A forest NEP simulation method and system based on three-branch EMA-CKAN and ecological constraints

Abstract

The present application relates to a kind of forest NEP simulation method and system based on three-branch EMA-CKAN and ecological constraint, constructs three-branch EMA-CKAN network, including three parallel EMA-CKAN subnet, forest carbon flux is decomposed into top canopy total primary productivity, bottom canopy total primary productivity and ecological respiration three sub-problems are modeled respectively, both retain the physical independence of each process and can be realized by shared representation and coupling mechanism mutual restraint, after embedding EMA multi-scale attention mechanism in the first two layers of three-dimensional CKAN of each subnet, then stack two layers of three-dimensional CKAN, more deep level semantic extraction and fusion are carried out to the feature enhanced by attention;And based on ecological constraint function, NEP is synthesized in the form of differentiable inside network, so that model is physically consistent in pursuit of statistical fitting while being constrained, and training stability, noise resistance and physical interpretability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forest carbon resource assessment, and particularly relates to a forest NEP simulation method and system based on three-branch EMA-CKAN and ecological constraints. BACKGROUND

[0002] As the core component of the terrestrial ecosystem, the forest ecosystem plays an irreplaceable role in regulating global carbon cycle, slowing down the increase of greenhouse gas concentration and mitigating the trend of climate warming. Among them, the NEP of the forest directly represents its net carbon exchange capacity, which is a key indicator to determine the carbon source / sink state of the ecosystem. However, due to the complex coupling of light, temperature, water and soil nutrients and other environmental factors affecting the carbon exchange process of the forest ecosystem, the current common light energy utilization rate model simplifies the calculation process to a certain extent, but it is difficult to accurately express the nonlinear relationship and heterogeneous response mechanism of the forest ecosystem to environmental factors, and the simulation accuracy and generalization ability in the complex forest canopy are significantly limited. SUMMARY

[0003] The purpose of the present application is to provide a forest NEP simulation method and system based on three-branch EMA-CKAN and ecological constraints to solve the problems in the background art.

[0004] To solve the above technical problems, the present application adopts the following technical scheme: a forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints, comprising:

[0005] Constructing a three-branch EMA-CKAN network, including three parallel EMA-CKAN subnetworks, decomposing the forest carbon flux into three sub-problems of the total primary productivity GPP of the top canopy, the total primary productivity GPP of the bottom canopy and the ecological respiration Re, and synthesizing NEP in a differentiable form inside the network based on an ecological constraint function, wherein the ecological constraint function is:

[0006] ;

[0007] wherein, is the NEP prediction deviation, is the mass constraint, is the energy constraint, is the temperature constraint, is the total weight, and the weight is between 0 and 0.2, , , are all non-negative numbers, and + + =1.

[0008] and

[0009] EMA multi-scale attention mechanism is embedded after the first two layers of three-dimensional CKAN in each subnetwork, and two layers of three-dimensional CKAN are stacked after the EMA multi-scale attention mechanism for deeper semantic extraction and fusion of the attention-enhanced features.

[0010] Preferably, the three-branch EMA-CKAN network is optimized, comprising:

[0011] The data set is constructed, including the required forest type flux site in-situ measurement data, the corresponding site 10km range surface area remote sensing data, meteorological data and soil data;

[0012] The data set is divided into a training data set, a test data set and a validation data set;

[0013] The remote sensing data, meteorological data and soil data are used as the input of the three-branch EMA-CKAN network, the correlation between the top and bottom canopy gross primary productivity, ecological respiration Re and remote sensing data, meteorological data and soil data is learned, the physical model parameter calibration process is converted into an end-to-end neural network training task, the ecological constraint function is calculated, and the training is repeated multiple times until the validation data set accuracy of the three-branch EMA-CKAN network converges;

[0014] The coefficient of determination and root mean square error indicators are used to check the fitting performance of the three-branch EMA-CKAN network on the training samples, and the optimal three-branch EMA-CKAN network is obtained.

[0015] Further preferably, before dividing the data set, the data set is preprocessed, comprising:

[0016] The photosynthetically active radiation intercepted by the top and bottom canopies of the forest at the site is calculated;

[0017] ;

[0018] ;

[0019] is the photosynthetically active radiation for the top canopy, is the photosynthetically active radiation for the bottom canopy; is the ground albedo; is the scattered photosynthetically active radiation below the canopy; is the internal radiation scattering term of the canopy; is the leaf inclination angle; is the solar zenith angle; and are the diffuse and direct components of incident solar radiation, respectively, estimated from incident solar radiation and clearness index. is the leaf area index of the top canopy layer, is the leaf area index of the bottom canopy layer; is the vegetation aggregation index, which depends on the forest type;

[0020] The day-scale flux data disclosed by the site is superimposed on the 8-day scale, and part of the abnormal data is deleted, including parameter observation vacancy, inconsistent with plant growth mode, and observation value anomaly;

[0021] The domain remote sensing data, meteorological data and soil data are resampled to 500m by bilinear interpolation method;

[0022] The remote sensing data, meteorological data and soil data use cubic spline interpolation to change the time resolution, so as to align with the 8-day of the flux data.

