Processing method and device of carbon flux data, equipment and medium
By processing carbon flux data using a deep neural network model, the problems of missing and anomalies were solved, the accuracy and intelligence of carbon flux data were improved, and accurate prediction and interpolation for different ecosystems and vegetation types were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- RAINROOT SCI LTD
- Filing Date
- 2026-02-09
- Publication Date
- 2026-06-02
Smart Images
Figure CN122132769A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of environmental science and artificial intelligence technology, and in particular to a method, apparatus, equipment and medium for processing carbon flux data. Background Technology
[0002] Carbon flux refers to the rate at which carbon (mainly in the form of carbon dioxide CO2, methane CH4, etc.) flows between different reservoirs in an ecosystem per unit time. Observing carbon flux is a crucial tool for studying the carbon cycle in ecosystems and is of great significance for understanding global climate change. Currently, traditional carbon flux observation methods, such as eddy covariance data processing, suffer from several drawbacks. Firstly, equipment, weather, and environmental conditions can lead to missing and anomalies in the acquired and calculated carbon flux data. Secondly, the calculation, processing, and interpolation of missing carbon flux data rely heavily on human experience for threshold setting, resulting in low levels of automation and a lack of consideration for environmental factors. This is particularly problematic when dealing with specific ecosystem types, leading to poor accuracy in the acquired carbon flux data.
[0003] Therefore, how to design a method to improve the quality and reliability of the acquired carbon flux data and get rid of the threshold setting and judgment based on human experience introduced in the traditional derivation method has become an urgent problem to be solved. Summary of the Invention
[0004] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a carbon flux data processing method, apparatus, equipment and medium. The main purpose is to solve the problems that the current carbon flux data has missing and abnormalities, and that the calculation process is based on human experience, resulting in poor carbon flux data accuracy and low level of intelligence in the processing process.
[0005] To achieve the above objectives, the main technical solutions adopted in this application include:
[0006] In a first aspect, embodiments of this application provide a method for processing carbon flux data, including:
[0007] Obtain raw observation data for the target area;
[0008] The raw observation data is input into a pre-trained deep neural network model to determine the primary category of the target region and the secondary categories subdivided under the primary category;
[0009] The carbon flux prediction formula corresponding to the secondary category is called to calculate the carbon flux prediction value; the vegetation physiological parameters in the carbon flux prediction formula corresponding to the secondary category are obtained by the deep neural network model from the historical raw observation data under the secondary category, and each secondary category has a corresponding carbon flux prediction formula;
[0010] Based on the output category of the deep neural network model, a physical mechanism verification equation is constructed and the theoretical value of carbon flux is calculated; the physical mechanism verification equation is constructed by the deep neural network model based on the energy conservation relationship;
[0011] Target carbon flux data is generated based on the predicted carbon flux value and the theoretical carbon flux value.
[0012] Optionally, the deep neural network model includes at least a time feature extraction layer and a correlation parameter construction layer; the primary category is an ecosystem category, and the secondary category is a vegetation category;
[0013] The raw observation data is input into the deep neural network model to determine the primary category of the target area and the secondary categories subdivided under the primary category. This includes: using the time feature extraction layer, extracting features from the raw observation data according to the time scale to obtain the seasonal variation features of the raw observation data; using the correlation parameter construction layer, performing correlation analysis on the raw observation data and the seasonal variation features of the raw observation data to identify the ecosystem category corresponding to the target area and the vegetation category under the ecosystem category.
[0014] Optionally, the step of constructing a correlation layer using the correlation parameters to perform correlation analysis on the original observation data and the seasonal variation characteristics of the original observation data, and to identify the ecosystem category corresponding to the target area and the vegetation category under the ecosystem category, includes: using the deep neural network model to determine the ecosystem category of the target area based on the vegetation index in the original observation data; and using the deep neural network model to analyze the leaf spectral characteristics and the seasonal variation characteristics of the original observation data to determine the vegetation category of the target area.
[0015] Optionally, the vegetation physiological parameters include at least light energy utilization rate, temperature response coefficient, and optimum temperature;
[0016] The steps for constructing the carbon flux prediction formula include: using the deep neural network model and employing the Bayesian optimization method based on historical raw observation data under the second-level category, obtaining the light energy utilization rate, the temperature response coefficient, and the optimal temperature; and using the deep neural network model to construct the carbon flux prediction formula based on the light energy utilization rate, the temperature response coefficient, and the optimal temperature.
[0017] Optionally, the deep neural network model further includes an uncertainty modeling layer;
[0018] Before constructing a carbon flux prediction formula based on the light energy utilization rate, the temperature response coefficient, and the optimum temperature using the deep neural network model, the method further includes: determining the confidence range of each of the light energy utilization rate, the temperature response coefficient, and the optimum temperature using the uncertainty modeling layer; and constructing a carbon flux prediction formula based on the light energy utilization rate, the temperature response coefficient, and the optimum temperature when all of the light energy utilization rate, the temperature response coefficient, and the optimum temperature fall within their respective confidence ranges.
[0019] Optionally, generating target carbon flux data based on the predicted carbon flux value and the theoretical carbon flux value includes: using the theoretical carbon flux value to perform outlier detection on the target carbon flux data sequence; removing carbon flux data from the target carbon flux data sequence if the difference between the carbon flux data and the theoretical carbon flux value is greater than a preset threshold; and interpolating the removed carbon flux data from the target carbon flux data sequence based on the carbon flux prediction formula.
[0020] Optionally, the method further includes: when a carbon flux deficiency is detected in the target time period, retrieving historical raw observation data with similar environmental conditions; the similar environmental conditions include at least one of the following: the same missing time period, the same temperature range, the same humidity level, and the same light intensity; based on the historical raw observation data with similar environmental conditions, using it as input to the deep neural network model, obtaining carbon flux prediction values as interpolation data, and interpolating the missing values.
[0021] Secondly, embodiments of this application provide a carbon flux data processing apparatus, comprising:
[0022] The acquisition unit is configured to acquire raw observation data of the target area;
[0023] The input unit is configured to input the raw observation data into a pre-trained deep neural network model to determine the primary category of the target region and the secondary category subdivided under the primary category;
[0024] The calculation unit is configured to call the carbon flux prediction formula corresponding to the secondary category to calculate the carbon flux prediction value; the vegetation physiological parameters in the carbon flux prediction formula corresponding to the secondary category are obtained by the deep neural network model from the historical raw observation data under the secondary category, and each secondary category has a corresponding carbon flux prediction formula.
[0025] The construction unit is configured to construct a physical mechanism verification equation and calculate the theoretical value of carbon flux based on the output category of the deep neural network model; the physical mechanism verification equation is constructed by the deep neural network model based on the energy conservation relationship;
[0026] The generation unit is configured to generate target carbon flux data based on the predicted carbon flux value and the theoretical carbon flux value.
