A method for decomposing CO2 flux observed by eddy covariance based on hybrid deep neural networks
By constructing a hybrid deep neural network module to independently estimate GPP and ER, the problems of expensive equipment and high professional skill requirements in existing technologies are solved, achieving high-precision CO2 flux splitting and providing convenient carbon cycle data.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-23
- Publication Date
- 2026-03-13
AI Technical Summary
Existing technologies cannot directly measure the CO2 uptake (GPP) and CO2 release (ER) from plant photosynthesis and respiration, and common separation methods require expensive equipment and highly specialized skills, and fail to fully consider the influence of various environmental factors.
A hybrid deep neural network was used to construct four independent deep neural network modules, which independently estimated GPP and ER using meteorological, soil and vegetation data, and achieved high-precision decomposition by training the neural network model.
It achieves high-precision estimation of GPP and ER without the need for expensive equipment and professional skills. The calculation process does not depend on specific assumptions and provides efficient and convenient carbon cycle change data.
Smart Images
Figure CN117347556B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ecological environment monitoring technology, specifically a method for decomposing CO2 flux based on eddy covariance observation using a hybrid deep neural network. Background Technology
[0002] Eddy covariance (EC) technology is widely used in the observation of surface water, heat, and CO2 fluxes. Observational data obtained through EC technology helps us understand regional land-atmosphere exchange and energy balance processes. CO2 flux observed through EC technology is the net ecosystem exchange (NEE) of CO2 between the ecosystem and the atmosphere. NEE mainly consists of the difference between total CO2 released by respiration (ER) and gross primary productivity (GPP) of CO2 absorbed by plants through photosynthesis; that is, the difference in CO2 flux between plants and the atmosphere. Estimating GPP and ER can provide a better understanding of ecosystem function and also provide quantitative data for the construction and validation of land surface carbon cycle models. Since EC technology cannot directly measure GPP and ER, current methods use experimental methods or empirical formulas to decompose the NEE measured by EC into GPP and ER.
[0003] The experimental method mainly refers to simultaneously measuring GPP or ER using other instruments while using EC technology to measure NEE. For example, using a box-type soil / plant respiration chamber to measure ER, or using carbon isotope technology to determine the proportion of GPP in NEE. The advantage of the experimental method is that it can accurately measure GPP or ER. The disadvantage is that after purchasing the EC system, it is still necessary to purchase expensive experimental equipment, and it also requires the experimental operators to have a high level of professional skills. Common empirical formulas for CO2 flux splitting include: fitting a nonlinear function of nighttime temperature and reference respiration to determine the nighttime ER value, and assuming that the functional relationship obtained by this method also applies to the daytime, so the GPP value is the difference between the daytime ER and NEE; or using daytime and nighttime data to fit a light and temperature driven model to estimate ER and GPP; or the eddy variance similarity theory method based on water vapor and carbon dioxide concentration data at a frequency of 10-20Hz, etc. However, the disadvantages of these methods are: 1) the use of the above splitting methods all require meeting specific assumptions. 2) Empirical functional relationships have been well applied at the individual plant level, but whether this method can be extended to the ecosystem scale remains to be clarified. 3) The dynamic changes in ER or GPP fluxes are not only related to temperature and radiation, but are also regulated by other variables such as soil moisture content, soil temperature, vegetation height, and leaf temperature. However, the currently established functional relationships do not fully consider the influence of multiple environmental factors. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention proposes a method for decomposing CO2 flux from eddy covariance observations based on a hybrid deep neural network. This method utilizes deep learning algorithms to construct four sub-deep neural network modules, employing different combinations of environmental driving factors to independently estimate GPP and ER, thereby achieving more accurate estimates of GPP and ER. Applying this method helps researchers better understand changes in the atmospheric carbon cycle.
[0005] The technical solution of this invention to solve the aforementioned technical problem is: designing a method for decomposing CO2 flux observed by eddy covariance based on a hybrid deep neural network, characterized in that the method includes the following steps:
[0006] S1: Acquire relevant meteorological, vegetation, and soil data for the target area with vegetation cover. This relevant data is obtained through monitoring every half hour. The meteorological data includes air temperature T. a Net radiation R a , saturated vapor pressure difference VPD, horizontal wind speed W s Wind direction W d Carbon dioxide flux (NEE) m EC technology observation of nighttime ecosystem respiration ER nRelevant soil data include topsoil moisture content (SM) and soil temperature (T). s Vegetation-related data include leaf area index (LAI) and leaf temperature (T). v Plant height h, vegetation transpiration T m The dataset obtained at each monitoring time point is divided into a data sample, and n data samples are obtained based on the total monitoring duration; NEE in the data samples m With T m Observations were obtained using EC technology;
[0007] S2: Construct a neural network estimation model
[0008] The neural network estimation model is a hybrid deep neural network model, comprising four independent deep neural networks (DNNs). a DNN b DNN p DNN T Deep Neural Networks (DNNs) a DNN b DNN p DNN T Each network contains an input layer, hidden layers, and an output layer. The number of hidden layers and the number of neurons in each layer are not shared, and the weights and biases during data transmission are also not shared.
