay coefficient is obtained.

[0013] Specifically, the carbon flux relationship of the multi-carbon pool coupled recursive calculation framework in step S4 is as follows: the tree layer and the shrub and grass layer input carbon flux to the litter layer through littering. After the litter is decomposed by microorganisms, it is input into the soil organic carbon layer in proportion. The soil organic carbon layer simultaneously receives input from root exudates and outputs carbon flux through heterotrophic respiration. The carbon storage of each carbon pool is updated iteratively every month.

[0014] Specifically, the feature alignment mechanism of the vertical federated transfer learning in step S5 is as follows: the source domain and the target domain nodes each map their local features to a unified latent space, optimize the feature mapping parameters through adversarial training to make the feature distribution of the two domains tend to be consistent, and then combine a small number of samples from the target domain to fine-tune the regression layer parameters to complete the model transfer.

[0015] Specifically, the channel attention mechanism is embedded in the output layer of the feature mapping network. It outputs the attention weights of each feature channel through global average pooling and a fully connected network, and statistically obtains the carbon sink contribution ratios of three driving factors: vegetation structure, meteorology, and soil, by grouping by category.

[0016] Specifically, the multi-temporal remote sensing image data in step S1 includes Sentinel-1 SAR images, Sentinel-2 optical images, and LiDAR point cloud data, covering the entire growth cycle in time, and uniformly resampled to a spatial resolution of 30 meters.

[0017] Specifically, the meteorological monitoring data in step S1 includes monthly average temperature, monthly precipitation, number of extreme high-temperature days and number of consecutive days without precipitation, and the soil property data includes soil organic carbon content, bulk density, pH value and texture type.

[0018] Specifically, the forestry carbon sink prediction results output in step S5 include the annual total carbon sink in the target area, the stratified carbon sink in each carbon pool, the monthly carbon sink dynamic curve, and the contribution report of driving factors.

[0019] Specifically, the method also includes an accuracy verification step: comparing the predicted results with the measured carbon sink in ground sample plots, using the coefficient of determination and root mean square error as evaluation indicators, and iteratively optimizing the model parameters.

[0020] The beneficial effects of this invention are: First, the accuracy of understory vegetation feature extraction is significantly improved. Through adaptive clutter suppression and coherent accumulation processing, non-coherent noise is effectively filtered out and weak echo signals from the understory are enhanced. Compared with traditional interpolation correction methods, the feature extraction accuracy of understory shrubs and litter layers is improved, reducing the systematic estimation bias of the understory carbon pool from the data source and improving the accuracy of multi-carbon pool measurement.

[0021] Second, the prediction stability under extreme weather conditions is significantly enhanced. The attenuation correction model based on the time-temperature equivalence principle quantifies the cumulative lag effect of climate stress. Compared with the static parameter model, the carbon sink prediction error in extreme weather years is reduced, improving the interannual comparability and reliability of the prediction results.

[0022] Third, it has strong cross-regional generalization ability and ensures data security. It adopts a vertical federated transfer learning mechanism to achieve cross-regional transfer of model knowledge without sharing the original data, improves the prediction accuracy of small sample areas, and avoids the risk of privacy leakage of forestry resource data, making it suitable for practical application.

[0023] Fourth, the prediction results are highly interpretable and meet verification requirements. By quantifying the carbon sink contribution of various driving factors through the channel attention mechanism, the prediction process is traceable and the results are interpretable, which meets the core requirements of carbon sink trading: "measurable, reportable, and verifiable." It can be directly applied to the measurement and verification of formal forestry carbon sink projects.

[0024] Fifth, multi-carbon pool coupled calculations are more in line with ecological realities. The four-carbon pool dynamic coupling framework considers the flow of carbon between vegetation, litter, and soil. Compared with the method of summing up the independent estimates of each carbon pool, the overall accuracy of the total carbon sink prediction is improved, which is more in line with the real carbon cycle of forest ecosystems. Attached Figure Description

[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Figure 1 This is a flowchart illustrating the overall process of the forestry carbon sink prediction method based on big data as described in this invention. Figure 2 This is a schematic diagram of the adaptive clutter suppression and coherence accumulation remote sensing preprocessing workflow described in this invention; Figure 3 This is a schematic diagram of the carbon sink attenuation correction model based on the time-temperature equivalence principle described in this invention. Figure 4 This is a schematic diagram of the four-carbon library coupled recursive calculation framework described in this invention; Figure 5 This is a diagram of the vertical federated transfer learning and feature attention correction architecture described in this invention. Detailed Implementation

