A forest evapotranspiration splitting and correction method and system based on liquid flow- stem flow- vorticity collaborative constraint

Abstract

The present application belongs to the technical field of ecological hydrology and land ecosystem carbon and water cycle observation, and discloses a forest evapotranspiration splitting and correction method and system based on liquid flow- stem flow- vortex cooperation constraint, which solves the problems of low physical credibility of evapotranspiration component splitting, non-closed energy of vortex flux, difficult correction of system error and lack of assimilation framework of multi-source observation data existing in the prior art of forest evapotranspiration observation technology. Firstly, the present application unifies the time and space datum of multi-source observation data, obtains the true value of stand transpiration based on the upscaled liquid flow data, inverses the canopy interception amount based on the precipitation redistribution balance equation, constructs a physical neural network embedded with energy balance and evapotranspiration mechanism, fuses the measured true value soft constraint to construct a composite loss function, realizes the collaborative splitting and flux correction of three types of evapotranspiration components, i.e. forest canopy transpiration, canopy interception evaporation and soil evaporation, and completes the quantification of result uncertainty in combination with the Monte Carlo method.

Description

Technical Field

[0001] This invention belongs to the field of ecohydrology and terrestrial ecosystem carbon and water cycle observation technology. Specifically, it relates to a method and system for forest evapotranspiration decomposition and correction based on sap flow-stem flow-eddy co-constraint. It is used for forest ecosystem eddy flux error correction and the quantification of total forest evapotranspiration into three components: canopy evapotranspiration, canopy intercepted evaporation, and understory soil evaporation. It can be widely applied to forest hydrological mechanism research, terrestrial carbon and water cycle simulation, regional forest water resource assessment, and the construction of surface parameterization models for global climate change. Background Technology

[0002] Against the backdrop of in-depth research on global climate change and terrestrial ecosystem carbon and water cycles, forest evapotranspiration, as a core flux connecting ecosystem water cycles and energy balance, has become a key issue in ecohydrology and global change research due to the need for precise component analysis and quantification. Forest evapotranspiration consists of multiple components, including canopy transpiration, understory evapotranspiration, and canopy intercepted evaporation. The response mechanisms of each component to environmental changes differ significantly, and precise analysis is crucial for understanding forest water use strategies and assessing ecosystem climate feedback.

[0003] Existing studies generally use a single observation method for analysis, lacking a standardized and physical data fusion bridge between the true value of evapotranspiration at the individual scale and the total eddy flux at the stand scale. This results in the inability to simultaneously guarantee the accuracy of evapotranspiration component separation and the reliability of eddy flux correction, specifically including:

[0004] (1) Weak physical basis for splitting evapotranspiration components: Existing splitting schemes mostly use fixed empirical coefficients and simplified empirical models to split total evapotranspiration, without incorporating measured constraints of the canopy precipitation redistribution process, resulting in extremely high uncertainty in the estimation results of canopy interception evaporation.

[0005] (2) Flux correction and component separation are separated: Traditional energy non-closed correction relies only on statistical residual correction and energy balance ratio parameter adjustment, without introducing the physical process constraint of canopy transpiration, and cannot use the measured true value of liquid flow to constrain the eddy system deviation. Nighttime flux is underestimated and systematic positive and negative deviations are difficult to identify and compensate.

[0006] (3) There is no collaborative assimilation framework for multi-source observation data: the four types of observation data, namely sap flow, trunk stem flow, forest penetration rain and eddy flux, are processed and calculated offline. There is no correlation assimilation system based on the physical laws of water and energy conservation. The closed-loop technical process of "observation and acquisition - physical constraints - flux correction - component splitting" is missing, which limits the development of multi-source flux fusion algorithms.

[0007] The shortcomings of the aforementioned traditional methods result in three major bottlenecks in the accurate quantification of forest evapotranspiration:

[0008] First, the component breakdown results lack verification of the true values ​​of individual wood samples, making it impossible to quantitatively distinguish the contribution ratio of the three types of evapotranspiration components to total moisture loss and the environmental driving mechanism.

[0009] Secondly, the eddy flux energy non-closed system error has not been effectively eliminated, and distorted flux data will distort the spatiotemporal distribution characteristics of regional forest carbon cycle and reduce the accuracy of climate model surface parameterization schemes.

[0010] Third, the existing data processing flow is offline and independent, and it does not realize an intelligent assimilation algorithm that deeply couples physical mechanism equations with neural networks, so it cannot simultaneously complete eddy flux correction and dynamic separation of multi-component evapotranspiration. Summary of the Invention

[0011] The technical problem to be solved by this invention is to provide a method and system for forest evapotranspiration decomposition and correction based on sap flow-stem flow-eddy current co-constraint, which solves the problems of low physical reliability of evapotranspiration component decomposition, non-closed energy of eddy current flux, difficulty in correcting systematic errors, and lack of assimilation framework for multi-source observation data in existing forest evapotranspiration observation technologies.

[0012] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0013] On the one hand, this invention provides a method for forest evapotranspiration decomposition and correction based on sap flow-stem flow-eddy current co-constraints, comprising the following steps:

[0014] S1. Collect sap flow data, trunk stem flow data, canopy penetration rainfall data, external rainfall data, and total eddy evapotranspiration observation data of the target forest and perform preprocessing;

[0015] S2. Using the liquid flow data as the true value of canopy transpiration at the individual scale, and combining it with the diameter at breast height or leaf area parameters, perform upscaling calculations to obtain the stand transpiration.

