Farmland net carbon emission prediction method, system and device

Abstract

The invention provides a farmland net carbon emission prediction method, a farmland net carbon emission prediction system and a farmland net carbon emission prediction device. The farmland net carbon emission prediction method comprises the following steps: acquiring crop related data, nitrous oxide emission measured data, methane emission measured data and carbon sequestration efficiency measured data of different crops in different planting areas; constructing a nitrous oxide emission prediction model according to the crop related data and nitrous oxide emission measured data; constructing a methane emission prediction model according to the crop related data and the methane emission measured data; constructing a carbon sequestration efficiency prediction model according to the crop related data and the carbon sequestration efficiency measured data; calculating a methane emission predicted value, a nitrous oxide emission predicted value and a carbon sequestration efficiency predicted value; and calculating a net carbon emission predicted value according to the nitrous oxide emission predicted value, the methane emission predicted value, the carbon sequestration efficiency predicted value and a preset net carbon emission model. According to the invention, high-precision and regionalized prediction and evaluation of farmland net carbon emission can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of agricultural carbon emission prediction, and specifically relates to a method, system and device for predicting net carbon emissions from farmland. Background Art

[0002] As the issue of greenhouse gas emissions from agricultural systems has received increasing attention, the research on farmland carbon emission assessment systems has become a technical focus. The system usually includes two core links: one is the simulation of greenhouse gas emission processes, especially the high temporal and spatial resolution estimation of key gases such as N2O and CH4; the other is the quantification of carbon sequestration processes, which mainly involves the contribution of measures such as straw return to the field, organic fertilizer or biochar application to the increase in soil organic carbon. Due to the high complexity of farmland ecological processes, emissions and carbon sequestration processes are subject to the interactive influence of various management measures (such as planting density, fertilization methods and types, tillage methods) and environmental factors (soil type, climatic conditions, etc.), forming a nonlinear, multivariable coupled feedback mechanism, which places high demands on model structure design, variable processing capabilities and dynamic response capabilities. The prediction methods in existing technologies have the following deficiencies or defects: First, when it comes to greenhouse gas emissions modeling, traditional approaches often rely on IPCC default emission factors or empirical statistical models. While these methods can provide rough estimates, they struggle to capture the complex, nonlinear interactions between farmland management practices and environmental factors at a finer scale. For example, nitrous oxide (N2O) emissions are influenced not only by fertilizer application and crop type, but also by the coupling between management practices like tillage methods and fertilization timing, climate, and soil properties. These variables exhibit highly nonlinear and multi-layered interactions, making accurate modeling difficult with fixed coefficients or single-variable functions alone.

[0003] Furthermore, the multi-source heterogeneity of variables (the coexistence of categorical and continuous variables), scale differences (the coexistence of field operations and long-term climate averages), and temporal dynamics further exacerbate modeling difficulties. While recent studies have attempted to improve model prediction accuracy by introducing machine learning methods such as neural networks and random forests, most modeling schemes remain "black box tools," failing to effectively integrate agronomic process mechanisms, lacking cross-regional generalization and the ability to explain management responses, making them difficult to achieve the goals of farmland carbon reduction and precise regulation.

[0004] Second, in carbon sequestration modeling, existing studies mostly use empirical parameters or fixed coefficients, ignoring the interactions between farmland management practices (such as tillage methods, fertilization intensity, water regulation, and crop type) and environmental soil and climate factors. For example, the same straw return policy may have completely different carbon stability and sequestration potential in hot and humid southern regions and arid and cold northern regions. Traditional methods generally lack the ability to model the "management × environment" interaction, making it impossible to simulate carbon sequestration efficiency tailored to local conditions.

[0005] Furthermore, while patent CN113513269A proposes a farmland carbon emissions prediction system based on "resource estimation," integrating multi-source data and modeling methods, its core remains based on the static application of emission factors and lacks a coupled learning mechanism between resource input and carbon emissions. The so-called "prediction model" is actually a combination of traditional models nested together, making it difficult to address the dynamic response of emissions processes to changes in management measures. Overall, it resembles a data-filling inventory tool rather than an intelligent system with generalization capabilities and decision support capabilities.

[0006] In summary, existing farmland carbon emission and sequestration assessment technologies generally suffer from low estimation accuracy, insufficient responsiveness, and poor system integration. Emission models often rely on fixed emission factors, which fail to reflect the complex interactions between management practices and environmental factors. Sequestration assessments often use a unified efficiency coefficient, ignoring the differences between different management practices. Furthermore, while some data-driven methods have improved predictive performance, they generally lack agronomic support, resulting in limited model interpretability and regional generalization, making them unable to meet the current demands for intelligent and refined agricultural carbon accounting. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method, system and device for predicting net carbon emissions from farmland, which can achieve high-precision, regionalized prediction and evaluation of net carbon emissions from farmland.

