Ground surface evapotranspiration estimation method combining data assimilation and machine learning

Abstract

The invention discloses a surface evapotranspiration estimation method combining data assimilation and machine learning, and belongs to the field of data assimilation algorithm improvement and the technical field of surface evapotranspiration simulation. A deviation correction model and an observation operator are coupled to a data assimilation method through a machine learning method, so that accurate estimation of surface evapotranspiration is realized, soil moisture transmission, vegetation dynamic and photosynthetic processes of key physical processes of the model are optimized in real time, and finally improved surface evapotranspiration is obtained. According to the method, the existing site observation soil moisture data and remote sensing observation leaf area index and sunlight induced chlorophyll fluorescence data are effectively utilized to optimize Noah-MP model soil moisture transmission, vegetation dynamic and photosynthetic processes, and a high-precision space-time continuous region surface evapotranspiration data set which can be used in multiple fields is produced; errors caused by observation data missing or model parameters of a traditional model are reduced.

Description

Technical Field

[0001] The invention belongs to the field of data assimilation algorithm improvement and surface evapotranspiration simulation technology, and relates to a surface evapotranspiration estimation method combining data assimilation and machine learning. Background Art

[0002] Surface evapotranspiration is a very important process of material and energy conversion and transport in the soil-vegetation-atmosphere system, which is related to the earth's water, energy and carbon cycles. Surface evapotranspiration mainly includes evaporation of intercepted water bodies, soil and vegetation surfaces, as well as transpiration of vegetation. Accurately estimating surface evapotranspiration and soil evaporation and vegetation transpiration component values ​​can not only meet the needs of rational management of global limited water resources and optimal irrigation decision-making projects for farmland, but also provide important response measures for potential changes in the global water cycle under any climate change situation.

[0003] Ground observation, remote sensing estimation, process model simulation and data assimilation are the main methods for obtaining evapotranspiration. The existing carbon and water flux observation network provides high-precision, real-time data support and reference basis for flux state analysis at point scale or small area. However, the spatial representation range of ground observation is small, it is difficult to expand spatially, and the instrument maintenance cost is high, so it cannot meet the research of large-scale evapotranspiration spatiotemporal pattern. In order to solve this problem, remote sensing estimation method came into being. This estimation method can provide large-scale evapotranspiration simulation data, but the remote sensing estimation method is easily affected by the lack of remote sensing data source, long distance or quality problems, and the observations obtained are instantaneous values ​​at the time of satellite transit, and its spatiotemporal resolution is rough, and it cannot provide high-precision results. The process model is a model used to simulate the physical mechanism process in nature, which can obtain spatiotemporal continuous state variable simulation results. In the study of terrestrial water cycle and carbon cycle, process model has become a basic application method. The current process model is relatively complete, but the complex structure and numerous parameters of the model also introduce uncertainty.

[0004] Data assimilation is a common method for obtaining better model parameters and state variables in data-model fusion. Data assimilation integrates multi-source, multi-scale and discontinuously distributed observation data into the process model to automatically update and adjust the model trajectory. It can improve the reliability and accuracy of the model to better reflect the real situation. Therefore, data assimilation is a very valuable method. Most of the current data assimilation only assimilates a single observation variable, and does not fully utilize the site and multi-source remote sensing observation information at the same time. In order to make more full use of multi-source observation data, it is necessary to combine machine learning and data assimilation to develop a robust evapotranspiration estimation method. Summary of the invention

[0005] In view of the problems in the prior art, the present invention provides a method for estimating surface evapotranspiration by combining data assimilation and machine learning.

[0006] A method for estimating evapotranspiration by combining data assimilation and machine learning, which couples a bias correction model and an observation operator into a data assimilation method through a machine learning method, thereby achieving accurate estimation of evapotranspiration, real-time optimization of soil moisture transport, vegetation dynamics, and photosynthetic processes of key physical processes of the model, and finally obtaining improved evapotranspiration, includes the following steps: initializing a model step, predicting a model state, observing an operator step, and updating state variables and parameters.

