Soil carbon dynamic prediction method and system

Abstract

The invention discloses a soil carbon dynamic prediction method and system, and relates to the technical field of soil carbon reserve prediction. According to the method, after covariable data including plant data, soil data and climate data in a to-be-predicted area are obtained, data, of which the correlation with the dynamic change of soil carbon in the to-be-predicted area exceeds a preset standard, in the covariable data are extracted from the covariable data to serve as soil carbon dynamic influence factors; inputting other data except the soil carbon dynamic influence factor in the covariable data into a CENTURY model to obtain a first soil carbon dynamic prediction result of the to-be-predicted area; and finally, inputting the soil carbon dynamic influence factor and the first soil carbon dynamic prediction result into a trained converter model to obtain a soil carbon dynamic optimization prediction result of the to-be-predicted region. Therefore, the problem that the CENTURY model cannot dynamically and accurately simulate and predict the energy plant soil carbon is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of soil carbon dynamics prediction, and more specifically, to a method and system for predicting soil carbon dynamics. Background Art

[0002] The soil carbon pool accounts for more than 90% of the terrestrial ecosystem carbon pool, which is 3-4 times that of the vegetation carbon pool and 2-3 times that of the atmospheric carbon pool. The global soil organic carbon sequestration potential is 2.3 billion - 5.5 billion tons of carbon dioxide equivalent per year. Therefore, giving full play to the soil carbon sink capacity has great benefits for emission reduction. As energy plants used as green fuel raw materials, they will cause changes in the soil carbon pool during their planting stage and enhance the soil carbon sequestration ability, so they have great potential to mitigate climate change.

[0003] Currently, process simulation methods are usually used to estimate the change rate of the soil carbon pool. It constructs a mechanism model based on influencing factors such as climate, soil physical and chemical properties, and agroforestry management measures to estimate the change rate of the soil carbon pool. However, as a traditional soil carbon dynamics model, CENTURY4.0 has the problem of inaccurate simulation. For example, when using the CENTURY4.0 model to simulate and predict the soil carbon dynamics of sugarcane planting in Brazil, it is found that the results are different under different soil management modes. It can be seen that the existing methods cannot accurately simulate and predict the soil carbon dynamics of energy plants. Therefore, there is an urgent need for a soil carbon dynamics prediction method with high prediction accuracy and strong generalization ability. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for predicting soil carbon dynamics to solve the problem that the CENTURY model cannot accurately simulate and predict the soil carbon dynamics of energy plants.

[0005] In the first aspect, the present invention provides a method for predicting soil carbon dynamics, including:

[0006] Obtaining covariate data of the area to be predicted, where the covariate data includes: plant data, soil data, and climate data within the area to be predicted;

[0007] Extracting soil carbon dynamics influencing factors from the covariate data, where the soil carbon dynamics influencing factors are data in the covariate data whose correlation with the soil carbon dynamics change in the area to be predicted exceeds a preset standard;

[0008] Inputting the remaining data in the covariate data except the soil carbon dynamics influencing factors into the CENTURY model to obtain a first soil carbon dynamics prediction result for the area to be predicted;

[0009] Input the augmented data soil carbon dynamics impact factor and the first soil carbon dynamics prediction result into the trained transformer model to obtain an optimized prediction result of the soil carbon dynamics for the area to be predicted.

[0010] In a preferred embodiment, the plant data includes: normalized difference vegetation index data, net primary productivity of vegetation data, leaf area index data, optimal growth temperature, and maximum temperature for growth inhibition; the soil data includes: bulk density data, field capacity, permanent wilting point, total phosphorus data, total nitrogen data, pH data, clay data, soil organic matter data, and soil water evaporation; the climate data includes: monthly maximum temperature, monthly minimum temperature, monthly precipitation, annual accumulated temperature, relative humidity, and surface net radiation intensity.

[0011] In a preferred embodiment, the CENTURY model is the calibrated and validated CENTURY4.0 model;

[0012] Among them, calibrating and validating the CENTURY model includes:

[0013] Extract the target data from the covariate data, and the target data is the historical climate data, plant data, and soil data within the area to be predicted;

[0014] Input the target data into the CENTURY4.0 initial model to obtain a second soil carbon dynamics prediction result for the area to be predicted;

[0015] Based on the second soil carbon dynamics prediction result and the actual result of the soil carbon dynamics corresponding to the target data, calibrate and validate the CENTURY4.0 initial model to obtain the CENTURY model.

[0016] In a preferred embodiment, before inputting the remaining data in the covariate data except the soil carbon dynamics impact factor into the CENTURY model, it further includes:

[0017] Based on the Delta downscaling method, downscale the future climate data in the climate data;

[0018] Based on the Taylor diagram, screen the climate patterns of the monthly precipitation, monthly minimum temperature, and monthly maximum temperature in the future climate data;

[0019] Based on the screened climate patterns, screen the future climate data.

[0020] In a preferred embodiment, extracting the soil carbon dynamics impact factor from the covariate data includes:

[0021] Based on the Pearson correlation coefficient method, determine the correlation between the covariate data and the dynamic changes of soil carbon, and obtain the correlation coefficients between each piece of the covariate data and the dynamic changes of soil carbon;

[0022] Based on the correlation coefficients, determine the soil carbon dynamic impact factors in the covariate data.

[0023] In a preferred embodiment, the soil carbon dynamic impact factors include: annual accumulated temperature, relative humidity, net radiation intensity, and normalized difference vegetation index.

