A global surface maximum relative humidity inversion method based on physical constraints and machine learning

Abstract

The present invention provides a global maximum surface relative humidity inversion method based on physical constraints and machine learning, relating to the fields of hydrology and computer science. The method comprises the following steps: constructing a candidate dataset and obtaining the measured latent heat flux and synchronous meteorological elements at flux observation sites; inputting the candidate dataset and synchronous meteorological elements into a physical constraint model, calculating the simulated latent heat flux, and selecting the value with the smallest error to determine the optimal maximum surface relative humidity; training a machine learning inversion model using the optimal humidity and synchronous meteorological elements as samples; and merging the trained model with remotely sensed surface temperature, land cover type, and reanalysis meteorological data to infer and output a global gridded result. This method achieves the inversion of maximum surface relative humidity on a global scale, providing a new method for quantification.

Description

Technical Field

[0001] The present invention relates to the fields of hydrology and computer science, and in particular to a global surface maximum relative humidity inversion method based on physical constraints and machine learning. Background Art

[0002] Maximum surface relative humidity, by limiting the threshold effect of surface saturation vapor pressure, directly determines the distribution of extreme values ​​and diurnal variations in surface vapor pressure. Constrained by thermal conditions such as surface temperature and air pressure, it exhibits significant temporal and spatial heterogeneity due to factors such as topography, vegetation cover, and urbanization. Therefore, it regulates the exchange of matter and energy between the land and atmosphere at multiple levels and scales. Accurately quantifying maximum surface relative humidity not only provides a key parameter for revealing the underlying mechanisms of global-scale precipitation distribution and atmospheric circulation patterns, but also lays a theoretical foundation for assessing the frequency and intensity of extreme droughts, heavy rains, and other climate events under global warming, as well as for analyzing their causes.

[0003] In existing research, estimates of maximum surface relative humidity often rely on empirical values: in areas with dense vegetation or high soil moisture, a value of 0.8–1.0 is often assumed; in regions with low vegetation cover or arid conditions, a value of 0.6 or lower is often used. Although this method is easy to implement, it is highly subjective, has limited applicability, and ignores the interplay of multiple key factors, such as vegetation physiological conditions, soil moisture dynamics, and ambient temperature. Consequently, it is difficult to obtain accurate and reliable estimates of maximum surface relative humidity. In reality, maximum surface relative humidity is a fundamental parameter in the water and energy cycles, significantly influencing turbulent mixing in the atmospheric boundary layer, hydrological processes, precipitation formation, and surface evapotranspiration. However, currently, no method exists that both accounts for the coupling mechanisms of the large-scale atmosphere-vegetation-soil system and enables high-precision quantification under all-weather, all-time, and all-space conditions. This gap not only limits our understanding of the mechanisms of these processes but also, to a certain extent, hinders the improvement of the accuracy of environmental monitoring and ecohydrological models. Therefore, proposing a high-precision quantitative method for the maximum relative humidity of the surface that takes into account environmental factors, has wide applicability and all-weather operation capabilities has become a scientific problem that needs to be solved urgently. Summary of the Invention

[0004] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide a global surface maximum relative humidity inversion method based on physical constraints and machine learning. The dynamic interaction relationship between the maximum surface relative humidity and evapotranspiration is used as the physical constraint, and the machine learning algorithm is used as the inversion model construction method to realize the inversion of the maximum surface relative humidity on a global scale, which can provide a new method for accurately quantifying the maximum surface relative humidity.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A global surface maximum relative humidity inversion method based on physical constraints and machine learning, including:

[0007] Constructing a maximum surface relative humidity candidate data set; the maximum surface relative humidity candidate data set includes a plurality of candidate values ​​of the maximum surface relative humidity;

[0008] Obtain the measured latent heat flux and synchronous meteorological elements at the flux observation sites, and obtain remotely sensed surface temperature, land cover type and reanalysis meteorological data;

[0009] Inputting the surface maximum relative humidity candidate dataset and the synchronized meteorological elements into a physical constraint model, calculating the simulated latent heat flux corresponding to each candidate value, and comparing the simulated latent heat flux with the measured latent heat flux to select the candidate value with the smallest error in the simulated latent heat flux, and determining the candidate value with the smallest error as the optimal surface maximum relative humidity at the flux observation station;

[0010] According to the surface cover type, the optimal surface maximum relative humidity and synchronous meteorological elements are used as samples to train a machine learning inversion model;

[0011] The trained machine learning inversion model is integrated with the remote sensing surface temperature, the surface cover type and the reanalysis meteorological data to calculate the maximum relative humidity of the surface and output a global gridded result.

