Data-driven actual evapotranspiration estimation method incorporating vegetation biophysical characteristics

Abstract

The application discloses a data-driven actual evapotranspiration estimation method fusing vegetation biophysical characteristics, first, collecting the measured meteorological and remote sensing data of a flux tower site, performing quality control and energy balance correction to obtain an actual evapotranspiration reference value; then, based on the processed data, a PMLV2 model is constructed station by station to simulate vegetation biophysical characteristic variables such as stomatal conductance; these characteristics are summarized with meteorological and remote sensing data to form an input factor set, and the training set and test set of each station are summarized according to vegetation types; finally, taking the factor set as input, an LSTM actual evapotranspiration estimation model is constructed by using the training set, and the accuracy thereof is evaluated through the test set; the application fuses the vegetation biophysical characteristic variables simulated by the PMLV2 model, enhances the adaptability and generalization ability of the data-driven model under the change of environmental conditions, and effectively alleviates the overfitting and performance decline problems of the traditional data-driven method under the condition of sample scarcity or insufficient representation.

Description

Technical Field

[0001] This invention belongs to the field of hydrology and remote sensing application technology, and in particular relates to a data-driven method for estimating actual evapotranspiration that integrates vegetation biophysical characteristics. Background Technology

[0002] Actual evapotranspiration, as the primary pathway for water loss in terrestrial ecosystems, is a crucial link connecting the biosphere and the atmosphere in the exchange of water, carbon flux, and energy. Against the backdrop of global climate change and intensified human activities, accurately understanding the spatiotemporal distribution of actual evapotranspiration is of paramount importance for sustainable regional water resource management, optimized agricultural irrigation, and ecological security maintenance. Currently, the industry primarily obtains actual evapotranspiration data through two methods: ground-based observations and model simulations. While ground-based observation methods, such as eddy covariance, can provide high-precision local measured data, their sparse station coverage, high observation costs, and limited spatial representativeness make them unsuitable for large-scale, continuous dynamic monitoring of actual evapotranspiration. Therefore, model-based estimation has become the mainstream direction in current evapotranspiration research and applications.

[0003] Currently, methods for estimating actual evapotranspiration are mainly divided into two categories: physical process-based models and data-driven models. Physical process models are based on energy balance and water transport theories, simulating the dynamic processes of vegetation transpiration and soil evaporation by constructing complex process equations and parameterization schemes. Data-driven models mainly utilize remote sensing data and meteorological observations or reanalysis data, employing algorithms such as machine learning to establish a nonlinear mapping relationship between input features and evapotranspiration.

[0004] However, existing technologies still have significant shortcomings in practical applications. While physical process models have clear mechanisms, their structures are complex and highly sensitive to input parameters. In areas with strong spatial heterogeneity or where observational data is scarce, parameters are difficult to obtain and calibrate accurately, leading to significant uncertainties in simulation results. Although data-driven models can effectively fit nonlinear relationships, their performance is highly dependent on the quantity and quality of training samples. Furthermore, vegetation-related input features are usually limited to variables directly observed by remote sensing (such as leaf area index and normalized difference vegetation index), lacking a deep expression of vegetation physiological and ecological processes. This results in insufficient generalization ability of the models under changing environmental conditions and weak mechanistic interpretability. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a data-driven method for estimating actual evapotranspiration that integrates vegetation biophysical characteristics, which can both capture nonlinear relationships using large-scale data and introduce vegetation biophysical process mechanisms to improve generalization ability and interpretability.

[0006] Technical solution: The data-driven method for estimating actual evapotranspiration described in this invention includes the following steps:

[0007] S1. Collect measured meteorological data from flux tower stations, perform quality control and energy balance correction preprocessing on the measured meteorological data to obtain actual evapotranspiration reference values, and extract corresponding remote sensing image data based on the geographical location of the flux tower stations, and interpolate the remote sensing image data to a daily scale.

[0008] S2. Based on the meteorological and remote sensing data processed in step S1, construct the PMLV2 remote sensing evapotranspiration model station by station to simulate the vegetation biophysical characteristic variables during the evapotranspiration process.

[0009] S3. Summarize the vegetation biophysical characteristic variables obtained in step S2 with the preprocessed meteorological data and remote sensing data obtained in step S1 to form a model input factor set.

