Intelligent agricultural crop demand driven multi-source linkage method and water and fertilizer intelligent allocation system

Abstract

This invention relates to the field of smart agriculture technology, and discloses a multi-source linkage method and an intelligent water and fertilizer allocation system driven by crop demand in smart agriculture. The multi-source linkage method includes: constructing an adaptive sparse sensor network and planning mobile inspection paths; using a Bayesian framework and introducing a data credibility weighting mechanism, and employing variational Kriging interpolation to achieve unified spatiotemporal resolution; achieving crop stress state identification adaptively by using a health status correction factor and a multi-state classifier; generating a crop water and fertilizer demand prediction scheme that considers yield loss risk by combining a risk value model and future weather forecasts; generating an optimal water and fertilizer allocation strategy; realizing multi-source linkage closed-loop feedback control, and continuously optimizing system parameters through an online learning mechanism to achieve adaptive evolution. This invention reduces the number of sensor nodes while ensuring monitoring coverage through adaptive sparse sensor network construction technology.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and more specifically, to a multi-source linkage method driven by crop demand in smart agriculture and a smart water and fertilizer allocation system. Background Technology

[0002] The North China Plain is an important grain-producing area in my country, with a winter wheat planting area of ​​over 15 million hectares. However, the region faces a severe water shortage problem. The annual precipitation is only 400-600 mm, and it is unevenly distributed in time and space. Rainfall is scarce during the spring wheat greening-up to heading stage (March-May), while evaporation is as high as 200-300 mm, creating a significant contradiction between water supply and demand.

[0003] In recent years, with the development of the Internet of Things, big data, and artificial intelligence technologies, smart agriculture technology has gradually emerged, and various precision agricultural management systems based on information technology have been widely researched and applied. However, existing technologies still have the following technical shortcomings in practical applications:

[0004] Existing precision agriculture systems typically require high-density sensor deployment, generally necessitating a grid density of 25m × 25m. This necessitates the deployment of numerous sensor nodes across large-scale farmlands, resulting in high initial investment costs. Furthermore, sensors suffer from high failure rates in harsh farmland environments, leading to substantial maintenance costs and severely hindering the large-scale application of smart agriculture technologies. Farmland environmental monitoring involves multiple data sources, including soil sensors (point measurements), remote sensing imagery (area measurements), and meteorological data (regional measurements). These sources suffer from issues such as mismatched spatiotemporal scales, varying sampling frequencies, and inconsistent data quality. Existing fusion algorithms lack dynamic evaluation mechanisms for data reliability, resulting in low accuracy of fusion results and failing to provide reliable support for precise decision-making. Different crop varieties exhibit varying physiological characteristics and water response mechanisms. Existing water stress identification models are mostly based on single-variety calibration, lacking variety adaptability. Additionally, agricultural production faces various uncertainties such as pests and diseases, nutrient deficiencies, weather changes, equipment failures, and water source fluctuations. Existing systems lack effective risk assessment and emergency response mechanisms, making them prone to decision-making biases or execution failures under abnormal conditions, thus failing to guarantee crop production safety. Therefore, there is an urgent need for a smart agricultural water and fertilizer precision allocation system that can solve the above-mentioned technical problems, achieve low-cost and high-reliability farmland environment perception, improve the accuracy of multi-source data fusion, enhance the adaptability of crop demand identification, and have risk perception and emergency response capabilities, so as to meet the actual needs of modern precision agricultural production. Summary of the Invention

[0005] This invention provides a multi-source linkage method and an intelligent water and fertilizer allocation system driven by crop demand in smart agriculture, which solves the technical problems in related technologies such as high sensor deployment costs, insufficient accuracy of multi-source data fusion, lack of variety adaptability in crop stress identification, and poor system robustness.

[0006] This invention provides a multi-source linkage method driven by crop demand in smart agriculture, including:

[0007] An adaptive sparse sensor network is constructed and a mobile inspection path is planned using a layout optimization model and a dynamic programming algorithm.

[0008] Based on the Bayesian framework and by introducing a data credibility weighting mechanism, multi-source heterogeneous data collected by an adaptive sparse sensor network are fused, and the spatiotemporal resolution is unified by using the variational kriging interpolation method.

[0009] A variety-specific crop water stress index model was constructed using unified data, and crop stress state identification with variety adaptation was achieved through health status correction factor and multi-state classifier.

[0010] Based on the identified stress state probabilities, a crop water and fertilizer demand forecasting scheme that takes into account the risk of yield loss is generated by combining the value at risk model and future weather forecasts.

[0011] For crop water and fertilizer demand prediction schemes, multi-objective robust optimization is performed to generate the optimal water and fertilizer allocation strategy;

[0012] The optimal water and fertilizer allocation strategy is input into the adaptive PID controller to achieve multi-source linkage closed-loop feedback control, and the system parameters are continuously optimized through an online learning mechanism to achieve adaptive evolution.