[0007] The advantages of the present invention are: to effectively utilize the existing soil moisture data observed by the site and the leaf area index and sunlight-induced chlorophyll fluorescence data observed by remote sensing, to optimize the soil moisture transmission, vegetation dynamics and photosynthetic process of the Noah-MP model, and to produce a high-precision spatiotemporal continuous regional surface evapotranspiration data set that can be used in multiple fields, thereby reducing the errors caused by the lack of observation data or model parameters in traditional models. Although the origins and applications of data assimilation and machine learning are different, both use various observations to extract and interpret information. Both can be regarded as methods of inverting observation models. From the perspective of Bayesian theory, data assimilation and machine learning are equivalent. This algorithm is formally committed to solving how to combine data assimilation and machine learning algorithms, and optimize the key parameters and physical processes of the model by fusing site and multi-source remote sensing data, thereby realizing the spatiotemporal simulation of surface evapotranspiration.

[0008] The evapotranspiration simulation method of the present invention combines data assimilation and machine learning to overcome the limitations of traditional assimilation methods in model improvement and high-precision simulation by innovatively applying data assimilation and artificial intelligence technology to evapotranspiration simulation. The present invention significantly improves the accuracy and timeliness of evapotranspiration simulation results by real-time integration of multi-source observation information, and can adapt to complex environmental changes. The method of the present invention is widely applicable to many fields such as agricultural irrigation, water resources management, climate change research and ecological restoration. The innovation of the present invention lies in its integrated advantages of data assimilation and machine learning. By fusing multi-source data, the applicability and accuracy of evapotranspiration simulation are effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. As shown in the figure:

[0010] Figure 1The present invention provides a combined data assimilation and machine learning (DA-ML) method for estimating land surface evapotranspiration.

[0011] FIG. 2( a ) is a diagram showing the land cover types and observation station distribution in the Heihe River Basin of the present invention.

[0012] FIG. 2( b ) is a second diagram of the land cover types and observation station distribution in the Heihe River Basin of the present invention.

[0013] FIG. 2( c ) is three diagrams showing the land cover types and the distribution of observation sites in the Heihe River Basin according to the present invention.

[0014] FIG2( d ) is a fourth diagram of the land cover types and observation station distribution in the Heihe River Basin of the present invention.

[0015] FIG. 3( a ) is a diagram showing a site verification of the Noah-MP model and DA-ML method for simulating surface evapotranspiration in the Heihe River Basin of the present invention.

[0016] FIG. 3( b ) is a second diagram of the site verification of the Noah-MP model and DA-ML method for simulating surface evapotranspiration in the Heihe River Basin of the present invention.

[0017] FIG3( c ) is three graphs showing the site verification of the Noah-MP model and DA-ML method for simulating surface evapotranspiration in the Heihe River Basin of the present invention.

[0018] Figure 4 The Noah-MP model and DA-ML method of the Heihe River Basin of the present invention simulate the spatial distribution of soil moisture.

[0019] Figure 5 The Noah-MP model and DA-ML method of the Heihe River Basin of the present invention simulate the spatial distribution of surface evapotranspiration. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Example 1: Figure 1 , Figure 2(a), Figure 2(b), Figure 2(c), Figure 2(d), Figure 3(a), Figure 3(b), Figure 3(c), Figure 4 and Figure 5As shown in the figure, a surface evapotranspiration estimation method combining data assimilation and machine learning is designed to improve the accuracy and timeliness of evapotranspiration simulation by real-time fusion of site soil moisture observations, remote sensing leaf area index observations, and sunlight-induced chlorophyll fluorescence observations.

[0022] A method for estimating surface evapotranspiration by combining data assimilation and machine learning, comprising the following steps:

[0023] Step S1: Construct a machine learning bias correction model to achieve bias correction of soil moisture.

[0024] Step S2: Construct a machine learning observation operator to improve the photosynthetic process by sunlight-induced chlorophyll fluorescence.

[0025] Step S3: Construct a data assimilation system to integrate site soil moisture, remote sensing leaf area index and sunlight-induced chlorophyll fluorescence to optimize the model's physical processes and parameters.

[0026] Step S4: simulating and calculating the evapotranspiration using the optimized model to obtain a corrected evapotranspiration amount.

[0027] like Figure 1 As shown in FIG. 1 , a method for estimating surface evapotranspiration by combining data assimilation and machine learning also includes the following steps:

[0028] Step 1: Establish a machine learning bias correction model based on site observation data (soil moisture profile), including soil moisture bias modeling based on machine learning.