[0010] S4. Statistically analyze the vegetation types of the flux tower stations. For each vegetation type, divide the data of all stations into training and test sets according to the time series within the station at a preset ratio, and summarize them separately.

[0011] S5. Using the model input factor set formed in step S3 as input, train the Long Short-Term Memory (LSTM) network model using the training set to construct an actual evapotranspiration estimation model; apply the trained actual evapotranspiration estimation model to the test set, output the actual evapotranspiration estimation value, and compare the actual evapotranspiration estimation value with the actual evapotranspiration reference value after preprocessing in step S1, and use statistical indicators to quantitatively evaluate the estimation accuracy of the model at all sites.

[0012] This invention integrates measured meteorological data from flux tower sites with multi-source remote sensing data, and performs rigorous quality control and scale matching to provide reliable and spatiotemporally consistent driving data and validation benchmarks for the model. Furthermore, by coupling the PMLV2 remote sensing evapotranspiration model, it accurately simulates key vegetation biophysical process variables such as stomatal conductance and canopy available energy, thus introducing physical mechanisms into the subsequent data-driven model. Based on this, the aforementioned mechanistic variables are fused with the original observation and remote sensing data to construct a comprehensive input factor set, greatly enriching the information on vegetation evapotranspiration physiological processes that the model can capture. Subsequently, training and testing data from each site are summarized according to vegetation type, laying the foundation for the model to learn evapotranspiration behavior patterns for specific vegetation types. Finally, a Long Short-Term Memory (LSTM) network is used to process its temporal dependence characteristics, enabling it to deeply mine the complex nonlinear relationships between meteorological, remote sensing, and biophysical variables and evapotranspiration from large-scale data. The introduction of vegetation biophysical variables with clear physical meaning significantly improves the model's generalization ability and mechanistic interpretability under different regions and conditions, achieving a high degree of accuracy and robustness in actual evapotranspiration estimation.

[0013] Preferably, the measured meteorological data collected in step S1 includes daily average temperature, precipitation, wind speed, saturated vapor pressure difference, air pressure, incident shortwave radiation, incident longwave radiation, and carbon dioxide concentration; the extracted remote sensing image data includes leaf area index, surface reflectance, and surface emissivity.

[0014] By systematically collecting multi-source diurnal meteorological data covering thermal, dynamic, and radiative factors, and combining it with remote sensing parameters that reflect key information about vegetation structure, surface properties, and atmospheric environment, this step provides a comprehensive and physically meaningful driving foundation for subsequent model construction. This ensures that the model can accurately characterize the complex environmental processes affecting evapotranspiration from multiple dimensions such as energy balance, water supply and demand, and biogeochemical cycles.

[0015] Preferably, the preprocessing of the measured meteorological data for quality control and energy balance correction in step S1 includes:

[0016] Dates with latent heat flux quality control markers below a preset first threshold, precipitation days, and dates with negative latent heat flux or incident shortwave radiation are excluded.

[0017] For records where the energy balance ratio is lower than a preset second threshold or higher than a preset third threshold, the following formula is used for correction:

[0018]

[0019] In the formula, For energy balance ratio, For latent heat flux, For sensible heat flux, Net radiation, This is the corrected latent heat flux.

[0020] By implementing strict quality control on measured meteorological data, unreliable data points caused by instrument errors, abnormal weather, or non-ideal observation conditions were effectively eliminated. Furthermore, to address the common problem of unclosed energy budgets on the Earth's surface, a physical correction method based on the energy balance ratio was adopted to calibrate latent heat flux. This significantly improved the physical consistency and reliability of flux tower observation data, especially the actual evapotranspiration reference values ​​used as the benchmark for model training and validation, laying a solid data quality foundation for subsequent model construction.

[0021] Preferably, the PMLV2 remote sensing evapotranspiration model described in step S2 is a surface evapotranspiration estimation framework constructed based on the Penman-Monteith formula. The PMLV2 remote sensing evapotranspiration model characterizes the total daily evapotranspiration as the sum of canopy evaporation, soil evaporation, and intercepted evaporation.

[0022] The PMLV2 model employs the SCE-UA optimization algorithm for parameter calibration at each site. The calibration process aims to maximize the Nash-Sutcliffe efficiency coefficient between the estimated and measured evapotranspiration values ​​for that site.