[0027] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0028] like Figures 1-5 As shown, the forestry carbon sink prediction method based on big data described in this invention includes five core steps: data acquisition, remote sensing data preprocessing and feature enhancement, dynamic correction of carbon sink attenuation, multi-carbon pool coupled recursive calculation, federated migration generalization correction, and result output. The specific implementation methods of each step are as follows: Multi-source forestry big data collection involves collecting four types of basic data from the target area to construct a forestry carbon sink big data dataset. First, multi-temporal remote sensing image data, including Sentinel-1 SAR images, Sentinel-2 optical images, and airborne / spaceborne LiDAR point cloud data, covering a complete growth cycle of more than 12 consecutive months. The spatial resolution is unified to a 30-meter scale through resampling and projected onto a unified coordinate system to extract the vertical structure and spectral features of vegetation. Second, meteorological monitoring data, including monthly observation data from national meteorological stations and regional automatic weather stations, with specific indicators including monthly average temperature, monthly precipitation, monthly extreme maximum temperature, number of extreme high temperature days, and number of consecutive days without precipitation, which are used for quantitative calculation of the degree of climate stress. Thirdly, soil property data, including indicators such as organic carbon content, bulk density, pH value, and texture type of the top 0-30cm and 30-60cm soil profiles, are used to obtain a raster dataset that matches the remote sensing data through spatial interpolation. Fourthly, forestry quadrat survey data, including ground-based measured data such as tree diameter at breast height, tree height, stand density, shrub and grass biomass, litter dry weight, and soil organic carbon within fixed quadrats, are used for model training, parameter calibration, and accuracy verification.

[0029] Adaptive clutter suppression and coherent accumulation for remote sensing feature enhancement address the challenge of accurately extracting weak echo signals from forest understory. This involves denoising and feature enhancement of multi-temporal remote sensing data. The specific process is as follows: First, the multi-temporal SAR images and LiDAR point cloud data are accurately spatially registered, and the UTM projection coordinate system is uniformly adopted. The data is then resampled to a spatial resolution of 30m. The vertical structure is divided into layers at 1m intervals along the vertical direction. The radar backscatter signal and LiDAR point cloud quantity of each grid pixel in each vertical layer are extracted to form a multi-temporal vertical signal sequence.

[0030] Adaptive clutter suppression employs local adaptive threshold filtering: a 7×7 neighborhood window is selected centered on the target pixel, and the mean signal within the window is calculated. with standard deviation Set the noise threshold as follows: ,in, For adaptive threshold coefficients, a value of 2.0 is used for flat terrain, 2.3 for mountainous terrain with a slope greater than 25°, and 2.5 for snow-covered areas. Pixels with signal strength exceeding the threshold T within the window are identified as incoherent clutter points and replaced with the mean of the coherent signals in the neighborhood to filter out random noise caused by terrain shadows, cloud scattering, and snow reflection.

[0031] Coherent accumulation processing is applied to multi-temporal signals within the same vertical layer: First, using the first phase data as a baseline, phase offset correction is performed on the signals from the remaining phases to eliminate phase deviations caused by differences in satellite orbits; then... The phase-aligned coherent signals are then amplitude-superimposed, resulting in a signal-to-noise ratio improvement of [number] times. The study focuses on enhancing the signals of shrubs, grasses, and litter in the 0-2m height layer under the forest canopy. Finally, three types of structural features are extracted from each vertical layer: point cloud density, average height, and vegetation coverage. Combined with spectral features such as leaf area index and enhanced vegetation index extracted from optical images, a multi-dimensional vegetation structural feature set covering the tree layer, shrub layer, and litter layer is formed.

[0032] Based on the time-temperature equivalence principle, a dynamic carbon sequestration correction model is constructed to address the problem of large prediction biases in carbon sequestration under climate stress. The specific process is as follows: First, determine the baseline carbon accumulation rate: Screen quadrat observation data from at least three consecutive years without extreme weather events, group them by stand type, age group, and site grade, and fit the data to obtain the monthly average carbon accumulation rate of each group under suitable temperature and humidity conditions, which will serve as the baseline rate. The unit is tC·ha -1 ·month -1 .

[0033] Introducing the time-temperature equivalence principle, using equivalent aging time As an intermediate variable, the actual stress duration t under different stress intensities is converted into the equivalent aging time at the reference temperature. The conversion relationship follows the Arrhenius equation: ,in: The apparent activation energy for carbon sequestration is 65 kJ / mol. The value of is the ideal gas constant, which is 8.314 J / (mol·K). The reference thermodynamic temperature is 298K (corresponding to 25℃). The actual monthly average thermodynamic temperature, in K.

[0034] When the actual temperature is higher than the reference temperature, the equivalent aging time is longer than the actual time, and the carbon sink decay effect is amplified.

[0035] To address drought stress, a soil relative humidity correction factor is introduced: when soil relative humidity... When the water content is below 60% of field capacity, the drought stress coefficient is calculated using the following formula: That is, for every 10 percentage point decrease in humidity, the equivalent aging time increases by 8%; when hour, There is no drought stress attenuation. The final equivalent aging time used in the calculation is the high-temperature equivalent aging time multiplied by the drought stress coefficient.

[0036] The final monthly carbon sequestration equivalent decay coefficient is calculated using the following formula: ,in, The monthly carbon sequestration equivalent decay coefficient ranges from 0 to 1, with a smaller value indicating a more significant decay. The actual carbon accumulation rate under stress, in tC·ha. -1 ·month -1 ; The aging time constant is taken as the baseline for carbon sequestration function, and is set at 120~240 days based on the stand's stress resistance.

