Multi-objective dynamic optimization method for paddy field water threshold during the growing season

Abstract

This invention discloses a multi-objective dynamic optimization method for paddy field moisture thresholds during the growth period. The method includes: acquiring field environment, weather forecast, and crop apparent data; dynamically assimilating and correcting the crop mechanism model to output a comprehensive physiological state vector at the current moment; decoupling this physiological state vector, weather forecast, and multi-dimensional offline response surface into a dynamic objective weight vector; determining the resource budget upper limit and the baseline trigger threshold range based on the physiological state and reverse retrieval on the response surface at the transition time of the growth stage; calculating the dynamic initiation threshold and executing the control state machine judgment at the regular decision time; and solving the weighted objective function on the response surface based on the objective weights and budget upper limit when the trigger condition is met, outputting the optimal control variable instruction. This invention solves the risks of underlying mechanism divergence and non-convergence in the optimization process, achieving high-frequency and accurate multi-objective optimization.

Description

Technical Field

[0001] This invention relates to the field of agricultural informatization and precision irrigation control technology, and in particular to a multi-objective dynamic optimization method for paddy field moisture threshold during the growing season. Background Technology

[0002] Paddy field water management is a technical aspect of regulating the physiological state of crop populations and yield formation. Precise water regulation not only requires responding to short-term weather and soil environmental disturbances, but also necessitates long-term consideration of the physiological water requirements of crops at different developmental stages and the overall water resource allocation of the system. Establishing a dynamic decision-making mechanism based on the integration of crop growth mechanisms and real-time field sensing data has engineering application significance for overcoming the spatiotemporal mismatch of water supply caused by traditional static experience threshold control, improving the overall system's optimization accuracy, computational efficiency, and robustness of underlying control.

[0003] Current mainstream paddy field irrigation decision-making systems typically rely on static sensor threshold control or forward simulation using crop models. Regarding state awareness, existing technologies mostly replace corresponding state variable parameters in the crop mechanism model through periodic field sampling or remote sensing inversion, correcting numerical biases introduced by long-term simulations. In terms of operational decision-making, existing solutions generally use preset empirical water limits as irrigation trigger rules, or employ heuristic algorithms to perform offline iterative optimization for a single water quantity target within a given parameter space, and then issue the calculated irrigation quota to the field microcontroller for execution via one-dimensional conditional judgment commands.

[0004] However, the aforementioned existing technologies still suffer from the following technical defects when dealing with complex and multi-constrained environments, making the decision-making system highly susceptible to instability at the edge. Specifically: In the model state assimilation stage, existing solutions force numerical overwriting of the crop's organ dry matter library, but fail to consider the dynamic constraints based on energy flow and material conservation within the crop growth model, which can easily lead to divergence and collapse in subsequent photosynthetic and respiration calculations; When facing multi-variable collaborative decision-making scenarios such as combined water and fertilizer scheduling, the alternating dimensionality reduction optimization algorithm used in existing solutions can induce zipper effect oscillations at the ridges of the response surface with large mixed second-order partial derivatives, causing the optimization process of the underlying controller to fail to converge. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-objective dynamic optimization method for the water threshold in paddy fields during the growing season, in order to solve the above-mentioned problems.

[0006] Technical solution: A multi-objective dynamic optimization method for paddy field moisture threshold during the growth period, comprising:

[0007] Acquire real-time field environmental data, weather forecast data, and crop apparent growth indicators;

[0008] The crop mechanism model is dynamically assimilated and corrected based on field environmental data and crop apparent growth indicators to obtain the crop comprehensive physiological state vector at the current moment.

[0009] Independent driving components are extracted from crop integrated physiological state vector, meteorological forecast data and pre-constructed multidimensional offline response surface, and a dynamic target weight vector is synthesized.

[0010] At the moment of crop growth stage transition, a reverse retrieval is performed on the multidimensional offline response surface based on the crop comprehensive physiological state vector to determine the upper limit of resource budget and the range of benchmark trigger threshold for the current growth stage.

[0011] At regular decision-making moments, the dynamic start threshold is calculated based on field environmental data, dynamic target weight vector and benchmark trigger threshold range, and the control state machine triggers the decision.

[0012] When the triggering condition is met, the weighted objective function is solved on the multidimensional offline response surface based on the dynamic target weight vector and the upper limit of the resource budget to obtain the optimal control variable instruction.

[0013] Beneficially, to address the problem of mechanistic model divergence and collapse caused by state assimilation disrupting carbon mass closure, the proposed solution introduces a forced compensation mechanism for whole-plant carbon mass conservation. By utilizing the non-structural carbohydrate reservoir in the main stem as a buffer, the absolute mass difference in dry matter of aboveground organs resulting from directional adjustments is subject to equal and inverse zero-sum subtraction or accumulation. This method strictly maintains the energy conservation law of photosynthesis and respiration within the system, eliminating the risk of divergence in subsequent calculations of the underlying kinetic equations.

[0014] To address the potential needs of multivariate water and fertilizer collaborative decision-making scenarios, the solution also provides an advanced implementation method for two-dimensional decoupled optimization based on the joint gradient ascent method using analytical partial derivatives. By extracting the partial derivatives of water and nitrogen from the continuous tensor product surface, a two-dimensional joint gradient vector is reconstructed, and the instructions are updated synchronously and in parallel along the steepest ascent direction. This approach avoids the zipper effect oscillations caused by traditional coordinate alternating dimensionality reduction algorithms at the ridges of strongly coupled variable response surfaces from the underlying mathematical architecture, improving solution efficiency and global convergence. Under the basic water single-control mode, the solution combines the analytical partial derivatives of the multidimensional offline response surface with the golden section search algorithm to efficiently solve for the final irrigation instruction quantity that maximizes the weighted net benefit under resource budget constraints, achieving high-frequency and accurate optimization. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of the multi-objective dynamic optimization method for paddy field moisture threshold during the growing season in the embodiments of this application.

[0016] Figure 2This is a flowchart illustrating the steps for obtaining the crop's comprehensive physiological state vector at the current moment in this embodiment of the application.

[0017] Figure 3 This is a flowchart illustrating the steps for obtaining the dynamic target weight vector in an embodiment of this application.

[0018] Figure 4 This is a flowchart illustrating the steps of the mutually exclusive and exhaustive state machine decision-making logic in the high-frequency decision layer of this application embodiment. Detailed Implementation

[0019] Example 1, such as Figure 1 As shown in the figure, this embodiment elaborates in detail the dual-frequency scheduling global macro framework of the multi-objective dynamic optimization method for paddy field moisture threshold during the growth period, as well as the system environment and data perception process.

[0020] Step 101: Obtain real-time field environmental data, weather forecast data, and crop apparent growth indicators.

[0021] Specifically, field environmental data includes parameters reflecting the current physical state of the paddy field, such as real-time collected soil moisture and field surface water depth. In practical deployment, multiple layers of soil moisture sensors can be deployed at different depths in the root zone of the target field to continuously and automatically record the volumetric water content at each depth. Simultaneously, water level gauges are deployed on the field surface to synchronously record the field surface water depth. The system connects to local weather stations to obtain real-time meteorological data such as temperature, radiation, and precipitation, and obtains short-term weather forecast data from the regional meteorological service platform, extracting forecast confidence indices. Crop apparent growth indicators are external morphological data of crops obtained through field measurements or remote sensing. For example, during the rice's greening and tillering stages, the canopy leaf area index is obtained through field measurements using a canopy analyzer or by inverting multispectral images; the number of effective tillers per unit area is obtained through quadrat counting; or the main stem height is obtained through ruler measurement.

[0022] In some optional implementations, after acquiring the above multi-source data, the system will integrate the water content and water depth of each layer after aligning them with timestamps, and attach the observation time and growth stage label to the crop apparent growth index, and uniformly encode it for subsequent model assimilation.

[0023] Step 102: Dynamically assimilate and correct the crop mechanism model based on field environmental data and crop apparent growth indicators to obtain the crop comprehensive physiological state vector at the current moment.

[0024] The crop mechanism model is a digital model that simulates the dynamics of crop growth. In this embodiment, the model can be a mechanistic process model with daily step-by-step progression capability, such as the ORYZA series of rice growth simulation systems, the CERES-Rice module of the integrated crop environment and resources simulation-rice model, or the rice parameterized version of the WOFOST model of the World Food Research-crop growth simulation. This type of model includes sub-modules such as developmental processes, photosynthetic production, and assimilate allocation. The crop integrated physiological state vector is a set of current physiological characteristics that integrates apparent observation data and model-calculated data. During operation, pure forward model simulation will continuously accumulate deviations due to parameter errors and environmental disturbances. Therefore, the system uses measured field environmental data as boundary conditions to drive the model, and when new crop apparent growth indicators are obtained, it extracts the simulated apparent quantities from the model and compares them with the measured indicators, using the deviation between the two to make directional adjustments to the dry matter pool or rate parameters within the model. After the correction is completed, internal physiological variables that cannot be directly measured, such as root zone water absorption rate and panicle differentiation process index, are extracted from the model. These variables are then combined with the measured apparent indicators and normalized to obtain the crop's comprehensive physiological state vector at the current moment.

[0025] Furthermore, to ensure the consistency of the normalization process, the system obtains the value range reference of each physiological component through stress-free free growth simulation statistics in the offline stage. In the online operation stage, it only needs to call the fixed normalization parameter set to perform forward normalization.

[0026] Step 103: Extract independent driving components based on crop integrated physiological state vector, meteorological forecast data and pre-constructed multidimensional offline response surface, and synthesize dynamic target weight vector.