[0016] S3. Construct a precipitation redistribution balance equation and deduce the canopy interception by combining canopy penetration rainfall data, trunk stem flow data, and extraforest rainfall data.

[0017] S4. Construct a physical information neural network that integrates surface energy balance and Penman-Monteith evapotranspiration mechanism. Use the forest stand evapotranspiration and canopy interception as soft constraints to embed the network composite loss function. Use total eddy evapotranspiration observation data, meteorological driving variables and soil moisture state variables as inputs to complete the training of the physical information neural network.

[0018] S5. The trained physical information neural network outputs the corrected three types of evapotranspiration components: canopy transpiration, canopy intercepted evaporation, and soil evaporation.

[0019] In this scheme, based on field observation data of sap flow, stem flow, rainfall distribution, and eddy flux, vegetation physiological true value constraints and forest hydrological true value constraints are constructed respectively. At the same time, surface energy balance and Penman-Monteith evapotranspiration physical mechanism constraints are coupled to form a composite correction mechanism of "data fitting + physiological true value constraints + hydrological true value constraints + energy mechanism constraints". The nonlinear mapping between environmental driving characteristics and evapotranspiration components is realized through physical information neural network. The component estimation bias is corrected by dual measured true value soft constraints. The rationality of the model derivation process is constrained by physical conservation rules. The total evapotranspiration flux is aligned with field observations, and each component of evapotranspiration is aligned with the objective laws of vegetation physiology and forest hydrology. This effectively improves the reliability and physical consistency of the estimation of the three types of forest evapotranspiration components.

[0020] Furthermore, in step S1, the methods for collecting sap flow data, stem flow data, canopy penetration rainfall data, external rainfall data, and total eddy evapotranspiration observation data of the target forest include:

[0021] Liquid flow data of individual logs were collected using a thermal diffusion liquid flow probe;

[0022] Tree stem flow data were collected using a tipping bucket flow meter;

[0023] Canopy penetration rainfall data were collected using a forest understory rain gauge array.

[0024] Total rainfall data outside the forest were collected using a standard rain gauge.

[0025] The total vorticity evapotranspiration observation data were obtained by collecting water vapor flux using an vorticity correlation system and converting it.

[0026] In this scheme, various data acquisition devices are adapted to the observation characteristics of individual tree transpiration physiological processes, canopy hydrological redistribution processes, and regional-scale energy flux exchange processes. They can obtain raw observation information from multiple dimensions, including individual vegetation, canopy hydrology, and ecosystem fluxes, ensuring that each type of input data corresponds to the physical process of evapotranspiration decomposition. This guarantees the relevance, effectiveness, and high accuracy of the input data from the source of observation, providing a reliable raw data foundation for subsequent truth value calculation, mechanism constraints, and model training.

[0027] Furthermore, in step S1, the preprocessing includes:

[0028] Missing value imputation, outlier removal, component error correction, unit unification, and aggregation with natural days as the time window are used to obtain a daily-scale dataset.

[0029] The sub-error corrections include: nighttime zero-flow baseline correction and temperature drift correction for sap flow data; splashing error correction and installation tilt correction for trunk stem flow and canopy penetration rainfall data; and coordinate rotation, WPL density effect correction, and frequency response correction for total eddy evapotranspiration observation data.

[0030] This scheme achieves effective matching and fusion of multi-source heterogeneous observation data by unifying the time scale, data format and units, ensuring that the sap flow, hydrology and flux data strictly correspond on the daily time scale and that the physical processes are consistent. It avoids the transmission of original observation noise and systematic errors to the downstream true value calculation and model training stages, and ensures the accuracy and stability of subsequent stand evapotranspiration, canopy interception inversion results and model training samples.

[0031] Furthermore, in step S2, the sap flow data is used as the true value of canopy transpiration at the individual scale, and combined with the diameter at breast height (DBH) or leaf area parameters, upscaling calculation is performed to obtain the stand transpiration rate, specifically including:

[0032] Obtain the diameter at breast height (DBH) of each individual tree within the sample plot. and liquid flow rate Calculate the cross-sectional area of ​​a single piece of wood. ;

[0033] Using the cross-sectional area weighted method, according to the formula Calculate forest stand evapotranspiration ;

[0034] Alternatively, the leaf area index upscaling method can be used, according to the formula. Calculate forest stand evapotranspiration ,in, It represents the leaf area index of a single tree.

[0035] This scheme achieves a scale transformation from individual tree sap flow observations to overall stand transpiration based on forest community structural characteristics. It uses either the basal area reflecting tree transpiration capacity or the leaf area index characterizing canopy photosynthetic transpiration coverage as weighting factors, fully considering the spatial heterogeneity of individual tree size and canopy coverage within the stand. Both upscaling algorithms are constructed based on the physiological correlation characteristics of vegetation transpiration, effectively mitigating the problem of insufficient spatial representativeness in single-point, single-tree observations. This accurately reconstructs the actual daily transpiration total of the entire stand, providing the physical information neural network with transpiration truth value constraints that have clear physiological significance and closely reflect the true state of the stand.

[0036] Furthermore, in step S3, a precipitation redistribution balance equation is constructed, and the canopy interception is derived by combining canopy penetration rainfall data, trunk stem flow data, and extraforest rainfall data. Specifically, this includes:

[0037] Constructing the precipitation redistribution equilibrium equation ,in, This represents the total rainfall outside the forest on a daily scale. Canopy penetration rainfall on a daily scale; This refers to the daily trunk stem flow rate; Canopy interception on a daily scale;

[0038] Based on daily total rainfall outside the forest Daily scale canopy penetration rainfall Daily-scale trunk stem flow By combining the precipitation redistribution balance equation, the daily-scale canopy interception was calculated. .