[0008] The present invention provides a method for predicting net carbon emissions from farmland, comprising the following steps: S1. Obtain crop-related data, measured nitrous oxide emission data, measured methane emission data, and measured carbon sequestration efficiency data for different crops in different planting areas, and divide the data into training and validation sets; the crop-related data include crop type data, management measures data, soil data, and climate data; S2. Constructing a nitrous oxide emission prediction model based on machine learning based on the crop-related data and the measured nitrous oxide emission data; S3. Constructing a methane emission prediction model based on machine learning based on the crop-related data and the measured methane emission data; S4. Constructing a carbon sequestration efficiency prediction model based on machine learning based on the crop-related data and the measured carbon sequestration efficiency data; S5. Calculating a predicted methane emission value, a predicted nitrous oxide emission value, and a predicted carbon sequestration efficiency value, respectively, based on the crop-related data, the nitrous oxide emission prediction model, the methane emission prediction model, and the carbon sequestration efficiency prediction model; S6. Calculate a net carbon emission prediction value based on the nitrous oxide emission prediction value, the methane emission prediction value, the carbon sequestration efficiency prediction value, and a preset net carbon emission model.

[0009] Furthermore, the S1 also includes: preprocessing the crop-related data, performing integer encoding processing on the categorical variables, normalizing all variables, and using the processed data as basic data for dividing the training set and the validation set.

[0010] Furthermore, the S2 includes the following steps: S201. Construct a nonlinear emission response function based on nitrogen input, N2O emission value, natural N2O emission coefficient under no nitrogen input conditions, and sensitivity factor of N2O emission to nitrogen application rate; S202, constructing a twin neural network, and using the crop-related data and the measured nitrous oxide emission data in the training set and the validation set to determine the values ​​of the natural N2O emission coefficient and the sensitivity factor; S203: Substitute the values ​​of the natural N2O emission coefficient and the sensitivity factor into the nonlinear emission response function to obtain the nitrous oxide emission prediction model.

[0011] Furthermore, the expression of the nonlinear emission response function is: ; in, Indicates N2O emission value, represents the nitrogen input value, a represents the natural N2O emission coefficient under the condition of no nitrogen input, and b represents the sensitivity factor of N2O emission to the amount of nitrogen applied.

[0012] Furthermore, the S3 further includes the following steps: S301. Setting a candidate methane emission algorithm and its key parameters; S302, performing cross-validation training on the candidate methane emission algorithms based on the crop-related data and the measured methane emission data in the training set; S303: Using preset evaluation indicators to determine the best methane emission candidate algorithm as the methane emission prediction model.

[0013] Furthermore, the S4 includes the following steps: S401. Setting a candidate carbon sequestration efficiency algorithm and its key parameters; S402, performing cross-validation training on the carbon sequestration efficiency candidate algorithms based on the crop-related data and the measured carbon sequestration efficiency data in the training set; S403. Using preset evaluation indicators, determine the best candidate algorithm as the carbon sequestration efficiency prediction model.

[0014] Furthermore, the S6 includes the following steps: S601. Calculate a greenhouse gas carbon emission prediction value based on the nitrous oxide emission prediction value and the methane emission prediction value; S602, calculating a carbon fixation prediction value based on the carbon fixation efficiency prediction value and the carbon input value; S603: Calculate a net carbon emission prediction value based on the greenhouse gas carbon emission prediction value and the carbon fixation prediction value.

[0015] Furthermore, the calculation formula for the predicted net carbon emissions value is: ; The calculation formula for the predicted greenhouse gas carbon emissions is: ; The formula for calculating the predicted carbon sequestration value is: ; in, represents the predicted value of nitrous oxide emissions, represents the predicted value of methane emissions, represents the predicted value of carbon sequestration efficiency, Represents the carbon input value.

[0016] Furthermore, the method further includes: S7, generating a recommendation plan based on the predicted nitrous oxide emission value, the predicted methane emission value, the predicted carbon fixation value, the predicted net carbon emission value and the corresponding weights, including the following steps: S701. For any variety in any planting area, construct a scenario matrix based on continuous variables, categorical variables, predicted nitrous oxide emissions, predicted methane emissions, predicted carbon sequestration, and predicted net carbon emissions; the continuous variables include carbon input and nitrogen input, and the categorical variables include tillage method, number of fertilizations, fertilization method, fertilizer type, and planting density; S702: performing positive indicator normalization processing on the predicted nitrous oxide emissions, predicted methane emissions, predicted carbon fixation, and predicted net carbon emissions of any scenario in the scenario matrix to obtain four indicators; S703. Construct a comprehensive scoring model based on the four indicators and their weights; S704: Generate a recommendation plan with a corresponding weight combination based on the maximum comprehensive score.

[0017] The present invention also provides a farmland net carbon emission prediction system, comprising: An acquisition module is used to obtain crop-related data, measured nitrous oxide emission data, measured methane emission data, and measured carbon sequestration efficiency data for different crops in different planting areas, and divide the data into training sets and validation sets; the crop-related data includes crop type data, management measures data, soil data, and climate data; A first model building module is used to build a nitrous oxide emission prediction model based on machine learning according to the crop-related data and the measured nitrous oxide emission data; A second model building module is used to build a methane emission prediction model based on machine learning according to the crop-related data and the measured methane emission data; A third model building module is used to build a carbon sequestration efficiency prediction model based on machine learning according to the crop-related data and the measured carbon sequestration efficiency data; A prediction module, configured to calculate a methane emission prediction value, a nitrous oxide emission prediction value, and a carbon sequestration efficiency prediction value, respectively, based on the crop-related data, the nitrous oxide emission prediction model, the methane emission prediction model, and the carbon sequestration efficiency prediction model; A calculation module is used to calculate the net carbon emission prediction value based on the nitrous oxide emission prediction value, the methane emission prediction value, the carbon sequestration efficiency prediction value and a preset net carbon emission model.

