Regional soil moisture dynamic distribution prediction method and system

Abstract

The invention provides a regional soil moisture dynamic distribution prediction method, belongs to the field of soil moisture content prediction, and provides a sequential hybrid drive method (mRestart-EnKF-ML), the method combines a modified ensemble Kalman filter (mRestart-EnKF) and a machine learning model (ML), and the dynamic distribution of the regional soil moisture is predicted. And dynamically updating soil hydraulic parameters by assimilating data of a plurality of observation stations in real time, and generating a soil moisture profile data set with any temporal-spatial resolution. Through machine learning model training, the method can accurately predict the soil water content SMC at no observation point, and has high adaptability and expandability. Therefore, the technology provided by the invention has remarkable advantages in the aspects of improving SMC prediction precision, expanding temporal-spatial resolution, coping with complex terrains and meteorological conditions and the like, and an efficient and practical solution can be provided for dynamic monitoring and management of soil moisture.

Description

Technical Field

[0001] The invention belongs to the technical field of soil moisture prediction, and in particular relates to a method and system for predicting the dynamic distribution of regional soil moisture. Background Art

[0002] The spatiotemporal distribution of soil moisture content (SMC) plays a vital role in understanding the cycles and interactions of terrestrial water, energy, and carbon fluxes. Changes in SMC not only affect the surface evapotranspiration process, soil moisture infiltration and runoff, but are also directly related to the feedback mechanism of vegetation growth, agricultural production, and climate change. Therefore, long-term and accurate observation and budgeting of SMC at the regional scale is an important requirement in practical applications such as agriculture and water resources management. However, the existing SMC data observation methods are still significantly insufficient at the regional scale. Although the International Soil Moisture Network (ISMN) provides high-resolution soil moisture observation data through 3148 stations in 80 networks around the world, the spatial distribution of these data is relatively discrete, which makes it difficult to meet the needs of accurate soil moisture management at the field scale. On the other hand, although satellite remote sensing technology can provide spatially continuous SMC data, its limitations are also obvious. First, satellite data can only observe the moisture content of a few centimeters (about 5 cm) of the soil surface, and cannot effectively characterize the moisture conditions of deep soil. Second, the spatial resolution of satellites is low, usually about 25 kilometers, which makes it difficult to apply to finer scales. In addition, satellite observations are difficult to obtain accurate data under conditions such as complex terrain, dense vegetation, and cloud cover, especially in areas where the soil is frozen or covered with snow, where the problem of missing data is more serious.

[0003] In addition to direct observations, various soil moisture evolution models are also widely used to infer SMC in areas without observations. According to different modeling mechanisms, these models are generally divided into three categories. The first type of model calibrates the soil moisture diagnostic equation at sites with observation data, and then uses the similarity of hydroclimatic characteristics and soil properties to extrapolate it to sites without observations. Although this method can expand the application scope of observational data to a certain extent, its effect depends on the similarity between the calibration area and the extrapolation area, so it performs poorly in areas with strong heterogeneity. The second type of model is based on the physical processes of the vadose zone, mainly including hydrological models (such as MIKE SHE and SWAT) and the Richardson-Richards equation. This type of model infers soil moisture distribution by describing hydrological processes (such as infiltration, runoff, and evaporation). However, since these models require a large number of parameters that are difficult to obtain and are computationally expensive at the regional scale, their feasibility in large-scale applications is limited. To overcome this problem, data assimilation methods have emerged to combine observational data with physical models to improve the estimation accuracy of SMC. For example, surface soil moisture data from satellite observations can be integrated with land surface models to improve the SMC status of the surface and root zone; at the same time, point-scale observation data (such as SMC, soil head and groundwater level) can be used to assimilate physical models (such as Richards equations) to update soil hydraulic property parameters. However, although physical mechanism models have achieved some success, the widespread application of these models still faces many challenges due to the inherent nonlinearity of unsaturated flow and its heterogeneity. The third type of model uses machine learning algorithms to directly predict SMC by inputting a large amount of auxiliary data and meteorological variables. Machine learning models can derive highly nonlinear relationships between SMC and complex surface and atmospheric characteristics through a large amount of training data. Purely data-driven machine learning methods have quickly become popular in SMC prediction because they do not rely on governing equations or physical processes (such as energy balance) and causal relationships. However, due to the lack of physical constraints, these models often perform poorly when they are beyond the calibration range, especially in the prediction of extreme events, with low prediction accuracy and reliability. Therefore, how to combine physical models with machine learning methods and give full play to their respective advantages has become a key issue in the current field of soil moisture research.

[0004] In recent years, hybrid models that combine the advantages of physical models and data-driven methods have gradually attracted widespread attention. Among them, the Physics-Informed Neural Network (PINN) and its variants are the most common methods. PINN embeds the physical control equations into the loss function of the neural network, introduces physical consistency through the training process of the neural network, and ensures that the model prediction results conform to the physical laws. However, existing PINN studies are mostly based on virtual data, and there are still significant challenges when facing complex and noisy data in the real world. The performance is unstable, which makes it easy to introduce bias when predicting. Another common hybrid-driven method is to use the simulated data generated by the physical model as the training data for machine learning to incorporate some physical knowledge constraints. For example, Yang et al. proposed Kriging and CoKriging methods based on physical constraints to reconstruct sparsely observed steady-state heat transport problems. However, when there are parameter or structural uncertainties in the soil moisture model, the simulation results may introduce biases, thereby affecting the prediction performance of the machine learning model on a spatiotemporal scale.