[0027] The multidimensional offline response surface is a pre-constructed continuous surface representing the mapping relationship between the environment, control variables, and yield. It reflects the expected impact of applying different irrigation amounts on the final yield under specific physiological states and soil moisture conditions. The dynamic objective weight vector is a dynamic control coefficient used to balance multiple objectives such as yield preservation and water conservation. Since the crop's yield preservation requirements and water conservation space are constantly changing under different growth stages and different weather forecast confidence levels, the system decouples the control requirements of three independent dimensions by extracting the physiological period sensitivity driving component, the forecast uncertainty driving component, and the marginal benefit decay driving component. Subsequently, the three components are synthesized according to specific rules to output a dynamic objective weight vector that can reflect the current optimal resource allocation tendency in real time.

[0028] As an example, when the value of the dynamic target weight vector approaches the upper limit, it indicates that the current period is a high-yield-sensitive period or the forecast uncertainty is extremely high, and the system decision will strongly tend to irrigate fully to ensure production; when the weight value is low, it indicates that the marginal yield return of the current increased irrigation is limited, and the system decision will tend to actively conserve water.

[0029] Step 104: At the moment of crop growth stage transition, perform reverse retrieval on the multidimensional offline response surface based on the crop comprehensive physiological state vector to determine the upper limit of resource budget and the baseline trigger threshold range of the current growth stage.

[0030] In this embodiment, the system employs a nested low-frequency and high-frequency scheduling logic. The low-frequency decision layer is triggered when the crop's growth stage transitions. The resource budget ceiling refers to the maximum allowable amount of water or fertilizer to be consumed within a single growth stage; the baseline trigger threshold range is the basic water boundary for triggering actions calculated based on economic benefits. When triggering at the low-frequency layer, the system uses the latest crop comprehensive physiological state vector as input to perform long-term global planning on a multi-dimensional offline response surface. Through reverse retrieval, the system estimates the minimum amount of water resources required to maintain the predetermined yield target within the current growth stage under the current state, and uses this, combined with the total remaining available water, to define the resource budget ceiling for that stage. Simultaneously, the system searches for an economic break-even point on the surface to determine a rough baseline trigger threshold range, thus defining the search range for precise optimization at the high-frequency layer.

[0031] Understandably, this low-frequency layer budget planning mechanism effectively resolves the contradiction between long-term, all-season water budgets and short-term water stress, preventing the risk of excessive consumption in the early stages leading to a lack of water available during critical periods.

[0032] Step 105: At the regular decision-making moment, calculate the dynamic start threshold based on field environment data, dynamic target weight vector and benchmark trigger threshold range, and execute the control state machine trigger determination.

[0033] Correspondingly, the high-frequency decision-making layer executes at each regular decision-making moment. The dynamic activation threshold is a judgment boundary refined by combining real-time weights with the benchmark threshold given by the low-frequency layer. In this embodiment, the system reads the measured effective water content of the root zone and the water depth of the field surface at the current moment, and uses the dynamic target weight vector generated at the current moment to accurately calculate the dynamic activation threshold within the benchmark trigger threshold interval that makes the weighted marginal yield increase equal to the weighted marginal water use cost. Specifically, the system uses the lower and upper bounds of the benchmark trigger threshold interval as the search interval, and searches for the soil water content value within this interval that makes the following equation true as the dynamic activation threshold: , where w t This represents the dynamic target weight vector at the current moment. Let g be the partial derivative of the multidimensional offline response surface with respect to soil moisture content θ under the current physiological state. critical The system uses a binary search method to iteratively approximate the moisture content value that satisfies the above equation within the search interval. Subsequently, the system substitutes the current measured moisture data into the control state machine and performs a rigorous logical comparison with the calculated dynamic start threshold to determine whether the subsequent control variable optimization process truly needs to be triggered at the current moment.

[0034] Step 106: When the triggering condition is met, the weighted objective function is solved on the multidimensional offline response surface based on the dynamic target weight vector and the upper limit of the resource budget to obtain the optimal control variable instruction.

[0035] The optimal control variable instruction is the irrigation amount or combined water and nitrogen application amount ultimately issued to the physical execution end. Once the high-frequency state machine determines that the current soil moisture is below the preset safety threshold, meeting the trigger condition, the system enters the optimization solution phase. At this time, the system constructs a weighted objective function with the control variable as the independent variable. This objective function comprehensively considers the expected yield benefit and resource consumption cost given by the multidimensional offline response surface, and uses a dynamic objective weight vector to weight these two items. During the solution process, the system is strictly constrained by the resource budget upper limit defined by the low-frequency layer, and finds the optimal control variable instruction that maximizes the weighted net benefit through an optimization algorithm. Finally, this instruction is encapsulated and issued to the field control equipment for execution.

[0036] Furthermore, before the generated instructions are issued, they can be processed by the minimum executable quantity filtering rule. If the calculated instruction quantity is lower than the minimum executable physical threshold of the irrigation system, the instruction quantity is set to zero and recorded in the under-irrigation accumulation register to avoid frequent start-up and shutdown of field irrigation and drainage facilities.

[0037] Example 2: This example details the algorithm of the perception layer, explains how the model is corrected through observation data, and demonstrates the anti-divergence physical base.

[0038] Step 201: Using field environmental data as the water boundary condition, drive the crop mechanism model to perform a day-by-day forward simulation to obtain the internal state trajectory of the crop mechanism model at the current moment.

[0039] In context, field environmental data specifically refers to real-time multi-layered soil moisture content in the root zone and the depth of the field surface water layer. The water boundary condition refers to the underlying constraints when the model solves the partial differential equations of energy balance and mass exchange. Specifically, starting from the day after transplanting, the model reads in this water boundary condition daily, calculates the canopy transpiration demand and soil evaporation based on the energy balance equation, and then calculates the daily net dry matter increment based on the photosynthesis and respiration submodules. The dry matter increment is distributed among the four sinks—leaf, stem sheath, root system, and panicle—according to the allocation rules corresponding to the current developmental stage, maintaining a complete rice growth trajectory within the model. In some optional implementations, the daily forward progression step size can be adjusted to hourly progression based on the frequency of meteorological data acquisition to capture transient stomatal closure effects under extreme high temperatures.

[0040] Step 202: Extract the simulated apparent quantities of the crop mechanism model at the current moment, and subtract them from the crop apparent growth indicators to obtain the observation simulation deviation.

[0041] Simulated apparent values ​​refer to the numerical values ​​of current plant morphological characteristics calculated by crop mechanism models based on historical weather-driven natural evolution without any external experimental intervention. Crop apparent growth indices are data obtained through field sampling or UAV multispectral remote sensing. After acquiring two sets of parallel data, the system extracts the simulated leaf area index, effective tiller number, and plant height at the current moment, respectively, and performs algebraic difference operations with the corresponding measured apparent indices to obtain three independent observation-simulation biases. This embodiment focuses on calculating the systematic errors accumulated over time in purely theoretical forward simulations due to environmental perturbations and varietal parameter shifts.

[0042] Step 203: Based on the preset causal transmission relationship between each internal state variable and the crop apparent growth index, the corresponding dry matter pool and development rate parameters are adjusted in a directional manner according to the proportion of the observed simulation deviation using the gain coefficient calculated based on the error variance.

[0043] Causal transmission relationships refer to the clearly defined organ growth mapping logic in plant physiology. For example, deviations in leaf area index are mainly attributed to shifts in leaf dry matter pool or specific leaf area, thus requiring adjustments to the leaf dry matter pool; deviations in the number of effective tillers reflect a systematic shift in tillering rate or mortality rate, thus requiring correction of the phased rate parameters of the tillering dynamics submodule. To prevent single corrections from causing instability within the model, the adjustment magnitude is strictly controlled by a gain coefficient. This gain coefficient favors observed values ​​when observation reliability is high, and favors model simulation values ​​when observed data are significantly affected by field sampling errors. The gain coefficient is calculated using a simplified form of the classical Kalman gain, with the specific formula as follows:

[0044] K=(σ m 2) / (σ m 2 +σ0 2 );

[0045] Where K is the gain coefficient, σ m 2 The prior error variance σ0 is the variance of the model's prediction of this variable. 2 This represents the variance of measurement errors from field observations. Furthermore, the prior error variance can be statistically obtained from historical simulation residuals and fixed offline, while the measurement error variance is rigorously determined by the instrument's physical precision or the quadrat sampling scheme. This mechanism objectively calculates the reliability of both parties, avoiding erroneous and significant alterations to the mechanistic model parameters due to extreme sensor noise.

[0046] Step 204: Calculate in real time the difference in absolute dry matter mass of aboveground organs before and after a single adjustment.

[0047] Immediately after executing the directional adjustment, the system triggers a verification mechanism for the underlying kinetic equations. At this point, the system iterates through the total dry matter of the leaves, stems, sheaths, and panicles before adjustment, and then calculates the difference in the absolute mass of the aboveground organs' dry matter. This embodiment aims to prevent a physical paradox. The underlying crop mechanism model relies on rigorous carbon-mass and energy balance kinetic equations. If the leaf dry matter pool is adjusted in the same direction according to the deviation ratio, biomass would be created or destroyed out of thin air, which is physically impossible and would cause the model to diverge and collapse when calculating the photosynthetic distribution and respiration consumption for the next day.

[0048] Step 205: When the observed simulation deviation indicates an increase in the dry matter pool and the absolute mass difference is greater than zero, the absolute mass difference is deducted in equal and reverse amounts from the non-structural carbohydrate pool of the main stem in the crop mechanism model.

[0049] This indicates that the external field observation value is greater than the model prediction value, and the system has just artificially injected additional virtual biomass into the leaves or stem sheaths. To activate the forced compensation mechanism for the conservation of carbon mass throughout the plant, the system locates the non-structural carbohydrate storage pool in the main stem, which acts as an energy buffer within the plant. The system converts the injected additional biomass into an equivalent amount of carbohydrate consumption and subtracts it from the buffer pool. Specifically, assuming that the leaf area index measured by multispectral remote sensing on a certain day is higher than the model prediction, the system adjusts the leaf dry matter pool accordingly, calculating an absolute mass difference of +150 grams per square meter. At this point, the system will forcibly deduct 150 grams per square meter from the non-structural carbohydrate storage pool in the main stem.