[0029] Step 2: Perform model simulation bias correction (multi-layer soil moisture) based on the Noah-MP model and the machine learning bias correction model, build a machine learning bias correction model to achieve soil moisture bias correction, and then go to step 6.

[0030] Step 3, remote sensing observation data (leaf area index), go to step 6,

[0031] Step 4, remote sensing observation data (sunlight-induced chlorophyll fluorescence), go to step 6,

[0032] Step 5: construct a machine learning observation operator to realize the improvement of photosynthetic process by sunlight-induced chlorophyll fluorescence. The machine learning observation operator step includes modeling sunlight-induced chlorophyll fluorescence based on machine learning and remote sensing observation of sunlight-induced chlorophyll fluorescence.

[0033] Step 6: Perform data assimilation algorithm to build a data assimilation algorithm to integrate site soil moisture, remote sensing leaf area index and sunlight-induced chlorophyll fluorescence to optimize the model's physical processes and parameters.

[0034] Step 7: Update model state variables and parameters (leaf biomass, multi-layer soil temperature and humidity, representative specific leaf area c g and V cmax25 ),

[0035] Step 8, surface evapotranspiration simulation (soil moisture, leaf area index, evapotranspiration), use the optimized model to simulate and calculate evapotranspiration to obtain the corrected evapotranspiration.

[0036] Example 2: Figure 1 , Figure 2(a), Figure 2(b), Figure 2(c), Figure 2(d), Figure 3(a), Figure 3(b), Figure 3(c), Figure 4 and Figure 5 As shown in Figure 1, a method for estimating surface evapotranspiration by combining data assimilation and machine learning contains the following steps:

[0037] 1. Initialize the model:

[0038] If T is defined as the initial value of the model state variable, at the beginning of the Ensemble Kalman Filter (EnKF) method, the initial value (T 0 ) is used to create a set of state vectors by adding normally distributed random errors (with mean zero and specified variance):

[0039] T j,0 =T 0 +μ j (1)

[0040] Where, T j,0 Represents the analysis variable at the initial moment. j is the mean zero background error.

[0041] 2. Model status prediction:

[0042] In the model state prediction stage, the prediction equation is:

[0043]

[0044] In the formula, represents the forecast variable at the k+1th moment; the superscript “f” indicates the forecast variable; represents the analysis variable at the kth moment; M represents the process model (also called the model operator), and F k+1 and α k+1 Represents the atmospheric driving data and model parameters at the k+1th moment. ω j is the model structural error of a Gaussian distribution with a mean of zero. y(-) is the bias correction model built based on machine learning methods (where x and P are its input and hyperparameters, respectively). The bias correction model is expressed as:

[0045] y(x,P)=ML(X 1 ,X 2 ,X 3 ,....) (3)

[0046] Machine learning methods are used to establish the explanatory variables (X 1 , X 2 , X 3 …) and soil moisture simulation bias. By constructing a soil moisture bias correction model, the soil moisture bias correction in the Noah-MP model data assimilation process is realized, thereby improving the multi-layer soil moisture simulation.

[0047] 3. Observation operator:

[0048] The observation operator can be expressed as:

[0049]

[0050] In the formula, represents the observed value at the k+1th moment of the simulation; H(–) represents the observation operator corresponding to the model state variable; η j It is an observation error with a Gaussian distribution of zero mean and variance R. In the process of assimilation of new remote sensing data, since the process model cannot directly simulate the remote sensing signal, this brings challenges to the assimilation of new remote sensing data.

[0051] Based on machine learning methods, an observation operator was constructed based on new remote sensing data (sunlight-induced chlorophyll fluorescence, SIF) and its influencing factor observation data. On this basis, the key parameters of photosynthesis (V cmax25 ) and improve the simulation accuracy of photosynthetic process. The relationship between SIF and influencing variables is:

[0052] SIF=ML(X 1 ,X 2 ,X 3 ,...) (6)

[0053] Machine learning methods are used to establish the explanatory variables (X 1 , X 2 , X 3 …) and SIF.

