Regional soil moisture content monitoring system optimization design method and system
By training the Gaussian process (GPR) model and quantifying data value using relative entropy, the soil moisture monitoring scheme is optimized, which solves the problems of insufficient applicability of soil moisture movement models in real environments and high data collection costs, and achieves high-precision and low-cost soil moisture monitoring.
Patent Information
- Application Number
- CN202510861756.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-25
- Publication Date
- 2025-09-23
AI Technical Summary
Existing soil moisture movement models are not applicable enough in real heterogeneous environments, and the description of biophysical processes is incomplete, resulting in low soil moisture movement prediction accuracy. In addition, the unplanned collection of massive data leads to high monitoring costs and deteriorated accuracy.
Based on historical data from multiple soil moisture monitoring stations, a Gaussian process (GPR) model was trained to generate prediction sample sets for alternative monitoring schemes. The data value was quantified using relative entropy, and the monitoring scheme was optimized to reduce costs and improve accuracy.
Without reducing the accuracy of soil moisture prediction, the monitoring cost can be reduced by optimizing the monitoring plan, so as to achieve the accuracy and economy of regional soil moisture monitoring.
Smart Images
Figure CN120688364A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of soil moisture monitoring, and in particular to an optimization design method and system for a regional soil moisture monitoring system. Background Art
[0002] The spatiotemporal distribution of soil moisture holds significant scientific significance and application value in the study of terrestrial water-energy-carbon cycle interactions. In agricultural production, this parameter is not only a key control factor for optimizing irrigation schedules, implementing precision fertilization, and increasing crop yields, but also a core indicator for building modern agricultural water resource management systems. Currently, physical-mechanism-based soil water movement models (such as HYDRUS, FEFLOW, and MIKE SHE) mathematically describe hydrological processes such as rainfall infiltration, evapotranspiration, and root water uptake by coupling mass conservation equations with Darcy's law. However, existing theoretical frameworks still have significant limitations in describing soil water movement. First, the applicability of underlying theoretical assumptions (such as the simplified soil water potential model of the Richards equation and the laminar flow assumption of Darcy's law) to real heterogeneous environments remains questionable. Second, the mathematical representation of key biophysical processes (such as the dynamics of root hydraulic structure and its feedback loop with soil moisture) remains incomplete. This incomplete description of physical mechanisms makes it difficult for current soil water physics models to accurately predict soil water movement in real environments. With the development of the Internet of Things (IoT) and the improvement of remote sensing observation capabilities, data-driven approaches based on machine learning have provided a new research paradigm for simulating soil moisture dynamics. Current research focuses on constructing meteorological factor-soil moisture response models (at the ground-based site scale) or joint backscatter coefficient-vegetation index inversion models (at the remote sensing observation scale), using algorithms such as Gaussian process regression and deep neural networks to establish nonlinear mapping relationships. Numerous studies have demonstrated that these purely data-driven approaches can be used to measure spatial and temporal variations in soil moisture, with accuracy comparable to that of methods based on physical models.
[0003] From traditional soil water physical models to today's purely data-driven soil water models, our reliance on data has intensified with the advent of the big data era. However, it's worth noting that the value of different data often varies. Unplanned collection and integration of massive amounts of data will not only dramatically increase monitoring costs but also introduce additional unknown errors into the soil water simulation system, thereby deteriorating the accuracy of soil water simulation. Therefore, given economic and time constraints, it is necessary to pre-assess the potential value of data before collection. Once the value of future observations is quantified, the most beneficial monitoring plans for soil water simulation can be selected and collected. This reduces monitoring costs while minimizing the introduction of excessive uncertainty, allowing for the accurate construction of purely data-driven soil water models and precise simulation and prediction of soil water movement.
[0004] When designing a future soil moisture monitoring system, data is unknown, so a priori analysis is necessary. The basic idea is to initially construct a soil moisture model based on existing prior knowledge (or prior data), generate corresponding virtual observation samples, and then perform Bayesian analysis on each virtual sample to obtain utility function values. Ultimately, the average of all utility function values is defined as the expected data value of the monitoring scheme. Previous researchers have typically used the information gain of the data as a utility function, such as Shannon entropy difference, relative entropy, and signal degrees of freedom. This classic framework is known as the Bayesian data value assessment framework. However, previous studies on the value of soil moisture data have been based on physical models, typically focusing only on the potential value of future multi-source observation systems in identifying physical processes and their parameters. The unplanned and costly collection of massive amounts of data has led to high monitoring costs for existing monitoring schemes. Summary of the Invention
[0005] The purpose of the present invention is to provide a method and system for optimizing the design of a regional soil moisture monitoring system in order to solve the problems in the prior art.