[0023] Preferably, the 8-day scale remote sensing data, meteorological data and soil data after preprocessing are combined with the 8-day scale flux data to make time alignment; according to the proportion of 6:2:2, they are divided into training set, validation set and test set; and each site data appears in the training set, validation set and test set for more than 2 months.

[0024] Further preferably, the in-situ measured data of the flux site includes net ecosystem productivity NEP, ecological respiration Re, and gross primary productivity GPP data;

[0025] The remote sensing data includes leaf area index LAI, enhanced vegetation index EVI, normalized vegetation index NDVI, chlorophyll fluorescence SIF, and chlorophyll concentration LCC;

[0026] The meteorological data includes air temperature Ta, saturated water vapor pressure difference VPD, and solar radiation SW;

[0027] The soil data includes soil nitrogen content Sn, soil acid-base pH, soil organic carbon concentration SOC, and soil moisture SM;

[0028] The air temperature Ta, enhanced vegetation index EVI, soil moisture SM, soil organic carbon concentration SOC, and soil acid-base pH constitute the Re estimation data;

[0029] The top canopy photosynthetically active radiation , air temperature Ta, saturated water vapor pressure difference VPD, SW, soil moisture SM, chlorophyll fluorescence SIF, chlorophyll concentration LCC, normalized vegetation index NDVI, enhanced vegetation index EVI, soil nitrogen content Sn, and soil moisture SM constitute the estimation data;

[0030] The bottom canopy photosynthetically active radiation air temperature Ta, vapor pressure deficit VPD, SW, chlorophyll fluorescence SIF, chlorophyll concentration LCC, normalized difference vegetation index NDVI, enhanced vegetation index EVI, soil nitrogen content Sn and soil moisture SM estimated data.

[0031] Preferably, in a sub-network, based on Re estimated data, calculate the ecosystem respiration Re:

[0032] ;

[0033] In a sub-network, based on estimated data, calculate the top canopy gross primary productivity :

[0034] ;

[0035] In a sub-network, based on estimated data, calculate the bottom canopy gross primary productivity :

[0036] ;

[0037] wherein, , represent the upper canopy of the forest, and the lower canopy of the forest EMA-CKAN sub-network output results.

[0038] Preferably, based on the ecosystem respiration Re, the top canopy gross primary productivity and the bottom canopy gross primary productivity calculate the simulated value of the forest ecosystem NEP :

[0039] ;

[0040] ;

[0041] wherein, , represent the simulated value of the ecosystem respiration;

[0042] Preferably, all input variables, including remote sensing, meteorological and soil data, are normalized using min-max before inputting into the deep learning model, linearly mapping the original numerical values to the range of 0-1, improving the stability and convergence speed of model training; the network optimizer selects the Adam optimizer; modify the weights in the entire network after obtaining the loss, and train multiple times until the accuracy of the validation dataset of the three-branch EMA-CKAN network converges, and save the model parameters.

[0043] Further preferably, the coefficient of determination (R2 ) to verify the predicted value-observation linear correlation, the root mean square error (RMSE) is used to quantify the absolute deviation; the calculation is as follows:

[0044] ;

[0045] .

[0046] Preferably, each sub-network has a consistent overall structure, and their first layer is an input layer, which adjusts the data to adapt the input data to the input network and connects with the subsequent network layer; then a 2-layer 3D CKAN layer is connected, the convolution kernel size is 3*3*3, the filters are 16 and 32 respectively, and the activation function uses the leak relu function, and the α in leak relu is selected in the range of 0.01-0.3;

[0047] The EMA multi-scale attention mechanism divides the channels into 4 groups, and uses parallel depthwise-separable convolution for each group to calculate the multi-scale spatial response with a 1*3 convolution kernel; the scale responses are spliced and subjected to 1*1 convolution, then normalized by BatchNorm to obtain the spatial attention map, and then subjected to 1*1 convolution projection to restore the channel dimension, and finally the original feature is added back with a residual coefficient γ; the EMA multi-scale attention mechanism enhances the mapping relationship between multi-scale space and multi-source ecological information, and improves the response ability of the model to the nonlinear ecological relationship of multi-source ecological information;

[0048] Preferably, the stacked two-layer 3D CKAN convolution kernel size is 3*3*3, the filters are 64 and 128 respectively, and the activation function uses the leak relu function; after the stacked two-layer 3D CKAN, the following are added in turn:

[0049] A reshaping layer (reshap) is added to convert the three-dimensional features to two-dimensional features;

[0050] Four layers of two-dimensional CKAN network, the convolution kernel size is 3*3, the activation function uses the leak relu function, the α in leak relu is selected in the range of 0.01-0.3, and the filters are 128, 128, 64 and 32 respectively;

[0051] A 2*2 size pooling layer is added;

[0052] A reshaping layer is added;

[0053] Three fully connected layers, the number of neural networks in the fully connected layers is 32, 16 and 1 respectively.