[0018] Furthermore, it also includes: A recommendation module is used to generate a recommendation plan based on the nitrous oxide emission prediction value, the methane emission prediction value, the carbon fixation prediction value, the net carbon emission prediction value and the corresponding weights.

[0019] The present invention also provides a device for predicting net carbon emissions from farmland, comprising a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the steps of the above method.

[0020] The beneficial effects brought by the present invention are: 1. Universality and Adaptability of Multi-Agent Modeling by Crop and Region: This paper proposes a multi-agent modeling method based on crops and regions. It constructs corresponding agent models based on the growth patterns and phenological characteristics of different crops, as well as differences in regional climate and soil conditions, and implements collaborative simulation at the regional scale. This method effectively captures the impact of spatial heterogeneity on carbon emissions and carbon sequestration processes, ensuring high universality and regional adaptability across crops and regions, overcoming the limitation of single models, which often suffer from reduced prediction accuracy under different environmental conditions.

[0021] 2. High-precision prediction capabilities fused with multi-dimensional input data: During the modeling process, this paper incorporates multi-dimensional input data covering soil physical and chemical properties, climate factors, carbon and nitrogen inputs, tillage management, crop management, and fertilization management. This approach leverages the interpretability of farmland process mechanisms with the nonlinear fitting advantages of data-driven models to form a hybrid prediction framework that integrates mechanism and data. This framework significantly improves prediction accuracy and stability across multiple crop systems and management scenarios, ensuring that carbon budget assessment results meet decision-making requirements in terms of both spatiotemporal resolution and numerical reliability.

[0022] 3. Precision Management and Policy Support Based on Model Output: Leveraging the highly accurate and adaptable prediction results described above, this invention enables intelligent farmland management plan recommendations based on model output. Users and government management departments can quickly screen and optimize combinations of management measures, such as tillage systems, fertilization strategies, and carbon and nitrogen inputs, based on different carbon reduction and sequestration targets, to achieve precise, locally applicable management. This functionality also provides a quantitative basis for the formulation of agricultural carbon management policies, supporting the implementation of differentiated and precise carbon reduction and sequestration strategies across different regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a schematic diagram of the method in S2 of the present invention; Figure 3 This is a comparison chart of the output results of the nitrous oxide emission prediction model of the present invention and existing machine learning; Figure 4 Schematic diagram of the output results of the methane emission prediction model in the present invention; Figure 5 This is a schematic diagram of the output results of the carbon sequestration efficiency prediction model in the present invention. DETAILED DESCRIPTION

[0024] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0025] As 1 to Figure 5 As shown, taking a certain planting area as an example, the present invention provides a method for predicting net carbon emissions from farmland, including the following steps: S1. Obtain crop-related data, measured nitrous oxide emissions, measured methane emissions, and measured carbon sequestration efficiency data for different crops in different growing areas, and divide them into training and validation sets with an 8:2 ratio. Crop-related data includes crop type data, management practice data, soil data, and climate data.

[0026] Crop type data include corn, wheat, and rice.

[0027] Management measure data include the number of fertilizations (single, multiple), fertilization method (surface application, deep application), fertilizer type (urea, inorganic fertilizer, high-efficiency nitrogen fertilizer, organic fertilizer, green manure, biochar, straw), tillage method (no-till, conventional tillage), planting density (low density, medium density, high density), nitrogen input, and carbon input.

[0028] Soil data include soil pH, bulk density, organic carbon content, total nitrogen content, clay content, and cation exchange capacity.

[0029] Climate data include average annual temperature and average annual rainfall.

[0030] The above data contains categorical variables. In order to use it as input data for model training, the crop-related data is preprocessed. Specifically, the categorical variables are integer-encoded. The processing results are shown in the following table:

[0031] After all variables are represented numerically, they are uniformly normalized using the Min-Max normalization method, scaling them to the range [0, 1]. This process is performed uniformly across the entire dataset to ensure that each feature dimension is within a relatively balanced numerical range within the model input. Subsequently, the processed samples are divided into a training set and a validation set, with a ratio of 80% training and 20% validation, ensuring that the model can achieve effective learning while maintaining generalization capabilities.

[0032] Considering the significant spatial heterogeneity of farmland carbon cycling across natural regions and agricultural ecosystems, this paper introduces a sub-modeling strategy based on crop-region combinations during the data modeling phase to further enhance the model's regional adaptability and predictive representativeness. Specifically, for maize-growing regions, data were categorized into four sub-regions based on their spatial distribution across China: the Northeast, Northwest, Huanghuaihai Plain, and Southern China. For wheat-growing systems, the farming systems and soil-climate conditions of each ecological zone were combined to sub-categorize them into the Northern Dryland Region, the Huanghuaihai Winter Wheat-producing Region, the Southwest Hilly Transition Region, and the Middle and Lower Yangtze River Irrigated Region. Rice-growing systems were further categorized into the Northern Rice-growing Region, the Yangtze River Basin Rice-growing Region, the Southwest Mountainous Region, and the South China Tropical Rice-growing Region based on differences in hydrothermal conditions and cropping patterns. Predictive models were constructed for each of these crop-growing regions, effectively capturing the inter-regional differential impacts of key influencing factors (such as climate, soil, and management practices) on carbon emissions and sequestration. This allows for more regionally tailored intelligent management recommendation recommendations, enhancing the practicality and application of model outputs in practical agricultural decision-making.