[0037] The calculated monthly equivalent decay coefficient is used to dynamically correct the carbon accumulation rate in subsequent multi-carbon pool calculations.

[0038] A multi-carbon pool coupled recursive calculation framework is constructed, comprising a four-carbon pool coupled recursive calculation framework including a tree layer, a shrub and grass layer, a litter layer, and a soil organic carbon layer. Based on the law of conservation of mass, dynamic coupled calculations of each carbon pool are achieved. The specific process is as follows: The forest ecosystem was divided into four carbon pool subsystems, with a time step of one month. =1 month, the general formula for recursively estimating the carbon storage of each carbon pool is: ,in, Let represent the carbon storage of the i-th carbon pool at time t, in tC. This represents the net carbon change of the i-th carbon pool in the current month, in tC.

[0039] The specific calculation methods for net carbon change and carbon flux in each carbon pool are as follows: (1) For the tree layer, the net carbon accumulation and the flux discharged to the litter layer in the current month are respectively: ,in, Baseline carbon accumulation rate of the tree layer, unit: tC·ha -1 ·month -1 ; The corresponding forest stand area is expressed in ha. The monthly carbon flux from tree litter is expressed in tC / month. The monthly shedding coefficient for trees is 0.008~0.012 for evergreen species and 0.015~0.025 for deciduous species.

[0040] (2) For the shrub and grass layer, the net carbon accumulation and the fallout flux to the litter layer in the current month are respectively: ,in, The baseline carbon accumulation rate of the shrub and grass layer is obtained by fitting the understory vegetation structure characteristics, in units of tC·ha. -1 ·month -1 ; The monthly carbon flux from the shrubland layer is expressed in tC / month. The monthly shedding coefficient for shrubs is taken as 0.02~0.03.

[0041] (3) Litter layer, the input flux is the sum of the litter volume of trees and shrubs. The output flux is the litter decomposition flux, calculated using the following formula: ,in, Monthly carbon flux from litter decomposition, unit: tC / month; The carbon storage in the litter layer for the current month, in tC; The baseline decomposition rate of litter is 0.003–0.005 months for temperate forests. -1 Subtropical forests were sampled at 0.006~0.008 months. -1 ; The temperature correction function for the decomposition rate is Q. 10 Model form; The decomposition rate is a moisture correction function, exhibiting a parabolic variation with soil relative humidity. Carbon produced from litter decomposition is input into the soil organic carbon pool and emitted into the atmosphere in a 7:3 ratio; that is, the input flux into the soil organic carbon pool is: .

[0042] (4) Soil organic carbon layer, input fluxes include litter decomposition inputs. With root exudate input The amount of root exudates input accounts for 18% of the net primary productivity of trees, that is: The output flux is the soil heterotrophic respiration carbon flux, calculated using the following formula: ,in, Monthly carbon flux output from soil heterotrophic respiration, unit: tC / month; The amount of soil organic carbon in the current month is expressed in tC. Reference rate for soil heterotrophic respiration, unit: month -1 .

[0043] The monthly net carbon change in the soil organic carbon layer is as follows: After initializing the initial carbon storage of each carbon pool, the four carbon pools are dynamically coupled and recursively calculated through the above carbon flux relationship. The annual carbon sink of each carbon pool is calculated iteratively month by month, and the sum is used to obtain the predicted value of the regional baseline total carbon sink.

[0044] Vertical federated transfer learning and feature attention correction address the issues of poor generalization and insufficient interpretability in small sample regions. Vertical federated transfer learning is used to achieve cross-regional model generalization, while an attention mechanism is employed to output factor contributions. The specific process is as follows: A vertical federated transfer learning framework is adopted, which includes source domain data nodes and target domain data nodes. The two nodes only exchange encrypted feature mapping results and do not transmit the original sample data. The specific steps are as follows: (1) Unified feature space: Both parties pre-agree on a unified multi-carbon pool feature dimension, including 3 categories of 12-16 dimensions of vegetation structure features, meteorological features, and soil features, and each party completes feature extraction and standardization locally.

[0045] (2) Feature mapping alignment: Both parties construct their own feature mapping networks to map the local original features to a unified latent feature space; the parameters of the mapping network are optimized through gradient backpropagation and adversarial training. The domain discriminator aims to distinguish the latent features of the source domain from the target domain, while the feature mapping network aims to confuse the domain discriminator. The two parties conduct adversarial iterations to eventually make the latent feature distributions of the source domain and the target domain more consistent, thus completing feature alignment.

[0046] (3) Model transfer and correction: The parameters of the carbon sink prediction regression layer trained in the source domain are fine-tuned by combining a small number of labeled samples in the target domain to obtain the final prediction model of the target domain, and the benchmark carbon sink prediction value is generalized and corrected.