[0050] Step 206: When the balance of the non-structural carbohydrate storage bank in the main stem is insufficient, the remaining amount is transferred to the root dry matter bank for deduction.

[0051] When plants experience severe growth restriction or excessive consumption, their main stem buffer pool may be on the verge of depletion. Continuing the calculation example above, if, when preparing to deduct 150 grams per square meter, the system finds that the current balance of the main stem's non-structural carbohydrate storage pool is only 100 grams per square meter, the deduction would result in a negative storage capacity, leading to a calculation overflow error. In this case, the system will zero out the storage pool, deducting 100 grams per square meter, and then transfer the remaining 50 grams per square meter mass deficit through the material distribution network to the root dry matter pool in the underground part of the plant for further deduction. The physiological basis for this downward transfer deduction is that when vigorous growth in the aboveground parts depletes the free carbon source, the plant will sacrifice the dry matter accumulation in the underground root system to compensate.

[0052] Step 207: When the observed simulation deviation indicates a reduction in the dry matter pool and the absolute mass difference is less than zero, the reduced absolute mass is added to the non-structural carbohydrate pool of the main stem in equal amounts.

[0053] This embodiment addresses a situation where external observations indicate crop growth is worse than the model's expectations. The system artificially reduces the virtual biomass of leaves or stem sheaths. To adhere to the law of conservation of mass, the lost mass cannot be erased by the computer. The system converts the absolute mass reduction into an equivalent amount of carbohydrates and stores it back in the non-structural carbohydrate reservoir of the main stem. Through a two-way mechanism of forward subtraction and reverse accumulation, the system ensures that the total carbon mass of the whole plant remains absolutely constant before and after a single assimilation correction; that is, the change in total mass is zero, maintaining the forward-progressing photosynthetic and respiration energy balance law of the model.

[0054] Step 208: Extract internal physiological variables from the crop mechanism model after state assimilation correction, and merge the internal physiological variables with the crop's apparent growth indicators to obtain the crop's comprehensive physiological state vector at the current moment, such as... Figure 2 As shown.

[0055] Internal physiological variables refer to deep-seated mechanistic parameters that cannot be directly obtained through non-destructive field measurements but are indispensable for irrigation decisions and have been calibrated and constrained by measured data. These include root zone water absorption rate, reflecting the actual soil water extraction intensity due to current canopy transpiration pull and root distribution; dry matter allocation coefficients, indicating whether rice is in the vegetative or reproductive growth stage; panicle differentiation progress index, characterizing relative progress; developmental stage index, reflecting accumulated temperature progress; and canopy nitrogen nutrient status estimate. At the same decision-making moment, the system merges the model's calculated output with the measured apparent growth indicators, thus forming a unified multidimensional array encompassing both apparent indicators and deep physiological states. This array represents the crop's comprehensive physiological state vector at the current moment. As an improvement to the above scheme, the extracted components can also be mapped to a unified zero-to-one interval using the min-max normalization method, eliminating bias caused by dimensional differences and significant numerical magnitude disparities during subsequent high-dimensional space optimization.

[0056] Example 3: This example elaborates on the operational logic of multi-objective weight generation and its underlying algebraic function constraints, demonstrating that the dynamic objective weight vector is accurately calculated within a preset absolute boundary interval through three-component decoupling.

[0057] Step 301: Extract physiological period sensitivity driving components based on the developmental stage index and population physiological indicators in the crop integrated physiological state vector.

[0058] Specifically, the degree of damage to final yield varies depending on the stage of growth when subjected to the same amount of water deficit. The system constructs a sensitivity baseline curve with a developmental stage index as the independent variable. This curve rises rapidly from the jointing to the booting stage, peaks at the heading and flowering stage, and gradually declines in the late grain-filling stage. To capture sensitivity changes caused by population differences within the same growth stage, the system further refines the baseline curve using population physiological indicators such as the leaf area index and dry matter allocation coefficient from the crop's comprehensive physiological state vector. For example, when the normalized leaf area index indicates vigorous population growth and is in the active panicle differentiation window, the system multiplies the baseline sensitivity coefficient by an amplification factor greater than one, allocating more resources to maintaining the yield of high-yielding populations. The values ​​after this refinement constitute the physiological period sensitivity driving component.

[0059] Step 302: Extract the forecast confidence index from the meteorological forecast data and generate the forecast uncertainty driving component by combining it with the crop integrated physiological state vector mapping.

[0060] In agricultural decision-making, the reliability of short-term weather forecasts determines the degree of conservatism in decision-making. The more uncertain the forecast, the more the decision should favor production preservation, because the penalty for irreversible yield loss due to underestimating water demand in an uncertain environment far outweighs the water resource cost of overestimating water demand. Therefore, the system needs to convert the ensemble forecast dispersion or skill score provided by weather forecasting agencies into a forecast confidence index, and calculate the risk premium based on this.

[0061] Step 303: The forecast confidence index is processed using a decreasing logistic function that has been scaled and translated to calculate the forecast uncertainty driving component. The decreasing logistic function is configured with a preset theoretical scaling upper limit, decay center point and shape parameter.

[0062] To eliminate the black-box effect of weight generation and ensure that edge computing terminals can output conservative increments without jumps in extremely low time complexity, the system employs a translated and scaled decreasing logistic function to implement the mapping logic. The specific computer mathematical expression is as follows:

[0063] ;

[0064] Where, α uncer t(p t ) represents the driving component of the forecast uncertainty in the output, p t The input forecast confidence index, whose value range is set to zero to one, α max Here, is the theoretical scaling upper limit, e is the base of the natural logarithm, p0 is the inflection point of the curve (i.e., the attenuation center), and k is the shape parameter controlling the steepness of the curve. During system initialization, the preset theoretical scaling upper limit is 0.2, the inflection point is set to 0.65, and the shape parameter is set to 7.32. Furthermore, the system queries the multidimensional offline response surface to obtain the meteorological uncertainty amplitude of the yield under the current situation. If this amplitude value is in the high percentile of the historical distribution, the output of the above formula is multiplied by an amplification factor to achieve interactive adjustment between meteorological uncertainty and yield sensitivity.

[0065] Step 304: On the multidimensional offline response surface, based on the crop integrated physiological state vector and field environment data, query the initial marginal gradient at the zero control variable in the current state, and map it to generate the marginal benefit decay driving component.

[0066] This embodiment aims to evaluate the marginal yield increase return of additional irrigation. The system uses the current physiological state and current soil moisture as query indices, extracting the partial derivative value at which the irrigation application amount is zero from the multidimensional offline response surface as the initial marginal gradient. This initial marginal gradient characterizes the marginal yield increase effect brought by the first millimeter of irrigation water in the state before irrigation. If this initial marginal gradient approaches zero or is negative, it indicates that the current moisture condition is already close to sufficient or excessively wet; continuing to prioritize yield preservation would only waste water resources or even cause waterlogging.

[0067] Step 305: Compare the initial marginal gradient with the percentile constant of the historical marginal gradient distribution of the fertility stage, and use a continuous piecewise linear function with hard boundary truncation to calculate the marginal benefit decay driving component.

[0068] To ensure the continuity of the gradient response across the entire domain and avoid weight cross-boundary oscillations caused by small perturbations, the system employs a continuous piecewise linear function with hard boundary truncation to perform the mapping. The specific algebraic constraint model is calculated across four logical intervals:

[0069] When g0 is greater than or equal to g 75 At that time, α margin =0.15;

[0070] When g0 is strictly greater than g 25 And strictly less than g 75 At that time, α margin =0.05+0.10*(g0-g 25 ) / (g 75 -g 25 );

[0071] When g0 is strictly greater than 0 and less than or equal to g 25 At that time, α margin =-0.10+0.15*g0 / g 25 ;

[0072] When g0 is less than or equal to 0, α margin =-0.10.

[0073] Where, α margin To calculate the marginal benefit decay driving component of the output, g0 is the initial marginal gradient obtained from the query, g 25 With g 75 These are the 25th and 75th percentile constants of the historical marginal gradient for this reproductive stage in the offline simulation database, respectively. The piecewise function described above provides fixed and strong support for production conservation when the gradient is extremely high, forcibly triggers water-saving and flood control constraints when water saturation causes the gradient to become negative, and maintains an absolutely linear and smooth transition in the intermediate region.

[0074] Step 306: When performing additive rule synthesis, the value range of the physiological period sensitivity driving component is linearly compressed and mapped to reserve adjustment ranges for the forecast uncertainty driving component and the marginal benefit decay driving component.

[0075] To ensure that the three driving components have effective regulatory capacity at all stages of the entire growth period, a range-balanced design is necessary. The system repositions the physiological period-sensitive driving component as a contributing component of the base weights, and linearly compresses its original value range from zero to one to an effective range of 0.15 to 0.7. This compression operation retains sufficient regulatory margin in the synthesized weight space, allowing the forecast uncertainty driving component and the marginal benefit decay driving component to still have a substantial influence on the final synthesized weights, even during the most sensitive heading and flowering stages.

[0076] Step 307: The physiological period sensitivity driving component, the forecast uncertainty driving component, and the marginal benefit decay driving component are synthesized according to the addition rule. After boundary truncation and temporal smoothing filtering, the dynamic target weight vector is obtained, as follows: Figure 3 As shown.

[0077] This embodiment abandons the multiplicative synthesis rule because the physical meanings of the three components mentioned above belong to mutually independent adjustment dimensions. The additive structure ensures that extreme fluctuations in any single component will not propagate through multiplication and cause the synthesis result to go out of control. After the synthesis calculation is completed, the system applies hard upper and lower bounds to the result, setting the lower bound to 0.1 to retain the minimum production considerations, and setting the upper bound to 0.95 to prevent the water-saving target from failing due to the lack of constraints.

[0078] Step 308: Substitute the original weight values ​​obtained after compression mapping and additive synthesis, along with the dynamic target weight vector from the previous decision time, into the exponentially weighted moving average filter function for cross-time update to obtain the dynamic target weight vector.