[0005] In summary, although the current hybrid drive model (such as PINN) embeds the physical control equations into the neural network training process, it is mainly based on virtual data sets. When dealing with complex and real water and soil environments, it faces problems such as noise processing and model uncertainty, which makes it easy to introduce deviations in prediction. Summary of the invention

[0006] In order to solve the above problems, the present invention provides a method for predicting the dynamic distribution of regional soil moisture.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A method for predicting regional soil moisture dynamic distribution, comprising:

[0009] Obtain the real-time soil moisture content measured values ​​of multiple monitoring stations and the spatiotemporal information, meteorological factors and soil texture data of the corresponding monitoring stations;

[0010] The physical model of soil moisture movement at the monitoring site is constructed using the Richards equation. The spatiotemporal information, meteorological factors and soil texture data of each monitoring site are input into the physical model of soil moisture movement to generate multiple soil moisture content prediction values.

[0011] At time t, the modified ensemble Kalman filter mRestart-EnKF is used to assimilate the actual soil moisture content at the current moment. Specifically, the error characteristic analysis is used to weight multiple soil moisture content prediction values ​​and the actual soil moisture content at the current moment, and the optimal estimated value of the soil moisture content at the current moment is obtained. Multiple parameters of the soil moisture movement physical model are updated according to the optimal estimated value.

[0012] The updated parameter mean is returned to time 0, the soil moisture movement physical model is run to the current time t, and the soil moisture content prediction values ​​of multiple stations with arbitrary spatiotemporal resolution are used as the soil moisture content training data set; the spatiotemporal information, meteorological factors and soil texture data of the monitoring stations are used as input, and the soil moisture content training data set generated from time 0 to t is used as output to train the machine learning model;

[0013] The spatiotemporal information, meteorological factors and soil texture data of any unobserved site are input into the trained machine learning model to output the soil moisture content at any depth of the corresponding site.

[0014] Preferably, the spatiotemporal information of the monitoring site includes the time of data collection, as well as the longitude, latitude, altitude and vertical depth of the site, the environmental factors include daily precipitation and daily average temperature, and the soil texture data includes the proportion of sand, silt and clay.

[0015] Preferably, at time t, the modified ensemble Kalman filter mRestart-EnKF is used to assimilate the actual soil moisture content at the current moment, specifically by using error characteristic analysis to weight multiple soil moisture content prediction values ​​and the actual soil moisture content at the current moment, to obtain the optimal estimate of the soil moisture content at the current moment, and to update multiple parameters of the soil moisture movement physical model according to the optimal estimate, specifically including

[0016] At time step t, the observed soil moisture content and the corresponding true soil moisture content are expressed as and To express, the relationship between the two is described by the following equation:

[0017]

[0018] Where: The matrix H represents the observation operator connecting the state and the observation, and its dimension is N s ×N p ; ε t Represents the observation error vector at time t, with a mean of 0 and a covariance matrix of R t Gaussian distribution of

[0019] In the prediction phase of mRestart-EnKF, the updated soil moisture content at time t-1 The predicted soil moisture content at time t is obtained by calling the soil moisture movement physical model F

[0020]

[0021] Where: f and a represent the prediction step and analysis update step respectively;

[0022] In the update phase of mRestart-EnKF, if the observed soil moisture content at time t is By introducing the concept of Kalman gain, the weight distribution of the predicted soil moisture content and the actual measured soil moisture content at the current moment is performed to obtain the updated predicted soil moisture content:

[0023]

[0024] Where: K t represents the Kalman gain at time t;

[0025] After completing the observation of soil moisture content at the current time t After assimilation, mRestart-EnKF will update the parameters of the model The mean value of is returned to time 0, and the soil moisture movement physical model is re-run. The expression is as follows:

[0026]

[0027] Where: F 0→t Represents the simulation of the physical model of soil moisture movement from time 0 to time t; the superscript re means rerun; Indicates based on The mean is the mean soil moisture content obtained after running the soil moisture movement physical model again; s0 represents the initial soil moisture content at time 0.

[0028] Preferably, returning the updated parameter mean to time 0, running the soil moisture movement physical model to the current time t, and using the generated soil moisture content prediction values ​​of multiple stations at any temporal and spatial resolution as the soil moisture content training data set specifically includes:

[0029] According to the model parameters of each site and the soil moisture content s0 at time 0, re-run the soil moisture movement physical model from 0 to t; for any site q, generate the mean soil moisture content of all its historical time steps The mean soil moisture content of all sites constitutes the soil moisture content of all available sites. Construct a training data set.

[0030] Preferably, the machine learning model is trained by taking the spatiotemporal information, meteorological factors and soil texture data of the monitoring site as input and the soil moisture training data set generated from time 0 to time t as output, specifically:

[0031] Utilize soil moisture content at all available sites The corresponding input vector Train a machine learning model together to obtain a soil moisture machine learning model at the current time t; then predict the soil moisture content at any plane position and any depth based on the trained model;

[0032] For each site, based on the saved vector and the re-simulated Reconstructing the soil moisture content at the current moment Updated expansion vector It is used as the prior vector of the next time step (t+1) to continue running the soil moisture movement physical model forward;

[0033] The steps from obtaining the model parameters and soil moisture content to updating the expansion vector are repeated until all historical observation data points of all monitoring stations have been assimilated.

