Net ecosystem exchange simulation method based on double-disturbance convolutional neural network

Abstract

The invention provides a net ecosystem exchange simulation method based on a double-disturbance convolutional neural network, and relates to the technical field of ecosystem carbon cycle simulation, and the method comprises the steps: obtaining meteorological environment parameters and vegetation biophysical indexes, and constructing a feature variable set; a double-disturbance convolutional neural network model is constructed, the double-disturbance convolutional neural network model comprises a double-disturbance module and a convolution module, the double-disturbance module generates disturbance data of characteristic variables through a double-disturbance mechanism of Gaussian noise and Gaussian blur, and disturbance intensity is dynamically adjusted through a dynamic mask matrix; the convolution module is provided with two layers of one-dimensional convolution operations, and outputs through a linear layer; and training and verifying the model by using the characteristic variable set to obtain a trained model, and outputting a net ecosystem exchange simulation result. According to the invention, accurate simulation of the NEE change process under different ecological systems is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of ecosystem carbon cycle simulation, and in particular to a method for simulating net ecosystem exchange based on a double perturbation convolutional neural network. Background Art

[0002] Net ecosystem exchange (NEE) is a key indicator for measuring the carbon exchange situation of an ecosystem, reflecting the difference between the carbon fixation process of plants and the carbon release process of plant respiration within the ecosystem. Essentially, NEE represents the difference between the total amount of carbon input and the total amount of carbon output within the ecosystem. The accurate simulation of NEE is of great significance for evaluating the carbon balance of the ecosystem, exploring the carbon cycle mechanism of the ecosystem, and formulating strategies to address climate change. However, due to the diversity and spatial heterogeneity of land cover types in terrestrial ecosystems, accurately simulating NEE remains challenging.

[0003] In recent years, the NEE of an ecosystem can be measured by using micrometeorological techniques (i.e., the eddy covariance method or the eddy correlation method). The eddy correlation method is one of the current standard methods for measuring NEE and is widely regarded as a classic method in micrometeorology. Although the eddy correlation technology has achieved a major breakthrough in carbon flux observation technology and directly observed NEE from a meteorological perspective for the first time, this method still has many limitations, such as strict requirements for terrain, susceptibility to meteorological conditions, easy damage of sensors in bad weather, and complex data correction and interpolation methods. Moreover, the eddy correlation technology can only perform NEE observation data at the site scale, and the currently established global carbon flux observation network still cannot meet the research on NEE estimation at large spatio-temporal scales. Therefore, the research on NEE based on data-driven methods is an important research direction at present.

[0004] The research on NEE can be mainly divided into the eddy covariance system method, the process-based model method, and the data-driven model method. The eddy covariance (EC) technology can directly observe the CO flux between the land-atmosphere ecosystem, and can obtain continuous flux data without disturbing the microenvironment of the soil surface at the research site. At present, the eddy covariance technology has been widely used in wetland, grassland, farmland, forest, and shrub ecosystems. However, the observation range of the eddy covariance technology is very limited, and there are large uncertainties in the observed data, so data revision is required. Process models usually rely on the prior knowledge of domain experts for construction and adjustment, which may make the models affected by human factors. Some complex real-world problems are difficult to accurately model through simple mathematical formulas and rules. Since process models rely on specific domain knowledge and formulas, their generalization ability is usually limited, and the performance of the models may be relatively poor when facing new situations or data. The data-driven method plays a crucial role in NEE simulation. More and more studies use machine learning and deep learning models to simulate NEE.

[0005] Although deep learning has obvious advantages in time series data modeling, it is still in its infancy in NEE simulation. Traditional machine learning models still dominate, and the modeling potential of deep neural networks for carbon flux data has not been fully explored. Existing NEE simulation models are mostly specific to a certain location and have poor adaptability to different ecosystems. This method is effective within a local range, but it cannot meet the model requirements of scalability and generalizability. Although the data-driven method improves the numerical simulation accuracy, its "black box" characteristic leads to insufficient analysis of the feature importance in the simulation process. Summary of the Invention

[0006] The purpose of the present invention is to: in order to solve the problem of the lack of NEE simulation models for different types of ecosystems, a net ecosystem exchange simulation method based on a double perturbation convolutional neural network is proposed, including the following steps:

[0007] S1. Obtain meteorological environment parameters and vegetation biophysical indicators, and construct a feature variable set;

[0008] S2. Construct a double perturbation convolutional neural network model, including a double perturbation module and a convolutional module. The double perturbation module generates perturbation data of the feature variables through a double perturbation mechanism of Gaussian noise and Gaussian blur, and dynamically adjusts the perturbation intensity using a dynamic mask matrix; the convolutional module sets two layers of one-dimensional convolutional operations and outputs through a linear layer;

[0009] S3. Use the feature variable set to train and validate the model, obtain the trained model, and output the net ecosystem exchange simulation result.