[0033] S2. Construct a nitrous oxide emission prediction model based on machine learning based on crop-related data and measured nitrous oxide emission data. Specifically, Figure 2 As shown, S2 includes the following steps: S201. Construct a nonlinear emission response function based on nitrogen input, N2O emission value, natural N2O emission coefficient under no nitrogen input conditions, and sensitivity factor of N2O emission to nitrogen application rate. The expression of the nonlinear emission response function is: ; in, Indicates N2O emission value, represents the nitrogen input value, a represents the natural N2O emission coefficient under no nitrogen input conditions, and b represents the sensitivity factor of N2O emissions to nitrogen application rate. This functional form effectively reflects the nonlinear variation pattern of N2O emissions, with relatively flat emissions at low nitrogen application levels and rapidly increasing emissions at high nitrogen application levels.

[0034] S202. Since both a and b are unknown quantities, two parallel fully connected feedforward neural networks are constructed: subnetwork A is used to estimate a, and subnetwork B is used to estimate b. The natural N2O emission coefficient and sensitivity factor are determined using crop-related data and measured N2O emission data from the training and validation sets.

[0035] Specifically, the input data for subnetwork A is soil and climate data. Because background emissions are less affected by management measures, management variables are not included. Subnetwork A consists of five layers, with the number of neurons being 64, 128, 256, 128, and 64, respectively. This gradually enhances the network's ability to fit the nonlinear relationship between environmental variables and background emissions. The inputs for subnetwork B are soil, climate, and management measures data. Subnetwork B maintains the same structure as subnetwork A, also a five-layer fully connected neural network with the number of neurons being 64, 128, 256, 128, and 64, ensuring consistency in structural comparison and comparability in model responses.

[0036] To optimize model performance, this system uses a unified loss function for end-to-end training, with the goal of minimizing the mean square error (MSE) between predicted and measured emissions: ; where y i is the measured value of N2O emission, is the predicted N2O emission value, and i is the sample index. Training is performed using the Adam optimizer, with gradients backpropagated through the subnetwork and exponential function structure to simultaneously optimize the two intermediate parameters a and b and the overall model prediction accuracy. The training iterations were set to 2000 to ensure sufficient model convergence and stability.

[0037] S203. Substitute the values ​​of the natural N2O emission coefficient and the sensitivity factor into the nonlinear emission response function to obtain a nitrous oxide emission prediction model.

[0038] like Figure 3 As shown in the figure, the model (KDL-N2O) constructed in this step is compared with three traditional machine learning methods: random forest (RF), extreme gradient boosting (XGBoost), and standard neural network (NN). It can be seen that the model is superior to the above methods in terms of prediction accuracy and nonlinear trend capture, among which the overall prediction R 2 The R value of the method is 0.79, which is significantly higher than that of other methods. 2The range of the RI is 0.65 to 0.76. In terms of emission curve fitting, the model also shows stronger consistency and stability.

[0039] It can be seen that this step achieves accurate and highly interpretable modeling of nitrous oxide emissions from farmland through the coupled modeling structure of "climate-soil-management input to parameter estimation and then to nonlinear emission prediction", which makes the model have both the interpretability of agronomic formulas and the generalization ability of neural networks. This structure breaks through the technical limitations of fixed parameters and rigid responses in traditional models, and achieves functional enhancement of the model structure rather than just variable stacking. The model input system covers management measures such as fertilizer amount, tillage method, fertilization time, and planting density, as well as environmental factors such as soil type and climatic conditions, and realizes fusion modeling through a unified feature space, which improves the model's response accuracy to changes in the combination of management measures and solves the problem of insufficient adaptability of previous models to single variables or fixed scenarios.

[0040] S3: Construct a machine learning-based methane emission prediction model based on crop-related data and measured methane emission data. Specifically, S3 also includes the following steps: S301. Set candidate methane emission algorithms and their key parameters. Candidate methane emission algorithms include ridge regression algorithm, support vector regression algorithm, feedforward neural network algorithm, residual neural network algorithm, gradient boosting tree algorithm, and random forest algorithm.

[0041] The main hyperparameter of the ridge regression algorithm is the strength of the L2 regularization term, or the regularization coefficient, which ranges from 0.1 to 10 with a step size of 0.1. This controls the degree to which the model penalizes the magnitude of the parameter. Since ridge regression uses a closed-form solution and does not require iterative training, there is no need to set a stopping condition.

[0042] The key parameters of the support vector regression algorithm include: the penalty coefficient search range is 1 to 100, with a step size of 10; the error tolerance range is set to 0.01 to 0.2, with a step size of 0.01. To control the training time, the maximum number of iterations is set to 1000. If the loss function changes within an iteration by less than 1×10- 3 Stop early.

[0043] The parameter space of the feedforward neural network algorithm includes: hidden layer structure, set to single hidden layer (32, 64, 128) or double hidden layer ([64, 32], [128, 64]); learning rate range from 0.001 to 0.01, with a step size of 0.001. The training process is performed for a maximum of 100 epochs, and an early stopping mechanism is introduced. If the validation error does not decrease significantly within 10 consecutive epochs (the change is less than 1×10- 4 ) means terminating the training early.

[0044] To accommodate deeper model training, the residual neural network algorithm uses the following parameters: number of residual blocks ∈ [2, 3, 4]; number of neurons per block ∈ [32, 64, 128]; and learning rate, also ranging from 0.001 to 0.01 with a step size of 0.001. The maximum number of training epochs is 150, and early stopping is enabled, using the same criteria as for feedforward neural networks.