[0012] Preferably, the candidate values ​​of the maximum surface relative humidity in the maximum surface relative humidity candidate data set are generated by discretizing within a preset range.

[0013] Preferably, the expression of the physical constraint model includes:

[0014]

[0015]

[0016]

[0017]

[0018] in, h s,max is the maximum relative humidity of the surface; e s is the surface vapor pressure; Ω is the land-atmosphere coupling factor; e a is the actual water vapor pressure; e s *is the surface saturated vapor pressure; Δ is the slope of the saturated water vapor pressure curve; γ is the wet-bulb constant; r s is the surface impedance; r ae is the aerodynamic impedance; ρ is the air density; C p is the specific heat of air; T s is the surface temperature; T a is the air temperature; R n is the net radiation from the surface; G is the soil heat flux; ET is the latent heat flux.

[0019] Preferably, the surface maximum relative humidity candidate dataset and the synchronous meteorological elements are input into a physical constraint model, the simulated latent heat flux corresponding to each candidate value is calculated, and the simulated latent heat flux is compared with the measured latent heat flux to select the candidate value with the smallest error in the simulated latent heat flux, and the candidate value with the smallest error is determined as the optimal surface maximum relative humidity of the flux observation station, including:

[0020] Combining the candidate maximum relative humidity value with the synchronized meteorological element to form a multi-dimensional input feature array;

[0021] Inputting the multidimensional input feature array into the physical constraint model to calculate the simulated latent heat flux corresponding to each candidate value using the physical constraint model;

[0022] Comparing the simulated latent heat flux with the measured latent heat flux one by one to calculate the error of each candidate value;

[0023] The candidate value with the smallest error is selected and determined as the optimal maximum surface relative humidity for the flux observation site.

[0024] Preferably, the optimal maximum relative humidity on the surface and the synchronous meteorological elements are used as samples to train a machine learning inversion model, including:

[0025] Classifying the samples composed of the optimal surface maximum relative humidity and synchronous meteorological elements according to the surface cover type to form a sample set of multiple categories;

[0026] Pairing the optimal maximum surface relative humidity and the corresponding synchronous meteorological elements with the sample set of each category to construct a feature vector;

[0027] Performing data enhancement on the feature vector, and dividing the feature vector after data enhancement into a training set and a validation set;

[0028] The training set is used to fit the preset initial model, and the validation set is used to perform model evaluation to tune the model parameters to obtain the final machine learning inversion model.

[0029] Preferably, the machine learning inversion model includes any one of a random forest, a support vector machine and a neural network.

[0030] Preferably, the model evaluation method is a cross-validation method.

[0031] Preferably, the land cover types include: forests, grasslands, cities, farmlands and water bodies.

[0032] Preferably, the validation indicators of the machine learning inversion model include root mean square error and coefficient of determination.

[0033] Preferably, the numerical range of the candidate dataset of maximum relative humidity on the surface is [0, 1].

[0034] The present invention discloses the following technical effects:

[0035] Based on the physical constraint relationship between the maximum surface relative humidity and latent heat flux, the present invention uses the constructed surface maximum relative humidity candidate dataset and observation site data to calculate and screen the optimal surface maximum relative humidity; and uses a machine learning algorithm to construct an inversion model at the site scale. It further combines remotely sensed surface temperature data, surface cover type remotely sensed data and reanalysis meteorological data products, and uses the inversion model to realize the inversion of the global surface maximum relative humidity, which can provide a new method for accurate quantification, precise monitoring and simulation of the global surface water-gas exchange process. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] 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. 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 paying any creative work.

[0037] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of the technical route provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0040] The purpose of this invention is to provide a global surface maximum relative humidity inversion method based on physical constraints and machine learning. The dynamic interaction relationship between the surface maximum relative humidity and evapotranspiration is used as the physical constraint, and the machine learning algorithm is used as the inversion model construction method to realize the inversion of the surface maximum relative humidity on a global scale, which can provide a new method for accurately quantifying the surface maximum relative humidity.