[0054] 4. Update of state variables and parameters:

[0055] The updating process of model parameters and state variables in the data assimilation method is as follows:

[0056]

[0057] K=Pf H T (HP f H T +R) -1 (8)

[0058]

[0059] In the formula, and Represent the analyzed value and predicted value of the state variable at the k+1th moment respectively; K is the Kalman gain, is the observed variable at time k+1, P f is the background error covariance matrix at time k+1, represents the predicted mean of the state variables at time k+1, and the superscript T represents the transpose of the matrix.

[0060] Example 3: Figure 1 , Figure 2(a), Figure 2(b), Figure 2(c), Figure 2(d), Figure 3(a), Figure 3(b), Figure 3(c), Figure 4 and Figure 5 As shown, a method for estimating land surface evapotranspiration by combining data assimilation and machine learning is provided. The embodiment of the present invention constructs a complete joint framework of data assimilation and machine learning, realizes the spatiotemporal simulation of land surface evapotranspiration, and includes the following steps:

[0061] 1. Construction of model operators

[0062] The Noah-MP (Noah Multi-Parameterization) model is an advanced land surface model widely used in climate and hydrological studies. Through a variety of parameterization schemes, the model can simulate soil moisture, energy exchange, and plant physiological processes, providing an important tool for studying water and energy cycles on the land surface. In the Noah-MP model, the calculation of evapotranspiration (ET) is a key link and is significantly affected by soil moisture and photosynthesis. Soil moisture directly determines the water availability of plants and affects their evapotranspiration capacity. Adequate soil moisture enables plants to photosynthesize effectively, thereby promoting growth and increasing evapotranspiration. On the contrary, when soil moisture is insufficient, plants will close their stomata to reduce water loss, which will cause actual evapotranspiration to be lower than potential evapotranspiration. By improving soil moisture and plant photosynthesis, the Noah-MP model can more accurately simulate the evapotranspiration process, thereby improving the reliability of climate models.

[0063] In the Noah-MP model, the energy exchange process between the canopy and the soil is considered separately, where the canopy energy exchange process is expressed as:

[0064] S W,C =LW,C +H C +LE C +ΔH C (11)

[0065] Among them, S W,C and L W,C represent the net shortwave radiation and longwave radiation absorbed by the canopy (W m -2 ). C and LE C Represents the sensible and latent heat fluxes of the canopy (W m -2 ). ΔH C Represents the change in canopy heat storage (W m -2 ). The energy exchange process of soil is expressed as,

[0066] S W,G =L W,G +H G +LE G +G 0 (12)

[0067] Among them, S W,G and L W,G represent the net shortwave radiation and longwave radiation absorbed by the soil (W m -2 ). G and LE G represent the sensible and latent heat fluxes of the soil (W m -2 ), G 0 is the soil heat flux (W m -2 ).

[0068] The sensible heat flux H can be calculated according to the overall transport method,

[0069] H=ρ a C h C p U(T g -T a ) (13)

[0070] Among them, ρ a is the air density (kg m -3 ), C p is the specific heat of air at constant pressure (J kg -1 K -1 ), U represents wind speed (ms -1 ). g and T a Represents the surface temperature and air temperature (K) respectively. h is the heat transfer coefficient and is calculated using the Monin–Obukhov similarity theory.

[0071] In the Noah-MP model, surface evapotranspiration is mainly composed of soil evaporation and vegetation transpiration. Vegetation transpiration can be expressed as:

[0072] E C =f veg ρ a C w C p (e s(Tv) -e c ) (14)

[0073] Among them, f veg is the vegetation cover and is calculated from LAI and stem area index (SAI). c represents the actual water vapor pressure, e s(Tv) Indicates the saturated water vapor pressure at vegetation temperature. C w Represents the transpiration coefficient from leaves to canopy.

[0074] Soil evaporation is expressed as,

[0075]

[0076] Among them, e s(Tg) represents the saturated water vapor pressure at the surface temperature, r a and r soil represent aerodynamic impedance and soil impedance (sm -1 ). g is the relative humidity of the air on the soil surface.

[0077] 2. Combined data assimilation and machine learning techniques

[0078] A combined data assimilation and machine learning evapotranspiration estimation method gives full play to the advantages of data assimilation and machine learning methods, with data assimilation system as the main method and machine learning method as the auxiliary method to achieve deep integration of multi-source data and Noah-MP model.