[0013] Furthermore, the objective function of the sensor layout optimization model for the adaptive sparse sensor network is to minimize the weighted sum of deployment cost, monitoring error, and maintenance difficulty, where the cost weight is 40%, the monitoring error weight is 40%, and the maintenance difficulty weight is 20%, while satisfying the constraints that the monitoring coverage is not less than 90% and the redundancy is not less than 2.

[0014] Furthermore, the data credibility weighting mechanism is expressed as follows:

[0015] ;

[0016] in The posterior probability distribution of the fused parameters. , , They represent the 1st, 2nd, and 3rd respectively. Observational data from multiple data sources; For the first Likelihood function of each data source; The parameter is a prior distribution; For the first Data credibility weights for each data source; Proportional to the sign; The symbol for multiplication; This represents the number of data sources.

[0017] Furthermore, the variety-specific crop water stress index is specifically manifested as follows:

[0018] ;

[0019] in The water stress index is a variety-specific index. For the blade temperature; Air temperature; This serves as the lower limit baseline for variety specificity. This serves as the upper limit baseline for variety specificity. For effective accumulated temperature; This refers to the atmospheric water vapor pressure difference. A health status correction factor;

[0020] Introducing a lower limit baseline for variety specificity and the upper limit of variety specificity baseline The stress assessment criteria are dynamically adjusted by considering the physiological characteristics of different varieties; at the same time, a health status correction factor is incorporated. This enabled the correction for non-moisture stress factors.

[0021] Furthermore, the constraints of the multi-objective robust optimization include that the hydraulic uniformity coefficient of the irrigation system is not less than 0.85, the total amount of fertilizer applied in a single application does not exceed 110% of the maximum absorption threshold of the crop, and the total operating cost is within the preset budget range.

[0022] Furthermore, the multi-source linkage closed-loop feedback control constructs an adaptive proportional, integral, and derivative controller to achieve precise water and fertilizer regulation in each control zone. The control output calculation formula is as follows:

[0023] ;

[0024] in For the first Each partition at time Control output; For the first Each partition at time Proportional control gain; For the first Each partition at time Integral control gain; For the first Each partition at time The differential control gain; For the first Each partition at time Control error; For integration variables; From 0 to The definite integral of .

[0025] Furthermore, the continuous optimization of system parameters through an online learning mechanism to achieve adaptive evolution includes:

[0026] An online incremental learning mechanism is established, and an improved stochastic gradient descent algorithm is used to update the model parameters online, allowing the model to be continuously updated when new data is obtained without retraining.

[0027] An exponential decay mechanism is used to ensure that new data has a higher weight, while the weight of historical data decays over time.

[0028] Establish a multi-dimensional indicator system for model performance evaluation, including water use efficiency, prediction accuracy, system reliability, and economic benefits;

[0029] The multi-armed slot machine algorithm is used for model selection, and the confidence upper bound algorithm is used to make the optimal choice under uncertain conditions.

[0030] A long-term optimization strategy based on the Q-learning reinforcement learning framework is established. The optimal strategy is achieved by learning the state-action value function, thereby realizing the continuous optimization and adaptive evolution of the system.

[0031] Furthermore, the stress state is divided using a Softmax classifier, resulting in four states:

[0032] when When, it is a state of no coercion; when At that time, it was a state of mild stress; when At that time, it was a state of moderate stress; when At that time, it was a state of severe stress.

[0033] Furthermore, the stress state analysis uses a hidden Markov model to model the time series of crop stress state and predict the evolution trend of stress state in the next 24 hours.

[0034] This invention provides a smart agricultural crop demand-driven intelligent water and fertilizer allocation system for executing the aforementioned smart agricultural crop demand-driven multi-source linkage method, including:

[0035] The data acquisition module is used to acquire multi-source heterogeneous data through an adaptive sparse sensor network;

[0036] The data fusion module is used to perform high-precision fusion and uncertainty quantification of the multi-source heterogeneous data based on a Bayesian framework and data credibility weights.

[0037] The stress identification module is used to accurately identify the stress status of different crop varieties by using a variety-adaptive crop water stress index model combined with health status correction.

[0038] The demand forecasting module is used to generate crop water and fertilizer demand forecasting schemes that take risks into account, based on the value at risk model and combined with stress conditions and weather forecasts.

[0039] The optimization decision module is used to perform multi-objective robust optimization using the NSGAIII genetic algorithm to generate the optimal water and fertilizer allocation strategy.

[0040] The linkage control and learning module is used to execute closed-loop feedback control according to the optimal strategy through an adaptive PID controller, and continuously optimize the model parameters and control strategy in the system through an online incremental learning mechanism based on the actual execution effect.

[0041] The beneficial effects of this invention are as follows: by using adaptive sparse sensor network construction technology, the number of sensor nodes is reduced while ensuring monitoring coverage. At the same time, the intelligent maintenance system can predict equipment failures and achieve preventive maintenance, effectively reducing the initial investment and subsequent operation and maintenance costs of the system, and laying an economic foundation for the large-scale promotion of intelligent agricultural technology.