[0039] This scheme follows the water conservation mechanism of forest canopy precipitation redistribution, precisely decomposing extra-forest rainfall into three coupled components: throughfall, stem flow, and canopy interception. By directly closing the water balance relationship through observational data, it achieves a mechanistic and precise inversion of canopy interception. The obtained interception is entirely derived from objective forest hydrological processes, accurately reflecting the forest canopy's ability to retain, absorb, and redistribute rainfall, providing stable and reliable true-value constraints for subsequent model training based on canopy interception and evaporation hydrological data.

[0040] Furthermore, step S3 also includes: the total rainfall outside the forest on a daily scale. When observations are missing, canopy penetration rainfall on a daily scale is used. Daily-scale trunk stem flow The sum of these values ​​can be used as a proxy for total precipitation. And introduce a correction factor Corrected daily-scale canopy interception inversion values: The correction coefficient The rate was determined by the measured value of canopy interception in independent sample plots.

[0041] In this scheme, a physically adaptable precipitation proxy and interception correction mechanism is constructed to address the practical scenario of missing field rainfall observation data. This effectively solves the problem of rainfall data gaps caused by extreme weather and equipment failures, which prevent the calculation of canopy interception. By calibrating correction coefficients through measured sample plots, the water volume deviation caused by proxy precipitation data can be accurately compensated, ensuring that the canopy interception inversion results during periods of missing data still conform to the actual hydrological patterns, thus improving the overall data tolerance of the method.

[0042] Furthermore, in step S4, the physical information neural network includes an input layer, a shared hidden layer, a physical constraint layer, and an output layer connected in sequence.

[0043] The input layer is used to receive daily-scale total eddy evapotranspiration, meteorological driving variables and soil moisture state variables as input features, and to send all input features to the shared hidden layer.

[0044] The shared hidden layer is composed of multiple stacked fully connected layers. It is used to receive the feature vector output by the input layer and complete the unified nonlinear feature mapping to extract deep features that fuse environmental and throughput information, and then pass the deep features to the physical constraint layer.

[0045] The physical constraint layer incorporates the surface energy balance equation and the Penman-Monteith evapotranspiration equation. During network forward propagation, this layer receives the deep feature vectors output from the shared hidden layer, adapts and transforms the deep features, and then directly transmits them downstream to the multi-branch output layer for component prediction. During the model training phase, this layer introduces two types of measured true values: stand evapotranspiration and canopy interception. Combined with the surface energy balance constraint conditions, it constructs evapotranspiration true value constraint terms, interception true value constraint terms, and energy balance constraint terms, which together constitute a composite loss function.

[0046] The output layer is provided with three independent output branches that do not interfere with each other. These branches receive the features transmitted from the physical constraint layer and perform calculations on them respectively, outputting the corrected canopy transpiration, canopy intercepted evaporation, and soil evaporation.

[0047] In this scheme, the network architecture achieves a deep integration of data-driven feature learning and physical mechanism constraints. It aggregates multiple environmental and flux-driving factors through the input layer, mines nonlinear correlation features of the evapotranspiration process based on the shared hidden layer, and embeds mature evapotranspiration and energy balance mechanisms through the physical constraint layer, overcoming the deficiency of traditional pure data networks lacking physical priors. During the training phase, multiple soft constraints—including vegetation physiological truth values, hydrological truth values, and energy mechanisms—are applied simultaneously. Combined with a multi-branch independent output structure, this achieves decoupling and decomposition of the three types of evapotranspiration components, ensuring that the decomposition results possess both data fitting accuracy and physical rationality at the network structure level.

[0048] Furthermore, the composite loss function is expressed as:

[0049] ;

[0050] in, Total loss; The mean square error data fitting term between the total evapotranspiration and eddy covariance total evapotranspiration observation data output by the physical information neural network is obtained by summing the corrected canopy evapotranspiration, canopy intercepted evapotranspiration, and soil evapotranspiration. For the evaporation truth constraint term, ,in, This is the corrected canopy transpiration. Forest stand transpiration; For energy balance constraints, ,in, This is the net radiation level. For soil heat flux, For sensible heat flux, Latent heat flux; To retain the truth value constraint term, ,in, This is the corrected canopy interception evaporation. The inverse daily-scale canopy interception; , , These are the weight coefficients corresponding to the transpiration true value constraint, the energy balance constraint, and the retention true value constraint, respectively.

[0051] In this scheme, the composite loss function constructs a multi-dimensional collaborative constraint training and correction mechanism. The mean square error data fitting term ensures that the overall evapotranspiration flux of the model remains consistent with field eddy current observations. The estimation accuracy of the two core evapotranspiration components is constrained based on true evapotranspiration loss and true cutoff loss, respectively. The energy balance loss forces the model output to satisfy the physical law of surface energy conservation. These multiple constraint terms cooperate and complement each other, simultaneously correcting the model's total bias, component bias, and mechanistic bias, achieving comprehensive collaborative optimization of observational data, vegetation physiology, forest hydrology, and energy mechanisms.