[0079] In this method, a soil moisture deviation correction model based on machine learning methods and site soil moisture profile observation data was constructed to achieve the deviation correction of the Noah-MP model and the optimization of multi-layer soil moisture on heterogeneous surfaces. The soil moisture deviation correction process is as follows:

[0080]

[0081] In the formula, represents the soil moisture forecast variable at the k+1th moment; represents the soil moisture analysis variable at the kth moment; y(-) is the soil moisture deviation correction model built based on machine learning methods. Normalized soil texture (ST), land cover type (LC), temperature (Ta ), wind speed (U), precipitation (P), relative humidity (RH), incident shortwave and longwave radiation (R s and R l ) observations were used as input to the machine learning method. The differences between the multi-layer soil moisture observations and the Noah-MP estimates were used as the output of the machine learning method to generate a bias-corrected model:

[0082] y(x,P)=ML(ST,LC,Ta,U,P,RH,Rs,R l ) (17)

[0083] Machine learning methods were used to establish the explanatory variables (ST, LC, T a ,U,P,RH,R s and R l ) and soil moisture simulation bias. By constructing a soil moisture bias correction model, the soil moisture bias correction in the Noah-MP model data assimilation process is realized, thereby improving the multi-layer soil moisture simulation.

[0084] On this basis, the data assimilation system assimilates the LAI inverted by remote sensing into the unbiased Noah-MP model, and optimizes the model state variables (leaf biomass) and model parameters (specific leaf area, c g ), improve the model's simulation accuracy of vegetation dynamics. In the process of leaf area index (LAI) assimilation, LAI and specific leaf area (c g ) and leaf biomass are related as follows:

[0085] LAI=c g ×B g (18)

[0086] In the formula, c g Represents specific leaf area (m 2 g -1 ), B g Represents the biomass of the leaf part (gm -2 ).

[0087] In addition, a SIF observation operator was constructed based on machine learning methods and remote sensing SIF observation data, realizing the direct simulation of SIF data by the mechanism model. By coupling the SIF observation operator to the multi-source data assimilation system, the remote sensing SIF data can be further assimilated; and then by optimizing the key parameters of photosynthesis (V cmax25 ), further improving the simulation accuracy of photosynthetic process in Noah-MP model. The relationship between S IF and influencing variables is:

[0088] SIF=ML(ST,LC,ET,GPP,LAI) (19)

[0089] Machine learning methods were used to establish the association between the explanatory variables (ST, LC, ET, GPP, LAI) and SIF.

[0090] In summary, this method improves soil moisture, vegetation dynamics simulation and photosynthetic process in the land surface process model by combining data assimilation and machine learning methods, ultimately improving the simulation accuracy of surface evapotranspiration.

[0091] Example 4: Figure 1 , Figure 2(a), Figure 2(b), Figure 2(c), Figure 2(d), Figure 3(a), Figure 3(b), Figure 3(c), Figure 4 and Figure 5 As shown in Figure 1, a method for estimating surface evapotranspiration by combining data assimilation and machine learning is shown in Figure 2. The Heihe River Basin is located in the middle of the Hexi Corridor (37.7°~42.7°N, 97.1°~102.0°E). The upper reaches of the Heihe River Basin originate from Qilian in Qinghai, and the middle reaches flow through Zhangye and Jiuquan in Gansu, and finally flow into the Juyan Lake in Ejina Banner, Inner Mongolia. It is a region with rich landforms, complex and changeable terrain, and high-altitude cold and arid areas. The Heihe River Basin is located in the interior of the Eurasian continent, far away from the ocean, and has a dry climate. The upper reaches of the Heihe River are the cryosphere of the Qilian Mountains, with an altitude of 1700-5564m, an annual average temperature of about 1.5-2.0℃, and an annual precipitation of about 400mm, belonging to the mountain climate zone; the middle reaches are artificial oasis areas + deserts, with an altitude of 1352-1700m, an annual average temperature of about 2.8-7.6℃, and a precipitation of 100-200mm, belonging to the arid climate zone; the lower reaches are natural oasis + Gobi desert, with an altitude of 912-1249m, an annual average temperature of about 8℃, and a precipitation of less than 50mm, belonging to the extremely arid climate zone. In recent years, climate warming has led to increased precipitation in the Heihe River basin and increased runoff due to the melting of glaciers, snow and permafrost; at the same time, the ecological water transfer project has redistributed water resources in the middle and lower reaches of the Heihe River and restored the ecosystem of the lower reaches of the Heihe River, but the expansion of cultivated land area and groundwater exploitation in the middle reaches have aggravated local ecological degradation.