[0006] The present invention specifically provides the following technical solutions: A method for optimizing the design of a regional soil moisture monitoring system, comprising: The historical data of multiple soil moisture monitoring stations are used as input, and the soil moisture data are used as output to train the Gaussian process (GPR) model; the historical data include spatiotemporal information, environmental factors, and soil geological data of the soil moisture monitoring stations; The trained Gaussian process (GPR) model is used to generate a prediction sample set for the i-th alternative soil moisture monitoring scheme and a probability distribution for predicting the soil moisture conditions in the entire region. A new Gaussian process (GPR) model is trained jointly with the prediction sample set and the historical data from multiple monitoring stations. The i-th alternative soil moisture monitoring scheme is the soil moisture data observed at the i-th soil moisture monitoring station for several days in the future. The soil moisture conditions in the entire region are the soil moisture contents of all plots within the monitoring range of the multiple soil moisture monitoring stations. The new Gaussian process (GPR) model was used to predict the probability distribution of soil moisture in the entire region again. The relative entropy was used to quantify the difference in the predicted probability distribution of soil moisture in the entire region before and after the joint prediction sample set. Multiple relative entropies of the i-th alternative soil moisture monitoring scheme were obtained, and the data value of the i-th alternative soil moisture monitoring scheme was obtained by averaging the multiple relative entropies. Repeat the steps of training a new Gaussian process GPR model and obtaining data value until the relative entropy values of all future alternative soil moisture monitoring schemes are obtained. By comparing their sizes, the alternative soil moisture monitoring scheme with the largest relative entropy value is selected as the optimal observation scheme.
[0007] Preferably, the method of using historical data from multiple soil moisture monitoring stations as input and soil moisture data as output to train a Gaussian process (GPR) model includes: The mean function and covariance function of soil moisture data are defined by hyperparameters. When the historical data of multiple soil moisture monitoring stations meet the Gaussian distribution, the hyperparameter probability distribution is obtained by the log-likelihood function of the observation vector under given observation values. Based on the probability distribution of the hyperparameters, the hyperparameters of the mean function and covariance function are calculated using partial derivatives to obtain the hyperparameters that maximize the log marginal likelihood in the hyperparameter probability distribution. The final mean function and covariance function are obtained through the hyperparameters with the maximum log marginal likelihood. The trained Gaussian process GPR model is obtained through the final mean function and covariance function.
[0008] Preferably, the method of generating a prediction sample set of the i-th alternative soil moisture monitoring scheme using the trained Gaussian process GPR model includes: When the target variable is the probability distribution of soil moisture in the entire region at a specified spatial resolution, the entire region is divided into N g grids, and obtain N g Data input ; Based on input The trained Gaussian process GPR model is used to predict the mean soil moisture in the entire region. and variance , and use the mean value of soil moisture in the whole region and variance Generate a prediction sample set for the i-th alternative soil moisture monitoring scheme that satisfies the Gaussian distribution.
[0009] Preferably, the joint training of a new Gaussian process GPR model based on historical data of multiple monitoring sites and a prediction sample set includes: Using historical data from multiple monitoring stations and the generated j The Gaussian process GPR model is trained again with the predicted samples to obtain the expanded input and output. The specific expressions are: ; ; Where: Indicates the i A soil moisture monitoring program based on the predicted mean and variance The generated j soil moisture samples; Represents The corresponding Gaussian process GPR model input, j = 1, 2, ..., N r , N r Indicates the maximum number of prediction samples; use and As input and output, the Gaussian process GPR model is trained to obtain a new Gaussian process Model.
[0010] Preferably, the difference in the predicted probability distribution of the soil moisture in the entire region before and after the joint prediction sample set is quantified using relative entropy to obtain multiple relative entropies of the i-th alternative soil moisture monitoring scheme, and the data value of the i-th alternative soil moisture monitoring scheme is obtained by averaging the multiple relative entropies, specifically: The jth sample of the i-th alternative soil moisture monitoring scheme is used for joint training. Model, will Substitute new Model, get the new mean value of soil moisture in the whole region and variance ; Under the premise that both the prior and posterior probability density functions satisfy the n-dimensional Gaussian distribution, the new mean of the soil moisture in the entire region is obtained. and variance Obtain the model fitting error and prior log likelihood, and define the relative entropy of the jth sample of the i-th alternative soil moisture monitoring scheme by the sum of the model fitting error and the prior log likelihood Based on the i-th alternative soil moisture monitoring scheme N r samples, and the relative entropy of the samples Take the average as the relative entropy of the i-th alternative soil moisture monitoring scheme .
