Dry farmland crop precision irrigation control method and system based on digital twinning

Abstract

This invention discloses a method and system for precision irrigation control of dryland crops based on digital twins, relating to the field of crop irrigation technology. It includes: loading multi-source sensing data; obtaining the posterior distribution of soil hydraulic parameters through Bayesian inference; performing probability field modeling on the initial moisture content; obtaining the probability set of weather forecast variables; integrating the above probability distributions to drive Monte Carlo simulation to obtain the probability distribution of soil moisture content within the forecast period; calculating the probability of water shortage accordingly; determining a water shortage probability threshold based on the user's risk preference; and generating irrigation volume instructions based on the corresponding quantiles when the water shortage probability is greater than or equal to the threshold. This invention achieves differentiated intelligent irrigation decision-making under multiple risk preferences, improving the field adaptability of irrigation decisions.

Description

Technical Field

[0001] This invention relates to the field of crop irrigation technology, specifically to a method and system for precision irrigation control of dryland crops based on digital twins. Background Technology

[0002] Current farmland irrigation decision-making methods based on digital twins mainly fall into two categories. One category, such as CN121386434A, trains an agent in a digital twin environment using reinforcement learning, directly outputting deterministic irrigation control commands to achieve automated decision-making. This type of method treats uncertainties in the real-world environment as training noise, neglecting risk quantification and interpretation in the decision-making process, making it difficult for users to trust the system's output. The other category, such as CN120689164A, although using probabilistic modeling through lag indices in the soil moisture prediction stage, fails to extend its uncertainty quantification to the irrigation decision-making level. The decision-making stage still degenerates into deterministic threshold judgment, unable to provide users with differentiated strategies based on different risk preferences.

[0003] In summary, existing technologies generally treat predictive uncertainty as an internal implicit defect, lacking a complete mechanism to explicitly quantify uncertainty and use it to drive risk alerts and irrigation decisions. Users cannot know the credibility of the decisions, which limits the promotion and application of digital twin systems in actual agricultural production. Summary of the Invention

[0004] This invention provides a method and system for precision irrigation control of dryland crops based on digital twins, aiming to solve the technical problem that existing irrigation decision-making technologies cannot provide users with differentiated strategies under different risk preferences.

[0005] In view of the above problems, the present invention provides a method and system for precision irrigation control of dryland crops based on digital twins.

[0006] In a first aspect, the present invention provides a method for precision irrigation control of dryland crops based on digital twins, comprising:

[0007] Load multi-source sensing data of the target farmland, wherein the multi-source sensing data includes at least measured soil moisture data and weather forecast data; Based on the measured soil moisture data, Bayesian inference is performed on the key soil hydraulic parameters in the digital twin model to obtain the posterior probability distribution of the key soil hydraulic parameters. Based on the multi-source sensing data, the model input conditions of the digital twin model are probabilistically modeled to obtain the probability distribution field of the initial soil moisture content. Based on the meteorological forecast data, obtain the probability set of the meteorological forecast variables; The posterior probability distribution of the key soil hydraulic parameters, the probability distribution field of the initial soil moisture content, and the probability set of the meteorological forecast variables are integrated to drive the digital twin model to perform Monte Carlo simulation and obtain the probability distribution of soil moisture content during the prediction period. The probability of water shortage is calculated based on the probability distribution of soil moisture content during the predicted period; Read the user-configured risk preference pattern and determine the water shortage probability threshold based on the risk preference pattern; When the water shortage probability is greater than or equal to the water shortage probability threshold, an irrigation decision is triggered, and an irrigation instruction is generated based on the quantile corresponding to the risk preference pattern in the posterior distribution of irrigation volume.

[0008] Secondly, the present invention provides a precision irrigation control system for dryland crops based on digital twins, comprising: A multi-source data loading module is used to load multi-source sensing data of the target farmland, wherein the multi-source sensing data includes at least measured soil moisture data and weather forecast data. The Bayesian inference module is used to perform Bayesian inference on key soil hydraulic parameters in the digital twin model based on the measured soil moisture data, and obtain the posterior probability distribution of the key soil hydraulic parameters. The initial probability field modeling module is used to perform probability modeling on the model input conditions of the digital twin model based on the multi-source sensing data, and obtain the probability distribution field of the initial soil moisture content. The meteorological probability set acquisition module is used to acquire the probability set of meteorological forecast variables based on the meteorological forecast data. The Monte Carlo simulation module is used to integrate the posterior probability distribution of the key soil hydraulic parameters, the probability distribution field of the initial soil moisture content, and the probability set of the meteorological forecast variables to drive the digital twin model to perform Monte Carlo simulation and obtain the probability distribution of soil moisture content during the prediction period. The water shortage probability calculation module is used to calculate the water shortage probability based on the probability distribution of soil moisture content during the predicted period. The risk threshold determination module is used to read the user-configured risk preference pattern and determine the water shortage probability threshold based on the risk preference pattern. An irrigation instruction generation module is used to trigger an irrigation decision when the water shortage probability is greater than or equal to the water shortage probability threshold, and to generate an irrigation instruction based on the quantile corresponding to the risk preference pattern in the posterior distribution of irrigation volume.

[0009] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides a method and system for precision irrigation control of dryland crops based on digital twins. By integrating Bayesian inference, probability field modeling, meteorological probability ensembles, and Monte Carlo simulation, it achieves explicit quantification of uncertainty throughout the entire process from parameter inversion to soil moisture prediction. This overcomes the shortcomings of existing methods that treat uncertainty as noise or only model it in the prediction stage, leading to decision-making degenerates into threshold judgment. The method calculates the probability of water shortage based on the predicted moisture content probability distribution and dynamically determines the water shortage probability threshold according to the user-configured conservative, balanced, or risk-averse risk preference mode. When the water shortage probability reaches the threshold, a differentiated irrigation instruction is generated by selecting the corresponding quantile from the posterior distribution of irrigation volume. Ultimately, the method outputs the decision result and provides decision credibility, allowing users to flexibly adjust irrigation strategies according to their own risk tolerance, achieving an adaptive balance between water conservation and yield protection. This enhances the interpretability and practical application value of the digital twin irrigation method. Attached Figure Description

[0010] Figure 1 A flowchart illustrating the precision irrigation control method for dryland crops based on digital twins provided in an embodiment of the present invention; Figure 2 A schematic diagram of the overall logic of the precision irrigation control method for dryland crops based on digital twins provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of a precision irrigation control system for dryland crops based on digital twins provided in an embodiment of the present invention; The components represented by each number in the attached diagram are explained below: The module includes: multi-source data loading module 11, Bayesian inference module 12, initial probability field modeling module 13, meteorological probability set acquisition module 14, Monte Carlo simulation module 15, water shortage probability calculation module 16, risk threshold determination module 17, and irrigation instruction generation module 18. Detailed Implementation

[0011] This invention provides a method and system for precision irrigation control of dryland crops based on digital twins, which addresses the technical problem that existing irrigation decision-making technologies cannot provide users with differentiated strategies under different risk preferences.

[0012] Example 1, as Figure 1 , Figure 2 As shown, this invention provides a method for precision irrigation control of dryland crops based on digital twins, the method comprising: S100: Load multi-source sensing data of the target farmland, wherein the multi-source sensing data includes at least measured soil moisture data and weather forecast data.

[0013] In precision irrigation control of dryland crops, relying solely on measured soil moisture data is insufficient to comprehensively characterize farmland water transport and crop water consumption. Measured soil moisture data only reflects the local water content at sensor locations and cannot cover the spatial heterogeneity of the field. Meteorological forecast data provides atmospheric driving conditions but lacks information on soil thermal state, salinity stress, and crop growth status, leading to incomplete initial boundaries for the digital twin model and affecting the accuracy of subsequent parameter inference and prediction. Therefore, step S100 requires loading multi-source sensing data covering soil, crops, meteorology, and remote sensing to provide sufficient input information for the digital twin model.

[0014] Step S100 in the method provided in this embodiment of the invention includes: Load measured soil temperature data, measured soil salinity data, leaf area index and canopy temperature from satellite remote sensing imagery, as well as soil property data and crop variety data of the target farmland.

[0015] First, the measured soil moisture data and weather forecast data are loaded. The measured soil moisture data refers to the time series of soil volumetric water content obtained directly from frequency domain reflectance (FDR) or time domain reflectance (TDR) sensors deployed at different soil depths in the target farmland, according to a preset sampling frequency. This data is expressed as a percentage, reflecting the volume proportion of water in the soil pores. The weather forecast data refers to numerical weather prediction products obtained from meteorological departments or third-party meteorological service interfaces, including hourly precipitation, air temperature, wind speed, solar radiation, and reference crop evapotranspiration for the next 168 hours.

[0016] Specifically, data collected by soil moisture sensors is aggregated to a data logger via a wireless sensor network. The data logger then uploads the data to a data storage server in a JavaScript object abbreviation format using a unified timestamp format. Weather forecast data is retrieved via an application programming interface (API) using the Hypertext Transfer Protocol (HTTP), and the returned data is in a hierarchical data format or a general geoscience data format file.

[0017] For example, a potato field in the dryland farming region of Northwest China was selected as the target farmland. Three representative sampling points were selected in the farmland. At each sampling point, a frequency domain reflectance method soil moisture sensor (FDR-100 model) was installed at depths of 20cm, 40cm, and 60cm. The sampling interval was set to 30 minutes, and data was collected continuously for 30 days to obtain a sequence of measured soil moisture data. For example, the soil volumetric water content at a depth of 20cm at a certain moment was 28.5%. Meteorological forecast data were obtained from the China Meteorological Administration's ensemble forecast system, using the 3rd edition of the meteorological forecast system. The initial forecast time was 8:00 AM on April 10, 2026, and the forecast period was 168 hours. The predicted variables included: hourly precipitation, air temperature at 2 meters above the ground, wind speed at 10 meters above the ground, and downward shortwave radiation.

[0018] Secondly, the measured soil temperature and soil salinity data are loaded. The measured soil temperature data refers to the soil temperature value obtained by a thermistor temperature sensor deployed at the same location as the soil moisture sensor. The measured soil salinity data refers to the soil salt concentration obtained by converting the conductivity of the soil solution extract or in-situ soil pore water measured by a conductivity sensor.

[0019] Specifically, soil temperature and soil moisture sensors are integrated onto the same probe to synchronously collect temperature data. Soil salinity data are obtained using a four-electrode conductivity sensor directly buried in the soil. An AC voltage is applied, and the current response is measured to calculate the apparent soil conductivity. This conductivity is then corrected for soil texture and temperature coefficient to a standard conductivity value at 25°C, and finally converted to soil salinity using an empirical formula. All data are stored together with soil moisture data using the same timestamp.

[0020] For example, thermistor temperature sensors were integrated and installed at the same depths of 20cm, 40cm, and 60cm to synchronously record soil temperature. For instance, the soil temperature at a certain moment at a depth of 20cm was 18.5℃. Four-electrode conductivity sensors were buried at the same locations to measure the apparent conductivity of the soil. After calibration, the conductivity at a standard of 25℃ was found to be 0.35dS / m, which translates to a soil salinity of 0.8g / kg, indicating that the plot belongs to non-saline soil.

[0021] Next, leaf area index (LAI) and canopy temperature from satellite remote sensing imagery data are loaded. Leaf area index (LAI) is the ratio of the total area of ​​plant leaves per unit land area to the land area; it is a dimensionless parameter used to characterize crop canopy structure and photosynthetically active radiation interception capacity. Canopy temperature refers to the infrared radiation temperature at the top of the crop canopy, reflecting the crop's water stress status. Satellite remote sensing imagery data refers to multispectral or thermal infrared imagery products acquired from satellite platforms such as the Medium Resolution Imaging Spectroradiometer (MRISC), Landsat, or Sentinel-2.

[0022] Specifically, atmospherically and geometrically corrected remote sensing image products were downloaded from publicly available Earth observation data interfaces. For leaf area index (LAI), the normalized vegetation index (NDI) was calculated using satellite multispectral bands, and then pixel-by-pixel LAI values ​​were obtained by inverting the NDI with an empirical statistical model of NDI and LAI. For canopy temperature, land surface temperature was extracted from thermal infrared imagery, and the average radiation temperature of all topsoil pixels within the target farmland area was taken as the canopy temperature, taking into account the crop masking area. The data was loaded in a geotagged raster format, and the remote sensing pixel values ​​were mapped to the grid cells of the digital twin model using spatial interpolation methods.

[0023] For example, a Sentinel-2 optical image with a spatial resolution of 10m was selected from a date close to the actual soil moisture measurement data. After atmospheric correction, the reflectance in the red and near-infrared bands was calculated, yielding a Normalized Difference Vegetation Index (NDV) of 0.72. Based on the linear regression formula between the NPV and leaf area index (LAI): LAI = 2.5 × NPV - 0.3, the LAI was calculated to be 1.5. The thermal infrared band of Landsat-8, which passed over the same day, was selected to extract the surface temperature within the target farmland area. After atmospheric moisture correction, the canopy temperature was found to be 24.2℃.

[0024] Simultaneously, soil property data and crop variety data of the target farmland are loaded. Soil property data includes soil texture, soil bulk density, field water holding capacity, wilting coefficient, and saturated hydraulic conductivity. Crop variety data includes crop type, variety name, growing season days, maximum root depth, crop coefficient curve, and water sensitivity coefficient.