[0023] By adopting the PMLV2 model based on the Penman-Monteith formula and distinguishing different evaporation components, a robust physical framework for the evapotranspiration process mechanism was introduced for subsequent data-driven models. At the same time, the SCE-UA optimization algorithm was used to calibrate the parameters station by station, ensuring that the model could adaptively capture the unique underlying surface and climate characteristics of different stations. This generated a set of high-precision vegetation biophysical characteristic variables that could reflect universal physical laws and have site adaptability, greatly enhancing the physical consistency and representativeness of these intermediate variables for subsequent data-driven modeling.

[0024] Preferably, the vegetation biophysical characteristic variables mentioned in step S2 include canopy available energy. Pore ​​conductance and canopy conductance .

[0025] By introducing key vegetation biophysical characteristic variables such as canopy available energy, stomatal conductance, and canopy conductance simulated by the PMLV2 model, the energy driving conditions and water vapor exchange capacity between the canopy and the atmosphere on which the vegetation transpiration process depends are quantified mechanistically. This transforms the core physical processes controlling evapotranspiration into intrinsic driving factors that the LSTM model can directly learn, significantly enhancing the model's ability to mechanistically describe and generalize the response of vegetation physiological activities to environmental changes.

[0026] Preferably, the canopy can be powered by energy. The calculation formula is:

[0027]

[0028] In the formula, It is the net ground radiation, calculated from the variables of incident shortwave radiation, incident longwave radiation, surface reflectivity, and surface emissivity; It is the effective radiation absorption ratio of the soil, calculated from the leaf area index variable.

[0029] By calculating the net ground radiation using surface reflectance and emissivity data retrieved through coupled remote sensing, and combining this with the soil effective radiation absorption ratio driven by the leaf area index, the available energy of the canopy is accurately quantified. This formula realizes the physical transformation from multi-source remote sensing observation to vegetation canopy energy-driven processes, thus providing an energy input for evapotranspiration models that can dynamically reflect changes in surface cover and radiation processes, and enhancing the mechanistic and accurate simulation of energy balance processes.

[0030] Preferably, the porosity The calculation formula is:

[0031]

[0032] In the formula, It is the leaf photosynthetic rate, calculated from the carbon dioxide concentration variable; It is the concentration of carbon dioxide in the air on the leaf surface; It is the water vapor pressure difference; and These are model parameters, namely, porosity coefficient and... Standard water vapor pressure difference at that time.

[0033] This formula incorporates key environmental factors such as leaf photosynthesis, air carbon dioxide concentration, and atmospheric humidity into the quantitative process of stomatal behavior by coupling measured carbon dioxide concentration with meteorological data. This mechanistically characterizes the physiological response strategy of vegetation to dynamically adjust stomatal opening in order to balance carbon absorption and water consumption, providing an intrinsic control variable for evapotranspiration models that can profoundly reflect the physiological and ecological constraints of vegetation.

[0034] Preferably, the canopy conductance The calculation method is as follows:

[0035] Using Beer's law of radiation attenuation Pore ​​conductance Upscale to canopy conductance :

[0036]

[0037]

[0038] In the formula, It is the leaf area index per unit area of ​​the entire canopy; It represents the highest photosynthetic rate when light and carbon dioxide concentrations reach saturation, and is the maximum carboxylation capacity. Half of; It is the initial slope of the assimilation efficiency versus light response curve; It is the initial slope of the assimilation efficiency versus CO2 response curve; It is the flux density of effective photosynthetic radiation (PAR) at the top of the canopy; It is the extinction coefficient of shortwave radiation.

[0039] By applying Beer's law of radiation decay and comprehensively considering the response mechanisms of vegetation to light intensity and carbon dioxide concentration, the stomatal conductance at the leaf scale is physically and reasonably scaled up to the canopy level. This allows for precise quantification of the overall capacity of the vegetation canopy to exchange water vapor with the atmosphere, enabling the model to more realistically reflect the key impact of the spatial heterogeneity of radiative transmission and physiological processes within the canopy on total evapotranspiration.

[0040] Preferably, the actual evapotranspiration reference value mentioned in step S5 needs to be converted using the latent heat flux observed at the flux tower site. The conversion formula is as follows:

[0041]

[0042] In the formula, Latent heat flux, unit: , This refers to actual evaporation, in units of , The average daily temperature is expressed in °C.