[0027] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the carbon flux data processing method described in the first aspect.
[0028] Fourthly, this application provides an electronic device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the computer program to implement the carbon flux data processing method described in the first aspect.
[0029] Using the above technical solution, this application provides a method for processing carbon flux data. First, raw observation data of the target area is acquired. Then, the raw observation data is input into a pre-trained deep neural network model to determine the primary category of the target area and the secondary categories subdivided under the primary category. The carbon flux prediction formula corresponding to the secondary category is called to calculate the predicted carbon flux value. The vegetation physiological parameters in the carbon flux prediction formula corresponding to the secondary category are learned by the deep neural network model from historical raw observation data under the secondary category, and each secondary category has a corresponding carbon flux prediction formula. Based on the output category of the deep neural network model, a physical mechanism verification equation is constructed and the theoretical carbon flux value is calculated. The physical mechanism verification equation is constructed by the deep neural network model based on the energy conservation relationship. Finally, the target carbon flux data is generated based on the predicted carbon flux value and the theoretical carbon flux value. Compared to related technologies that derive carbon flux data based on human experience, this application utilizes a deep neural network model to accurately identify the community categories of each plant species in the target area and pre-constructs a dedicated carbon flux prediction formula for each community category. This allows for targeted carbon flux prediction based on different community categories, avoiding the traditional strategy of applying a single formula to all categories. Furthermore, by verifying the equations through physical mechanisms to determine the theoretical carbon flux values for the target area, the rationality and accuracy of the carbon flux predictions are validated, thereby improving the accuracy of the carbon flux data. Attached Figure Description
[0030] Figure 1 A flowchart illustrating a method for processing carbon flux data provided in an embodiment of this application;
[0031] Figure 2A comparison diagram of traditional carbon flux data processing methods provided in this application embodiment;
[0032] Figure 3 A schematic diagram of the architecture of a deep neural network model provided in an embodiment of this application;
[0033] Figure 4 This is a schematic diagram of a carbon flux data processing device provided in an embodiment of this application. Detailed Implementation
[0034] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.
[0035] As mentioned in the background, current traditional methods for observing carbon flux data typically employ eddy covariance, where eddy covariance refers to the irregular, random motion of air. Traditional eddy covariance methods for acquiring carbon flux data generally include several key steps: raw observation data acquisition, data preprocessing and quality control, coordinate rotation (to align the measured data with the actual Earth's surface), covariance calculation and frequency correction, density correction, and quality control. These key steps may vary slightly in different schemes. However, in the aforementioned chain of derivation from the acquired raw observation data to the calculated carbon flux data, some standardized processing procedures rely too heavily on mechanized thresholds and idealized assumptions, which can easily lead to poor accuracy in the final carbon flux data.
[0036] For example, turbulence refers to an irregular and random state of motion of a fluid (air or water), and the determination of turbulence development (also known as turbulence screening) relies excessively on a single threshold of friction velocity (u*). Current mainstream data screening strategies (i.e., u* filtering) assume that when the friction velocity exceeds a certain fixed threshold, NEE no longer changes with increasing turbulence, thus marking low u* data as "invalid" and discarding it. However, this strategy mechanically assumes a universal saturation relationship between turbulent mixing intensity and ecosystem respiration, ignoring the differences in nonlinear responses of different ecosystems (such as farmland, forests, and wetlands) under stable boundary layers. Studies have shown that a fixed u* threshold cannot capture the interaction effects of temperature, soil moisture, and turbulence, leading to a systematic underestimation or overestimation of nighttime respiration.
[0037] Secondly, the net ecosystem exchange (NEE) is usually calculated using NEE = -GPP + R_eco (GPP is total primary productivity, i.e., carbon flux, and R_eco is ecosystem respiration, usually negative for carbon sinks and positive for carbon sources). However, the calculation of NEE is based on overly simplified physiological assumptions. For example, the currently widely used night-time approach and day-time approach both implicitly assume two core assumptions: (1) the temperature sensitivity of ecosystem respiration (R_eco) during the day is consistent with that at night; (2) GPP (gross-primary-productivity) is zero when there is no photosynthesis during the day, and the effect of light inhibition on leaf respiration is not considered. These artificial assumptions ignore the inhibitory effect of mitochondrial respiration in plants under light conditions, leading to a systematic overestimation of R_eco and total primary productivity, which in turn distorts the seasonal dynamics of light use efficiency (LUE) and water use efficiency (WUE).
[0038] Third, data quality control (QC) relies on fixed statistical thresholds and lacks dynamic adaptability. Current standards typically use constant z-values (z=4 at night, z=5.5 during the day) to remove outliers based on median absolute deviation (MAD), or set fixed data integrity thresholds (e.g., if more than 10% of data is missing within a 30-minute window, the entire segment is removed). This "one-size-fits-all" statistical threshold fails to consider the differences in turbulence characteristics across different underlying surfaces (e.g., complex terrain, heterogeneous vegetation), potentially incorrectly removing genuine extreme carbon exchange events (e.g., pulsed respiration after heavy rain) while retaining data that violates the basic assumptions of eddy covariance (e.g., advection, non-steady-state).
[0039] In summary, current carbon flux data processing paradigms suffer from artificial assumptions and a tendency toward "oversimplification" in at least three key stages: quality control, turbulence screening, and flux segmentation. These include using fixed statistical thresholds to circumvent complex biophysical processes, employing steady-state assumptions to avoid the challenges of handling non-stationary turbulence, and using a simplistic day-night dichotomy to avoid the dynamic coupling of physiological processes. While these examples serve as illustrations, it's clear that while these simplifications ensure the operability and cross-site comparability of data processing, they may introduce difficult-to-quantify systematic biases when applied to precise carbon measurement or extreme climate response assessments.
[0040] Therefore, to improve the above-mentioned problems, this application proposes a carbon flux data processing method. This method can be applied to various ecological monitoring systems in fields such as agricultural management, forestry resource assessment, and carbon resource assessment. During operation, it can execute any of the carbon flux data processing methods mentioned below. Figure 1 As shown, the method includes:
[0041] S101, Obtain the raw observation data of the target area.
[0042] In S101, the raw observation data are time-series observation data labeled with characteristics such as year, season, region, and climate, specifically including carbon dioxide-related data, energy flux data, meteorological data, and vegetation data for the target area.
[0043] As a feasible implementation method, the raw observation data is collected from multiple observation stations, providing multidimensional data including: CO2 flux data: continuous time-series data, sampled at a frequency of 10-30 minutes, characterizing carbon exchange between the ecosystem and the atmosphere; energy flux data: sensible heat flux, latent heat flux, and soil heat flux, reflecting the distribution of energy in the ecosystem; meteorological data: environmental factors such as temperature, humidity, wind speed, solar radiation, and precipitation; spectral remote sensing data: vegetation index data such as NDVI, EVI, and LAI; phenological data: the timing of events such as the onset of vegetation growth, flowering, fruiting, and leaf fall; soil data: soil type, organic matter content, pH value, nitrogen, phosphorus, and potassium content; and topographic data: geographical features such as altitude, slope, and aspect.