[0024] In a preferred embodiment, the soil carbon dynamic prediction method further includes: a method for training an initial transformer model to obtain the trained transformer model;

[0025] The method for training the initial transformer model to obtain the trained transformer model includes:

[0026] Based on a preset time series generation model, obtain augmented data of the sample data. The sample data includes: soil carbon dynamic impact factors of a sample area, prediction results of the soil carbon dynamics of the sample area obtained by inputting the remaining data in the covariate data of the sample area except the soil carbon dynamic impact factors into the CENTURY model, and soil carbon dynamic observation values of the sample area;

[0027] Divide the augmented data into a training set and a test set;

[0028] Input the training set into the initial transformer model for regression training to obtain a preliminary transformer model;

[0029] Input the test set into the preliminary transformer model to obtain the soil carbon dynamic prediction results to be verified corresponding to the test set;

[0030] Based on the soil carbon dynamic prediction results to be verified, adjust the parameters of the preliminary transformer model through grid search and cross-validation methods to obtain the trained transformer model.

[0031] In a preferred embodiment, before dividing the augmented data into a training set and a test set, it further includes:

[0032] Perform dimensionality reduction processing on the sample data and the augmented data to obtain the dimensionality-reduced data of the sample data and the augmented data respectively;

[0033] Perform result visualization processing on the dimensionality-reduced data to obtain the feature distribution maps of the sample data and the augmented data;

[0034] Perform feature correlation analysis on the feature distribution maps to obtain the correlation between the sample data and the augmented data;

[0035] Based on the correlation, screen the augmented data.

[0036] In a preferred solution, the dimensionality reduction processing of the sample data and the augmented data includes:

[0037] Based on the principal component analysis method, perform dimensionality reduction processing on the sample data and the augmented data.

[0038] In a second aspect, the present invention provides a soil carbon dynamics prediction system, which is characterized by including:

[0039] An acquisition unit for acquiring covariate data of a region to be predicted, where the covariate data includes: plant data, soil data, and climate data within the region to be predicted;

[0040] A first processing unit for extracting soil carbon dynamics influencing factors from the covariate data, where the soil carbon dynamics influencing factors are data in the covariate data whose correlation with the soil carbon dynamics change in the region to be predicted exceeds a preset standard;

[0041] A second processing unit for inputting the remaining data in the covariate data except the soil carbon dynamics influencing factors into the CENTURY model to obtain a first soil carbon dynamics prediction result for the region to be predicted;

[0042] A prediction unit for inputting the soil carbon dynamics influencing factors and the first soil carbon dynamics prediction result into a trained transformer model to obtain an optimized soil carbon dynamics prediction result for the region to be predicted.

[0043] To achieve the above object, the soil carbon dynamics prediction method and system provided by the present invention first obtain covariate data including plant data, soil data, and climate data within the region to be predicted, then extract data in the covariate data whose correlation with the soil carbon dynamics change in the region to be predicted exceeds a preset standard as soil carbon dynamics influencing factors, and after inputting the remaining data in the covariate data except the soil carbon dynamics influencing factors into the CENTURY model to obtain a first soil carbon dynamics prediction result for the region to be predicted, input the soil carbon dynamics influencing factors and the first soil carbon dynamics prediction result into a trained transformer model to obtain an optimized soil carbon dynamics prediction result for the region to be predicted. Thus, by utilizing the non-linear prediction advantage of the transformer model, the generalization ability of the traditional soil carbon mechanics model CENTURY is greatly improved, and the accuracy of the soil carbon dynamics prediction for the region to be predicted is significantly enhanced. Description of the Drawings

[0044] Figure 1Flow chart of a soil carbon dynamics prediction method provided by an embodiment of the present invention.

[0045] Figure 2 Thermodynamic diagram of the correlation coefficients between the factors obtained by using the Pearson correlation coefficient method to extract the soil carbon dynamics influencing factors provided by an embodiment of the present invention.

[0046] Figure 3 (a) Effect evaluation diagram after correcting the simulation of soil carbon dynamics by the CENTURY 4.0 initial model provided by an embodiment of the present invention.

[0047] Figure 3 (b) Effect evaluation diagram after verifying the simulation of soil carbon dynamics by the corrected CENTURY 4.0 initial model provided by an embodiment of the present invention.

[0048] Figure 4 (a) Schematic diagram for evaluating the downscaling results of future climate data using a Taylor diagram provided by an embodiment of the present invention.

[0049] Figure 4 (b) Schematic diagram for evaluating the downscaling results of future climate data using a Taylor diagram provided by an embodiment of the present invention.

[0050] Figure 4 (c) Schematic diagram for evaluating the downscaling results of future climate data using a Taylor diagram provided by an embodiment of the present invention.

[0051] Figure 5 Schematic diagram of the visualization result of evaluating the original data and the generated data using the principal component analysis method provided by an embodiment of the present invention.

[0052] Figure 6 (a) Result comparison diagram of soil carbon dynamics prediction and the actual soil carbon dynamics of the calibration plot using the CENTURY 4.0 model and the soil carbon dynamics prediction method provided by the present invention in an embodiment of the present invention.

[0053] Figure 6 (b) Result comparison diagram of soil carbon dynamics prediction and the actual soil carbon dynamics of the verification plot using the CENTURY 4.0 model and the soil carbon dynamics prediction method provided by the present invention in an embodiment of the present invention.

[0054] Figure 7 (a) Taylor diagram of the topsoil organic carbon prediction of the calibration plot using the CENTURY 4.0 model and the soil carbon dynamics prediction method provided by the present invention in an embodiment of the present invention.

[0055] Figure 7(b) is the Taylor diagram of the topsoil organic carbon prediction of the verification plot using the CENTURY 4.0 model and the soil carbon dynamics prediction method provided by the present invention.

[0056] Figure 8 (a) is the change diagram of the soil organic carbon density of Arundo donax planted in a certain area predicted by the CENTURY 4.0 model.

[0057] Figure 8 (b) is the change diagram of the soil organic carbon density of Arundo donax planted in a certain area predicted by the method provided by the embodiment of the present invention.