[0009] Deep Neural Networks (DNNs) a Used to obtain estimated vegetation canopy respiratory flux (ER) a Deep Neural Networks (DNNs) b Used to obtain estimated soil respiration flux ER b Deep Neural Networks (DNNs) p Deep neural networks (DNNs) are used to estimate the photosynthetic flux (GPP) of plants. T Used to obtain estimated vegetation transpiration T p ;
[0010] Deep Neural Networks (DNNs) a The input is T a T v SM, W s W d LAI, h and ER n ;
[0011] Deep Neural Networks (DNNs) b The input is T a T s SM, W s and ER n ;
[0012] Deep Neural Networks (DNNs)p The input is T a R a SM, W s W d VPD, LAI, and h;
[0013] Deep Neural Networks (DNNs) T The inputs are GPP and T. a R a SM, VPD, W s , LAI and h;
[0014] Deep Neural Networks (DNNs) a With Deep Neural Networks (DNNs) b The sum of the outputs is the estimated ecosystem respiration ER, using a deep neural network (DNN). p The sum of the output and the estimated ER is the estimated NEE;
[0015] The output of the neural network estimation model is constrained by carbon dioxide flux and vegetation transpiration obtained from eddy covariance techniques. The loss function of the model is set as follows:
[0016]
[0017] In the formula, NEE p and NEE m These are carbon dioxide flux estimates from a hybrid model and carbon dioxide flux observed from eddy currents, respectively; T p and T m W1 and W2 represent the estimated vegetation transpiration and the eddy covariance-observed vegetation transpiration, respectively; W1 and W2 represent the weighting coefficients of the estimation errors of carbon dioxide flux and vegetation transpiration, respectively, with values ranging from 0 to 1; Loss represents the total estimation error; n represents the total number of data samples.
[0018] S3: Training the neural network to estimate the model
[0019] Set up a deep neural network (DNN) a DNN b DNN p DNN T The number of hidden layers and the number of neurons in each layer are determined, and the weights and biases in the four deep neural networks are initialized using a random assignment method; T in a data sample in S1 is... a T v SM, W s W d LAI, h and ER n Input into a deep neural network (DNN) a In the middle, T a T sSM, W s and ER n Input into a deep neural network (DNN) b In the middle, T a R a SM, W s W d VPD, LAI, and h are input into a deep neural network (DNN). p The estimated ER values were obtained respectively. a Estimated ER b 1. Estimated GPP; 2. The estimated GPP and the T in this data sample a R a SM, VPD, W s LAI and h are input into a deep neural network (DNN). T In this process, the estimated vegetation transpiration T was obtained. p The estimated ER a Estimated ER b The estimated carbon dioxide flux (NEE) is obtained by adding the estimated GPP and the estimated GPP. p ;
[0020] The n data samples in S1 are sequentially input into the neural network estimation model. Based on the estimation results and the eddy current observations of carbon dioxide flux and vegetation evapotranspiration in the data samples, the value of the loss function is calculated.
[0021]
[0022] In the formula, NEE p and NEE m These are estimated carbon dioxide flux and carbon dioxide flux observed by eddy current, respectively; T p and T m W1 and W2 represent the estimated vegetation transpiration and the eddy covariance-observed vegetation transpiration, respectively; W1 and W2 represent the weighting coefficients of the estimation errors of carbon dioxide flux and vegetation transpiration, respectively, with values ranging from 0 to 1; Loss represents the total estimation error; n represents the total number of data samples.
[0023] Backpropagation is performed based on the loss function value to update the weights and biases of the four deep neural networks once, completing one iteration of training for the neural network estimation model. Then, the n data samples from S1 are input into the neural network estimation model again, and the weights and biases are updated again based on the loss function value. Then, the n data samples from S1 are input again for the next iteration of training. This process is repeated until the loss function value no longer decreases, resulting in a well-trained neural network estimation model.