[0034] The present invention also provides a regional soil moisture dynamic distribution prediction system, comprising:

[0035] A data acquisition module is used to obtain the real-time measured values ​​of soil moisture content at multiple monitoring stations and the temporal and spatial information, meteorological factors and soil texture data of the corresponding monitoring stations;

[0036] The historical prediction module is used to construct a physical model of soil moisture movement at the monitoring site using the Richards equation, input the spatiotemporal information, meteorological factors and soil texture data of each monitoring site into the physical model of soil moisture movement, and generate multiple soil moisture content prediction values;

[0037] The data updating module is used to assimilate the actual soil moisture content at the current moment by using the modified ensemble Kalman filter mRestart-EnKF at the moment t. Specifically, the error characteristic analysis is used to weight the multiple predicted soil moisture content values ​​and the actual soil moisture content at the current moment, and the optimal estimated value of the soil moisture content at the current moment is obtained. The multiple parameters of the soil moisture movement physical model are updated according to the optimal estimated value.

[0038] The model training module is used to return the updated parameter mean to time 0, run the soil moisture movement physical model to the current time t, and use the generated soil moisture content prediction values ​​of multiple stations with arbitrary spatiotemporal resolution as the soil moisture content training data set; the machine learning model is trained with the spatiotemporal information, meteorological factors and soil texture data of the monitoring station as input and the soil moisture content training data set generated from time 0 to t as output;

[0039] The target prediction module is used to input the spatiotemporal information, meteorological factors and soil texture data of any unobserved site into the trained machine learning model and output the soil moisture content at any depth of the corresponding site.

[0040] The present invention also provides a computer device, including a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement any one of the steps in the method for predicting the dynamic distribution of regional soil moisture.

[0041] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is loaded by a processor, it can execute any one of the steps in the method for predicting the dynamic distribution of regional soil moisture.

[0042] The regional soil moisture dynamic distribution prediction method provided by the present invention has the following beneficial effects:

[0043] The present invention innovatively proposes a sequential hybrid drive method combining a modified ensemble Kalman filter (mRestart-EnKF) and a machine learning model, by obtaining the real-time soil moisture content measured values ​​of multiple monitoring sites and the spatiotemporal information, meteorological factors and soil texture data of the corresponding sites, and inputting the spatiotemporal information, meteorological factors and soil texture data of each monitoring site into the soil moisture movement physical model, generating multiple soil moisture content prediction values, and then auditing the multiple soil moisture content prediction values ​​and the current moment soil moisture content measured values ​​through the modified ensemble Kalman filter to update the parameters of the soil moisture movement physical model, and obtaining multiple soil moisture content prediction values ​​as soil moisture content training data sets through cyclic updating, and training the machine learning model. The method of the present invention can dynamically update the soil moisture movement physical model parameters by assimilating the data of multiple observation sites in real time, and generate high-resolution soil moisture content prediction values ​​as soil moisture content training data sets, rather than using virtual data sets. The original data comes from multiple monitoring sites, which can cover complex and real water and soil environments. The performance of the trained machine learning model is stable, and it can achieve accurate prediction of the spatiotemporal dynamics of soil moisture content (SMC). BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiment of the present invention and its design scheme, the following briefly introduces the drawings required for this embodiment. 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.

[0045] Figure 1 It is a flow chart of the method for predicting the dynamic distribution of regional soil moisture in Example 1 of the present invention;

[0046] Figure 2 This is the spatial location distribution map of the 14 sites in the XMS-CAT network;

[0047] Figure 3 It is a time series data graph of daily rainfall and multi-depth soil moisture content at 14 stations in the XMS-CAT network;

[0048] Figure 4 The proportion of sand, silt and clay at different depths in the 14 stations of the XMS-CAT network;

[0049] Figure 5 RMSE of the average soil moisture content inferred by the mRestart-EnKF-ML, OL-ML, and F-ML methods for the 14 stations in the XMS-CAT network. DETAILED DESCRIPTION

[0050] In order to enable those skilled in the art to better understand the technical solution of the present invention and implement it, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention.

[0051] The present invention mainly utilizes multi-site in-situ monitoring to accurately simulate and predict soil moisture profiles at regional scales such as irrigation areas and even watersheds. The mRestart-EnKF-ML proposed in the present invention utilizes a modified variant of the ensemble Kalman filter algorithm to assimilate multi-site soil moisture profile observations in real time, calibrate the soil moisture movement physical model in real time, and generate a multi-site data set, and then train the machine learning model through the multi-site data set, and finally predict the regional three-dimensional soil moisture dynamic distribution through the trained machine learning model. The vertical direction in the present invention refers to vertically downward from the surface of the earth, specifically refers to the physical model of soil moisture movement only simulates the movement process of soil moisture in the vertical direction, and does not consider lateral flow. The three-dimensional in the present invention refers to three directions and any spatial point, that is, the present invention is to obtain the soil moisture content at any point in three-dimensional space.