[0092] This method selected the Heihe River Basin for method construction, site verification and regional spatial pattern analysis. The data of the comprehensive observation network of surface processes in the Heihe River Basin were selected for method construction, where Figure 2(a) shows the site distribution of the comprehensive observation network of surface processes in the Heihe River Basin. Figures 2(b), 2(c) and 2(d) are the locations and surface types of the three verification stations, Arou, Daman and Sidaoqiao, respectively. The observation data of the eddy covariance instrument at the three stations were used to verify the corresponding pixels within the observation range of the eddy covariance instrument. The spatial pattern of the evapotranspiration simulation results was mainly analyzed. The rationality of the spatial pattern of the data assimilation algorithm results is explained.

[0093] 1. Site Verification

[0094] Figures 3(a), 3(b), and 3(c) show the observed values ​​at the Arou, Daman, and Sidaoqiao stations and the daily evapotranspiration time series estimated by the Noah-MP model and the DA-ML method.

[0095] As shown in Figure 3(a), Figure 3(b), and Figure 3(c), the ET estimates of the Noah-MP model at the three sites are significantly lower than the observed values. This is because the Noah-MP model does not consider the effects of dense root systems and high soil organic carbon content on evapotranspiration. In addition, the Noah-MP model does not fully consider the effects of irrigation events and groundwater levels. Changes in midstream and shallow groundwater levels lead to changes in the soil moisture transport process. In contrast, the ET simulation values ​​of the DA-ML method are closer to the observed values ​​and are better than the Noah-MP model. This shows that the joint data assimilation and machine learning method can make full use of the effective information in the site soil moisture, remote sensing LAI, and SIF to provide high-precision ET estimates.

[0096] 2. Regional application

[0097] Figure 4 and 5 The seasonal changes of soil moisture and evapotranspiration in the Heihe River Basin from May to September are shown respectively.

[0098] like Figure 4 and 5As shown in the figure, the Noah-MP model underestimates the evapotranspiration in the middle and lower reaches of the Heihe River Basin in the oasis area. This is because the effects of irrigation in the oasis area in the middle reaches of the Heihe River and the lateral flow of shallow groundwater in the downstream riparian forest area on soil moisture and LAI are not considered in the Noah-MP model. Lower soil moisture and LAI reduce the simulation of evapotranspiration in the Noah-MP model. Compared with the Noah-MP model, the DA-ML method improves the simulation of soil moisture and vegetation dynamics in the oasis area in the middle and lower reaches of the Heihe River Basin by integrating multi-source remote sensing LAI, SIF and site soil moisture observation data, and further improves the estimation of evapotranspiration. The surface evapotranspiration simulated by the DA-ML method is compared with the remote sensing evapotranspiration product (ETMap). The results show that the DA-ML method can improve the simulation accuracy of evapotranspiration and keep it consistent with ETMap. In general, compared with the Noah-MP model, the DA-ML method improves the simulation accuracy of evapotranspiration, especially in the vegetation cover area.

[0099] The present invention improves the simulation accuracy of soil moisture transport, vegetation dynamics and photosynthetic process in the Noah-MP model and ultimately improves the simulation accuracy of evapotranspiration by fusing remote sensing LAI and SIF data with soil moisture site observation data.

[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for estimating surface evapotranspiration by combining data assimilation and machine learning, characterized in that ,The bias correction model and observation operator are coupled to the data assimilation method through machine learning methods, thereby achieving accurate estimation of surface evapotranspiration, real-time optimization of soil moisture transfer, vegetation dynamics, and photosynthetic processes of the key physical processes of the model, and finally obtaining improved surface evapotranspiration, which contains the following steps: initialization model step, model state prediction step, observation operator step, state variable and parameter update step.