[0025] Specifically, soil property data were obtained through laboratory methods for soil physicochemical analysis. Topsoil and deep undisturbed soil samples were collected from the target farmland using a grid method. These samples were then brought back to the laboratory where soil texture was determined using the pipette method or a laser particle size analyzer; soil bulk density was determined using the ring sampler method; field water holding capacity and wilting coefficient were determined using the pressure membrane method; and saturated hydraulic conductivity was determined using the constant head method. Crop variety data were obtained from variety description documents provided by local agricultural extension departments or seed suppliers and entered into the database.

[0026] For example, undisturbed soil samples were collected from the 0-60cm soil layer. Laboratory analysis showed the following results: sand content 20%, silt content 72%, clay content 8%, and soil texture was loess. The measured soil bulk density was 1.19 g / cm³. The field water holding capacity was 21.2%, the wilting coefficient was 7.2%, and the saturated water content was 28.6%. The crop variety was Longshu 10, a late-maturing variety with a total growth period of approximately 120 days and a maximum root depth of 60cm. The crop coefficient curves for the initial, mid-, and late growth stages were 0.4, 1.15, and 0.65, respectively. The water sensitivity coefficients for the tuber formation and tuber enlargement stages were 0.75 and 0.90, respectively.

[0027] In this embodiment of the invention, comprehensive loading of multi-source sensing data of the target farmland was completed, including soil moisture, soil temperature, soil salinity, weather forecasts, satellite remote sensing leaf area index and canopy temperature, soil properties, and crop variety data. This data covers key information elements of the soil-vegetation-atmosphere continuum, providing prior parameters and conditional information for subsequent Bayesian inference, a spatial interpolation basis for initial water content probability field modeling, and a raw data source for constructing the probability set of weather forecast variables, thus ensuring the integrity of the input conditions for the digital twin model.

[0028] S200: Based on the measured soil moisture data, perform Bayesian inference on the key soil hydraulic parameters in the digital twin model to obtain the posterior probability distribution of the key soil hydraulic parameters.

[0029] In this embodiment of the invention, the digital twin model includes several key soil hydraulic parameters, such as saturated hydraulic conductivity, field capacity, wilting coefficient, and shape parameters. These parameters exhibit significant spatial heterogeneity across different fields and soil depths, and their precise values ​​cannot be obtained through direct measurement. Traditional methods assign a fixed value to each parameter, ignoring their uncertainty and leading to discrepancies between model predictions and actual soil moisture dynamics. Bayesian inference combines prior knowledge with time series data of measured soil moisture to calculate the posterior probability distribution of each parameter given the measured data. This posterior distribution quantifies the parameter's uncertainty, rather than providing a single fixed value, thus offering a parameter input set that reflects actual variation for subsequent Monte Carlo simulations.

[0030] Step S200 in the method provided in this embodiment of the invention includes: Set prior probability distributions for each key soil hydraulic parameter; The time series of the measured soil moisture data is used as conditional information; The posterior distribution is updated batch by batch using the differential evolution Markov chain Monte Carlo algorithm. It also automatically triggers a post-hoc update whenever new data is received.

[0031] First, a prior probability distribution is established for each key soil hydraulic parameter. A prior probability distribution refers to the initial probability distribution set for each key soil hydraulic parameter based on existing soil property data, empirical values ​​from literature, or regional statistical patterns before any measured soil moisture data are observed. The prior distribution reflects the prior knowledge of the reasonable range of parameter values. Common types of prior distributions include uniform distribution, normal distribution, or log-normal distribution. The parameters of the prior distribution are determined based on the empirical range of values ​​for soil texture classification.

[0032] Specifically, for each key soil hydraulic parameter, based on the soil texture and bulk density obtained in S100, the empirical range of values ​​for that parameter is obtained by consulting a soil hydraulic property database or using a soil transfer function. When setting the prior probability distribution, if the parameter value range is finite and has no obvious central trend, a uniform distribution is adopted; if the parameter has a known typical mean and small variation, a normal distribution is adopted. The prior distribution of each parameter is stored in the form of a probability density function as the starting point for Bayesian inference.

[0033] For example, based on the loess soil and soil bulk density of 1.19 g / cm³ 3 Prior information for four key soil hydraulic parameters was obtained from the Rosetta database. The first key parameter is saturated hydraulic conductivity K_s, with an empirical range of 0.5 cm / h to 8 cm / h, set as uniformly distributed, denoted as K_s~Uniform(0.5,8.0). The second key parameter is field capacity θ_fc, with an empirical range of 0.20 to 0.35, set as uniformly distributed, denoted as θ_fc~Uniform(0.20,0.35). The third key parameter is the wilting coefficient θ_wp, with an empirical range of 0.08 to 0.15, set as uniformly distributed, denoted as θ_wp~Uniform(0.08,0.15). The fourth key parameter is the van Gernühten model shape parameter n, with an empirical range of 1.1 to 1.8, set as normally distributed, with a mean of 1.45 and a standard deviation of 0.15, denoted as n~Normal(1.45,0.15). 2 ).

[0034] Secondly, the time series of the measured soil moisture data is used as conditional information. This time series data serves as conditional information for Bayesian inference, updating the prior probability distribution. The conditional information provides discrepancies between model predictions and actual observations; the smaller the discrepancy, the more reasonable the parameter values. Bayesian inference is a parametric statistical method based on Bayes' theorem. It utilizes prior knowledge and measured data, employing Markov chain Monte Carlo sampling to calculate the probability distribution of parameters under given data.

[0035] Specifically, the time series of measured soil moisture data recorded in S100 is obtained. For each set of candidate parameter values, the digital twin model is driven to simulate the change in soil moisture content within the same time period, and the error between the simulated value and the measured value is calculated. Assuming that the error follows a normal distribution with a mean of 0, a likelihood function is constructed: likelihood value = product of the error probability density at each time step. This likelihood value reflects the probability that the digital twin model can reproduce the measured data sequence under the current parameter values.

[0036] For example, a time series of measured soil moisture data at a depth of 20 cm for seven consecutive days was selected, with a sampling interval of 30 minutes, resulting in 336 data points. The specific sequence segment is as follows: Volumetric water content measured at 8:00 AM on day 1 was 28.5%, at 8:30 AM it was 28.4%, ..., and at 8:00 PM on day 7 it was 27.2%. Using this time series as conditional information, for a set of candidate parameters, such as K_s=2.0 cm / h, θ_fc=0.28, θ_wp=0.12, n=1.45, a digital twin model was run to simulate the water content changes at a depth of 20 cm within the same seven days, yielding the simulated sequence: 28.3% at 8:00 AM on day 1, 28.2% at 8:30 AM, ..., and 27.5% at 8:00 PM on day 7. The difference between the simulated and measured values ​​at each time point was calculated, resulting in an error sequence: -0.2%, -0.2%, ..., +0.3%. Assuming the errors follow a normal distribution with a mean of 0 and a standard deviation of 1%, the likelihood value of this set of parameters is 0.032 when the probability density values ​​corresponding to each error are multiplied together.

[0037] Secondly, the posterior distribution is updated batch-wise using the Differential Evolutionary Markov Chain Monte Carlo algorithm. The posterior probability distribution refers to the probability distribution of key soil hydraulic parameters given a time series of measured soil moisture data. According to Bayes' theorem, the posterior probability distribution is proportional to the prior probability distribution multiplied by the likelihood function. The Differential Evolutionary Markov Chain Monte Carlo algorithm is an optimized sampling method for drawing samples from the posterior distribution. This algorithm runs multiple Markov chains simultaneously, with candidate samples for each chain generated from the current samples of other chains through differential evolution. The acceptance of candidate samples is determined based on the Metropolis acceptance criterion. The algorithm can efficiently explore high-dimensional parameter spaces and obtain a sample set of the posterior distribution.

[0038] Specifically, the number of parallel Markov chains is set to the chain number. A set of parameter samples is initialized for each chain, randomly drawn from a predefined prior distribution. For the current sample of each chain, samples from two other chains are randomly selected, and a difference vector is calculated. This difference vector is multiplied by a scaling factor and added to the current sample to generate candidate samples. The likelihood value corresponding to the candidate sample is calculated using a digital twin model, and the posterior probability ratio is calculated by combining it with the prior probability density value. The posterior probability ratio of the candidate sample and the current sample is compared with a uniformly distributed random number in the range [0,1]. If the ratio is greater than the random number, the candidate sample is accepted; otherwise, it is rejected. This process is repeated until the preset number of iterations is reached. The first half of the iterations is discarded as a warm-up period, and the samples from the second half of the iterations are retained as valid samples for the posterior distribution.

[0039] For example, we set the number of parallel Markov chains to 10, with a parameter dimension of 4 (K_s, θ_fc, θ_wp, n). The initial samples of the first chain are drawn from the prior distribution: K_s = 1.2 cm / h, θ_fc = 0.26, θ_wp = 0.10, n = 1.50. The initial samples of the second chain are: K_s = 3.5 cm / h, θ_fc = 0.30, θ_wp = 0.13, n = 1.40. In each iteration, for the current chain, we randomly select the current samples of the other two chains, calculate the difference vector, multiply it by a scaling factor of 0.5, and add it to the current sample of the first chain to generate candidate samples. Assume the candidate samples are K_s = 1.8 cm / h, θ_fc = 0.27, θ_wp = 0.11, n = 1.48. We call the digital twin model to calculate the likelihood value of the candidate samples as 0.045. Calculate the prior probability density: K_s=1.8 has a density value of 1 / 7.5≈0.1333 on Uniform(0.5,8.0), θ_fc=0.27 has a density value of 1 / 0.15≈6.6667 on Uniform(0.20,0.35), θ_wp=0.11 has a density value of 1 / 0.07≈14.2857 on Uniform(0.08,0.15), and n=1.48 has a density value of 1 / 7.5≈0.1333 on Uniform(0.20,0.35). 2 The density value on the first chain is 2.65. The product of the prior probability densities is approximately 33.6 (0.1333 × 6.6667 × 14.2857 × 2.65). The posterior probability is 1.512 (33.6 × 0.045). The posterior probability of the current sample is similarly calculated to be 1.202. The posterior probability ratio is 1.512 / 1.202 = 1.258, which is greater than the random number's 0.3, therefore the candidate sample is accepted. This process is repeated 5000 times, retaining the samples from the last 2500 iterations, resulting in 10 chains × 2500 = 25000 samples, forming an approximation of the posterior distribution.

[0040] Finally, a posterior update is automatically triggered each time new data is received. A data increment monitoring mechanism is set up in the data loading module. When the number of newly added measured soil moisture data records reaches a preset threshold, the posterior update process is automatically triggered. The obtained current posterior distribution sample set is used as the new prior distribution: for each parameter, a kernel density estimate is fitted from the samples as the new prior. The newly added data time series is used as conditional information, and the differential evolution Markov chain Monte Carlo algorithm is used to perform sampling again, generating a new sample set from the posterior distribution. After the update is completed, the original posterior distribution sample set is replaced for subsequent Monte Carlo simulations.

[0041] For example, after the initial Bayesian inference is completed on day 30, 25,000 posterior samples are obtained. From day 31 to day 33, the soil moisture sensor continues to collect data every 30 minutes, accumulating 144 new data points. The data loading module detects that the number of new data points has reached 144 and automatically triggers the posterior update. From the K_s posterior samples obtained on day 30, a new prior distribution of K_s is generated using kernel density estimation. This distribution approximates a truncated normal distribution with a mean of 2.1 cm / h and a standard deviation of 0.4 cm / h. Similarly, new prior distributions of θ_fc, θ_wp, and n are obtained. Using the 144 newly added measured soil moisture data points as conditional information, the differential evolution Markov chain Monte Carlo algorithm is re-executed, with 10 chains and 3000 iterations. The last 1500 iterations are retained, resulting in 15,000 updated posterior samples. The updated K_s posterior mean becomes 2.3 cm / h, and the standard deviation shrinks to 0.3 cm / h, indicating that the parameter uncertainty decreases as data accumulates.

[0042] In this embodiment of the invention, Bayesian inference of key soil hydraulic parameters in a digital twin model was completed. First, a prior probability distribution based on soil texture was established for each parameter. Then, time series data of measured soil moisture were used as conditional information. A differential evolution Markov chain Monte Carlo algorithm was employed to update the posterior distribution batch by batch. The posterior update was automatically triggered each time new data was received. Finally, the posterior probability distribution of each key soil hydraulic parameter was obtained. This posterior distribution quantifies the uncertainty of the parameters, replacing the single fixed-value parameter input method, and providing a random sample source that conforms to the true variability of the parameters for subsequent Monte Carlo simulations. The automatic triggering update mechanism ensures that the digital twin model can adaptively optimize as real-time monitoring data increases, improving the model's responsiveness to changes in field conditions.

[0043] S300: Based on the multi-source sensing data, perform probabilistic modeling of the model input conditions of the digital twin model to obtain the probability distribution field of the initial soil moisture content.