[0052] Next, the relevant contents involved in the hybrid drive algorithm mRestart-EnKF-ML proposed in the present invention are first described:

[0053] 1. Unsaturated flow control equation

[0054] The physical model in the present invention only considers the vertical one-dimensional soil moisture movement, and uses the Richards equation as the physical model of soil moisture movement. The Richards equation is:

[0055]

[0056] Wherein: θ represents the soil volume moisture content (referred to as soil moisture content in the present invention); h represents the pressure head; t represents time; z represents the vertical coordinate, with downward being positive; and K represents the unsaturated hydraulic conductivity.

[0057] There are three unknowns in the above formula: θ, h and K. The constitutive relationship of these three is often characterized by the van Genuchten-Mualem model:

[0058]

[0059] Where: θ s and θ r represent saturated and residual moisture contents respectively; α and n represent the shape parameters of soil moisture characteristic curve respectively; K s Indicates saturated hydraulic conductivity; S e Indicates effective saturation.

[0060] The present invention adopts the Ross method, which is a non-iterative numerical scheme, for the mRestart-EnKF part in the proposed hybrid method to quickly solve the one-dimensional Richards equation of equation (1).

[0061] 2. Modified Ensemble Kalman Filter Variant (mRestart-EnKF)

[0062] The traditional Kalman filter EnKF cannot ensure the physical consistency in the nonlinear system after updating, which means that the updated soil moisture and soil hydraulic parameters may no longer conform to the Richards equation. This inconsistency may cause a significant decrease in data assimilation performance when dealing with strong nonlinear soil moisture problems. In the present invention, a modified ensemble Kalman filter variant mRestart-EnKF is used to solve this inconsistency problem while maintaining a low computational cost. As the first step of the mRestart-EnKF-ML method, mRestart-EnKF has two main purposes in the present invention: first, assimilate the real-time observation data of soil moisture content at the current moment to achieve the joint update of parameters and soil moisture content; second, return the updated parameter mean to the initial moment, and re-run the forward soil moisture movement physical model (i.e., Richards equation), so as to obtain simulated "historical" data of arbitrary spatiotemporal resolution, providing a rich data set for subsequent machine learning training.

[0063] In the data assimilation of soil water physics models, the unknown soil hydraulic parameters and the state variables of interest are generally estimated jointly. Therefore, similar to the traditional EnKF, mRestart-EnKF first expands the soil moisture content s into the model parameter p to obtain the augmented vector y = [p, s] T Assume that the number of unknown parameters and state variables are N s and N p , N p is the number of soil moisture contents to be estimated at a single site. If you want to estimate the soil moisture content at three depths of 0.05cm\0.10cm\0.20cm, then N p =3. Then the vectors s and p are of dimension N s and N p A one-dimensional vector of; if the size of the sample set is N, then the augmented vector set can be expressed as:

[0064] Y={y1,y2,…,y N} (6)

[0065] At time step t, the measured soil moisture content and the corresponding true value are expressed as and To express it, the relationship between the two can be described by the following equation:

[0066]

[0067] Where: The matrix H represents the observation operator connecting the state and the observation, and its dimension is N s ×N p ; ε tRepresents the observation error vector at time t, with a mean of 0 and a covariance matrix of R t Gaussian distribution.

[0068] In the prediction phase of mRestart-EnKF, the updated soil moisture content at time t-1 The predicted soil moisture content at time t is obtained by calling the soil moisture movement physical model F (here refers to the Richards equation)

[0069]

[0070] Where: f and a represent the prediction step and analysis (update) step respectively.

[0071] In the update phase of mRestart-EnKF, if there is a soil moisture observation at time t By introducing the concept of Kalman gain, the model prediction and the soil moisture observation value are weighted and the updated soil moisture prediction value is obtained:

[0072]

[0073] Where: K t represents the Kalman gain at time t, which can be defined as:

[0074]

[0075] Where: The predicted covariance matrix representing the soil moisture content at time t can be approximated by sample statistics:

[0076]

[0077] In the formula: <*> represents the mean of *.

[0078] Different from the traditional EnKF, after completing the observation of soil moisture content at the current time t After assimilation, mRestart-EnKF will update the parameters of the model The mean value of is returned to time 0, and the soil moisture movement physical model is re-run. The expression is as follows:

[0079]

[0080] Where: F 0→t Represents the simulation of the physical model of soil moisture movement from time 0 to time t; the superscript re means rerun; Indicates based on Mean is the mean soil moisture content obtained after running the model again; s0 represents the initial soil moisture content at time 0.

[0081] Then the current soil moisture sample set can be reconstructed through the following two steps:

[0082] Step 1: Calculate and save the soil moisture content before returning to the run The error distribution

[0083]

[0084] Where: Represents the i-th soil moisture sample before the parameter sample mean is returned; N represents the number of samples.