[0010] The present invention also provides a computer-readable storage medium storing a computer program, which when executed by a processor, implements the above-mentioned net ecosystem exchange simulation method based on a dual-perturbation convolutional neural network.

[0011] The present invention also provides an electronic device, including a processor and a memory, the processor being connected to the memory, wherein the memory is configured to store a computer program including computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the above-mentioned net ecosystem exchange simulation method based on a dual-perturbation convolutional neural network.

[0012] The present invention also provides a computer program product, including a computer program / instructions, characterized in that the computer program / instructions, when executed by a processor, implement the steps of the above-mentioned net ecosystem exchange simulation method based on a dual-perturbation convolutional neural network.

[0013] The beneficial effects brought by the technical solution provided by the present invention are as follows:

[0014] The present invention proposes an NEE simulation model based on a dual-perturbation convolutional neural network. Based on deep learning technology, by introducing a dual-perturbation mechanism of Gaussian noise and Gaussian blur, the on-site measured data and remote sensing data are organically integrated, comprehensively considering vegetation indices, meteorological indicators, and other relevant variables, and combining a learnable dynamic mask matrix and a convolutional neural network structure, accurately simulating the NEE change process under different ecosystems, which is a unified model applicable to multiple ecosystems, providing a more comprehensive and reliable decision-making basis for ecosystem carbon cycle research and management. Description of the Drawings

[0015] Figure 1 is a flowchart of the net ecosystem exchange simulation method based on a dual-perturbation convolutional neural network according to an embodiment of the present invention;

[0016] Figure 2 is a comparison chart of predicted values and true values of the test set of the model DPA-CNN of the present invention and other benchmark models (random forest, LightGBM, RNN, LSTM, CNN-LSTM, Transformer) in a mixed data scenario;

[0017] Figure 3 is the RMSE, MAE, and R of the model DPA-CNN of the present invention and other benchmark models (random forest, LightGBM, RNN, LSTM, CNN-LSTM, Transformer, SPA-CNN) in a mixed data scenario 2Result comparison radar chart; among them, Figure A is the radial distribution of RMSE (blue line) and MAE (orange line) of different models, and Figure B is the distribution of R 2 value;

[0018] Figure 4 It is the comparison of R2 performance of the DPA-CNN model of the present invention and multiple deep learning models (LSTM, Transformer and traditional CNN) on two time scales of daily data and monthly data;

[0019] Figure 5 It is the SHAP important feature analysis honeycomb chart of all stations when the DPA-CNN model in the embodiment of the present invention simulates monthly NEE;

[0020] Figure 6 It is the block diagram of an electronic device in an exemplary embodiment of Embodiment 1 of the present invention. Detailed implementation manners

[0021] To make the objectives, technical solutions and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0022] The flowchart of the net ecosystem exchange simulation method based on the double perturbation convolutional neural network (CNN) in the embodiment of the present invention is as Figure 1 , and specifically includes the following steps:

[0023] S1. Construct a multi-dimensional data framework, obtain meteorological environment parameters and vegetation biophysical indicators, construct a set of characteristic variables, and obtain a data set for NEE simulation.

[0024] The selected characteristic variables include meteorological factors: precipitation P, temperature Ta, humidity Ma; radiation parameters: shortwave radiation Rd, net radiation Rn, photosynthetically active radiation PAR, photosynthetically active radiation fraction F par ; soil property: soil temperature T soil ; and vegetation indices: normalized difference vegetation index NDVI, enhanced vegetation index EVI. Using the data preprocessing method, filter out the complete and available characteristic variable time series. The specific process is as follows: First, clarify the screening criteria and requirements, determine the data quality requirements according to the research objectives and data needs, and then, according to the predefined screening criteria and data characteristic judgments, screen the data point by time point.

[0025] Adopt a hierarchical quality control system for the above remote sensing data to solve the problem of inconsistent spatio-temporal resolution between data from different sources. Specifically:

[0026] (1) Through the spatial resampling algorithm of the Google Earth Engine platform, heterogeneous data such as NDVI (250 meters) and FPAR (500 meters) are unified to a 1000-meter standard grid.

