A Global Surface Water Vapor Pressure Inversion Method Based on Surface Energy Distribution Constraints
By constructing a water vapor pressure inversion model based on surface energy distribution constraints, eliminating aerodynamic impedance terms, and combining machine learning techniques, the inaccuracy problem of surface water vapor pressure estimation in existing technologies has been solved, achieving continuous inversion and improved accuracy of global surface water vapor pressure.
Patent Information
- Application Number
- CN202611244562.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-17
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-17
AI Technical Summary
Existing methods for estimating surface water vapor pressure rely on empirical assumptions and aerodynamic impedance parameterization, resulting in inaccurate inversion results on a global scale and making it difficult to reflect the spatiotemporal differences in land cover type and climate background.
Based on the constraints of surface energy distribution, a physical constraint inversion model for surface water vapor pressure is constructed. The aerodynamic impedance term is eliminated by using the sensible heat flux, aerodynamic impedance expression and surface energy balance relationship. Combined with global flux observation station data and remote sensing data, water vapor pressure inversion under different surface cover types is realized through machine learning inversion model.
It enables continuous inversion of surface water vapor pressure in unobserved areas, reduces the influence of aerodynamic impedance parameterization, improves the accuracy and applicability of the inversion results, and can reflect the differences in water vapor pressure and meteorological elements under different surface cover conditions.
Smart Images

Figure CN122778960B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hydrological remote sensing inversion technology, and in particular to a global surface water vapor pressure inversion method based on surface energy distribution constraints. Background Technology
[0002] Surface water vapor pressure is an important physical quantity that characterizes the state of surface water vapor, evaporation capacity, and the intensity of water vapor exchange between land and atmosphere.
[0003] Existing methods for estimating surface water vapor pressure mostly rely on empirical assumptions or indirect parameterization. One type of method typically makes approximate estimates based on actual air water vapor pressure, near-surface relative humidity, or surface saturation water vapor pressure, which is insufficient to fully characterize the spatiotemporal differences under different land cover types, climate backgrounds, and energy distribution conditions. Another type of method is based on the theory of land-atmosphere turbulent exchange, estimating through impedance models or gradient diffusion relationships, but usually requires pre-determining aerodynamic impedance. Since aerodynamic impedance is influenced by factors such as wind speed, surface roughness, atmospheric stability, canopy structure, and topographic conditions, it is difficult to accurately describe it on a global scale using uniform empirical parameters. Its parameterization error directly affects the reliability and applicability of surface water vapor pressure inversion. Summary of the Invention
[0004] The purpose of this invention is to provide a global surface water vapor pressure inversion method based on surface energy distribution constraints, which can solve the problems of existing surface water vapor pressure estimation relying on empirical assumptions, large uncertainty in aerodynamic impedance parameterization, and limited spatial coverage of flux observation stations, and realize continuous inversion of surface water vapor pressure in unobserved areas.
[0005] To achieve the above objectives, the present invention provides the following solution: A global surface water vapor pressure inversion method based on surface energy distribution constraints includes: Acquire global flux observation station data, remote sensing surface temperature data, land cover type data, and reanalysis meteorological data; Based on the sensible heat flux expression, the aerodynamic impedance expression, and the surface energy balance relationship, the aerodynamic impedance term is eliminated, and a physical constraint inversion model of surface water vapor pressure is constructed. Based on flux data and synchronous meteorological elements from global flux observation stations, the effective value of surface water vapor pressure at the scale of global flux observation stations is obtained using a physical constraint inversion model of surface water vapor pressure. Based on land cover type data, the samples composed of effective values of surface water vapor pressure and synchronous meteorological elements are classified to form multiple land cover type sample sets; Multiple input feature vectors were constructed for sample sets of different surface cover types, and the effective value of surface water vapor pressure corresponding to each input feature vector was used as the training label. Establish machine learning inversion models corresponding to each land cover type, and use each input feature vector and corresponding training label to train and optimize the machine learning inversion models to obtain the final machine learning inversion models corresponding to each land cover type. Based on land cover type data, determine the land cover type of each pixel to be inverted in the unobserved area; For each pixel to be inverted, a pixel input feature vector with the same input variables as the model training phase is constructed based on remote sensing surface temperature data and reanalysis meteorological data. The pixel input feature vector is input into the final machine learning inversion model corresponding to the land cover type of the pixel to be inverted, and the surface water vapor pressure of the pixel to be inverted is obtained. The surface water vapor pressure of each pixel to be inverted in the unobserved area is collected to obtain the global surface water vapor pressure results.
[0006] Preferably, based on the sensible heat flux expression, the aerodynamic impedance expression, and the surface energy balance relationship, the aerodynamic impedance term is eliminated, and a surface water vapor pressure physical constraint inversion model is constructed, including: Based on the surface sensible heat exchange relationship, establish the expression for sensible heat flux:
[0007] Based on the aerodynamic impedance relationship and the surface energy balance relationship, the aerodynamic impedance expression is established as follows:
[0008] By combining the expressions for sensible heat flux and aerodynamic impedance, and eliminating the aerodynamic impedance term, a physical constraint inversion model for surface water vapor pressure is obtained:
[0009] In the formula, This is the initial value of surface water vapor pressure. This represents the actual water vapor pressure in the air. The dry-bulb and wet-bulb constants are... For surface temperature, For air temperature, For latent heat flux, For sensible heat flux, air density, The specific heat of air at constant pressure. It is the aerodynamic impedance.