[0042] By combining Bayesian data fusion with variational Kriging interpolation, a data credibility weighting mechanism was established, which achieved high-precision fusion of multi-source heterogeneous data and provided scientific and reliable data support for subsequent decision-making through uncertainty quantification, thus solving the technical problem of insufficient fusion accuracy of traditional methods.

[0043] This invention proposes a variety-specific method for calculating crop water stress index. Through multi-state stress classification and health status correction mechanism, it realizes adaptive stress identification for different crop varieties, improves the accuracy of water stress judgment, and provides a reliable decision-making basis for precise water and fertilizer allocation.

[0044] A robust multi-objective optimization framework considering uncertainty was established, and closed-loop feedback control with multi-source linkage was realized. This enabled the system to maintain stable operation when facing abnormal situations such as weather changes and equipment failures, effectively reducing the risk of yield loss under extreme conditions and improving the safety and reliability of agricultural production.

[0045] Through online incremental learning mechanisms and distributed collaborative control technology, the system can continuously optimize model parameters and control strategies based on actual operating results, thereby achieving continuous improvement in prediction accuracy and rapid execution of control response, providing technical support for the long-term stable operation of intelligent agricultural systems. Attached Figure Description

[0046] Figure 1 This is a flowchart of the multi-source linkage method driven by crop demand in smart agriculture in this invention. Detailed Implementation

[0047] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0048] At least one embodiment of the present invention discloses a multi-source linkage method driven by crop demand in smart agriculture, such as... Figure 1 As shown, it includes:

[0049] Step 1: Using a layout optimization model and dynamic programming algorithm, construct an adaptive sparse sensor network and plan a mobile inspection path;

[0050] The specific inputs for this step include: farmland topography data. Three-dimensional topographic information of farmland was obtained through GPS measurement and drone aerial photography. The data format is raster data with a spatial resolution of 1m×1m; soil type distribution map. Spatial distribution maps of soil texture, pH value, organic matter content, and other parameters obtained through soil sampling analysis are stored in vector polygon format; historical yield spatial distribution data... Wheat yield data from the past 3-5 years, including GPS coordinates and corresponding yield values, with a data accuracy requirement of 100 kg / ha; economic constraint parameters. The total budget for sensor deployment set by the user, in yuan, with a range of 50,000 to 500,000 yuan.

[0051] The processing procedure specifically includes:

[0052] To construct an adaptive sparse sensor network, this invention first establishes a sensor layout optimization model.

[0053] The core objective of the sensor layout optimization model is to achieve optimal monitoring performance within a limited budget. The sensor layout optimization function in this invention aims to find the optimal sensor layout scheme by minimizing the weighted sum of cost, error, and maintenance difficulty.

[0054] Specifically, the sensor layout optimization model comprehensively evaluates the deployment cost of each potential sensor location, the accuracy of its monitoring range, and the ease of subsequent maintenance. It balances these three factors using a set of weighted coefficients (cost 40%, monitoring error 40%, and maintenance difficulty 20%) calibrated by expert experience and historical data. This allows it to find the optimal solution within the total budget while ensuring coverage of at least 90% and redundancy of at least 2 (i.e., each point is covered by at least two sensors).

[0055] To further improve the flexibility and efficiency of the monitoring system, this invention also optimizes the inspection path for mobile sensor systems (such as drones) based on the fixed sensor layout. The system assigns different priorities and urgency levels to different monitoring points according to the crop growth stage and real-time risk assessment results, and then uses a dynamic programming algorithm to calculate the inspection path with the lowest total cost (considering both flight distance and monitoring importance).

[0056] The drone inspection path is optimized using a dynamic programming algorithm:

[0057] ;

[0058] in The optimal inspection path; The parameters are used to find the minimum value; For the set of all possible paths; The summation symbol; This represents the total number of path segments; For the first Flight distance along a segment of the path; For the first Priority of monitoring points along the path segment; For the first The urgency coefficient of the path segment.

[0059] To ensure the long-term stable operation of the entire monitoring system, this invention also designs an equipment health status assessment model.

[0060] The equipment health status assessment model continuously evaluates the health status of each sensor by analyzing multiple indicators such as sensor measurement accuracy, data stability, remaining power, and communication success rate in real time, and combines them with weights obtained through machine learning training, so as to carry out timely maintenance or replacement.

[0061] Specifically, the equipment health status assessment adopts a multi-index fusion model, which comprehensively considers four indicators: measurement accuracy, stability, battery status, and communication status, with weights of 0.4, 0.3, 0.2, and 0.1, respectively. These weights are obtained through machine learning training on six months of operating data from 100 sensor nodes.

[0062] Based on the health status assessment results, this invention constructs a complete intelligent maintenance system, specifically as follows:

[0063] Health status assessment indicators: Measurement accuracy, deviation from standard value <5% is normal; Data stability, coefficient of variation <10% over 24 consecutive hours; Battery status, power >20% is normal, <10% is an emergency; Communication status, success rate >95% within 24 hours is normal.