[0052] Furthermore, in step S5, the trained physical information neural network outputs the corrected three types of evapotranspiration components: canopy transpiration, canopy intercepted evaporation, and soil evaporation, specifically including:

[0053] The input data to be predicted is fed into the trained physical information neural network for inference. During the inference process, the Dropout mechanism is continuously enabled, and forward propagation inference is repeatedly executed to obtain multiple sets of evapotranspiration component prediction samples. Based on the multiple sets of prediction samples, the predicted mean and prediction variance of canopy evaporation, canopy intercepted evaporation, and soil evaporation are calculated respectively. The predicted mean is used as the corrected canopy evaporation, canopy intercepted evaporation, and soil evaporation outputs, and the prediction variance is used to characterize the uncertainty corresponding to the estimation results of each evapotranspiration component.

[0054] In this scheme, a Monte Carlo Dropout random sampling mechanism is introduced during the model inference stage. Multiple sets of independent prediction samples are generated through multiple forward inferences, and optimal results and uncertainty quantification are achieved based on the statistical mean and variance. Using the sample mean as the final evapotranspiration composition result effectively weakens the prediction bias caused by network random perturbations and improves the stability of the output results. Quantifying the uncertainty of each component estimation using the prediction variance accurately reflects the comprehensive error level caused by data noise, physical constraint adaptation bias, and model structure perturbations. This ensures that the final output evapotranspiration composition not only has high-precision correction results but also includes credibility evaluation information, improving the scientific rigor and engineering applicability of the evapotranspiration separation results.

[0055] On the other hand, the present invention also provides a forest evapotranspiration decomposition and correction system based on sap flow-stem flow-eddy current co-constraints, for implementing the above method, the system comprising:

[0056] The data acquisition module is used to acquire and preprocess sap flow data, trunk stem flow data, canopy penetration rainfall data, external rainfall data, and total eddy evapotranspiration observation data of the target forest.

[0057] The transpiration truth value calculation module is used to take the liquid flow data as the individual-scale truth value of canopy transpiration, and combine it with the diameter at breast height or leaf area parameters to perform upscaling calculation to obtain the stand transpiration.

[0058] The precipitation redistribution inversion module is used to construct the precipitation redistribution balance equation and invert the canopy interception by combining canopy penetration rainfall data, trunk stem flow data, and extraforest rainfall data.

[0059] The physical information neural network training module is used to construct a physical information neural network that integrates surface energy balance and Penman-Monteith evapotranspiration mechanism. The forest stand evapotranspiration and canopy interception are used as soft constraints to embed the network composite loss function. The total eddy evapotranspiration observation data, meteorological driving variables and soil moisture state variables are used as inputs to complete the training of the physical information neural network.

[0060] The evapotranspiration component prediction module is used to feed the input data to be predicted into the trained physical information neural network to carry out inference and output the corrected three types of evapotranspiration components: canopy transpiration, canopy intercepted evaporation, and soil evaporation.

[0061] In this scheme, the system adopts a modular and hierarchical architecture design, with each functional module having corresponding functions and methods, fully covering the technical chain of data preprocessing, evapotranspiration truth value solving, truncation truth value inversion, mechanism network training, and evapotranspiration component prediction.

[0062] The beneficial effects of this invention are:

[0063] (1) Based on the true value of forest stand evaporation obtained by single-tree sap flow upscale and the true value of canopy interception obtained by precipitation redistribution balance inversion, this invention implements precise soft constraints on evaporation and interception evaporation components respectively. At the same time, it couples the surface energy balance and the Penman-Monteith evapotranspiration mechanism to physically regulate the model extrapolation process, so that the evapotranspiration components output by the model not only fit the overall observation results of eddy flux, but also strictly conform to the objective laws of forest vegetation physiological metabolism and canopy hydrological cycle, fundamentally improving the physical interpretability and estimation accuracy of evapotranspiration decomposition results.

[0064] (2) The physical constraint layer of the physical information neural network in this invention undertakes the functions of mounting all physical rules and truth constraints and gradient control. During the training phase, the network weight parameters are optimized in reverse through a multi-constraint composite loss function, realizing the deep integration of "data adaptive learning + physical mechanism domestication". It not only retains the strong fitting ability of the neural network to the nonlinear coupling relationship of complex environment, but also completely solves the problems of energy non-closure, hydrological process distortion and inconsistent component estimation bias that are easy to occur in pure data model, realizing the complementary unity of data advantage and mechanism advantage.

[0065] (3) The present invention introduces the Monte Carlo Dropout uncertainty quantification mechanism in the evapotranspiration prediction process. While outputting high-precision corrected evapotranspiration components, it simultaneously quantifies the estimated dispersion and confidence level of each component, which can provide a refined evapotranspiration dataset with both high precision and credibility for forest eco-hydrological process analysis, water resource assessment and ecological model simulation. Attached Figure Description

[0066] Figure 1 This is a flowchart of the forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy co-constraint in this invention.

[0067] Figure 2 This is a structural diagram of the forest evapotranspiration splitting and correction system based on sap flow-stem flow-eddy co-constraint in this invention. Detailed Implementation

[0068] This invention aims to provide a method and system for forest evapotranspiration decomposition and correction based on sap flow-stem flow-eddy current co-constraints, solving the problems of low physical reliability of evapotranspiration component decomposition, non-closed energy of eddy current flux, difficulty in correcting systematic errors, and lack of assimilation framework for multi-source observation data in existing forest evapotranspiration observation technologies. Its core idea is to integrate multi-source heterogeneous data, including sap flow vegetation physiological observations, stem flow rainfall hydrological observations, and eddy flux ecological observations, to construct a collaborative correction framework of "measured true value constraints + physical mechanism constraints + flux data fitting". Utilizing the forest precipitation redistribution water balance law and vegetation transpiration physiological characteristics, high-precision measured true values ​​of stand evapotranspiration and canopy interception are obtained by inversion, and these are used as soft constraints embedded in a physical information neural network. Simultaneously, it nests the core physical mechanisms of surface energy balance and Penman-Monteith evapotranspiration, constructing a composite loss function containing data fitting terms, physiological true value constraints, hydrological true value constraints, and energy balance constraints. Through gradient backpropagation during the training phase, the network weight parameters are optimized, ensuring the model's ability to fit field observation data while achieving physical mechanism regulation and systematic error correction for evapotranspiration component decomposition. Furthermore, it includes a data missing fault tolerance compensation mechanism and an inference uncertainty quantification mechanism to complete multi-source data assimilation, model constraint training, accurate evapotranspiration component decomposition, and error assessment, achieving high-precision, high-reliability, and high-physical-consistency refined estimation of forest evapotranspiration components.