[0010] Preferably, based on flux data and synchronous meteorological elements from global flux observation station data, the effective value of surface water vapor pressure at the global flux observation station scale is obtained using a physical constraint inversion model, including: Quality screening was performed on daily-scale flux data and corresponding daily-scale meteorological data from global flux observation stations. Daily-scale records with unqualified quality labels, missing sensible heat flux, latent heat flux, air temperature, surface temperature or actual air vapor pressure, sensible heat flux less than or equal to zero, or inconsistent signs of the difference between surface temperature and air temperature and sensible heat flux were removed. Based on the station number and date, the sensible heat flux, latent heat flux, air temperature, surface temperature and actual air vapor pressure of the same global flux observation station on the same day are matched accordingly. The sensible heat flux, latent heat flux, air temperature, surface temperature, and actual air vapor pressure are input into the physical constraint inversion model of surface vapor pressure to obtain the initial values of surface vapor pressure at the flux observation station. The initial values of surface water vapor pressure at flux observation stations are constrained by physical range based on the saturated water vapor pressure corresponding to surface temperature. The initial values of surface water vapor pressure at flux observation stations that meet the physical range constraints are determined as the effective values of surface water vapor pressure at the global flux observation station scale.
[0011] Preferably, the first The saturated vapor pressure corresponding to the surface temperature of each global flux observation station is calculated using the following formula:
[0012] The physical range constraints are:
[0013] In the formula, For the first Initial values of surface water vapor pressure at 10 global flux observation stations For the first Saturated vapor pressure at a global flux observation station For the first Surface temperature at a global flux observation station It is an exponential function with the natural constant as the base. The unit of surface temperature is degrees Celsius, and the units of the initial value of surface water vapor pressure and saturated water vapor pressure are kilopascals.
[0014] Preferably, based on land cover type data, the samples composed of effective values of surface water vapor pressure and synchronous meteorological elements are classified to form multiple land cover type sample sets, including: The land cover type of each global flux observation station was determined based on land cover type data; A sample is formed by combining the effective value of surface water vapor pressure and synchronous meteorological elements corresponding to each global flux observation station. Based on the land cover type of each global flux observation station, each sample is assigned to the corresponding land cover type sample set; Land cover types include forest, farmland, grassland, and savanna and shrubland.
[0015] Preferably, multiple input feature vectors are constructed for each land cover type sample set, and the effective value of land vapor pressure corresponding to each input feature vector is used as the training label, including: Surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation are extracted from the synchronous meteorological elements of each sample. An input feature vector is constructed based on the surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation corresponding to each sample. The effective value of surface water vapor pressure corresponding to each sample is used as the training label corresponding to the input feature vector.
[0016] Preferably, a machine learning inversion model corresponding to each land cover type is established, and the machine learning inversion model is trained and optimized using each input feature vector and its corresponding training label to obtain the final machine learning inversion model corresponding to each land cover type, including: The input feature vector and corresponding training label in each land cover type sample set are divided into a training set, a validation set, and an independent test set. Regarding the first Establish machine learning inversion models based on land cover types:
[0017] To minimize the error between the predicted surface water vapor pressure and the training labels, a loss function is established:
[0018] The model parameters of the machine learning inversion model are fitted using the training set, and the model parameters are optimized using the validation set to obtain the final machine learning inversion model. The final machine learning inversion model was validated using an independent test set, with validation metrics including root mean square error and coefficient of determination. In the formula, For the first Predicted surface water vapor pressure corresponding to different land cover types For the first Machine learning inversion models corresponding to land cover types For the first The input feature vector corresponding to the land cover type. For the first Model parameters corresponding to land cover types, For the first The loss function corresponding to the land cover type, For the first The number of samples in the training set for each land cover type. For the first The first training set of land cover types The input feature vector of each sample, To and The corresponding training labels.
[0019] Preferably, for the first A machine learning inversion model is established for land cover types. The model parameters are fitted using the training set, and the model parameters are optimized using the validation set to obtain the final machine learning inversion model, including: The machine learning inversion model was determined to be a gradient boosting decision tree model. Multiple regression trees are built sequentially. The residuals between the training labels and the predicted surface water vapor pressure in the previous round are fitted using the regression trees built in each round. The prediction results of each regression tree are accumulated according to the learning rate. The number of regression trees, learning rate, maximum tree depth, and minimum number of leaf nodes were determined as the model parameters to be optimized. A grid search method was used to search for parameter combinations for each model parameter, and the minimum root mean square error of the validation set was used as the parameter selection criterion. The candidate values for the number of regression trees were 80, 110, 140, 170, and 200; the candidate values for the learning rate were 0.01, 0.03, 0.05, and 0.10; the candidate values for the maximum tree depth were 2, 3, 4, and 5; and the candidate values for the minimum number of leaf nodes were 5, 8, 10, 15, and 20. Based on the parameter selection criteria, the number of regression trees was set to 140, the learning rate to 0.05, the maximum tree depth to 2, and the minimum number of samples per leaf node to 8.