[0011] The present invention provides a regional soil moisture monitoring system optimization design system, comprising: An initial training module is used to train a Gaussian process (GPR) model using historical data from multiple soil moisture monitoring stations as input and soil moisture data as output; the historical data includes spatiotemporal information, environmental factors, and soil geological data of the soil moisture monitoring stations; a joint training module for generating a prediction sample set for an i-th alternative soil moisture monitoring scheme and a probability distribution of the soil moisture conditions in the entire region using the trained Gaussian process (GPR) model, and jointly training a new Gaussian process (GPR) model based on the historical data of multiple monitoring stations and the prediction sample set; the i-th alternative soil moisture monitoring scheme is the soil moisture data observed at the i-th soil moisture monitoring station for several days in the future; the soil moisture conditions in the entire region are the soil moisture contents of all plots within the monitoring range of the multiple soil moisture monitoring stations; The value calculation module is used to use the new Gaussian process GPR model to predict the probability distribution of soil moisture in the entire region again, and use relative entropy to quantify the difference in the predicted probability distribution of soil moisture in the entire region before and after the joint prediction sample set, to obtain multiple relative entropies of the i-th alternative soil moisture monitoring scheme, and to average the multiple relative entropies to obtain the data value of the i-th alternative soil moisture monitoring scheme; The observation scheme acquisition module is used to repeatedly train the new Gaussian process GPR model and obtain the data value steps until the relative entropy values of all future alternative soil moisture monitoring schemes are obtained. By comparing the sizes, the alternative soil moisture monitoring scheme with the largest relative entropy value is selected as the optimal observation scheme.
[0012] The present invention provides a computer device, comprising a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of the above-mentioned method for optimizing the design of a regional soil moisture monitoring system.
[0013] The present invention provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method for optimizing the design of a regional soil moisture monitoring system are realized.
[0014] Compared with the prior art, the present invention has the following significant advantages: The present invention trains a Gaussian process GPR model based on existing site data, generates a prediction sample set of alternative soil moisture monitoring schemes, and predicts the probability distribution of soil moisture in the entire region. The soil moisture monitoring scheme can be pre-optimized before soil moisture data is collected, and the Gaussian process GPR model is jointly trained based on the existing available site data and the generated prediction sample set. The probability distribution of soil moisture in the entire region is predicted again using multiple new Gaussian process GPR models, and the relative entropy and data value are obtained by the difference between the previous and subsequent probability distributions. The data value analysis framework is organically combined with the machine learning model, providing methodological guidance for the optimal design of the regional soil moisture monitoring system. The monitoring cost can be reduced as much as possible without reducing the accuracy of soil moisture prediction, while taking into account the accuracy and economy of regional soil moisture monitoring. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 This is a technical flow chart of an optimization design method for a regional soil moisture monitoring system in the present invention; Figure 2 The site distribution map provided by the present invention; Figure 3 Soil texture (percentages of clay, sand, and silt) at the sites provided for this invention; Figure 4 The altitude of the site provided by the present invention; Figure 5 The present invention provides a graph of daily rainfall and daily average temperature for a station; wherein, Figure 5 (a) is the diagram of the BCLL site, Figure 5 (b) is the graph of the CLA site. Figure 5 (c) is the diagram of the EBX site. Figure 5 (d) is the map of the LCA site, Figure 5 (e) is the map of the LLA site, Figure 5 (f) is a map of the LCS site; Figure 6 The present invention provides two graphs of daily rainfall and daily average temperature at the site; wherein, Figure 6 (a) is the diagram of the PSN site. Figure 6 (b) is the diagram of the RDS site. Figure 6 (c) is the graph of the SCA site, Figure 6 (d) is the map of the BDS site, Figure 6 (e) is a diagram of the CDPA site; Figure 7 Soil moisture content of the site provided by the present invention; Figure 8 The average relative entropy of the soil moisture monitoring program for the site from day 101 to day 110 provided by the present invention. DETAILED DESCRIPTION
[0016] The following is a clear and complete description of the technical solutions of the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.
[0017] This paper proposes a method for optimizing the design of a regional soil moisture monitoring system (GPR-BDW) based on information theory and machine learning algorithms. First, a Gaussian process regression model for regional soil moisture is constructed based on a historical observation dataset from agricultural production (including spatiotemporal information of sampling points, environmental factors, soil texture, and corresponding soil moisture data). Second, a Monte Carlo method is used to generate a prediction sample dataset for alternative soil moisture monitoring scenarios. The relative entropy information gain of each alternative soil moisture monitoring scenario is calculated through Bayesian posterior analysis. Finally, a decision-making model using relative entropy as a quantitative indicator is established to select the optimal monitoring scenario that balances economic efficiency and prediction accuracy. The core innovation of this method lies in the organic integration of a Bayesian data value analysis framework with a machine learning model. This not only develops a theoretical framework for data-driven models but also provides methodological guidance for the optimized design of regional soil moisture monitoring systems, which has important practical significance for advancing the development of digital agricultural infrastructure.