[0058] Figure 9 It is the spatial distribution diagram of the soil carbon dynamics of Arundo donax planted in a certain area predicted by the method provided by the embodiment of the present invention under the future climate scenario mode SSP126.

[0059] Figure 10 It is the spatial distribution diagram of the soil carbon dynamics of Arundo donax planted in a certain area predicted by the method provided by the embodiment of the present invention under the future climate scenario mode SSP245.

[0060] Figure 11 It is the spatial distribution diagram of the soil carbon dynamics of Arundo donax planted in a certain area predicted by the method provided by the embodiment of the present invention under the future climate scenario mode SSP370.

[0061] Figure 12 It is the spatial distribution diagram of the soil carbon dynamics of Arundo donax planted in a certain area predicted by the method provided by the embodiment of the present invention under the future climate scenario mode SSP585.

[0062] Figure 13 For Figures 9 to 12 The spatial distribution diagram of the set average value of the optimized and corrected results of the soil carbon dynamic changes under the four climate scenarios shown.

[0063] Figure 14 (a) is the violin diagram of the annual increment distribution of the soil organic carbon density of Arundo donax planted in this area predicted by the CENTURY 4.0 model under different future climate scenarios.

[0064] Figure 14 (b) is the violin diagram of the annual increment distribution of the soil organic carbon density of Arundo donax planted in this area predicted by the method provided by the embodiment of the present invention under different future climate scenarios.

[0065] Figure 15 It is the structural diagram of a soil carbon dynamics prediction system provided by the embodiment of the present invention.

[0066] Figure 16 It is the structural schematic diagram of an electronic device provided by the embodiment of the present invention. Detailed implementation manners

[0067] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation manners.

[0068] The soil carbon dynamics prediction method provided by the present invention is executed on an electronic device, which may be an intelligent terminal device such as a laptop computer, a personal computer, and a tablet computer, or a server on the network side.

[0069] As Figure 1 shown, the soil carbon dynamics prediction method provided by the embodiments of the present invention mainly includes the following steps:

[0070] 110. Obtain the covariate data of the area to be predicted.

[0071] Among them, the covariate data includes: plant data, soil data, and climate data within the area to be predicted.

[0072] In this embodiment, according to the planted species of energy plants and soil types within the area to be predicted, the plant data and soil data of the area to be predicted are obtained.

[0073] Among them, the planted species of energy plants include: all plant species that can be planted among the first-generation biomass energy raw materials and the second-generation biomass energy raw materials. For example: soybeans, corn, sugarcane, beets, olives, etc. among the first-generation biomass raw materials; Arundo donax, Phragmites australis, Jatropha curcas, Miscanthus sinensis, Panicum virgatum, etc. among the second-generation biomass energy raw materials.

[0074] In an alternative embodiment, the plant data includes: normalized difference vegetation index data, vegetation net primary productivity data, leaf area index data, optimal growth temperature, and maximum growth-inhibiting temperature; the soil data includes: bulk density data, field capacity, permanent wilting point, total phosphorus data, total nitrogen data, pH data, clay data, soil organic matter data, and soil water evaporation; the climate data includes: monthly maximum temperature, monthly minimum temperature, monthly precipitation, annual accumulated temperature, relative humidity, and surface net radiation intensity.

[0075] In some possible embodiments, the covariate data of the area to be predicted may be sourced from the Climatic Research Unit (CRU) or the Coupled Model Intercomparison Project Phase 6 (CMIP6), the Harmonized World Soil Database (HWSD), the Big Earth Data Platform for Three Poles (BEDPTP), and the Resource and Environment Science Data Center (RESDC), etc. Specifically, the covariate data can be obtained with reference to Table 1.

[0076] Table 1 Example table of covariate data sources

[0077]

[0078]

[0079] Further, after obtaining data from CRU, CMIP6, etc., it is necessary to first fill in missing values and standardize the data, and then apply it as covariate data to subsequent steps.

[0080] 120. Extract the soil carbon dynamics impact factors from the covariate data.

[0081] Among them, the soil carbon dynamics impact factors are the data in the covariate data whose correlation with the soil carbon dynamics change in the area to be predicted exceeds a preset standard.

[0082] Further, in an optional embodiment, extracting the soil carbon dynamics impact factors from the covariate data includes:

[0083] Based on the Pearson correlation coefficient method, determine the correlation between the covariate data and the soil carbon dynamics change, and obtain the correlation coefficients between each covariate data and the soil carbon dynamics change;

[0084] Based on the correlation coefficients, determine the soil carbon dynamics impact factors in the covariate data.

[0085] Furthermore, the soil carbon dynamics impact factors include: annual accumulated temperature, relative humidity, net radiation intensity, and normalized difference vegetation index.

[0086] In this embodiment, the Pearson correlation coefficient method is used to extract the soil carbon dynamics impact factors from the obtained covariate data of the area to be predicted.

[0087] Specifically, the Pearson correlation coefficient (r) is often used to analyze the degree of correlation between variables, and its calculation formula is:

[0088]

[0089] Among them, the range of r is between -1 and 1. When it is greater than 0, it indicates a positive correlation, and when it is less than 0, it indicates a negative correlation. Taking variables such as annual accumulated temperature AT, soil moisture evaporation SME, and normalized difference vegetation index data NDVI as examples, through the Pearson correlation coefficient method, the correlation coefficient heat map between these variables is as Figure 2 shown. The correlation strength between variables and soil organic carbon can be judged by the value range of the absolute value of the correlation coefficient. Among them, 0.8 - 1.0 is a very strong correlation, 0.6 - 0.8 is a strong correlation, 0.4 - 0.6 is a medium - strength correlation, 0.2 - 0.4 is a weak correlation, and 0 - 0.2 is a very weak correlation or no correlation.

[0090] 130. Input the remaining data in the covariate data except the soil carbon dynamics impact factors into the CENTURY model to obtain the first soil carbon dynamics prediction result for the area to be predicted.