[0079] Because sudden changes in weather forecasts or sensor observation noise can cause short-term pulse fluctuations in the driving components between adjacent decision-making moments, directly passing weights can lead to repeated on / off actions of field water valves. Therefore, the system applies an exponentially weighted moving average filter to the truncated original weight values. Specifically, the algebraic relationship is that the dynamic target weight vector at the current moment equals the dynamic target weight vector at the previous moment multiplied by a memory coefficient, plus the original weight value at the current moment multiplied by one minus the memory coefficient. In some preferred embodiments, the memory coefficient is set to 0.3, balancing the system's rapid response to new weather information with the stability of cross-time decision-making. The final output dynamic target weight vector degenerates into a scalar form in the dual-objective decision-making mode, representing the absolute compromise tendency of the current decision step in the competition between the dual objectives of production preservation and water conservation.

[0080] Example 4: This example details the process of achieving dimensionality reduction of computing power through reverse water volume binary search, focusing on the underlying logic of physical dimension conversion, as well as the specific execution mechanism of budget profit and loss carry-over and smooth scheduling when crop growth stages change.

[0081] Step 401: On the multidimensional offline response surface, obtain the peak yield potential corresponding to the crop's comprehensive physiological state vector, and calculate the absolute target yield by combining it with the preset target yield protection ratio.

[0082] Specifically, along the marginal product gradient field of the analytical surface, the system searches for the point where the marginal product gradient equals zero as the peak production potential for that stage. This peak production potential represents the upper bound of the production that can still be achieved under the current state. Subsequently, the system calculates the absolute target production by combining the target production maintenance ratio determined by the dynamic target weights. The calculation formula is as follows:

[0083] Y target =Y pot *η;

[0084] Among them, Y target For the absolute target output, Y pot Let η represent the peak production potential, and η represent the target production guarantee ratio. In practical applications, when the dynamic target weight is high, the target production guarantee ratio can be set to 95%, corresponding to a higher estimated irrigation demand; when the dynamic target weight is low, the target production guarantee ratio drops to 85%, corresponding to a lower estimated irrigation demand.

[0085] Step 402: Set the lower and upper bounds of the search interval for the control variables, use the binary search numerical approximation algorithm to calculate the midpoint of the current interval, and input the midpoint of the current interval into the multidimensional offline response surface to calculate the corresponding predicted output.

[0086] Since the multidimensional offline response surface is a unidirectional forward model for predicting yield based on input water quantity, requiring the minimum water quantity to be derived from the input target yield is mathematically essentially a problem of finding the root of a forward nonlinear surrogate model. Based on the physical constraint that the agronomic response curve exhibits a strictly monotonically increasing trend before reaching the peak yield potential, the system sets the lower bound of the water requirement search to zero and the upper bound to the historical extreme water quantity for that period. The system enters a loop iteration, calculates the midpoint of the current interval, and inputs this midpoint as a probe parameter into the multidimensional offline response surface to quickly calculate the corresponding predicted yield. This design compresses the complex nonlinear inversion into a minimally simplistic forward query.

[0087] Step 403: When the predicted output is less than the absolute target output, discard the left half of the interval; when the predicted output is greater than or equal to the absolute target output, discard the right half of the interval. Continue iterating until the search interval width is less than the preset tolerance threshold.

[0088] The system compares the predicted output calculated forward with the absolute target output. If the predicted output is less than the absolute target output, it indicates that the current calculated water volume is insufficient. The system discards the left half of the interval and updates the lower bound to the current interval's midpoint. If the predicted output is greater than or equal to the absolute target output, it indicates that the water volume has reached or exceeded the target. To approximate the minimum water volume, the system discards the right half of the interval and updates the upper bound to the current interval's midpoint. The system continuously checks whether the width of the current search interval is less than or equal to a preset water meter approximation tolerance threshold. This tolerance threshold can be set to one millimeter. If the threshold is met, the iteration terminates; otherwise, it continues the halving approximation.

[0089] Step 404: The right boundary value of the interval that meets the tolerance condition is taken as the absolute minimum cumulative demand, and the resource budget ceiling for the current reproductive stage is determined by combining it with the total remaining available resources currently being tracked and maintained.

[0090] When the search interval width meets the tolerance condition, the system extracts the current right boundary value as the estimated absolute minimum cumulative demand to meet the target production ratio. The system allocates the currently available total remaining resources among the remaining stages according to the proportion of their respective absolute minimum cumulative demands to obtain the resource budget ceiling for the current reproductive stage. As an improvement and alternative to the above scheme, when estimating the water demand of each remaining stage in the low-frequency layer, to avoid the current short-term extreme values ​​of weights affecting long-term planning, for more distant subsequent stages, the system can use the typical historical average weights of each stage saved in the offline calibration stage as substitute values ​​for the expected weights, reducing the excessive influence of dynamic weights at a single moment on long-term budget allocation.

[0091] Step 405: Obtain the pre-configured unit resource economic cost and expected output price.

[0092] In this embodiment, the unit resource economic cost refers to the local irrigation water price, and the expected output price refers to the grain price of the target crop. These two economic parameters are the basic input variables for calculating the irrigation economic break-even point.

[0093] Step 406 introduces a volume-area depth physical conversion operator to map the ratio of unit resource economic cost to expected output price into a physical metric isomorphic to the marginal product gradient field, and calculates the economic break-even gradient.

[0094] The physical metric of the marginal product gradient field derived from the multidimensional offline response surface analysis is kilograms per hectare per millimeter. Dividing the water price by the grain price yields a metric of kilograms per cubic meter. Without this physical dimensional transformation, the calculated critical threshold will exhibit an error of tenfold, leading to complete decision failure. To ensure absolute alignment of the physical dimensions on both sides of the equation, the system performs a rigorous volume-area-depth physical conversion operator on the economic indicators. The calculation formula is as follows:

[0095] gcritical =(Price water / Price grain )*C v ;

[0096] Among them, g critical For the economic break-even gradient, Price water Price is the economic cost per unit of resource input. grain C represents the expected output price. v This refers to the physical conversion operator for volume, area, and depth. (C) v The value of is fixed at ten, which physically means converting the volume of a one-millimeter water layer depth over a one-hectare area into ten cubic meters. Through this operator conversion, the dimensions of the obtained economic break-even gradient are rigorously unified, eliminating the risk of magnitude collapse caused by engineering conversions. The system saves this economic break-even gradient for subsequent high-frequency decision-making when calculating the dynamic activation threshold.

[0097] Step 407: Within the current growth stage, based on the upper and lower endpoints of the expected range of change of the crop comprehensive physiological state vector obtained from historical simulation statistics of this stage, respectively, reverse search is performed on the soil moisture critical value that makes the marginal yield gradient of the multidimensional offline response surface equal to the economic break-even gradient, forming the benchmark trigger threshold interval.

[0098] Because the physiological state of crops continuously changes during their growth stages, a single threshold cannot cover the needs of the entire stage. The system extracts the upper and lower endpoints of the expected range of change in the physiological state vector within each stage, and queries the soil moisture critical values ​​that match the economic break-even gradient on the multidimensional offline response surface. These two critical values ​​correspond to the activation thresholds at the least sensitive and most sensitive moments within the stage, respectively, together forming the baseline trigger threshold range passed to the high-frequency layer. Furthermore, to avoid drastic jumps between the baseline trigger threshold range of the new stage and the actual threshold at the end of the previous stage, which could cause oscillations in the water valve switching pulse, the system applies a linear, gradual, and smooth transition to the dynamic activation thresholds for the first three decision moments of the new stage. Specifically, the threshold on the first day is equal to the threshold at the end of the previous stage plus one-third of the difference between the old and new thresholds; on the second day, it is equal to two-thirds; and from the third day onwards, it transitions to the new stage threshold, ensuring the continuity of irrigation management across stages.

[0099] Step 408: When a transition to a reproductive stage occurs, read the actual resource consumption continuously accumulated by the system in the previous stage, calculate the difference between the actual resource consumption and the upper limit of the resource budget in the previous stage, and determine the budget surplus or deficit status of the previous stage.

[0100] The system is triggered to execute a carryover mechanism each time a new reproductive phase begins. The system reads the total amount of water actually distributed and consumed in the previous phase and compares it with the resource budget ceiling initially set for the previous phase. If the actual consumption is lower than the budget ceiling, it is determined to be a budget surplus; if the actual consumption exceeds the budget ceiling, it is determined to be a budget deficit.

[0101] Step 409: When a budget surplus is determined, the surplus resources are allocated and carried over to the upper limit of the resource budget for each subsequent reproductive stage according to a preset ratio.

[0102] Budget surpluses typically arise from reduced irrigation demand due to rainy weather. Instead of leaving these surplus funds idle, the system injects them into the subsequent budget pool. Specifically, 40% of the surplus can be allocated to the current, upcoming stage, while the remaining 60% is weighted and allocated according to the sensitivity-driven components of subsequent stages, ensuring sufficiency for the current stage while considering the overall situation.

[0103] Step 410: When a budget deficit is determined, the resource budget ceiling for each subsequent fertility stage is deducted in order of increasing water sensitivity priority based on the preset water sensitivity priority of each fertility stage, so as to ensure the overall water balance.

[0104] Budget deficits typically arise from earlier droughts exceeding expectations. To absorb this excess water, the system establishes a deduction order based on the inherent water sensitivity priority of each growth stage. The stage with the lowest sensitivity receives the largest deficit deduction first, while the most sensitive critical period is addressed last, thus maximizing water supply during critical periods. When implementing deficit deductions, the system mandates that the minimum reserved budget for the early and middle stages of grouting must not be less than 60% of the estimated demand for that stage, ensuring that the grouting period is not deprived of water security due to earlier over-irrigation.

[0105] Example 5, as Figure 4 As shown, this embodiment elaborates in detail the mutually exclusive and exhaustive state machine decision-making logic in the high-frequency decision layer, as well as the nonlinear optimization process for different control variable dimensions.