[0043] By converting the latent heat flux observed by the flux tower into a reference value of actual evapotranspiration expressed in terms of water depth using the daily average temperature, this conversion formula achieves physical unification between energy units and hydrological units. This ensures that the target variables used in subsequent model training and validation are directly comparable in terms of dimensions, providing a benchmark true value with clear physical meaning and consistent with the needs of practical applications for data-driven models.

[0044] Preferably, the statistical indicators mentioned in step S5 are the Nash-Sutcliffe efficiency coefficient (NSE) and the root mean square error (RMSE), as shown in the following formula:

[0045]

[0046] Where n is the length of the data sequence; and This represents the actual evapotranspiration measured and estimated at time i. This represents the average actual evapotranspiration measured for the entire sequence.

[0047] By comprehensively employing the Nash-Sutcliffe efficiency coefficient and root mean square error, we can conduct a comprehensive and rigorous quantitative evaluation of the accuracy and reliability of the actual evapotranspiration estimation model from two complementary dimensions: the consistency of the dynamic trends between the model's simulated values ​​and the measured values, and the magnitude of the absolute deviation. This objectively reveals the model's overall performance in capturing the temporal variation patterns of evapotranspiration and controlling the absolute estimation error.

[0048] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. By integrating vegetation biophysical characteristic variables simulated by the PMLV2 model, the adaptability and generalization ability of the data-driven model under changing environmental conditions are enhanced, effectively alleviating the overfitting and performance degradation problems of traditional data-driven methods when samples are scarce or insufficiently representative; 2. By introducing vegetation biophysical characteristics based on physical mechanisms, the model can not only capture the nonlinear relationship between environmental factors and actual evapotranspiration, but also has the ability to deeply express vegetation physiological and ecological processes, making up for the shortcomings of pure data-driven models in terms of mechanism interpretation; 3. By combining the mechanism clarity of the physical process model (PMLV2) with the powerful fitting ability of the data-driven model (LSTM), while maintaining high accuracy, the adaptability and reliability of the model under different vegetation types and climatic conditions are improved, providing more robust evapotranspiration estimation data for regional water resource management and agricultural applications. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0050] Figure 2 The results of vegetation biophysical characteristic variables simulated by PMLV2 in Embodiment 1 of the present invention;

[0051] Figure 3 The above is the simulation result of hyperparameter optimization using LSTM with grid search in Embodiment 1 of the present invention;

[0052] Figure 4 This refers to the actual evapotranspiration estimation results of the model training set and test set in Embodiment 1 of the present invention;

[0053] Figure 5 This is a comparison of the actual evapotranspiration estimation accuracy during the test period between the data-driven method considering vegetation biophysical characteristics in Embodiment 2 of the present invention and the data-driven method based on physical processes and driven only by environmental factors. Detailed Implementation

[0054] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0055] Example 1:

[0056] This embodiment uses 103 flux tower sites widely distributed globally as an example. These sites cover a variety of climates and ecosystems, including 10 major vegetation types: 12 sites belong to croplands (CRO), 11 sites to deciduous broadleaf forests (DBF), 7 sites to evergreen broadleaf forests (EBF), 21 sites to evergreen needleleaf forests (ENF), 23 sites to grasslands (GRA), 6 sites to mixed forests (MF), 4 sites to open shrublands (OSH), 7 sites to savannas (SAV), 6 sites to wetlands (WET), and 6 sites to woody savannas (WSA). These sites are suitable as typical research objects for this invention considering vegetation biophysical processes.

[0057] This embodiment utilizes daily meteorological and water flux observation data from eddy covariance observation flux towers provided by the FLUXNET2015 database, along with MODIS leaf area index, surface albedo, and surface emissivity datasets provided by NASA. A PMLV2 model is constructed station-by-station to simulate vegetation biophysical process variables. Based on this, meteorological and remote sensing data representing environmental conditions, along with vegetation biophysical characteristic variables, are used as the input factor set for the data-driven model. Training and testing sets are divided for different vegetation types to train and test the LSTM model, constructing a data-driven actual evapotranspiration estimation model that considers vegetation physiological characteristics. The accuracy is then compared with the evapotranspiration reference values ​​measured by the flux towers.