[0044] In dryland farming, measured soil moisture data comes from only a limited number of discrete sensor sampling points, while digital twin models require assigning initial soil moisture content values ​​to every grid cell across the entire target farmland. Directly using inverse distance weighting or Kriging interpolation methods can only obtain moisture content estimates for each location, failing to assess the reliability of the interpolation results. If the initial moisture content exhibits significant spatial variation, a single interpolation estimate introduces uncertainty, which propagates and amplifies in subsequent simulations. Gaussian process regression, a nonparametric probabilistic regression method, can predict soil moisture content at any unsampled location based on discrete sampling points, simultaneously outputting the predicted mean and variance. Combining the predicted mean and variance for all unsampled locations generates a probability distribution field of initial soil moisture content covering the entire field, providing probabilistic input for subsequent Monte Carlo simulations.

[0045] Step S300 in the method provided in this embodiment of the invention includes: Gaussian process regression was used, with the coordinates and depth of discrete sampling points of the sensor as input features and the measured soil moisture content at the corresponding points as the target variable, to train the Gaussian process regression model; Using the trained Gaussian process regression model, predictions are made for any unsampled location in the target farmland, and the predicted mean and variance of soil moisture content at the corresponding location are output. The predicted mean and predicted variance of all unsampled locations are combined to generate a probability distribution field of the initial soil moisture content covering the entire field area.

[0046] First, Gaussian process regression is employed, using the coordinates and depth of discrete sampling points of the sensors as input features and the measured soil moisture content at the corresponding points as the target variable to train the Gaussian process regression model. Gaussian process regression is a nonparametric regression method based on Bayesian theory, assuming that the target variable corresponding to any finite number of input points follows a joint Gaussian distribution. This model describes the similarity between different input points by defining a covariance function: the closer the input points are in space and depth, the stronger the correlation of their target variables. Discrete sampling points of the sensors refer to the geographical coordinates and burial depth of the soil moisture sensors actually deployed in the target farmland. Training the Gaussian process regression model involves using the input features and measured values ​​of the sampling points to optimize the hyperparameters in the covariance function by maximizing the logarithmic marginal likelihood function, thereby enabling the model to capture the spatial and depth-related variations in soil moisture content.

[0047] Specifically, the three-dimensional coordinates of all sensor sampling points are collected: a relative coordinate system is established with the southwest corner of the target farmland as the origin, labeled x (horizontal coordinate), y (vertical coordinate), and z (depth coordinate). The three-dimensional coordinates of all sampling points are combined into an input feature matrix, denoted as: Input matrix = [x, y, z]. The measured soil moisture content values ​​at corresponding points are combined into a target vector. The covariance function is chosen as the squared exponential kernel function with a noise term. The form of the squared exponential kernel function is: Kernel function value = signal variance × exp(-(squared distance between two input points) / (2 × squared length scale)), where the distance between two input points... The signal variance controls the magnitude of the predicted value variation, while the length scale controls the range of influence of input variations on the output correlation. The gradient ascent method is used to maximize the logarithmic marginal likelihood function, solving for the optimal values ​​of the signal variance and length scale. After training, the Gaussian process regression model determines the posterior distribution for predicting soil moisture content at any new input point.

[0048] For example, the target farmland is rectangular (200m × 100m), with the southwest corner as the origin (0,0). Three sensor sampling points are set up with planar coordinates of (50,30), (120,60), and (170,80), respectively. Each sampling point has three depths: 20cm, 40cm, and 60cm, for a total of nine sampling points. The measured soil moisture content at each point, from shallowest to deepest, is as follows: Point 1 (28.5%, 27.8%, 26.2%), Point 2 (27.2%, 26.5%, 25.0%), and Point 3 (26.8%, 25.9%, 24.5%). A quadratic exponential kernel function is used, with an initial assumption of a signal variance of 10 and a length scale of 50. Optimization is performed using the gradient ascent method, and the optimal signal variance is found to be 8.5 and the optimal length scale is 35 after convergence. The optimized kernel function can better explain the spatial correlation between the sampling point data.

[0049] Secondly, using the trained Gaussian process regression model, predictions are made for any unsampled location in the target farmland, outputting the predicted mean and variance of soil moisture content at that location. Unsampled locations refer to any spatial point in the target farmland where no soil moisture sensors are deployed, including planar grid nodes and different depth levels. The predicted mean is the best estimate of soil moisture content output by the Gaussian process regression model at that location. The predicted variance quantifies the uncertainty of this estimate; a larger variance indicates a less reliable prediction at that location. The predicted mean is calculated as a weighted average of the target values ​​from the training samples, with the weights proportional to the kernel function values ​​from the unsampled location to each sampling point. The predicted variance is calculated by subtracting the variance reduction term resulting from the training sample information from the prior variance of the kernel function.

[0050] Specifically, the target farmland is divided into regular grids, for example, a planar grid spacing of 5m × 5m. Each grid node is further considered with several depth layers, for example, every 10cm, from the surface to a maximum root depth of 100cm. For each unsampled location, the kernel function value vector between that location and all training sampled points is calculated. According to the standard prediction formula of Gaussian process regression, the prediction mean = transpose of the kernel function value vector × inverse of the kernel function matrix × training target value vector. The prediction variance = kernel function autocovariance (sum of signal variance and noise variance) - transpose of the kernel function value vector × inverse of the kernel function matrix × kernel function value vector. The prediction mean and prediction variance are stored in arrays corresponding to the grid locations.

[0051] For example, take an unsampled location in the target farmland with planar coordinates (100, 50) and a depth of 40 cm. Calculate the kernel function value between this location and the 9 training sampling points. The distance from this location to point 1 (50, 30, 20) is calculated to be 57.45. Similarly, the kernel function value vectors to other sampling points are calculated as [2.21, 2.68, 1.92, 1.85, 2.50, 1.76, 1.52, 2.10, 1.48]. The kernel function matrix between training sampling points is a 9×9 matrix. Multiplying the kernel function value vector by the inverse of the kernel function matrix, and then multiplying by the training target value vector, the predicted mean is calculated to be 26.8%. The prediction variance is calculated as follows: kernel function autocovariance = signal variance 8.5 + noise variance 0.5 = 9.0. Substituting into the formula, the prediction variance is 7.8, and the prediction standard deviation is sqrt(7.8) = 2.79%. Therefore, at the position (100, 50, 40), the predicted mean of soil moisture content is 26.8%, and the prediction variance is 7.8.

[0052] Next, the predicted mean and predicted variance of all unsampled locations are combined to generate a probability distribution field of initial soil moisture content covering the entire field. This probability distribution field represents the soil moisture content at each spatial location within the target farmland using a probability distribution, rather than a single numerical value. This probability distribution is typically assumed to be normally distributed, with the predicted mean as the mean and the predicted variance as the variance. The predicted mean and predicted variance of all grid nodes are organized into a three-dimensional array by spatial index to form the probability distribution field. This distribution field serves as the initial condition for Monte Carlo simulations using the digital twin model: each simulation independently samples from the probability distribution of one location to obtain a set of initial spatial distributions of moisture content.

[0053] Specifically, following the defined grid, all planar grid nodes and preset depth layers are traversed. For each grid node, the predicted mean m(x,y,z) and predicted variance v(x,y,z) are obtained from the Gaussian process regression model. The mean and variance are stored in two three-dimensional arrays of the same dimension: a mean array M and a variance array V, with the array dimensions being the number of grid rows × the number of grid columns × the number of depth layers. The probability distribution field is represented as: the initial soil moisture content at any location (x,y,z) ~Normal(M(x,y,z),V(x,y,z)). This probability distribution field is saved as a binary or network-standard data format file for subsequent steps in S500. Simultaneously, a visualization image can be generated to help users understand the spatial distribution of uncertainty in the initial moisture content.

[0054] For example, a 200m × 100m farmland is divided into planar grids with a spacing of 5m, resulting in 41 grid nodes in the row direction (0m, 5m, 10m, ..., 200m) and 21 grid nodes in the column direction (0m, 5m, 10m, ..., 100m), for a total of 861 planar nodes. The depth layer is set from the surface to 60cm, with each layer being 10cm deep, for a total of 7 layers. This results in a total of 861 × 7 = 6027 spatial locations. For each location, a Gaussian process regression model is used to calculate the predicted mean and predicted variance. For example, at planar coordinates (100, 50) and a depth of 20cm, the predicted mean is 27.5%, and the predicted variance is 6.2. For example, at planar coordinates (100, 50) and a depth of 60cm, the predicted mean is 25.3%, and the predicted variance is 8.1. The mean and variance of all 6027 locations are stored in the mean array M

[41]

[21] [7] and the variance array V

[41]

[21] [7], respectively. Finally, a probability distribution field of the initial soil moisture content covering the entire field area is generated.

[0055] In this embodiment of the invention, probabilistic modeling of the initial soil moisture content in the input conditions of a digital twin model was completed. First, Gaussian process regression was used, with the three-dimensional coordinates of discrete sampling points of the sensor as input features and the measured soil moisture content at the corresponding points as the target variable. The Gaussian process regression model was trained and the kernel function hyperparameters were optimized. Then, the trained model was used to predict any unsampled location in the target farmland, simultaneously outputting the predicted mean and predicted variance. Finally, the predicted mean and predicted variance of all unsampled locations were combined to generate a probability distribution field covering the entire field area for the initial soil moisture content. This probability distribution field not only provides the best estimate for each location but also quantifies the uncertainty introduced by spatial interpolation, providing initial state random samples that conform to actual spatial variability for Monte Carlo simulation. Compared to single-value interpolation methods, the probability distribution field can more realistically reflect the uncertainty of the initial conditions, thereby improving the reliability of subsequent soil moisture content predictions.

[0056] S400: Based on the meteorological forecast data, obtain the probability set of the meteorological forecast variables.

[0057] Meteorological forecast data are typically presented in deterministic form, but actual atmospheric evolution exhibits chaotic characteristics, and a single deterministic forecast cannot reflect the uncertainty of the forecast. Ensemble forecast systems generate multiple parallel forecast members by applying small perturbations to the initial field, boundary conditions, or model physical processes; the set of differences among these members reflects the uncertainty of atmospheric evolution. However, ensemble forecast systems from different sources have different forecast biases and dispersion characteristics. Directly using ensemble forecasts from a single source, or averaging them with equal weights, cannot fully utilize multi-source information and cannot quantitatively characterize the reliability differences between forecast sources. Furthermore, raw ensemble forecasts often contain biases and require post-processing calibration. Therefore, step S400 requires acquiring ensemble forecast data for precipitation, temperature, wind speed, and radiation from multiple source ensemble forecast systems, performing weighted fusion using dynamic weights based on information entropy, and then performing post-processing calibration using quantile mapping. Finally, a probability set of calibrated meteorological forecast variables is obtained, providing meteorological driving inputs with reliable uncertainty representation for subsequent Monte Carlo simulations.

[0058] Step S400 in the method provided in this embodiment of the invention includes: Accompanied forecast data for precipitation, temperature, wind speed, and radiation are obtained from ensemble forecast systems of multiple sources in the aforementioned meteorological forecast data. We employ dynamic weights based on information entropy to perform weighted fusion of ensemble forecast data from different forecast sources, where forecast sources with greater forecast uncertainty have lower weights. The quantile mapping method is used to perform post-processing calibration on the weighted and fused ensemble forecast data to obtain the probability set of the calibrated meteorological forecast variables.

[0059] First, ensemble forecast data for precipitation, temperature, wind speed, and radiation are obtained from multiple ensemble forecast systems within the aforementioned meteorological forecast data. Forecast data is acquired from interfaces of two or more ensemble forecast systems. For each forecast source, all its forecast members are obtained. Forecast variables include precipitation, temperature, wind speed, and radiation. The values ​​of each member at each time point within the forecast period are extracted according to a uniform temporal and spatial resolution. The data from all forecast sources are organized into a multidimensional array according to variable category, member number, and forecast time.

[0060] For example, two ensemble forecast sources are selected: the European Centre for Medium-Range Weather Forecasts (ECMWFEPS) ensemble forecast system based on the fifth-generation reanalysis data, and the China Meteorological Administration's Global Ensemble Forecast System (CMAGEPS). Each source acquires hourly forecast data for the next 168 hours. ECMWFEPS provides a total of 51 members, with the first member predicting 0.2 mm of precipitation, 18.5°C, 2.1 m / s wind speed, and 320 W / m² radiation for the first hour. 2 The second member had a precipitation of 0.0 mm, a temperature of 18.8℃, a wind speed of 2.3 m / s, and a radiation of 305 W / m. 2 This continues until the 51st member. CMAGEPS provides a total of 31 members, with a similar structure. All data is stored according to variable categories: for example, the precipitation array has a dimension of 2 times the number of forecast sources × the number of members × 168 hours, and the temperature, wind speed, and radiation arrays are stored similarly.

[0061] Secondly, a dynamic weighting method based on information entropy is used to perform weighted fusion of ensemble forecast data from different forecast sources, where the forecast source with greater forecast uncertainty has a lower weight. Forecast uncertainty refers to the deviation between weather forecasts and the actual future state. This deviation cannot be directly calculated and is usually estimated through the dispersion among ensemble members: the greater the dispersion, the higher the forecast uncertainty. Information entropy is an indicator of uncertainty; a higher value indicates a higher degree of disorder in the probability distribution, i.e., greater forecast uncertainty. The dynamic weighting based on information entropy means dynamically calculating the weight of each forecast source according to the ensemble dispersion of each forecast source in the current forecast period. The forecast source with greater uncertainty has a lower weight, and the forecast source with less uncertainty has a higher weight. Weighted fusion means merging all members of multiple forecast sources into a large probability set according to the calculated weights.