[0018] In response to the high cost of multi-source monitoring in existing pure data-driven soil water models, the purpose of the present invention is to propose an optimization design method for regional soil moisture monitoring systems (GPR-BDW) based on information theory and machine learning algorithms. Compared with the unplanned and high-cost collection of massive data in traditional pure data-driven soil water modeling, the technical framework proposed in this invention can pre-evaluate the data value of future alternative soil moisture monitoring schemes based on information indicators (relative entropy) before data collection, and then optimize the design of future soil moisture monitoring schemes by only collecting observations with higher data value, thereby achieving the goal of significantly reducing monitoring costs.
[0019] The GPR-BDW method of the present invention is based on input data such as spatiotemporal information, environmental factors, soil texture, and the corresponding output data of soil moisture content from multiple sites. It uses the Gaussian process regression (GPR) method to construct a purely data-driven soil water model, and couples the model to the Bayesian data value assessment framework (BDW). It introduces the relative entropy indicator to quantitatively assess the data value, and then reduces the monitoring cost by selecting the monitoring plan with the greatest data value. The corresponding processing method flow is as follows: Figure 1 shown.
[0020] like Figure 1 As shown, in this embodiment, a method for optimizing the design of a regional soil moisture monitoring system includes the following steps: Step S1: Take historical data of multiple soil moisture monitoring stations as input and soil moisture data as output to train a Gaussian process (GPR) model; the historical data includes spatiotemporal information, environmental factors, and soil geological data of the soil moisture monitoring stations.
[0021] In the field of agricultural production, obtaining N sThe spatial and temporal information, environmental factors, soil texture and other historical data of each monitoring station are used as the input X of the GPR model. The corresponding historical soil moisture data are obtained simultaneously as the output y of the GPR model. The GPR model is trained based on (X, y).
[0022] The input X and output y of the GPR model are expressed as: (1); (2); Where: x represents the input vector of the GPR model, including spatiotemporal information, environmental factors, soil texture, etc.; N s represents the number of stations with available historical data; t represents the length of time for which historical data are available, in days; y represents the output of the GPR model, i.e., soil moisture; y represents the vector of all available historical soil moisture data.
[0023] The mean function and covariance function of soil moisture data are defined by hyperparameters, and when the historical data of multiple soil moisture monitoring stations meet the Gaussian distribution, the hyperparameter probability distribution is obtained by the log-likelihood function of the observation vector under given observation values. Specifically: the GPR model is completely composed of the mean function m(x) and the covariance function To define these two functions, the hyperparameters was introduced. Among them, is the linear coefficient vector; Controls the marginal variance of the model output; Then it represents the correlation length between each dimension of x. In the present invention, the linear mean function and the isotropic square exponential covariance function are specified:
[0024] (3); (4); Based on the probability distribution of hyperparameters, the hyperparameters of the mean function and covariance function are calculated using partial derivatives to obtain the hyperparameters that maximize the logarithmic marginal likelihood in the hyperparameter probability distribution, and the final mean function and covariance function are obtained through the hyperparameters with the largest logarithmic marginal likelihood. Specifically: Considering that the prior knowledge of hyperparameters is very limited, it is necessary to use the current training data to calculate the hyperparameters. Hyperparameters Assume that the training data satisfies the Gaussian distribution, and the probability distribution of the hyperparameters can be obtained. The specific expression is:
[0025] (5); Where: Represents the prior mean of y; covariance matrix It is calculated based on the covariance function specified by formula (4), and its i-th row and j-th column element is .
[0026] Then, the hyperparameters are derived based on the easily derived partial derivatives. The specific expression is: (6); (7); Where: and They represent the hyperparameters of the mean function and covariance function, namely and .
[0027] Based on equations (6) and (7), we seek the hyperparameters that can maximize the logarithmic marginal likelihood in equation (5). The final mean function and covariance function are obtained through the hyperparameters with the largest logarithmic marginal likelihood. The trained Gaussian process GPR model is obtained through the final mean function and covariance function, and then the trained soil water GPR model is obtained.
[0028] Step S2: Using the trained Gaussian process (GPR) model, a prediction sample set for the i-th alternative soil moisture monitoring scenario and a probability distribution for predicting the soil moisture content for the entire region are generated. A new Gaussian process (GPR) model is trained based on the historical data from multiple monitoring stations and the prediction sample set. The i-th alternative soil moisture monitoring scenario is the soil moisture content data observed at the i-th soil moisture monitoring station for several days in the future. The regional soil moisture content is the soil moisture content of all plots within the monitoring range of the multiple soil moisture monitoring stations.