[0058] like Figure 1 As shown, the specific method in this embodiment is as follows:

[0059] Step S1: Download daily meteorological and water flux observation data for the selected flux tower from the FLUXNET2015 database, including daily average temperature (Ta), precipitation (P), wind speed (WS), saturated vapor pressure difference (VPD), air pressure (Pa), incident shortwave radiation (Rs_in), incident longwave radiation (Rl_in), carbon dioxide concentration (Ca), sensible heat flux (H), and latent heat flux (LE). All station data are strictly screened using the following criteria: excluding dates with an LE quality control mark ≤ 0.75, excluding dates with precipitation P > 0, and excluding dates with negative LE or Rs_in values ​​to ensure effective energy input. In addition, to reduce errors caused by energy non-closure, records with an energy balance ratio (EBR) less than 80% or greater than 120% are corrected using the Bowen ratio method. Download MODIS leaf area index data MCD15A3H.006, surface albedo data MCD43A3.006, and surface emissivity data MOD11A2.006, and extract them according to the geographical location of the flux tower site using the nearest neighbor interpolation method. Products with low temporal resolution (such as 4-day, 8-day, or monthly data) are unified to the daily scale through linear interpolation.

[0060] Step S2: Based on the data processed and collected in Step S1, construct PMLV2 remote sensing evapotranspiration models for each station; optimize the model parameters using the SCE-UA parameter optimization algorithm, with the objective function being the Nash-Sutcliffe efficiency coefficient (NSE) of the actual evapotranspiration at the station; simulate the time series values ​​of key variables in the biophysical process for different vegetation types represented by different stations, including canopy available energy Ac, stomatal conductance gs, and canopy conductance Gc.

[0061] Figure 2 This section describes the distribution of biophysical characteristic variables under different vegetation types simulated by the PMLV2 model. Stomatal conductance represents the rate of water vapor exchange at the leaf scale and typically reflects the vegetation's water regulation strategy. In farmland (CRO) and wetlands (WET), its value is significantly higher than in other vegetation types. Due to the larger leaf area and more complex canopy structure of forest vegetation types, the values ​​of canopy conductance and canopy effective energy are significantly higher than in non-forest vegetation types. Because of differences in stomatal regulation mechanisms, canopy structure, and other vegetation physiological characteristics, there is significant heterogeneity in water regulation and energy distribution during evapotranspiration among different vegetation types, highlighting the necessity of using key physiological variables from physical process simulations as auxiliary inputs for data-driven models.

[0062] Step S3: Summarize the vegetation biophysical characteristic variables obtained in Step S2 with the original variables obtained in Step S1 to form a model input factor set, including daily average temperature, precipitation, wind speed, saturated vapor pressure difference, air pressure, incident shortwave radiation, incident longwave radiation, leaf area index, surface reflectance, surface emissivity, carbon dioxide concentration, canopy available energy, stomatal conductance, and canopy conductance.

[0063] Step S4: Divide the data of all stations under 10 vegetation types in the 103 flux tower stations into a 7:3 ratio according to the time series within the station, and summarize them into training set and test set respectively; the total number of valid samples of all stations is 49910, with an average of 3494 training samples and 1497 test samples under each vegetation type.

[0064] Step S5: Based on the input factor set from Step S3, train a Long Short-Term Memory (LSTM) network model using the training set to construct a practical evapotranspiration estimation model. During training, the average Nash-Sutcliffe efficiency coefficient (NSE) of flux tower sites is used as the objective function for optimization. Key hyperparameters are determined through a grid search method to obtain the optimal prediction performance.

[0065] Figure 3 This test was conducted to evaluate the possible values ​​of key hyperparameters of the LSTM (number of hidden units: 40, 80, 120, 160, 200; dropout rate: 0.01, 0.03, 0.05, 0.1, 0.2, 0.4; time series length: 20, 50, 100, 150, 200). The final LSTM model structure consists of two stacked LSTM layers, each containing 120 hidden units, followed by a fully connected output layer. The model uses a 100-day time series length, with a dropout rate of 3% to reduce the risk of overfitting.

[0066] The trained model is applied to the test set to output an estimated value of actual evapotranspiration, which is then compared with the reference value of actual evapotranspiration after preprocessing in step S1. Before comparison, the latent heat flux observed at the flux tower site needs to be converted using the following formula:

[0067]

[0068] In the formula, Latent heat flux, unit: , This refers to actual evaporation, in units of , The average daily temperature is expressed in °C.