[0054] Preferably, The quality constraint is calculated as:

[0055] ;

[0056] Quality constraints are used to constrain the joint simulation accuracy of the three components of carbon flux by the three-branch EMA-CKAN network, requiring the predicted NEP, gross primary productivity (GPP) and ecosystem respiration (Re) to satisfy the consistency of material balance at both sample level and summary level, and the constraint definition is:

[0057] ;

[0058] Where n is the number of samples used for calculation; , , respectively represent the predicted net ecosystem productivity, gross primary productivity and ecosystem respiration by the three-branch EMA-CKAN, , , are the corresponding site observation values;

[0059] is the calculation of energy constraints to constrain whether the utilization efficiency of light energy by the canopy conforms to the known physical mechanism, and the loss is defined as:

[0060] ;

[0061] Where n is the number of samples used for calculation; represents the predicted net ecosystem productivity by the three-branch EMA-CKAN, and is the calculation of the total canopy gross primary productivity based on the relationship between the absorbed light energy by the forest canopy and the fixed biomass energy, and the calculation method is as follows:

[0062] ;

[0063] In the formula, , and respectively are the correction factors of maximum VPD and minimum photosynthetic temperature, indicating the influence of saturated water vapor pressure difference and minimum air temperature on vegetation photosynthesis; and are obtained by bilinear regression of flux data. The minimum air temperature, and soil moisture correction factor are calculated as:

[0064] ;

[0065] ;

[0066] In the formula, , , , is a parameter related to forest type, which is obtained by analyzing the time series variation of site flux data

[0067] For temperature constraint to depict the sensitivity of ecological respiration to temperature and constrain the prediction of Re, the constraint introduces a physical reference model of temperature response, the parameters of which can be initialized by site data of flux observation station or by ecological class, and can be fine-tuned in the training according to the physically acceptable range, and the constraint expression is

[0068] ;

[0069] wherein is the predicted value of ecological respiration, is the corresponding physical reference value of ecological respiration considering temperature, and the calculation mode is as shown below

[0070] ;

[0071] wherein 10°C or 15°C is selected, representing the reference temperature, is the reference respiration intensity, is the ecological respiration intensity coefficient, which is obtained by selecting the measured and S by regression estimation, and the formula is as shown below:

[0072] ;

[0073] ;

[0074] ;

[0075] In the formula, T is the measured data of the flux station.

[0076] The application also discloses a forest NEP simulation system based on three-branch EMA-CKAN and ecological constraint, comprising:

[0077] An acquisition sample data unit: acquiring required in-situ measurement data of forest type flux station, corresponding station 10km range surface remote sensing data, meteorological data and soil data, and matching with flux observation scale;

[0078] A data preprocessing unit: calculating top canopy photosynthetically active radiation and bottom canopy photosynthetically active radiation The system unifies data from different sources and with different time resolutions to the 8-day time scale required by the model, and performs outlier cleaning and imputation; it overlays or averages the daily throughput data to the 8-day scale according to rules, while removing abnormal training data; it outputs aligned multi-source image patch data, performs min-max normalization on it, and divides it into training set, validation set and test set in a 6:2:2 ratio.

[0079] Model training unit: Construct and train a three-branch EMA-CKAN model, combine ecological constraint functions to estimate forest NEP in real end-to-end, use Adam optimizer to update weights during training, save model weights and training history after training, and finally use the coefficient of determination and root mean square error index on an independent test set to check the fitting performance of the three-branch EMA-CKAN network to the training samples, and save the optimal three-branch EMA-CKAN network.

[0080] NEP simulation unit: Configured with an optimal three-branch EMA-CKAN network for batch inference and spatialized output of the target region.

[0081] Preferably, remote sensing, meteorological, and soil data of the target area are acquired and resampled using the same preprocessing procedure as during training to obtain an 8-day scale prediction dataset, which is then input into the model pixel-by-pixel or in blocks for calculation. , The NEP spatiotemporal sequence is synthesized with Re and finally generated to produce NEP data for the target forest region.

[0082] Beneficial Effects: This invention employs a three-branch structure to decompose forest carbon flux into three sub-problems: top canopy GPP, bottom canopy GPP, and ecological respiration (Re). These sub-problems are modeled separately, preserving the physical independence of each process while enabling mutual constraints through shared representations and coupling mechanisms. This structurally reduces feature confounding between processes and improves the estimation accuracy of sub-processes and the overall NEP. Secondly, the EMA-CKAN sub-network embeds an EMA multi-scale attention mechanism after the first two layers of 3D CKAN to capture features that have not been over-downsampled, enhancing the discriminability of local photosynthetic and respiration-related features. Subsequently, two more layers of 3D CKAN are stacked to perform deeper semantic extraction and fusion of the attention-enhanced features. Furthermore, this invention constructs an ecological constraint system comprised of the synergistic effects of mass, energy, and temperature constraints, and integrates it as a differentiable loss function into the training process of the three-branch EMA-CKAN network. This allows the model to pursue statistical fitting while being constrained by physical consistency, synergistically improving training stability, noise resistance, and physical interpretability. Attached Figure Description

[0083] Figure 1This is a flowchart of the forest NEP simulation method based on the three-branch EMA-CKAN and ecological constraints of the present invention;

[0084] Figure 2 This is a diagram of the three-branch EMA-CKAN network of the present invention;

[0085] Figure 3 This is a diagram verifying the simulation accuracy of the three-branch EMA-CKAN model in this invention; Detailed Implementation

[0086] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe a forest NEP simulation method and system based on a three-branch EMA-CKAN and ecological constraints, or several specific implementations of this invention, and does not strictly limit the scope of protection specifically claimed by this invention.