[0106] Step 501: Extract the current measured root zone water content and field surface water depth from the field environmental data.

[0107] During routine decision-making, the system needs to perceive the physical state of the field in real time for high-frequency scheduling. From the field environmental data reported by the underlying sensor network, the system analyzes the current measured root zone water content, which characterizes the soil's moisture sufficiency, and the surface water depth, which characterizes the state of flooding. These two variables constitute the input parameters for subsequent state machine execution trigger decisions.

[0108] Step 502: Establish mutually exclusive and exhaustive Boolean logic decision branches, including decision branch A and decision branch B.

[0109] To prevent convergence issues or control flow dangling in the optimization process of industrial control code on edge devices, the system constructs a complete truth table based on the principle of mutual exclusion and exhaustiveness. Mutual exclusion and exhaustiveness ensures that regardless of the extreme combination of measured root zone water content and surface water depth, the system can and will only fall into a defined logical branch. This design addresses a software defect in conventional conditional statements that easily misses exits when handling multi-dimensional boundary overlaps.

[0110] Step 503: When the current measured root zone water content is strictly greater than the dynamic start threshold, and at the same time the field surface water depth is greater than zero or the current growth stage does not require maintaining the field surface physical water layer, the decision branch A is entered, the triggering condition is not met, and the control variable is forcibly set to zero.

[0111] Decision branch A represents the safe exit logic for not irrigating. When the system determines that the current measured root zone moisture content is higher than the dynamic activation threshold, it indicates that not only is the soil moisture sufficient, but the microclimate environment of the canopy is also safe. Specifically, as long as there is still a physical water layer on the field surface, or the crop is in a physiological stage that clearly requires drying out, such as the late tillering stage, the system will enter decision branch A. At this time, the system determines that the triggering conditions are not met, skips all optimization calculations, forcibly sets the irrigation and fertilization commands to zero, and enters a silent budget tracking state.

[0112] Step 504: When the current measured root zone water content is less than or equal to the dynamic start threshold, or when the current measured root zone water content is strictly greater than the dynamic start threshold but the field surface water depth is zero and the current growth stage requires maintaining the field surface physical water layer, the decision branch B is entered, and the triggering condition is met.

[0113] Branch B represents the emergency logic exit point for deficit-triggered irrigation. The system will immediately trigger branch B as soon as the soil moisture content falls below the critical threshold for water balance. Furthermore, even if the soil moisture is still sufficient, if the crop is in an extremely sensitive stage, such as the heading and flowering stage, where a water layer must be maintained for physical temperature defense, and the surface water layer has already dried up, the system will also trigger branch B. If either of these two sub-conditions is met, the system determines that the current environment poses a substantial threat to crop yield, activating the subsequent control variable optimization process.

[0114] Step 505: Using irrigation volume as the sole independent variable, and combining the dynamic target weight vector with the preset water-yield economic equivalent conversion rate, construct a one-dimensional weighted objective function.

[0115] When the system is in a standard single-control window for water usage, it enters a one-dimensional univariate optimization process. The system constructs a continuous function with irrigation amount as the sole independent variable, aiming to maximize the weighted net benefit. The linear algebraic equation of the one-dimensional weighted objective function is expressed as:

[0116] J(I)=w t *Y expected -(1-w t )*λ*I;

[0117] Where J(I) is the constructed one-dimensional weighted objective function, I is the irrigation amount as the independent variable, and w t Y is the currently generated dynamic target weight vector. expected Let λ be the expected yield obtained from the multidimensional offline response surface, and let λ be a preset water-yield economic equivalent conversion rate. The physical dimension of this conversion rate must be strictly set to kilograms per hectare per millimeter to ensure that the polynomials on both sides of the minus sign are absolutely isomorphic in the physical metric space.

[0118] Step 506: Extract the analytical first-order partial derivative of the multidimensional offline response surface with respect to irrigation amount in the current state, and establish a first-order necessary condition equation that makes the weighted marginal increase in production equal to the weighted marginal water cost.

[0119] Because the expected yield's response to irrigation volume exhibits a concave function characteristic with diminishing marginal returns, the maximum point of the objective function must occur where its first derivative is zero. The system utilizes the mathematical properties of the tensor product continuous surface to extract the analytic first-order partial derivatives, establishing a balance-of-payments equation in the sense of physical economics. The first-order necessary condition equation is expressed as:

[0120] w t *dY expected / dI=(1-w t )*λ;

[0121] Among them, dY expected / dI is the analytical first-order partial derivative of the expected output with respect to irrigation amount. The left side of the equation represents the marginal increase in output after weighting, and the right side of the equation represents the marginal cost of water consumption after weighting.

[0122] Step 507: Using the upper limit of the resource budget for the current reproductive stage as the upper bound of the search interval, the golden section search algorithm is used to optimize the first-order necessary condition equation and solve for the final irrigation instruction amount.

[0123] The system sets the lower bound of the optimization search to zero, and the upper bound to be the resource budget ceiling for the current reproductive stage defined by the low-frequency layer. Within a closed search interval, the system uses the golden section search algorithm to continuously narrow the interval surrounding the optimal solution. Due to the strict unimodality of the first-order necessary condition equation, the golden section search algorithm can converge geometrically within a finite number of iterations, efficiently solving for the precise irrigation amount that perfectly offsets the marginal benefits and costs. This value is the final irrigation command amount to be issued and executed.

[0124] Step 508: Combine the dynamic target weight vector, water cost coefficient, and nitrogen application cost coefficient, which is strictly defined as a scalar without physical dimensions, to construct a two-dimensional weighted composite objective function containing two variables: water and nitrogen.

[0125] As an advanced implementation of the aforementioned water-only control scheme, a water-nitrogen dual control mechanism is triggered when the system is within the rice fertilization window. The system expands its optimization space from one dimension to two dimensions, maximizing the synergistic benefits of integrated water and fertilizer management. The constructed two-dimensional weighted composite objective function is expressed as follows:

[0126] J(I,N)=w Y *Y(I,N)-w W *λ W *Iw N *λ N *N;

[0127] Where J(I,N) is a two-dimensional weighted composite objective function, I is the irrigation amount, N is the nitrogen application rate, and w Y As the weight of the output target, w W As the weight for water conservation targets, w N Let λ be the weight of the nitrogen reduction target, and Y(I,N) be the expected yield driven by the bivariate model. W Let λ be the water cost coefficient with physical dimensions, and λ be the water cost coefficient. N Let be the nitrogen application cost coefficient. To ensure the validity of the calculation, the nitrogen application cost coefficient is strictly defined as a physical dimensionless scalar, so that the product of the penalty terms naturally remains on the base dimension of kilograms per hectare.

[0128] To eliminate the impact of the dimensional heterogeneity between irrigation amount (mm) and nitrogen application rate (kg / ha) on the joint gradient calculation and iteration step size determination, a normalization operation is introduced before performing the gradient ascent iteration. I is defined as the irrigation direction reference scale. ref Define the nitrogen-oriented reference scale N as the upper limit of the resource budget (mm) for the current reproductive stage. ref Let I' be the upper limit of the nitrogen application budget (kg / ha) for the current growth stage. Let the normalized irrigation variable I' = I / I ref Normalized nitrogen application variable N' = N / NN refThen I' and N' are both dimensionless scaling factors, with values ​​ranging from zero to one. Subsequent joint gradient ascent iterations are performed in this normalized space.

[0129] Step 509: Reconstruct the two-dimensional joint gradient vector in the normalized variable space. According to the chain rule, the partial derivative of the objective function with respect to the normalized irrigation variables is equal to the partial derivative of the objective function with respect to the physical irrigation quantity multiplied by the irrigation direction reference scale, that is:

[0130]

[0131] Similarly, the partial derivative of the objective function with respect to the normalized nitrogen application variable is equal to the partial derivative of the objective function with respect to the physical nitrogen application rate multiplied by the nitrogen direction reference scale, that is:

[0132]

[0133] in, and These are the partial derivatives of the water and nitrogen directions of the multidimensional offline response surface at the current state coordinates, respectively. After scaling to the aforementioned reference scale, both normalized gradient components have a unified dimension of kg / ha, ensuring consistency in the physical dimensions of subsequent L2 calculations, search direction determination, and scalar step size selection. The system combines these two components into a two-dimensional joint gradient vector in the normalized space.

[0134] Step 510: Perform joint gradient ascent iterations in the normalized space. The system initializes both the normalized irrigation variable and the normalized nitrogen application variable to zero, corresponding to the zero irrigation and zero fertilization states in physical space. In each iteration, the system calculates the normalized two-dimensional joint gradient vector at the current coordinates, and then calculates the L2 norm of this gradient vector:

[0135]

[0136] Since both components are in kg / ha dimension, the aforementioned L2 norm is a dimensionally compatible kg / ha scalar. The system uses this gradient vector as the steepest ascent direction and employs a backtracking search technique to adaptively determine the dimensionless scalar step size α. k Its physical dimension is ha / kg, making α k Multiplying by the gradient component in kg / ha yields the dimensionless increment of the normalized variable. The normalized variable is then updated synchronously using the following formula:

[0137]

[0138] When the gradient 2 norm When the marginal adjustment space is less than the preset uniform approximation tolerance ε, for example, ε = 10 kg / ha, indicating that the economic significance threshold has been reached, the iteration terminates. The system then performs an inverse transformation of the converged normalized coordinates back to the physical dimension space.

[0139]

[0140] income This is the final irrigation instruction amount (mm). This is the nitrogen application rate (kg / ha).

[0141] Since the entire iterative process proceeds along the true steepest ascent direction in the dimensional isomorphic normalized space, the moisture and nitrogen variables are updated synchronously and in parallel at each step, avoiding the oscillations caused by the zipper effect at the ridge of the response surface in the alternating dimensionality reduction algorithm.