[0045] The gradient boosting tree algorithm was used as the ensemble learning model. Its parameters included the following: number of weak learners ∈ [100, 200, ..., 1000], with a step size of 100; maximum tree depth ∈ [3, 5, ..., 10], with a step size of 1; and subsampling ratio ∈ [0.6, 0.7, ..., 1.0], with a step size of 0.1. Training was performed using an early stopping mechanism; if the validation set error did not decrease within 20 consecutive epochs, the iteration was terminated.

[0046] The parameters of the random forest algorithm include the number of decision trees, ranging from 100 to 1000, with a step size of 100; the size of the feature subset, ranging from 1 to 20, with a step size of 1.

[0047] S302. Cross-validation training is performed on the candidate methane emission algorithms based on the crop-related data and measured methane emission data in the training set. Specifically, in the process of hyperparameter optimization, in order to uniformly evaluate the comprehensive performance of the model under different parameter combinations, a ten-fold cross-validation strategy is adopted, and the key performance indicators are normalized and weighted scoring is performed. The specific process is as follows: First, for each set of candidate parameter combinations, a ten-fold cross-validation is performed: the training set is divided into ten equal sub-sets, one of which is used as the validation set in turn, and the remaining nine are used for model training. After completing ten rounds, the average is taken to obtain three evaluation indicators: root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 ).

[0048] S303, using the preset evaluation index to determine the best methane emission candidate algorithm as the methane emission prediction model. Since the three indicators have different measurement scales and inconsistent optimization directions, in order to ensure comparability, normalization processing must be performed first. RMSE and MAE use the reverse normalization method, R 2 Using the positive normalization method, we get the following results: RMSE_norm=(RMSE_max-RMSE) / (RMSE_max-RMSE_min); MAE_norm=(MAE_max-MAE) / (MAE_max-MAE_min); R 2 _norm=(R 2 -R 2 _min) / (R 2 _max-R2 _min).

[0049] The weight parameter ratio of each indicator is set to 1:1:1, and the comprehensive scoring function is constructed as follows: Score=RMSE_norm+MAE_norm+R 2 _norm.

[0050] Combining the cross-validation evaluation results with the model complexity, the algorithm with the highest comprehensive score under multiple indicators was selected from the candidate methane emission algorithms as the final methane emission prediction model.

[0051] like Figure 4 As shown, in this example, the random forest algorithm was selected as the final methane emissions prediction model due to its strong nonlinear fitting capabilities, robust feature selection, and overall stability. After multiple rounds of validation, the final parameters selected were: 100 decision trees and a feature subset size of 12. All other parameters remained at their default settings, and after model training, the prediction and evaluation process proceeded directly. It not only demonstrated low error during the training phase but also demonstrated good generalization on an independent test set.

[0052] The final model can output the methane emission intensity per unit area of ​​rice fields under different combinations of management measures, achieve high-precision estimation of the methane contribution of rice areas to the carbon emission system, and provide technical support for intelligent carbon accounting and optimization of field management measures.

[0053] This example employs a nonlinear modeling approach centered on random forests, integrating the collaborative modeling of farmland management practices and environmental variables. By systematically constructing an input feature system encompassing key factors such as cropping systems, fertilization strategies, tillage practices, carbon and nitrogen inputs, soil properties, and climatic conditions, the model effectively identifies the nonlinear regulatory mechanisms of CH4 emissions under different combinations of management and environmental variables. Furthermore, by extracting the model's impact weights for each input variable and combining them with a visual analysis of variable-emission response trends, it is possible to identify the driving direction and extent of the impact of key management practices on CH4 emissions.

[0054] In summary, this example breaks through the problems of low variable fusion and weak interactive processing capabilities of traditional methods in CH4 emission modeling, constructs a random forest model that integrates key management measures and environmental variables, significantly improves the model's prediction accuracy, application adaptability, and management interpretability, and has good scalability and practical application value.

[0055] S4. Constructing a carbon sequestration efficiency prediction model based on machine learning based on crop-related data and measured carbon sequestration efficiency data. Further, S4 includes the following steps: S401: Setting carbon sequestration efficiency candidate algorithms and their key parameters: The carbon sequestration efficiency candidate algorithms in this step have the same candidate algorithm type and key parameters as those in S301, so they are not described in detail.

[0056] S402: Cross-validation training is performed on the candidate carbon sequestration efficiency algorithms based on the crop-related data in the training set and the measured carbon sequestration efficiency data. The training method in this step is the same as that in S302, so it will not be repeated here.

[0057] S403, using the preset evaluation index to determine the best candidate algorithm as the carbon sequestration efficiency prediction model. The evaluation index in this step is the same as the evaluation index in S303, so it will not be repeated. Figure 5 As shown, in this example, the random forest algorithm was selected as the final carbon sequestration efficiency prediction model due to its strong nonlinear fitting capabilities, robust feature selection, and overall stability. After multiple rounds of validation, the final parameters selected were: 100 decision trees and a feature subset size of 8. All other parameters remained at their default settings, and after model training, the prediction and evaluation process proceeded directly. It not only demonstrated low error during the training phase but also demonstrated good generalization on an independent test set.

[0058] The final model can output the carbon sequestration efficiency per unit area of ​​farmland under different combinations of management measures, providing technical support for intelligent carbon sequestration accounting and optimization of field management measures.