[0062] Specifically, for each forecast source and each meteorological variable, the ensemble mean and ensemble standard deviation of that forecast source at each moment within the forecast period are calculated. The ensemble standard deviation serves as a measure of forecast uncertainty. The average of the ensemble standard deviations for all moments of the forecast source yields the overall uncertainty index U_i for that source. The weight of each forecast source is calculated: W_i = (1 / U_i) / (sum of (1 / U) for all forecast sources). This formula ensures that the weights are inversely proportional to the uncertainty and that the sum is 1. Then, members are randomly selected from the original member set of each forecast source according to their weight proportions to form a new weighted fusion set. The total number of members in the new set is equal to the pre-defined target set size, where the number of members from the i-th forecast source is approximately the target set size multiplied by W_i. Member resampling is used to preserve the probability distribution: all members of each forecast source are merged into a pool, and the probability of each member being selected is proportional to the weight W_i of its respective forecast source divided by the number of members in that forecast source. Sampling with replacement is performed until the target number of members is reached.

[0063] For example, the forecast uncertainty of precipitation variables for the next 168 hours is calculated for ECMWFEPS and CMAGEPS. The mean standard deviation of precipitation at each time point within the 168-hour period is calculated for the 51 members of ECMWFEPS, yielding an uncertainty index U_EC = 0.15 mm; the mean standard deviation of precipitation at each time point within the 168-hour period is calculated for the 31 members of CMAGEPS, yielding an uncertainty index U_CMA = 0.25 mm. According to the formula, the weights... W_CMA=0.375. The target fused set size is set to 100 members. Therefore, approximately 63 members are sampled with replacement from the 51 members of ECMWFEPS, and 37 members are sampled with replacement from the 31 members of CMAGEPS, forming a weighted fused set of 100 members. For the temperature variable, U_EC=0.8℃, U_CMA=1.2℃, weights W_EC=0.6, W_CMA=0.4, and similar sampling and fusion are performed. Wind speed and radiation variables are weighted and sampled and fused in the same way. Finally, fused sets of 100 members each are obtained for precipitation, temperature, wind speed, and radiation.

[0064] Next, the quantile mapping method is applied to the weighted ensemble forecast data for post-processing calibration, yielding a calibrated probability set of weather forecast variables. Quantile mapping is a statistical post-processing method used to correct biases in ensemble forecasts. This method establishes a mapping relationship between the forecast cumulative distribution function and the observed cumulative distribution function based on historical ensemble forecast data and corresponding station-measured data. Specifically, for a forecast quantile, the corresponding value in the observed data is found, and then a mapping function is constructed: calibrated forecast value = observed quantile function (forecast quantile position). Applying the mapping function to each member of the weighted ensemble forecast data yields the calibrated set, whose statistical distribution characteristics are more consistent with historical observed distributions.

[0065] Specifically, weighted fusion of forecast data from various sources for the same forecast lead time over the past 90 days is collected, along with measured precipitation, temperature, wind speed, and radiation data from meteorological stations near the target farmland within the corresponding lead time. For each variable, a paired sample of forecast and measured values ​​is established. The empirical cumulative distribution function of the forecast set is calculated, which involves sorting the forecast member values ​​for all training periods and calculating the cumulative probability corresponding to each value. The empirical cumulative distribution function of the measured value sequence is also calculated. For any new forecast value x_pred, the quantile τ corresponding to x_pred in the forecast cumulative distribution function is first determined, and then the measured value y_obs corresponding to that quantile τ is found in the measured cumulative distribution function. The calibrated value is y_obs. The above mapping operation is performed on each member in the weighted fusion set to obtain the calibrated probability set. If the data volume is sufficient, variants of quantile mapping can be used, such as quantile matching or methods based on equidistant cumulative distribution function matching.

[0066] For example, the training set is obtained by weighted fusion of daily ECMWFEPS and CMAGEPS data from the past 90 days for the next day's precipitation, along with the corresponding measured daily precipitation data. During the training period, the empirical distribution of the predicted precipitation set is as follows: the cumulative probability for a predicted value of 0.0 mm is 0.3, for 1.0 mm it is 0.6, and for 3.0 mm it is 0.9. The empirical distribution of the measured precipitation is as follows: the cumulative probability for a measured value of 0.0 mm is 0.2, for 0.5 mm it is 0.5, for 2.0 mm it is 0.85, and for 5.0 mm it is 0.95. Now, a new forecast member has a precipitation value of 1.0 mm. The cumulative probability of this forecast value in the forecast distribution is 0.6. The measured value corresponding to the cumulative probability of 0.6 in the measured distribution is found to be between 0.5 mm and 2.0 mm. Linear interpolation yields a calibration value of 0.5 + (2.0 - 0.5) × (0.6 - 0.5) / (0.85 - 0.5) ≈ 0.93 mm. Therefore, the calibrated precipitation for this member is 0.93 mm. Quantile mapping is then performed on each of the 100 weighted and fused members to obtain a set of calibrated precipitation probabilities, totaling 100 values. Similarly, mapping functions are established and calibrated for temperature, wind speed, and radiation variables. The final probability sets of the calibrated weather forecast variables are represented as follows: precipitation set P_set=[p1,p2,...,p100]; temperature set T_set=[t1,t2,...,t100]; wind speed set W_set=[w1,w2,...,w100]; radiation set R_set=[r1,r2,...,r100].

[0067] In this embodiment of the invention, ensemble forecast data for precipitation, temperature, wind speed, and radiation are first acquired from multiple ensemble forecast systems. Then, a dynamic weighting method based on information entropy is used to weight and fuse the ensemble forecast data from different sources, resulting in lower weights for forecast sources with greater forecast uncertainty. Finally, the quantile mapping method is applied to the weighted and fused ensemble forecast data for post-processing calibration, yielding a probability set of calibrated meteorological forecast variables. This probability set expands the meteorological forecast from a single deterministic value to multiple equally likely scenarios, with each member representing a possible future meteorological evolution path. The dispersion of the set quantitatively expresses the forecast uncertainty, and the quantile mapping method corrects for systematic biases. This probability set serves as a crucial input for the Monte Carlo simulation in step S500, providing meteorological driving conditions that conform to the characteristics of true uncertainty for the probability distribution of soil moisture content during the prediction period.

[0068] S500: Integrate the posterior probability distribution of the key soil hydraulic parameters, the probability distribution field of the initial soil moisture content, and the probability set of the meteorological forecast variables to drive the digital twin model to perform Monte Carlo simulation and obtain the probability distribution of soil moisture content during the prediction period.

[0069] In this embodiment of the invention, the input of the digital twin model includes three types of uncertainty sources: key soil hydraulic parameters have posterior probability distributions; initial soil moisture content has a spatial probability distribution field; and meteorological forecast variables have a probability set. If the model is run once using only a single estimate of the three types of inputs, a single predicted soil moisture content value is obtained, completely ignoring the joint uncertainty of the inputs, and the prediction result lacks credibility. Monte Carlo simulation, by randomly sampling a large number of samples from the probability distributions of the three types of inputs, combines them into multiple complete sets of model input conditions, and drives the digital twin model to run multiple times, obtaining multiple sets of predicted soil moisture content value sequences. Statistical analysis of multiple sets of predicted values ​​allows estimation of the probability distribution of soil moisture content at each moment within the prediction period. This probability distribution quantifies the transmission effect of input uncertainty to the predicted output, providing a foundation for subsequent calculation of water shortage probability.

[0070] Step S500 in the method provided in this embodiment of the invention includes: The first set of samples is drawn from the posterior probability distribution of the key soil hydraulic parameters; the second set of samples is drawn from the probability distribution field of the initial soil moisture content; and the third set of samples is drawn from the probability set of the weather forecast variables. The first group of samples, the second group of samples, and the third group of samples are combined into complete model input conditions; The digital twin model is driven to run once with the complete model input conditions, and a set of predicted soil moisture content values ​​are output. Repeat the above soil moisture content prediction process to obtain multiple sets of predicted soil moisture content values; Statistical analysis was performed on the multiple sets of predicted soil moisture content values ​​to estimate the probability distribution of soil moisture content within the prediction period.

[0071] First, a first set of samples is drawn from the posterior probability distribution of the key soil hydraulic parameters; a second set of samples is drawn from the probability distribution field of the initial soil moisture content; and a third set of samples is drawn from the probability set of the weather forecast variables.

[0072] The posterior probability distribution of key soil hydraulic parameters refers to the sample set of each parameter obtained through Bayesian inference in step S200. Each sample contains a set of parameter values ​​such as K_s, θ_fc, θ_wp, and n. The first set of samples refers to a combination of parameters randomly selected from this sample set.

[0073] The probability distribution field of initial soil moisture content refers to the mean and variance arrays generated in step S300, covering every grid node in the entire field, with the moisture content of each grid node following a normal distribution. The second set of samples refers to an initial moisture content value obtained by independently and randomly sampling each spatial grid node from this probability distribution field, thus forming a complete spatial distribution map of initial moisture content.

[0074] The probability set of meteorological forecast variables refers to the set of precipitation, temperature, wind speed, and radiation after calibration in the S400 step. The third set of samples refers to randomly selecting one member from each variable's probability set, such as selecting the i-th member from the precipitation set, the j-th member from the temperature set, etc., to form a complete meteorological driving time series.

[0075] First, set the total number of Monte Carlo simulations to N, for example, N=1000. For the k-th simulation: use a random number generator to draw a sample from the posterior distribution sample set without replacement or randomly. If the sample size is insufficient, repeat the sampling. This sample is denoted as the first sample θ_k. For example, set the total number of Monte Carlo simulations to N=1000. For the first simulation (k=1): randomly select the 1203rd sample from the 25000 posterior samples obtained from S200. Its parameters are: saturated hydraulic conductivity K_s=2.2cm / h, field capacity θ_fc=0.27, wilting coefficient θ_wp=0.11, and van Gnoherten shape parameter n=1.46. This sample is designated as the first sample θ_1.

[0076] Secondly, for the initial moisture content distribution field, all spatial grid nodes are traversed. At each node, a random value is independently generated using a normal distribution random number generator, with the predicted mean as the mean and the predicted variance as the variance. The random values ​​of all nodes are combined into a two-dimensional or three-dimensional array, denoted as the second sample S_k. For example, for the initial moisture content probability distribution field, each node is traversed. Taking the plane coordinates (100, 50) and depth 20cm as an example, the predicted mean of this node is 27.5%, and the predicted variance is 6.2. A random value is generated using a normal distribution, for example, 27.9%. Taking the plane coordinates (100, 50) and depth 60cm as an example, the predicted mean is 25.3%, and the predicted variance is 8.1. A random value is generated as 24.8%. All 6027 nodes are sampled independently to obtain a complete initial moisture content spatial distribution map S_1.

[0077] Secondly, for meteorological forecast variables, one member is randomly selected from the precipitation probability set, one member is independently and randomly selected from the temperature probability set, one member is independently and randomly selected from the wind speed probability set, and one member is independently and randomly selected from the radiation probability set, forming a 168-hour hourly meteorological time series, denoted as the third sample M_k. During the sampling process, the precipitation, temperature, wind speed, and radiation sets are sampled independently to reflect the possible independence between the uncertainties of each meteorological variable. For example, for meteorological variables, the 34th member is randomly selected from the precipitation probability set, which provides the hourly precipitation series for the next 168 hours, such as 0.0 mm in the 1st hour, 0.1 mm in the 2nd hour, ..., 0.0 mm in the 168th hour; the 72nd member is independently and randomly selected from the temperature set to obtain the hourly temperature series; the 15th member is independently and randomly selected from the wind speed set; and the 89th member is independently and randomly selected from the radiation set. The four sequences were aligned by time and combined into a meteorological driving time series M_1.

[0078] Then, the first set of samples, the second set of samples, and the third set of samples are combined to form the complete model input conditions. Complete model input conditions refer to the entire set of input data required to uniquely drive a digital twin model to perform a complete simulation. The first set of samples, the second set of samples, and the third set of samples together constitute the basic conditions for a simulation.

[0079] Specifically, the first set of samples θ_k, the second set S_k, and the third set M_k are stored in a data structure, such as a dictionary object, with keys of soil_hydraulic_params, initial_water_content, and meteorology, respectively. Ensure the data format is completely consistent with the input interface defined in the digital twin model. For S_k, it needs to be organized according to the spatial discretization grid of the digital twin model, with spatial interpolation resampling performed if necessary. For M_k, a sequence needs to be provided according to the time step required by the model. After combination, this input condition is passed to the running function of the digital twin model.