[0047] The feature attention mechanism is embedded in the output layer of the feature mapping network, employing a channel attention structure: global average pooling is performed on the input features across all dimensions, followed by two fully connected layers and a sigmoid activation function to output the attention weights for each feature channel. The weight range is 0~1, and satisfies The total contribution percentages of the three driving factors—vegetation structure, meteorology, and soil—are obtained by grouping and summing according to feature categories and output synchronously with the prediction results.

[0048] The final output includes the annual total carbon sink of the target region, the stratified carbon sink of each carbon pool, the monthly dynamic curve of carbon sink, and a report on the contribution of driving factors.

[0049] Example 1: Forestry Carbon Sequestration Prediction in Planted Chinese Cunninghamia lanceolata Forests in Southern China: This example focuses on a concentrated distribution area of ​​planted Chinese Cunninghamia lanceolata forests in a province in southern my country. The total area is approximately 500 square kilometers, dominated by middle-aged planted Chinese Cunninghamia lanceolata with an average age of 18 years. The understory includes a small amount of broadleaf shrubs, grasses, and litter. The terrain is primarily low hills with an average slope of 15°. Thirty-six fixed ground monitoring plots were established within the area for model parameter calibration and accuracy verification. This example uses the method described in this invention to predict forestry carbon sequestration for 2023. The specific implementation process is as follows: The first step was multi-source forestry data collection. Twelve Sentinel-1 SAR images, 24 Sentinel-2 optical images, and one airborne LiDAR point cloud data covering the entire study area from January to December 2023 were collected. All remote sensing data were unified to a 30-meter spatial resolution through resampling and projected onto the WGS84 / UTM50N coordinate system. Monthly observation data from three national meteorological stations and twelve regional automatic weather stations were collected, extracting four meteorological indicators: monthly average temperature, monthly precipitation, number of days with a daily maximum temperature ≥35℃, and number of consecutive days without precipitation. Among these, data from July 2023 onwards... August saw 22 consecutive days of high temperatures and drought, with an extreme maximum temperature of 39℃. Data on organic carbon content, bulk density, and texture of the 0-30cm soil layer were collected from 100 soil profiles in the region, and a 30-meter resolution soil property raster dataset was obtained through Kriging interpolation. Data from 36 fixed quadrats were collected at the end of 2023, including tree diameter at breast height and height, shrub and grass biomass, litter dry weight, and soil organic carbon content. Data from 24 quadrats were used for parameter calibration, and data from 12 quadrats were used for independent accuracy verification.

[0050] The second step is remote sensing data preprocessing and feature extraction. First, precise spatial registration is performed between SAR imagery and LiDAR point clouds, with the registration error controlled within one pixel. The area is divided into 20 structural layers at 1-meter intervals along the vertical direction. The radar backscatter signal and point cloud quantity sequence corresponding to each layer are extracted for each pixel. An adaptive threshold clutter suppression algorithm is employed. Considering the region's average slope of 15°, a threshold coefficient k=2.0 is set. The noise threshold is calculated using the signal mean and standard deviation of a 7×7 neighborhood window surrounding each pixel, filtering out incoherent clutter noise caused by mountain shadows and cloud scattering, achieving a noise removal rate of 89%. Phase alignment and coherent accumulation are performed on 12 SAR signals from three core layers: 0-2m, 2-10m, and above 10m. The 0-2m layer corresponds to understory shrubs and litter; after coherent accumulation, the signal-to-noise ratio is improved by 17dB, effectively enhancing the identifiability of weak signals under the forest canopy. Finally, three types of structural features—point cloud density, average height, and coverage—are extracted from each layer. These are combined with two types of spectral features—leaf area index and enhanced vegetation index—calculated from optical images, to form a 12-dimensional multi-dimensional vegetation structure feature set.

[0051] The third step is to calculate the carbon sink depletion coefficient. Based on sample plot data from 2018-2021 without extreme weather events, the baseline carbon accumulation rate for the region is obtained by fitting data according to the middle-aged Chinese fir forest group. 0.85 tons of carbon per hectare per month; activating energy for carbon sequestration function. reference temperature Reference aging time constant The drought stress correction coefficient is set to increase the equivalent time by 8% for every 10% decrease in humidity. By inputting monthly meteorological and soil moisture data for 2023, the carbon sink attenuation coefficient for the high-temperature and drought period in July and August of that year was calculated to be 0.72 through the time-temperature equivalent equation. The attenuation coefficients for the other months are distributed between 0.92 and 0.98, forming a complete monthly attenuation coefficient sequence.