[0029] Generate a sample set of predictions for alternative soil moisture monitoring scenarios. Use the trained GPR model to predict the soil moisture probability distribution for each of the N alternative future soil moisture monitoring scenarios, obtaining the predicted mean and variance. Simultaneously, use the trained GPR model to predict the probability distribution of soil moisture for the entire region, calculating the mean and variance.
[0030] The GPR model trained in step S1 can be used to predict the soil moisture probability distribution corresponding to any alternative soil moisture monitoring scheme. Here, we take the i-th (i=1, 2, …, N) soil moisture monitoring scheme as an example, assuming that its corresponding input is , then the corresponding output The mean and variance They are:
[0031] (8); (9); Where: express The prior mean of ; express and The covariance matrix of ; express The prior covariance matrix of .
[0032] At the same time, assuming that the target variable is the probability distribution of soil moisture in the entire region at a specified spatial resolution, the entire region is first divided into N g grid, then the corresponding input for: (10); Similarly, based on the input , enter replace Substituting into formula (8-9), the mean soil moisture of the entire region can be predicted. and variance , referred to as and , and use the mean value of soil moisture in the whole region and variance Generate a prediction sample set of alternative soil moisture monitoring schemes that satisfies Gaussian distribution.
[0033] Taking the i-th (i=1,2,…,N) alternative soil moisture monitoring scheme as an example, the mean and variance of the soil moisture content generated in step S2 are used to generate N Gaussian distributions. r prediction samples; the available site data in step S1 and the prediction samples of the i-th alternative soil moisture monitoring scheme generated in step S2 are used as training sets at the same time, and a new GPR model is obtained by joint training to predict the probability distribution of soil moisture in the entire region again.
[0034] Using equations (8-9) in step S2, the predicted mean value of the i-th (i=1, 2, …, N) soil moisture monitoring scheme can be obtained: and variance Based on this mean and variance, the monitoring solution is generated to meet the Gaussian distribution of N r samples; the available site data is the historical data of multiple monitoring sites. Using the available site data in step S1 and the generated j-th (j=1, 2, …, N r ) prediction samples to train the GPR model again to obtain the expanded input and output. The specific expression is:
[0035] (11); (12); Where: Indicates the i-th (i=1, 2, …, N) soil moisture monitoring scheme based on the predicted mean and variance The jth (j=1, 2, …, Nr) soil moisture sample generated; Represents The corresponding Gaussian process GPR model input, j = 1, 2, ..., N r , N r Indicates the maximum number of prediction samples.
[0036] use and Replace X and y respectively, substitute into equations (3) to (7), train the Gaussian process GPR model, and obtain the new Gaussian process Model. Substitute new Model, using equations (8) and (9) to obtain the new mean value of soil moisture in the entire region and variance Considering that there are N alternative soil moisture monitoring schemes, each alternative soil moisture monitoring scheme generates N r prediction samples, then N*N are trained in this step. r GPR model and predict N*N r The mean soil moisture of the whole region and variance .
[0037] Step S3: Use the new Gaussian process GPR model to predict the probability distribution of soil moisture in the entire region again, and use relative entropy to quantify the difference in the predicted probability distribution of soil moisture in the entire region before and after the joint prediction sample set, and obtain multiple relative entropies of the i-th alternative soil moisture monitoring scheme. Take the average of multiple relative entropies to obtain the data value of the i-th alternative soil moisture monitoring scheme.
[0038] Relative entropy is introduced as a quantitative indicator of data value. The relative entropy is used to quantify the difference in the probability distribution of soil moisture conditions across the entire region generated in steps S2 and S3, respectively. The data value of the i-th (i=1, 2, …, N) alternative soil moisture monitoring scheme is calculated (using relative entropy as an indicator).
[0039] In step S2, only the available station data can be trained to predict the mean soil moisture of the entire region. and variance , respectively referred to as and In step S2, the existing data and the prediction samples of the alternative soil moisture monitoring scheme are jointly trained to obtain N*N r The mean soil moisture of the whole region and variance For the sake of convenience, the present invention refers to the mean and variance of the soil moisture in the whole region obtained by joint training of the jth sample of the i-th alternative soil moisture monitoring scheme as and As a comprehensive indicator, relative entropy (RE) provides an indicator to measure the information content of the (pre) posterior probability density function (pdf) relative to the background or prior pdf. Under the premise that both the prior and posterior pdfs satisfy the n-dimensional Gaussian distribution, the mean of the new soil moisture in the whole region is obtained. and variance Obtain the model fitting error and prior log likelihood, and define the relative entropy of the jth sample of the i-th alternative soil moisture monitoring scheme by the sum of the model fitting error and the prior log likelihood ; The specific expression is:
[0040] (13); (14); (15); Where: represents the relative entropy value corresponding to the jth sample of the i-th alternative soil moisture monitoring scheme; det(*) represents the determinant of the matrix *; Tr (*) represents the trace of the matrix *, is the model fitting error, is the prior log-likelihood.