[0020] Preferably, for each pixel to be inverted, based on remote sensing surface temperature data and reanalysis meteorological data, a pixel input feature vector is constructed that uses the same input variables as in the model training phase, including: Obtain the surface temperature of the pixel to be inverted from remote sensing surface temperature data; The atmospheric pressure, water vapor pressure deficit, and net radiation of the pixels to be retrieved are determined based on reanalysis meteorological data. The pixel input feature vector is constructed based on the surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation of the pixel to be inverted, so that the pixel input feature vector uses the same input variables as the input feature vector in the model training stage.
[0021] Preferably, the pixel input feature vector is input into the final machine learning inversion model corresponding to the land cover type of the pixel to be inverted, to obtain the surface water vapor pressure of the pixel to be inverted, including: Based on the land cover type of the pixel to be inverted, select the final machine learning inversion model corresponding to the land cover type of the pixel to be inverted from the final machine learning inversion models corresponding to forest, farmland, grassland, savanna and shrubland respectively. The pixel input feature vector of the pixel to be inverted is input into the selected final machine learning inversion model, and the surface water vapor pressure of the pixel to be inverted is calculated using the selected final machine learning inversion model.
[0022] The present invention discloses the following beneficial effects: This invention eliminates the aerodynamic impedance term based on the sensible heat flux expression, the aerodynamic impedance expression, and the surface energy balance relationship, constructing a physically constrained inversion model for surface water vapor pressure. This reduces the impact of empirical parameterization of aerodynamic impedance on the inversion results. It calculates the effective value of surface water vapor pressure using data from global flux observation stations, and forms sample sets according to land cover type, training corresponding machine learning inversion models to reflect the differences between surface water vapor pressure and meteorological elements under different land cover conditions. Furthermore, it constructs input feature vectors for the pixels to be inverted based on remotely sensed surface temperature data and reanalysis meteorological data, and calls the final machine learning inversion model corresponding to the land cover type for pixel-by-pixel calculation. This solves the problem of limited spatial distribution of flux observation stations and difficulty in directly covering unobserved areas, enabling continuous inversion of surface water vapor pressure in unobserved areas. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 A flowchart of the method provided in an embodiment of the present invention; Figure 2 Scatter plot of farmland surface water vapor pressure inversion results provided in this embodiment of the invention; Figure 3 Scatter plot of forest surface water vapor pressure inversion results provided in this embodiment of the invention; Figure 4 A scatter plot of grassland surface water vapor pressure inversion results provided in this embodiment of the invention; Figure 5 This is a scatter plot of the surface water vapor pressure inversion results for sparse grasslands and shrublands provided in an embodiment of the present invention. Figure 6 A grid diagram of the surface water vapor pressure inversion results provided in an embodiment of the present invention. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0027] Figure 1 A flowchart of the method provided in the embodiments of the present invention, such as Figure 1 As shown, this embodiment provides a global surface water vapor pressure inversion method based on surface energy distribution constraints, including: Step 101: Acquire global flux observation station data, remote sensing surface temperature data, land cover type data, and reanalysis meteorological data; Step 102: Based on the sensible heat flux expression, the aerodynamic impedance expression, and the surface energy balance relationship, eliminate the aerodynamic impedance term and construct a surface water vapor pressure physical constraint inversion model; Step 103: Based on flux data and synchronous meteorological elements from global flux observation stations, obtain the effective value of surface water vapor pressure at the scale of global flux observation stations using the physical constraint inversion model of surface water vapor pressure. Step 104: Based on the land cover type data, classify the samples composed of effective values of surface water vapor pressure and synchronous meteorological elements to form multiple land cover type sample sets; Step 105: Construct multiple input feature vectors for each land cover type sample set, and use the effective value of land vapor pressure corresponding to each input feature vector as the training label; Step 106: Establish machine learning inversion models corresponding to each land cover type, and train and optimize the machine learning inversion models using each input feature vector and the corresponding training labels to obtain the final machine learning inversion models corresponding to each land cover type. Step 107: Based on the land cover type data, determine the land cover type of each pixel to be inverted in the unobserved area; Step 108: For each pixel to be inverted, construct a pixel input feature vector based on remote sensing surface temperature data and reanalysis meteorological data, using the same input variables as in the model training phase; Step 109: Input the pixel input feature vector into the final machine learning inversion model corresponding to the land cover type of the pixel to be inverted, and obtain the surface water vapor pressure of the pixel to be inverted; Step 110: Collect the surface water vapor pressure of each pixel to be inverted in the unobserved area to obtain the global surface water vapor pressure results.
[0028] The available energy at the Earth's surface is mainly distributed as sensible heat flux and latent heat flux. Sensible heat flux characterizes the heat exchange between the surface and near-surface air driven by temperature differences, while latent heat flux characterizes water vapor transport and energy consumption during evaporation and transpiration. Based on the expressions for sensible heat flux, aerodynamic impedance, and surface energy balance, relationships can be established between surface water vapor pressure and actual air water vapor pressure, surface temperature, air temperature, latent heat flux, and sensible heat flux, eliminating the aerodynamic impedance term during the derivation process.