[0080] For example, continuing from the first simulation. The first set of samples θ_1 = (K_s = 2.2 cm / h, θ_fc = 0.27, θ_wp = 0.11, n = 1.46) is converted into a parameter input file recognizable by the digital twin model. The initial moisture content values ​​of the second set of samples S_1 (6027 grid nodes) are reorganized into a three-dimensional array according to the soil profile stratification and planar grid used in the digital twin model. The first dimension has 41 nodes in the x-direction, the second dimension has 21 nodes in the y-direction, and the third dimension has 7 nodes in the depth direction. The array values ​​are moisture content percentages. The hourly sequences of precipitation, temperature, wind speed, and radiation in the third set of samples M_1 are converted into the meteorological input format required by the model, such as hourly precipitation, average temperature, average wind speed, and average radiation. These three parts are saved as a complete input condition folder or a single input file, named sim_001_input.

[0081] Next, the digital twin model is run once using the complete set of model input conditions, outputting a set of predicted soil moisture content values. A digital twin model is a soil moisture transport simulation model built based on physical mechanisms. It can simulate the change of soil moisture content over time at different depths and locations within a given prediction period, based on given soil hydraulic parameters, initial moisture content distribution, and meteorological driving data. Running the model once means feeding a complete set of model input conditions into the model for calculation, and the model outputs the soil moisture content distribution at each output time point within the prediction period. A set of predicted soil moisture content values ​​refers to the total output results generated in one simulation, typically including moisture content values ​​for each time step and each spatial location.

[0082] Specifically, the executable program or function interface of the digital twin model is called, passing in the combined input conditions. The model performs numerical solutions, starting from the initial time t0, iteratively calculating the soil moisture redistribution and evapotranspiration processes at each time step according to the meteorological-driven time step. For each output time, the model outputs the soil moisture content across the entire spatial grid. The target soil layer at each output time is recorded, such as the average moisture content of the main root layer (0-60cm) or the moisture content at a specified depth. The output results are stored in array form, for example, for a prediction period of T time steps, recording an average moisture content value or a spatial distribution field at each time step. The resulting set of predicted values ​​is the result Y_k of the k-th simulation.

[0083] For example, continue from the first simulation. Call the HYDRUS-1D model. Input the complete input conditions for the first simulation. The model runs to calculate for 168 hours from 8:00 AM on April 10, 2026 to 8:00 AM on April 17, 2026, with a time step of 1 hour. Output the average volumetric water content of the 0-60cm soil layer at each hour. Obtain the output sequence Y_1: water content 27.5% in the first hour, 27.3% in the second hour, ..., 24.8% in the 168th hour. Store this sequence as an array Y_1

[168] .

[0084] Then, the soil moisture content prediction process described above is repeated to obtain multiple sets of predicted soil moisture content values. In each cycle, the first, second, and third sets of samples are independently extracted, combined, and used to drive the digital twin model to obtain a new sequence of predicted soil moisture content values. Multiple sets of predicted values ​​refer to collecting the output sequences of all N simulations together to form an N-row, T-column matrix, where N is the number of simulations and T is the number of output times within the prediction period.

[0085] Specifically, set the loop counter k from 1 to N. In each loop, randomly select three sets of samples, combine the inputs, run the model, and store the output. Note that the samples selected each time should be independent of each other, that is, the current selection should not affect the probability distribution of the next selection. To avoid the periodicity problem of pseudo-random numbers, a Mason twisting algorithm random number generator can be used, with different random seeds set. Append all output results Y_k (k=1..N) to a three-dimensional array or list with dimensions [N,T] or more complex dimensions. Record the input sample index for each simulation for subsequent analysis.

[0086] For example, N=1000 simulations are set. After the first simulation is completed, a second simulation is performed: another set of parameters θ_2 is randomly drawn from the posterior sample; S_2 is obtained by independently sampling again from the initial water content probability distribution field; precipitation, temperature, wind speed, and radiation members are independently drawn from the meteorological probability set and combined into M_2; the model is driven to obtain the Y_2 sequence. This process is repeated until 1000 simulations are completed. Finally, 1000 sets of output sequences are obtained, each containing 168 hourly water content values. All outputs are stored in a 1000-row, 168-column matrix Y, where the k-th row and t-th column represents the root layer average volumetric water content in the t-th hour of the k-th simulation.

[0087] Finally, statistical analysis is performed on the multiple sets of predicted soil moisture content values ​​to estimate the probability distribution of soil moisture content within the prediction period. Statistical analysis refers to calculating the empirical probability distribution characteristics at a given moment for the N sets of predicted values ​​generated by the Monte Carlo simulation. Estimating the probability distribution of soil moisture content within the prediction period means that, for each prediction moment t, based on N simulation results, the cumulative distribution function or probability density function of soil moisture content at that moment is obtained. Commonly used parameters include: mean, standard deviation, quantiles, and the probability of exceeding or falling below a certain threshold.

[0088] Specifically, for each time point t, the moisture content values ​​at that time point are extracted from N simulations, forming a one-dimensional array moisture content_t = [Y_1[t], Y_2[t], ..., Y_N[t]]. The mean and standard deviation of this array are calculated. Quantities are calculated: the moisture content_t is sorted in ascending order, and the p-th percentile is the value at index p×(N-1) / 100+1 after sorting. Commonly used quantiles are the 10th, 25th, 50th, 75th, and 90th percentiles. Additionally, if drought risk needs to be assessed, the probability of moisture content being lower than the wilting coefficient can be calculated: P_low_t = (number of moisture contents less than 7.2% in_t) / N. The statistical results from all times are organized into a time series format to obtain a probability distribution description of soil moisture content within the prediction period.

[0089] For example, consider 1000 simulations. Take 1000 moisture content values ​​from the 24th hour, for example, a minimum of 22.5% and a maximum of 29.8%. Calculate the mean to be 25.6% and the standard deviation to be 1.8%. Sort the 1000 values; the 10th percentile is 23.4%, the 50th percentile is 25.5%, and the 90th percentile is 27.8%. The wilting coefficient is 7.2%, and since all 1000 values ​​are greater than 7.2%, the probability of water shortage is 0. Next, take 1000 values ​​from the 168th hour; the mean is 24.2%, the standard deviation is 2.1%, the 10th percentile is 21.5%, the 50th percentile is 24.0%, and the 90th percentile is 26.5%. If any individual simulated value is lower than 7.2%, calculate the proportion to obtain the probability of water shortage. Plot the mean and quantiles of all 168 time points into a curve to obtain the probability distribution of soil moisture content during the prediction period. Store this probability distribution in the form of a table or data file for use in the water shortage probability calculation in step S600.

[0090] In this embodiment of the invention, multi-source probabilistic input integration and Monte Carlo simulation of a digital twin model were completed. First, a first set of samples was extracted from the posterior probability distribution of key soil hydraulic parameters, a second set from the probability distribution field of initial soil moisture content, and a third set from the probability set of weather forecast variables. Then, the three sets of samples were combined into complete model input conditions. This drove the digital twin model to run once, outputting a set of predicted soil moisture content values. The above process was repeated to obtain multiple sets of predicted values. Finally, statistical analysis was performed on the multiple sets of predicted values ​​to estimate the probability distribution of soil moisture content at each moment within the prediction period. This probability distribution quantitatively expresses the uncertainty in soil moisture prediction caused by the heterogeneity of soil parameters, the spatial variability of initial moisture content, and the uncertainty of weather forecasts, providing a scientific basis for calculating the probability of water shortage and avoiding decision-making errors that may result from a single deterministic prediction.

[0091] S600: Calculate the probability of water shortage based on the probability distribution of soil moisture content during the predicted period.

[0092] In dryland farming, crop water shortage status is typically determined by whether soil moisture content falls below a certain critical threshold. Traditional deterministic methods compare a single predicted value to a threshold; if the mean is higher than the threshold, water shortage is considered not an issue, and if the mean is lower, water shortage is considered. Traditional methods ignore the uncertainty distribution of predicted moisture content: even if the mean is higher than the threshold, there is still a probability that the actual moisture content is lower; conversely, even if the mean is lower than the threshold, there is still a probability that the actual moisture content is higher. Step S500 has obtained the probability distribution of soil moisture content at each moment within the prediction period, which fully describes the uncertainty of future moisture content values. The probability of water shortage refers to the probability that, under this probability distribution, the soil moisture content at a certain moment or within a certain period within the prediction period will fall below the crop water shortage critical value. Calculating the probability of water shortage directly transforms prediction uncertainty into a decision risk indicator, providing a quantitative basis for subsequent steps to trigger irrigation based on user risk preferences.

[0093] Specifically, obtain the soil moisture content sample set for each time t within the prediction period obtained from the Monte Carlo simulation in step S500: moisture content _t=[Y_1[t],Y_2[t],...,Y_N[t]], where N is the total number of Monte Carlo simulations. Set a crop water shortage critical value θ_critical, which is usually taken as the lower limit of soil moisture allowed by the crop, such as the wilting coefficient θ_wp or a certain percentage of field capacity, such as 50% field capacity. For each time t, calculate the number of samples in the sample set whose soil moisture content is lower than θ_critical, denoted as N_low_t. Then the water shortage probability at that time is _t=N_low_t / N. If it is necessary to calculate the water shortage probability for several consecutive hours, a water shortage event can be defined as at least one time within the prediction period where the moisture content is lower than θ_critical or the average moisture content is lower than θ_critical, selected according to actual irrigation needs. A common calculation method is to take the maximum value of the water shortage probability at each time point within the prediction period, or to calculate a weighted average water shortage probability. The calculated water shortage probability is then output in time series format.

[0094] For example, let's set the crop water shortage critical value θ_critical for potatoes during tuber formation as the wilting coefficient θ_wp = 7.2%. Taking 1000 moisture content samples at the 24th hour, the number of samples below 7.2% is 0, therefore the water shortage probability at the 24th hour = 0 / 1000 = 0%. Taking 1000 samples at the 168th hour, 35 samples are found to be below 7.2%, so the water shortage probability at the 168th hour = 35 / 1000 = 3.5%. Taking the moment with the highest water shortage probability in the next 168 hours, if 120 samples at the 68th hour are below 7.2%, the maximum water shortage probability is 12%. If a water shortage event is defined as a moisture content below 7.2% at any time within the next 168 hours, then the water shortage probability = the number of simulations in 1000 simulations where at least one sample has a moisture content below 7.2% divided by 1000. Statistical analysis shows that in 380 out of 1000 simulations, the water shortage probability fell below 7.2% at some point, indicating a water shortage probability of 38%. This water shortage probability value is then passed as output to step S700. If hourly output is required, a 168-hour water shortage probability time series is generated, for example, 0% for the first hour, 0% for the second hour, ..., 12% for the 68th hour, ..., 3.5% for the 168th hour.

[0095] In this embodiment of the invention, the probability of water shortage is calculated based on the probability distribution of soil moisture content within the predicted time period obtained in step S500. This probability quantifies the likelihood of future crop water stress and is a risk indicator ranging from 0% to 100%. Compared to traditional methods that rely solely on a single predicted value and threshold for rigid judgment, the probability of water shortage fully considers the impact of the triple uncertainties of soil hydraulic parameters, spatial variability of initial moisture content, and weather forecasts on the prediction results, reflecting the true risk level of moisture content potentially falling below the critical value. This indicator provides a direct and comparable basis for determining the water shortage probability threshold in step S700 by combining user risk preferences, upgrading irrigation decisions from a deterministic "irrigate or not irrigate" approach to a risk-driven assessment of whether the risk is acceptable, thus improving the scientific nature of the decision and its robustness to uncertainty.

[0096] S700: Read the user-configured risk preference mode and determine the water shortage probability threshold based on the risk preference mode.

[0097] In actual irrigation, different users have varying risk tolerance for irrigation decisions: conservative users tend to irrigate earlier to avoid crop drought, even if it may waste a small amount of water; risk-taking users tend to reduce the frequency of irrigation to conserve water and are willing to bear some risk of crop water shortage; balanced users fall somewhere in between. The probability of water shortage is a value ranging from 0% to 100%, which itself does not directly trigger irrigation but needs to be compared with a decision threshold. If the probability of water shortage exceeds the threshold, irrigation is triggered; otherwise, it is not. The size of this threshold directly determines the frequency and risk level of irrigation decisions. Furthermore, crops have different sensitivities to water at different growth stages. For example, the water sensitivity coefficient is high during tuber formation and tuber enlargement, and low during maturity. The same user may want to adopt different risk preference modes at different stages. Therefore, it is necessary to read the user-configured risk preference mode, determine the water shortage probability threshold according to preset rules or user-defined values, and support automatic switching according to growth stages.

[0098] Step S700 in the method provided in this embodiment of the invention includes: Read the user's configured risk preference mode, which is selected from conservative mode, balanced mode or risk-taking mode; When the user's risk preference mode is configured as conservative, the water shortage probability threshold is set to a value lower than the water shortage probability threshold corresponding to the balanced mode. When the user's risk preference mode is set to the risk-taking mode, the water shortage probability threshold is set to a value greater than the water shortage probability threshold corresponding to the balanced mode. When the user's risk preference mode is set to the balanced mode, the water shortage probability threshold is set to a value that is greater than the water shortage probability threshold corresponding to the conservative mode and less than the water shortage probability threshold corresponding to the risky mode. Receive user's custom input on the water shortage probability threshold, and use the custom input to override the preset water shortage probability threshold; It receives risk preference modes configured by users for different growth stages of crops and automatically switches the corresponding water shortage probability threshold for different growth stages.