[0041] Based on the i-th alternative soil moisture monitoring scheme N r samples, and the relative entropy of the samples Take the average as the relative entropy of the i-th alternative soil moisture monitoring scheme , the specific expression is: (16); Step S4: Repeat the steps of training a new Gaussian process GPR model and obtaining data value until the relative entropy values of all future alternative soil moisture monitoring schemes are obtained. By comparing the sizes, the alternative soil moisture monitoring scheme with the largest relative entropy value is selected as the optimal observation scheme.
[0042] The following uses the soil water content (SWC) of 11 stations in the XMS-CAT network of the International Soil Moisture Network (ISMN, https: / / ismn.earth / en / ) from December 1, 2021 to March 20, 2022, as well as spatiotemporal information (time, longitude, latitude, altitude), environmental factors (daily rainfall, daily average temperature, such as Figure 5 and Figure 6 As shown, red represents temperature and blue represents rainfall), soil texture (percentage of clay, sand and silt) data, such as Figure 2 As shown, the 11 sites are named BCLL, CLA, EBX, LCA, LLA, LCS, PSN, RDS, SCA, BDS and CDPA, and the soil texture is as follows Figure 3 As shown, the site elevation is Figure 4 As shown, the soil moisture content of the site is as follows Figure 7 As shown, combined with the process Figure 1 , the overall processing flow of the regional soil moisture monitoring system optimized by the present invention is introduced and verified.
[0043] Step 1: Based on N s = Historical data of 11 stations from December 1, 2021 to March 10, 2022 (t = 100 days), and use equations (1) to (7) to train the Gaussian process model GPR, where: x Including time, longitude, latitude, altitude, daily rainfall, daily average temperature, clay ratio, sand ratio and silt ratio, output y is the corresponding soil water content, with a total of N s *t=11*100=1100 sets of training data; the mean function uses the linear mean function, and the covariance function uses the isotropic square exponential covariance function.
[0044] Step 2: Assume that only one of the 11 stations can be observed from the 101st to the 110th day. Now we need to select the station with the greatest data value for observation. That is, the number of alternative soil moisture monitoring schemes N=11. The i-th (i=1, 2, …, N) alternative soil moisture monitoring scheme is to observe the soil moisture time series of the i-th station from the 101st to the 110th day. Assume that the ultimate goal is to predict the soil moisture distribution of the entire region on the 110th day. Use the GPR model trained in step 1 to predict the soil moisture probability distribution corresponding to the i-th alternative soil moisture monitoring scheme. The corresponding input is , using formula (8-9), the soil moisture prediction corresponding to the i-th alternative soil moisture monitoring scheme can be obtained: The mean and variance At the same time, the GPR model trained in step 1 is used to predict the soil moisture distribution on the 110th day. The corresponding input is , x* includes time, longitude, latitude, altitude, rainfall, daily rainfall, daily average temperature, clay ratio, sand ratio and silt ratio. Similarly, using formula (8) and formula (9), the soil moisture distribution of the whole area on the 110th day can be predicted: The mean and variance .
[0045] Step 3: Taking the i-th alternative soil moisture monitoring scheme as an example, use the soil moisture content generated in step 2 The mean and variance , generate N that satisfies Gaussian distribution r = 1000 prediction samples: ; For the jth (j=1, 2, …, N r ) samples, and jointly train them with the historical data of Ns=11 stations t=100 days in step 1 to obtain a new Model, whose input and output They are: ; ; Using the new training The model predicts the soil moisture distribution on the 110th day again, and the corresponding input is still , and obtain the mean soil moisture of the entire region and variance Considering that there are N alternative soil moisture monitoring schemes, each alternative soil moisture monitoring scheme generates N r prediction samples, then N*N are trained in this step. r =11*1000=11000 GPR models, and predict N*N r The mean soil moisture of the whole region and variance .
[0046] Step 4: Train only N in step 2 s = 100 days of historical data from 11 stations can be used to predict the average soil moisture conditions in the entire region and variance , respectively referred to as and In step 3, the joint training of existing data and alternative soil moisture monitoring scheme prediction samples can predict N*N r The mean soil moisture of the whole region and variance For the sake of convenience, the present invention refers to the mean and variance of the soil moisture in the whole region obtained by joint training of the jth sample of the i-th alternative soil moisture monitoring scheme as and For the i-th alternative soil moisture monitoring scheme, Nr relative entropy values can be calculated using equations (13) to (15), and then the average relative entropy value of the alternative soil moisture monitoring scheme can be calculated using equation (16).