[0029] In step 101, the global flux observation station data uses daily flux and meteorological data provided by EuropeFlux and AmeriFlux, covering the period from 2000 to 2021. The global flux observation station data includes at least sensible heat flux, latent heat flux, air temperature, surface temperature, and actual atmospheric vapor pressure, and also includes soil heat flux, net radiation, relative humidity, specific humidity, wind speed, and atmospheric pressure for subsequent feature screening.
[0030] The data from global flux observation stations underwent quality screening, excluding diurnal records with substandard quality labels, missing sensible heat flux, latent heat flux, air temperature, surface temperature, or actual vapor pressure, sensible heat flux less than or equal to 0, and diurnal records where the difference between surface temperature and air temperature did not match the sign of the sensible heat flux. After quality screening, 16,969 valid diurnal samples from 33 global flux observation stations covering forests, farmland, grasslands, savannahs, and shrublands were obtained.
[0031] Remote sensing surface temperature data were obtained using the MODIS MOD11A1 Collection 6.1 diurnal product with a raw spatial resolution of 1 km. Land cover type data were obtained using the ESA WorldCover 2021 v200 product with a raw spatial resolution of 10 m. Reanalysis meteorological and radiation data were obtained using NASA POWER MERRA-2 data and CERES SYN1deg data.
[0032] Before global pixel-by-pixel inversion, remote sensing land surface temperature data, reanalysis meteorological data, and radiation data were unified to a daily scale, and remote sensing land surface temperature data, land cover type data, reanalysis meteorological data, and radiation data were unified to a 0.025° spatial grid. Land surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation are continuous variables, and bilinear interpolation was used for continuous variables; land cover type is a discrete categorical variable, and mode resampling was used for land cover type. After data processing, the same spatial grid has corresponding land surface temperature, atmospheric pressure, water vapor pressure deficit, net radiation, and land cover type for the same date.
[0033] In step 102, the physical constraint inversion model for surface water vapor pressure is expressed as formula (1): (1) This represents the initial value of surface water vapor pressure; This is the actual water vapor pressure in the air; These are the wet-bulb and dry-bulb constants; It refers to the surface temperature; Air temperature; Latent heat flux; This refers to the sensible heat flux.
[0034] Formula (1) is derived by combining the expressions for sensible heat flux, aerodynamic impedance, and surface energy balance. The expression for sensible heat flux is given by formula (2): (2) The aerodynamic impedance expression is given by formula (3): (3) air density; The specific heat of air at constant pressure; It is the aerodynamic impedance.
[0035] Substituting formula (3) into formula (2), we get:
[0036] Simplify the above equation to eliminate the aerodynamic resistance term. Formula (1) is obtained. Therefore, subsequent calculations do not require setting separate empirical parameters of aerodynamic impedance determined by wind speed, surface roughness, atmospheric stability, canopy structure and topographic conditions.
[0037] In step 103, the global flux observation station data are mapped according to the station number and date. For the same global flux observation station on the same date, sensible heat flux, latent heat flux, air temperature, surface temperature, and actual air vapor pressure are combined to form a multidimensional input feature array.
[0038] The multidimensional input feature array is input into the physical constraint inversion model of surface water vapor pressure. For a valid diurnal sample of the i-th global flux observation station, the initial value of surface water vapor pressure at the flux observation station is expressed as formula (4): (4) Let be the initial value of the surface water vapor pressure at the flux observation station corresponding to the i-th global flux observation station; Let be the actual water vapor pressure of the air at the i-th global flux observation station; Let be the wet-bulb constant corresponding to the i-th global flux observation station; Let be the surface temperature of the i-th global flux observation station; Let be the air temperature at the i-th global flux observation station; Let be the latent heat flux of the i-th global flux observation station; Let be the sensible heat flux at the i-th global flux observation station.
[0039] Calculate the corresponding saturated vapor pressure based on the surface temperature of the i-th global flux observation station: (5) Let be the saturated vapor pressure corresponding to the surface temperature of the i-th global flux observation station; in formula (5) Using Celsius, The unit is kPa.
[0040] Physical range constraints are applied to the initial values of surface water vapor pressure at flux observation stations based on saturated water vapor pressure. (6) The initial surface water vapor pressure values of flux observation stations that satisfy formula (6) are determined as the effective surface water vapor pressure values. Samples that do not satisfy formula (6) are marked as outliers or removed from subsequent training samples. After completing the calculations for all global flux observation stations, a set of effective surface water vapor pressure values at the global flux observation station scale is formed.
[0041] In step 104, the land cover category identifier for each global flux observation station is read based on the land cover type data. Land cover category identifiers include forest, farmland, grassland, savanna, and shrubland. According to the land cover category identifiers, the effective values of surface water vapor pressure and synchronous meteorological elements for the same date corresponding to each global flux observation station are classified to form multiple land cover type sample sets.
[0042] The complete land cover type sample set is represented as follows: (7) It is a collection of samples of all land cover types; This represents the total number of land cover types. This is the sample set for the t-th type of land cover. In this embodiment, =4.
[0043] The sample set of land cover type t is represented as follows: (8) Let be the input feature vector of the k-th sample in the t-th land cover type; To and The corresponding effective value of surface water vapor pressure; Let be the number of samples for the t-th land cover type.