[0044] S102, input the raw observation data into a pre-trained deep neural network model to determine the primary category of the target region and the secondary categories subdivided under the primary category.
[0045] Deep neural network models can employ various existing mature deep learning models, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). In one feasible implementation, the deep neural network model specifically includes a feature extraction layer, a correlation parameter construction layer, and an uncertainty modeling layer. The feature extraction layer uses a Temporal Convolutional Network (TCN) to extract time-scale features from the original observation data, providing a rich data foundation for the parameters in the subsequent S103 carbon flux prediction formula. The correlation parameter construction layer is used to perform correlation analysis on multiple related parameters with preset correlation relationships. For example, based on the correlation analysis of related parameters, the category of the target region can be determined. Then, when determining the parameters in the carbon flux prediction formula, the aforementioned carbon dioxide-related data and meteorological data, as well as seasonal and temporal characteristics, are considered. This allows for comprehensive analysis of data under different weather conditions (such as rainy days and snowy days), different seasons, and different time periods, thereby improving the accuracy of the parameters in the carbon flux prediction formula. For another example, in S104, when constructing the physical mechanism verification equation, the physical mechanism verification equation is constructed using energy flux data, carbon dioxide data, and meteorological data from the original observation data.
[0046] The uncertainty modeling layer is used to perform uncertainty analysis on the results of the correlation analysis of the correlation parameter construction layer, in order to generate vegetation physiological parameters and confidence intervals for constructing the carbon flux prediction formula and energy conservation parameters and confidence intervals for constructing the physical mechanism verification equation. Specifically, the uncertainty modeling layer uses a Bayesian neural network to quantify the uncertainty of each parameter estimate. This can be achieved through steps such as approximating the posterior distribution of the parameters using variational distribution, using Monte Carlo random discarding for uncertainty estimation during the inference stage, and combining multiple prediction results to improve stability.
[0047] S103, call the carbon flux prediction formula corresponding to the secondary category to calculate the carbon flux prediction value.
[0048] In S103, the vegetation physiological parameters in the carbon flux prediction formulas corresponding to the secondary categories are learned by a deep neural network model from historical raw observation data under the secondary categories, and each secondary category has a corresponding carbon flux prediction formula. First, the primary categories are ecosystem categories, including but not limited to forest ecosystems, farmland ecosystems, grassland ecosystems, and aquatic ecosystems. These four ecosystem categories can cover most ecosystem types in the observation area. The secondary categories are specific vegetation categories, including but not limited to coniferous forests (spruce, pine, etc.), broadleaf forests (oak, poplar, etc.), and mixed forests (a hybrid model of the above two categories) under forest ecosystems; C3 crop categories (wheat, rice, etc.) and C4 crop categories (corn, sugarcane, etc.) under farmland ecosystems; and tall grass, low grass, and desert shrub categories under grassland ecosystems.
[0049] This embodiment utilizes a deep neural network model to construct a corresponding carbon flux prediction formula for each of the aforementioned secondary categories, used to predict the carbon flux of the target region within a target time period. The parameters for the carbon flux prediction are obtained by learning from historical raw observation data under the secondary categories using the deep neural network model; that is, they are determined through continuous learning based on a large amount of raw observation data under the same secondary categories. Taking photosynthetic parameters as an example, these parameters include light energy utilization efficiency α, temperature coefficient β, optimum temperature T_opt, and critical vapor pressure difference VPD_critical. These parameters are used in the carbon flux prediction formula to obtain the predicted carbon flux value. The parameters vary for different categories. For example, C4 crops: 0.025-0.040 μmol CO2 / J; coniferous forests: 0.015-0.025 μmol CO2 / J; broadleaf forests: 0.020-0.030 μmol CO2 / J; grasslands: 0.018-0.028 μmol CO2 / J. Vapor pressure differential sensitivity reflects the different degrees of response of different plants to drought stress. For example, grassland plants are highly sensitive, with VPD_critical = 1.0-1.5 kPa; forest plants are moderately sensitive, with VPD_critical = 2.0-2.5 kPa.
[0050] S104: Based on the output category of the deep neural network model, construct the physical mechanism verification equation and calculate the theoretical value of carbon flux.
[0051] The physical mechanism verification equations were constructed by a deep neural network model based on the energy conservation principle. The energy conservation principle of the physical mechanism includes energy conservation and mass conservation, that is, the carbon cycle and energy cycle in a certain ecosystem have a conservation relationship. Taking a specific value as an example, the oxygen consumption and production should be equal.
[0052] The output category of the deep neural network model required to construct the physical mechanism verification equation includes at least one of a primary category and a secondary category. Specifically, when constructing the physical mechanism verification equation, it is necessary to determine at least the carbon uptake and carbon emissions in an ecosystem. For example, the physical mechanism verification equation may specifically include the following types:
[0053] (1) Net Ecosystem Exchange Equation (NEE Balance Equation):
[0054] NEE=−(GPP−R_eco−Rsoil−Rwater)
[0055] Where NEE is the net ecosystem carbon exchange, GPP is the gross primary productivity, R_eco is ecosystem respiration, R_soil is soil respiration, and R_water is waterborne carbon emissions. This equation describes the mass conservation relationship of carbon in an ecosystem and is applicable to all ecosystem types.
[0056] (2) Closed-loop energy balance equation:
[0057] Rn=H+LE+G
[0058] Where Rn is the net radiation flux, H is the sensible heat flux, LE is the latent heat flux, and G is the soil heat flux. This equation describes the energy budget balance of the ecosystem and is the core basis for verifying the physical rationality of carbon flux data.
[0059] (3) General equation for ecosystem respiration:
[0060] R_eco = Rref × Q10(T−Tref) / 10
[0061] Where R_eco is the ecosystem respiration rate, R_ref is the respiration rate at the reference temperature, Q10 is the temperature sensitivity coefficient, T is the current temperature, and T_ref is the reference temperature. This equation describes the general response of respiration to temperature.
[0062] S105 generates target carbon flux data based on predicted and theoretical carbon flux values.
[0063] In S105, generating target carbon flux data involves various scenarios. For example, if the original carbon flux observation data is missing, the target carbon flux data can be generated based on a certain ratio of predicted and theoretical values to impute the missing data. Alternatively, outliers can be identified and marked in the carbon flux data obtained from the current traditional scheme, and finally, a high-quality carbon flux dataset can be output.