[0024] S4: Estimating GPP and ER using a neural network estimation model
[0025] Obtain the air temperature T at a specific monitoring time point in the target area of S1. a Net radiation R a , saturated vapor pressure difference VPD, horizontal wind speed W s Wind direction W d EC technology observation of nighttime ecosystem respiration ER n Surface soil moisture content (SM) and soil temperature (T) s Leaf area index (LAI) and air temperature (T) a Blade temperature T v The plant height h is input into the neural network estimation model trained in S3, and then processed by the deep neural network (DNN). a and deep neural networks (DNN) b The output yields the estimated ER at that time point, generated by a deep neural network (DNN). p The output yields the estimated GPP at that time point.
[0026] Compared with existing technologies, the advantages of this invention are as follows: The proposed method for decomposing CO2 flux based on eddy covariance observation employs four sub-deep neural network modules, utilizing different combinations of environmental driving factors to independently estimate GPP and ER. After the network model is trained, GPP and ER can be estimated with high accuracy using only some meteorological, soil, and vegetation data. This method differs from empirical formula methods in that the calculation process does not require specific prerequisites or the operator to possess precision instrument operation skills. Applying this method helps researchers obtain atmospheric carbon cycle change data more efficiently and conveniently. Attached Figure Description
[0027] Figure 1 This is a schematic diagram illustrating the structure and principle of a neural network estimation model for an embodiment of a method for decomposing CO2 flux based on eddy covariance observation according to the present invention.
[0028] Figure 2 This is an evaluation graph showing the results obtained using different network models and experimental methods. Figure 2 (a) in the figure represents the evaluation results of the ER value obtained based on the daytime temperature data fitting method and the ER value obtained using the neural network estimation model of the present invention; Figure 2 (b) shows the ER value obtained based on the daytime temperature data fitting method and the ER value obtained using the neural network estimation model, which lacks a DNN. b Evaluation results of ER values obtained from network estimation; Figure 2 (c) shows the ER value obtained based on the daytime temperature data fitting method and the ER value obtained using the neural network estimation model, which lacks a DNN. T Evaluation results of ER values obtained from network estimation; Figure 2 In the figure, (d) represents the evaluation result of the GPP value obtained based on the daytime temperature data fitting method and the GPP value obtained using the neural network estimation model of the present invention; Figure 2 (e) represents the GPP value obtained based on the daytime temperature data fitting method and the value obtained using the neural network estimation model, which lacks a DNN. b Evaluation results of GPP values obtained from network estimation; Figure 2 In the figure, (f) represents the GPP value obtained based on the daytime temperature data fitting method and the value obtained by the neural network estimation model, which lacks a DNN. T Evaluation results of GPP values obtained from network estimation. Detailed Implementation
[0029] The present invention will now be described in further detail.
[0030] This invention provides a method for decomposing CO2 flux from eddy covariance observations based on a hybrid deep neural network. The method includes the following steps:
[0031] S1: Acquire relevant meteorological, vegetation, and soil data for the target area with vegetation cover. This data is obtained through monitoring every half hour. The meteorological data includes air temperature T. a Net radiation R a (used to characterize the effect of radiation on the excitation energy of photosynthesis), saturated vapor pressure difference VPD, horizontal wind speed W s Wind direction W d Carbon dioxide flux (NEE) m EC technology observation of nighttime ecosystem respiration ER n Relevant soil data include topsoil moisture content (SM) and soil temperature (T). s Vegetation-related data include leaf area index (LAI) and leaf temperature (T). v The data includes plant height (h) and vegetation transpiration (T). The total monitoring time for these data is no less than 72 consecutive hours. The dataset obtained at each monitoring time point is divided into one data sample. Based on the total monitoring time, n data samples are obtained, where n is no less than 144. The NEE in the data samples... m With T m Observations were obtained using EC technology.
[0032] S2: Construct a neural network estimation model
[0033] Carbon dioxide flux NEE is defined as the difference between ecosystem respiration flux (ER) and total primary productivity (GPP) (i.e., plant photosynthetic flux), and the calculation formula is shown below:
[0034] NEE = ER - GPP (2)
[0035] To more accurately estimate ER, ER can be further broken down into plant canopy respiration ER. a (i.e., vegetation canopy respiration flux) and soil respiration ER b (i.e., soil respiration flux), as shown in (3):
[0036] ER = ER a +ER b (3)
[0037] Numerous studies have shown that ER a ER b It has different environmental drivers than GPP.