2. The method for estimating evapotranspiration by combining data assimilation and machine learning according to claim 1, characterized in that: It also contains the following steps: Step 1: Establish a machine learning bias correction model based on site observation data (soil moisture profile), including soil moisture bias modeling based on machine learning. Step 2: Perform model simulation bias correction (multi-layer soil moisture) based on the Noah-MP model and the machine learning bias correction model, build a machine learning bias correction model to achieve soil moisture bias correction, and then go to step 6. Step 3, remote sensing observation data (leaf area index), go to step 6, Step 4, remote sensing observation data (sunlight-induced chlorophyll fluorescence), go to step 6, Step 5: construct a machine learning observation operator to realize the improvement of photosynthetic process by sunlight-induced chlorophyll fluorescence. The machine learning observation operator step includes modeling sunlight-induced chlorophyll fluorescence based on machine learning and remote sensing observation of sunlight-induced chlorophyll fluorescence. Step 6: Perform data assimilation algorithm to build a data assimilation algorithm to integrate site soil moisture, remote sensing leaf area index and sunlight-induced chlorophyll fluorescence to optimize the model's physical processes and parameters. Step 7: Update model state variables and parameters (leaf biomass, multi-layer soil temperature and humidity, representative specific leaf area c g and V cmax25 ), Step 8, surface evapotranspiration simulation (soil moisture, leaf area index, evapotranspiration), use the optimized model to simulate and calculate evapotranspiration to obtain the corrected evapotranspiration.

3. The method for estimating surface evapotranspiration by combining data assimilation and machine learning according to claim 1, characterized in that: It also contains a step to initialize the model, defining T as the initial value of the model state variables. At the beginning of the Ensemble Kalman Filter (EnKF) method, the initial value (T0) is used to create a set of state vectors by adding normally distributed random errors (with zero mean and specified variance): T j,0 =T0+μ j (1) Where, T j,0 represents the analysis variable at the initial moment, μ j is the mean zero background error.

4. The method for estimating evapotranspiration by combining data assimilation and machine learning according to claim 1, characterized in that: It also contains the model state prediction step, In the model state prediction stage, the prediction equation is: In the formula, represents the forecast variable at the k+1th moment; the superscript "f" indicates the forecast variable; represents the analysis variable at the kth moment; M represents the process model (also called model operator), F k+1 and α k+1 represents the atmospheric driving data and model parameters at the k+1th moment, ω j is the model structural error of a Gaussian distribution with a mean of zero, y(-) is a bias correction model built based on machine learning methods (where x and P are its input and hyperparameters, respectively), and the bias correction model is expressed as: y(x,P)=ML(X1,X2,X3,....) (3) Machine learning methods are used to establish the relationship between explanatory variables (X1, X2, X3...) and soil moisture simulation bias. By constructing a soil moisture bias correction model, soil moisture bias correction is achieved in the Noah-MP model data assimilation process, thereby improving multi-layer soil moisture simulation.

5. The method for estimating evapotranspiration by combining data assimilation and machine learning according to claim 1, characterized in that: It also contains the observation operator step: The observation operator is expressed as: In the formula, represents the observed value at the k+1th moment of the simulation; H(–) represents the observation operator corresponding to the model state variable; η j is the observation error of a Gaussian distribution with mean zero and variance R, Based on machine learning methods, an observation operator was constructed based on new remote sensing data (sunlight-induced chlorophyll fluorescence, SIF) and its influencing factor observation data, and the key parameters of photosynthesis (V cmax25 ) and improve the simulation accuracy of the photosynthetic process. The relationship between SIF and influencing variables is: SIF=ML(X1,X2,X3,...) (6) Machine learning methods are used to establish the relationship between the explanatory variables (X1, X2, X3…) and SIF.

6. The method for estimating evapotranspiration by combining data assimilation and machine learning according to claim 1, characterized in that: It also contains state variables and parameter update steps: The updating process of model parameters and state variables in the data assimilation method is as follows: K=P f H T (HP f H T +R) -1 (8) In the formula, and Represent the analyzed value and predicted value of the state variable at the k+1th moment respectively; K is the Kalman gain, is the observed variable at time k+1, P f is the background error covariance matrix at time k+1, represents the predicted mean of the state variables at time k+1, and the superscript T represents the transpose of the matrix.