[0064] In this embodiment, raw observation data of the target area is first acquired, and then input into a pre-trained deep neural network model to determine the primary category and the secondary categories subdivided under the primary category. The carbon flux prediction formula corresponding to the secondary category is called to calculate the predicted carbon flux value. The vegetation physiological parameters in the carbon flux prediction formula corresponding to the secondary category are learned by the deep neural network model from the historical raw observation data of the secondary category, and each secondary category has a corresponding carbon flux prediction formula. Based on the output category of the deep neural network model, a physical mechanism verification equation is constructed and the theoretical carbon flux value is calculated. The physical mechanism verification equation is constructed by the deep neural network model based on the energy conservation relationship. Finally, the target carbon flux data is generated based on the predicted and theoretical carbon flux values. Compared with the related technologies that derive carbon flux data based on human experience, this embodiment uses a deep neural network model to accurately identify the community categories of each plant in the target area and pre-constructs a dedicated carbon flux prediction formula for each community category. This allows for targeted prediction of carbon flux data based on different community categories, avoiding the traditional strategy of using a single formula applicable to all categories. On the other hand, by verifying the equations based on physical mechanisms to determine the theoretical value of carbon flux in the target area, the rationality and accuracy of the predicted carbon flux value are verified, thereby improving the accuracy of carbon flux data.
[0065] Optionally, the deep neural network model includes at least a time feature extraction layer and a correlation parameter construction layer; the primary category is the ecosystem category and the secondary category is the vegetation category.
[0066] The raw observation data is input into a deep neural network model to determine the primary category of the target area and the secondary categories subdivided under the primary category. This includes: using a time feature extraction layer to extract features from the raw observation data according to the time scale to obtain the seasonal variation characteristics of the raw observation data; and using a correlation parameter construction layer to perform correlation analysis on the raw observation data and the seasonal variation characteristics of the raw observation data to identify the ecosystem category corresponding to the target area and the vegetation category under the ecosystem category.
[0067] Furthermore, by constructing a layer based on correlation parameters, correlation analysis is performed on the original observation data and the seasonal variation characteristics of the original observation data to identify the ecosystem category corresponding to the target area and the vegetation category under the ecosystem category. This includes: using a deep neural network model to determine the ecosystem category of the target area based on the vegetation index in the original observation data; and using a deep neural network model to analyze the leaf spectral characteristics and the seasonal variation characteristics of the original observation data to determine the vegetation category of the target area.
[0068] This embodiment describes the process of determining the primary and secondary categories. For example, the deep neural network model first determines the basic ecosystem type of the target area based on vegetation indices (such as NDVI, EVI, etc.) and energy distribution characteristics (determined based on the ratio of sensible heat flux and latent heat flux) in the original observation data. In one feasible implementation, the NDVI time series for a continuous year is analyzed: if the NDVI remains high throughout the year (>0.6) and exhibits obvious seasonal fluctuations, while the latent heat flux accounts for more than 50% of the total energy flux, it is identified as a forest ecosystem; if the NDVI shows a clear periodicity of rapid growth-maturity-harvest, and sensible heat flux dominates during the day, it is identified as an agricultural ecosystem; if the NDVI value is moderate (0.3-0.5) and changes gradually, with a relatively balanced energy distribution, it is identified as a grassland ecosystem.
[0069] After determining the primary categories, vegetation is further subdivided. Taking forest ecosystems as an example, a deep neural network model analyzes leaf spectral characteristics: by analyzing the reflectance spectra in the visible-near-infrared band (400-900nm), coniferous forests exhibit high reflectance in the near-infrared band (750-900nm) with gentle seasonal changes; broadleaf forests, on the other hand, show a steep reflectance slope in the red-edge band (680-750nm), and their NDVI decreases significantly in autumn (a decrease of >40%). By extracting these spectral features through a convolutional neural network and combining them with seasonal characteristics (such as the start date of the growing season and the date of leaf fall), the deep neural network model can accurately distinguish between coniferous forests, broadleaf forests, and mixed types of the two. Through correlation analysis of the original observation data using the deep neural network model, different vegetation categories under different ecosystems are accurately identified first, so that targeted carbon flux prediction formulas can be used for subsequent predictions for different vegetation categories, thereby avoiding manual calculation of carbon flux data based solely on a single data point or a fixed method.
[0070] Optionally, vegetation physiological parameters include at least light energy utilization rate, temperature response coefficient, and optimum temperature;
[0071] The steps for constructing the carbon flux prediction formula include: using a deep neural network model and employing a Bayesian optimization method based on historical raw observation data under the second-level category, obtaining light energy utilization rate, temperature response coefficient, and optimum temperature; and using a deep neural network model to construct the carbon flux prediction formula based on light energy utilization rate, temperature response coefficient, and optimum temperature.
[0072] In this embodiment, optimal parameters for each vegetation category are learned from historical raw observation data using Bayesian optimization methods. Taking photosynthetic parameters as an example, firstly, the relationship between photosynthetically active radiation, air temperature, vapor pressure difference, and actual photosynthetic yield in historical data is analyzed. Then, mathematical optimization methods are used to find the parameter combination that best explains the observation data. Simultaneously, the model also considers the variation patterns of parameters throughout the four seasons and the trend changes over many years, ultimately establishing a complete parameter environment dependency database for each vegetation category. Then, using the Bayesian optimization framework, the prior distribution of light energy utilization efficiency α is set to N(0.020, 0.005²), the prior distribution of temperature response coefficient β is N(0.065, 0.015²), and the prior distribution of optimum temperature T_opt is N(20, 3²). A surrogate model from the parameter space to the model performance is established using Gaussian process regression, and the expected improvement (EI) is used as the acquisition function to iteratively search for the optimal parameter combination. In each iteration, the predicted GPP is calculated using candidate parameters and compared with the actual observed values, and the negative log-likelihood is calculated as the loss function. After 50-100 iterations, the parameters converge to the combination that minimizes the loss function, which is the optimal parameter learned, thus constructing the carbon flux prediction formula.
[0073] Optionally, the deep neural network model may also include an uncertainty modeling layer;
[0074] Before constructing a carbon flux prediction formula based on light energy utilization rate, temperature response coefficient, and optimum temperature using a deep neural network model, the method further includes: determining the confidence range of light energy utilization rate, temperature response coefficient, and optimum temperature using an uncertainty modeling layer; and constructing a carbon flux prediction formula based on light energy utilization rate, temperature response coefficient, and optimum temperature when all three fall within their respective confidence ranges.
[0075] In this embodiment, before constructing the carbon flux prediction formula, the reliability of the learned parameters is verified through an uncertainty modeling layer. Specifically, the uncertainty modeling layer integrates the following physical constraints: parameter range constraints: each parameter has a reasonable range in biology and physics; correlation constraints between parameters: such as the biological relationship between photosynthetic efficiency and respiration coefficient; energy conservation constraints: ensuring that the parameter estimation results satisfy the energy balance of the ecosystem; and mass conservation constraints: ensuring the conservation of mass in the carbon cycle process. The accurate estimation of these parameters directly determines the prediction accuracy and practicality of the formula.