[0027] (2) In the time series data synthesis stage, the maximum value synthesis method is used to remove cloud and fog noise. Its monthly-scale synthesis strategy smooths the interference caused by short-term environmental fluctuations while retaining the vegetation growth peak.

[0028] (3) The meteorological and radiation data of the remaining data sources are cropped and resampled (bilinear) to achieve matching and alignment in terms of spatial scale.

[0029] (4) For NEE data, missing values are inserted, duplicate values are removed, outliers are processed, and units are unified using adjacent values in time or space.

[0030] (5) Based on Pearson correlation analysis, a feature correlation matrix is constructed to explore the influence mechanism of the selected feature variables on the net ecosystem exchange and the associations between variables.

[0031] (6) In the NEE simulation, time window slicing is performed independently for each site, data from all sites are merged, non-target variables are globally normalized, and spatial arrangement biases are eliminated by randomly shuffling the sample order to obtain a dataset for NEE simulation.

[0032] S2. Construct a double perturbation convolutional neural network model, including a double perturbation module and a convolutional module. The double perturbation module generates perturbation data of feature variables through a double perturbation mechanism of Gaussian noise and Gaussian blur, and dynamically adjusts the perturbation intensity using a dynamic mask matrix; the convolutional module sets two one-dimensional convolutional operations and outputs through a linear layer. Among them, a ReLU activation function is also set after the one-dimensional convolution.

[0033] The core idea of the double perturbation convolutional neural network model is to enhance the model's ability to identify significant features by introducing two complementary data perturbation mechanisms and generating perturbation data through the double perturbation mechanism of Gaussian noise and Gaussian blur.

[0034] Gaussian noise perturbation forces the model to learn insensitivity to instantaneous noise by simulating errors or environmental fluctuations. Given the input time series data matrix where B is the batch size, T is the time step, and D is the feature dimension, the Gaussian noise perturbation is defined as:

[0035]

[0036] where, X noise represents the set of feature variables after Gaussian noise perturbation, X represents the set of feature variables, ∈ represents the probability density function of Gaussian noise, It is indicated that ∈ follows a Gaussian distribution with a mathematical expectation of 0 and a variance of where σ1 controls the noise intensity and is used to simulate errors or environmental fluctuations. Such perturbations force the model to learn insensitivity to instantaneous noise.

[0037] The Gaussian blur perturbation extracts the smooth trend in the time dimension through a convolution operation. This operation strengthens the model's perception of the continuous change of monthly meteorological elements by weighted averaging the information of neighboring time steps. The expression is:

[0038]

[0039] where, X blur represents the set of feature variables after Gaussian blur perturbation, X represents the set of feature variables, and * represents the convolution operation. represents a one-dimensional Gaussian kernel with a standard deviation of σ2. represents a one-dimensional Gaussian kernel with a standard deviation of σ2 for the variable t, and t represents a random variable.

[0040] The learnable dynamic mask matrix is used to dynamically adjust the perturbation intensity, collaboratively optimize the feature importance and perturbation pattern selection. A learnable dynamic mask matrix M ∈ [0, 1] T×D is introduced to achieve dynamic weight allocation. Here, T represents the time step of the time series data matrix input to the double-perturbation convolutional neural network model, D represents the feature dimension of the time series data matrix input to the double-perturbation convolutional neural network model, and the mixed feature matrix is calculated in the following form:

[0041]

[0042] where, represents the mixed feature matrix, ⊙ represents element-wise multiplication, X noise represents the set of feature variables after Gaussian noise perturbation, and X blur represents the set of feature variables after Gaussian blur perturbation.

[0043] The mask value M t,d is normalized by the sigmoid function, representing the dependence degree of the corresponding spatio-temporal position on the noise perturbation features. At the initial stage, M = 0.5 is set to achieve uniform mixing of noise and blurred features, and it is dynamically adjusted through gradient descent driven by training data during the training process. This design enables each time step-feature pair to have an independent mixing weight. For example, when M t,d → 1, the model tends to retain the Gaussian noise features at this position, while when M t,d → 0, it relies on the blurred trend information.

[0044] The perturbed feature set is weighted and mixed according to the mask value to generate the final input features for subsequent convolutional neural network modeling.