[0059] This example integrates a collaborative modeling system for farmland greenhouse gas (CH4, N2O) emissions and soil organic carbon (SOC) sequestration. During the construction phase, data connectivity and structural alignment between models were achieved. Interactions between management measures and environmental factors were incorporated into each submodule, significantly improving the estimation accuracy and regional adaptability of each indicator. The system ultimately outputs the net carbon emission intensity per unit area of ​​farmland under a given management combination (emissions minus sequestration), supporting differentiated carbon footprint calculations down to the crop, field, region, and management strategy level.

[0060] S5. Calculate the predicted methane emission value, the predicted nitrous oxide emission value, and the predicted carbon sequestration efficiency value based on crop-related data, the nitrous oxide emission prediction model, the methane emission prediction model, and the carbon sequestration efficiency prediction model, respectively.

[0061] S6. Calculate the predicted net carbon emissions value based on the predicted nitrous oxide emissions value, the predicted methane emissions value, the predicted carbon sequestration efficiency value, and a preset net carbon emissions model. S6 includes the following steps: S601. Calculate the greenhouse gas carbon emission forecast value based on the nitrous oxide emission forecast value and the methane emission forecast value. The calculation formula for the greenhouse gas carbon emission forecast value is: .

[0062] in, represents the predicted value of nitrous oxide emissions, Represents the predicted value of methane emissions.

[0063] S602. Calculate the predicted carbon fixation value based on the predicted carbon fixation efficiency and the carbon input value. The calculation formula for the predicted carbon fixation value is: .in, represents the predicted value of carbon sequestration efficiency, Represents the predicted value of carbon input.

[0064] S603. Calculate the predicted net carbon emissions value based on the predicted greenhouse gas carbon emissions value and the predicted carbon fixation value. The calculation formula for the predicted net carbon emissions value is: .

[0065] This step can ultimately output comprehensive assessment results including CH4 emissions, N2O emissions, carbon fixation values, and net carbon emissions. It can also support agricultural carbon management and visualization at the county, watershed, and even national scales, assisting in the formulation of carbon reduction policies and the optimization of implementation paths.

[0066] S7. Generate a recommended solution based on the predicted values ​​of nitrous oxide emissions, methane emissions, carbon sequestration, and net carbon emissions, as well as their corresponding weights. This step systematically generates a scenario combination space for carbon and nitrogen inputs and management measures, using farmland in the planting area as the analysis unit. Combined with the output of the regionally differentiated prediction model, this step achieves personalized low-carbon farmland management optimization through a scoring mechanism driven by multi-objective standardization and user preference functions. Specifically, S7 includes the following steps: S701. For any variety in any planting area, construct a scenario matrix based on continuous variables, categorical variables, predicted values ​​of nitrous oxide emissions, predicted values ​​of methane emissions, predicted values ​​of carbon fixation, and predicted values ​​of net carbon emissions to comprehensively simulate the carbon and nitrogen responses under different inputs and management strategies.

[0067] Continuous variables included carbon input and nitrogen input, and categorical variables included tillage method, number of fertilization applications, fertilization method, fertilizer type, and planting density.

[0068] In this example, the carbon input variable is set to 200–4000 kg·C·ha⁻¹·yr⁻¹, with a step size of 200 kg, for a total of 20 levels. The nitrogen fertilizer input is set to 10–300 kg·N·ha⁻¹·yr⁻¹, with a step size of 10 kg, for a total of 30 gradients. Regarding management measures, the system defines five typical controllable variables: tillage method (no-till / conventional tillage), fertilization frequency (single / multiple), fertilization method (surface application / deep application), fertilizer type (urea, inorganic fertilizer, high-efficiency nitrogen fertilizer, organic fertilizer, green manure, biochar, straw), and planting density (low density, medium density, high density). Using a multidimensional Cartesian product method, a total of 168 combined management scenarios are generated. Based on these settings, the system forms a complete input variable combination space, containing 168 × 30 × 20 = 100,800 management scenarios. All management scenario simulations are based on the climate, soil and crop types of the user-specified plots, fixing external natural variables to ensure that the simulation results reflect the impact of management behavior itself on the indicators.

[0069] Based on the aforementioned scenario combination space, the system invokes regionally customized carbon and nitrogen indicator prediction models to simulate all indicators for each scenario. Considering the significant differences in climatic conditions, soil types, and cropping systems among agricultural ecosystems in different geographical locations, the system established regionally differentiated model libraries for four indicators: nitrous oxide emissions, methane emissions, carbon sequestration efficiency, and net carbon emissions. Specifically, for the same prediction indicator, the model structure, input variable importance, and parameter settings were independently trained and optimized across different agricultural regions, fully reflecting the heterogeneity of regional ecological responses.

[0070] For example, when constructing a nitrous oxide emission prediction model, the system distinguished between multiple ecological types, such as the Northeast corn region, the Huanghuai wheat-maize rotation region, and the southern rice-growing region. An independent neural network model was constructed for each region, and its structural depth and weight settings were trained based on the measured data input from the region. Similarly, the methane emission prediction model also uses an independent random forest model structure within the region based on the distribution of rice fields, significantly improving its adaptability to methane emissions in rice-growing areas. Through the system's built-in model scheduling mechanism, the system can automatically match the corresponding indicator model according to the region to which the plot belongs, ensuring that all scenario simulation results have regional adaptability and practical reliability.