[0087] Traditional forest NEP estimation models have shortcomings in their mechanism representation: existing methods mostly rely on data-driven regression, treating NEP as a single objective, while ignoring... , The physiological processes associated with respiration (Re) make it difficult to guarantee a reasonable decomposition and cross-regional generalization of carbon budget processes. Furthermore, traditional methods typically rely on single loss functions such as RMSE to optimize parameters, failing to explicitly incorporate key ecological mechanisms such as mass conservation, energy balance, and temperature dependence into the training process, thus exhibiting deficiencies in physical consistency and physiological rationality. To overcome these problems, this invention proposes an ecologically constrained three-branch EMA-CKAN network. This method decomposes forest carbon flux into three branches: top canopy GPP, bottom canopy GPP, and ecological respiration. A multi-scale attention module (EMA) is introduced during convolutional feature extraction to fully preserve the spatial information of multi-source ecological variables that have not yet been weakened by pooling, enhancing the model's responsiveness to key environmental features. Simultaneously, an ecological constraint loss function is introduced into the optimization objective, comprehensively considering mass conservation, energy balance, and temperature effects. This allows the model to maintain high-precision fitting while adhering to the fundamental physical laws of ecosystem processes, achieving a more accurate and generalized simulation of forest NEP.

[0088] Example 1: As Figure 1 As shown, a forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints includes the following steps:

[0089] S1. Obtain in-situ measurement data of flux stations for the required forest types, and obtain area remote sensing, meteorological and soil data within 10km of the corresponding stations;

[0090] The in-situ measurement data from flux stations are publicly available data on net ecosystem productivity (NEP), ecological respiration (Re), and gross primary productivity (GPP) measured at flux stations.

[0091] The required remote sensing data include: leaf area index (LAI), enhanced vegetation index (EVI), normalized difference vegetation index (NDVI), chlorophyll fluorescence (SIF), and chlorophyll concentration (LCC).

[0092] The required remote sensing data include: air temperature (Ta), vapor pressure difference (VPD), and solar radiation (SW).

[0093] The required soil data include: soil nitrogen content (Sn), soil pH, soil organic carbon concentration (SOC), and soil moisture (SM).

[0094] The required remote sensing, meteorological, and soil data for the site locations need to be extracted from spatial data within a 10km radius.

[0095] In a specific case, flux data came from 54 flux sites in the China Flux Consortium and the FLUX2015 dataset. Specifically, LAI data used MODIS MOD15A2H; NDVI and EVI data used MODIS MCD12C1; SIF data used GOSIF; LCC data used the 2002-2019 global 8-day 500-meter leaf chlorophyll content product dataset; SIF data used the GOSIF dataset; soil moisture (SM), soil organic carbon (SOC), soil pH, and soil total nitrogen (Sn) data used the SoilGrids dataset; SM data was downloaded from the GLEAM dataset; shortwave incident radiation (SR), air temperature (Ta), and dew point temperature (Td) came from the ERA5 dataset; VPD was obtained from atmospheric temperature and dew point temperature.

[0096] S2. Calculate the photosynthetically active radiation intercepted by the top and bottom canopies of the forest at the site. , Furthermore, the remote sensing, meteorological, and soil data around the station obtained in step S1 are preprocessed to generate a specified k*k image patch with a spatial resolution of 500m and a temporal resolution of 8-day, and the time is aligned with the flux data; thus, the preprocessed flux, remote sensing, meteorological, and soil datasets are obtained.

[0097] Regarding the setting of k, in this embodiment k=7, in other embodiments k can be 9 or 11;

[0098] Preprocessing includes:

[0099] Data within a 10km range is resampled to 500m using bilinear interpolation. If the spatial resolution of the dataset used is exactly 500m, resampling is not necessary.

[0100] The day-scale flux data published by the site were overlaid on the 8-day scale, while some abnormal data were deleted, including parameter observation gaps, data that did not conform to crop growth patterns, and data with abnormal observation values.

[0101] Remote sensing data, meteorological data, and soil moisture are adjusted using cubic spline interpolation to change the temporal resolution, aligning the 8-day data of their flux.