[0052] The fourth step is the coupled recursive calculation of multiple carbon pools. Using January 2023 as the initial time step, the initial carbon storage of each carbon pool is initialized: the initial carbon storage of the tree layer is calculated based on the biomass data of the quadrat at the end of 2022, averaging 72 tons of carbon per hectare; the initial carbon storage of the shrub and grass layer is obtained by fitting the characteristics of the understory vegetation, averaging 2.8 tons of carbon per hectare; the initial carbon storage of the litter layer is calculated by multiplying the dry weight of the litter at the end of the previous year by 0.5 (carbon content), averaging 4.2 tons of carbon per hectare; the initial carbon storage of soil organic carbon is calculated based on soil profile data, with an average of 85 tons of carbon per hectare in the 0-30cm layer. A monthly litter coefficient for Chinese fir is set. Monthly withering coefficient of shrubs Litter baseline decomposition rate The calculations are performed recursively in monthly increments: Monthly carbon accumulation in the tree layer is obtained by multiplying the baseline rate by the monthly decay coefficient and then by the corresponding stand area; monthly carbon accumulation in the shrub and grass layer is calculated by multiplying the monthly growth rate of shrub and grass biomass by the decay coefficient; monthly input flux in the litter layer is the monthly litter volume of trees and shrubs, and output flux is the litter decomposition volume, calculated using a first-order kinetic model corrected for temperature and moisture; input flux in the soil organic carbon layer includes litter decomposition carbon and root exudate carbon, where root exudate accounts for 18% of the net primary productivity of trees, and output flux is soil heterotrophic respiration consumption, which is positively correlated with soil temperature. After 12 months of recursive calculation, the annual carbon sink of each carbon pool is obtained, and these are summed to obtain the predicted regional baseline total carbon sink.

[0053] Step 5: Federated Migration Correction and Result Output. A Chinese fir plantation area with sufficient samples from neighboring provinces was selected as the source domain. Without exchanging original data, a feature mapping network was constructed between the source and target domains based on a unified 12-dimensional feature space. Through 50 rounds of adversarial training, the feature distributions of the source and target domains were adapted. The feature extraction layer parameters of the source domain model were transferred to the target domain model. The regression layer parameters were fine-tuned using 24 local calibration plots to generalize and correct the baseline carbon sink prediction. Simultaneously, the contribution of each driving factor was calculated using a channel attention mechanism: vegetation structure factor accounted for 62%, meteorological factor for 23%, and soil factor for 15%. The final output for this region in 2023 was 8.2 tons of carbon sink per hectare, with a total annual carbon sink of 410,000 tons.

[0054] Accuracy was verified using measured data from 12 independent validation quadrats. The coefficient of determination R of the prediction results in this embodiment was calculated.2 The mean square error (RMSE) was 0.91, and the root mean square error (RMSE) was 0.68 tons per hectare.

[0055] To demonstrate the technical effectiveness of each step, three control examples were set up for comparison: Control Example 1-1: Carbon sink prediction was performed using a traditional random forest model, with only optical image features and meteorological data input, without forest understory signal enhancement and multi-carbon pool coupling calculations. The verification result R... 2 The value was 0.73, and the RMSE was 1.32 tons per hectare. Comparative Examples 1-2: The adaptive clutter suppression and coherent accumulation steps were omitted; the traditional cloud removal interpolation method was used to process the remote sensing data. The remaining steps were consistent with this embodiment. The verification result R... 2 The value was 0.82, and the RMSE was 0.97 tons per hectare. Comparative Examples 1-3: The time-temperature equivalent decay correction step was omitted, a fixed carbon accumulation rate was used, and the remaining steps were consistent with this example. The verification result R... 2 The value was 0.84, and the RMSE was 0.91 tons per hectare.

[0056] Example 2: Forestry Carbon Sequestration Prediction in Natural Broadleaf Secondary Forest Areas in Northern China: This example uses a natural broadleaf secondary forest in a forest area of ​​Northeast my country as the research object. The total area is approximately 800 square kilometers, dominated by Mongolian oak and birch, with an average forest age of 35 years. The understory litter layer is relatively thick, and the soil organic carbon storage is high. This area experienced severe spring drought and extreme summer heat events in 2023, exhibiting significant climate stress characteristics. There was a four-month snow cover period in winter. Forty-two fixed ground plots were established in the area, of which 28 were used for parameter calibration and 14 for independent validation. This example uses the method described in this invention to predict forestry carbon sequestration in 2023. The specific implementation process is as follows: The first step was multi-source data acquisition. Twelve Sentinel-1 SAR images and twenty-two Sentinel-2 optical images from January to December 2023, along with ground LiDAR point cloud data covering the sampling points, were collected and unified to a spatial resolution of 30 meters through resampling. Monthly observation data from five national meteorological stations in the region were collected, including indicators such as monthly average temperature, monthly precipitation, number of consecutive days without precipitation, and number of extreme low-temperature days. Specifically, there were 42 consecutive days without precipitation in April and May 2023, the extreme maximum temperature in July reached 38℃, and November to March of the following year was the snow cover period. Data on organic carbon, bulk density, and pH values ​​in the 0-30cm and 30-60cm soil layers from 120 soil profiles were collected, and interpolation yielded a global soil property raster. Finally, measured data from 42 fixed sample plots at the end of 2023 were collected, and training / calibration and validation sets were separated.