[0045] Feature extraction and modeling are performed through the convolutional neural network part to complete end-to-end time series modeling, specifically as follows:

[0046] (1) The input data undergoes two layers of one-dimensional convolutional operations. The number of channels of the convolutional kernel in the first layer is expanded to 32 dimensions, and the second layer is further expanded to 64 dimensions, and the non-linear activation function is used to achieve the gradual abstraction of local features layer by layer;

[0047] (2) After the feature tensor is flattened, the dimensionality reduction is completed through a fully connected layer, and finally the single-value regression simulation result is output. This structure can effectively capture the local dependence relationship in the time dimension and achieve end-to-end time series modeling.

[0048] S3. Use the set of feature variables to train and validate the model, obtain the trained model, and output the net ecosystem exchange simulation result.

[0049] To improve the generalization ability and interpretability of the model, double regularization constraints are introduced into the loss function:

[0050]

[0051] Among them, represents the loss function, λ1∥M∥1 represents the L1 regularization term, and the L1 regularization term promotes the sparsity of the mask, guiding the model to focus on key features; λ2∥MM T -I∥ F represents the orthogonal regularization term. The orthogonal regularization term avoids the coupling effect between feature dimensions by constraining the orthogonality of the mask matrix; λ1 and λ2 are weight parameters, ∥∥1 represents the L1 norm, M represents the dynamic mask matrix, ∥∥ F represents the F norm, and I represents the identity matrix.

[0052] The double perturbation hybrid strategy generates new training samples by introducing noise perturbation and Gaussian blur, enhancing the robustness of the model to different degradation modes. As a learnable part, the mask perturbation enables the model to dynamically adjust the perturbation intensity, collaboratively optimize the feature importance and perturbation mode selection, and helps the model identify more important features in a complex perturbation environment.

[0053] In the embodiment of the present invention, a three-dimensional time series data set is constructed by using a sliding time window (window_size = 8) in the data loading stage. The AdamW optimizer (initial learning rate 10 -4)With the cosine annealing schedule (T_max = 4000, eta_min = 0), it not only ensures fast convergence in the initial stage but also enables fine-tuning of parameters in the later stage. The loss function incorporates a regularization term on the basis of MSE to form a composite optimization objective. The design of an extremely long training cycle of 4000 epochs takes into account the need for sufficient iteration due to the parameter coupling between the CNN feature extractor and the saliency module. The weight decay coefficient of 10 -4 is set to effectively control the model complexity. The optimal hyperparameters are shown in Table 1.

[0054] Table 1

[0055] Hyperparameter Category Parameter Setting Random Seed seed = 42 Time Window time_window = 8 Batch Size batch_size = 128 Initial Learning Rate <![CDATA[lr=10 -4 > Training Epochs epochs = 4000 Optimizer Configuration <![CDATA[AdamW + Weight Decay 10 -4 > Learning Rate Scheduling Cosine Annealing (η_min = 0)

[0056] The model evaluation metrics Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) are selected to verify the model performance. In addition, the coefficient of determination (R2) is introduced to measure the degree of agreement between NEE prediction and observation. The trained model is obtained and the NEE simulation results are output.

[0057] A hybrid training set is constructed and multi-model comparison experiments are carried out, covering 7 benchmark models including traditional machine learning (Random Forest (RF), LightGBM), classical deep learning (RNN, LSTM, CNN), and hybrid architectures (CNN-LSTM, Transformer). At the same time, an ablation experiment (SPA-CNN) is introduced to explore the contribution of the dual perturbation module. SPA-CNN is a model formed by embedding the SPA (Spatial Pyramid Attention Network) module into CNN. In the hybrid data scenario, the experimental results of all models are shown in Table 2.

[0058] Table 2

[0059] Model Name RMSE MAE R2 RNN 55.1089 28.2236 0.7372 LSTM 54.6859 26.9516 0.7412 CNN 41.7884 26.1281 0.8422 CNN-LSTM 53.7128 31.0190 0.7504 Transformer 51.9842 29.2532 0.7662 DPA-CNN 41.0179 25.1465 0.8544 SPA-CNN 42.4143 26.8158 0.8443 RF 54.1777 27.8581 0.7247 LightGBM 58.3051 29.9748 0.6812

[0060] The performance comparison of the model DPA-CNN of the present invention with other benchmark models (Random Forest, LightGBM, RNN, LSTM, CNN-LSTM, Transformer) in the hybrid data scenario is as Figure 2 shown.