[0044] In step 105, a correlation test is performed for each land cover type sample set. Candidate variables for the correlation test include air temperature, land surface temperature, soil heat flux, net radiation, actual air vapor pressure, relative humidity, specific humidity, vapor pressure deficit, wind speed, and atmospheric pressure.
[0045] The correlation coefficients between the effective value of surface water vapor pressure and each candidate variable were calculated. Based on the correlation test results, variables with low correlation were eliminated, and finally, surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation were retained as model input variables.
[0046] The input feature vector corresponding to the t-th land cover type is represented as: (9) , … These are the input variables used for surface water vapor pressure inversion; This represents the number of input variables. In this embodiment, =4, with the four input variables being surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation, respectively.
[0047] For the k-th sample in the t-th land cover type, construct the input feature vector based on the corresponding land surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation. and the effective value of surface water vapor pressure corresponding to the same sample As training labels.
[0048] In step 106, an initial machine learning inversion model is established for each land cover type: (10) The predicted surface water vapor pressure corresponds to the t-th type of land cover. This is the initial machine learning inversion model corresponding to the t-th land cover type; Let be the input feature vector corresponding to the t-th type of land cover; These are the model parameters.
[0049] Machine learning algorithms can employ gradient boosting decision trees, random forests, support vector regression, or neural network regression. This embodiment uses the gradient boosting decision tree algorithm. The gradient boosting decision tree algorithm sequentially builds multiple regression trees. Each regression tree built in each round is used to fit the residual between the training labels and the predicted surface water vapor pressure from the previous round, and the prediction results of each regression tree are accumulated according to the learning rate.
[0050] The gradient boosting decision tree model is represented as follows: (11) This represents the number of regression trees; The learning rate; This represents the prediction result of the q-th regression tree corresponding to the t-th land cover type; This includes the number of regression trees, maximum tree depth, learning rate, minimum number of samples per leaf node, subsampling ratio, and regularization parameters.
[0051] The target proportions for the training, validation, and independent test sets were set at 70%, 15%, and 15%, respectively. The sample sets for each land cover type were randomly partitioned, and adjustments were made based on the actual number of valid samples for each land cover type. Ultimately, the sample sizes for the farmland training, validation, and independent test sets were 3013, 623, and 709, respectively; for forest, 3903, 1532, and 1316; for grassland, 2210, 343, and 547; and for savanna and shrubland, 1325, 221, and 1227. The total sample size for all datasets was 16969.
[0052] The training set is used to fit the model parameters of the machine learning inversion model. The optimization objective of the model parameters is to minimize the error between the predicted surface water vapor pressure and the corresponding training label. The loss function for the t-th land cover type is expressed as: (12) Let be the loss function corresponding to the t-th type of land cover; The number of samples used in the current model training; Let be the input feature vector of the k-th training sample in the t-th land cover type; These are the corresponding training labels.
[0053] A grid search algorithm was used to optimize the number of regression trees, learning rate, maximum tree depth, and minimum number of samples per leaf node. Candidate values for the number of regression trees were 80, 110, 140, 170, and 200; for the learning rate, 0.01, 0.03, 0.05, and 0.10; for the maximum tree depth, 2, 3, 4, and 5; and for the minimum number of samples per leaf node, 5, 8, 10, 15, and 20.
[0054] Machine learning inversion models were trained using each candidate parameter combination, and the root mean square error (RMSE) of the corresponding models on the validation set was calculated. Using the minimum RMSE on the validation set as the parameter selection criterion, the final settings were determined as follows: 140 regression trees, a learning rate of 0.05, a maximum tree depth of 2, and a minimum number of samples per leaf node of 8.
[0055] Let the optimal model parameters be denoted as... The final machine learning inversion model corresponding to the t-th land cover type is represented as: (13) This is the final machine learning inversion model corresponding to the t-th land cover type; Let be the optimal model parameters for the t-th land cover type. For the input feature vector... The final predicted surface water vapor pressure calculated by the machine learning inversion model is: .
[0056] During training, when the value of the loss function is continuously If there is no significant improvement within a certain number of rounds, or if the number of training rounds reaches the preset maximum, training should be stopped. The number of rounds is a pre-defined positive integer.
[0057] After parameter optimization, each final machine learning inversion model is validated using an independent test set. Validation metrics include root mean square error and coefficient of determination.
[0058] Figure 2Scatter plot of farmland surface water vapor pressure inversion results provided in an embodiment of the present invention. Figure 2 The horizontal axis represents the effective value of surface water vapor pressure, and the vertical axis represents the predicted surface water vapor pressure. Both axes are in kPa, with a range of 0 to 3.5 kPa and a scale interval of 0.5 kPa. The dashed line extending upwards and to the right from the origin is a 1:1 reference line where the predicted surface water vapor pressure equals the effective value. Each scatter point in the figure represents a sample from an independent farmland test set, and the color gradient of the scatter points indicates the distribution and clustering areas of the sample points. Figure 2 The upper left corner indicates the number of independent test set samples (n=709) and the coefficient of determination. =0.940 and root mean square error RMSE=0.097 kPa.