7. The method for estimating evapotranspiration by combining data assimilation and machine learning according to claim 4, characterized in that: It also contains the following steps: In the Noah-MP model, the energy exchange process between the canopy and the soil is considered separately, where the canopy energy exchange process is expressed as: S W,C =L W,C +H C +LE C +ΔH C (11) Among them, S W,C and L W,C represent the net shortwave radiation and longwave radiation absorbed by the canopy (W m -2 ), H C and LE C Represents the sensible and latent heat fluxes of the canopy (W m -2 ), ΔH C Represents the change in canopy heat storage (W m -2 ), the energy exchange process of soil is expressed as, S W,G =L W,G +H G +THE G +G0 (12) Among them, S W,G and L W,G represent the net shortwave radiation and longwave radiation absorbed by the soil (W m -2 ), H G and LE G represent the sensible and latent heat fluxes of the soil (W m -2 ), G0 is the soil heat flux (W m -2 ), The sensible heat flux H is calculated according to the overall transport method, H=ρ a C h C p U(T g -T a ) (13) Among them, ρ a is the air density (kg m -3 ), C p is the specific heat of air at constant pressure (J kg -1 K -1 ), U represents wind speed (ms -1 ), T g and T a Represent the surface temperature and air temperature (K), C h is the heat transfer coefficient and is calculated using the Monin–Obukhov similarity theory.

8. The method for estimating surface evapotranspiration by combining data assimilation and machine learning according to claim 7, characterized in that: It also contains the following steps: In the Noah-MP model, surface evapotranspiration is mainly composed of soil evaporation and vegetation transpiration, where vegetation transpiration is expressed as, E C =f veg ρ a C w C p (e s(Tv) -e c ) (14) Among them, f veg is the vegetation cover and is calculated from LAI and stem area index (SAI), e c represents the actual water vapor pressure, e s(Tv) represents the saturated water vapor pressure at vegetation temperature, C w represents the transpiration coefficient from leaves to canopy, Soil evaporation is expressed as, Among them, e s(Tg) represents the saturated water vapor pressure at the surface temperature, r a and r soil represent aerodynamic impedance and soil impedance (sm -1 ), h g is the relative humidity of the air on the soil surface.

9. The method for estimating surface evapotranspiration by combining data assimilation and machine learning according to claim 4, characterized in that: It also contains the following steps: The deviation correction process of soil moisture is as follows: In the formula, represents the soil moisture forecast variable at the k+1th moment; represents the soil moisture analysis variable at the kth moment; y(-) is the soil moisture deviation correction model built based on machine learning methods, normalizing soil texture (ST), land cover type (LC), temperature (T a ), wind speed (U), precipitation (P), relative humidity (RH), incident shortwave and longwave radiation (R s and R l ) observations are used as input to the machine learning approach, and the differences between multi-layer soil moisture observations and Noah-MP estimates are used as output of the machine learning approach to generate a bias-corrected model: y(x,P)=ML(ST,LC,Ta,U,P,RH,Rs,R l ) (17) Machine learning methods were used to establish the explanatory variables (ST, LC, T a ,U,P,RH,R s and R l ) and soil moisture simulation bias. By constructing a soil moisture bias correction model, the soil moisture bias correction in the Noah-MP model data assimilation process is realized, thereby improving the multi-layer soil moisture simulation.

10. The method for estimating evapotranspiration by combining data assimilation and machine learning according to claim 8, characterized in that: It also contains the following steps: The LAI retrieved from remote sensing is assimilated into the unbiased Noah-MP model, and the model state variables (leaf biomass) and model parameters (specific leaf area, c g ), improve the model's simulation accuracy of vegetation dynamics, in the process of leaf area index (LAI) assimilation, LAI and specific leaf area (c g ) and leaf biomass are related as follows: LAI=c g ×B g (18) In the formula, c g Represents specific leaf area (m 2 g -1 ), B g Represents the biomass of the leaf part (gm -2 ), Based on machine learning methods and remote sensing SIF observation data, a SIF observation operator was constructed to achieve direct simulation of SIF data by the mechanism model. By coupling the SIF observation operator to the multi-source data assimilation system, the remote sensing SIF data was further assimilated. cmax25 ), further improving the simulation accuracy of photosynthetic process in Noah-MP model, the relationship between SIF and influencing variables is: SIF=ML(ST,LC,ET,GPP,LAI) (19) Machine learning methods were used to establish the association between the explanatory variables (ST, LC, ET, GPP, LAI) and SIF.

Images