[0038] The neural network estimation model is a hybrid four sub-DNNs model (HFSD), which includes four independent deep neural networks (DNNs). a DNN b DNN p DNN T .
[0039] Deep Neural Networks (DNNs) are nonlinear adaptive dynamic systems widely used in nonlinear regression, classification, and optimal fitting. They consist of an input layer, hidden layers, and an output layer. The number of neurons in the input layer represents the number of input variables, and the number of neurons in the output layer represents the number of output variables. The number of hidden layers and the number of neurons in each layer are the trainable parameters of the network model. Neurons in different layers of a DNN are connected layer by layer through weights, biases, and activation functions. Data transfer in a DNN is expressed by formula (1).
[0040] y = ReLU(wx + b) (1)
[0041] Here, "x" and "y" represent the input and output variables, respectively, and "w" and "b" represent the weights and biases, respectively. "ReLU" represents the ReLU activation function. In a DNN, the output variable y is the data from a neuron in the current layer of the network, and the input variable x is the weighted sum of the data from all neurons in the previous layer of the network.
[0042] Deep Neural Networks (DNNs) a DNN b DNN p DNN T Each network contains an input layer, hidden layers, and an output layer. The number of hidden layers and the number of neurons in each layer are not shared, and the weights and biases during data transmission are also not shared.
[0043] Deep Neural Networks (DNNs)a Used to obtain estimated vegetation canopy respiratory flux (ER) a Deep Neural Networks (DNNs) b Used to obtain estimated soil respiration flux ER b Deep Neural Networks (DNNs) p Deep neural networks (DNNs) are used to estimate the photosynthetic flux (GPP) of plants. T Used to obtain estimated vegetation transpiration T p .
[0044] Deep Neural Networks (DNNs) a The input is T a T v SM, W s W d LAI, h and ER n T v It can characterize the respiration intensity of plant leaves.
[0045] Deep Neural Networks (DNNs) b The input is T a T s SM, W s and ER n T s It can characterize the impact of soil microbial activity on ER b The impact.
[0046] Deep Neural Networks (DNNs) p The input is T a R a SM, W s W d , VPD, LAI and h. Among them R a Used to characterize the effect of radiation on the excitation energy of photosynthesis. T a It is a prerequisite for photosynthesis and plays an important role in biochemical reactions. s and W d Information on flux footprints of the EC system is used. VPD and SM are used to characterize the constraint of vegetation stomata on GPP flux. Finally, LAI and h reflect vegetation morphology.
[0047] Deep Neural Networks (DNNs) T The inputs are GPP and T. a R a SM, VPD, W s , LAI and h.
[0048] Deep Neural Networks (DNNs) a With Deep Neural Networks (DNNs) b The sum of the outputs is the estimated ecosystem respiration ER, using a deep neural network (DNN).p The sum of the output and the estimated ER is the estimated NEE.
[0049] The output of the HFSD model is mainly constrained by carbon dioxide flux and vegetation transpiration obtained from eddy covariance techniques. The loss function of the HFSD model is set as follows:
[0050]
[0051] In the formula, NEE p and NEE m These are carbon dioxide flux estimates from a hybrid model and carbon dioxide flux observed from eddy currents, respectively; T p and T m W1 and W2 represent the estimated vegetation transpiration and the eddy current-observed vegetation transpiration, respectively; W1 and W2 represent the weighting coefficients of the estimation errors of carbon dioxide flux and vegetation transpiration, respectively, with values ranging from 0 to 1; Loss represents the total estimation error; and n represents the total number of data samples.
[0052] S3: Training the neural network to estimate the model
[0053] Set up a deep neural network (DNN) a DNN b DNN p DNN T The number of hidden layers and the number of neurons in each layer are determined, and the weights and biases in the four deep neural networks are initialized using a random assignment method; T in a data sample in S1 is... a T v SM, W s W d LAI, h and ER n Input into a deep neural network (DNN) a In the middle, T a T s SM, W s and ER n Input into a deep neural network (DNN) b In the middle, T a R a SM, W s W d VPD, LAI, and h are input into a deep neural network (DNN). p The estimated ER values were obtained respectively. a Estimated ER b The estimated GPP. This estimated GPP and the T values in this data sample... a R a SM, VPD, W s LAI and h are input into a deep neural network (DNN).T In this process, the estimated vegetation transpiration T was obtained. p The estimated ER a Estimated ER b The estimated carbon dioxide flux (NEE) is obtained by adding the estimated GPP and the estimated GPP. p .