[0071] S702. For any scenario in the scenario matrix, the predicted values ​​of nitrous oxide emissions, methane emissions, carbon fixation, and net carbon emissions are normalized as positive indicators to obtain four indicators. Since the four indicators N2O, CH4, Carbon_Sequestration, and Net_GHG have significant differences in dimension, direction, and scale, the system normalizes all simulation outputs and uniformly maps them to the [0,1] interval. In order to adapt to the optimization algorithm, all indicators are standardized as positive indicators with the principle of "the larger the value, the better". In specific operations, for N2O, CH4, and Net_GHG, the reverse normalization method is used, and for Carbon_Sequestration, the forward normalization is directly performed. For the Carbon_Sequestration indicator, the normalization formula is: Carbon_Sequestration norm =(XX min ) / (Xmax-X min For the other three indicators (taking N2O as an example), the standardized formula is: N2O norm =(X max -X) / (X max -X min ). Among them, X is the predicted value of the current indicator, X min is the minimum value of all predicted values ​​of the current indicator, X max is the maximum value of all predicted values ​​of the current indicator. According to the normalization process, the normalized values ​​of the four indicators are obtained respectively: N2O norm 、CH 4norm 、Carbon_Sequestration norm 、Net_GHG norm Through this standardization mechanism, the system achieves the comparability and weighting of multi-objective indicators under a unified evaluation system, laying the foundation for subsequent personalized optimization.

[0072] S703. Construct a comprehensive scoring model based on four indicators and their weights. After completing all scenario simulations and indicator normalization, further introduce a user-driven target preference function to achieve personalized optimization recommendations under multiple objectives. Users can set the weights of each indicator according to actual management needs, including N2O emission weight w1, CH4 emission weight w2, Carbon_Sequestration weight w3, and Net_GHG weight w4. The weight value range is [0,1] and the sum is 1. Therefore, the expression of the comprehensive scoring model is: .

[0073] S704: Generate a recommended solution with the corresponding weighted combination based on the maximum comprehensive score. Each management scenario is scored according to the aforementioned scoring function, with higher scores representing better performance under the current preference settings. Ultimately, the highest-scoring scenario is selected from all 100,800 management scenarios and recommended as the optimal management solution for the current plot and objectives, achieving an organic integration of regional adaptability and user-orientation.

[0074] It should be noted that for different planting areas, due to different relevant data, the relevant models and their parameters also have different situations, but since the prediction process is the same, they will not be elaborated on.

[0075] Based on the same inventive concept, the present invention also provides a farmland net carbon emission prediction system, comprising: The acquisition module is used to obtain crop-related data, nitrous oxide emission measured data, methane emission measured data and carbon sequestration efficiency measured data of different crops in different planting areas, and divide them into training sets and validation sets; crop-related data include crop type data, management measures data, soil data and climate data.

[0076] The first model building module is used to build a nitrous oxide emission prediction model based on crop-related data and actual nitrous oxide emission data.

[0077] The second model building module is used to build a methane emission prediction model based on machine learning based on crop-related data and actual methane emission data.

[0078] The third model construction module is used to construct a carbon sequestration efficiency prediction model based on machine learning based on crop-related data and measured carbon sequestration efficiency data.

[0079] The prediction module is used to calculate the methane emission prediction value, the nitrous oxide emission prediction value and the carbon sequestration efficiency prediction value respectively based on crop-related data, the nitrous oxide emission prediction model, the methane emission prediction model and the carbon sequestration efficiency prediction model.

[0080] The calculation module is used to calculate the net carbon emission prediction value based on the nitrous oxide emission prediction value, the methane emission prediction value, the carbon sequestration efficiency prediction value and the preset net carbon emission model.

[0081] The recommendation module is used to generate a recommendation plan based on the predicted value of nitrous oxide emissions, the predicted value of methane emissions, the predicted value of carbon fixation, the predicted value of net carbon emissions and their corresponding weights.

[0082] Based on the same inventive concept, the present invention also provides a device for predicting net carbon emissions from farmland, comprising a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the steps of the above method.

[0083] The above are only preferred embodiments of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for predicting net carbon emissions from farmland, characterized in that: The steps include: S1. Obtain crop-related data, measured nitrous oxide emission data, measured methane emission data, and measured carbon sequestration efficiency data for different crops in different planting areas, and divide the data into training and validation sets; the crop-related data include crop type data, management measures data, soil data, and climate data; S2. Constructing a nitrous oxide emission prediction model based on machine learning based on the crop-related data and the measured nitrous oxide emission data; S3. Constructing a methane emission prediction model based on machine learning based on the crop-related data and the measured methane emission data; S4. Constructing a carbon sequestration efficiency prediction model based on machine learning based on the crop-related data and the measured carbon sequestration efficiency data; S5. Calculating a predicted methane emission value, a predicted nitrous oxide emission value, and a predicted carbon sequestration efficiency value, respectively, based on the crop-related data, the nitrous oxide emission prediction model, the methane emission prediction model, and the carbon sequestration efficiency prediction model; S6. Calculate a net carbon emission prediction value based on the nitrous oxide emission prediction value, the methane emission prediction value, the carbon sequestration efficiency prediction value, and a preset net carbon emission model.

2. A farmland net carbon emission prediction method according to claim 1, characterized in that: The S1 further includes: preprocessing the crop-related data, performing integer encoding processing on the categorical variables, normalizing all variables, and using the processed data as basic data for dividing the training set and the validation set.