[0091] It is understandable that the accuracy of the CENTURY model determines the accuracy of the first soil carbon dynamics prediction results for the area to be predicted. Based on this, in an alternative embodiment, the CENTURY model is the calibrated and validated CENTURY4.0 model, thereby improving the accuracy of the CENTURY model.

[0092] Furthermore, calibrating and validating the CENTURY model includes:

[0093] Extracting the target data from the covariate data.

[0094] Among them, the target data is the historical climate data, plant data, and soil data within the area to be predicted;

[0095] Inputting the target data into the CENTURY4.0 initial model to obtain the second soil carbon dynamics prediction results for the area to be predicted;

[0096] Based on the second soil carbon dynamics prediction results and the actual soil carbon dynamics results corresponding to the target data, calibrating and validating the CENTURY4.0 initial model to obtain the CENTURY model.

[0097] In this embodiment, since the soil carbon dynamics results corresponding to the historical climate data, vegetation data, and soil data are known. Therefore, by inputting the historical climate data, plant data, etc. in the covariate data into the CENTURY4.0 initial model, the model parameters of the CENTURY4.0 initial model can be calibrated and validated. Refer to the Figure 3 effect evaluation of calibrating and validating the simulation of soil carbon dynamics by the CENTURY 4.0 initial model shown in (a) and (b), which effectively improves the accuracy of the CENTURY model for predicting soil carbon dynamics in the area to be predicted.

[0098] Furthermore, the accuracy of the covariate data is also an important factor affecting the accuracy of soil carbon dynamics prediction. Based on this, in an alternative embodiment, before inputting the remaining data in the covariate data except the soil carbon dynamics influencing factors into the CENTURY model, it further includes:

[0099] Based on the Delta downscaling method, downscaling the future climate data in the climate data;

[0100] Based on the Taylor diagram, screening the climate patterns of monthly precipitation, monthly minimum temperature, and monthly maximum temperature in the future climate data;

[0101] Based on the screened climate patterns, screening the future climate data.

[0102] It should be noted that the future climate data currently obtained is generally high - resolution climate scenario data of global climate models (GCMs) recognized by CMIP6. The assessment of climate change usually adopts the method of combining global climate models GCMs with future shared socioeconomic pathways SSPs. For example, the SSPs involved include SSP126, SSP245, SSP370, and SSP585, etc.

[0103] In some possible embodiments, for future climate data such as future temperature and precipitation downloaded from CMIP6, it is first downscaled by the Delta method and then adopted with the same resolution (a resolution of 0.5°×0.5°) as the historical climate data downloaded from CRU.

[0104] Among them, the downscaling formula for future temperature is as follows:

[0105] T f = T o +(T Gf - T Go ) (2)

[0106] Among them, T f is the future temperature matrix regenerated by the Delta method, T Gf is the future temperature matrix predicted by GCMs, T Go is the average temperature matrix of the reference period simulated by GCMs, and T o is the average temperature matrix actually observed in the reference period.

[0107] The downscaling formula for future precipitation data is as follows:

[0108]

[0109] Among them, P f is the future precipitation matrix regenerated by the Delta method, P Gf is the future precipitation matrix predicted by GCMs, P Go is the average precipitation matrix of the reference period simulated by GCMs, and P o is the average precipitation matrix actually observed in the reference period.

[0110] Furthermore, assuming that the GCMs involved include BCC - CSM2 - MR, CAS - ESM2 - 0, INM - CM4 - 8, MRI - ESM2 - 0, and CanESM5 - 1, the results of evaluating the performance of each GCM using Taylor diagrams are as shown in Figure 4 (a), (b), and (c). Among them, Figure 4 (a) represents the monthly average precipitation pr in each GCM, Figure 4(b) Characterize the monthly minimum temperature tmin in each GCM, Figure 4 (c) Characterize the monthly maximum temperature tmax in each GCM.

[0111] It can be understood that the Taylor diagram evaluates the accuracy changes of different climate models by calculating the correlation coefficient R, the centered root mean square error RMSE, and the standard deviation (σ f simulation value deviation, σ r observation value deviation) of the simulation value matrix f and the observation value matrix r, that is, as shown in the following formula:

[0112]

[0113] where, and are the average values of the simulation and observation data matrices respectively, and N is the total number of sample points in the study area. All data in the Taylor diagram are standardized. The observation field is located at the position with a standard deviation of 1. The larger the spatial correlation coefficient between the simulation data and the observation data shown on the Taylor diagram, the smaller the centered root mean square error between the two, and the closer the standard deviation is to 1, indicating that the simulation results of this model are better. Therefore, the advantages and disadvantages of the simulation results of different models can be seen more comprehensively from the Taylor diagram.

[0114] 140. Input the soil carbon dynamic impact factor and the first soil carbon dynamic prediction result into the trained transformer model to obtain the optimized prediction result of the soil carbon dynamic for the area to be predicted.

[0115] Specifically, the transformer model Transformer is essentially an encoder-decoder architecture. When the encoder processes each time point in the sequence, it calculates the attention scores through the self-attention mechanism to obtain a weighted vector to determine the degree to which the value of the time point is affected by its own or the values of other time points. The decoder gradually restores the low-dimensional representation of the target sequence obtained by the encoder to the target sequence. The formula of the above self-attention mechanism is as follows:

[0116]

[0117] where, Z is the output of the self-attention mechanism, Q, K, and V are obtained after the input vector undergoes different linear transformations respectively, T is the vector transpose operation, and d k is the scaling factor, which can make the gradient more stable during training. And in order for the self-attention mechanism to make full use of the sequential information of the time series data, Transformer also adds a position encoding mechanism, and the position encoding is realized by the following formula:

[0118] PE (pos,2i) = sin(pos / 100002i / dmodel ) (9)

[0119] PE (pos,2i+1) = cos(pos / 100002i / dmodel) (10)

[0120] where pos is the position in the input sequence and i is the index of the embedding value.