[0064] Maintenance decision rules: Preventive maintenance is triggered when the health score is <0.7; Emergency maintenance is triggered immediately when any key indicator is abnormal; Planned maintenance is performed periodically according to the equipment's life cycle.

[0065] Execution process: Fault diagnosis, automatically identifying fault type and location; maintenance scheduling, optimizing maintenance path and time arrangement; resource allocation, rationally allocating maintenance personnel and spare parts; effect evaluation, evaluating the performance recovery status after maintenance.

[0066] To ensure the accuracy and comparability of the model, the input parameters need to be standardized before establishing the sensor layout optimization model:

[0067] Topographic data, such as elevation, slope, and aspect, are standardized using Z-scores to eliminate dimensional differences.

[0068] Soil parameters, such as pH, organic matter content, and bulk density, were normalized to the [0, 1] interval using Min-Max. Historical yield data were processed using quantile normalization to avoid the influence of extreme values. Economic constraint parameters were normalized by applying regional average cost to the budget amount.

[0069] Through the above processing steps, the system first obtains an optimized sensor layout scheme based on the sensor layout optimization model. This scheme includes key information such as the GPS coordinates, sensor type, and installation parameters of each sensor node, stored in JSON format. Simultaneously, based on the training results of the equipment health status assessment model, a system capable of real-time evaluation of equipment health status is obtained. The model is implemented using a multi-index fusion machine learning method (specifically, a random forest algorithm model trained on 6 months of running data from 100 sensor nodes) and saved in pickle format for subsequent equipment status monitoring and maintenance decisions.

[0070] Step 2: Based on the Bayesian framework and by introducing a data credibility weighting mechanism, multi-source heterogeneous data collected by the adaptive sparse sensor network are fused, and the spatiotemporal resolution is unified by using the variational kriging interpolation method.

[0071] The specific inputs for this step include: fixed sensor data. Real-time monitoring data from the fixed sensor network deployed in step 1, including parameters such as soil moisture, temperature, and conductivity, is collected hourly; mobile sensor data... Periodic monitoring data from sensors mounted on drones, including multispectral imagery and rapid soil detection data, are collected 2-3 times per week; satellite remote sensing data. Multispectral and hyperspectral remote sensing images from satellites such as Landsat and Sentinel, with spatial resolution of 10–30 m and temporal resolution of 5–16 days; meteorological station data. Meteorological data, including temperature, humidity, precipitation, and wind speed, are collected from regional meteorological stations at a frequency of once per hour; equipment health status. The sensor health status assessment results from step 1.

[0072] The processing procedure specifically includes:

[0073] The core of this step is to establish a multi-source data fusion framework based on Bayesian theory. Through the Bayesian probability model, multi-source heterogeneous data is scientifically fused. It mainly considers sensor data, UAV data, satellite remote sensing data and weather station data. Unlike traditional methods, the multi-source data fusion framework introduces an improved data credibility weighting mechanism, which dynamically allocates weights by comprehensively evaluating the three key factors of reliability, freshness and accuracy of each data source.

[0074] The data credibility weighting mechanism is represented as follows:

[0075] ;

[0076] in The posterior probability distribution of the fused parameters. , , They represent the 1st, 2nd, and 3rd respectively. Observational data from multiple data sources; For the first Likelihood function of each data source; The parameter is a prior distribution; For the first Data credibility weights for each data source; Proportional to the sign; The symbol for multiplication; Number of data sources;

[0077] Data credibility weight The calculation is performed using a normalization method. The specific implementation is as follows:

[0078] ;

[0079] in For the first Data credibility weights for each data source; and The first The first data source and the first The reliability of each data source; and The first The first data source and the first The freshness of each data source; and The first The first data source and the first The accuracy of each data source; The summation symbol; Index the data source; This represents the total number of data sources.

[0080] Using the aforementioned Bayesian data fusion framework and data credibility weighting mechanism, the system will use fixed sensor data. Mobile sensor data Satellite remote sensing data and weather station data Integrate the systems while taking into account the health status of the equipment. The impact on data reliability ultimately leads to the generation of a spatiotemporally continuous farmland environment monitoring dataset. This dataset contains continuous distribution information on key parameters such as soil moisture, temperature, and nutrients throughout the farmland.

[0081] The specific integration process is as follows: Time alignment and spatial registration are performed on each data source to unify data with different spatiotemporal resolutions into the same reference system; the reliability of each data source is assessed based on the sensor health status, while also considering the freshness of the data and the accuracy of historical verification; the comprehensive weight of each data source is calculated based on the three factors of reliability, freshness, and accuracy; Bayesian theory is used to fuse the data sources according to their weights to generate a preliminary fused dataset; the confidence interval of the fusion result is calculated to evaluate the reliability of the data.