[0061] The results of RMSE, MAE and R of the model DPA-CNN of the present invention and other benchmark models (Random Forest, LightGBM, RNN, LSTM, CNN-LSTM, Transformer, SPA-CNN) in the hybrid data scenario are compared as 2 shown. Figure 3 shown. Figure 3 The radar chart inFigure 3 In Figure A of Figure 3 , the radial distributions of RMSE (blue line) and MAE (orange line) of different models clearly show the differences in simulation errors among the models. The closer the line is to the center of the circle, the better the error index performance of the model. The starting value of the radial scale of this figure is 10, and the interval is 10 units. And 2 Figure B in 2 shows the distribution of the R 2 value (blue line). The closer the R 2 value is to the outer circle, the better the fitting effect of the model on the data. The starting value of the radial scale of this figure is 0.1, and the interval is 0.1 unit. Compared with the LightGBM model with the worst effect, the simulation results of DPA-CNN are better. RMSE is reduced by 29.45%, MAE is reduced by 16.10%, and R 2 is increased by 25.29%. These results indicate that the DPA-CNN model is superior to the LightGBM model in capturing complex nonlinear relationships and dependencies in time series data. Compared with LSTM, the RMSE of DPA-CNN is reduced by 24.99%, the MAE is reduced by 6.70%, and R 2 is increased by 15.27%. In addition, compared with CNN-LSTM representing the hybrid model, the RMSE of DPA-CNN is reduced by 23.64%, the MAE is reduced by 18.93%, and R Figure 2 is increased by 13.84%. This shows that the introduction of the double perturbation module enables DPA-CNN to better perform global modeling while considering local features and time series dependencies. Although the Transformer model is slightly lower than CNN-LSTM in various indicators, it still shows stronger capabilities compared to models such as RNN, LSTM, RandomForest, and LightGBM. The Random Forest model has been widely used in previous NEE simulation studies. Compared with it, DPA-CNN also shows obvious advantages, where RMSE is reduced by 24.37%, MAE is reduced by 9.73%, and R Figure 2 is increased by 17.35%. With its deep learning architecture and double perturbation module, DPA-CNN can better capture complex patterns in the data and further improve the prediction accuracy. Figure 2Each subfigure shows the comparison between the test set predictions and true values of different comparison models. The scatter points in each subfigure represent the corresponding relationship between the actual simulation values and true values. The red dashed line represents the y = x line. If all points are on this line, it indicates a high degree of fitting between the simulation values and true values. The results prove that compared with traditional machine learning simulations, deep learning models, especially DPA-CNN, have better accuracy improvement in the NEE simulation task, demonstrating the potential of deep learning models in the ecological field.

[0062] To reveal the ability of deep learning models to capture ecological characteristics at multiple time granularities, based on data of four types of ecosystems (forest, grassland, farmland, shrub), the present invention systematically compares the simulation performance differences of the double perturbation convolutional neural network (DPA-CNN), time series models (LSTM, Transformer), and traditional CNN of the present invention at daily and monthly scales. Figure 4 Shows the R 2 performance comparison of the DPA-CNN model of the present invention and multiple deep learning models (LSTM, Transformer, and traditional CNN) at two time scales. The results show that DPA-CNN shows obvious advantages in the daily-monthly cross-scale simulation. Its R 2 in the grassland monthly scale reaches 0.9643, but in the forest daily scale, LSTM exceeds DPA-CNN (0.6091) with an R 2 of 0.6385, revealing the law of the model performance being regulated by the "time-ecology" two-dimensionality.

[0063] Figure 5 is the SHAP (SHapley Additive exPlanations) important feature analysis honeycomb diagram of all stations when the DPA-CNN model of the present invention simulates monthly NEE. It can be seen that the SHAP value distribution ranges of the normalized difference vegetation index NDVI and enhanced vegetation index EVI features are relatively wide, indicating that they have a greater impact on the model's simulation of NEE. When the EVI and NDVI values are high (red), the SHAP values are mostly negative, indicating that high vegetation indices will reduce the NEE values simulated by the model, which is consistent with the role of vegetation in the carbon cycle. On the other hand, the SHAP value distributions of some features such as Rn are relatively concentrated and close to 0, indicating that these features have relatively little impact on the model output.

[0064] In an exemplary embodiment, it includes a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned net ecosystem exchange simulation method based on a double perturbation convolutional neural network.