[0121] The vectors generated by the positional encoding are input into n identical encoder layers. Each encoder layer contains two sub-layers: a multi-head attention sub-layer and a feed-forward network sub-layer. After each sub-layer, there is a residual connection and a layer normalization operation. After the stacked input and output of n encoder layers, the last encoder layer will output an encoded matrix for input to the encoder layer.

[0122] The decoder has a structure similar to that of the encoder. The difference is that when performing the second attention mechanism calculation, the inputs of K and V come from the output encoded matrix of the encoder, while the Q matrix is calculated from the output of the previous decoder layer (if it is the first decoder layer, the input matrix of the encoder is used). Then, through the feed-forward neural network module, the target prediction sequence is mapped out.

[0123] Furthermore, in an optional embodiment, the soil carbon dynamics prediction method further includes: a method for training an initial Transformer model to obtain a trained Transformer model, that is:

[0124] Based on a preset time series generation model, augmented data of the sample data is obtained.

[0125] Among them, the sample data includes: the soil carbon dynamics impact factors in the sample area, the soil carbon dynamics prediction results for the sample area obtained by inputting the remaining data in the covariate data of the sample area except the soil carbon dynamics impact factors into the CENTURY model, and the soil carbon dynamics observation values of the sample area;

[0126] The augmented data is divided into a training set and a test set;

[0127] The training set is input into the initial Transformer model for regression training to obtain a preliminary Transformer model;

[0128] The test set is input into the preliminary Transformer model to obtain the soil carbon dynamics prediction results to be verified corresponding to the test set;

[0129] Based on the soil carbon dynamics prediction results to be verified, the preliminary Transformer model is tuned through the grid search method and the cross-validation method to obtain a trained Transformer model.

[0130] Specifically, the TimeGAN time series generation model is improved based on the lack of learning of the time series correlation in the Generative Adversarial Network (GAN), and it can associate the feature distributions at different time points and the complex relationships between time variables. The training of TimeGAN is a joint training, which relies on three different loss functions to train the autoencoder and the generative adversarial network.

[0131] That is, first, the input sequence is encoded into a low-dimensional vector through the embedding function in the autoencoder, which reduces the computational complexity on the one hand and extracts static and dynamic features on the other hand. The embedding function and the recovery function form a reconstruction loss function LR:

[0132]

[0133] Then the generator learns the static and dynamic features of the original data and generates new data. The unsupervised loss function LU for the generator and the discriminator is defined as:

[0134]

[0135] However, it can be understood that relying solely on the unsupervised training of the encoding-decoding network cannot accurately capture its dynamic characteristics in the time dimension. Therefore, it is required that while generating data, the generator can predict the data at the next time step T when the input data at the given 1-(T-1) moment is provided, and its prediction loss function LS is:

[0136]

[0137] In an optional embodiment, before dividing the augmented data into a training set and a test set, it further includes:

[0138] Performing dimensionality reduction processing on the sample data and the augmented data to obtain the dimensionality-reduced data of the sample data and the augmented data respectively;

[0139] Performing result visualization processing on the dimensionality-reduced data to obtain the feature distribution maps of the sample data and the augmented data;

[0140] Performing feature correlation analysis on the feature distribution maps to obtain the correlation between the sample data and the augmented data;

[0141] Based on the correlation, screening the augmented data.

[0142] In this embodiment, by performing dimensionality reduction processing on the sample data and the augmented data and visualizing the results, the feature distribution maps of the sample data, that is, the original data and the augmented data, are drawn, and then the correlation between the generated data and the original data is evaluated.

[0143] Further, in an optional embodiment, dimensionality reduction processing is performed on the sample data and the augmented data, including:

[0144] Based on the principal component analysis method, dimensionality reduction processing is performed on the sample data and the augmented data.

[0145] Specifically, referring to Figure 5 , the principal component analysis method PCA is used to perform dimensionality reduction processing on the data generated by TimeGAN and the original data, and the results are visualized, so that the correlation between the generated data and the original data can be seen.

[0146] In summary, for the soil carbon dynamics prediction method provided by the embodiments of the present invention, on the one hand, by using the TimeGAN time series generation model to augment the original data such as the soil carbon dynamics influencing factors for training the Transformer model, the problem of model overfitting caused by few data samples is effectively solved. By utilizing the non-linear prediction advantage of the Transformer model, the generalization ability of the traditional soil carbon mechanics model CENTURY is greatly improved. That is, by combining the TimeGAN and Transformer models, the prediction accuracy of the soil carbon dynamics in the area to be predicted is significantly improved.

[0147] On the other hand, when using the soil carbon dynamics prediction method provided by the embodiments of the present invention to predict the soil carbon dynamics in the area to be predicted, the time and capital cost investment for regional soil carbon dynamics monitoring can be effectively saved. That is, by simulating the soil carbon dynamics data in the area where the energy plant to be predicted is planted, the need for scientific researchers to conduct long-term soil experiments on the target area for many years can be reduced, thereby avoiding the consumption of a large amount of manpower, financial resources and material resources.

[0148] On yet another hand, by using the soil carbon dynamics prediction method provided by the embodiments of the present invention, the future soil carbon dynamics prediction of energy plants can be realized, and it can also be used to evaluate the carbon sequestration potential of unknown types of energy plant areas for planting. That is, without changing the regional vegetation cover type, the soil carbon dynamics changes caused by the long-term large-scale planting of a certain energy plant in the target area can be simulated and predicted, so as to help relevant personnel select the optimal energy plant species for planting in the target area.

[0149] Next, from the perspective of examples, the advantages of using the soil carbon dynamics prediction method provided by the embodiments of the present invention compared with using the traditional CENTURY for soil carbon dynamics prediction are described.