[0082] However, due to the inconsistent spatial resolution of the input data sources, Spatial discontinuities still exist. To address this issue, an improved variational kriging interpolation method is employed:

[0083] ;

[0084] in For prediction points The interpolation result; For the first The weight of each observation point; For the first Observed values ​​at each observation point; For prediction points The random error term; The summation symbol; This represents the number of observation points; This is for predicting the location of the point.

[0085] The constraints are:

[0086] ;

[0087] in For the first The weight of each observation point; It is a variogram; It is a Lagrange multiplier; For the first The location of each observation point; For the first The location of each observation point; To predict the location of the point; The summation symbol; This represents the number of observation points.

[0088] To improve interpolation accuracy, this invention modifies the traditional variogram function by taking into account the influence of data density.

[0089] ;

[0090] in It is a variogram; For the nugget effect; This is the value of the partial sill. For distance; These are the range parameters; This is a data density correction function.

[0091] Data density correction function The specific implementation is as follows: First, calculate the given distance. The density of observation points within the specified range is compared with a preset reference density to obtain the base density ratio. This ratio is then mapped to a range of 0.5 to 1.5 using an exponential transformation, forming a density adjustment coefficient. When the observation point density is higher than the reference density, the correction coefficient is greater than 1, enhancing the variogram value and reflecting subtle variations within high-density areas. When the observation point density is lower than the reference density, the correction coefficient is less than 1, reducing the variogram value and avoiding over-interpretation of variability in sparse data areas. Furthermore, the function includes a smooth transition mechanism, using distance weight decay to ensure a smooth transition of the variogram between different density regions, preventing abrupt changes. This function effectively solves the accuracy problem of traditional Kriging interpolation under non-uniform sampling conditions, enabling the interpolation results to better adapt to the complex spatial distribution characteristics of farmland environments.

[0092] After completing spatial interpolation, for Uncertainty quantification was performed on each monitoring parameter. An error propagation analysis method was used, comprehensively considering errors in the original data, weight calculation, and interpolation model. The 95% confidence interval for each monitoring parameter was calculated using Monte Carlo simulation. Specifically, thousands of random perturbation simulations are performed for each prediction point, and the upper and lower limits of the confidence interval are determined based on the distribution characteristics of the perturbation results. Finally, the system is evaluated based on multiple indicators. The overall credibility score was determined by weighted calculation. The evaluation process considered factors such as the original credibility weights of each data source, the prediction error of interpolation validation points, the spatiotemporal coverage of the data, and the stability of the fusion process. The value range is 0-1.

[0093] Based on the above processing steps, the farmland environment monitoring dataset Confidence interval Overall credibility score of data fusion results Together, they constitute the complete monitoring and assessment results of the farmland environment by the system.

[0094] Step 3: Construct a variety-specific crop water stress index model using the unified data, and achieve variety-adaptive crop stress state identification through health status correction factors and multi-state classifiers.

[0095] The input for this step specifically includes: the farmland environmental monitoring dataset from step 2. Wheat variety information Information including variety name, physiological characteristic parameters, drought resistance level, etc., is input by the user or obtained from a crop variety database; growth stage data. Crop growth and development stage information based on accumulated temperature calculations; pest and disease monitoring data. Data on the occurrence of pests and diseases were obtained through image recognition and expert investigation.

[0096] The processing procedure specifically includes:

[0097] Traditional methods for calculating the Crop Water Stress Index (CWSI) often neglect physiological differences among varieties, resulting in a lack of specificity in stress assessment. To address this issue, this invention proposes a variety-adaptive Crop Water Stress Index calculation model:

[0098] ;

[0099] in The water stress index is a variety-specific index. For the blade temperature; Air temperature; This serves as the lower limit baseline for variety specificity. This serves as the upper limit baseline for variety specificity. For effective accumulated temperature; This refers to the atmospheric water vapor pressure difference. A health status correction factor.

[0100] The unique feature of this model is that it introduces a variety-specific baseline function for the first time. and The stress assessment criteria are dynamically adjusted by considering the physiological characteristics of different varieties; at the same time, a health status correction factor is innovatively added. This allows for correction of non-moisture stress factors such as pests and diseases. This makes stress identification more accurate and adaptable to different varieties.

[0101] in The water stress index is a variety-specific index. For the blade temperature; Air temperature; This serves as the lower limit baseline for variety specificity. This serves as the upper limit baseline for variety specificity. For effective accumulated temperature; This refers to the atmospheric water vapor pressure difference. A health status correction factor.