[0085] Step 2: Based on the mean value of the soil moisture sample at the current moment obtained after re-running The error distribution saved by formula (12) reconstructs the posterior soil moisture content at the current moment

[0086] At the next time step (t+1), the calculated and Substitute into equation (8), and repeat the similar process until the simulation ends. It is worth noting that in addition to providing the prior vector for the next time step through the above equation, mRestart-EnKF in the present invention is also responsible for generating the “historical” data set S required for machine learning modeling at the current time step. t :

[0087]

[0088] 3. Build data-driven models based on machine learning algorithms

[0089] In the mRestart-EnKF step of the mRestart-EnKF-ML method, N s The assimilation steps of the vertical one-dimensional physical model and soil moisture profile observations of each station are independent and synchronous. Therefore, in fact, N s Set of simulated data sets:

[0090]

[0091] Where: Indicates N from 1 to t s A collection of “historical” data from sites as the output of subsequent machine learning models. (p=1,2,…,k; q=1,2,…,N s ) has a dimension of Ng (i.e., the number of nodes specified in the vertical soil profile), the total number of training data at time step k is N s ×k×N g .and The corresponding machine learning model input is:

[0092]

[0093] Where: x p,q (p=1,2,…,t; q=1,2,…,N s ) represents the input vector of the qth station at the pth time step; Represents all N from 1 to t s In the soil water problem of the present invention, x includes not only the main environmental factors related to soil water movement and the characteristic information of the soil itself, but also the location information of the corresponding site, such as longitude, latitude and altitude.

[0094] In addition, in the hybrid drive algorithm mRestart-EnKF-ML proposed in the present invention, the soil moisture movement physical model will be rebuilt sequentially at each time step as the real-time soil moisture content measured data is continuously collected. and After all are prepared, the machine learning model can be trained. The machine learning methods used in the present invention include convolutional neural network (CNN), random forest (RF) and support vector machine (SVM). Considering that these three machine learning methods are relatively mature, the relevant technical details are not repeated here.

[0095] Example 1

[0096] Based on this, the present invention provides a method for predicting the dynamic distribution of regional soil moisture. Traditional soil moisture prediction models often estimate the soil moisture content at any position (x, y) on the surface, and do not pay attention to the soil moisture content at different vertical depths z at the station (x, y). The method provided in the embodiment of the present invention can predict the moisture content at any depth z at any station (x, y), that is, the soil moisture content at (x, y, z). Therefore, the embodiment of the present invention predicts three-dimensional soil moisture.

[0097] Specific as Figure 1 As shown, the method comprises the following steps:

[0098] Step 1: At any time step t, from N s The real-time observation data (i.e., the measured values ​​of soil moisture content) of each monitoring station are independently and synchronously assimilated to achieve the joint update of the physical model parameters and soil moisture content of each station, so that the following equations (8)-(11) are obtained: Then, the parameters of the model The ensemble mean and soil moisture content The aggregate sample volatility It can be statistically obtained and saved through formulas (12) and (13).

[0099] Step 2: Based on each site’s and s0, and re-run the soil moisture movement physical model from 0 to t through equation (12). For any station q, the mean soil moisture content of all its historical time steps can be generated, that is,

[0100] Step 3: Soil moisture content at all available sites The corresponding input vector are used together to train the machine learning model to obtain the soil moisture machine learning model at the current time t. The soil moisture profile at any plane position can then be predicted based on the trained model. It is worth noting that the soil moisture content s re The dimension (i.e., the number of nodes in the vertical profile of soil moisture content) can be manually set according to the research objectives and machine learning modeling requirements of the present invention.

[0101] Step 4: For each site, based on the saved vector and the re-simulated The soil moisture content at the current moment is reconstructed by formula (14): Updated expansion vector It is used as the prior vector of the next time step (t+1) to continue running the soil moisture movement physical model forward.

[0102] Step 5: Repeat steps 1 to 4 until all historical observation data points of all monitoring stations have been assimilated.

[0103] The regional soil moisture dynamic distribution prediction method provided by the present invention has the following advantages:

[0104] 1. Sequential hybrid drive method (mRestart-EnKF-ML): This paper innovatively proposes a sequential hybrid drive method that combines the modified ensemble Kalman filter (mRestart-EnKF) with a machine learning model to achieve accurate prediction of the spatiotemporal dynamics of soil moisture content (SMC). This method can dynamically update soil hydraulic parameters by assimilating data from multiple observation sites in real time, generate high-resolution soil moisture profile data sets and soil moisture (soil moisture content) profiles of arbitrary spatiotemporal resolution, and provide a reliable solution for SMC prediction without observation points.

[0105] 2. Dynamic update of soil hydraulic parameters: The mRestart-EnKF method can update the physical model parameters through real-time assimilation within the time step to ensure physical consistency. At the same time, it effectively solves the performance degradation problem of the traditional ensemble Kalman filter method EnKF method when dealing with nonlinear problems. The addition of machine learning models enhances the accuracy and robustness of predictions.

[0106] 3. Improved computational efficiency and scalability: Compared with the traditional three-dimensional soil moisture physical model, this invention combines the scalability of one-dimensional and three-dimensional models, utilizes the dynamic update of the one-dimensional Richards model, and sequentially deduce the three-dimensional soil moisture dynamics, avoiding the high computational cost and difficulty in convergence of the traditional three-dimensional model. At the same time, it sequentially reconstructs the SMC dynamics, supports the integration of real-time data and the update of soil hydraulic parameters, adapts to the needs of online data collection and real-time prediction, and greatly improves computational efficiency. This scalable solution meets the needs of soil moisture management at regional and even global scales.

[0107] 4. Enhanced adaptability and robustness: The present invention organically combines physical models with machine learning, which not only ensures the constraints of the physical model and avoids the problem of lack of physical consistency in pure machine learning methods, but also improves the model's predictive adaptability and robustness in complex terrain and extreme meteorological conditions through the nonlinear mapping capability of machine learning, and enhances the model's predictive ability for complex areas and extreme situations.