[0054] The n data samples in S1 are sequentially input into the neural network estimation model. Based on the estimation results and the eddy current observations of carbon dioxide flux and vegetation evapotranspiration in the data samples, the value of the loss function is calculated.
[0055]
[0056] In the formula, NEE p and NEE m These are estimated carbon dioxide flux and carbon dioxide flux observed by eddy current, respectively; T p and T m W1 and W2 represent the estimated vegetation transpiration and the eddy current-observed vegetation transpiration, respectively; W1 and W2 represent the weighting coefficients of the estimation errors of carbon dioxide flux and vegetation transpiration, respectively, with values ranging from 0 to 1; Loss represents the total estimation error; and n represents the total number of data samples.
[0057] Backpropagation is performed based on the loss function value to update the weights and biases of the four deep neural networks once, completing one iteration of training for the neural network estimation model. Then, the n data samples from S1 are input into the neural network estimation model again, and backpropagation is performed again based on the loss function value to update the weights and biases once more. This process is repeated until the loss function value no longer decreases, resulting in a well-trained neural network estimation model.
[0058] To evaluate the accuracy of the neural network estimation model designed in this invention, based on the data samples in S1, the ER value was obtained using a daytime temperature data fitting method (a measured method). Furthermore, the ER value estimated using the neural network estimation model designed in this invention was obtained, and the neural network estimation model lacked a DNN. b The ER value estimated by the network, and the lack of a DNN in the neural network estimation model. T ER value obtained from network estimation;
[0059] Furthermore, based on the data samples in S1, the GPP value (a measured method) is obtained using a daytime temperature data fitting method. Simultaneously, the GPP value estimated by the neural network estimation model designed in this invention is also obtained; this neural network estimation model lacks a DNN. b The GPP value obtained by the network estimation and the lack of DNN in the neural network estimation modelT The GPP value obtained from network estimation.
[0060] Using the coefficient of determination (R) 2 The evaluation, based on the results of fitting daytime temperature data and estimating the results of the network model, uses the square of the correlation coefficient and the root mean square error (RMSE) as the basis for assessment. The evaluation results are as follows: Figure 2 As shown, where, Figure 2 (a), (b), and (c) show the evaluation results of ER values obtained through different methods. Figure 2 (e), (d), and (f) represent the evaluation results of GPP values obtained through different methods.
[0061] from Figure 2 As can be seen, compared to other network models, the neural network estimation model designed in this invention achieves higher R values in estimating GPP and ER. 2 The highest values were 0.92 and 0.89, respectively, while the RMSE was 1.38 umol / s / m. 2 and 2.82umol / s / m 2 The value is the smallest, therefore the neural network estimation model designed in this invention has the highest accuracy.
[0062] S4: Estimating GPP and ER using a neural network estimation model
[0063] Obtain the air temperature T at a specific monitoring time point in the target area of S1. a Net radiation R a , saturated vapor pressure difference VPD, horizontal wind speed W s Wind direction W d EC technology observation of nighttime ecosystem respiration ER n Surface soil moisture content (SM) and soil temperature (T) s Leaf area index (LAI) and air temperature (T) a Blade temperature T v The plant height h is input into the neural network estimation model trained in S3, and then processed by the deep neural network (DNN). a and deep neural networks (DNN) b The output yields the estimated ER at that time point, generated by a deep neural network (DNN). p The output yields the estimated GPP at that time point.
[0064] Any aspects not covered in this invention are applicable to existing technologies.