[0142] Example 6: This example details how, in an agricultural industrialization application scenario, the system optimizes scheduling by performing forward-looking corrections based on short-term weather forecasts, prevents control instability through a dual-pricing risk stripping mechanism, and ultimately safeguards the overall safety baseline of water resources through an abnormal scenario safety interlocking mechanism.

[0143] Step 601: Based on meteorological forecast data and current measured root zone water content, simulate forward to obtain the predicted soil moisture trajectory.

[0144] Specifically, this step forms the basis of the look-ahead correction mechanism. The system inputs the current measured root zone water content as the initial state into the water balance submodule of the crop mechanism model. Based on this, the system uses short-term weather forecast data covering the next three to five days as atmospheric driving conditions, and, assuming no irrigation is performed at the current decision point, uses the model to extrapolate the dynamic consumption and replenishment process of water. This extrapolation process outputs a time-series curve reflecting the evolution of soil moisture over the next few days; this curve is defined as the predicted soil moisture trajectory. Through this trajectory, the system possesses the ability to predict short-term water surplus and deficit trends.

[0145] Step 602: When the predicted soil moisture trajectory meets the preset precipitation substitution prediction conditions, the determination that the triggering conditions are met is forcibly postponed and corrected to not triggering for the time being; when the drought advance prediction conditions are met, the determination that the triggering conditions are not met is forcibly advanced and corrected to meet the triggering conditions.

[0146] Specifically, the precipitation substitution prediction condition requires that the cumulative forecast precipitation within the next three days exceeds the effective precipitation threshold of 15 mm, and the predicted soil moisture trajectory shows that soil moisture can naturally recover to above the dynamic activation threshold after precipitation. When this condition is met, even if the current soil moisture has fallen to the critical line for triggering irrigation, the system will use future precipitation to replace artificial irrigation, forcibly postponing the current irrigation action and conserving water resources. Conversely, the drought advance prediction condition requires that there is no effective precipitation within the next three days and the predicted soil moisture trajectory will fall below the dynamic activation threshold. In this case, even if the current soil moisture is slightly above the threshold, the system will issue an irrigation instruction in advance to prevent crops from suffering irreversible drought stress before the next decision cycle.

[0147] Furthermore, to prevent crop drought caused by delayed irrigation due to inaccurate weather forecasts, the system introduces a protection detail where the safety margin is dynamically adjusted based on confidence level. The safety margin is dynamically mapped inversely to the forecast confidence index. When the forecast confidence index falls below 0.7, the system dynamically increases the safety margin. The calculation logic for the dynamic safety margin uses the following linear formula:

[0148] Margin dynamic =Margin initial +Factor*(0.7-p t ) / 0.3;

[0149] Among them, Margin dynamic For the calculated dynamic safety margin, Margin initial The initial safety margin is the standard preset value, Factor is the preset penalty amplification factor, and p t This is the confidence index for the currently acquired weather forecast. In the forward simulation, if any point in the predicted soil moisture trajectory falls below the sum of the permanent wilting point moisture content and the aforementioned dynamic safety margin, the delayed correction immediately fails, and the system reverts to the conventional decision of immediate irrigation.

[0150] Based on the above embodiments, an alternative scheme for compensating for delayed irrigation is provided. In actual engineering operation, if the system performs a delay correction, but the actual rainfall in the following days does not reach 50% of the forecast rainfall, it is determined that the rainfall delay has failed. To compensate for the actual water demand lost by the crop, the system explicitly adds the full amount of irrigation that should have been issued but was delayed in the previous moment to the under-irrigation accumulation register at the next decision moment. The accumulated value in this register will be used as an additional penalty compensation item, added to the objective function of the optimal control variable instruction for the current decision, forcing the system to prioritize compensatory water replenishment, forming a complete fault-tolerant closed loop.

[0151] Step 603: At the decision moment when the forward correction mechanism takes effect, the forecast uncertainty driving component containing climate risk factors is forced to be set to zero when synthesizing the dynamic target weight vector, thus removing the double pricing of the same meteorological fluctuation risk.

[0152] In the underlying architecture of this invention, the dynamic target weight vector already contains an uncertainty-driven component generated based on weather forecasts. This component causes the system to favor production preservation and increased irrigation when the forecast is unclear. However, when the system simultaneously performs forward delays or forward adjustments based on short-term weather forecasts, it is effectively applying the same weather risk information to irrigation decisions twice. This double pricing can lead to an overreaction to weather fluctuations, causing instability in the irrigation logic. To resolve this theoretical paradox, the system deploys a risk stripping mechanism. At the same decision moment when the system determines that the delay or forward adjustment conditions are met and effective, the system forcibly intervenes in the weight synthesis process, absolutely resetting the value of the forecast uncertainty-driven component to zero. Through this stripping operation, the dynamic target weight vector is synthesized only from two definite driving components: physiological period sensitivity and marginal benefit decay, ensuring the purity of the multi-objective optimization architecture and the stability of the control output.

[0153] Step 604: Analyze field environmental data and crop apparent growth indicators in real time. When an abnormal scenario of high temperature stress or continuous deep drought matching the preset conditions is identified, bypass the control state machine to trigger the judgment and forcibly issue an emergency irrigation command.

[0154] In agricultural industrialization control, the safety interlocking mechanism to prevent devastating disasters must possess the highest execution authority, exceeding conventional economic benefit optimization. The system establishes an abnormal scenario identification layer independent of conventional dual-frequency scheduling. For example, for an acute high-temperature stress abnormal scenario, the matching conditions require that the current daily maximum temperature exceed 35 degrees Celsius, the crop be in the extremely sensitive heading and flowering stage, and the field surface water depth be zero. Once these three boundary conditions are simultaneously met, the system immediately cuts off the logical flow of the conventional state machine, bypassing all marginal benefit calculations based on response surfaces, and issues an emergency irrigation command. The specific water volume of this emergency irrigation command is strictly set to the physical volume of water required to quickly establish a shallow water layer of three to five centimeters on the field surface. This aims to forcibly lower the canopy microenvironment temperature through the physical cooling effect of water evaporation, protecting rice pollen viability from high-temperature damage. Similarly, when the root zone water content falls below the wilting threshold for three consecutive decision cycles, the system triggers a persistent deep drought abnormal scenario, forcibly issuing a restorative water replenishment command.

[0155] Step 605: Continuously accumulate the total amount of water from emergency irrigation commands issued throughout the entire reproductive period. When the accumulated total reaches the upper limit of the total preset resource budget for the whole season, intercept and block the subsequent emergency irrigation commands and convert them into early warning suggestions to ensure the safety of overall water resources under constraints.

[0156] While emergency irrigation commands safeguard the survival of crops in extreme conditions, their lack of individual budget constraints makes them highly susceptible to uncontrolled water resource depletion in years of abnormal weather, potentially leading to reservoir depletion or disrupting the water balance of the entire irrigation district. To address this, the system deploys a global emergency red line limit at the highest level. The system uses an accumulator to continuously track the water volume of every emergency irrigation command issued due to abnormal scenarios throughout the entire growth period. The system sets a percentage limit on the total pre-set resource budget for the entire season; in this embodiment, this is specifically set as a 15% global emergency red line. When the total water volume in the accumulator reaches this absolute limit, the system triggers a physical isolation interlock. Thereafter, regardless of how frequently the abnormal scenario identification layer issues emergency irrigation commands, the system will intercept and block them before they reach the execution mechanism. Correspondingly, the blocked commands are converted into visual red alert suggestions and output to the management console, returning final scheduling authority under extreme water shortage conditions to human decision-making. This constructs a final engineering firewall between safeguarding crop lifelines and protecting the overall water resource depletion limit.

[0157] Example 7: This example details how, before the system goes online, a multi-dimensional offline response surface with anti-vibration characteristics is constructed in an offline environment, and how qualitative agronomic experience is transformed into a computational cost function to achieve automated closed-loop calibration of weight parameters.

[0158] Step 701: Obtain historical meteorological scenarios and simulated yield sample data containing discrete moisture step penalty features.

[0159] During the offline construction phase, the system needs to accumulate massive amounts of virtual experimental data as the underlying support for surface fitting. The system utilizes crop mechanism models to perform batch forward simulations under a large number of historical meteorological sequences. During the simulation process, when encountering scenarios such as the disappearance of the field surface water layer and extremely low root zone water content from the booting to heading stage, the model applies a step-down penalty to the crop fertilization rate, leaving discrete water step-down penalty features in the output simulated yield sample data. Furthermore, in the initial state loading stage for generating such sample data, the system needs to perform an inverse mapping operation to restore the normalized physiological vectors to physical dimensions. To ensure that the restored physiological state does not produce internal conflicts, the system enforces an inverse mapping consistency projection mechanism. Specifically, using the algebraic constraints of the state variables built into the mechanism model as a reference, if the leaf area index obtained from the inverse mapping is found to be mismatched with the leaf dry matter pool, the system, prioritizing the absolute constancy of the developmental stage index and the total dry matter pool of each organ, will forcibly project the component deviating from the constraints to the nearest feasible domain boundary to ensure the biological consistency of the underlying data.

[0160] Step 702: The relative water content of the root zone and the depth of the field surface water layer in the sample data are merged and mapped into a single continuous equivalent water storage of the root zone, eliminating the topological break in the metric space caused by the difference in physical measures.

[0161] This embodiment addresses the incompatibility issue in the measurement of the basic data dimension. In the original sample data, the indicator reflecting the water state uses the root zone relative water content, ranging from zero to one, when the field surface is dry, while using the field surface water depth in centimeters when the field surface is wet. This abrupt change in physical semantics can cause topological breaks at the saturation point in subsequent continuous surface fitting algorithms, leading to oscillations without physical meaning. To address this, the system introduces a unified measurement method to construct the equivalent water storage in the root zone. The specific calculation formula adopts the following linear format:

[0162] W eq =θ rel *I dry +(1+h / h ref )*I wet ;

[0163] Among them, W eq To calculate the equivalent storage capacity of a single continuous root zone, θ rel I represents the relative water content of the root zone in the original sample data. dry A Boolean variable indicating the absence of a water layer on the field surface, where h is the depth of the water layer on the field surface. ref I is the preset water layer reference depth. wet This is a Boolean variable indicating the presence of a water layer on the field surface. After the above operator transformation, this variable achieves continuity and monotonically increasing across the entire domain, eliminating the potential for topological breaks.