[0065] Please refer to Figure 6, in an exemplary embodiment, it further includes an electronic device, which includes at least one processor, at least one memory, and at least one communication bus.

[0066] Among them, a computer program is stored on the memory. The computer program includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the above-mentioned net ecosystem exchange simulation method based on the double-disturbance convolutional neural network.

[0067] In an exemplary embodiment, it further includes a computer program product, including a computer program / instructions, characterized in that when the computer program / instructions are executed by a processor, the steps of the above-mentioned net ecosystem exchange simulation method based on the double-disturbance convolutional neural network are implemented.

[0068] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for simulating net ecosystem exchange based on a double-disturbance convolutional neural network, characterized in that It includes the following steps: S1. Obtain meteorological environment parameters and vegetation biophysical indicators, and construct a set of characteristic variables; S2. Construct a double-perturbation convolutional neural network model, including a double-perturbation module and a convolutional module. The double-perturbation module generates perturbation data of the characteristic variables through a double-perturbation mechanism of Gaussian noise and Gaussian blur, and dynamically adjusts the perturbation intensity by using a dynamic mask matrix; the convolutional module sets two layers of one-dimensional convolutional operations and outputs through a linear layer; S3. Use the set of characteristic variables to train and validate the model, obtain the trained model, and output the simulated results of net ecosystem exchange.

2. The method for simulating net ecosystem exchange based on a double-disturbance convolutional neural network according to claim 1, wherein The characteristic variables include: meteorological factors, radiation parameters, soil properties, and vegetation indices.

3. A net ecosystem exchange simulation method based on a double-disturbance convolutional neural network according to claim 1, characterized in that, The Gaussian noise perturbation is expressed as: Among them, X noise represents the set of feature variables after Gaussian noise perturbation, X represents the set of feature variables, and ∈ represents the probability density function of Gaussian noise. represents that ∈ follows a Gaussian distribution with a mathematical expectation of 0 and a variance of ​ 4. A net ecosystem exchange simulation method based on a double perturbation convolutional neural network according to claim 1, characterized in that The Gaussian blur perturbation is expressed as: Among them, X blur represents the set of feature variables after Gaussian blur perturbation, X represents the set of feature variables, and * represents the convolution operation. represents a one-dimensional Gaussian kernel with a standard deviation of σ2. represents a one-dimensional Gaussian kernel with a standard deviation of σ2 for the variable t, where t represents a random variable.

5. A method for simulating net ecosystem exchange based on a double-disturbance convolutional neural network according to claim 1, characterized in that, The dynamic mask matrix is represented as: M ∈ [0, 1] T×D , where T represents the time step of the time series data matrix input to the double-perturbation convolutional neural network model, and D represents the feature dimension of the time series data matrix input to the double-perturbation convolutional neural network model; Dynamically adjusting the perturbation intensity by using a dynamic mask matrix is expressed as: Among them, represents the mixed feature matrix, ⊙ represents element-wise multiplication, and X noise represents the set of feature variables after Gaussian noise perturbation, and X blur represents the set of feature variables after Gaussian blur perturbation.

6. A net ecosystem exchange simulation method based on a double-disturbance convolutional neural network according to claim 1, characterized in that, The loss function for model training is: Among them, denotes the loss function, λ1∥M∥1 denotes the L1 regularization term, and λ2∥MM T -I∥ F denotes the orthogonal regularization term, λ1 and λ2 are weight parameters, ∥∥1 denotes the L1 norm, M denotes the dynamic mask matrix, and ∥∥ F denotes the F norm, and I denotes the identity matrix.

7. A net ecosystem exchange simulation system based on a dual-perturbation convolutional neural network, characterized in that, It includes: A characteristic variable construction module, which is used to obtain meteorological environment parameters and vegetation biophysical indicators and construct a set of characteristic variables; A model construction module, which is used to construct a double-perturbation convolutional neural network model, generate perturbation data through a double-perturbation mechanism of Gaussian noise and Gaussian blur, and dynamically adjust the perturbation intensity by using a dynamic mask matrix; A result output module, which is used to train and validate the model by using the set of characteristic variables, obtain the trained model, and output the simulated results of net ecosystem exchange.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method according to any one of claims 1-6.

9. An electronic device, characterized in that, It includes a processor and a memory, and the processor is connected to the memory. Among them, the memory is used to store a computer program, the computer program includes computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the method according to any one of claims 1-6.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by a processor, it implements the steps of the method according to any one of claims 1-6.

Images