[0102] , The APAR absorbed by the top and bottom canopies of a forest canopy is calculated as follows:

[0103] ;

[0104] ;

[0105] In the above formula, It is the surface albedo; This refers to the scattered photosynthetically effective radiation below the canopy; This refers to the radiation scattering term within the canopy; The leaf tilt angle is set to 60°. The solar zenith angle; and These are the diffuse and direct components of the incident PAR, estimated from the incident PAR and the clear sky index. For the top canopy LAI, For the bottom canopy LAI; It is a vegetation aggregation index, which depends on the forest type;

[0106] S3. Combine the flux, remote sensing, meteorological, and soil datasets obtained in step S2, and divide them into training, testing, and validation datasets; specifically:

[0107] The preprocessed 8-day scale meteorological, remote sensing, and soil data obtained in step S2 are combined with the 8-day scale flux data to achieve time alignment; Ta, EVI, SM, SOC, and pH constitute the Re estimation data, with a data structure of k*k*5; APARsu, Ta, VPD, SW, SIF, LCC, NDVI, EVI, Sn, and SM constitute The estimated data has a k*k*10 structure; where APARsh, Ta, VPD, SW, SIF, LCC, NDVI, EVI, Sn, and SM constitute the data. The estimated data has a k*k*10 structure and is divided into training, validation, and test sets in a 6:2:2 ratio to obtain the training data for the three-branch EMA-CKAN model. Simultaneously, it should be ensured that data from each site appears in the training, validation, and test sets for at least two months.

[0108] S4. Construct a three-branch EMA-CKAN network, refer to... Figure 2 As shown, the three-branch EMA-CKAN network comprises three branch subnets: Branch (input is k×k×10) Branch (input is k×k×10) and Re branch (input is k×k×5);

[0109] and The calculation method is as follows:

[0110] ;

[0111] ;

[0112] in, , The output results of the EMA-CKAN subnetworks representing the upper and lower canopies of the forest are provided by... , , , , , , , and The simulation yielded the following results; and The actual forest top and bottom canopy GPP.

[0113] Another independent EMA-CKAN subnetwork estimates Re by combining Ta, EVI, SM, SOC, and pH parameters; among these variables, Ta represents the influence of climatic factors, SM, SOC, and pH are used to characterize the influence of soil factors, and EVI is used to characterize the influence of remote sensing factors; the Re expression is shown below:

[0114] ;

[0115] After obtaining GPPsu, GPPsh, and Re through the three EMA-CKAN subnetworks mentioned above, the estimated value of forest ecosystem NEP was calculated. :

[0116] ;

[0117] ;

[0118] in, This is the simulated value for NEP; , The sum of the GPP of the top and bottom canopies of the forest is: ; Simulated values ​​representing ecological respiration;

[0119] Each sub-network has the same overall structure. Their first layer is the input layer, which adjusts the input data to fit the network and connects with subsequent network layers. Two 3D CKAN layers are then connected, with 3x3x3 kernels and 16 and 32 filters respectively. The Leak ReLU activation function is used, with α ranging from 0.01 to 0.3.

[0120] Subsequently, the EMA multi-scale attention mechanism is embedded, dividing the channels into 4 groups. For each group, parallel depthwise-separable convolutions with 1*3 kernels are used to calculate the multi-scale spatial response. The responses at each scale are concatenated and then subjected to 1*1 convolutions, followed by BatchNorm normalization to obtain the spatial attention map. The channel dimension is then restored by 1*1 convolution projection, and finally the residual coefficients γ are added back to the original features. EMA enhances the mapping relationship between multi-scale spatial and multi-source ecological information, thereby improving the model's response capability to the nonlinear ecological relationships of multi-source ecological information.

[0121] Subsequently, two 3D CKAN layers are connected, with convolutional kernel size of 3*3*3 and filters of 64 and 128 respectively. The activation function used is the Leak ReLU function, with α ranging from 0.01 to 0.3. After this, a reshaping layer is connected to convert the 3D features into 2D features. Then, four 2D CKAN networks are added, with convolutional kernel size of 3*3 and the activation function used is the Leak ReLU function, with α ranging from 0.01 to 0.3. In this embodiment, α is 0.1, and filters are 128, 128, 64, and 32 respectively. After this, a 2*2 pooling layer is added. Then, a reshaping layer is added, followed by three fully connected layers with 32, 16, and 1 neural network in each layer.

[0122] The two-dimensional and three-dimensional CKAN network layers mentioned above are both constructed using KAN (Kolmogorov-Arnold Network). Compared with conventional CNNs, they can achieve essentially the same accuracy with fewer parameters.

[0123] Optimization of the three-branch EMA-CKAN network includes:

[0124] The training and validation datasets obtained in steps S3 and S4 are imported into the three-branch EMA-CKAN network. This model uses remote sensing, meteorological, and soil data as input, and then uses the three-branch EMA-CKAN network to learn the total primary productivity (GPP) of the top and bottom canopies. s ᵤ、GPP sh Using Re and remote sensing data, meteorological data, and soil data, the physical model parameter calibration process is transformed into an end-to-end neural network training task, combined with ecological constraint functions:

[0125] ;

[0126] The training process is repeated multiple times until the accuracy of the validation dataset for the three-branch EMA-CKAN network converges. The coefficient of determination (R²) is then used. 2 The root mean square error (RMSE) metric is used to check the fitting performance of the three-branch EMA-CKAN network to the training samples, and the trained three-branch EMA-CKAN network is saved. In this embodiment, λ is 0.1. A higher overall weight will increase the physical generalization of the network, but will make the network difficult to converge. m It is 0.6, w e The value is 0.3, w t It is 0.1.