Claims
1. A method for disentangling eddy covariance CO2 fluxes based on hybrid deep neural networks, characterized in that, The method comprises the following steps: S1: Acquire relevant meteorological, vegetation, and soil data for the target area with vegetation cover. This relevant data is obtained through monitoring every half hour. The meteorological data includes air temperature T. a Net radiation R a , saturated vapor pressure difference VPD, horizontal wind speed W s Wind direction W d Carbon dioxide flux (NEE) m EC technology observation of nighttime ecosystem respiration ER n Relevant soil data include topsoil moisture content (SM) and soil temperature (T). s Vegetation-related data include leaf area index (LAI) and leaf temperature (T). v Plant height h, vegetation transpiration T m The dataset obtained at each monitoring time point is divided into a data sample, and n data samples are obtained based on the total monitoring duration; NEE in the data samples m With T m Observations were obtained using EC technology; S2: constructing a neural network estimation model The neural network estimation model is a hybrid deep neural network model, which comprises four independent deep neural networks DNN a , DNN b , DNN p , DNN T ; each of the deep neural networks DNN a , DNN b , DNN p , DNN T comprises an input layer, a hidden layer and an output layer, the number of layers of the hidden layer and the number of neurons of each layer in each network are not shared, and the weight and bias in the data transmission process are not shared. Deep neural network DNN a For obtaining an estimated vegetation canopy respiration flux ER a , deep neural network DNN b For obtaining an estimated soil respiration flux ER b , deep neural network DNN p For obtaining an estimated plant photosynthesis flux GPP, deep neural network DNN T For obtaining an estimated vegetation transpiration T p ; Deep neural network, DNN a The input to the DNN is T a , T v , SM, W s , W d , LAI, h and ER n ; Deep neural network, DNN b The input to the DNN is T a , T s , SM, W s , and ER n ; Deep neural network, DNN p The input to T a , R a , SM, W s , W d , VPD, LAI and h; Deep neural network, DNN T The input to the DNN is GPP, T a , R a , SM, VPD, W s , LAI and h; deep neural network DNN a the sum of the outputs of the deep neural network DNN b with the estimated ecosystem respiration ER, the sum of the outputs of the deep neural network DNN p with the estimated ER is the estimated NEE; The output result of the neural network estimation model is constrained by the carbon dioxide flux and the vegetation transpiration obtained by the vorticity technique observation, and the loss function of the model is set as: where NEE p and NEE m are the estimated and observed carbon dioxide fluxes, respectively; T p and T m are the estimated and observed vegetation transpiration, respectively; W1 and W2 are the weight coefficients of the estimated errors of carbon dioxide flux and vegetation transpiration, respectively, both of which are between 0 and 1; Loss is the total error of the estimation; and n is the total number of data samples. S3: training the neural network estimation model Setting up deep neural networks DNN a , DNN b , DNN p , DNN T The number of layers of hidden layers and the number of neurons of each layer in the DNN are set, and the random assignment method is used to initialize the weights and biases in the four DNNs; the T a , T v , SM, W s , W d , LAI, h and ER n in a data sample in S1 are input into the DNN a , the T a , T s , SM, W s and ER n are input into the DNN b , the T a , R a , SM, W s , W d , VPD, LAI and h are input into the DNN p , and the estimated ER a , the estimated ER b , and the estimated GPP are obtained, respectively; the estimated GPP and the T a , R a , SM, VPD, W s , LAI and h in the data sample are input into the DNN T , and the estimated vegetation transpiration T p is obtained; the estimated ER a , the estimated ER b , and the estimated GPP are added, and the estimated carbon dioxide flux NEE p is obtained. The n data samples in S1 are sequentially input into the neural network estimation model, and the value of the loss function is calculated according to the estimation result and the carbon dioxide flux and the vegetation transpiration observed by the vorticity in the data samples: where NEE p and NEE m are the estimated and observed carbon dioxide fluxes, respectively; T p and T m are the estimated and observed vegetation transpiration, respectively; W1 and W2 are the weight coefficients of the estimated errors of carbon dioxide flux and vegetation transpiration, respectively, both of which are between 0 and 1; Loss is the total estimated error; and n is the total number of data samples. According to the value of the loss function, the weights and biases in the four deep neural networks are updated once, and the neural network estimation model completes one iteration training; then, the n data samples in S1 are input into the neural network estimation model again, and the weights and biases are updated once according to the value of the loss function again, and then the n data samples in S1 are input again for the next iteration training; constantly repeat until the value of the loss function no longer decreases, and the trained neural network estimation model is obtained; S4: estimating GPP and ER by using the neural network estimation model Air temperature T at a monitoring time point of a target region in S1 a , net radiation R a , saturated vapor pressure difference VPD, horizontal wind speed W s , wind direction W d , night-time ecosystem respiration ER observed by EC technology n , surface soil moisture SM, soil temperature T s , vegetation leaf area index LAI, air temperature T a , leaf temperature T v , plant height h, input into the trained neural network estimation model in S3, the estimated ER at the time point is obtained from the output of the deep neural network DNN a , and the output of the deep neural network DNN b , the estimated GPP at the time point is obtained from the output of the deep neural network DNN p .
2. The method of claim 1, wherein the method is based on a hybrid deep neural network. In S1, n is not less than 144.