[0076] In one feasible implementation, the Monte Carlo Dropout method is used: during the inference phase of the trained Bayesian neural network, Dropout is kept on for T=100 forward propagations. For each parameter (such as light energy utilization α), 100 sample values are collected, and their mean and standard deviation are calculated.
[0077] The confidence range is set to the mean ± 2 standard deviations. For example, if the mean of α after 100 samplings is 0.021 and the standard deviation is 0.002, then its confidence range is [0.017, 0.025]. All learned parameters are checked to ensure they fall within their biologically reasonable ranges: α ∈ [0.015, 0.045], β ∈ [0.04, 0.10], T_opt ∈ [5, 35]. Only when all parameters fall within the confidence range of the sampling distribution and meet biological constraints are these parameters used to construct the formal carbon flux prediction formula.
[0078] The following are examples and explanations of several carbon flux estimation formulas:
[0079] Formula specific to coniferous forests:
[0080] GPP_conifer=α_conifer×PAR×exp(-β_conifer×(T-T_opt_conifer)² / 400)×f(VPD_conifer)×(1-exp(-γ_conifer×LAI))
[0081] Wherein, α_conifer is light energy utilization (the range is determined to be 0.018±0.003 after the parameter confidence interval is determined by the uncertainty modeling layer), β_conifer is temperature response coefficient (the range is determined to be 0.065±0.010), T_opt_conifer is optimum temperature (20±3°C), γ_conifer is leaf area index effect (0.4±0.1), LAI is leaf area index, PAR is photosynthetically active radiation, and VPD is vapor pressure difference.
[0082] C3 crop-specific formula:
[0083] GPP_C3=α_C3×PAR×(1+260×Ci / Ca)×f(T_C3)×f(VPD_C3)×f(LAI_C3)
[0084] Where Ci / Ca is the ratio of intercellular to atmospheric CO2 concentration (0.8±0.1), α_C3 is the light energy utilization rate (0.028±0.004), and T_opt_C3 is the optimum temperature (25±4°C).
[0085] Furthermore, in addition to constructing carbon flux prediction formulas for different vegetation categories within various ecosystems, the deep neural network model also builds a dynamically updated formula library to manage, update, and validate formulas for different vegetation categories. Formula management includes continuously optimizing formula parameters, such as light energy utilization and optimum temperature, as raw observation data accumulates. Simultaneously, it mines and adjusts the influence relationships on formula parameters based on seasonal variations and climate conditions (based on raw observation data with temporal and seasonal characteristics). Finally, the accuracy is validated using an uncertainty modeling layer.
[0086] Optionally, target carbon flux data is generated based on predicted and theoretical carbon flux values, including: using theoretical carbon flux values to detect outliers in the target carbon flux data sequence; removing carbon flux data from the target carbon flux data sequence if the difference between the carbon flux data and the theoretical carbon flux value is greater than a preset threshold; and interpolating the removed carbon flux data from the target carbon flux data sequence based on the carbon flux prediction formula.
[0087] In this embodiment, outlier detection employs a method combining physical mechanism constraints and statistical tests. In one feasible implementation, firstly, for each 30-minute time interval, the observed carbon flux is simultaneously calculated using the carbon flux prediction formula, and the theoretical carbon flux value is calculated using the physical mechanism verification equation. Let the observed value be F_obs, the predicted value be F_pred, and the theoretical value be F_phy.
[0088] Two discrepancy indices are calculated: D1 = |F_obs - F_pred| / σ_pred (standardized residual), where σ_pred represents the uncertainty of the predicted value; and D2 = |F_obs - F_phy|. A preset threshold is set: when D1 > 3 and D2 > 10 μmol / m² / s, it is considered an outlier. For example, at a coniferous forest site, F_obs = 15 μmol / m² / s (strong respiration) was observed during the nighttime, but the predicted F_pred based on temperature conditions was 5 ± 1 μmol / m² / s, and the maximum reasonable value of F_phy based on energy balance was 8 μmol / m² / s. In this case, D1 = 10 and D2 = 7, which is considered an outlier and removed from the data sequence. Subsequent data interpolation can be performed at the locations of the removed data, thereby improving the accuracy of carbon flux data.
[0089] Optionally, the method further includes: when a carbon flux deficiency is detected in the target time period, retrieving historical raw observation data with similar environmental conditions; similar environmental conditions include at least one of the following: the same missing time period, the same temperature range, the same humidity level, and the same light intensity; based on the historical raw observation data with similar environmental conditions, using it as input to a deep neural network model, obtaining carbon flux prediction values as interpolation data, and interpolating the missing values.
[0090] In this embodiment, a dual imputation strategy of "environmental similarity matching combined with mechanistic model prediction" is adopted for the detected carbon flux missing periods. In a feasible implementation, when carbon flux data is found to be missing for a certain period (e.g., 10:00-10:30 on July 15, 2025), the available environmental variables for that period are first extracted, including air temperature (T=25.3°C), relative humidity (RH=65%), photosynthetically active radiation (PAR=1200μmol / m² / s), wind speed (u=2.1m / s), etc.
[0091] Then, a search was conducted in the historical database for time periods with similar environmental conditions. The similarity criteria were: same time period (±1 hour), temperature difference <2°C, PAR difference <15%, and humidity difference <10%. It is possible to find the time period from 10:30 to 11:00 on June 12, 2023, with environmental conditions of: T=24.8°C, PAR=1180 μmol / m² / s, and RH=68%. Complete observation data are available for this time period (GPP=18.5 μmol / m² / s).
[0092] Next, these similar environmental conditions are input into a deep neural network model, and the predicted carbon flux is calculated using a pre-learned formula and parameters specific to coniferous forests. Simultaneously, a weighted average is calculated based on actual observations from similar historical periods. The final interpolated value is obtained by combining the mechanistic model prediction (weight 0.7) and historical similarity values (weight 0.3), for example: interpolated value = 0.7 × 19.2 + 0.3 × 18.5 = 18.99 μmol / m² / s. The uncertainty of the interpolated value is also calculated, mainly considering three factors: environmental matching degree, historical data quality, and model prediction uncertainty. Finally, a complete interpolation result and quality assessment report are provided.
[0093] For outliers that have been removed, the interpolation process is initiated. First, check for any instrument malfunction records or extreme weather events. If there are no special reasons, interpolation is performed based on the carbon flux prediction formula: use the environmental variables (temperature, PAR, etc.) actually measured during the period to input into the dedicated formula, calculate the theoretical prediction value as the interpolation value, and provide a 95% confidence interval, for example, interpolation value = 5.2 ± 1.3 μmol / m² / s.