[0150] Specifically, the experimental monitoring data of the soil for Arundo donax cultivation in the experimental site of the University of Naples in Italy by researchers such as M. Fagnano for many years are used as the training data for the soil organic carbon of the sample soil and the calibration of the CENTURY 4.0 initial model. The verification data of the CENTURY 4.0 and the soil carbon dynamics prediction method provided by the embodiments of the present invention jointly use the simulation results of the Arundo donax cultivation in the Garriga di Podenzano test site in the northwest of Italy by researchers such as Enrico Martani.

[0151] Figure 6 (a) and (b) are respectively the results of predicting the soil carbon dynamics of the calibration plot and the verification plot using the traditional CENTURY model, the method (Transformer) of the embodiments of the present invention, and the actual observation value (Observation). Figure 6 It can be seen that whether it is the calibration plot or the verification plot, due to the limitation of the CENTURY 4.0 model to the inherent inputs in the model and the lack of the inputs of other actual environmental impact factors, there are large errors in the simulation results. And when the simulation is expanded from the drawing scale to a larger scale, there will be greater uncertainty, that is, the predicted values obtained by the CENTURY 4.0 model are only the same as the true values in the overall trend, but there are large errors in the numerical values. The predicted results obtained by using the method provided by the embodiments of the present invention are more in line with the actual situation.

[0152] Furthermore, in order to more intuitively reflect the optimization effect, referring to the method of evaluating the simulation effects of different GCMSs, the Taylor diagram is also used to evaluate the optimization effect. Then the Taylor diagrams of the topsoil organic carbon prediction of the two plots are respectively as Figure 7 (a) and (b) shown. Figure 7 It can be seen that whether it is the calibration plot or the verification plot, the TimeGAN-Transformer provided by the embodiments of the present invention has a correlation coefficient similar to that of the CENTURY 4.0 model, but has a lower root mean square error and standard deviation.

[0153] Even further, assuming that the future shared socioeconomic pathways SSPs are adopted for climate change assessment, and the SSPs involved include SSP126, SSP245, SSP370, and SSP585. Then Figure 8 (a) and (b) are respectively the change diagrams of the soil organic carbon density of Arundo donax cultivation in a certain area predicted by using the CENTURY 4.0 model and the method provided by the embodiments of the present invention, Figures 9 to 12 and are respectively the spatial distribution diagrams of the soil carbon dynamics of Arundo donax cultivation in this area predicted by using the method provided by the embodiments of the present invention under four different future climate scenario modes of SSP126, SSP245, SSP370, and SSP585, Figure 13Spatial distribution map of the set average of the optimized and corrected results of soil carbon dynamic changes under the above four climate scenarios Figure 14 (a) and (b) are violin plots showing the annual increment distribution of soil organic carbon density in the region when planting giant reed under different future climate scenarios predicted by the CENTURY 4.0 model and the method provided by the embodiment of the present invention respectively.

[0154] Then from Figure 8 (a), it can be seen that the soil organic carbon density simulated by CENTURY basically shows a linear growth trend. From Figure 8 (b), it can be seen that after optimization using the method provided by the embodiment of the present invention, it shows a fluctuating upward trend, and the soil organic carbon density increases slowly. It can also be noted that the soil organic carbon density increases the fastest in the first three years predicted by the method provided by the embodiment of the present invention, which is similar to the research results of Nocentini and Monti's "Land-use change from poplar to switchgrass and giant reed increases soil organic carbon", that is, after changing the land use from poplar to giant reed, the soil organic carbon increases instead, and the growth rate of soil organic carbon density in the first three years is faster than that in the following years. In fact, a large number of poplars were also planted in the area studied by Nocentini and Monti before, and later they were cleared due to damage to the ecological environment. The main reason for the relatively fast growth of soil organic carbon density in the early stage of planting giant reed is the decomposition of the carbon source accumulated by the previous poplars. Therefore, the planting history of their research area is similar to that of the area verified in this example and has reference value.

[0155] From Figure 14 (a), it can be seen that the average growth rate of soil organic carbon density simulated by CENTURY from 2025 to 2050 is 0.351 MgC / ha / yr, and the fluctuation range is: -0.386 - 1.074 MgC / ha / yr. From Figure 14(b) It can be seen that after optimization using the method provided in the embodiments of the present invention, the average growth rate is 0.146 MgC / ha / yr, and the fluctuation range is: -1.401 - 2.763 MgC / ha / yr. Regarding the research results on wetland carbon sequestration, it is found that the geometric mean of the carbon sequestration rate of North American freshwater mineral wetland soil is: 0.17 MgC / ha / yr, which is similar to the predicted result after optimization using the method provided in the embodiments of the present invention. And there is research indicating that 8 - 10 years is the minimum experimental frame to avoid the deviation of organic carbon estimation of perennial biomass crops. If only the change of topsoil organic carbon in the first 8 years is studied, after optimization using the method provided in the embodiments of the present invention, the soil organic carbon density increases by 0.427 MgC / ha / yr on average per year, which is closer to the carbon increase rate of 0.6 MgC / ha / yr of Arundo donax planted for 10 years in this research. Therefore, whether it is short-term or long-term planting of Arundo donax, the soil organic carbon density simulated by the CENTURY model is closer to the actual situation after being optimized by the method provided in the embodiments of the present invention.

[0156] It can be seen from Figures 9 to 12 that the average values of the soil organic carbon density in this area under different future climate scenario models, SSP126, SSP245, SSP370, and SSP585 are 3.68 MgC / ha, 3.74 MgC / ha, 3.67 MgC / ha, and 3.47 MgC / ha respectively. It can be seen from Figure 13 that the set average value of the optimized correction results of soil carbon dynamic changes under the four climate scenarios is: 3.64 MgC / ha. The area of this area is calculated by ArcMap10.5 to be 39145 ha. From this, it is deduced that the soil carbon sequestration potential in this area during the simulation period is: 0.143 TgC, and the average annual carbon sequestration potential is: 0.0057 TgC / yr.