3. A method for predicting net carbon emissions from farmland according to claim 1, characterized in that: The S2 comprises the following steps: S201. Construct a nonlinear emission response function based on nitrogen input, N2O emission value, natural N2O emission coefficient under no nitrogen input conditions, and sensitivity factor of N2O emission to nitrogen application rate; S202, constructing a twin neural network, and using the crop-related data and the measured nitrous oxide emission data in the training set and the validation set to determine the values ​​of the natural N2O emission coefficient and the sensitivity factor; S203: Substitute the values ​​of the natural N2O emission coefficient and the sensitivity factor into the nonlinear emission response function to obtain the nitrous oxide emission prediction model.

4. A method for predicting net carbon emissions from farmland according to claim 3, characterized in that: The expression of the nonlinear emission response function is: ; in, Indicates N2O emission value, represents the nitrogen input value, a represents the natural N2O emission coefficient under the condition of no nitrogen input, and b represents the sensitivity factor of N2O emission to the amount of nitrogen applied.

5. The method for predicting net carbon emissions from farmland according to claim 1, wherein: The S3 further comprises the following steps: S301. Setting a candidate methane emission algorithm and its key parameters; S302, performing cross-validation training on the candidate methane emission algorithms based on the crop-related data and the measured methane emission data in the training set; S303: Using preset evaluation indicators to determine the best methane emission candidate algorithm as the methane emission prediction model.

6. A method for predicting net carbon emissions from farmland according to claim 1, characterized in that: The S4 comprises the following steps: S401. Setting a candidate carbon sequestration efficiency algorithm and its key parameters; S402, performing cross-validation training on the carbon sequestration efficiency candidate algorithms based on the crop-related data and the measured carbon sequestration efficiency data in the training set; S403. Using preset evaluation indicators, determine the best candidate algorithm as the carbon sequestration efficiency prediction model.

7. A method for predicting net carbon emissions from farmland according to claim 1, characterized in that: The S6 comprises the following steps: S601. Calculate a greenhouse gas carbon emission prediction value based on the nitrous oxide emission prediction value and the methane emission prediction value; S602, calculating a carbon fixation prediction value based on the carbon fixation efficiency prediction value and the carbon input value; S603: Calculate a net carbon emission prediction value based on the greenhouse gas carbon emission prediction value and the carbon fixation prediction value.

8. A method for predicting net carbon emissions from farmland according to claim 7, characterized in that: The calculation formula for the predicted net carbon emissions is: ； The calculation formula for the predicted greenhouse gas carbon emissions is: ; The formula for calculating the predicted carbon sequestration value is: ; in, represents the predicted value of nitrous oxide emissions, represents the predicted value of methane emissions, represents the predicted value of carbon sequestration efficiency, Represents the carbon input value.

9. A method for predicting net carbon emissions from farmland according to claim 1, characterized in that: Also includes: S7. Generating a recommendation plan based on the predicted nitrous oxide emission value, the predicted methane emission value, the predicted carbon fixation value, the predicted net carbon emission value, and the corresponding weights, including the following steps: S701. For any variety in any planting area, construct a scenario matrix based on continuous variables, categorical variables, predicted nitrous oxide emissions, predicted methane emissions, predicted carbon sequestration, and predicted net carbon emissions; the continuous variables include carbon input and nitrogen input, and the categorical variables include tillage method, number of fertilizations, fertilization method, fertilizer type, and planting density; S702: performing positive indicator normalization processing on the predicted nitrous oxide emissions, predicted methane emissions, predicted carbon fixation, and predicted net carbon emissions of any scenario in the scenario matrix to obtain four indicators; S703. Construct a comprehensive scoring model based on the four indicators and their weights; S704: Generate a recommendation plan with a corresponding weight combination based on the maximum comprehensive score.

10. A farmland net carbon emission prediction system, characterized in that: include: An acquisition module is used to obtain crop-related data, measured nitrous oxide emission data, measured methane emission data, and measured carbon sequestration efficiency data for different crops in different planting areas, and divide the data into training sets and validation sets; the crop-related data includes crop type data, management measures data, soil data, and climate data; A first model building module is used to build a nitrous oxide emission prediction model based on machine learning according to the crop-related data and the measured nitrous oxide emission data; A second model building module is used to build a methane emission prediction model based on machine learning according to the crop-related data and the measured methane emission data; A third model building module is used to build a carbon sequestration efficiency prediction model based on machine learning according to the crop-related data and the measured carbon sequestration efficiency data; A prediction module, configured to calculate a methane emission prediction value, a nitrous oxide emission prediction value, and a carbon sequestration efficiency prediction value, respectively, based on the crop-related data, the nitrous oxide emission prediction model, the methane emission prediction model, and the carbon sequestration efficiency prediction model; A calculation module is used to calculate the net carbon emission prediction value based on the nitrous oxide emission prediction value, the methane emission prediction value, the carbon sequestration efficiency prediction value and a preset net carbon emission model.

11. A farmland net carbon emission prediction system according to claim 10, characterized in that: Also includes: A recommendation module is used to generate a recommendation plan based on the nitrous oxide emission prediction value, the methane emission prediction value, the carbon fixation prediction value, the net carbon emission prediction value and the corresponding weights.

12. A device for predicting net carbon emissions from farmland, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to implement the steps of the method according to claim 1.

Images