[0127] In the ecological constraint function, the mass constraint forces consistency of material budget when predicting each carbon flux component, preventing the model from reducing the overall error by adjusting a single component; the energy constraint drives the network's GPP prediction to be statistically consistent with this physical reference, incorporating the light energy-GPP ecological process into the training process to improve the model's generalization ability under different light conditions and canopy structures; the temperature constraint embeds the sensitivity of ecological respiration to temperature as a physical prior into the network, providing the physical reference of Re through a differentiable temperature response submodule.

[0128] All input variables of the network, including remote sensing, meteorological and soil data, are normalized using min-max before being input into the deep learning model, which linearly maps the original values ​​to the range of 0-1, improving the stability and convergence speed of model training; the Adam optimizer is selected as the network optimizer.

[0129] refer to Figure 3 As shown, the coefficient of determination (R²) 2 The root mean square error (RMSE) is used to test the linear correlation between predicted and observed values, and it is used to quantify the absolute deviation. The calculation is as follows:

[0130] ;

[0131] ;

[0132] like Figure 3 As shown, the NEP R2 of the four forest types (mixed forest in the upper left, deciduous broad-leaved forest in the upper right, evergreen coniferous forest in the lower left, and evergreen broad-leaved forest in the lower right) are all above 0.7. The scattered points are concentrated and closely distributed on both sides of the reference line, which shows that the model has good consistency and accuracy throughout the entire NEP value range, and does not show any obvious systematic overestimation or underestimation.

[0133] For the ecological constraint function:

[0134] ;

[0135] in, For NEP prediction bias, For quality constraints, For energy confinement, For temperature constraints, The overall weight λ ranges from 0 to 0.2. In this embodiment, the overall weight λ is 0.1. A higher overall weight increases the network's physical generalization ability but makes the network more difficult to converge. , , All are non-negative numbers, and + + =1;

[0136] The NEP prediction bias is represented by the RMSE loss function. This loss function penalizes larger errors more severely to avoid extreme errors, and is calculated as follows:

[0137] ;

[0138] Loss mass As a quality constraint, a variation of RMSE loss, it is used to constrain the joint simulation accuracy of the three-branch EMA-CKAN network for the three components of carbon flux. It requires that the predicted NEP, GPP, and respiration (Re) satisfy material budget consistency at both the sample and aggregate levels. The constraint is defined as follows:

[0139] ;

[0140] Where n is the number of samples used for calculation; , , These represent net ecosystem productivity, gross primary productivity, and ecological respiration, as predicted by the three-branch EMA-CKAN model. , , These are the corresponding station observations;

[0141] The energy-constrained calculation is used to determine whether the canopy's efficiency in utilizing light energy conforms to known physical mechanisms; its loss is defined as:

[0142] ;

[0143] Where n is the number of samples used for calculation; This represents the net ecosystem production predicted by the three-branch EMA-CKAN, while To correspond to the relationship between light energy absorbed by the forest canopy and its conversion into fixed biomass energy, the total primary productivity of the entire canopy is calculated using the following method:

[0144] ;

[0145] In the formula, , and are correction factors for maximum VPD and minimum photosynthetic temperature, respectively, representing the effects of saturated vapor-water pressure difference and minimum air temperature on vegetation photosynthesis; and The data was obtained through bilinear regression of flux data. Minimum temperature, The soil moisture content correction factor is calculated as follows:

[0146] ;

[0147] ;

[0148] In the formula, , , , These are parameters related to forest type, obtained through time-series variation analysis of station flux data.

[0149] Temperature constraints are used to characterize the sensitivity of ecological respiration to temperature and to constrain the prediction of Re. These constraints introduce a physical reference model for the temperature response. The parameters of this model can be initialized station-by-site or by ecological class using field data from flux observation stations, and can be fine-tuned during training within physically acceptable ranges. The constraint expression is as follows:

[0150] ;

[0151] in This is a predicted value for ecological respiration. The physical reference values ​​for ecological respiration, taking temperature into account, are calculated as follows:

[0152] ;

[0153] in Select 10°C or 15°C to represent the reference temperature. As the baseline breathing intensity, The eco-respiration intensity coefficient is measured by selecting flux observation stations. S is obtained through regression estimation, as shown in the following formula:

[0154] ;

[0155] ;

[0156] ;

[0157] In the formula, T represents the flux station measurement data;

[0158] S5. Acquire remote sensing, meteorological and soil data of the target forest area, preprocess all data of the target area using the method in step S2, and obtain a multi-source prediction dataset of the target area with the same format as the training dataset. Use the trained three-branch EMA-CKAN network saved in step S4 to simulate the NEP of the target forest area and obtain the forest NEP results.

[0159] In this embodiment, the selected study area is a global mixed forest, deciduous broad-leaved forest, evergreen coniferous forest and evergreen broad-leaved forest region, and the selected time period is the whole year of 2015.

[0160] Example 2: This invention also discloses a forest NEP simulation system based on a three-branch EMA-CKAN and ecological constraints, characterized in that it includes:

[0161] Sample Data Acquisition Unit: This unit is responsible for acquiring station flux data for training, as well as collecting remote sensing, soil, and meteorological data within a 10 km radius of the station to match the flux observation scale. The collected data includes LAI, EVI, NDVI, SIF, and LCC remote sensing data; Ta, VPD, and PAR meteorological variables; and Sn, pH, SOC, and SM soil variables.