[0108] 5. Real-time data integration and update capability: The mRestart-EnKF-ML method supports the collection and integration of real-time data, can dynamically update soil hydraulic parameters and reconstruct SMC dynamics, meet the needs of real-time prediction and management, and adapt to modern online data collection capabilities.

[0109] 6. Strong scalability: Through machine learning model training, this method can accurately predict SMC without observation points and has high adaptability and scalability.

[0110] 5. Ability to adapt to various terrain and climatic conditions: This method has the ability to handle soil moisture prediction in complex terrain, densely vegetated areas and extreme climatic conditions. It performs well in dealing with uncertainty and heterogeneity, significantly improving the reliability and accuracy of the prediction results.

[0111] Therefore, the method proposed in the present invention has significant advantages in improving SMC prediction accuracy, expanding spatiotemporal resolution, and coping with complex terrain and meteorological conditions, and is expected to provide an efficient and practical solution for soil moisture dynamic monitoring and management at regional and even global scales.

[0112] Next, experiments are conducted to verify the advantages of the method for predicting the dynamic distribution of regional soil moisture proposed in the present invention.

[0113] 1. Example verification data source

[0114] The validation data for this example come from the Catalan soil physical parameters monitoring station network (XMS-CAT), which was launched in 2015 by the Institute of Geography and Cartography of Catalonia (ICGC). This paper selected 14 monitoring stations ( Figure 2 ), which continuously recorded soil moisture content (SWC) and environmental parameters such as temperature and precipitation from January 1, 2021 to March 10, 2022 ( Figure 3 ) to verify the feasibility of the proposed mRestart-EnKF-ML method. These sites are located in the Mediterranean climate zone, covering an area of ​​10,565 km2 (41.80° to 42.59°N, 0.83° to 2.29°E), with altitudes ranging from 238 to 1238 m. Based on the soil texture information provided by ISMN, Figure 4 The proportions of sand, silt, and clay at different depths at each station are also summarized in the paper.

[0115] 2. Example verification model parameter setting

[0116] To describe the dynamics of vertical soil moisture movement, each site is represented as a one-dimensional soil column in this example. The height of these soil columns is 1.00 m and they are discretized into 53 uniformly sized grids. The upper boundary of all sites is set to the atmospheric boundary, while the lower boundary condition is determined based on long-term deep soil moisture content (SWC) data: sites BCLL, CDLP_2, PSN, RDS, SCA, CMN, and CDPA use a constant water content lower boundary condition, while sites CLA, EBX, LCA, LLA, LCS, BDS, and EML use a free drainage lower boundary condition. Based on the soil texture classification map, the initial soil parameters (including θ r ,θ s , α, n and K s ). Considering the large number of sites and their soil stratification, the specific soil hydraulic parameter values ​​are not shown here. In addition, α, n and K s The unknown parameters are assumed to be log-normally distributed, and the variance is set manually based on experience. Then, N = 50 initial samples are generated from the prior distribution. These parameter sets are gradually updated as the real-time soil moisture observation data is assimilated. In addition, the variance of the soil moisture measurement error is set to 0.02 2 .

[0117] In order to describe the spatiotemporal distribution of soil water content (SWC) in the study area, three machine learning models (including CNN, RF and SVM) were constructed and trained. The input x of the model contains three categories of information: spatiotemporal information (time, longitude, latitude, altitude and vertical depth), environmental factors (daily precipitation and daily average temperature) and soil texture data (proportion of sand, silt and clay), and the output is the SWC at the corresponding depth of each station. The CNN model first configures an input layer that matches the size of the training data, followed by a convolutional layer with a kernel size of [3, 1], 16 filters, and "samepadding" to keep the input size unchanged. The ReLU activation layer introduces nonlinearity and finally outputs through a single neuron fully connected layer. The regression layer is used to calculate the loss of the regression task. The training options use the Adam optimizer, the maximum number of iterations is 50, the mini-batch size is 16, and the learning rate adopts a piecewise descent strategy, and the learning rate is reduced by 0.05 times after every 10 iterations. For the RF model, the TreeBagger function is used for training, and the number of trees is set to 100. TreeBagger uses bootstrap sampling with replacement, considering one-third of the predictors at each split, and the minimum leaf node size is set to 1, which means that each leaf node contains at least one observation. In addition, the SVM model uses a radial basis function (RBF) kernel, automatic kernel function scaling, and a box constraint value of 1 to control the weight distribution of training and test errors. The ∈ value is set to 0.1 to define the training loss sensitivity of the ∈-tube, and the SVM solver is Sequential Minimal Optimization (SMO).

[0118] The "leave one out cross-validation" procedure was used for validation in all site test cases. Specifically, data from 13 sites were used in the mRestart-EnKF data assimilation and training of the machine learning model, while the remaining 1 site was used for model validation. This operation was repeated 14 times to ensure that each site had an opportunity to be used as an independent validation set. In addition, in order to verify the importance of Kalman update in the hybrid algorithm mRestart-EnKF-ML, an open-loop version (OL-ML) that does not consider the real-time update of physical model parameters was also introduced as a comparison. In OL-ML, multiple sites run the soil moisture movement physical model based on initial parameters and initial soil moisture content, and the generated historical data is directly used to train the machine learning model and make predictions. At the same time, in order to verify the effectiveness of the physical constraints provided by the physical model, the present invention also designs a model (F-ML) based entirely on machine learning as a comparison. At a given time point t, the F-ML model uses all historical measured data from multiple sites up to time step t to construct a pure data-driven model and predict the SMC at any depth at any unobserved site.