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Figure 1 A flow chart of the method provided in the embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a global surface maximum relative humidity inversion method based on physical constraints and machine learning, including:

[0043] Step 100: constructing a maximum surface relative humidity candidate dataset; the maximum surface relative humidity candidate dataset includes a plurality of candidate values ​​of the maximum surface relative humidity;

[0044] Step 200: Obtain the measured latent heat flux and synchronous meteorological elements of the flux observation site, and obtain remotely sensed surface temperature, land cover type and reanalysis meteorological data;

[0045] Step 300: Input the surface maximum relative humidity candidate dataset and the synchronized meteorological elements into the physical constraint model, calculate the simulated latent heat flux corresponding to each candidate value, and compare the simulated latent heat flux with the measured latent heat flux to select the candidate value with the smallest simulated latent heat flux error. The candidate value with the smallest error is determined as the optimal surface maximum relative humidity for the flux observation station.

[0046] Step 400: training a machine learning inversion model based on the land cover type and taking the optimal maximum surface relative humidity and synchronous meteorological elements as samples;

[0047] Step 500: Utilize the trained machine learning inversion model to fuse with remote sensing surface temperature, land cover type and reanalysis meteorological data to estimate the maximum relative humidity of the surface in the unobserved area and output the global gridded result.

[0048] Specifically, such as Figure 2As shown, the technical route of this embodiment includes the following steps:

[0049] Step 1: Construct a candidate dataset of maximum relative humidity on the surface;

[0050] Step 2: Obtain global flux observation site data, surface temperature remote sensing data, remote sensing land cover type data, and reanalysis meteorological data products;

[0051] Step 3: Based on steps 1 and 2, the optimal maximum surface relative humidity is calculated and screened using the physical constraint model between maximum surface relative humidity and latent heat flux, the candidate surface maximum relative humidity dataset, and observation site data.

[0052] Step 4: Based on the optimal maximum surface relative humidity result obtained in step 3, meteorological variables are screened by combining correlation analysis, and a machine learning algorithm is further used to construct an inversion model for the maximum surface relative humidity of each land cover type.

[0053] Step 5: Based on the inversion model of the maximum surface relative humidity for each land cover type obtained in step 4, the global maximum surface relative humidity can be obtained using the global surface temperature remote sensing data, remote sensing land cover type data and reanalysis meteorological data products collected in step 2.

[0054] Optionally, in step 1, the candidate dataset of maximum relative humidity of the surface is: a self-constructed dataset with a range of [0, 1] and a step size of 0.01 for candidate values ​​of maximum relative humidity of the surface.

[0055] Optionally, in step 2, the global flux observation site data, surface temperature remote sensing data and reanalysis meteorological data products include: site-scale latent heat flux data, and site- and global-scale air temperature, surface temperature, soil heat flux, net radiation, vapor pressure deficit, relative humidity, absolute humidity, specific humidity, wind speed and atmospheric pressure data, as well as global remote sensing surface cover type data, including forest, farmland, grassland and other types.

[0056] Furthermore, in step 3, the physical constraint model between the maximum relative humidity of the surface and the latent heat flux is as follows:

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] in, h s,maxis the maximum relative humidity of the surface; e s is the surface vapor pressure; Ω is the land-atmosphere coupling factor; e a is the actual water vapor pressure; e s * is the surface saturated vapor pressure; Δ is the slope of the saturated water vapor pressure curve; γ is the wet-bulb constant; r s is the surface impedance; r ae is the aerodynamic impedance; ρ is the air density; C p is the specific heat of air; T s is the surface temperature; T a is the air temperature; R n is the net radiation from the surface; G is the soil heat flux; ET is the latent heat flux.

[0062] Furthermore, in step 3, the optimal maximum relative humidity of the surface is obtained by calculating and screening the candidate data set of maximum relative humidity of the surface and the observation station data.

[0063] Furthermore, based on the calculation and screening, the optimal maximum relative humidity of the surface is obtained. The specific implementation steps are as follows:

[0064] Step 3011: Obtain input data; obtain candidate values ​​from the basic surface maximum relative humidity dataset, denoted as:

[0065]

[0066] in Indicates discrete generation in the closed interval [0,1] with a step size of 0.01. j indivual h s,max candidate value;

[0067] Latent heat flux data and meteorological data (air temperature, surface temperature, soil heat flux, net radiation, vapor pressure deficit, relative humidity, absolute humidity, specific humidity, wind speed and atmospheric pressure data) are obtained from global flux observation sites; denoted as:

[0068]

[0069] in , Respectively represent iLatent heat flux and meteorological data corresponding to each station; is the total number of sites.