[0099] First, the user's configured risk preference mode is read. This mode is selected from conservative, balanced, or risk-taking modes. The risk preference mode refers to the user's pre-defined strategy type regarding their acceptance of water shortage risk in irrigation decisions. A conservative mode indicates a user's preference for early irrigation to reduce water shortage risk, corresponding to a lower water shortage probability threshold. A risk-taking mode indicates a user's preference for conserving water and accepting some water shortage risk, corresponding to a higher water shortage probability threshold. A balanced mode falls between the two. User configuration refers to the mode options pre-selected through the human-computer interaction interface. The read operation is used to obtain the user's currently selected mode value.

[0100] Specifically, the risk preference mode identifier is stored in the user configuration file; for example, conservative, balanced, and adventurous correspond to conservative, balanced, and adventurous risk preferences, respectively. A configuration read function is set up to retrieve the user's most recently set mode value from a local database, parameter file, or network interface. If the read fails or the user has never configured a mode, the default mode is set to balanced. The retrieved mode value is stored in a memory variable for use in subsequent steps. Dynamic modification at runtime is supported: after the user modifies the mode, the configuration read function immediately returns the new value.

[0101] Secondly, when the user's risk preference mode is set to conservative mode, the water shortage probability threshold is set to a value lower than the water shortage probability threshold corresponding to balanced mode; when the user's risk preference mode is set to risky mode, the water shortage probability threshold is set to a value higher than the water shortage probability threshold corresponding to balanced mode; when the user's risk preference mode is set to balanced mode, the water shortage probability threshold is set to a value higher than the water shortage probability threshold corresponding to conservative mode and lower than the water shortage probability threshold corresponding to risky mode.

[0102] Specifically, irrigation is triggered when the probability of water shortage is greater than or equal to the water shortage probability threshold; it is not triggered when the probability is less than the threshold. The preset water shortage probability threshold is set as follows: a lower threshold corresponds to the conservative mode, a higher threshold to the risky mode, and a threshold in between for the balanced mode. Specific values ​​can be preset based on crop type, soil characteristics, and local water resource availability. Preset baseline values ​​are, for example: 20% for the conservative mode, 40% for the balanced mode, and 60% for the risky mode. These values ​​are merely examples and can be adjusted in actual use.

[0103] For example, the preset water shortage probability thresholds for three modes are: conservative 20%, balanced 40%, and risky 60%. The read risk preference mode is balanced, so the water shortage probability threshold is 40%. If the user switches to conservative mode, the water shortage probability threshold is 20%; if they switch to risky mode, the water shortage probability threshold is 60%.

[0104] Furthermore, the system receives user-defined input for a water shortage probability threshold, which overrides the preset threshold. Custom input refers to users actively typing or swiping to select a specific percentage value on the interactive interface if they are not satisfied with the preset threshold. Overriding means replacing the preset water shortage probability threshold with the new value input by the user. This function provides users with more precise control, allowing them to adjust the sensitivity of their decisions based on personal experience or the current year's water conditions.

[0105] Specifically, a custom threshold input field is added to the user configuration interface. This field accepts integers or floating-point numbers between 0 and 100. An "Enable Custom Threshold" checkbox is also enabled. When the user checks the "Enable Custom Threshold" checkbox and enters a valid value, the configuration read function prioritizes returning the custom value, ignoring the preset water shortage probability threshold. If the user unchecks the checkbox or deletes the custom input, the preset water shortage probability threshold is reverted to being used. The custom threshold can be stored simultaneously with the risk preference mode, but it has higher priority.

[0106] For example, after a user selects the balanced mode, the preset water shortage probability threshold is 40%. The user believes that 40% is still too high and is worried about crop drought, so they enter 25% in the custom input box and check "Enable custom threshold". The configuration read function detects that the custom threshold is valid and overwrites the water shortage probability threshold with 25%. At this point, although the risk preference mode is displayed as balanced, the actual water shortage probability threshold used is 25%. If the user unchecks the box, the water shortage probability threshold reverts to 40%.

[0107] Finally, the system receives the risk preference modes configured by the user for different crop growth stages and automatically switches the corresponding water shortage probability threshold for each stage. Different growth stages have varying sensitivities to water stress: the water shortage probability threshold should be lowered in sensitive stages and raised in insensitive stages. Users can configure a risk preference mode individually for each stage or directly configure the water shortage probability threshold for each stage. Automatic switching means that the digital twin model automatically applies the corresponding threshold based on the growth stage calculated by the current date or crop accumulated temperature, eliminating the need for daily manual adjustments by the user.

[0108] The user configuration interface provides a table or list where rows correspond to different crop growth stages, and columns correspond to the risk preference mode or direct threshold for each stage. Users select a mode for each stage, and the configuration is stored as a stage-threshold mapping table, for example, [Stage 1: Conservative, water shortage probability threshold 20%; Stage 2: Balanced, water shortage probability threshold 40%; Stage 3: Risky, water shortage probability threshold 60%] or directly stores the threshold value. When an irrigation decision is executed, the current growth stage is first calculated based on the current date and the crop sowing date. Then, the corresponding water shortage probability threshold is retrieved from the mapping table. If a stage is not configured by the user, the global default value is inherited. As time progresses, when the growth stage changes, the next decision automatically uses the threshold for the new stage.

[0109] For example, potato growth stages are divided into: seedling stage, tuber formation stage, tuber enlargement stage, starch accumulation stage, and maturity stage. Users can set risk preference modes for each stage in the configuration interface: conservative mode for the seedling stage, conservative mode for the tuber formation stage, balanced mode for the tuber enlargement stage, risky mode for the starch accumulation stage, and risky mode for the maturity stage. If the current stage is tuber formation, the conservative mode threshold of 20% is automatically switched to. When the crop enters the tuber enlargement stage, the balanced mode threshold of 40% is automatically switched to on the next decision. If the user has not configured a stage, that stage inherits the global default value of 40%.

[0110] In this embodiment of the invention, the user's risk preference mode and the determination of the water shortage probability threshold are completed. First, the user's configured conservative, balanced, or risky mode is read; then, the corresponding water shortage probability threshold is determined based on the mode; next, the user's custom input is received to override the preset threshold; finally, the user's risk preference mode configured for different crop growth stages is received, and the corresponding water shortage probability threshold is automatically switched at different growth stages. This step returns control of decision-making risk to the user, realizing personalized customization of irrigation strategies. The same digital twin model can serve users with different risk preferences and can adapt to the dynamic changes in water sensitivity during the crop's growth period, improving the flexibility and field adaptability of irrigation decisions.

[0111] S800: When the water shortage probability is greater than or equal to the water shortage probability threshold, an irrigation decision is triggered, and an irrigation instruction is generated based on the quantile corresponding to the risk preference pattern in the posterior distribution of irrigation volume.

[0112] In this embodiment of the invention, when the probability of water shortage reaches or exceeds the water shortage probability threshold, irrigation needs to be performed and the specific irrigation amount needs to be determined. The irrigation amount is not a fixed value, but depends on the future degree of soil moisture deficit and the user's tolerance for the risk of over-irrigation. The posterior distribution of irrigation amount refers to the probability distribution of irrigation amount that can restore soil moisture to the target state, derived backward through a digital twin model under the premise of water shortage conditions. Conservative users tend to irrigate more to ensure crops are not drought-prone and should choose a higher quantile in the posterior distribution of irrigation amount; risk-taking users tend to irrigate less to conserve resources and should choose a lower quantile; balanced users choose a middle quantile. Therefore, it is necessary to select the corresponding quantile from the posterior distribution of irrigation amount according to the risk preference pattern, and send the irrigation amount value corresponding to that quantile as an instruction to the execution device.

[0113] Step S800 in the method provided in this embodiment of the invention includes: When the risk preference mode is a conservative mode, the quantile with a value greater than or equal to the first quantile threshold is selected from the posterior distribution of irrigation volume as the target quantile. When the risk preference mode is a balanced mode, the quantiles whose quantile values ​​are greater than or equal to the second quantile threshold and less than the first quantile threshold are selected from the posterior distribution of irrigation volume as the target quantiles. When the risk preference mode is the risk-taking mode, the quantile with a value less than the second quantile threshold is selected from the posterior distribution of irrigation volume as the target quantile. Wherein, the first quantile threshold is higher than the second quantile threshold; The irrigation amount value corresponding to the target quantile is issued as the irrigation amount command.

[0114] First, a first quantile threshold and a second quantile threshold are pre-set to delineate the quantile intervals of the posterior distribution of irrigation amount. The first quantile threshold is higher than the second quantile threshold. These two thresholds divide the quantile axis of the posterior distribution of irrigation amount into three intervals: the lower quantile interval (below the second quantile threshold), the middle quantile interval (between the second and first quantile thresholds), and the upper quantile interval (above or equal to the first quantile threshold). For example, the second quantile threshold can be set to 30%, and the first quantile threshold to 70%. Specific values ​​can be adjusted according to local irrigation practices and water resource conditions.

[0115] The posterior distribution of irrigation amount refers to the probability distribution of the required irrigation amount obtained through reverse simulation using a digital twin model, under the conditions of the probability distribution of soil moisture content, the crop water shortage threshold, and the user-defined recovery target within a given prediction period. This distribution exists in the form of a sample set or quantiles, such as 1000 possible irrigation amount values. The target quantile refers to a percentile selected from this distribution based on a risk preference pattern.

[0116] First, determine the relationship between the water shortage probability calculated in step S600 and the water shortage probability threshold determined in step S700. If the water shortage probability is less than the water shortage probability threshold, irrigation is not triggered; if the water shortage probability is greater than or equal to the water shortage probability threshold, an irrigation decision is triggered, and the following branch continues to be executed: First, when the risk preference mode is conservative, quantiles with quantile values ​​greater than or equal to the first quantile threshold are selected from the posterior distribution of irrigation volume as target quantiles. The first quantile threshold can be selected itself or a higher value; specifically, it can be set to a fixed quantile corresponding to the first quantile threshold.

[0117] Secondly, when the risk preference pattern is a balanced pattern, the target quantile is selected from the posterior distribution of the water volume where the quantile value is greater than or equal to the second quantile threshold and less than the first quantile threshold. When the risk preference pattern is a balanced pattern, the target quantile is selected from the posterior distribution of the water volume where the quantile value is greater than or equal to the second quantile threshold and less than the first quantile threshold; typically, the median is used.

[0118] Furthermore, when the risk preference mode is the risk-taking mode, the target quantile is selected from the posterior distribution of the flooding volume where the quantile value is less than the second quantile threshold. The quantile corresponding to the second quantile threshold or lower can be selected.

[0119] For example, assume that the probability of water shortage in the next 168 hours calculated in step S600 is 45%, and the water shortage probability threshold determined in step S700 based on the user's balanced model is 40%. If 45% ≥ 40%, an irrigation decision is triggered. First, the posterior distribution of irrigation amount is obtained, which is obtained through reverse simulation using a digital twin model: when the water shortage condition is met, starting from the current soil moisture content distribution, the future soil moisture changes after adding different amounts of irrigation are simulated, and the irrigation amount required to restore the crop to a suitable moisture state is selected. This process is repeated 1000 times to obtain 1000 irrigation amount samples. Assume that the obtained posterior distribution of irrigation amount is: 10th percentile = 18mm, 20th percentile = 22mm, 30th percentile = 25mm, 50th percentile = 30mm, 70th percentile = 35mm, 80th percentile = 38mm, 90th percentile = 42mm. The user's risk preference mode is balanced. Based on the rules, a quantile greater than or equal to the second quantile threshold (30%) and less than the first quantile threshold (70%) is selected. The default selection is the median, i.e., the 50th quantile, corresponding to a water drop volume of 30mm. If the user is in a conservative mode, a quantile greater than or equal to 70% is selected, for example, the 70th quantile corresponds to 35mm; if the user is in an adventurous mode, a quantile less than 30% is selected, for example, the 20th quantile corresponds to 22mm.

[0120] Finally, the irrigation volume value corresponding to the target quantile is issued as an irrigation volume command. An irrigation volume command is a control signal sent to the irrigation execution equipment, typically containing the irrigation volume value and the irrigation start time, duration, or end time. Issuance refers to transmitting the command to the field controller via wired or wireless communication protocols, where the controller executes the actual irrigation operation. Based on the target irrigation volume, combined with the target farmland area and the irrigation system flow rate, the required irrigation duration is calculated: Irrigation duration = (Target irrigation volume × Target farmland area) / Irrigation system flow rate. An irrigation command message is generated and sent to the field controller via message queue telemetry transmission protocol or Hypertext Transfer Security protocol. Simultaneously, log information such as the command issuance time, target quantile, and risk preference mode is recorded.

[0121] For example, the target farmland area is 200m × 100m = 20000m². 2 Target irrigation amount = 30mm, total water requirement = 30mm × 20000m³ 2 =600m 3The irrigation system has a flow rate of 50 m³ / h, so the required irrigation duration is 600 / 50 = 12 hours. An irrigation command is generated: Start irrigation for 12 hours, total water volume 600 cubic meters. This command is sent to the solenoid valve controller in the field via a LoRa wireless module. After receiving the command, the controller opens the valve at the specified time to begin irrigation. Simultaneously, the decision log is recorded as follows: 2026-04-17 08:00:00, water shortage probability 45% ≥ threshold 40%, risk preference balanced, 50th percentile irrigation volume 30 mm, irrigation duration 12 hours, issued.