[0157] Next, a soil carbon dynamic prediction system provided by the embodiments of the present invention will be introduced. The soil carbon dynamic prediction system described below can be considered as a module architecture for implementing the soil carbon dynamic prediction method provided by the embodiments of the present invention; the content described below can be referred to each other with the above content.

[0158] Optionally, referring to Figure 15 , Figure 15 is a structural block diagram of a soil carbon dynamic prediction system provided by the embodiments of the present invention. The system may include:

[0159] A collection unit 10, configured to obtain covariate data of the area to be predicted, where the covariate data includes: plant data, soil data, and climate data within the area to be predicted;

[0160] The first processing unit 20 is used to extract soil carbon dynamic impact factors from the covariate data, where the soil carbon dynamic impact factors are the data in the covariate data with a correlation exceeding a preset standard with the soil carbon dynamic changes in the area to be predicted;

[0161] The second processing unit 30 is used to input the remaining data in the covariate data except the soil carbon dynamic impact factors into the CENTURY model to obtain a first soil carbon dynamic prediction result for the area to be predicted;

[0162] The prediction unit 40 is used to input the soil carbon dynamic impact factors and the first soil carbon dynamic prediction result into the trained transformer model to obtain an optimized soil carbon dynamic prediction result for the area to be predicted.

[0163] Optionally, the plant data includes: normalized difference vegetation index data, net primary productivity of vegetation data, leaf area index data, optimal growth temperature, and maximum temperature for growth inhibition; the soil data includes: bulk density data, field water holding capacity, permanent wilting point, total phosphorus data, total nitrogen data, pH data, clay data, soil organic matter data, and soil water evaporation; the climate data includes: monthly maximum temperature, monthly minimum temperature, monthly precipitation, annual accumulated temperature, relative humidity, and surface net radiation intensity.

[0164] Optionally, the CENTURY model is the CENTURY4.0 model after calibration and verification;

[0165] Among them, calibrating and verifying the CENTURY model includes:

[0166] Extracting the target data in the covariate data, where the target data is the historical climate data, plant data, and soil data in the area to be predicted;

[0167] Inputting the target data into the CENTURY4.0 initial model to obtain a second soil carbon dynamic prediction result for the area to be predicted;

[0168] Based on the second soil carbon dynamic prediction result and the actual soil carbon dynamic result corresponding to the target data, calibrating and verifying the CENTURY4.0 initial model to obtain the CENTURY model.

[0169] Optionally, the acquisition unit 10 is further used for:

[0170] Downscaling the future climate data in the climate data based on the Delta downscaling method;

[0171] Screening the climate patterns of monthly precipitation, monthly minimum temperature, and monthly maximum temperature in the future climate data based on the Taylor diagram;

[0172] Based on the selected climate models, screen the future climate data.

[0173] Optionally, the first processing unit 20 is specifically configured to:

[0174] Based on the Pearson correlation coefficient method, determine the correlation between the covariate data and the dynamic changes of soil carbon, and obtain the correlation coefficients between each covariate data and the dynamic changes of soil carbon;

[0175] Based on the correlation coefficients, determine the soil carbon dynamic impact factors in the covariate data.

[0176] Optionally, the soil carbon dynamic impact factors include: annual accumulated temperature, relative humidity, net radiation intensity, and normalized difference vegetation index.

[0177] Optionally, the prediction unit 40 is further configured to:

[0178] Based on a preset time series generation model, obtain the augmented data of the sample data. The sample data includes: the soil carbon dynamic impact factors of the sample area, the soil carbon dynamic prediction results of the sample area obtained by inputting the remaining data in the covariate data of the sample area except the soil carbon dynamic impact factors into the CENTURY model, and the soil carbon dynamic observation values of the sample area;

[0179] Divide the augmented data into a training set and a test set;

[0180] Input the training set into the initial transformer model for regression training to obtain a preliminary transformer model;

[0181] Input the test set into the preliminary transformer model to obtain the soil carbon dynamic prediction results to be verified corresponding to the test set;

[0182] Based on the soil carbon dynamic prediction results to be verified, adjust the parameters of the preliminary transformer model through the grid search method and the cross-validation method to obtain the trained transformer model.

[0183] Optionally, the prediction unit 40 is further configured to:

[0184] Perform dimensionality reduction processing on the sample data and the augmented data to obtain the dimensionality-reduced data of the sample data and the augmented data respectively;

[0185] Perform result visualization processing on the dimensionality-reduced data to obtain the feature distribution maps of the sample data and the augmented data;

[0186] Perform feature correlation analysis on the feature distribution maps to obtain the correlation between the sample data and the augmented data;

[0187] Based on the correlation, screen the augmented data.

[0188] Optionally, the prediction unit 40 is more specifically configured to:

[0189] Based on the principal component analysis method, dimensionality reduction processing is performed on the sample data and the augmented data.

[0190] Next, with reference to Figure 16 to describe the electronic device provided in the embodiment of the present application. The electronic device provided in this embodiment may include: at least one processor 100, at least one communication interface 200, at least one memory 300, and at least one communication bus 400;

[0191] In the embodiment of the present invention, the number of the processor 100, the communication interface 200, the memory 300, and the communication bus 400 is at least one, and the processor 100, the communication interface 200, and the memory 300 complete mutual communication through the communication bus 400; obviously, Figure 16 The communication connection schematic diagram shown by the processor 100, the communication interface 200, the memory 300, and the communication bus 400 is only optional;

[0192] Optionally, the communication interface 200 may be an interface of a communication module, such as an interface of a GSM module; the processor 100 may be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiment of the present invention.

[0193] The memory 300 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory.