[0069] In terms of specific implementation, the implementation process of the forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy current co-constraint provided by this invention is described in [link to relevant documentation]. Figure 1 It includes the following steps:

[0070] S1. Acquisition and preprocessing of multi-source eco-hydrological observation data:

[0071] In this step, sap flow data, stem flow data, canopy penetration rainfall data, external rainfall data, and total eddy evapotranspiration observation data of the target forest are collected and preprocessed. The purpose is to integrate and acquire multi-dimensional raw observation data on forest vegetation evapotranspiration, canopy precipitation redistribution, and ecosystem evapotranspiration flux. Standardized preprocessing methods such as error correction, anomaly removal, scale unification, and missing data repair are used to eliminate systematic errors and random noise from field observations, constructing a high-quality daily-scale dataset. This provides reliable basic data support for subsequent stand evapotranspiration truth calculations, canopy interception inversion, and physical information neural network model training in this invention, ensuring the accuracy of evapotranspiration decomposition from the data source.

[0072] In one exemplary implementation, sap flow data, stem flow data, canopy penetration rainfall data, total rainfall outside the forest, and total evapotranspiration flux data measured by the eddy covariance system are obtained from the target forest plots over a complete hydrological year. The multi-source data are then subjected to preprocessing such as missing value imputation, outlier removal, component error correction, and unit unification, and aggregated on a daily scale to generate a dataset with a consistent spatiotemporal reference.

[0073] In terms of data acquisition, a thermal diffusion sap flow probe was used to collect sap flow data of individual trees; a tipping bucket flow meter was used to collect stem flow data of tree trunks; a forest rain gauge array was used to collect canopy penetration rainfall data; a standard rain gauge outside the forest was used to collect total rainfall data outside the forest; and an eddy covariance system was used to collect water vapor flux and convert it into total eddy evapotranspiration observation data.

[0074] The sub-error corrections include: nighttime zero-flow baseline correction and temperature drift correction for sap flow data; splash error correction and installation tilt correction for trunk stem flow and canopy penetration rainfall data; and coordinate rotation, WPL density effect correction, and frequency response correction for total eddy evapotranspiration observation data.

[0075] The diurnal-scale aggregation uses an arithmetic average or cumulative summation method with the day boundary as a unified time window to ensure that the four data sources—sap flow, stem flow, rainfall, and eddy current—are strictly aligned in the time dimension.

[0076] S2. Upscaling solution of true stand transpiration based on individual tree sap flow observations:

[0077] In this step, the sap flow data is used as the true value of canopy transpiration at the individual scale. Combined with parameters such as diameter at breast height (DBH) or leaf area, upscaling is performed to obtain the stand transpiration rate. The aim is to construct a true physiological benchmark for canopy transpiration based on high-precision individual tree sap flow observation data. By combining this data with community structure parameters such as DBH and leaf area, a weighted upscaling conversion is performed, overcoming the spatial representativeness limitations of individual tree observations and obtaining a daily-scale stand transpiration rate that truly reflects the growth status of the entire stand. This provides a true-value constraint for the training of the physical information neural network, representing the physiological dimension of vegetation.

[0078] In one exemplary implementation, this step is performed as follows:

[0079] Obtain the diameter at breast height (DBH) of each individual tree within the sample plot. and liquid flow rate Calculate the cross-sectional area of ​​a single piece of wood. ;

[0080] Using the cross-sectional area weighted method, according to the formula Calculate forest stand evapotranspiration ;

[0081] Alternatively, the leaf area index upscaling method can be used, according to the formula. Calculate forest stand evapotranspiration ,in, It represents the leaf area index of a single tree.

[0082] The final calculated forest stand evaporation The physical truth labels for transpiration components will be embedded in the loss function of the physical information neural network as transpiration truth constraint terms. ,in The weighting coefficients for the evaporation truth value constraint term.

[0083] S3. Canopy interception inversion based on precipitation redistribution water balance:

[0084] In this step, a precipitation redistribution balance equation is constructed, and canopy interception is derived by combining canopy penetration rainfall data, stem flow data, and extraforest rainfall data. The aim is to accurately derive daily canopy interception based on the forest canopy precipitation redistribution balance mechanism, by closing the water balance relationship through separate observation data of extraforest rainfall, penetration rainfall, and stem flow, and to construct the true values ​​of the forest hydrological dimension.

[0085] In one exemplary implementation, this step is performed as follows:

[0086] Constructing the precipitation redistribution equilibrium equation ,in, This represents the total rainfall outside the forest on a daily scale. Canopy penetration rainfall on a daily scale; This refers to the daily trunk stem flow rate; Canopy interception on a daily scale;

[0087] Based on daily total rainfall outside the forest Daily scale canopy penetration rainfall Daily-scale trunk stem flow By combining the precipitation redistribution balance equation, the daily-scale canopy interception was calculated. .