[0122] In this embodiment of the invention, it is first determined whether the water shortage probability reaches a water shortage probability threshold. If not, irrigation is not triggered; if it does, a corresponding quantile is selected from the posterior distribution of irrigation volume based on the user's risk preference pattern: a high quantile for conservative users, a middle quantile for balanced users, and a low quantile for risk-taking users. Then, the irrigation volume value corresponding to this quantile is converted into a specific irrigation command and sent to the execution device. This step achieves a closed loop from risk perception to risk response: it not only decides whether to irrigate but also determines the irrigation volume based on user preferences, enabling the irrigation volume to adaptively balance the contradiction between yield preservation and water conservation. Conservative users reduce crop drought risk by irrigating more, risk-taking users conserve water resources by irrigating less, and balanced users achieve a compromise between the two. Compared to traditional methods that output a fixed irrigation volume, this method effectively improves the precision of irrigation decisions and user satisfaction.

[0123] The method provided in this embodiment of the invention also includes a safe rollback mechanism: A traditional irrigation rule base based on experience thresholds is pre-built into the system; When the amount of missing sensor data exceeds a preset proportion or the width of the model prediction confidence interval exceeds a preset width threshold, the digital twin model is determined to be unreliable, and the system is downgraded from the digital twin driving mode to the traditional irrigation mode to execute the traditional irrigation rules. Based on the probability distribution of soil moisture content during the predicted period, the predicted soil moisture stress state is determined. When the predicted soil moisture stress exceeds the crop's physiological tolerance limit, the system forcibly switches from the current risk preference mode to a conservative mode and executes irrigation decisions based on the water shortage probability threshold and irrigation volume quantile corresponding to the conservative mode. The relationship between the traditional irrigation mode and the conservative mode is as follows: the conservative mode has a higher priority than the traditional irrigation mode. When both triggering conditions are met simultaneously, the system will execute the conservative mode first.

[0124] In this embodiment of the invention, the digital twin model can provide high-precision probabilistic predictions and irrigation decisions under ideal conditions. However, in actual farmland environments, it may face situations such as sensor failure, communication interruption, insufficient model input data, or model failure due to extreme weather. For example, soil moisture sensors may lose data due to battery depletion, or the dispersion among ensemble forecast members may be too large, causing the prediction confidence interval to exceed the acceptable range. In such cases, the prediction results of the digital twin model are unreliable, and continuing to rely on model-driven decision-making may lead to misjudgments. Furthermore, even if the model is operating normally, when the predicted soil moisture stress exceeds the crop's physiological tolerance limit, the most conservative irrigation strategy must be adopted immediately to save the crop, rather than continuing to follow the user's original risk-taking or equilibrium preferences.

[0125] Therefore, a safe fallback mechanism needs to be established, including two levels of degradation protection: first, when the model is unreliable, the model should be downgraded from the digital twin-driven mode to the traditional irrigation mode; second, when the predicted stress exceeds the crop's tolerance limit, the model should be forcibly switched to the conservative mode, with the latter having higher priority than the former.

[0126] First, a traditional irrigation rule base based on empirical thresholds is pre-built into the system. This traditional irrigation rule base refers to a set of empirical rules that make irrigation decisions based solely on fixed soil moisture thresholds and crop growth stages, without relying on probabilistic predictions from a digital twin model. This rule base serves as a backup plan in case the digital twin model fails. Each rule in the rule base includes a trigger condition and a corresponding irrigation amount. A common form of trigger condition is: irrigation is triggered when the sensor-measured soil moisture content is below a fixed threshold at a certain depth for a certain period of time. The irrigation amount is set as a fixed value based on crop type, growth stage, and local experience. The rule base is stored in tabular form in the system's internal memory.

[0127] Specifically, irrigation experience parameters recommended by local agricultural technology extension departments are collected, or summarized from historical irrigation records. The rule base is compiled into a set of condition-action statements, for example: IF 20cm depth moisture content < 20% AND growth stage IN [tuberous enlargement stage, starch accumulation stage] THEN irrigation amount = 25mm. The rule base can be modified by users through configuration files. The version and update time of the rule base are recorded together.

[0128] Secondly, when the amount of missing sensor data exceeds a preset proportion or the width of the model prediction confidence interval exceeds a preset width threshold, the digital twin model is deemed unreliable, and the system downgrades from the digital twin-driven mode to the traditional irrigation mode, executing traditional irrigation rules. The amount of missing sensor data refers to the percentage of missing records for key variables such as measured soil moisture and measured soil temperature in the multi-source sensing data loaded in step S100, relative to the total number of records to be collected, for example, 30%. The model prediction confidence interval width refers to the difference between the upper and lower limits of a certain confidence level in the probability distribution of soil moisture content within the prediction period obtained in step S500. The preset width threshold can be set to 10% for volumetric moisture content; that is, when the interval width exceeds 10%, the model is considered unreliable. The digital twin-driven mode refers to using the probabilistic decision-making process from S100 to S800. The traditional irrigation mode refers to directly executing rules in the rule base. After downgrading, the system no longer performs Monte Carlo simulation and probabilistic decision-making, but instead directly outputs irrigation instructions based on real-time sensor data or timed rules.

[0129] Specifically, at the beginning of each decision cycle, the integrity of sensor data is first checked: the total number of soil moisture data entries that should have been collected in the past 24 hours is counted, and the ratio with the actual number of entries entered into the database is calculated. The missing proportion is then determined. If the missing proportion is greater than 0.3, a downgrade is triggered. Simultaneously, the width of the 90% prediction confidence interval for each moment in the next 168 hours in step S500 is calculated, which is the difference between the 95th percentile and the 5th percentile, and the maximum width is taken. If the maximum width is greater than 1, and the volumetric water content is 0%, a downgrade is also triggered. If any condition is met, the model is deemed unreliable. At this point, a flag `model_reliable=false` is set, steps S600 to S800 are skipped, and the matching rules in the traditional irrigation rule base are directly called to generate and issue irrigation commands.

[0130] Next, based on the probability distribution of soil moisture content within the predicted time period, the predicted soil moisture stress state is determined. Soil moisture stress state refers to an index of the severity of crop water stress calculated based on the probability distribution of soil moisture content within the predicted time period. When the predicted soil moisture content is lower than the critical soil water potential or effective water content threshold required for a specific stage of crop growth, the crop is determined to be in a water stress state. This determination can be made hourly or based on the worst-case scenario within the predicted time period. The critical threshold is usually taken as the lower limit of soil moisture allowed for the crop, such as 50% of field capacity or the wilting coefficient. The stress state can be quantified into levels such as no stress, mild stress, moderate stress, and severe stress.

[0131] Specifically, from the probability distribution of soil moisture content within the prediction period obtained in step S500, for each time t, the probability P_deficit_t that the moisture content is below the critical threshold θ_critical is calculated, or the 10th percentile of the moisture content is directly calculated. Stress level classification criteria are set: when the 10th percentile ≥ θ_critical + 5%, it is determined to be no stress; when the 10th percentile is in the interval [θ_critical, θ_critical + 5%), it is determined to be mild stress; when the 10th percentile is below θ_critical but above θ_wp, it is determined to be moderate stress; when the 10th percentile is below θ_wp, it is determined to be severe stress. The most severe stress level within the prediction period is output.

[0132] Furthermore, when the predicted soil moisture stress exceeds the crop's physiological tolerance limit, the system forcibly switches from the current risk-averse mode to a conservative mode, and executes irrigation decisions based on the water shortage probability threshold and irrigation quantile corresponding to the conservative mode. The crop's physiological tolerance limit refers to the most severe level of water stress a crop can withstand at a certain growth stage without permanent damage or a significant drop in yield. For example, during the tuber enlargement stage, soil moisture content below the wilting coefficient for more than 24 hours may lead to a yield reduction of more than 20%, which is considered exceeding the tolerance limit. When the determined stress state reaches or exceeds this limit, the system no longer follows the user's original risk-averse or equilibrium preference, but forcibly switches to the conservative mode. The conservative mode corresponds to a lower water shortage probability threshold and a higher irrigation quantile. After the forced switch, the decision logic in steps S700 and S800 is executed according to the conservative mode.

[0133] Specifically, the system configuration presets crop physiological tolerance limits for each growth stage. For example, during the tuber enlargement stage, if the soil moisture content remains below the wilting coefficient for more than 12 hours, it is considered exceeding the tolerance limit. During each decision-making process, based on the prediction results, if the stress state exceeds the preset limit, the internal variable `forced_conservative_flag` is set to true, overriding the user's currently configured `risk_mode`, forcing the use of the conservative mode. Subsequent selections of the water shortage probability threshold and irrigation quantiles use conservative values. This forced switching continues until the current irrigation decision is completed, and is reassessed in the next decision cycle.

[0134] The relationship between the traditional irrigation mode and the conservative mode is as follows: the conservative mode has higher priority than the traditional irrigation mode. When both triggering conditions are met simultaneously, the system prioritizes the conservative mode. This priority means that when the digital twin model is unreliable and the predicted soil moisture stress exceeds the crop's physiological tolerance limit, the system should prioritize the conservative mode over the traditional irrigation mode. This is because, although the conservative mode relies on the digital twin model, the model's uncertainty is masked by the severity of the stress; while the traditional irrigation mode is based on a fixed threshold and may not be able to dynamically adjust the irrigation amount according to the severity of the stress. Prioritizing the conservative mode means that the system does not degrade and still attempts to use the digital twin model, making decisions based on a conservative preference.

[0135] Specifically, a priority judgment logic is set at the beginning of the decision-making process. First, it checks whether the forced switching condition is met. If it is, regardless of the model's reliability, it directly enters the conservative mode decision-making process. Only when the forced switching condition is not met is the digital twin model reliability condition checked; if the digital twin model is unreliable, it enters the traditional irrigation mode. Therefore, when both conditions are met simultaneously, the forced switching condition takes precedence, and the conservative mode is executed.

[0136] In this embodiment of the invention, a safety fallback mechanism is implemented to provide robustness and security under abnormal conditions. First, a pre-built traditional irrigation rule base based on experience thresholds ensures that basic irrigation can be degraded if the digital twin model becomes unreliable due to missing data or overly broad prediction confidence, preventing complete loss of control. Second, the soil moisture stress state is determined based on the predicted probability distribution of soil moisture content. When the stress exceeds the crop's physiological tolerance limit, a forced switch from the user's current risk preference mode to a conservative mode is initiated to ensure crop safety in the safest way. Finally, the conservative mode is explicitly prioritized over the traditional irrigation mode to prevent strategy confusion when multiple anomalies are triggered simultaneously. This mechanism enhances the security and fault tolerance of the digital twin irrigation control system in actual farmland deployment, ensuring the survival baseline of crops under extreme conditions.

[0137] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides a method and system for precision irrigation control of dryland crops based on digital twins. It constructs a complete closed loop of uncertainty quantification and risk-driven irrigation decision-making by loading multi-source sensing data, obtaining the posterior distribution of soil hydraulic parameters through Bayesian inference, generating an initial water content distribution field through probability field modeling, constructing a meteorological probability set, predicting the probability distribution of soil water content through Monte Carlo simulation, calculating the probability of water shortage, determining adaptive thresholds based on user risk preference patterns, and generating differentiated instructions based on the posterior quantiles of irrigation volume. Simultaneously, a safety fallback mechanism is introduced, downgrading to traditional irrigation rules when data is missing or predictions are unreliable, and forcibly switching to a conservative mode when stress exceeds the crop's tolerance limit. This invention explicitly transmits prediction uncertainty to the decision-making layer, allowing users to flexibly adjust irrigation strategies according to their risk tolerance, achieving an adaptive balance between water conservation and yield protection, and effectively improving decision credibility, field adaptability, and system robustness.

[0138] Example 2, as Figure 3 As shown, this invention provides a precision irrigation control system for dryland crops based on digital twins, the system comprising: The multi-source data loading module 11 is used to load multi-source sensing data of the target farmland, wherein the multi-source sensing data includes at least measured soil moisture data and weather forecast data. The Bayesian inference module 12 is used to perform Bayesian inference on key soil hydraulic parameters in the digital twin model based on the measured soil moisture data, and obtain the posterior probability distribution of the key soil hydraulic parameters. The initial probability field modeling module 13 is used to perform probability modeling on the model input conditions of the digital twin model based on the multi-source sensing data, and obtain the probability distribution field of the initial soil moisture content. The meteorological probability set acquisition module 14 is used to acquire the probability set of meteorological forecast variables based on the meteorological forecast data. The Monte Carlo simulation module 15 is used to integrate the posterior probability distribution of the key soil hydraulic parameters, the probability distribution field of the initial soil moisture content, and the probability set of the meteorological forecast variables to drive the digital twin model to perform Monte Carlo simulation and obtain the probability distribution of soil moisture content during the prediction period. The water shortage probability calculation module 16 is used to calculate the water shortage probability based on the probability distribution of soil moisture content during the prediction period. Risk threshold determination module 17 is used to read the user-configured risk preference mode and determine the water shortage probability threshold according to the risk preference mode; The irrigation instruction generation module 18 is used to trigger an irrigation decision when the water shortage probability is greater than or equal to the water shortage probability threshold, and to generate an irrigation instruction based on the quantile corresponding to the risk preference pattern in the posterior distribution of irrigation volume.