[0102] The specific implementation is as follows: First, a baseline model is established. By analyzing the linear relationship between the minimum temperature difference between leaves and air and the atmospheric vapor pressure difference under non-water stress conditions, the slope and intercept parameters are obtained. Then, the baseline parameters are adjusted according to the physiological characteristics of different varieties (such as stomatal conductance, leaf structure, root development, etc.), while introducing an effective accumulated temperature factor to reflect the influence of growth stages on the baseline. The specific adjustment process involves calculating the variety correction coefficient using genotype characteristic data from the variety parameter database, multiplying this coefficient by the effective accumulated temperature correction function to obtain the total adjustment amount, and finally applying the total adjustment amount to the baseline model. This method ensures that, under the same atmospheric vapor pressure difference, drought-resistant varieties have lower baseline values, reflecting their stronger water retention capacity; and as the effective accumulated temperature increases (as the growth stage progresses), the baseline will dynamically adjust to adapt to the changing patterns of crop growth and development. This function realizes the adaptive adjustment of the water stress determination criteria to different varieties and growth stages.

[0103] The specific implementation is as follows: Based on this, a variety-specific stress threshold offset is added. This offset is calculated by multiplying the variety's maximum drought tolerance parameter, the stress sensitivity coefficient of the current growth stage, and the intensity of environmental stress, ensuring that different varieties have different stress judgment criteria under the same environmental conditions.

[0104] The specific implementation is as follows: First, the disease index, pest index, and nutrient stress index are normalized. Then, a comprehensive health loss index is calculated by weighted summation, with weights of 0.5 for disease, 0.3 for pests, and 0.2 for nutrient stress (determined based on the degree of influence of water stress identification). Finally, the formula is modified. The health status is converted into a correction factor to ensure that the worse the health status, the lower the sensitivity of water stress identification.

[0105] Variety-specific parameters are obtained through transfer learning. These parameters refer to the unique physiological characteristics exhibited by different crop varieties during their growth and development. Specifically, the parameters are first established based on the calibration parameters of known varieties (such as Jimai 22), and then adaptive adjustments are made according to the characteristics of the new variety.

[0106] The specific adjustment method is as follows: basic parameters are added to varietal difference parameters, which are calculated through a deep learning network. This network takes varietal characteristics (such as plant height and root depth), climate data (such as temperature and precipitation), and soil properties (such as texture and pH) as input, and outputs the parameter differences between varieties. This method can effectively utilize existing varietal knowledge to quickly obtain parameters for new varieties, avoiding the need for repeated large-scale field trials for calibration. It also considers the influence of environmental factors, improving the adaptability of the parameters. To achieve this goal, this invention uses a Multi-Layer Perceptron (MLP) neural network as the specific implementation of the deep learning model. The MLP network structure consists of three fully connected layers, with 64 and 32 neurons in the hidden layers, respectively. The activation function is ReLU, and the output layer uses linear activation. Variety characteristics include 12 morphological indicators such as plant height, root depth, and leaf area index.

[0107] The specific structure of a multi-layer perceptron (MLP) neural network:

[0108] The input layer consists of variety feature vectors with 12 dimensions (including morphological indicators such as plant height, root depth, and leaf area index).

[0109] Hidden layer 1 has 64 neurons and the activation function is ReLU; hidden layer 2 has 32 neurons and the activation function is ReLU; the output layer is a variety parameter difference vector with a dimension of 8 and the activation function is linear; the dropout layer is added after the hidden layers with a dropout rate of 0.3.

[0110] Training parameter settings:

[0111] Learning rate, 0.001, using Adam optimizer; batch size, 32; number of training epochs, 200; loss function, mean squared error (MSE); regularization, L2 regularization, weight decay coefficient is 0.0001; early stopping policy, training stops when the validation set loss does not decrease for 10 consecutive epochs.

[0112] Subsequently, the system not only determines whether the crop is under stress, but also uses a Softmax classifier to classify it into:

[0113] when At that time, it was a state of no coercion;

[0114] when At that time, it was a state of mild stress;

[0115] when At that time, the state was considered moderate stress.

[0116] when At that time, it was a state of severe stress.

[0117] The state probability distribution is calculated using the Softmax function:

[0118] ;

[0119] ;

[0120] in This is the probability distribution vector of the stress state; It is a normalized exponential function; The probability of a state without stress; The probability of a mild stress state; The probability of a moderate stress state; The probability of a severe stress state; For the first The probability of a stress state; For the first The transpose weight matrix of each state; This is a variety-specific crop water stress index; Normalized Difference Vegetation Index; Soil moisture content; It is a meteorological stress index; Information on phenological periods; For the first Bias vectors for each state.

[0121] Finally, a Hidden Markov Model (HMM) is introduced to analyze the temporal evolution of stress states. For example, the model knows that "mild stress" is more likely to develop into "moderate stress" rather than jumping directly to "no stress." This temporal analysis capability makes state predictions more consistent with natural laws.