[0069] Statistical indicators were used to quantitatively evaluate the accuracy of the model's actual evapotranspiration estimation across all sites. The statistical indicators used were the Nash-Sutcliffe efficiency coefficient (NSE) and root mean square error (RMSE). NSE measures the goodness of fit between the model's simulated values ​​and observed values; its value typically ranges from negative infinity to 1. A value closer to 1 indicates a better simulation and higher consistency between observations and estimates. RMSE reflects the average deviation between estimated and observed values; a smaller RMSE value indicates lower error and higher model accuracy. The formulas for calculating these indicators are as follows:

[0070]

[0071] Where n is the length of the data sequence; and The actual evapotranspiration measured and estimated at time i ( ); The average of the actual evapotranspiration measured for the entire sequence ( ).

[0072] Figure 4 This paper compares the estimated evapotranspiration values ​​with the measured values ​​using this method. The results show that this method exhibits good performance in estimating actual evapotranspiration; the linear fit between the estimated and measured reference values ​​is close to a 1:1 line. The NSE values ​​for the training and testing periods are 0.918 and 0.913, respectively, and the RMSE values ​​are 0.387. and 0.396 .

[0073] Example 2:

[0074] The mainstream PMLV2 remote sensing evapotranspiration model was selected as the representative physical process model for comparison, and the LSTM model driven solely by environmental factors was selected as the representative pure data-driven model for comparison. The proposed method was compared from both physical process model and data-driven model perspectives. To ensure a fair comparison between the models, the PMLV2 and LSTM models used the exact same 11 multi-source time series variables as used in the vegetation biophysical characteristic simulation in this method, and constructed using the same training and test sets. The difference between this method and the LSTM model that relies solely on environmental factor inputs is that it uses vegetation biophysical characteristic variables as additional inputs.

[0075] Figure 5 The accuracy of different models in estimating actual evapotranspiration during the test period is shown, with accuracy metrics categorized and statistically analyzed according to the vegetation type of the flux towers. Our method (LSTM considering vegetation biophysical characteristics) showed the best performance in estimating actual evapotranspiration, with median NSE and RMSE of 0.932 and 0.206 for all flux tower sites, respectively. Secondly, for the LSTM model driven solely by environmental factors, the median NSE and RMSE for flux tower sites were 0.782 and 0.398, respectively. The PMLV2 model based on physical processes performed the worst, with median NSE and RMSE of 0.697 and 0.540 for flux tower sites, respectively. The advantages of this method are more pronounced in non-forest vegetation types. For example, in sparse vegetation (WSA), the median NSE for stations obtained by the PML model is negative (-0.076), while the median NSE estimated by this method is 0.911. Performance comparisons with physical process models and purely data-driven models demonstrate that by additionally considering vegetation biophysical characteristics in the input of data-driven models, the accuracy of the model's estimation of actual evapotranspiration and its applicability in complex vegetation types can be effectively improved.

[0076] Against the backdrop of increasingly significant impacts from climate change and human activities, accurate estimation of actual evapotranspiration is crucial for understanding regional water cycles, assessing water resource security, and optimizing hydrological and ecological models. This method, leveraging the nonlinear modeling capabilities and large-scale data processing power of data-driven models, fully considers the differences in biophysical characteristics such as water regulation and energy distribution within vegetation during evapotranspiration. It introduces key variables such as canopy available energy and stomatal conductance as additional inputs to the data-driven model, effectively improving the accuracy and applicability of actual evapotranspiration estimation for different vegetation types.