[0088] Specifically, the total rainfall outside the forest on that day. When observations are missing, canopy penetration rainfall on a daily scale is used. Daily-scale trunk stem flow The sum of these values ​​can be used as a proxy for total precipitation. And introduce a correction factor Corrected daily-scale canopy interception inversion values: The correction coefficient The values ​​were obtained through measured canopy interception values ​​in independent plots, ranging from 0.8 to 1.2.

[0089] S4. Training of a physical information neural network with multiple physical mechanisms and multiple truth values ​​in a collaborative manner:

[0090] In this step, a physical information neural network (PIN) integrating surface energy balance and the Penman-Monteith evapotranspiration mechanism is constructed. The forest stand evapotranspiration and canopy interception are used as soft-constraints embedded in the network's composite loss function. Eddy total evapotranspiration observation data, meteorological driving variables, and soil moisture state variables are used as inputs to complete the training of the PIN. The aim is to construct a PIN model architecture integrating energy balance and the Penman-Monteith evapotranspiration mechanism, using the true values ​​of forest stand evapotranspiration and canopy interception as soft-constraints embedded in the composite loss function, and completing iterative model training based on multi-source environmental, soil, and flux observation data. Through multi-constraint joint back gradient optimization of network parameters during the training phase, the model simultaneously satisfies the requirements of eddy flux observation fitting, vegetation evapotranspiration physiological laws, canopy hydrological distribution laws, and surface energy conservation laws, achieving deep coupling between data-driven feature learning and physical mechanism constraints, and correcting evapotranspiration component estimation biases at the model level.

[0091] In one exemplary implementation, this step is performed as follows:

[0092] The Physical Information Neural Network (PINN) is built upon the existing, well-known classical PINN architecture, with a fully connected multilayer perceptron as its backbone. This invention adds a physical constraint layer to this structure. The specific architecture of the PINN in this invention is as follows:

[0093] The input layer receives daily total eddy evapotranspiration, meteorological driving variables, and soil moisture state variables as input features, and then sends all input features to the shared hidden layer.

[0094] The shared hidden layer, composed of multiple stacked fully connected layers, is used to receive the feature vectors output by the input layer and complete the unified nonlinear feature mapping, extract deep features that fuse environmental and throughput information, and then pass the deep features to the physical constraint layer.

[0095] The physical constraint layer incorporates the surface energy balance equation and the Penman-Monteith evapotranspiration equation. During network forward propagation, this layer receives deep feature vectors from the shared hidden layer, adapts and transforms the deep features, and then directly transmits them downstream to the multi-branch output layer for component prediction. During model training, this layer introduces two types of measured true values: stand evapotranspiration and canopy interception. Combined with the surface energy balance constraint, it constructs true evapotranspiration constraint terms, true interception constraint terms, and energy balance constraint terms, which together form a composite loss function.

[0096] The output layer has three independent output branches that do not interfere with each other. They receive the features transmitted from the physical constraint layer and perform calculations on them respectively, outputting the corrected canopy transpiration, canopy intercepted evaporation, and soil evaporation.

[0097] It should be noted that the surface energy balance equation and Penman-Monteith evapotranspiration equation built into the physical constraint layer serve as the logic for constraining the network feature learning and component inference throughout the physical mapping rules and mechanistic boundary conditions. The Penman-Monteith evapotranspiration equation is used to limit the positive response mechanism between meteorological, soil, and vegetation factors and the three types of evapotranspiration components, ensuring that the evapotranspiration output by the network always conforms to the physical laws of classical evapotranspiration estimation. The surface energy balance equation is used to constrain the matching relationship between latent heat, sensible heat, and surface available energy, ensuring that the evapotranspiration components obtained by decomposition meet the regional energy closure conditions.

[0098] The constructed composite loss function is expressed as:

[0099] ;

[0100] in, Total loss; The mean square error data fitting term between the total evapotranspiration output by the physical information neural network and the observed data of eddy total evapotranspiration is given. The total evapotranspiration output by the physical information neural network is obtained by summing the corrected canopy evaporation, canopy intercepted evaporation, and soil evaporation.

[0101] For the evaporation truth constraint term, ,in, This is the corrected canopy transpiration. Forest stand transpiration; For energy balance constraints, ,in, The net radiation is obtained by actual measurement using the four-component net radiometer equipped with the eddy coherence observation tower, or by spatial downscaling and local correction using the MODIS / ERA5 remote sensing reanalysis product. Soil heat flux is obtained from actual measurements using a soil heat flux plate or estimated based on a soil-vegetation-atmosphere transport model combined with soil texture and moisture content data. For sensible heat flux, Latent heat flux; and All data were obtained from actual measurements using the eddy covariance system.

[0102] To retain the truth value constraint term, ,in, This is the corrected canopy interception evaporation. This is the inverse daily-scale canopy interception.

[0103] , , These are the weight coefficients corresponding to the transpiration true value constraint, the energy balance constraint, and the retention true value constraint, respectively.

[0104] During model training, total diurnal evapotranspiration was used. The inputs include meteorological driving variables and soil moisture state variables, among which the meteorological driving variables include net radiation. air temperature relative humidity Wind speed Soil moisture state variables include soil temperature. and soil volumetric moisture content The Adam optimizer was used, with an initial learning rate of 1×10⁻⁶. -3 The batch size was set to 32, the number of training iterations was set to 200, and an early stopping mechanism was introduced to avoid overfitting.