[0164] Step 703: Using the principal components of physiological state, the equivalent water storage in the root zone, and control variables as input dimensions, a monotonicity constraint penalty term is introduced to suppress the pseudo-gradient derivative oscillation caused by the water step penalty feature. The tensor product B-spline function is used for continuous fitting, and a multidimensional offline response surface is pre-constructed.

[0165] To control the computational complexity of surface construction, the system performs principal component analysis on the multidimensional physiological state variables. The system follows a strict dimensionality reduction criterion: when the cumulative variance explained by the extracted first principal component exceeds 85%, the system retains only this first principal component as a single comprehensive coordinate axis for the physiological state; at this point, the fitted response surface is a three-dimensional structure. After establishing the input dimension, the system uses a tensor product B-spline function for surface fitting. Due to the discrete step-descent penalty in the underlying sample data, conventional B-spline algorithms are prone to Gibbs freefall when fitting such discontinuous data, leading to alternating positive and negative pseudo-gradient derivative oscillations during analytical differentiation. To overcome this theoretical obstacle, the system forcibly introduces a monotonicity constraint penalty term into the fitting objective function, suppressing derivative oscillations from the underlying mathematical framework. This ensures that the final constructed multidimensional offline response surface is continuously differentiable, providing a smooth and reliable spatial foundation for analytical optimization in online high-frequency environments.

[0166] As an extension and alternative to the above scheme, if the variance explained by the first principal component cannot reach 85% in some complex varieties, the system can retain two or more principal components. In this case, the input dimension increases accordingly, the construction dimension of the tensor product B-spline expands accordingly, and spline nodes are independently arranged in the direction of each principal component.

[0167] Step 704: Obtain historical time series data covering typical meteorological years throughout the entire reproductive period.

[0168] In this step, the system selects meteorological sequences from typical historical years, including normal years, spring drought years, summer drought years, rainy years, and high-temperature years, to drive the rice growth model from transplanting to maturity. At each decision point, the system generates input features such as a comprehensive crop physiological state vector, meteorological forecast confidence index, and marginal yield gradient, and stores these features in a time series to form historical time series data for the entire growth period.

[0169] Step 705: The pre-defined multiple qualitative agronomic behavior criteria are transformed into a calculated agronomic behavior penalty cost function. The agronomic behavior criteria include at least the peak value criterion for heading and flowering options, the water control depression criterion for tillering stage, and the over-wetness prevention criterion.

[0170] This embodiment bridges the gap between agricultural expert experience and automated parameter tuning algorithms. The system eliminates manual trial and error, translating expert qualitative consensus into a rigorous system of algebraic equations. The overall agronomic behavior penalty cost function is calculated as follows:

[0171] Loss=ω1*L1+ω2*L2+ω3*L3+ω4*L4;

[0172] Wherein, Loss is the overall agronomic behavior penalty cost function, ω1, ω2, ω3, and ω4 are the preset weight coefficients corresponding to the four penalty terms, L1 is the penalty term corresponding to the peak value criterion of heading and flowering options, L2 is the penalty term corresponding to the water control and seedling depression criterion of tillering stage, L3 is the penalty term corresponding to the low confidence level of continuous forecast and counter-trend rebound criterion, and L4 is the penalty term corresponding to the overflow prevention criterion of excessive rainfall and wetness.

[0173] Furthermore, the system explicitly defines the four independent penalty terms algebraically. Taking the peak value criterion for heading and flowering options as an example, its calculation formula is as follows:

[0174] L1=max(0,0.85-max F (w t )) 2 ;

[0175] Where L1 is the penalty term of the peak weighting criterion for the heading and flowering option, max is a mathematical function that takes the maximum of the two, and 0.85 is the expected minimum peak weighting of the protected property rights in this stage. F (w t The value represents the maximum peak value of the dynamic target weight sequence within the heading and flowering time interval of the simulation. This formula ensures from the ground up that if a sufficiently high protection weight cannot be generated during the sensitive period, the system will impose a severe quadratic penalty.

[0176] Similarly, the calculation formula for the water control depression criterion during the tillering stage is:

[0177] L 2= max(0,min T (w t -0.45) 2 ;

[0178] Where L2 is the penalty term for the water control and seedling depression criterion during the tillering stage, max is a mathematical function that takes the maximum of the two, and min... T (w t ) represents the minimum value of the dynamic target weight sequence within the tillering time interval during simulation, and 0.45 is the maximum weight trough value allowed to fall below this threshold. This formula penalizes the conservative behavior of the system when it fails to effectively identify the acceptable water deficit window.

[0179] Step 706: With minimizing the agronomic behavior penalty cost function as the optimization objective, the particle swarm optimization algorithm is used to perform global iterative optimization of the mapping parameters based on historical time series data until the agronomic behavior penalty cost function converges to a preset tolerance range. The mapping parameters of the dynamic target weight vector are then predetermined and frozen.

[0180] In this step, the system utilizes a particle swarm optimization algorithm for global optimization. The algorithm encapsulates all mapping parameters to be calibrated as particles in a high-dimensional space. The system continuously inputs historical time-series data into the weighted evolution model to generate a simulated weight sequence, and continuously evaluates the fitness of each particle using the agronomic behavior penalty cost function calculated above. The particle swarm continuously updates its velocity and position vectors through an information-sharing mechanism, with minimizing the overall cost function as the sole optimization direction. The iteration terminates when the output value of the overall cost function decreases and stabilizes within a preset tolerance range close to zero. At this point, the system extracts the globally optimal particle with the best fitness, predetermines and freezes its parameter set, encapsulates it into a read-only parameter package, and deploys it to the online terminal, achieving unmanned operation and optimization closed-loop throughout the entire parameter tuning process.

[0181] Example 8: This example includes a high-dimensional space retrieval mechanism, instruction engineering constraints, and human-computer interaction system designed to ensure the efficient and stable operation of the system on edge physical devices.

[0182] Step 801: At the moment of crop growth stage transition, perform reverse retrieval on the multidimensional offline response surface based on the crop comprehensive physiological state vector to determine the upper limit of resource budget and the baseline trigger threshold range of the current growth stage.

[0183] In the reverse retrieval or forward query of the high-frequency layer in this step, the edge computing terminal needs to perform millisecond-level positioning in a massive offline surface database. The system adopts a query architecture based on spatial indexing. Specifically, the system uses the principal component coordinates of the physiological state vector and soil moisture value as index keys to pre-partition the tensor product spline node grid of the response surface and construct a balanced Kd-tree (K-dimensional tree) spatial index. Given a query point, the system can quickly locate the local support domain of that point in logarithmic time complexity, and only needs to extract and calculate a very small number of non-zero basis functions to return the target value. This tree-like partitioning retrieval mechanism eliminates the hardware computing power bottleneck of edge farmland equipment.

[0184] Furthermore, an alternative solution for degrading high-dimensional indexes in complex crop models is provided. When the number of principal components retained during principal component analysis of physiological state vectors increases, causing the input dimension of the multidimensional offline response surface to exceed five dimensions, traditional Kd-trees suffer from the curse of dimensionality when handling nearest neighbor searches in high-dimensional spaces, leading to a sharp decline in query efficiency. Under this extreme condition, the system automatically degrades to a hierarchical grid approximate nearest neighbor retrieval strategy based on locality-sensitive hashing. This alternative strategy maps high-dimensional feature vectors to low-dimensional hash buckets, sacrificing a tiny bit of retrieval accuracy in exchange for an exponential improvement in query speed, while ensuring strict real-time performance of control command output.

[0185] Step 802: When the triggering condition is met, the weighted objective function is solved on the multidimensional offline response surface based on the dynamic target weight vector and the upper limit of the resource budget to obtain the optimal control variable instruction.

[0186] After obtaining the continuous optimal control variable instructions mathematically, the theoretical calculation value cannot be blindly issued due to the significant mechanical limitations of field irrigation and drainage facilities. Therefore, the system introduces a rigorous three-layer physical engineering truncation constraint. The first layer is a budget constraint; if the calculated value exceeds the remaining resource budget for this stage, it is truncated to within the remaining budget. The second layer is a single irrigation project upper limit constraint; the system pre-configures the maximum single water supply capacity determined by the product of the channel cross-sectional flow rate and the maximum allowable irrigation duration. If the truncated calculated value still exceeds the project's physical upper limit, it is further forcibly truncated to the upper limit level. The unmet shortfall is recorded in a dedicated under-irrigation accumulation register. To accurately track the theoretical water resource deficit, the system updates this register using the following linear accumulation formula:

[0187] R next =R current +I calc -I exec ;

[0188] Among them, R next R is the value of the under-water accumulation register for the next time step. current I is the initial value of the under-watering accumulation register at the current moment. calc I represents the theoretically calculated optimal control variable command quantity. exec This represents the actual number of commands issued after being truncated by the physical engineering limit.

[0189] In the optimization operation at the next decision moment, the accumulated value in the register will be extracted and added to the weighted objective function as an additional production loss penalty, which will prompt the system to prioritize compensatory refilling from the underlying mathematical logic.

[0190] The third layer of constraint is the minimum executable quantity filtering constraint. Frequent micro-starting and stopping of large field pumps and gate facilities can lead to energy waste and mechanical wear. Therefore, the system sets a minimum executable quantity physical threshold, which can be configured as a fixed constant within the range of five to ten millimeters. If the final command quantity after the first two layers of truncation is lower than this minimum executable quantity physical threshold, the system will forcibly reset the irrigation command to zero, i.e., actively abandon this micro-irrigation action. Simultaneously, the system will add this portion of the water that should have been irrigated to the aforementioned under-irrigation accumulation register, waiting for it to accumulate to an effective scale before consolidating and executing it centrally. Through this set of engineering constraint filtering, the theoretically optimal control variable command is successfully and safely transformed into an effective operating command that conforms to the actual needs of agricultural machinery.