[0057] The second step is remote sensing data preprocessing. Spatial registration of multi-temporal SAR and LiDAR data is performed, dividing the data into 25 structural layers at 1m intervals along the vertical direction. To address signal clutter caused by winter snow cover in northern forest areas, an adaptive dynamic threshold clutter suppression algorithm is employed. The noise threshold coefficient is set to 2.5 during the snow-covered period and 2.0 during the non-snow-covered period, effectively filtering out incoherent noise caused by snow reflection and terrain shadows, achieving an 85% noise removal rate during the winter snow-covered period. Twelve coherent accumulation phases are performed on the signals from forest litter and shrub / grass layers at a depth of 0-3m, improving the signal-to-noise ratio by 15dB. Point cloud density, canopy height, and coverage features of each vertical layer are extracted, and combined with the normalized vegetation index and leaf area index extracted from optical imagery, a 14-dimensional vegetation structure feature set is constructed.

[0058] The third step is to calculate the carbon sink attenuation coefficient. Based on sample plot data from 2017-2020 without extreme weather events, the baseline carbon accumulation rate of natural broadleaf forests in the region was fitted to be 0.72 tons of carbon per hectare per month. A low-temperature stress conversion coefficient was added to the time-temperature equivalent model; when the temperature is below -10℃, the equivalent aging time increases by 5% for every 2℃ decrease in temperature. A baseline aging time constant was also set. It is well-suited to the strong resilience of natural forests. Using monthly meteorological data from 2023, the attenuation coefficients were calculated to be 0.75 for the dry season in April-May, 0.78 for the high-temperature season in July, and 0.88 for the low-temperature season in winter. The attenuation coefficients for the remaining months ranged from 0.93 to 0.97.

[0059] The fourth step involves coupled recursive calculations of multiple carbon pools. The carbon storage of each pool is initialized with a step size of January. The initial carbon storage of the litter layer is 7.5 tons per hectare, and the initial carbon storage of the soil organic carbon layer is 128 tons per hectare, significantly higher than that of plantations in southern China. A monthly litter coefficient for deciduous broad-leaved trees is also set. Monthly withering coefficient of shrubs 2. Baseline decomposition rate of litter The carbon sequestration rate is adapted to that of temperate forests. Calculations are performed recursively in monthly increments. Carbon accumulation in the tree layer is combined with a broadleaf tree diameter-at-breast growth model and a decay coefficient; carbon accumulation in the shrub and grass layer is fitted based on understory characteristics; litter decomposition uses first-order kinetic parameters of temperate forests; and soil organic carbon pools consider the inhibitory effect of winter freeze-thaw cycles on soil respiration, with soil respiration rates reduced by a factor of 0.3 during low-temperature periods. Each carbon pool is dynamically coupled through fluxes such as litter, decomposition, and root input, yielding a baseline carbon sink prediction after 12 months of recursion.

[0060] Step 5: Federated Transfer Correction and Result Output. A natural broad-leaved forest region with sufficient samples from the same climate zone was selected as the source domain. Longitudinal federated transfer learning was used for feature alignment and model parameter transfer, completing model generalization correction without sharing the original data. The contribution of each factor was obtained through a channel attention mechanism: vegetation structure factor accounted for 57%, meteorological factor for 29%, and soil factor for 14%. The final output for this region in 2023 was an annual carbon sink of 6.9 tons per hectare, with a total annual carbon sink of 552,000 tons.

[0061] The predicted result R in this embodiment was verified using measured data from 14 independent validation quadrats. 2 The value was 0.89, and the RMSE was 0.62 tons per hectare.

[0062] Three control groups were set up for comparison: Control 2-1: The traditional biomass conversion factor method was used, only calculating the carbon sink in the tree layer and ignoring the carbon pool in the understory and soil, to verify the result R. 2 The value was 0.65, and the RMSE was 1.58 tons per hectare. Comparative Example 2-2: A standard LSTM model was used for carbon sink prediction, inputting meteorological and remote sensing features without climate stress attenuation correction. The results were validated. 2 The value was 0.71, and the RMSE was 1.36 tons per hectare. Comparative Examples 2-3: The federal migration correction step was omitted, and the model was trained using only local samples. The remaining steps were consistent with this embodiment, and the verification result R... 2 The value was 0.78, and the RMSE was 1.05 tons per hectare.

[0063] Example 3: Forestry Carbon Sequestration Prediction in a Mixed Coniferous and Broadleaf Forest Area in Southwest China: This example uses a mixed coniferous and broadleaf forest area in a mountainous region of Southwest my country as the research object. The total area is approximately 500 square kilometers, with significant topographic relief, an elevation difference of up to 1200 meters, severe topographic obstruction, and complex forest stand types, including multiple tree species such as Yunnan pine and alpine oak. Only 18 fixed ground plots were set up, representing a typical small-sample complex terrain area, used to verify the cross-regional generalization ability and adaptability to complex terrain of the present invention. This example uses the method described in the present invention to conduct forestry carbon sequestration prediction for 2023. The specific implementation process is as follows: The first step was multi-source data acquisition. Twelve Sentinel-1 SAR images, twenty Sentinel-2 optical images, and spaceborne LiDAR data from January to December 2023 were collected and uniformly resampled to a 30-meter resolution. Simultaneously, 30-meter resolution digital elevation model data was collected for terrain correction. Monthly meteorological data from two national meteorological stations and six regional automatic weather stations within the region were collected, and temperature reduction correction was performed using altitude gradients, with a reduction rate of 0.6℃ for every 100 meters of elevation gain. Data from 60 soil profiles were collected, and soil property raster was obtained by interpolation using topographic factors such as slope and aspect. Measured data from 18 fixed sample plots were collected; all data were used for independent validation and were not included in local model parameter calibration.