Claims

1. A data-driven method for estimating actual evapotranspiration that integrates vegetation biophysical characteristics, characterized in that, Includes the following steps: S1. Collect measured meteorological data from flux tower stations, perform quality control and energy balance correction preprocessing on the measured meteorological data to obtain actual evapotranspiration reference values, and extract corresponding remote sensing image data based on the geographical location of the flux tower stations, and interpolate the remote sensing image data to a daily scale. The measured meteorological data include daily average temperature, precipitation, wind speed, saturated vapor pressure difference, air pressure, incident shortwave radiation, incident longwave radiation, and carbon dioxide concentration; the extracted remote sensing image data includes leaf area index, surface reflectance, and surface emissivity. S2. Based on the meteorological and remote sensing data processed in step S1, a PMLV2 remote sensing evapotranspiration model is constructed station by station to simulate the vegetation biophysical characteristic variables during the evapotranspiration process. The PMLV2 remote sensing evapotranspiration model is a surface evapotranspiration estimation framework based on the Penman-Monteith formula. The PMLV2 remote sensing evapotranspiration model characterizes the daily total evapotranspiration as the sum of canopy evaporation, soil evaporation, and intercepted evaporation. The PMLV2 remote sensing evapotranspiration model is calibrated for each station using the SCE-UA optimization algorithm. The calibration process aims to maximize the Nash-Sutcliffe efficiency coefficient between the estimated and measured evapotranspiration values ​​for that station. The vegetation biophysical characteristic variables include canopy available energy. Pore ​​conductance and canopy conductance ; S3. Summarize the vegetation biophysical characteristic variables obtained in step S2 with the preprocessed meteorological data and remote sensing data obtained in step S1 to form a model input factor set. S4. Statistically analyze the vegetation types of the flux tower stations. For each vegetation type, divide the data of all stations into training and test sets according to the time series within the station at a preset ratio, and summarize them separately. S5. Using the model input factor set formed in step S3 as input, train the Long Short-Term Memory (LSTM) network model using the training set to construct an actual evapotranspiration estimation model; apply the trained actual evapotranspiration estimation model to the test set, output the actual evapotranspiration estimation value, and compare the actual evapotranspiration estimation value with the actual evapotranspiration reference value after preprocessing in step S1, and use statistical indicators to quantitatively evaluate the estimation accuracy of the model at all sites.

2. The data-driven method for estimating actual evapotranspiration according to claim 1, characterized in that, Step S1, which involves preprocessing the measured meteorological data for quality control and energy balance correction, includes: Dates with latent heat flux quality control markers below a preset first threshold, precipitation days, and dates with negative latent heat flux or incident shortwave radiation are excluded. For records where the energy balance ratio is lower than a preset second threshold or higher than a preset third threshold, the following formula is used for correction: In the formula, For energy balance ratio, For latent heat flux, For sensible heat flux, Net radiation, This is the corrected latent heat flux.

3. The data-driven method for estimating actual evapotranspiration according to claim 1, characterized in that, The canopy has available energy The calculation formula is: In the formula, It is the net ground radiation, calculated from the variables of incident shortwave radiation, incident longwave radiation, surface reflectivity, and surface emissivity; It is the effective radiation absorption ratio of the soil, calculated from the leaf area index variable.

4. The data-driven method for estimating actual evapotranspiration according to claim 1, characterized in that, Pore ​​conductance The calculation formula is: In the formula, It is the leaf photosynthetic rate, calculated from the carbon dioxide concentration variable; It is the concentration of carbon dioxide in the air on the leaf surface; It is the water vapor pressure difference; and These are model parameters, namely, porosity coefficient and... Standard water vapor pressure difference at that time.

5. The data-driven method for estimating actual evapotranspiration according to claim 4, characterized in that, The canopy conductance The calculation method is as follows: Using Beer's law of radiation attenuation Pore ​​conductance Upscale to canopy conductance : In the formula, It is the leaf area index per unit area of ​​the entire canopy; It represents the highest photosynthetic rate when light and carbon dioxide concentrations reach saturation, and is the maximum carboxylation capacity. Half of; It is the initial slope of the assimilation efficiency versus light response curve; It is the initial slope of the assimilation efficiency versus CO2 response curve; It is the flux density of effective photosynthetic radiation (PAR) at the top of the canopy; It is the extinction coefficient of shortwave radiation.

6. The data-driven method for estimating actual evapotranspiration according to claim 1, characterized in that, The actual evapotranspiration reference value mentioned in step S5 needs to be converted using the latent heat flux observed at the flux tower site. The conversion formula is as follows: In the formula, Latent heat flux, unit: , This refers to actual evaporation, in units of , The average daily temperature is expressed in °C.

7. The data-driven method for estimating actual evapotranspiration according to claim 1, characterized in that, The statistical indicators mentioned in step S5 are the Nash-Sutcliffe efficiency coefficient (NSE) and the root mean square error (RMSE), as shown in the following formulas: Where n is the length of the data sequence; and This represents the actual evapotranspiration measured and estimated at time i. This represents the average actual evapotranspiration measured for the entire sequence.

Images