[0162] Data Preprocessing Unit: This unit calculates APARsu and APARsh, unifies data from different sources and with different time resolutions to the 8-day time scale required by the model, and performs outlier cleaning and imputation. Daily throughput data is superimposed or averaged to the 8-day scale according to rules, while outlier training data is removed. Finally, this unit outputs aligned multi-source image patch data, performs min-max normalization on it to improve training stability and convergence speed, and divides it into training, validation, and test sets in a 6:2:2 ratio.

[0163] Model Training Unit: This unit is responsible for building and training the three-branch EMA-CKAN model, combining ecological constraints with real-end-to-end forest NEP estimation, using the Adam optimizer for weight updates during training, saving model weights and training history after training, and finally providing quantitative indicators such as the coefficient of determination (R²) and RMSE on the independent test set as the basis for performance evaluation. The model parameters are saved after training.

[0164] NEP Simulation Unit: This unit primarily uses a pre-trained three-branch EMA-CKAN for batch inference and spatialization output of the target area. It acquires and resamples remote sensing, meteorological, and soil data of the target area using the same preprocessing procedure as during training, obtaining an 8-day scale prediction dataset. This dataset is then input into the model pixel-by-pixel or in blocks for computation. , The NEP spatiotemporal sequence is synthesized with Re and finally generated to produce NEP data for the target forest region.

[0165] The embodiments of the present invention have been described in detail above with reference to the examples. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints, characterized in that, include: A three-branch EMA-CKAN network was constructed, comprising three parallel EMA-CKAN subnetworks. Forest carbon flux was decomposed into three subproblems: top canopy total primary productivity (GPP), bottom canopy GPP, and ecological respiration (Re). Each subproblem was modeled separately, and NEP was synthesized within the network in a differentiable form based on an ecological constraint function. The ecological constraint function is as follows: ； in, For NEP prediction bias, For quality constraints, For energy confinement, For temperature constraints, This represents the overall weight, with a value between 0 and 0.

2. , , All are non-negative numbers, and + + =1; as well as An EMA multi-scale attention mechanism is embedded after the first two layers of 3D CKAN in each subnet, and two layers of 3D CKAN are stacked after the EMA multi-scale attention mechanism to perform deeper semantic extraction and fusion of attention-enhanced features. Optimizing the construction of the three-branch EMA-CKAN network includes: Construct a dataset that includes in-situ measurement data of flux stations for the required forest types, remote sensing data of the area within 10km of the corresponding stations, meteorological data, and soil data; The dataset is divided into training dataset, test dataset, and validation dataset; Using remote sensing data, meteorological data, and soil data as inputs to a three-branch EMA-CKAN network, the network learns the correlation between total primary productivity (TPP) and ecological respiration (Re) of the top and bottom canopies and the data from remote sensing, meteorological, and soil data. The physical model parameter calibration process is transformed into an end-to-end neural network training task. Ecological constraint functions are calculated, and the network is trained repeatedly until the accuracy of the validation dataset of the three-branch EMA-CKAN network converges. The coefficient of determination and root mean square error index are used to examine the fitting performance of the three-branch EMA-CKAN network to the training samples, and the optimal three-branch EMA-CKAN network is obtained. Before splitting the dataset, the dataset is preprocessed, including: The effective photosynthetically active radiation intercepted by the top and bottom canopies of the forest at the calculation site; ； ； For photosynthetically effective radiation from the top canopy, This provides effective photosynthetic radiation for the bottom canopy. It is the surface albedo; This refers to the scattered photosynthetically effective radiation below the canopy; This refers to the radiation scattering term within the canopy; For the leaf angle; The solar zenith angle; and These are the diffuse and direct components of the incident solar radiation, respectively, estimated from the incident solar radiation and the clear sky index; The leaf area index of the top canopy. The leaf area index of the bottom canopy; It is a vegetation aggregation index, which depends on the forest type; The day-scale flux data published by the site were overlaid on the 8-day scale, while some abnormal data were deleted, including parameter observation gaps, data that did not conform to crop growth patterns, and data with abnormal observation values. The domain remote sensing data, meteorological data, and soil data were resampled to 500m using bilinear interpolation. The temporal resolution of remote sensing data, meteorological data, and soil data was altered using cubic spline interpolation to align them with the 8-day flux data. The processed 8-day scale remote sensing data, meteorological data, and soil data were combined with the 8-day scale flux data to ensure time alignment; they were divided into training, validation, and test sets in a 6:2:2 ratio; and the data from each station appeared in the training, validation, and test sets for more than two months.