[0064] The second step is remote sensing data preprocessing. First, topographic radiometric correction is performed using DEM data. Then, spatial registration is performed on SAR and LiDAR data, dividing the data into 20 structural layers along the vertical direction. To address the large-area shadow clutter caused by mountainous terrain, an adaptive clutter suppression algorithm incorporating topographic factors is employed. The noise threshold coefficient for areas with slopes greater than 30° is increased to 2.3, effectively filtering out incoherent noise caused by topographic shadows, achieving a noise removal rate of 82% in steep slope areas. Coherent accumulation of multi-temporal signals from the 0-2 meter layer under the forest canopy is performed to compensate for signal attenuation caused by topographic shading, improving the signal-to-noise ratio by 14 dB. Vegetation structure features of each layer are extracted, and a 16-dimensional feature dataset is constructed by combining three types of topographic factors: altitude, slope, and aspect.

[0065] The third step is to calculate the carbon sink attenuation coefficient. A sufficient sample area with similar climate and forest stand type was selected as a reference, and the baseline carbon accumulation rate for mountain mixed coniferous and broad-leaved forests was fitted to be 0.78 tons of carbon per hectare per month. Differential baseline rates and stress conversion coefficients were set based on altitude gradients. For every 100 meters increase in altitude, the baseline rate decreased by 2%, stress resistance increased, and the aging time constant increased by 5%. Monthly meteorological data was input to calculate the monthly attenuation coefficients for different altitude gradients. The lowest attenuation coefficient was found in the low-altitude arid valley region (0.73), while the attenuation coefficients in the high-altitude region were mostly above 0.90.

[0066] The fourth step involves multi-carbon pool coupling and recursive calculation. The system is divided into three calculation units (low, medium, and high) based on altitude gradients. Parameters for each carbon pool are initialized separately. The low-altitude region is dominated by coniferous trees, while the proportion of broadleaf trees increases in the medium and high-altitude regions. The litter coefficient and decomposition rate for coniferous and broadleaf trees are set separately, and the parameters for mixed forests are calculated by weighting according to forest stand composition. The calculation is recursively performed on a monthly basis. For the tree layer, biomass conversion parameters for coniferous and broadleaf trees are used to weight and calculate the carbon sink of the mixed forest tree layer. The litter decomposition uses the weighted average decomposition coefficient of the mixed forest. The soil organic carbon pool considers the correction effect of mountain soil erosion and deposition on carbon flux, and soil carbon sink is reduced by a coefficient of 0.9 in areas with a slope greater than 25°. After recursion for each unit, the baseline carbon sink prediction value for the entire region is obtained by area weighting.

[0067] Step 5: Federated Transfer Correction and Result Output. Two regions with sufficient samples and similar climate and forest stands are selected as multi-source domains. Longitudinal federated transfer learning is used to align the features of the multi-source domains, transferring the model knowledge from the two source domains to the target region and generalizing the baseline carbon sink prediction values. Since there are no training samples in the target domain, only the source domain transfer model is used for prediction, without the need for local fine-tuning. The contribution of each factor is obtained through a channel attention mechanism: vegetation structure accounts for 54%, topography and meteorological factors combined account for 32%, and soil factors account for 14%. The final output for this region is an annual carbon sink per unit area of ​​7.5 tons per hectare in 2023, with a total annual carbon sink of 375,000 tons for the entire region.

[0068] The predicted result R in this embodiment was verified using measured data from 18 independent validation quadrats. 2 The value was 0.87, and the RMSE was 0.71 tons per hectare.

[0069] Three control examples were set up for comparison: Control example 3-1: The general carbon sink model for adjacent regions was directly applied without migration correction, and the verification result R was obtained. 2 The R² value was 0.52, and the RMSE was 1.87 tons per hectare. In contrast, Example 3-2 uses a traditional centralized transfer learning method, sharing the original source domain data for model transfer, and verifies the R² result. 2 The value was 0.76, and the RMSE was 1.12 tons per hectare. Comparative Example 3-3: The multi-carbon pool coupling step was omitted; only the tree layer carbon sink was calculated. The remaining steps were consistent with this embodiment. The verification result R... 2 The value was 0.72, and the RMSE was 1.24 tons per hectare.