[0059] Figure 3 Scatter plot of forest surface water vapor pressure inversion results provided in an embodiment of the present invention. Figure 3 The horizontal axis, vertical axis, coordinate range, coordinate scale, 1:1 reference line, and scatter plot representation are the same as those of the coordinate system. Figure 2 Consistent. Figure 3 The upper left corner indicates the number of independent test set samples (n=1316) and the coefficient of determination. =0.969 and root mean square error RMSE=0.060 kPa.
[0060] Figure 4 Scatter plot of grassland surface water vapor pressure inversion results provided in an embodiment of the present invention. Figure 4 The horizontal axis, vertical axis, coordinate range, coordinate scale, 1:1 reference line, and scatter plot representation are the same as those of the coordinate system. Figure 2 Consistent. Figure 4 The upper left corner indicates the number of independent test set samples (n=547) and the coefficient of determination. =0.877 and root mean square error RMSE=0.119 kPa.
[0061] Figure 5 This is a scatter plot of the surface water vapor pressure inversion results for sparse grasslands and shrublands provided in an embodiment of the present invention. Figure 5 The horizontal axis, vertical axis, coordinate range, coordinate scale, 1:1 reference line, and scatter plot representation are the same as those of the coordinate system. Figure 2 Consistent. Figure 5 The upper left corner indicates the number of independent test set samples (n=1227) and the coefficient of determination. =0.927 and root mean square error RMSE=0.099 kPa.
[0062] In step 107, based on the land cover type data that has undergone spatial resampling, the land cover type of each pixel in the unobserved area is determined pixel by pixel. The unobserved area refers to the area that lacks direct observation from global flux observation stations and requires the use of the final machine learning inversion model to calculate the surface water vapor pressure.
[0063] Pixels with land cover types of forest, farmland, grassland, savanna, and shrubland, and whose input data is valid, are identified as pixels to be inverted. Pixels representing oceans, inland waters, permanent snow and ice, and those with missing or substandard input data are not considered as pixels to be inverted.
[0064] In step 108, for each pixel to be inverted, the surface temperature for the corresponding date is obtained from remote sensing surface temperature data, and the atmospheric pressure, water vapor pressure deficit, and net radiation for the same date are obtained from reanalysis meteorological data and radiation data. Following the types and order of variables used in the model training phase, surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation are combined to form the pixel input feature vector.
[0065] The input variables used in the pixel input feature vector are consistent with the input variables used in the input feature vector of the training samples in step 105, so as to avoid inconsistencies in the types or order of variables during the model training stage and the global inference stage.
[0066] In step 109, based on the land cover type of the pixel to be inverted, the corresponding model is selected from the final machine learning inversion models corresponding to forest, farmland, grassland, savanna and shrubland respectively.
[0067] The final machine learning inversion model corresponding to forest is used for the inversion of forest pixels; the final machine learning inversion model corresponding to farmland pixels is used for the inversion of farmland pixels; the final machine learning inversion model corresponding to grassland pixels is used for the inversion of grassland pixels; and the final machine learning inversion model corresponding to savanna and shrub pixels is used for the inversion of savanna and shrub pixels.
[0068] The pixel input feature vector of the pixel to be inverted is input into the selected final machine learning inversion model to calculate the surface water vapor pressure of the pixel to be inverted on the corresponding date.
[0069] In step 110, surface water vapor pressure is collected according to the spatial location and corresponding date of each pixel to be inverted, forming the global surface water vapor pressure result: (14) This represents the global surface water vapor pressure results; Indicates spatial location ( , In time The corresponding surface water vapor pressure.
[0070] The spatial range of the global surface water vapor pressure results is 90°S to 90°N and 180°W to 180°E, with a spatial resolution of 0.025° and a temporal resolution of 1 day. The WGS84 coordinate system is used, and the coordinate reference system is identified as EPSG:4326. The unit of surface water vapor pressure is kPa.
[0071] Surface water vapor pressure inversion values are generated only for pixels in forests, farmland, grasslands, savannahs, and shrublands. Pixels representing oceans, inland waters, permanent snow and ice, and those with missing or substandard input data are uniformly assigned a value of -9999 and marked as NoData.
[0072] Figure 6 A grid diagram of the surface water vapor pressure inversion results provided in an embodiment of the present invention. Figure 6 The surface water vapor pressure results are illustrated using a 40×40 pixel local window and do not correspond to specific geographical regions or administrative divisions. The horizontal axis represents the grid column number, which is arranged from 1 to 40 from left to right; the vertical axis represents the grid row number, which is arranged from 1 to 40 from top to bottom; each grid corresponds to one spatial cell.
[0073] Figure 6 The right side displays a color scale for surface water vapor pressure, with units in kPa. Valid pixels are represented by colors ranging from dark purple, blue-green, green to yellow, according to their surface water vapor pressure values. Dark purple corresponds to lower surface water vapor pressure values within the map area, while yellow corresponds to higher values. Light gray grids represent water bodies, permanent ice and snow, and NoData pixels (those with missing or substandard input data). Global-scale results are generated using the same pixel-by-pixel calculation method, and the corresponding surface water vapor pressure is recorded according to spatial location and date.