[0119] 3. Example verification results

[0120] like Figure 5 The RMSE results of the average soil moisture content inferred by each station using mRestart-EnKF-ML, OL-ML and F-ML methods (here ML uses three machine learning models: CNN, RF and SVM) are shown. It can be found that compared with the OL-ML and F-ML methods, although the soil properties and climatic conditions of different stations are different, mRestart-EnKF-ML can effectively reduce the RMSE of soil moisture estimation at unobserved stations, that is, significantly improve the estimation accuracy of soil moisture profiles. In addition, it can be found that although the ML methods are different, their advantages remain robust after coupling into mRestart-EnKF-ML.

[0121] Based on the same inventive concept, the present invention also proposes a regional soil moisture dynamic distribution prediction system, comprising:

[0122] A data acquisition module is used to obtain the real-time measured values ​​of soil moisture content at multiple monitoring stations and the temporal and spatial information, meteorological factors and soil texture data of the corresponding monitoring stations;

[0123] The historical prediction module is used to construct a physical model of soil moisture movement at the monitoring site using the Richards equation, input the spatiotemporal information, meteorological factors and soil texture data of each monitoring site into the physical model of soil moisture movement, and generate multiple soil moisture content prediction values;

[0124] The data updating module is used to assimilate the actual soil moisture content at the current moment by using the modified ensemble Kalman filter mRestart-EnKF at the moment t. Specifically, the error characteristic analysis is used to weight the multiple predicted soil moisture content values ​​and the actual soil moisture content at the current moment, and the optimal estimated value of the soil moisture content at the current moment is obtained. The multiple parameters of the soil moisture movement physical model are updated according to the optimal estimated value.

[0125] The model training module is used to return the updated parameter mean to time 0, run the soil moisture movement physical model to the current time t, and use the generated soil moisture content prediction values ​​of multiple stations with arbitrary spatiotemporal resolution as the soil moisture content training data set; the machine learning model is trained with the spatiotemporal information, meteorological factors and soil texture data of the monitoring station as input and the soil moisture content training data set generated from time 0 to t as output;

[0126] The target prediction module is used to input the spatiotemporal information, meteorological factors and soil texture data of any unobserved site into the trained machine learning model, and output the three-dimensional soil moisture content at any depth of the corresponding site.

[0127] Each module in the above regional soil moisture dynamic distribution prediction system can be implemented in whole or in part by software, hardware and their combination. Each module can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each module above.

[0128] The present invention also provides a computer device, including a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps in the method embodiment for predicting the dynamic distribution of regional soil moisture. The specific implementation method can be found in the method embodiment, which will not be described in detail here.

[0129] Furthermore, the present invention also provides a non-temporary computer-readable storage medium containing instructions, and a computer program is stored on the storage medium. For example, a memory containing instructions, the above instructions can be executed by a processor of a computer device to complete the above method. For example, a non-temporary computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device. When the computer program is executed by the processor, the steps in the embodiment of the method for predicting the dynamic distribution of regional soil moisture can be implemented. The specific implementation method can be found in the method embodiment, which will not be repeated here.

[0130] It will be appreciated by those skilled in the art that embodiments of the present invention may provide methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0131] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as a combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0132] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0133] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0134] It should be pointed out that the specific implementation methods described above can enable those skilled in the art to understand the invention more comprehensively, but do not limit the invention in any way. Therefore, although the invention has been described in detail in this specification and embodiments, those skilled in the art should understand that the invention can still be modified or replaced by equivalents; and all technical solutions and improvements that do not deviate from the spirit and scope of the invention are included in the protection scope of the patent for the invention. Any figure mark in the claims should not be regarded as limiting the claims involved. Any simple change or equivalent replacement of the technical solution that can be obviously obtained by any technician familiar with the field within the technical scope disclosed in the present invention belongs to the protection scope of the present invention.

Claims

1. A method for predicting the dynamic distribution of regional soil moisture, characterized in that: include: Obtain the real-time soil moisture content measured values ​​of multiple monitoring stations and the spatiotemporal information, meteorological factors and soil texture data of the corresponding monitoring stations; The physical model of soil moisture movement at the monitoring site is constructed using the Richards equation. The spatiotemporal information, meteorological factors and soil texture data of each monitoring site are input into the physical model of soil moisture movement to generate multiple soil moisture content prediction values. At time t, the modified ensemble Kalman filter mRestart-EnKF is used to assimilate the actual soil moisture content at the current moment. Specifically, the error characteristic analysis is used to weight multiple soil moisture content prediction values ​​and the actual soil moisture content at the current moment, and the optimal estimated value of the soil moisture content at the current moment is obtained. Multiple parameters of the soil moisture movement physical model are updated according to the optimal estimated value. Return the updated parameter mean to time 0, run the soil moisture movement physical model to the current time t, and use the generated soil moisture content prediction values ​​of multiple stations at any temporal and spatial resolution as the soil moisture content training data set; The machine learning model is trained with the spatiotemporal information, meteorological factors and soil texture data of the monitoring site as input and the soil moisture training dataset generated from time 0 to t as output; The spatiotemporal information, meteorological factors and soil texture data of any unobserved site are input into the trained machine learning model to output the soil moisture content at any depth of the corresponding site.