[0105] The model's three outputs satisfy The conservation constraints, among which, For the corrected transpiration rate, The amount of evaporation to be retained for correction; Soil evaporation is corrected.

[0106] S5. Multi-component evaporation correction output and uncertainty quantification:

[0107] In this step, the trained physical information neural network outputs corrected evapotranspiration, canopy intercepted evaporation, and soil evaporation, and performs uncertainty quantification assessment. The aim is to invoke the converged physical neural network to perform evapotranspiration component inference, outputting three independent evapotranspiration components (canopy evapotranspiration, canopy intercepted evaporation, and soil evaporation) after correction for multiple physical constraints, and to quantify the uncertainty of the estimation results using the Monte Carlo Dropout mechanism.

[0108] In one exemplary implementation, this step is performed as follows:

[0109] The input data to be predicted is fed into the trained physical information neural network for inference. During the inference process, the Dropout mechanism is continuously enabled, and forward propagation inference is repeatedly executed to obtain multiple sets of evapotranspiration component prediction samples. Based on the multiple sets of prediction samples, the predicted mean and prediction variance of canopy evaporation, canopy intercepted evaporation, and soil evaporation are calculated respectively. The predicted mean is used as the corrected canopy evaporation, canopy intercepted evaporation, and soil evaporation outputs, and the prediction variance is used to characterize the uncertainty corresponding to the estimation results of each evapotranspiration component.

[0110] This invention also provides a forest evapotranspiration decomposition and correction system based on sap flow-stem flow-eddy current co-constraints, for implementing the above method, see [link to relevant documentation]. Figure 2 The system includes:

[0111] The data acquisition module is used to acquire and preprocess sap flow data, trunk stem flow data, canopy penetration rainfall data, external rainfall data, and total eddy evapotranspiration observation data of the target forest.

[0112] The transpiration truth value calculation module is used to take the liquid flow data as the individual-scale truth value of canopy transpiration, and combine it with the diameter at breast height or leaf area parameters to perform upscaling calculation to obtain the stand transpiration.

[0113] The precipitation redistribution inversion module is used to construct the precipitation redistribution balance equation and invert the canopy interception by combining canopy penetration rainfall data, trunk stem flow data, and extraforest rainfall data.

[0114] The physical information neural network training module is used to construct a physical information neural network that integrates surface energy balance and Penman-Monteith evapotranspiration mechanism. The forest stand evapotranspiration and canopy interception are used as soft constraints to embed the network composite loss function. The total eddy evapotranspiration observation data, meteorological driving variables and soil moisture state variables are used as inputs to complete the training of the physical information neural network.

[0115] The evapotranspiration component prediction module is used to feed the input data to be predicted into the trained physical information neural network to carry out inference and output the corrected three types of evapotranspiration components: canopy transpiration, canopy intercepted evaporation, and soil evaporation.

[0116] It should be noted that the functions of each functional module in the above system correspond to the steps in the forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy current co-constraint. Therefore, the specific implementation methods of the steps are also applicable to the corresponding functional modules in the system. The specific implementation of each functional module in the system will not be elaborated here.

[0117] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.

Claims

1. A method for forest evapotranspiration decomposition and correction based on sap flow-stem flow-eddy current co-constraints, characterized in that, Includes the following steps: S1. Collect sap flow data, trunk stem flow data, canopy penetration rainfall data, external rainfall data, and total eddy evapotranspiration observation data of the target forest and perform preprocessing; S2. Using the aforementioned sap flow data as the true value of canopy transpiration at the individual scale, and combining it with diameter at breast height (DBH) or leaf area parameters, perform upscaling calculations to obtain the stand transpiration rate: Obtain the diameter at breast height (DBH) of each individual tree within the sample plot. and fluid flow rate Calculate the cross-sectional area of ​​a single piece of wood. ; Using the cross-sectional area weighted method, according to the formula Calculate forest stand evapotranspiration ; Alternatively, the leaf area index upscaling method can be used, according to the formula. Calculate forest stand evapotranspiration ,in, The leaf area index of a single tree; S3. Construct a precipitation redistribution balance equation and, by combining canopy penetration rainfall data, trunk stem flow data, and extraforest rainfall data, deduce canopy interception: Constructing the precipitation redistribution equilibrium equation ,in, This represents the total rainfall outside the forest on a daily scale. Canopy penetration rainfall on a daily scale; This refers to the daily trunk stem flow rate; Canopy interception on a daily scale; Based on daily total rainfall outside the forest Daily scale canopy penetration rainfall Daily-scale trunk stem flow By combining the precipitation redistribution balance equation, the daily-scale canopy interception was calculated. ; Total rainfall outside the forest on that day When observations are missing, canopy penetration rainfall on a daily scale is used. Daily-scale trunk stem flow The sum of these values ​​can be used as a proxy for total precipitation. And introduce a correction factor Corrected daily-scale canopy interception inversion values: The correction coefficient The values ​​were determined by measuring the canopy interception in independent sample plots. S4. Construct a physical information neural network that integrates surface energy balance and Penman-Monteith evapotranspiration mechanism. Use the forest stand evapotranspiration and canopy interception as soft constraints to embed the network composite loss function. Use total eddy evapotranspiration observation data, meteorological driving variables and soil moisture state variables as inputs to complete the training of the physical information neural network. S5. The trained physical information neural network outputs the corrected three types of evapotranspiration components: canopy transpiration, canopy intercepted evaporation, and soil evaporation.