[0074] This embodiment eliminates the aerodynamic impedance term by using the sensible heat flux expression, the aerodynamic impedance expression, and the surface energy balance relationship. It calculates the effective value of surface water vapor pressure using data from global flux observation stations. Sample sets are constructed according to surface cover type, and corresponding final machine learning inversion models are trained. Based on remote sensing surface temperature data and reanalysis meteorological data, pixel input feature vectors of the pixels to be inverted are constructed to complete the pixel-by-pixel inversion of surface water vapor pressure in unobserved areas.
[0075] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0076] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.
Claims
1. A global surface water vapor pressure retrieval method based on surface energy distribution constraints, characterized in that, include: Acquire global flux observation station data, remote sensing surface temperature data, land cover type data, and reanalysis meteorological data; Based on the sensible heat flux expression, the aerodynamic impedance expression, and the surface energy balance relationship, the aerodynamic impedance term is eliminated, and a physical constraint inversion model of surface water vapor pressure is constructed. Based on flux data and synchronous meteorological elements from the global flux observation station data, the effective value of surface water vapor pressure at the global flux observation station scale is obtained using the physical constraint inversion model of surface water vapor pressure. Based on the land cover type data, the samples composed of the effective value of surface water vapor pressure and the synchronous meteorological elements are classified to form multiple land cover type sample sets; Multiple input feature vectors are constructed for each of the aforementioned land cover type sample sets, and the effective value of the land surface water vapor pressure corresponding to each of the aforementioned input feature vectors is used as the training label; Establish machine learning inversion models corresponding to each land cover type, and train and optimize the machine learning inversion models using the input feature vectors and the corresponding training labels to obtain the final machine learning inversion models corresponding to each land cover type. Based on the land cover type data, determine the land cover type of each pixel to be inverted in the unobserved area; For each pixel to be inverted, a pixel input feature vector with the same input variables as the model training stage is constructed based on the remote sensing surface temperature data and the reanalysis meteorological data. The pixel input feature vector is input into the final machine learning inversion model corresponding to the land cover type of the pixel to be inverted, to obtain the land surface water vapor pressure of the pixel to be inverted; The surface water vapor pressure of each pixel to be inverted in the unobserved area is collected to obtain the global surface water vapor pressure result; Based on the expressions for sensible heat flux, aerodynamic impedance, and surface energy balance, the aerodynamic impedance term is eliminated, and a surface water vapor pressure physical constraint inversion model is constructed, including: The expression for the sensible heat flux is established based on the surface sensible heat exchange relationship: The aerodynamic impedance expression is established based on the aerodynamic impedance relationship and the surface energy balance relationship: By combining the sensible heat flux expression and the aerodynamic impedance expression, and eliminating the aerodynamic impedance term, the physical constraint inversion model of surface water vapor pressure is obtained: wherein, is the initial value of the surface water vapor pressure, is the actual water vapor pressure of the air, is the dry-bulb constant, is the surface temperature, is the air temperature, is the latent heat flux, is the sensible heat flux, is the air density, is the air constant-pressure specific heat, is the air dynamic impedance.
2. The global surface water vapor pressure inversion method based on surface energy distribution constraints according to claim 1, characterized in that, Based on flux data and synchronous meteorological elements from the global flux observation station data, and using the physical constraint inversion model for surface water vapor pressure, the effective values of surface water vapor pressure at the global flux observation station scale are obtained, including: The daily-scale flux data and corresponding daily-scale meteorological data in the global flux observation station data are screened for quality, and daily-scale records with unqualified quality labels, missing sensible heat flux, latent heat flux, air temperature, surface temperature or actual air vapor pressure, sensible heat flux less than or equal to zero, or inconsistent signs of the difference between surface temperature and air temperature and sensible heat flux are removed. According to the station number and date, the sensible heat flux, latent heat flux, air temperature, surface temperature and actual air vapor pressure of the same global flux observation station on the same day are matched accordingly; The sensible heat flux, latent heat flux, air temperature, surface temperature, and actual air vapor pressure are input into the physical constraint inversion model of surface vapor pressure to obtain the initial value of surface vapor pressure at the flux observation station. The initial values of surface water vapor pressure at the flux observation stations are subject to physical range constraints based on the saturated water vapor pressure corresponding to the surface temperature. The initial values of surface water vapor pressure at the flux observation stations that satisfy the physical range constraints are determined as the effective values of surface water vapor pressure at the global flux observation station scale.
3. The global surface water vapor pressure inversion method based on surface energy distribution constraints according to claim 2, characterized in that, No. The saturated vapor pressure corresponding to the surface temperature of each global flux observation station is calculated using the following formula: The physical range constraint is: In the formula, For the first Initial values of surface water vapor pressure at 10 global flux observation stations For the first Saturated vapor pressure at a global flux observation station For the first Surface temperature at a global flux observation station The surface temperature is an exponential function with a base of a natural constant. The unit of the surface temperature is degrees Celsius. The units of the initial value of the surface water vapor pressure and the saturated water vapor pressure are both kilopascals.