2. The method for predicting the dynamic distribution of regional soil moisture according to claim 1, characterized in that: The spatiotemporal information of the monitoring site includes the time of data collection, as well as the longitude, latitude, altitude and vertical depth of the site; the environmental factors include daily precipitation and daily average temperature; and the soil texture data include the proportions of sand, silt and clay.

3. The method for predicting the dynamic distribution of regional soil moisture according to claim 3, characterized in that: At time t, the modified ensemble Kalman filter mRestart-EnKF is used to assimilate the actual soil moisture content at the current moment. Specifically, the error characteristic analysis is used to weight multiple soil moisture content prediction values ​​and the actual soil moisture content at the current moment to obtain the optimal estimated value of the soil moisture content at the current moment, and multiple parameters of the soil moisture movement physical model are updated according to the optimal estimated value, specifically including At time step t, the observed soil moisture content and the corresponding true soil moisture content are expressed as and To express, the relationship between the two is described by the following equation: Where: The matrix H represents the observation operator connecting the state and the observation, and its dimension is N s ×N p ; ε t Represents the observation error vector at time t, with a mean of 0 and a covariance matrix of R t Gaussian distribution of In the prediction phase of mRestart-EnKF, the updated soil moisture content at time t-1 The predicted soil moisture content at time t is obtained by calling the soil moisture movement physical model F Where: f and a represent the prediction step and analysis update step respectively; In the update phase of mRestart-EnKF, if the observed soil moisture content at time t is By introducing the concept of Kalman gain, the weight distribution between the predicted value of soil moisture content and the actual measured value of soil moisture content at the current moment is performed to obtain the updated predicted value of soil moisture content: Where: K t represents the Kalman gain at time t; After completing the observation of soil moisture content at the current time t After assimilation, mRestart-EnKF will update the parameters of the model The mean value of is returned to time 0, and the soil moisture movement physical model is re-run. The expression is as follows: Where: F 0→t Represents the simulation of the physical model of soil moisture movement from time 0 to time t; the superscript re means rerun; Indicates based on The mean is the mean soil moisture content obtained after running the soil moisture movement physical model again; s0 represents the initial soil moisture content at time 0.

4. The method for predicting the dynamic distribution of regional soil moisture according to claim 3, characterized in that: The updated parameter mean is returned to time 0, the soil moisture movement physical model is run to the current time t, and the generated soil moisture content prediction values ​​of multiple stations at any temporal and spatial resolution are used as the soil moisture content training data set, specifically including: According to the model parameters of each site and the soil moisture content s0 at time 0, re-run the soil moisture movement physical model from 0 to t; for any site q, generate the mean soil moisture content of all its historical time steps The mean soil moisture content of all sites constitutes the soil moisture content of all available sites. Construct a training data set.

5. The method for predicting the dynamic distribution of regional soil moisture according to claim 4, characterized in that: The machine learning model is trained by taking the spatiotemporal information, meteorological factors and soil texture data of the monitoring site as input and the soil moisture training data set generated from time 0 to time t as output, specifically: Utilize soil moisture content at all available sites The corresponding input vector Train a machine learning model together to obtain a soil moisture machine learning model at the current time t; then predict the soil moisture content at any plane position and any depth based on the trained model; For each site, based on the saved vector and the re-simulated Reconstructing the soil moisture content at the current moment Updated expansion vector It is used as the prior vector of the next time step (t+1) to continue running the soil moisture movement physical model forward; The steps from obtaining the model parameters and soil moisture content to updating the expansion vector are repeated until all historical observation data points of all monitoring stations have been assimilated.

6. A regional soil moisture dynamic distribution prediction system, characterized in that: include: A data acquisition module is used to obtain the real-time measured values ​​of soil moisture content at multiple monitoring stations and the temporal and spatial information, meteorological factors and soil texture data of the corresponding stations; The historical prediction module is used to construct a physical model of soil moisture movement at the monitoring site using the Richards equation, input the spatiotemporal information, meteorological factors and soil texture data of each monitoring site into the physical model of soil moisture movement, and generate multiple soil moisture content prediction values; The data updating module is used to assimilate the actual soil moisture content at the current moment by using the modified ensemble Kalman filter mRestart-EnKF at the moment t. Specifically, the error characteristic analysis is used to weight the multiple predicted soil moisture content values ​​and the actual soil moisture content at the current moment, and the optimal estimated value of the soil moisture content at the current moment is obtained. The multiple parameters of the soil moisture movement physical model are updated according to the optimal estimated value. The model training module is used to return the updated parameter mean to time 0, run the soil moisture movement physical model to the current time t, and use the generated soil moisture content prediction values ​​of multiple stations with arbitrary spatiotemporal resolution as the soil moisture content training data set; The machine learning model is trained with the spatiotemporal information, meteorological factors and soil texture data of the monitoring site as input and the soil moisture training dataset generated from time 0 to t as output; The target prediction module is used to input the spatiotemporal information, meteorological factors and soil texture data of any unobserved site into the trained machine learning model and output the soil moisture content at any depth of the corresponding site.

7. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is loaded into a processor, it can execute the steps of the method according to any one of claims 1 to 5.

Images