[0094] Furthermore, by combining the carbon flux data processing methods provided above, this approach addresses the shortcomings of current carbon flux data processing paradigms, which often suffer from artificial assumptions and a tendency towards "oversimplification" in the three key stages of quality control, turbulence screening, and flux segmentation. Specifically, this embodiment proposes a new method for acquiring carbon flux data. It utilizes a deep neural network model to determine the parameters of the carbon flux prediction formula and calculate the predicted carbon flux value based on raw observation data with temporal and seasonal characteristics. Simultaneously, it constructs a physical mechanism verification equation based on energy conservation to determine the theoretical carbon flux value. Finally, based on the predicted and theoretical carbon flux values, the carbon flux data for the target area is determined. Compared to related technologies that heavily rely on artificial assumptions or threshold settings, this approach offers a novel computational method. The entire process, including stages such as turbulence screening and flux segmentation, no longer involves artificial assumptions. Instead, it uses a model to analyze the relationships between data and refines the process down to at least the vegetation category level to construct carbon flux prediction formulas for each vegetation category. This eliminates the reliance on threshold settings and judgments based on human experience introduced in traditional derivation methods.
[0095] Combination Figure 2 As shown, the carbon flux data processing method provided by any of the above embodiments, in addition to the aforementioned schemes that derive carbon flux data based on human experience, utilizes a deep neural network model to accurately identify the community categories of each plant in the target area and pre-constructs a dedicated carbon flux prediction formula for each community category. This allows for targeted prediction of carbon flux data based on different community categories, avoiding the traditional strategy of applying a single formula to all categories. Furthermore, the theoretical carbon flux value of the target area is determined by verifying the equation through physical mechanisms to validate the rationality and accuracy of the carbon flux prediction value, thereby improving the accuracy of the carbon flux data. It also has at least the following technical effects:
[0096] This embodiment improves the use of deep neural network models for feature extraction from multidimensional carbon flux observation data; it automatically identifies key components of the ecological environment by combining time series analysis and spatial correlation analysis; and it discovers potential patterns and structures in the data through unsupervised learning methods. The core task is not simply to classify different ecosystems and vegetation types, but rather to estimate the parameters of the carbon flux prediction formula and the physical mechanism verification equation. This requires estimating multiple parameters from multidimensional carbon flux observation data, analyzing the physical and biological correlations between different parameters (e.g., α and T_opt are usually negatively correlated), and using the Bayesian square algorithm to provide confidence intervals for each parameter in a large amount of similar historical data, reflecting the uncertainty of the estimation. Finally, it also includes verification of physical rationality constraints.
[0097] Furthermore, the entire data processing workflow is automated, significantly reducing human intervention. It achieves automatic and accurate classification of vegetation types in complex ecosystems, refining classifications from coarse categories like "forest" and "farmland" to over 30 specific types, including coniferous forests, broad-leaved forests, C3 crops, and C4 crops. Each type has its unique photosynthetic mechanisms and parameter characteristics. Based on the accurately identified vegetation types, AI technology automatically constructs a dedicated carbon flux estimation formula for each type. This involves not only learning the formula structure but, more importantly, learning the specific parameters of each plant community, such as light energy utilization efficiency α, optimum temperature T_opt, and temperature response coefficient β. These parameters have clear biological significance and verifiability, completely solving the problem of the traditional, inefficient approach of applying a single formula to all ecosystems.
[0098] Throughout the process, the model also learns and mines a large number of nonlinear and time-varying relationships in the original observation data, enabling it to provide point estimates of unknowns and quantify the uncertainty of these estimates. This provides a reliable risk assessment basis for scientific research and policy making, significantly improving the accuracy of solving for unknowns. The integration of physical mechanism constraints ensures the scientific rigor and reliability of the data processing.
[0099] Finally, as a feasible implementation method, the specific model architecture of the deep neural network model described in this embodiment is explained. In this embodiment, a modular design is adopted, and information is transmitted between modules through standardized data interfaces. The specific model architecture is as follows: Figure 3 As shown, it includes:
[0100] Multi-scale feature extraction layer: A temporal convolutional network (TCN) is used to process time series data. By sliding dilated convolutions across multiple scales on the time axis, local and medium-term feature patterns of carbon flux data are extracted. TCN can simultaneously capture hourly-scale meteorological changes, daily-scale biological rhythms, seasonal-scale phenological changes, and annual-scale climate change trends.
[0101] Feature Representation Learning Layer: This layer uses a multilayer perceptron (MLP) to transform the extracted temporal features into high-dimensional feature representations, providing rich contextual information for parameter estimation. This layer captures complex nonlinear relationships through nonlinear transformations.
[0102] Formula parameter estimation layer: A specially designed multi-task regression network that simultaneously estimates multiple relevant formula parameters.
[0103] Shared hidden layers: learn common features and correlations among parameters;
[0104] Dedicated header for each parameter: A dedicated output header network is designed for each formula parameter;
[0105] Physical constraint unit: Integrates physical rationality constraints into the output layer to ensure that parameters are within the range that are biologically possible.
[0106] Bayesian uncertainty quantification layer: The uncertainty of parameter estimates is quantified using Monte Carlo Dropout and variational inference methods, providing confidence intervals and probability distributions for each parameter.
[0107] Physical constraint integration: Physical constraints are integrated into the parameter estimation process, including parameter range constraints, parameter correlation constraints, energy conservation constraints, mass conservation constraints, etc., to ensure the physical rationality of the parameter estimation results.
[0108] Furthermore, the comprehensive calculation of carbon absorption is achieved through four interrelated sub-modules, including:
[0109] Plant Community Carbon Absorption Module: Based on the fundamental principles of photosynthesis, this module treats the carbon absorption of a plant community as a nonlinear function of multiple environmental factors, including photosynthetically active radiation, temperature, humidity, and leaf area index. By measuring solar radiation intensity and considering temperature and humidity conditions, the photosynthetic efficiency of plant leaves is calculated. The leaf area index reflects the density of vegetation and is a crucial parameter affecting photosynthetic intensity. Ultimately, through the combined effect of these factors, the plant community's ability to absorb atmospheric CO2 is determined.
[0110] Soil Microbial Carbon Uptake Module: This module considers the impact of soil microbial respiration on the carbon cycle. Soil respiration is divided into two parts: surface respiration and deep respiration. Surface soil is in direct contact with the atmosphere, resulting in more active respiration; respiration in deep soil decreases exponentially with depth. By measuring the soil respiration rate at different depths and considering soil temperature and humidity conditions, the contribution of soil microorganisms to carbon flux is calculated.
[0111] Water Carbon Absorption Module: This module is specifically designed to handle carbon exchange processes in water bodies such as wetlands and lakes. CO2 exchange between water and the atmosphere follows the principles of physicochemical equilibrium; the mass transfer rate depends on the concentration difference and mass transfer coefficient at the water-air interface. By measuring the CO2 concentration difference between the water and air, and considering water surface conditions and temperature, the net carbon absorption of the water body is calculated.