2. The forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints according to claim 1, characterized in that: The in-situ measurement data from the flux stations include net ecosystem productivity (NEP), ecological respiration (Re), and total primary productivity (GPP) data. Remote sensing data include leaf area index (LAI), enhanced vegetation index (EVI), normalized difference vegetation index (NDVI), chlorophyll fluorescence index (SIF), and chlorophyll concentration (LCC). Meteorological data include: air temperature (Ta), saturated vapor pressure difference (VPD), and solar radiation (SW); Soil data include soil nitrogen content (Sn), soil pH, soil organic carbon concentration (SOC), and soil moisture (SM); The estimated data consist of air temperature Ta, enhanced vegetation index EVI, soil moisture SM, soil organic carbon concentration SOC, and soil pH Re. Effective photosynthetic radiation of the top canopy Composition of air temperature (Ta), saturated vapor pressure difference (VPD), solar radiation (SW), soil moisture (SM), chlorophyll fluorescence (SIF), chlorophyll concentration (LCC), normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), soil nitrogen content (Sn), and soil moisture (SM) Estimated data; Effective photosynthetic radiation of the bottom canopy Composition of air temperature (Ta), saturated vapor pressure difference (VPD), solar radiation (SW), chlorophyll fluorescence (SIF), chlorophyll concentration (LCC), normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), soil nitrogen content (Sn), and soil moisture (SM). Estimated data.

3. The forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints according to claim 2, characterized in that: In a subnet, calculate the ecorespiration Re based on Re estimation data: ； In a subnet, based on Estimate data to calculate the total primary productivity of the top canopy. : ； In a subnet, based on Estimate data to calculate total primary productivity of the bottom canopy. : ； in, , The output results represent the EMA-CKAN subnetworks for the upper and lower canopies of the forest.

4. The forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints according to claim 3, characterized in that: Based on ecological respiration (Re) and total primary productivity of the top canopy and total primary productivity of the bottom canopy Calculate simulated values ​​of forest ecosystem NEP : ； ； in, , The simulated value represents ecological respiration.

5. The forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints according to claim 1, characterized in that: The EMA multi-scale attention mechanism divides the channels into 4 groups. For each group, parallel depthwise-separable convolutions with 1*3 kernels are used to calculate the multi-scale spatial response. The responses of each scale are concatenated and then subjected to 1*1 convolutions, followed by BatchNorm normalization to obtain the spatial attention map. The channel dimension is restored by 1*1 convolution projection, and finally the residual coefficients γ are added back to the original features.

6. The forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints according to claim 5, characterized in that: The stacked two 3D CKAN convolutional layers have kernel sizes of 3*3*3, with 64 and 128 filters respectively, and use the leaky ReLU activation function. After the stacked two 3D CKAN layers, the following are added sequentially: A deformation layer converts three-dimensional features into two-dimensional features; A four-layer 2D CKAN network with a 3*3 kernel size, using the Leak ReLU activation function, and filters of 128, 128, 64 and 32 respectively; A 2x2 pooling layer; A deformable layer; The three fully connected layers have 32, 16, and 1 neural network layers, respectively.

7. The forest NEP simulation method based on three-branch EMA-CKAN and ecological constraints according to claim 6, characterized in that: In the ecological constraint function: Energy constraints Specifically: ； Where n is the number of samples used for calculation; This represents the net ecosystem productivity predicted by the three-branch EMA-CKAN network, while To correspond to the relationship between light energy absorbed by the forest canopy and its conversion into fixed biomass energy, the total primary productivity of the entire canopy was calculated using the following method: ； in, , and are correction factors for maximum VPD and minimum photosynthetic temperature, respectively, representing the effects of saturated vapor-water pressure difference and minimum air temperature on vegetation photosynthesis; and Obtained by bilinear regression of flux data; minimum temperature, The soil moisture content correction factor is calculated as follows: ； ； in, , , , These are parameters related to forest type, obtained through time-series variation analysis of station flux data; Temperature constraint Specifically: ； in, This is a predicted value for ecological respiration. The physical reference value for ecological respiration, taking temperature into account, is calculated as follows: ； in, As the reference temperature, As the baseline breathing intensity, The eco-respiration intensity coefficient is measured by selecting flux observation stations. S is obtained through regression estimation, as shown in the following formula: ； ； ； Where T represents flux station measurement data.

8. A system based on the forest NEP simulation method based on the three-branch EMA-CKAN and ecological constraints as described in any one of claims 1-7, characterized in that: include: Acquisition of sample data unit: Acquire the in-situ measurement data of the required forest type flux stations, the area remote sensing data within 10km of the corresponding stations, meteorological data and soil data, and match them with the flux observation scale; Data preprocessing unit: Calculates photosynthetically active radiation of the top canopy. and photosynthetically effective radiation of the bottom canopy The data from different sources and with different time resolutions were unified to the 8-day time scale required by the model, and outlier cleaning and imputation were completed. The daily throughput data is superimposed or averaged to an 8-day scale according to rules, while abnormal training data is removed. The output is aligned multi-source image patch data, min-max normalized, and divided into training, validation and test sets in a 6:2:2 ratio; Model training unit: Construct and train a three-branch EMA-CKAN model, combine ecological constraint functions to estimate forest NEP in real end-to-end, use Adam optimizer to update weights during training, save model weights and training history after training, and finally use the coefficient of determination and root mean square error index on an independent test set to check the fitting performance of the three-branch EMA-CKAN network to the training samples, and save the optimal three-branch EMA-CKAN network. NEP simulation unit: Configured with an optimal three-branch EMA-CKAN network for batch inference and spatialized output of the target region.

Images