[0047] Step 5: Repeat steps 3 and 4 until the relative entropy values of all future N=11 soil moisture scenarios are obtained. Figure 8 As shown, the gray dots represent the relative entropy of Nr=1000 samples, and the red dots represent the average relative entropy. By comparing the average relative entropy, it can be found that the data value of the EBX site is the greatest, so monitoring the soil moisture content of the EBX site from day 101 to day 110 is confirmed as the optimal observation plan.
[0048] The present invention proposes a regional soil moisture monitoring system optimization design system, which includes: an initial training module, a joint training module, a value calculation module and an observation scheme acquisition module.
[0049] Among them, the initial training module is used to take the historical data of multiple soil moisture monitoring stations as input and the soil moisture data as output to train the Gaussian process GPR model; the historical data includes the spatiotemporal information, environmental factors and soil geological data of the soil moisture monitoring stations; the joint training module is used to use the trained Gaussian process GPR model to generate the prediction sample set of the i-th alternative soil moisture monitoring scheme, as well as the probability distribution of the predicted soil moisture in the entire region, and jointly train a new Gaussian process GPR model based on the historical data of multiple monitoring stations and the prediction sample set; the i-th alternative soil moisture monitoring scheme is to observe the soil moisture data of the i-th soil moisture monitoring station in the next few days; the soil moisture in the entire region is multiple soil The soil moisture content of all plots within the monitoring range of the soil moisture monitoring station; the value calculation module is used to use the new Gaussian process GPR model to re-predict the probability distribution of soil moisture in the entire region, and use the relative entropy to quantify the difference in the predicted probability distribution of soil moisture in the entire region before and after the joint prediction sample set, to obtain multiple relative entropies of the i-th alternative soil moisture monitoring scheme, and average the multiple relative entropies to obtain the data value of the i-th alternative soil moisture monitoring scheme; the observation scheme acquisition module is used to repeatedly train the new Gaussian process GPR model and obtain the data value steps until the relative entropy values of all future alternative soil moisture monitoring schemes are obtained. By comparing the sizes, the soil moisture monitoring scheme with the largest relative entropy value is selected as the optimal observation scheme.
[0050] The present invention also provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor executes the steps of a method for optimizing the design of a regional soil moisture monitoring system.
[0051] According to the disclosed embodiments, a computing device may communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth communications, etc.), or with any device that enables a computing device to communicate with one or more other computing devices (e.g., routers, modems, etc.).
[0052] The present invention provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method for optimizing the design of a regional soil moisture monitoring system are realized.
[0053] The above content is a further detailed description of the present invention in combination with specific preferred embodiments. For those skilled in the art to which the present invention belongs, several simple deductions or replacements can be made without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.
Claims
1. A method for optimizing the design of a regional soil moisture monitoring system, characterized in that: include: The historical data of multiple soil moisture monitoring stations are used as input, and the soil moisture data are used as output to train the Gaussian process (GPR) model; the historical data include spatiotemporal information, environmental factors, and soil geological data of the soil moisture monitoring stations; The trained Gaussian process (GPR) model is used to generate a prediction sample set for the i-th alternative soil moisture monitoring scheme and a probability distribution for predicting the soil moisture conditions in the entire region. A new Gaussian process (GPR) model is trained jointly with the prediction sample set and the historical data from multiple monitoring stations. The i-th alternative soil moisture monitoring scheme is the soil moisture data observed at the i-th soil moisture monitoring station for several days in the future. The soil moisture conditions in the entire region are the soil moisture contents of all plots within the monitoring range of the multiple soil moisture monitoring stations. The new Gaussian process (GPR) model was used to predict the probability distribution of soil moisture in the entire region again. The relative entropy was used to quantify the difference in the predicted probability distribution of soil moisture in the entire region before and after the joint prediction sample set. Multiple relative entropies of the i-th alternative soil moisture monitoring scheme were obtained, and the data value of the i-th alternative soil moisture monitoring scheme was obtained by averaging the multiple relative entropies. Repeat the steps of training a new Gaussian process GPR model and obtaining data value until the relative entropy values of all future alternative soil moisture monitoring schemes are obtained. By comparing their sizes, the alternative soil moisture monitoring scheme with the largest relative entropy value is selected as the optimal observation scheme.
2. The method for optimizing the design of a regional soil moisture monitoring system according to claim 1, wherein: The method uses historical data from multiple soil moisture monitoring stations as input and soil moisture data as output to train a Gaussian process (GPR) model, including: The mean function and covariance function of soil moisture data are defined by hyperparameters. When the historical data of multiple soil moisture monitoring stations meet the Gaussian distribution, the hyperparameter probability distribution is obtained by the log-likelihood function of the observation vector under given observation values. Based on the probability distribution of the hyperparameters, the hyperparameters of the mean function and covariance function are calculated using partial derivatives to obtain the hyperparameters that maximize the log marginal likelihood in the hyperparameter probability distribution. The final mean function and covariance function are obtained through the hyperparameters with the maximum log marginal likelihood. The trained Gaussian process GPR model is obtained through the final mean function and covariance function.