[0139] In one embodiment, the multi-source data loading module 11 is further configured to: Load measured soil temperature data, measured soil salinity data, leaf area index and canopy temperature from satellite remote sensing imagery, as well as soil property data and crop variety data of the target farmland.

[0140] In one embodiment, the Bayesian inference module 12 is further configured to: Set prior probability distributions for each key soil hydraulic parameter; The time series of the measured soil moisture data is used as conditional information; The posterior distribution is updated batch by batch using the differential evolution Markov chain Monte Carlo algorithm. It also automatically triggers a post-hoc update whenever new data is received.

[0141] In one embodiment, the initial probability field modeling module 13 is further configured to: Gaussian process regression was used, with the coordinates and depth of discrete sampling points of the sensor as input features and the measured soil moisture content at the corresponding points as the target variable, to train the Gaussian process regression model; Using the trained Gaussian process regression model, predictions are made for any unsampled location in the target farmland, and the predicted mean and variance of soil moisture content at the corresponding location are output. The predicted mean and predicted variance of all unsampled locations are combined to generate a probability distribution field of the initial soil moisture content covering the entire field area.

[0142] In one embodiment, the meteorological probability set acquisition module 14 is further configured to: Accompanied forecast data for precipitation, temperature, wind speed, and radiation are obtained from ensemble forecast systems of multiple sources in the aforementioned meteorological forecast data. We employ dynamic weights based on information entropy to perform weighted fusion of ensemble forecast data from different forecast sources, where forecast sources with greater forecast uncertainty have lower weights. The quantile mapping method is used to perform post-processing calibration on the weighted and fused ensemble forecast data to obtain the probability set of the calibrated meteorological forecast variables.

[0143] In one embodiment, the Monte Carlo simulation module 15 is also used for: The first set of samples is drawn from the posterior probability distribution of the key soil hydraulic parameters; the second set of samples is drawn from the probability distribution field of the initial soil moisture content; and the third set of samples is drawn from the probability set of the weather forecast variables. The first group of samples, the second group of samples, and the third group of samples are combined into complete model input conditions; The digital twin model is driven to run once with the complete model input conditions, and a set of predicted soil moisture content values ​​are output. Repeat the above soil moisture content prediction process to obtain multiple sets of predicted soil moisture content values; Statistical analysis was performed on the multiple sets of predicted soil moisture content values ​​to estimate the probability distribution of soil moisture content within the prediction period.

[0144] In one embodiment, the risk threshold determination module 17 is further configured to: Read the user's configured risk preference mode, which is selected from conservative mode, balanced mode or risk-taking mode; When the user's risk preference mode is configured as conservative, the water shortage probability threshold is set to a value lower than the water shortage probability threshold corresponding to the balanced mode. When the user's risk preference mode is set to the risk-taking mode, the water shortage probability threshold is set to a value greater than the water shortage probability threshold corresponding to the balanced mode. When the user's risk preference mode is set to the balanced mode, the water shortage probability threshold is set to a value that is greater than the water shortage probability threshold corresponding to the conservative mode and less than the water shortage probability threshold corresponding to the risky mode. Receive user's custom input on the water shortage probability threshold, and use the custom input to override the preset water shortage probability threshold; It receives risk preference modes configured by users for different growth stages of crops and automatically switches the corresponding water shortage probability threshold for different growth stages.

[0145] In one embodiment, the irrigation instruction generation module 18 is further configured to: When the risk preference mode is a conservative mode, the quantile with a value greater than or equal to the first quantile threshold is selected from the posterior distribution of irrigation volume as the target quantile. When the risk preference mode is a balanced mode, the quantiles whose quantile values ​​are greater than or equal to the second quantile threshold and less than the first quantile threshold are selected from the posterior distribution of irrigation volume as the target quantiles. When the risk preference mode is the risk-taking mode, the quantile with a value less than the second quantile threshold is selected from the posterior distribution of irrigation volume as the target quantile. Wherein, the first quantile threshold is higher than the second quantile threshold; The irrigation amount value corresponding to the target quantile is issued as the irrigation amount command.

[0146] The system provided in this embodiment of the invention also includes a safe rollback mechanism: A traditional irrigation rule base based on experience thresholds is pre-built into the system; When the amount of missing sensor data exceeds a preset proportion or the width of the model prediction confidence interval exceeds a preset width threshold, the digital twin model is determined to be unreliable, and the system is downgraded from the digital twin driving mode to the traditional irrigation mode to execute the traditional irrigation rules. Based on the probability distribution of soil moisture content during the predicted period, the predicted soil moisture stress state is determined. When the predicted soil moisture stress exceeds the crop's physiological tolerance limit, the system forcibly switches from the current risk preference mode to a conservative mode and executes irrigation decisions based on the water shortage probability threshold and irrigation volume quantile corresponding to the conservative mode. The relationship between the traditional irrigation mode and the conservative mode is as follows: the conservative mode has a higher priority than the traditional irrigation mode. When both triggering conditions are met simultaneously, the system will execute the conservative mode first.

[0147] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for precision irrigation control of dryland crops based on digital twins, characterized in that, include: Load multi-source sensing data of the target farmland, wherein the multi-source sensing data includes at least measured soil moisture data and weather forecast data; Based on the measured soil moisture data, Bayesian inference is performed on the key soil hydraulic parameters in the digital twin model to obtain the posterior probability distribution of the key soil hydraulic parameters. Based on the multi-source sensing data, the model input conditions of the digital twin model are probabilistically modeled to obtain the probability distribution field of the initial soil moisture content. Based on the meteorological forecast data, obtain the probability set of the meteorological forecast variables; The posterior probability distribution of the key soil hydraulic parameters, the probability distribution field of the initial soil moisture content, and the probability set of the meteorological forecast variables are integrated to drive the digital twin model to perform Monte Carlo simulation and obtain the probability distribution of soil moisture content during the prediction period. The probability of water shortage is calculated based on the probability distribution of soil moisture content during the predicted period; Read the user-configured risk preference pattern and determine the water shortage probability threshold based on the risk preference pattern; When the water shortage probability is greater than or equal to the water shortage probability threshold, an irrigation decision is triggered, and an irrigation instruction is generated based on the quantile corresponding to the risk preference pattern in the posterior distribution of irrigation volume.

2. The method as described in claim 1, characterized in that, Loading multi-source sensing data of the target farmland also includes: Load measured soil temperature data, measured soil salinity data, leaf area index and canopy temperature from satellite remote sensing imagery, as well as soil property data and crop variety data of the target farmland.

3. The method as described in claim 1, characterized in that, Perform Bayesian inference on key soil hydraulic parameters to obtain their posterior probability distributions, including: Set prior probability distributions for each key soil hydraulic parameter; The time series of the measured soil moisture data is used as conditional information; The posterior distribution is updated batch by batch using the differential evolution Markov chain Monte Carlo algorithm. It also automatically triggers a post-hoc update whenever new data is received.

4. The method as described in claim 1, characterized in that, Probabilistic modeling is performed on the model input conditions to obtain the probability distribution field of the initial soil moisture content, including: Gaussian process regression was used, with the coordinates and depth of discrete sampling points of the sensor as input features and the measured soil moisture content at the corresponding points as the target variable, to train the Gaussian process regression model; Using the trained Gaussian process regression model, predictions are made for any unsampled location in the target farmland, and the predicted mean and variance of soil moisture content at the corresponding location are output. The predicted mean and predicted variance of all unsampled locations are combined to generate a probability distribution field of the initial soil moisture content covering the entire field area.

5. The method as described in claim 1, characterized in that, Obtain the probability set of weather forecast variables, including: Accompanied forecast data for precipitation, temperature, wind speed, and radiation are obtained from ensemble forecast systems of multiple sources in the aforementioned meteorological forecast data. We employ dynamic weights based on information entropy to perform weighted fusion of ensemble forecast data from different forecast sources, where forecast sources with greater forecast uncertainty have lower weights. The quantile mapping method is used to perform post-processing calibration on the weighted and fused ensemble forecast data to obtain the probability set of the calibrated meteorological forecast variables.

6. The method as described in claim 1, characterized in that, The digital twin model is driven to perform Monte Carlo simulations to obtain the probability distribution of soil moisture content within the predicted time period, including: The first set of samples is drawn from the posterior probability distribution of the key soil hydraulic parameters; the second set of samples is drawn from the probability distribution field of the initial soil moisture content; and the third set of samples is drawn from the probability set of the weather forecast variables. The first group of samples, the second group of samples, and the third group of samples are combined into complete model input conditions; The digital twin model is driven to run once with the complete model input conditions, and a set of predicted soil moisture content values ​​are output. Repeat the above soil moisture content prediction process to obtain multiple sets of predicted soil moisture content values; Statistical analysis was performed on the multiple sets of predicted soil moisture content values ​​to estimate the probability distribution of soil moisture content within the prediction period.

7. The method as described in claim 1, characterized in that, Read the user-configured risk preference pattern and determine the water shortage probability threshold based on the risk preference pattern, including: Read the user's configured risk preference mode, which is selected from conservative mode, balanced mode or risk-taking mode; When the user's risk preference mode is configured as conservative, the water shortage probability threshold is set to a value lower than the water shortage probability threshold corresponding to the balanced mode. When the user's risk preference mode is set to the risk-taking mode, the water shortage probability threshold is set to a value greater than the water shortage probability threshold corresponding to the balanced mode. When the user's risk preference mode is set to the balanced mode, the water shortage probability threshold is set to a value that is greater than the water shortage probability threshold corresponding to the conservative mode and less than the water shortage probability threshold corresponding to the risky mode. Receive user's custom input on the water shortage probability threshold, and use the custom input to override the preset water shortage probability threshold; It receives risk preference modes configured by users for different growth stages of crops and automatically switches the corresponding water shortage probability threshold for different growth stages.

8. The method as described in claim 1, characterized in that, Based on the quantiles in the posterior distribution of irrigation volume that correspond to the risk preference pattern, an irrigation volume instruction is generated, including: When the risk preference mode is a conservative mode, the quantile with a value greater than or equal to the first quantile threshold is selected from the posterior distribution of irrigation volume as the target quantile. When the risk preference mode is a balanced mode, the quantiles whose quantile values ​​are greater than or equal to the second quantile threshold and less than the first quantile threshold are selected from the posterior distribution of irrigation volume as the target quantiles. When the risk preference mode is the risk-taking mode, the quantile with a value less than the second quantile threshold is selected from the posterior distribution of irrigation volume as the target quantile. Wherein, the first quantile threshold is higher than the second quantile threshold; The irrigation amount value corresponding to the target quantile is issued as the irrigation amount command.

9. The method as described in claim 1, characterized in that, It also includes a safety rollback mechanism: A traditional irrigation rule base based on experience thresholds is pre-built into the system; When the amount of missing sensor data exceeds a preset proportion or the width of the model prediction confidence interval exceeds a preset width threshold, the digital twin model is determined to be unreliable, and the system is downgraded from the digital twin driving mode to the traditional irrigation mode to execute the traditional irrigation rules. Based on the probability distribution of soil moisture content during the predicted period, the predicted soil moisture stress state is determined. When the predicted soil moisture stress exceeds the crop's physiological tolerance limit, the system forcibly switches from the current risk preference mode to a conservative mode and executes irrigation decisions based on the water shortage probability threshold and irrigation volume quantile corresponding to the conservative mode. The relationship between the traditional irrigation mode and the conservative mode is as follows: the conservative mode has a higher priority than the traditional irrigation mode. When both triggering conditions are met simultaneously, the system will execute the conservative mode first.

10. A precision irrigation control system for dryland crops based on digital twins, characterized in that, The method for implementing the precision irrigation control of dryland crops based on digital twins as described in any one of claims 1-9 includes: A multi-source data loading module is used to load multi-source sensing data of the target farmland, wherein the multi-source sensing data includes at least measured soil moisture data and weather forecast data. The Bayesian inference module is used to perform Bayesian inference on key soil hydraulic parameters in the digital twin model based on the measured soil moisture data, and obtain the posterior probability distribution of the key soil hydraulic parameters. The initial probability field modeling module is used to perform probability modeling on the model input conditions of the digital twin model based on the multi-source sensing data, and obtain the probability distribution field of the initial soil moisture content. The meteorological probability set acquisition module is used to acquire the probability set of meteorological forecast variables based on the meteorological forecast data. The Monte Carlo simulation module is used to integrate the posterior probability distribution of the key soil hydraulic parameters, the probability distribution field of the initial soil moisture content, and the probability set of the meteorological forecast variables to drive the digital twin model to perform Monte Carlo simulation and obtain the probability distribution of soil moisture content during the prediction period. The water shortage probability calculation module is used to calculate the water shortage probability based on the probability distribution of soil moisture content during the predicted period. The risk threshold determination module is used to read the user-configured risk preference pattern and determine the water shortage probability threshold based on the risk preference pattern. An irrigation instruction generation module is used to trigger an irrigation decision when the water shortage probability is greater than or equal to the water shortage probability threshold, and to generate an irrigation instruction based on the quantile corresponding to the risk preference pattern in the posterior distribution of irrigation volume.

Images