[0191] After receiving the final valid operational instructions, the generated decision output is structurally encapsulated and presented by the system. All output information at each decision moment is packaged into a structured irrigation decision record. This record contains instruction fields that specify the exact irrigation amount down to the millimeter level and the recommended execution period, decision basis fields that cover data such as measured soil moisture, dynamic activation thresholds, and target weights, as well as budget tracking fields that track water consumption ratios and risk warning fields.

[0192] The aforementioned structured records are presented to agricultural managers in the form of a visual dashboard through the operating interface of the field decision-making terminal. The top area of ​​the interface highlights irrigation instructions and execution suggestions, while the middle area uses multiple line graphs to visually compare the trend of measured soil moisture values ​​with dynamic threshold trajectories at more than ten decision-making moments. The lower area uses gradient color-coded charts to display the historical evolution trajectory and future predicted trend of dynamic target weights throughout the entire growth period. The side panel dynamically displays a budget consumption progress bar and various weather warnings. Managers can use this interactive interface to accept the system's intelligent decisions and send digital control signals to the underlying controllers of physical gates or pump stations to execute automatic irrigation and drainage. In addition, the system retains a high-priority manual intervention channel, allowing operators to manually overwrite and modify irrigation amounts on the interface based on special on-site conditions and then force the issuance of the changes. The system will automatically re-include the scheduling deviations caused by manual modifications into the under-irrigation accumulation register or the global budget ledger for closed-loop tracking, ensuring seamless collaboration between the automated system and manual management.

[0193] It should be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

Claims

1. A multi-objective dynamic optimization method for paddy field moisture threshold during the growth period, characterized in that, include: Acquire real-time field environmental data, weather forecast data, and crop apparent growth indicators; The crop mechanism model is dynamically assimilated and corrected based on field environmental data and crop apparent growth indicators to obtain the crop's comprehensive physiological state vector at the current moment, including: Using field environmental data as water boundary conditions, the crop mechanism model is driven to perform day-by-day forward simulation to obtain the internal state trajectory of the crop mechanism model at the current moment. Extract the simulated apparent quantities of the crop mechanism model at the current moment, and subtract them from the crop apparent growth indicators to obtain the observation simulation bias. Based on the pre-defined causal transmission relationship between each internal state variable and the crop's apparent growth indicators, and using the gain coefficient calculated based on error variance, the corresponding dry matter pool and development rate parameters are adjusted proportionally to the observed simulation deviation. An embedded whole-plant carbon mass conservation forced compensation mechanism is implemented, specifically including: Real-time calculation of the absolute difference in dry matter mass of aboveground organs before and after a single adjustment; When the observed simulation deviation indicates an increase in the dry matter pool and the absolute mass difference is greater than zero, the absolute mass difference is deducted from the non-structural carbohydrate pool of the main stem in the crop mechanism model in an equal and reverse manner. When the balance of non-structural carbohydrate storage in the main stem is insufficient, the remaining amount will be transferred to the root dry matter bank for deduction. When the observed simulation deviation indicates a reduction in the dry matter pool and the absolute mass difference is less than zero, the reduced absolute mass will be added to the non-structural carbohydrate pool of the main stem in equal amounts. Internal physiological variables are extracted from the crop mechanism model after state assimilation correction, and the internal physiological variables are combined with crop apparent growth indicators to obtain the crop comprehensive physiological state vector at the current moment. Based on the crop integrated physiological state vector, meteorological forecast data, and pre-constructed multidimensional offline response surface, independent driving components are extracted and synthesized to obtain a dynamic target weight vector, including: Based on the developmental stage index and population physiological indicators in the crop integrated physiological state vector, physiological period sensitivity driving components are extracted. Extract forecast confidence indexes from meteorological forecast data and combine them with crop integrated physiological state vector mapping to generate forecast uncertainty driving components; On the multidimensional offline response surface, based on the crop integrated physiological state vector and field environment data, the initial marginal gradient at the zero control variable under the current state is queried and mapped to generate the marginal benefit decay driving component. The physiological period sensitivity driving component, the forecast uncertainty driving component, and the marginal benefit decay driving component are synthesized according to the addition rule. After boundary truncation and time series smoothing filtering, the dynamic target weight vector is obtained. At the moment of crop growth stage transition, a reverse retrieval is performed on the multidimensional offline response surface based on the crop comprehensive physiological state vector to determine the upper limit of resource budget and the range of benchmark trigger threshold for the current growth stage. At regular decision-making moments, the dynamic start threshold is calculated based on field environmental data, dynamic target weight vector and benchmark trigger threshold range, and the control state machine triggers the decision. When the triggering condition is met, the weighted objective function is solved on the multidimensional offline response surface based on the dynamic target weight vector and the upper limit of the resource budget to obtain the optimal control variable instruction.

2. The method according to claim 1, characterized in that, The uncertainty-driving component of mapping-generated forecasts and the marginal benefit decay-driving component of mapping-generated forecasts specifically include: The forecast confidence index is processed by a decreasing logistic function with range scaling and translation to calculate the forecast uncertainty driving components. The decreasing logistic function is configured with preset theoretical scaling upper limit, decay center point and shape parameters. By comparing the initial marginal gradient with the percentile constant of the historical marginal gradient distribution of the fertility stage, and using a continuous piecewise linear function with hard boundary truncation, the marginal benefit decay driving component is calculated. When performing additive rule synthesis, the range of values ​​of the physiological period sensitivity driving component is linearly compressed and mapped to reserve adjustment ranges for the forecast uncertainty driving component and the marginal benefit decay driving component. The original weight values ​​obtained by compression mapping and addition synthesis are substituted together with the dynamic target weight vector at the previous decision time into the exponentially weighted moving average filter function for cross-time update to obtain the dynamic target weight vector.

3. The method according to claim 1, characterized in that, When determining the resource budget ceiling for the current reproductive stage, the specific components include: On the multidimensional offline response surface, the peak yield potential corresponding to the crop's comprehensive physiological state vector is obtained, and the absolute target yield is calculated in combination with the preset target yield protection ratio. Set the lower and upper bounds of the search interval for the control variables, use the binary search numerical approximation algorithm to calculate the midpoint of the current interval, and input the midpoint of the current interval into the multidimensional offline response surface to calculate the corresponding predicted output. When the predicted output is less than the absolute target output, the left half of the interval is discarded; when the predicted output is greater than or equal to the absolute target output, the right half of the interval is discarded. The process continues iteratively until the search interval width is less than the preset tolerance threshold. The right boundary value of the interval that meets the tolerance condition is taken as the absolute minimum cumulative demand, and the resource budget ceiling for the current reproductive stage is determined by combining it with the total remaining available resources currently being tracked and maintained.

4. The method according to claim 3, characterized in that, When determining the baseline trigger threshold range, the economic break-even gradient is solved using the following method: Obtain the pre-configured unit resource economic cost and expected output price; By introducing a volume-area-depth physical conversion operator, the ratio of unit resource economic cost to expected output price is mapped to a physical metric isomorphic to the marginal product gradient field, and the economic break-even gradient is calculated. Within the current growth stage, based on the upper and lower endpoints of the expected range of change of the crop comprehensive physiological state vector obtained from historical simulation statistics of this stage, the soil moisture critical value that makes the marginal yield gradient of the multidimensional offline response surface equal to the economic break-even gradient is retrieved in reverse, forming the benchmark trigger threshold interval.

5. The method according to claim 1, characterized in that, The execution control state machine trigger determination includes: Extract the current measured root zone water content and field surface water depth from field environmental data; Establish mutually exclusive and exhaustive Boolean logic decision branches, including decision branch A and decision branch B; If the measured root zone water content is strictly greater than the dynamic start threshold, and the field surface water depth is greater than zero or the current growth stage does not require maintaining a physical water layer on the field surface, then the decision branch A is entered, the triggering condition is not met, and the control variable is forcibly set to zero. When the current measured root zone water content is less than or equal to the dynamic start threshold, or when the current measured root zone water content is strictly greater than the dynamic start threshold but the field surface water depth is zero and the current growth stage requires maintaining the field surface physical water layer, the decision falls into decision branch B, and the triggering condition is met.

6. The method according to claim 3, characterized in that, After determining the resource budget ceiling for the current reproductive stage, a cross-stage carryover mechanism is also included: When a transition occurs in the reproductive stage, the system reads the actual resource consumption that has been continuously accumulated in the previous stage, calculates the difference between the actual resource consumption and the upper limit of the resource budget in the previous stage, and determines the budget surplus or deficit status of the previous stage. When a budget surplus is determined, the surplus resources will be allocated and carried over to the upper limit of the resource budget for each subsequent reproductive stage according to a preset ratio. When a budget deficit is identified, the resource budget ceiling for each subsequent reproductive stage is deducted in order of increasing water sensitivity priority based on the preset water sensitivity priority of each reproductive stage, to ensure the overall water balance.

7. The method according to claim 5, characterized in that, When the triggering condition is met, and the optimal control variable instruction is limited to the final irrigation instruction amount, the weighted objective function is solved to obtain the optimal control variable instruction, including: Using irrigation amount as the sole independent variable, and combining the dynamic objective weight vector with the preset water-yield economic equivalent conversion rate, a one-dimensional weighted objective function is constructed. Extract the analytical first-order partial derivative of the multidimensional offline response surface with respect to irrigation amount under the current state, and establish a first-order necessary condition equation that makes the weighted marginal increase in production equal to the weighted marginal water cost. Using the upper limit of the resource budget for the current reproductive stage as the upper bound of the search interval, the golden section search algorithm is used to optimize the first-order necessary condition equation and solve for the final irrigation instruction amount.

Images