3. The method for optimizing the design of a regional soil moisture monitoring system according to claim 1, wherein: The method of using the trained Gaussian process GPR model to generate a prediction sample set for the i-th alternative soil moisture monitoring scheme includes: When the target variable is the probability distribution of soil moisture in the entire region at a specified spatial resolution, the entire region is divided into N g grids, and obtain N g Data input ; Based on input The trained Gaussian process GPR model is used to predict the mean soil moisture in the entire region. and variance , and use the mean value of soil moisture in the whole region and variance Generate a prediction sample set for the i-th alternative soil moisture monitoring scheme that satisfies the Gaussian distribution.
4. The method for optimizing the design of a regional soil moisture monitoring system according to claim 3, wherein: The new Gaussian process GPR model is trained based on the historical data of multiple monitoring sites and the prediction sample set, including: Using historical data from multiple monitoring stations and the generated j The Gaussian process GPR model is trained again with the predicted samples to obtain the expanded input and output. The specific expressions are: ； ； Where: Indicates the i A soil moisture monitoring program based on the predicted mean and variance The generated j soil moisture samples; Represents The corresponding Gaussian process GPR model input, j = 1, 2, ..., N r , N r Indicates the number of prediction samples; use and As input and output, the Gaussian process GPR model is trained to obtain a new Gaussian process Model.
5. The method for optimizing the design of a regional soil moisture monitoring system according to claim 1, wherein: The relative entropy is used to quantify the difference in the predicted probability distribution of the soil moisture in the entire region before and after the joint prediction sample set, and multiple relative entropies of the i-th alternative soil moisture monitoring scheme are obtained. The data value of the i-th alternative soil moisture monitoring scheme is obtained by averaging the multiple relative entropies, which is specifically: The jth sample of the i-th alternative soil moisture monitoring scheme is used for joint training. Model, will Substitute new Model, get the new mean value of soil moisture in the whole region and variance ; Under the premise that both the prior and posterior probability density functions satisfy the n-dimensional Gaussian distribution, the new mean of the soil moisture in the entire region is obtained. and variance Obtain the model fitting error and prior log likelihood, and define the relative entropy of the jth sample of the i-th alternative soil moisture monitoring scheme by the sum of the model fitting error and the prior log likelihood ; Based on the i-th alternative soil moisture monitoring scheme N r samples, and the relative entropy of the samples Take the average as the relative entropy of the i-th alternative soil moisture monitoring scheme .
6. A regional soil moisture monitoring system optimization design system, characterized in that: include: An initial training module is used to train a Gaussian process (GPR) model using historical data from multiple soil moisture monitoring stations as input and soil moisture data as output; the historical data includes spatiotemporal information, environmental factors, and soil geological data of the soil moisture monitoring stations; a joint training module for generating a prediction sample set for an i-th alternative soil moisture monitoring scheme and a probability distribution of the soil moisture conditions in the entire region using the trained Gaussian process (GPR) model, and jointly training a new Gaussian process (GPR) model based on the historical data of multiple monitoring stations and the prediction sample set; the i-th alternative soil moisture monitoring scheme is the soil moisture data observed at the i-th soil moisture monitoring station for several days in the future; the soil moisture conditions in the entire region are the soil moisture contents of all plots within the monitoring range of the multiple soil moisture monitoring stations; The value calculation module is used to use the new Gaussian process GPR model to predict the probability distribution of soil moisture in the entire region again, and use relative entropy to quantify the difference in the predicted probability distribution of soil moisture in the entire region before and after the joint prediction sample set, to obtain multiple relative entropies of the i-th alternative soil moisture monitoring scheme, and to average the multiple relative entropies to obtain the data value of the i-th alternative soil moisture monitoring scheme; The observation scheme acquisition module is used to repeatedly train the new Gaussian process GPR model and obtain the data value steps until the relative entropy values of all future alternative soil moisture monitoring schemes are obtained. By comparing the sizes, the alternative soil moisture monitoring scheme with the largest relative entropy value is selected as the optimal observation scheme.
7. A computer device, characterized in that: It includes a memory and a processor, wherein a program is stored in the memory, and when the program is executed by the processor, the processor executes the steps of the method for optimizing the design of a regional soil moisture monitoring system as described in any one of claims 1 to 5.
8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for optimizing the design of a regional soil moisture monitoring system according to any one of claims 1 to 5 are implemented.