[0112] Integrated Calculation and Verification Module: This module integrates the calculation results from the three sub-modules above and calculates the net ecological exchange volume according to the basic laws of ecological respiration. The final result needs to be verified by physical constraints such as energy balance and mass conservation to ensure the physical rationality of the calculation result.
[0113] Furthermore, fundamental laws of ecology and physics can be used as constraints and integrated into the learning process of AI models, including:
[0114] Energy conservation constraint: Ecosystems must adhere to the first law of thermodynamics, meaning the sum of sensible heat flux, latent heat flux, and soil heat flux should equal the net radiative energy received by the Earth's surface. Predictions that violate this physical law are penalized to ensure that the patterns learned by the AI model conform to fundamental physical principles.
[0115] Mass conservation constraint: The flow of carbon in the ecosystem must follow the law of conservation of mass, and the calculated CO2 flux cannot exceed the physically possible maximum value.
[0116] Boundary condition constraints: Combining ecological theory, reasonable boundary conditions are set for the output of the AI model.
[0117] The outlier detection module employs a multi-layered, multi-method integration strategy, including:
[0118] Statistical detection layer: Based on classical statistical theory, the Z-score method is used to calculate the deviation of each data point from the dataset mean. When the Z-score of a data point exceeds a preset threshold (usually 2-3 times the standard deviation), it is marked as a potential outlier. This method can effectively identify statistically significant outliers.
[0119] Machine learning detection layer: This layer trains various machine learning models (such as Isolation Forest and single-class Support Vector Machine) to learn pattern features from normal data. These models can identify anomalous data that does not conform to the learned patterns, making them particularly suitable for handling high-dimensional, non-linearly correlated data patterns.
[0120] Physical constraint verification layer: Based on ecological and physical constraints, this layer performs secondary verification on data that has passed the first two layers of testing. For example, it checks whether CO2 flux conforms to general ecological laws and whether the energy balance is closed. Only data points that pass multiple verifications simultaneously are confirmed as true outliers.
[0121] The comprehensive decision-making layer: This layer weights and fuses the detection results from the three levels mentioned above to arrive at the final outlier determination. A majority voting principle or a weighted average strategy is employed to ensure the accuracy and reliability of outlier detection.
[0122] The missing data imputation module includes:
[0123] Spatiotemporal Feature Construction Layer: First, the spatiotemporal patterns of the data are analyzed to construct a rich feature set. Time features include periodic information such as hours, seasons, and years; lag features reflect the historical dependencies of the data; and rolling statistical features capture short-term trend changes in the data. These features provide rich contextual information for subsequent interpolation algorithms.
[0124] Multiple interpolation strategy layer: Three different interpolation strategies are used in parallel, each with its own unique advantages.
[0125] Physical mechanism interpolation: Based on the carbon flux prediction formula constructed above, the missing carbon flux value is estimated using known environmental conditions (such as temperature, humidity, and light intensity). This method has clear physical significance, and the interpolation results conform to ecological principles.
[0126] Machine learning imputation: Using machine learning algorithms such as random forest and gradient boosting, complex relationship patterns between variables are learned from complete observation data, and then missing values are predicted based on the complete data around the missing points.
[0127] Spatiotemporal correlation interpolation: This method uses data from geographically nearby observation stations and employs spatial interpolation methods (such as inverse distance weighting and Kriging interpolation) to estimate missing data. It is suitable for situations involving multiple observation stations within the same region.
[0128] The ensemble optimization layer performs a weighted average of the interpolation results from the three strategies, automatically adjusting the weights based on the performance of each strategy under different conditions. For interpolation results with significant uncertainty, a corresponding confidence interval is provided to help users understand the reliability of the data.
[0129] Quality control layer: The final interpolation results are evaluated to check whether the interpolated data meets the physical constraints and whether it maintains reasonable continuity with the data at adjacent time points, thus ensuring the quality of interpolation.
[0130] Furthermore, as Figures 1 to 3 The specific implementation of the method shown in this embodiment provides a carbon flux data processing device, such as... Figure 4 As shown, the device includes:
[0131] Acquisition unit 401 is configured to acquire raw observation data of the target area;
[0132] The input unit 402 is configured to input the raw observation data into a pre-trained deep neural network model to determine the primary category of the target region and the secondary category subdivided under the primary category;
[0133] The calculation unit 403 is configured to call the carbon flux prediction formula corresponding to the secondary category to calculate the carbon flux prediction value; the vegetation physiological parameters in the carbon flux prediction formula corresponding to the secondary category are obtained by the deep neural network model from the historical raw observation data under the secondary category, and each secondary category has a corresponding carbon flux prediction formula.
[0134] Construction unit 404 is configured to construct a physical mechanism verification equation and calculate a theoretical value of carbon flux based on the output category of the deep neural network model; the physical mechanism verification equation is constructed by the deep neural network model based on the energy conservation relationship;
[0135] The generation unit 405 is configured to generate target carbon flux data based on the predicted carbon flux value and the theoretical carbon flux value.
[0136] In a specific application scenario, the input unit 402 is further configured to use the time feature extraction layer to extract features from the original observation data according to the time scale to obtain the seasonal change features of the original observation data; and to use the correlation parameter construction layer to perform correlation analysis on the original observation data and the seasonal change features of the original observation data to identify the ecosystem category corresponding to the target area and the vegetation category under the ecosystem category.
[0137] In specific application scenarios, the input unit 402 is further configured to, with the aid of the deep neural network model, determine the ecosystem category of the target area based on the vegetation index in the original observation data; and, with the aid of the deep neural network model, analyze the leaf spectral characteristics and the seasonal variation characteristics of the original observation data to determine the vegetation category of the target area.
[0138] In specific application scenarios, the computing unit 403 is further configured to, with the aid of the deep neural network model, use the Bayesian optimization method to obtain the light energy utilization rate, the temperature response coefficient, and the optimal temperature based on the historical raw observation data under the second-level category; and, with the aid of the deep neural network model, construct a carbon flux prediction formula based on the light energy utilization rate, the temperature response coefficient, and the optimal temperature.
[0139] In specific application scenarios, the computing unit 403 is further configured to, with the aid of the uncertainty modeling layer, determine the reliable numerical ranges corresponding to the light energy utilization rate, the temperature response coefficient, and the optimal temperature; and, when the light energy utilization rate, the temperature response coefficient, and the optimal temperature all fall within their respective reliable numerical ranges, construct a carbon flux prediction formula based on the light energy utilization rate, the temperature response coefficient, and the optimal temperature.
[0140] In a specific application scenario, the generation unit 405 is further configured to use the theoretical carbon flux value to perform outlier detection on the target carbon flux data sequence; if there is a carbon flux data in the target carbon flux data sequence whose difference from the theoretical carbon flux value is greater than a preset threshold, the carbon flux data in the target carbon flux data sequence is removed; and the removed carbon flux data in the target carbon flux data sequence is interpolated based on the carbon flux prediction formula.