[0122] Specifically, a 4×4 stress state transition matrix was constructed by analyzing historical monitoring data. Each element in the matrix ( ) indicates a state of coercion Transition to state The probability of transition is calculated based on the frequency statistics of state transitions in continuous observation data, and the data sparsity problem is handled using Bayesian smoothing techniques. The matrix construction employs a piecewise calculation method, estimating the transition probability according to the physiological characteristics of different crop growth stages to reflect the dynamic changes in the water response mechanism during growth. For example, during the jointing stage, the matrix shows a higher probability of upward transition (state deterioration), reflecting the high sensitivity to water stress at this stage; while during the grain-filling stage, it shows stronger state stability, reflecting the enhanced adaptability of mature crops. The system also introduces weather-related conditional probability adjustments, considering the impact of future meteorological conditions on the transition probability when predicting future stress states. This is achieved through the stress state transition matrix. It can predict short-term stress conditions, providing a forward-looking reference for irrigation decisions. The element values ​​in the matrix are updated online every 24 hours based on new observation data, ensuring that the prediction model can adapt to environmental changes and crop growth dynamics.

[0123] The Hidden Markov Model (HMM) is introduced as follows: The system first sets the crop water stress state into four hidden states (no stress, mild stress, moderate stress, and severe stress), and uses the actual observed monitoring indicators (such as leaf temperature, soil moisture, and vegetation index) as the observation sequence. The state transition matrix is ​​constructed through historical data mining. Specifically, it analyzes the evolution of stress states for different varieties under various weather conditions over the past three years, statistically analyzes the transition frequencies between states, and forms a 4×4 transition probability matrix. The system also establishes an observation probability distribution model, using a multivariate Gaussian mixture model to describe the joint distribution characteristics of various observed indicators under each stress state.

[0124] Before constructing a variety-adaptive CWSI model, the relevant parameters need to be preprocessed:

[0125] Variety information coding: One-Hot coding method is used to process variety categories (such as drought-resistant, high-yielding, high-quality, etc.), and the categorical variables are converted into numerical vectors;

[0126] Physiological parameter standardization: Z-score standardization was performed on variety-specific parameters such as leaf area index, root depth, and drought resistance coefficient;

[0127] Pest and disease index: Convert discrete ratings such as disease severity and pest density into continuous values ​​and normalize them to [0, 1].

[0128] Nutritional status indicators: Nutrient parameters such as nitrogen, phosphorus, and potassium content and chlorophyll index were normalized to avoid the influence of outliers.

[0129] Based on the above processing steps, the variety-specific water stress index is first obtained through the variety-adaptive CWSI calculation model. Then, the multi-state stress probability distribution is calculated using a Softmax classifier and the input feature vector. The system includes probability values ​​for four stress states. Finally, based on state transition analysis using a hidden Markov model, a 4×4 stress state transition matrix describing the transition probabilities between different stress states is generated. These output parameters together constitute the system's complete assessment of crop water stress status.

[0130] Step 4: Based on the identified stress state probabilities, combine the risk value model and future weather forecasts to generate a crop water and fertilizer demand forecasting scheme that takes into account the risk of yield loss.

[0131] The specific inputs for this step include: variety-specific water stress index. Results of water stress assessment from step 3; stress probability distribution Multi-state stress probability distribution from step 3; weather forecast data Weather forecast information for the next 7-15 days; soil moisture dynamics. Soil moisture variation trend calculated based on water balance model.

[0132] The processing procedure specifically includes:

[0133] This step fully considers various uncertainties when predicting the future water and fertilizer requirements of crops. Instead of using a fixed water stress coefficient, it creates a "fuzzy water stress coefficient" for water requirement prediction. This coefficient is obtained by weighted averaging the probabilities of occurrence of the four stress states (no, mild, moderate, and severe) calculated in step 3. This allows the water requirement prediction to dynamically reflect the current true health status of the crop.

[0134] Introducing fuzzy moisture stress coefficient and uncertainty quantification:

[0135] ;

[0136] in For water demand, For reference crop evapotranspiration, calculations were performed according to the Penman-Monteith formula; The crop-specific coefficient is the effective accumulated temperature. The function; The fuzzy water stress coefficient; For irrigated area; is the uncertainty coefficient.

[0137] Variety-specific crop coefficient This is an improvement on the traditional FAO crop coefficient, using a piecewise continuous function to describe the changes in water consumption characteristics of different crop varieties throughout their growth cycle. First, a baseline model is established based on the standard crop coefficient curve (three key points: early, middle, and late stages). Then, a variety adjustment factor is introduced for correction. Variety adjustment is based on three key physiological characteristics: leaf area development rate, root water uptake capacity, and stomatal regulation characteristics. The adjustment coefficient is automatically calculated based on morphological and physiological parameters from the variety database. The adjusted crop coefficient exhibits an S-shaped curve with changes in effective accumulated temperature, rising slowly in the early stage, reaching a peak in the middle stage, and gradually decreasing in the later stage.

[0138] Fuzzy water stress coefficient The stress coefficients were calculated using a state probability weighted average method. The stress coefficients for each of the four stress states were weighted and summed based on their probability of occurrence. The stress coefficients for different stress states are as follows:

[0139] The stress coefficient is 1.0 under no stress; 0.8 under mild stress; 0.6 under moderate stress; and 0.4 under severe stress.