2. The forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy co-constraint as described in claim 1, characterized in that, In step S1, the methods for collecting sap flow data, stem flow data, canopy penetration rainfall data, external rainfall data, and total eddy evapotranspiration observation data of the target forest include: Liquid flow data of individual logs were collected using a thermal diffusion liquid flow probe; Tree stem flow data were collected using a tipping bucket flow meter; Canopy penetration rainfall data were collected using a forest understory rain gauge array. Total rainfall data outside the forest were collected using a standard rain gauge. The total vortex evapotranspiration observation data were obtained by collecting water vapor flux using an vortex correlation system and converting it.

3. The forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy co-constraint as described in claim 2, characterized in that, In step S1, the preprocessing includes: Missing value imputation, outlier removal, component error correction, unit unification, and aggregation with natural days as the time window are used to obtain a daily-scale dataset. The sub-error corrections include: nighttime zero-flow baseline correction and temperature drift correction for sap flow data; splashing error correction and installation tilt correction for trunk stem flow and canopy penetration rainfall data; and coordinate rotation, WPL density effect correction, and frequency response correction for total eddy evapotranspiration observation data.

4. The forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy co-constraint as described in claim 1, characterized in that, In step S4, the physical information neural network includes an input layer, a shared hidden layer, a physical constraint layer, and an output layer connected in sequence. The input layer is used to receive daily-scale total eddy evapotranspiration, meteorological driving variables and soil moisture state variables as input features, and to send all input features to the shared hidden layer. The shared hidden layer is composed of multiple stacked fully connected layers. It is used to receive the feature vector output by the input layer and complete the unified nonlinear feature mapping to extract deep features that fuse environmental and throughput information, and then pass the deep features to the physical constraint layer. The physical constraint layer incorporates the surface energy balance equation and the Penman-Monteith evapotranspiration equation. During network forward propagation, this layer receives the deep feature vector output from the shared hidden layer, adapts and transforms the deep features, and then directly transmits them downstream to the multi-branch output layer for component prediction. During the model training phase, this layer introduces two types of measured true values: forest stand evapotranspiration and canopy interception. Combined with the surface energy balance constraint, true value constraint terms for evapotranspiration, true value constraint terms for interception, and energy balance constraint terms are constructed respectively, which together form a composite loss function. The output layer is provided with three independent output branches that do not interfere with each other. These branches receive the features transmitted from the physical constraint layer and perform calculations on them respectively, outputting the corrected canopy transpiration, canopy intercepted evaporation, and soil evaporation.

5. The forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy co-constraint as described in claim 4, characterized in that, The composite loss function is expressed as follows: ； in, Total loss; The mean square error data fitting term between the total evapotranspiration and eddy covariance total evapotranspiration observation data output by the physical information neural network is obtained by summing the corrected canopy evapotranspiration, canopy intercepted evapotranspiration, and soil evapotranspiration. For the evaporation truth constraint term, ,in, This is the corrected canopy transpiration. Forest stand transpiration; For energy balance constraints, ,in, This is the net radiation level. For soil heat flux, For sensible heat flux, Latent heat flux; To retain the truth value constraint term, ,in, This is the corrected canopy interception evaporation. The inverse daily-scale canopy interception; , , These are the weight coefficients corresponding to the transpiration true value constraint, the energy balance constraint, and the retention true value constraint, respectively.

6. A forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy current co-constraint as described in any one of claims 1 to 5, characterized in that, In step S5, the trained physical information neural network outputs the corrected three types of evapotranspiration components: canopy transpiration, canopy intercepted evaporation, and soil evaporation, specifically including: The input data to be predicted is fed into the trained physical information neural network for inference. During the inference process, the Dropout mechanism is continuously enabled, and forward propagation inference is repeatedly executed to obtain multiple sets of evapotranspiration component prediction samples. Based on the multiple sets of prediction samples, the predicted mean and prediction variance of canopy evaporation, canopy intercepted evaporation, and soil evaporation are calculated respectively. The predicted mean is used as the corrected canopy evaporation, canopy intercepted evaporation, and soil evaporation outputs, and the prediction variance is used to characterize the uncertainty corresponding to the estimation results of each evapotranspiration component.

7. A forest evapotranspiration decomposition and correction system based on sap flow-stem flow-eddy current co-constraints, used to implement the forest evapotranspiration decomposition and correction method based on sap flow-stem flow-eddy current co-constraints as described in any one of claims 1 to 6, characterized in that, The system includes: The data acquisition module is used to acquire and preprocess sap flow data, trunk stem flow data, canopy penetration rainfall data, external rainfall data, and total eddy evapotranspiration observation data of the target forest. The transpiration truth value calculation module is used to take the liquid flow data as the individual-scale truth value of canopy transpiration, and combine it with the diameter at breast height or leaf area parameters to perform upscaling calculation to obtain the stand transpiration. The precipitation redistribution inversion module is used to construct the precipitation redistribution balance equation and invert the canopy interception by combining canopy penetration rainfall data, trunk stem flow data, and extraforest rainfall data. The physical information neural network training module is used to construct a physical information neural network that integrates surface energy balance and Penman-Monteith evapotranspiration mechanism. The forest stand evapotranspiration and canopy interception are used as soft constraints to embed the network composite loss function. The total eddy evapotranspiration observation data, meteorological driving variables and soil moisture state variables are used as inputs to complete the training of the physical information neural network. The evapotranspiration component prediction module is used to feed the input data to be predicted into the trained physical information neural network to carry out inference and output the corrected three types of evapotranspiration components: canopy transpiration, canopy intercepted evaporation, and soil evaporation.

Images