[0070] Step 3012, model calculation; based on the physical constraint model (denoted as: ), for each , calculated under the same meteorological conditions The simulated latent heat flux is expressed as:

[0071]

[0072] Step 3013, error comparison; based on the latent heat flux simulation results corresponding to each candidate value , compared with the measured latent heat flux data from global flux observation sites Compare and calculate the absolute error between the two, which is recorded as:

[0073]

[0074] Step 3014, optimal value determination; based on the absolute error between the simulation results and the measured data , for the i sites, from all candidate values Selected The smallest item is the optimal maximum relative humidity of the surface, and finally forms the result set at the site scale, which is recorded as:

[0075]

[0076] Optionally, in step 4, the optimal maximum relative humidity value of the surface of each observation station is calculated. , corresponding meteorological data and site surface cover type labels , and modeling was performed using machine learning algorithms based on land cover types.

[0077] Furthermore, the machine learning algorithm is a gradient boosting decision tree (XGBoost, eXtreme GradientBoosting) algorithm.

[0078] Furthermore, an inversion model is constructed based on the algorithm, and the specific implementation steps are as follows:

[0079] Step 4011, site grouping; for each site i , read its surface cover category identification (such as forest, grassland, farmland, etc.) All sites are grouped according to coverage type to form several subsets, which are recorded as:

[0080]

[0081] in T is the total number of coverage types.

[0082] Step 4012: construct a grouped sample set; for each type t , construct a training set of this type, recorded as:

[0083]

[0084] in, .

[0085] Step 4013, feature selection; for each subset Correlation tests (Pearson coefficient) were performed on all the variables, some variables with low correlation were eliminated, and the main features were retained. middle and The correlation coefficient between them can be obtained by the following formula:

[0086]

[0087] in, n For subset the amount of data included; For subset Middle i The meteorological variable value corresponding to each data point; For this variable in the subset The mean of For subset The i-th data point corresponds to value; For subset middle Mean.

[0088] Furthermore, each subset Correlation tests were performed, i.e. calculation and (including correlation coefficients between air temperature, surface temperature, soil heat flux, net radiation, vapor pressure deficit, relative humidity, absolute humidity, specific humidity, wind speed and atmospheric pressure data).

[0089] Step 4014, model training: using the selected highly correlated meteorological variables, the data of the training set and the validation set are divided into a ratio of 7:3, and the gradient boosting decision tree (XGBoost) algorithm is used for each type. t The training surface maximum relative humidity inversion model is recorded as:

[0090]

[0091] in, Surface type t Corresponding XGBoost model; are hyperparameters, including but not limited to tree depth, learning rate, and subsampling rate.

[0092] Further, with | r xy |>0.5 is used as the judgment of meteorological variables and The level of correlation between them.

[0093] Step 4015: Model optimization. For each type of inversion model, XGBoost's core hyperparameters (including the learning rate, maximum tree depth, subsampling ratio, column sampling ratio, and number of weak learners) are optimized using the Bayesian optimization algorithm. Regularization terms (L1 / L2) are also introduced to control model complexity and reduce overfitting. Minimizing the loss function (mean squared error) improves model accuracy and generalization.

[0094] Specifically, the optimal hyperparameters of the inversion model for each type of land are determined based on the minimization of the mean square error. The form of the mean square error is:

[0095]

[0096] Step 4016, model output; for each type t , generate the final surface maximum relative humidity inversion model, recorded as:

[0097]

[0098] in, Corresponding surface type t The inversion model of the maximum relative humidity of the surface; For There are meteorological variables with high correlation; is the optimal hyperparameter.

[0099] Step 4017, model validation: Validate the inversion model using an independent validation dataset.

[0100] Furthermore, the root mean square error and the coefficient of determination were used as validation indicators.

[0101] Furthermore, the training process will stop when any of the following conditions are met:

[0102] When the value of the loss function is continuous Q When there is no significant improvement within 3 rounds, training is stopped.

[0103] When the value of the loss function is continuous R When the change within a round is less than the preset threshold, the training stops.

[0104] If the validation set loss is continuous Z When the number of rounds continues to rise, training stops.

[0105] When the number of training rounds reaches the preset maximum number of training rounds, the training stops.

[0106] Based on the physical constraint relationship between the maximum surface relative humidity and latent heat flux, the present invention uses the constructed surface maximum relative humidity candidate dataset and observation site data to calculate and screen the optimal surface maximum relative humidity; and uses a machine learning algorithm to construct an inversion model at the site scale. It further combines remotely sensed surface temperature data, surface cover type remotely sensed data and reanalysis meteorological data products, and uses the inversion model to realize the inversion of the global surface maximum relative humidity, which can provide a new method for accurate quantification, precise monitoring and simulation of the global surface water-gas exchange process.