4. The global surface water vapor pressure inversion method based on surface energy distribution constraints according to claim 1, characterized in that, Based on the land cover type data, the samples composed of the effective values of surface water vapor pressure and the synchronous meteorological elements are classified to form multiple land cover type sample sets, including: The land cover type of each global flux observation station is determined based on the land cover type data. The effective value of surface water vapor pressure and the synchronous meteorological elements corresponding to each of the global flux observation stations are combined to form a sample; Based on the land cover type of each of the global flux observation stations, each of the samples is assigned to the corresponding land cover type sample set; The land cover types include forest, farmland, grassland, and savanna and shrubland.
5. The global surface water vapor pressure inversion method based on surface energy distribution constraints according to claim 1, characterized in that, Multiple input feature vectors are constructed for each of the aforementioned land cover type sample sets, and the effective value of the land surface water vapor pressure corresponding to each of the aforementioned input feature vectors is used as the training label, including: Surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation are extracted from the synchronous meteorological elements of each sample. An input feature vector is constructed based on the surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation corresponding to each sample. The effective value of surface water vapor pressure corresponding to each sample is used as the training label corresponding to the input feature vector.
6. The global surface water vapor pressure inversion method based on surface energy distribution constraints according to claim 1, characterized in that, Establish machine learning inversion models corresponding to each land cover type, train and optimize the machine learning inversion models using the input feature vectors and corresponding training labels, and obtain the final machine learning inversion models corresponding to each land cover type, including: The input feature vector and the corresponding training label in each land cover type sample set are divided into a training set, a validation set, and an independent test set. Regarding the first The machine learning inversion model is established based on the land cover type: To minimize the error between the predicted surface water vapor pressure and the training labels, a loss function is established: The model parameters of the machine learning inversion model are fitted using the training set, and the model parameters are optimized using the validation set to obtain the final machine learning inversion model; The final machine learning inversion model is validated using the independent test set, and the validation metrics include root mean square error and coefficient of determination. In the formula, For the first Predicted surface water vapor pressure corresponding to different land cover types For the first Machine learning inversion models corresponding to land cover types For the first The input feature vector corresponding to the land cover type. For the first Model parameters corresponding to land cover types, For the first The loss function corresponding to the land cover type, For the first The number of samples in the training set for each land cover type. For the first The first training set of land cover types The input feature vector of each sample, To and The corresponding training labels.
7. The global surface water vapor pressure inversion method based on surface energy distribution constraints according to claim 6, characterized in that, Regarding the first The process involves establishing a machine learning inversion model based on land cover type, fitting the model parameters of the machine learning inversion model using the training set, optimizing the model parameters using the validation set, and obtaining the final machine learning inversion model, including: The machine learning inversion model was determined to be a gradient boosting decision tree model. Multiple regression trees are built sequentially. The residual between the training label and the predicted surface water vapor pressure in the previous round is fitted using the regression tree built in each round. The prediction results of each regression tree are accumulated according to the learning rate. The number of regression trees, learning rate, maximum tree depth, and minimum number of leaf node samples are determined as the model parameters to be optimized. A grid search method is used to search for parameter combinations for each of the model parameters, and the minimum root mean square error of the validation set is used as the parameter selection criterion. The candidate values for the number of regression trees are 80, 110, 140, 170, and 200; the candidate values for the learning rate are 0.01, 0.03, 0.05, and 0.10; the candidate values for the maximum tree depth are 2, 3, 4, and 5; and the candidate values for the minimum number of leaf node samples are 5, 8, 10, 15, and 20. Based on the parameter selection conditions, the number of regression trees is determined to be 140, the learning rate is determined to be 0.05, the maximum tree depth is determined to be 2, and the minimum number of samples per leaf node is determined to be 8.
8. The global surface water vapor pressure inversion method based on surface energy distribution constraints according to claim 5, characterized in that, For each pixel to be inverted, based on the remotely sensed surface temperature data and the reanalysis meteorological data, a pixel input feature vector is constructed that uses the same input variables as in the model training phase, including: The surface temperature of the pixel to be inverted is obtained from the remote sensing surface temperature data; Based on the reanalysis meteorological data, the atmospheric pressure, water vapor pressure deficit, and net radiation of the pixel to be inverted are determined; The pixel input feature vector is constructed based on the surface temperature, atmospheric pressure, water vapor pressure deficit, and net radiation of the pixel to be inverted, and the pixel input feature vector uses the same input variables as the input feature vector in the model training phase.
9. The global surface water vapor pressure inversion method based on surface energy distribution constraints according to claim 8, characterized in that, The pixel input feature vector is input into the final machine learning inversion model corresponding to the land cover type of the pixel to be inverted, to obtain the land surface water vapor pressure of the pixel to be inverted, including: Based on the land cover type of the pixel to be inverted, select the final machine learning inversion model corresponding to the land cover type of the pixel to be inverted from the final machine learning inversion models corresponding to forest, farmland, grassland, savanna and shrubland respectively. The pixel input feature vector of the pixel to be inverted is input into the selected final machine learning inversion model, and the surface water vapor pressure of the pixel to be inverted is calculated using the selected final machine learning inversion model.
Citation Information
Patent Citations
Remote sensing-based earth surface energy balance component estimation method and system
CN111814317A
Surface water heat flux integrated estimation method, system and equipment
CN120633485A