[0070] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A forestry carbon sink prediction method based on big data, characterized in that, Includes the following steps: S1. Collect multi-source forestry big data of the target area, including multi-temporal remote sensing image data, meteorological monitoring data, soil property data and forestry quadrat survey data; S2. Spatial registration and vertical layering of multi-temporal remote sensing image data are performed. Adaptive threshold clutter suppression is used to filter out non-coherent noise. Then, phase alignment and coherent accumulation are performed on multi-temporal reflection signals of the same vertical layer to enhance the weak echo signal of forest vegetation and extract multi-dimensional vegetation structure features of tree layer, shrub layer and litter layer. S3. Based on the time-temperature equivalence principle, using the equivalent aging time as an intermediate variable, the duration of high temperature and drought stress is converted into the equivalent stress amount under the baseline conditions. The monthly carbon sink equivalent decay coefficient of the forest in the target area under climate stress is calculated, and the baseline rate of carbon accumulation of vegetation growth is dynamically corrected. S4. Construct a multi-carbon pool coupling recursive calculation framework including tree layer, shrub and grass layer, litter layer and soil organic carbon layer. Based on the law of conservation of matter, connect each carbon pool through carbon flux such as litter, decomposition and root input. Combine vegetation structure characteristics and equivalent attenuation coefficient, recursively calculate the carbon sink of each carbon pool with a monthly step size, and couple to obtain the baseline carbon sink prediction value. S5. Employing a feature alignment mechanism based on vertical federated transfer learning, the feature distribution of the source domain and the target domain is adapted without sharing the original data. The knowledge of the source domain model is transferred to the target domain to generalize and correct the baseline carbon sink prediction value. At the same time, the carbon sink contribution of each driving factor is quantified through the channel attention mechanism, and finally the forestry carbon sink prediction results of the target area are output.

2. The forestry carbon sink prediction method based on big data according to claim 1, characterized in that: The specific process of adaptive clutter suppression in step S2 is as follows: a neighborhood window is selected with the target pixel as the center, the mean and standard deviation of the signal within the window are calculated, and the noise threshold coefficient is dynamically set in combination with the terrain and coverage conditions. Incoherent clutter points that exceed the threshold are replaced with the mean of the neighborhood coherent signal. Coherent accumulation is performed to superimpose multi-temporal signals for the 0-2 meter height layer under the forest to improve the signal-to-noise ratio of weak echoes.

3. The forestry carbon sink prediction method based on big data according to claim 1, characterized in that: The calculation process of the carbon sink equivalent decay coefficient in step S3 is as follows: First, the baseline rate of carbon accumulation in different forest stands is obtained by fitting the sample plot data based on historical years without extreme climate. Then, the equivalent aging time under high temperature stress is calculated by using the time-temperature equivalent equation in the form of Arrhenius. The drought stress effect is corrected by combining the relative humidity of the soil, and finally the monthly equivalent decay coefficient is obtained.

4. The forestry carbon sink prediction method based on big data according to claim 1, characterized in that: The carbon flux relationship in the multi-carbon pool coupling recursive calculation framework in step S4 is as follows: the tree layer and the shrub and grass layer input carbon flux to the litter layer through littering. After the litter is decomposed by microorganisms, it is input into the soil organic carbon layer in proportion. The soil organic carbon layer simultaneously receives input from root exudates and outputs carbon flux through heterotrophic respiration. The carbon storage of each carbon pool is updated iteratively every month.

5. The forestry carbon sink prediction method based on big data according to claim 1, characterized in that: The feature alignment mechanism in step S5 of the vertical federated transfer learning is as follows: the source domain and the target domain nodes each map their local features to a unified latent space, optimize the feature mapping parameters through adversarial training to make the feature distribution of the two domains tend to be consistent, and then combine a small number of samples from the target domain to fine-tune the regression layer parameters to complete the model transfer.

6. The forestry carbon sink prediction method based on big data according to claim 5, characterized in that: The channel attention mechanism is embedded in the output layer of the feature mapping network. It outputs the attention weights of each feature channel through global average pooling and a fully connected network. The carbon sink contribution ratios of three driving factors—vegetation structure, meteorology, and soil—are obtained by grouping and statistically analyzing them by category.

7. The forestry carbon sink prediction method based on big data according to claim 1, characterized in that: The multi-temporal remote sensing image data in step S1 includes Sentinel-1 SAR images, Sentinel-2 optical images, and LiDAR point cloud data, covering the entire growth cycle in time, and uniformly resampled to a spatial resolution of 30 meters.

8. The forestry carbon sink prediction method based on big data according to claim 1, characterized in that: The meteorological monitoring data in step S1 includes monthly average temperature, monthly precipitation, number of extreme high-temperature days and number of consecutive days without precipitation, and the soil property data includes soil organic carbon content, bulk density, pH value and texture type.

9. The forestry carbon sink prediction method based on big data according to claim 1, characterized in that: The forestry carbon sink prediction results output in step S5 include the annual total carbon sink in the target area, the stratified carbon sink in each carbon pool, the monthly carbon sink dynamic curve, and the contribution report of driving factors.

10. The forestry carbon sink prediction method based on big data according to claim 1, characterized in that: The method also includes an accuracy verification step: comparing the predicted results with the measured carbon sink in ground sample plots, using the coefficient of determination and root mean square error as evaluation indicators, and iteratively optimizing the model parameters.

Images