[0141] In a specific application scenario, the generation unit 405 is further configured to retrieve historical raw observation data with similar environmental conditions when a carbon flux deficiency is detected in the target time period; the similar environmental conditions include at least one of the following: the same missing time period, the same temperature range, the same humidity level, and the same light intensity; based on the historical raw observation data with similar environmental conditions, the carbon flux prediction value is obtained as the input of the deep neural network model and used as the interpolation data to interpolate the missing value.
[0142] It should be noted that other corresponding descriptions of the functional units involved in the carbon flux data processing device provided in this embodiment can be found in [reference needed]. Figures 1 to 3 The corresponding description in [the document] will not be repeated here.
[0143] Based on the above, Figures 1 to 3 Accordingly, this embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figures 1 to 3 The method shown.
[0144] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.
[0145] Based on the above, Figures 1 to 3 The method shown, and Figure 4 To achieve the above objectives, this application also provides an electronic device, which can be configured on a computer side, etc. The device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-described objectives. Figures 1 to 3 The method shown.
[0146] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.
[0147] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.
[0148] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.
[0149] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented in hardware. Compared with related technologies that derive carbon flux data based on human experience, this embodiment utilizes a deep neural network model to accurately identify the community categories of each plant in the target area and pre-constructs a dedicated carbon flux prediction formula for each community category. This allows for targeted prediction of carbon flux data based on different community categories, avoiding the traditional strategy of applying a single formula to all categories. Furthermore, by verifying the equations through physical mechanisms to determine the theoretical carbon flux value of the target area, the rationality and accuracy of the carbon flux prediction value are verified, thereby improving the accuracy of the carbon flux data.
[0150] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0151] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0152] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.
Claims
1. A method for processing carbon flux data, characterized in that, include: Obtain raw observation data for the target area; The raw observation data is input into a pre-trained deep neural network model to determine the primary category of the target region and the secondary categories subdivided under the primary category; The carbon flux prediction formula corresponding to the secondary category is called to calculate the predicted carbon flux value; The vegetation physiological parameters in the carbon flux prediction formula corresponding to the secondary category are obtained by the deep neural network model from the historical raw observation data under the secondary category, and each secondary category has a corresponding carbon flux prediction formula. Based on the output category of the deep neural network model, a physical mechanism verification equation is constructed and the theoretical value of carbon flux is calculated; the physical mechanism verification equation is constructed by the deep neural network model based on the energy conservation relationship; Target carbon flux data is generated based on the predicted carbon flux value and the theoretical carbon flux value.
2. The method according to claim 1, characterized in that, The deep neural network model includes at least a time feature extraction layer and a correlation parameter construction layer; the primary category is the ecosystem category, and the secondary category is the vegetation category; The raw observation data is input into the deep neural network model to determine the primary category of the target region and the secondary categories subdivided under the primary category, including: With the help of the time feature extraction layer, features are extracted from the original observation data according to the time scale to obtain the seasonal variation features of the original observation data; By using the correlation parameter construction layer, correlation analysis is performed on the original observation data and the seasonal variation characteristics of the original observation data to identify the ecosystem category corresponding to the target area and the vegetation category under the ecosystem category.
3. The method according to claim 2, characterized in that, The method of constructing a layer using the correlation parameters to perform correlation analysis on the original observation data and the seasonal variation characteristics of the original observation data, identifying the ecosystem category corresponding to the target area, and the vegetation category under the ecosystem category, includes: Using the deep neural network model, the ecosystem category of the target area is determined based on the vegetation index in the original observation data; By using the deep neural network model, the seasonal variation characteristics of leaf spectral features and raw observation data are analyzed to determine the vegetation category of the target area.
4. The method according to claim 1, characterized in that, The vegetation physiological parameters include at least light energy utilization rate, temperature response coefficient, and optimum temperature; The steps for constructing the carbon flux prediction formula include: Using the deep neural network model, the Bayesian optimization method is employed to obtain the light energy utilization rate, the temperature response coefficient, and the optimal temperature based on the historical raw observation data under the second-level category. Using the deep neural network model, a carbon flux prediction formula is constructed based on the light energy utilization rate, the temperature response coefficient, and the optimal temperature.
5. The method according to claim 4, characterized in that, The deep neural network model also includes an uncertainty modeling layer; Before constructing a carbon flux prediction formula based on the light energy utilization rate, the temperature response coefficient, and the optimum temperature using the deep neural network model, the method further includes: By using the uncertainty modeling layer, the reliable numerical ranges corresponding to the light energy utilization rate, the temperature response coefficient, and the optimal temperature are determined. When the light energy utilization rate, the temperature response coefficient, and the optimal temperature all fall within their respective reliable numerical ranges, a carbon flux prediction formula is constructed based on the light energy utilization rate, the temperature response coefficient, and the optimal temperature.
6. The method according to claim 1, characterized in that, The step of generating target carbon flux data based on the predicted carbon flux value and the theoretical carbon flux value includes: Outlier detection is performed on the target carbon flux data sequence using the aforementioned theoretical value of carbon flux. If the difference between the carbon flux data and the theoretical value of carbon flux in the target carbon flux data sequence is greater than a preset threshold, the carbon flux data in the target carbon flux data sequence will be removed. The carbon flux data removed from the target carbon flux data sequence is interpolated based on the carbon flux prediction formula.
7. The method according to claim 1, characterized in that, The method further includes: If a carbon flux deficiency is detected during the target period, historical raw observation data with similar environmental conditions are retrieved; the similar environmental conditions include at least one of the following: the same period of deficiency, the same temperature range, the same humidity level, and the same light intensity. Based on historical raw observation data with similar environmental conditions, the carbon flux prediction value is obtained as the input of the deep neural network model and used as the imputation data to imput the missing values.
8. A carbon flux data processing device, characterized in that, include: The acquisition unit is configured to acquire raw observation data of the target area; The input unit is configured to input the raw observation data into a pre-trained deep neural network model to determine the primary category of the target region and the secondary category subdivided under the primary category; The calculation unit is configured to call the carbon flux prediction formula corresponding to the secondary category to calculate the carbon flux prediction value; The vegetation physiological parameters in the carbon flux prediction formula corresponding to the secondary category are obtained by the deep neural network model from the historical raw observation data under the secondary category, and each secondary category has a corresponding carbon flux prediction formula. The construction unit is configured to construct a physical mechanism verification equation and calculate the theoretical value of carbon flux based on the output category of the deep neural network model; the physical mechanism verification equation is constructed by the deep neural network model based on the energy conservation relationship; The generation unit is configured to generate target carbon flux data based on the predicted carbon flux value and the theoretical carbon flux value.
9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.