[0194] Among them, the processor 100 is specifically configured to execute the application program in the memory to implement the steps of the above-mentioned soil carbon dynamic prediction method.

[0195] The above embodiments are only examples given for clear illustration, and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. And the obvious changes or modifications derived therefrom still fall within the protection scope of the present invention.

Claims

1. A soil carbon dynamics prediction method, characterized in that: include: Acquiring covariate data of the area to be predicted, wherein the covariate data includes: plant data, soil data and climate data in the area to be predicted; Extracting soil carbon dynamics influencing factors from the covariate data, wherein the soil carbon dynamics influencing factors are data in the covariate data whose correlation with the soil carbon dynamics change in the area to be predicted exceeds a preset standard; Inputting the remaining data in the covariate data except the soil carbon dynamics influencing factor into the CENTURY model to obtain a first soil carbon dynamics prediction result for the area to be predicted; The soil carbon dynamics influencing factor and the first soil carbon dynamics prediction result are input into a trained converter model to obtain an optimized prediction result of the soil carbon dynamics in the area to be predicted.

2. The soil carbon dynamic prediction method according to claim 1, characterized in that: The plant data include: normalized vegetation index data, vegetation net primary productivity data, leaf area index data, optimal growth temperature and maximum temperature for growth inhibition; the soil data include: bulk density data, field water holding capacity, permanent wilting point, total phosphorus data, total nitrogen data, pH data, clay data, soil organic matter data and soil water evaporation; the climate data include: monthly maximum temperature, monthly minimum temperature, monthly precipitation, annual accumulated temperature, relative humidity, and net surface radiation intensity.

3. The soil carbon dynamic prediction method according to claim 1 or 2, characterized in that: The CENTURY model is a calibrated and verified CENTURY 4.0 model; The CENTURY model is calibrated and verified, including: Extracting target data from the covariate data, wherein the target data is historical climate data, plant data, and soil data in the area to be predicted; Inputting the target data into the CENTURY4.0 initial model to obtain a second soil carbon dynamic prediction result for the area to be predicted; Based on the second soil carbon dynamics prediction result and the actual soil carbon dynamics result corresponding to the target data, the CENTURY 4.0 initial model is corrected and verified to obtain the CENTURY model.

4. The soil carbon dynamic prediction method according to claim 2, characterized in that: Before inputting the remaining data in the covariate data except the soil carbon dynamic influencing factor into the CENTURY model, the method further includes: Based on the Delta downscaling method, downscaling the future climate data in the climate data; Based on the Taylor diagram, the climate patterns of monthly precipitation, monthly minimum temperature and monthly maximum temperature in the future climate data are screened; The future climate data is filtered based on the climate model obtained by filtering.

5. The soil carbon dynamic prediction method according to claim 1, characterized in that: The extracting of soil carbon dynamic influencing factors from the covariate data includes: Based on the Pearson correlation coefficient method, the correlation between the covariate data and the dynamic changes of soil carbon is determined, and the correlation coefficient between each covariate data and the dynamic changes of soil carbon is obtained; Based on the correlation coefficient, the soil carbon dynamics influencing factor in the covariate data is determined.

6. The soil carbon dynamic prediction method according to claim 4, characterized in that: The soil carbon dynamic influencing factors include: annual accumulated temperature, relative humidity, net radiation intensity and normalized difference vegetation index.

7. The soil carbon dynamic prediction method according to claim 1, characterized in that: Also includes: A method for training an initial converter model to obtain the trained converter model; The method of training the initial converter model to obtain the trained converter model includes: Based on a preset time series generation model, expansion data of sample data is obtained, wherein the sample data includes: soil carbon dynamics influencing factors of the sample area, soil carbon dynamics prediction results of the sample area obtained by inputting the remaining data of the covariate data of the sample area except the soil carbon dynamics influencing factors into the CENTURY model, and soil carbon dynamics observation values ​​of the sample area; Dividing the augmented data into a training set and a test set; Inputting the training set into the initial converter model for regression training to obtain a preliminary converter model; Inputting the test set into the preparatory converter model to obtain a soil carbon dynamic prediction result to be verified corresponding to the test set; Based on the predicted results of soil carbon dynamics to be verified, the preparatory converter model is adjusted through grid search method and cross-validation method to obtain the trained converter model.

8. The soil carbon dynamic prediction method according to claim 7, characterized in that: Before dividing the expanded data into a training set and a test set, the method further includes: Performing dimensionality reduction processing on the sample data and the expanded data to obtain dimensionality reduction data of the sample data and the expanded data respectively; Performing result visualization processing on the dimension reduction data to obtain feature distribution diagrams of the sample data and the expanded data; Performing feature correlation analysis on the feature distribution graph to obtain the correlation between the sample data and the expanded data; Based on the correlation, the expanded data is screened.

9. The soil carbon dynamic prediction method according to claim 8, characterized in that: The performing dimensionality reduction processing on the sample data and the expanded data includes: Based on the principal component analysis method, dimension reduction processing is performed on the sample data and the expanded data.

10. A soil carbon dynamic prediction system, characterized in that: include: A collection unit, used for acquiring covariate data of the area to be predicted, wherein the covariate data includes: plant data, soil data and climate data in the area to be predicted; A first processing unit is used to extract a soil carbon dynamics influencing factor from the covariate data, wherein the soil carbon dynamics influencing factor is data in the covariate data whose correlation with the soil carbon dynamics change in the area to be predicted exceeds a preset standard; The second processing unit is used to input the remaining data of the covariate data except the soil carbon dynamics influencing factor into the CENTURY model to obtain a first soil carbon dynamics prediction result for the area to be predicted; The prediction unit is used to input the soil carbon dynamics influencing factor and the first soil carbon dynamics prediction result into the trained converter model to obtain the soil carbon dynamics optimization prediction result of the area to be predicted.

Images