[0107] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0108] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A global surface maximum relative humidity inversion method based on physical constraints and machine learning, characterized by: include: Constructing a maximum surface relative humidity candidate data set; the maximum surface relative humidity candidate data set includes a plurality of candidate values ​​of the maximum surface relative humidity; Obtain the measured latent heat flux and synchronous meteorological elements at the flux observation sites, and obtain remotely sensed surface temperature, land cover type and reanalysis meteorological data; Inputting the surface maximum relative humidity candidate dataset and the synchronized meteorological elements into a physical constraint model, calculating the simulated latent heat flux corresponding to each candidate value, and comparing the simulated latent heat flux with the measured latent heat flux to select the candidate value with the smallest error in the simulated latent heat flux, and determining the candidate value with the smallest error as the optimal surface maximum relative humidity at the flux observation station; According to the surface cover type, the optimal surface maximum relative humidity and synchronous meteorological elements are used as samples to train a machine learning inversion model; Utilizing the trained machine learning inversion model to fuse with the remote sensing surface temperature, the surface cover type and the reanalysis meteorological data, the maximum relative humidity of the surface is calculated, and a global gridded result is output; Inputting the surface maximum relative humidity candidate dataset and the synchronized meteorological elements into a physical constraint model, calculating the simulated latent heat flux corresponding to each candidate value, and comparing the simulated latent heat flux with the measured latent heat flux to select the candidate value with the smallest simulated latent heat flux error, and determining the candidate value with the smallest error as the optimal surface maximum relative humidity at the flux observation station, including: Combining the candidate value of the maximum relative humidity of the ground surface with the synchronous meteorological element to form a multi-dimensional input feature array; Inputting the multidimensional input feature array into the physical constraint model to calculate the simulated latent heat flux corresponding to each candidate value using the physical constraint model; Comparing the simulated latent heat flux with the measured latent heat flux one by one to calculate the error of the simulated latent heat flux corresponding to each candidate value; The candidate value with the smallest error is selected and determined as the optimal maximum surface relative humidity for the flux observation site.

2. The global surface maximum relative humidity inversion method based on physical constraints and machine learning according to claim 1 is characterized in that: The candidate values ​​of the maximum surface relative humidity in the maximum surface relative humidity candidate data set are generated by discretizing within a preset range.

3. The global surface maximum relative humidity inversion method based on physical constraints and machine learning according to claim 1 is characterized in that: The expression of the physical constraint model includes: in, h s,max is the maximum relative humidity of the surface; e s is the surface vapor pressure; Ω is the land-atmosphere coupling factor; e a is the actual water vapor pressure; e s * is the surface saturated vapor pressure; Δ is the slope of the saturated water vapor pressure curve; γ is the wet-bulb constant; r s is the surface impedance; r ae is the aerodynamic impedance; ρ is the air density; C p is the specific heat of air; T s is the surface temperature; T a is the air temperature; R n is the net radiation from the surface; G is the soil heat flux; ET is the latent heat flux.

4. The global surface maximum relative humidity inversion method based on physical constraints and machine learning according to claim 1 is characterized in that: The optimal maximum surface relative humidity and synchronous meteorological elements are used as samples to train a machine learning inversion model, including: Classifying the samples composed of the optimal surface maximum relative humidity and synchronous meteorological elements according to the surface cover type to form a sample set of multiple categories; Pairing the optimal maximum surface relative humidity and the corresponding synchronous meteorological elements with the sample set of each category to construct a feature vector; Performing data enhancement on the feature vector, and dividing the feature vector after data enhancement into a training set and a validation set; The training set is used to fit the preset initial model, and the validation set is used to perform model evaluation to tune the model parameters to obtain the final machine learning inversion model.

5. The global surface maximum relative humidity inversion method based on physical constraints and machine learning according to claim 1 is characterized in that: The machine learning inversion model includes any one of a random forest, a support vector machine and a neural network.

6. The global surface maximum relative humidity inversion method based on physical constraints and machine learning according to claim 1 is characterized in that: The model evaluation method is cross-validation method.

7. The global surface maximum relative humidity inversion method based on physical constraints and machine learning according to claim 1 is characterized in that: The land cover types include: forests, grasslands, cities, farmlands and water bodies.

8. The global surface maximum relative humidity inversion method based on physical constraints and machine learning according to claim 1 is characterized in that: The validation metrics of the machine learning inversion model include root mean square error and coefficient of determination.

9. The global surface maximum relative humidity inversion method based on physical constraints and machine learning according to claim 1 is characterized in that: The value range of the candidate dataset of maximum relative humidity on the surface is [0, 1].

Images