[0140] Simultaneously, the system also establishes a comprehensive risk assessment model, using a Bayesian network to dynamically update the probability of occurrence of various agricultural meteorological disasters such as drought, waterlogging, salinization, and frost damage. Ultimately, the system outputs not a single water and fertilizer requirement value, but an interval estimate with a 95% confidence level, such as "water requirement is 10-15 mm," along with a comprehensive risk index, providing more comprehensive information for subsequent decision optimization.

[0141] The uncertainty coefficient comprehensively considers multiple error sources and adopts the principle of variance synthesis, i.e., the uncertainty coefficient... By calculating weather forecast errors Model structure error Sensor measurement error and spatial variation error The square root of the sum of the squares of these four errors is obtained.

[0142] in, Due to weather forecast error, For model structure error, For sensor measurement error, This represents spatial variability error. Each error component was obtained through cross-validation analysis of historical data.

[0143] Based on the aforementioned uncertainty coefficient and fuzzy water stress coefficient, the system employs the Monte Carlo simulation method to estimate the water and fertilizer demand intervals. The specific implementation process is as follows: First, a parameter probability distribution model for water and fertilizer demand is constructed, and then... , , Key parameters are treated as random variables, and appropriate probability distribution functions such as normal distribution, triangular distribution, or beta distribution are assigned to them based on their historical data characteristics. Then, large-scale random sampling (usually 10,000 times) is performed. Each sampling obtains a set of random values ​​from each parameter distribution, which are then substituted into the water and fertilizer requirement calculation formulas to obtain a possible result value. This forms the probability distribution curves for water and fertilizer requirements. The system extracts the 2.5% and 97.5% quantile values ​​from these curves, which are used as the interval values ​​for water requirement calculation. and fertilizer requirement range The lower and upper limits are defined. This interval estimation method can comprehensively reflect the combined impact of various uncertainties, providing decision-makers with more objective demand assessment information and reducing the risk of water and fertilizer resource allocation. For short- to medium-term forecasts (3 to 7 days), the system will also dynamically adjust the interval range according to the weather forecast update frequency, making the forecast results increasingly refined over time. The forecast interval becomes narrower as it approaches the execution date, reflecting the improvement in forecast certainty. Among them, the water demand interval is calculated using a baseline value through an evapotranspiration model, and combined with uncertainties such as weather forecast errors and soil moisture monitoring errors, a probability distribution is generated using the Monte Carlo simulation method, with the 2.5% and 97.5% quantiles taken as the base values. and The fertilizer requirement range is calculated based on the baseline values ​​from the crop nutrient demand model. Considering the variability of soil nutrient supply capacity and fertilizer utilization rate, the probability distribution is obtained using Monte Carlo simulation, with the 2.5% and 97.5% quantiles taken as the baseline values. and .

[0144] Fertilizer requirement prediction is based on a model established using the principles of nutrient balance and crop absorption kinetics.

[0145] ;

[0146] in Indicates the amount of fertilizer required. This refers to the amount of nutrients absorbed by crops. For the amount of nutrients lost, For soil nutrient supply, Due to the uncertainty of fertilization.

[0147] Crop nutrient uptake The calculation is as follows: First, the target yield needs to be determined, based on historical yield data and current management practices. Then, the crop's nitrogen content characteristics are considered; for wheat, the nitrogen content in the grain is generally between 2.2% and 2.8%. The harvest index, the ratio of grain yield to aboveground biomass, must also be considered; for wheat, this index is typically between 0.45 and 0.50. Finally, nitrogen use efficiency must be taken into account; this efficiency value is generally between 0.3 and 0.6.

[0148] Taking all these factors into account, the total nitrogen uptake by crops can be calculated. Specifically, nitrogen uptake equals the target yield (in kilograms per hectare) multiplied by the grain nitrogen content (percentage) and the harvest index, then divided by the nitrogen use efficiency. The result calculated in this way is the total amount of nitrogen that crops need to absorb during their growth.

[0149] Risk assessment index A weighted summation method was adopted, taking into account four types of risks: drought, waterlogging, salinization, and frost damage, with weights of 0.4, 0.2, 0.2, and 0.2, respectively.

[0150] Each risk probability is dynamically updated via a Bayesian network, and the probability of occurrence for each type of risk is updated dynamically based on real-time observed environmental condition evidence. This environmental condition evidence includes meteorological data (such as temperature, humidity, and precipitation), soil conditions (such as water content and nutrient content), and crop growth status (such as growth vigor and pest and disease conditions). The system uses a Bayesian network approach to update the likelihood of each type of risk by calculating conditional probabilities. Specifically, for a certain type of risk, its probability of occurrence is equal to the conditional probability of that risk occurring given the observed environmental condition evidence. This conditional probability can be calculated using Bayes' theorem, which is the product of the probability of observing current evidence given the risk's occurrence and the prior probability of that risk, divided by the overall probability of observing current evidence.

[0151] Before conducting demand forecasting, the input variables need to be preprocessed:

[0152] Stress state probability: Ensure that the sum of the probabilities of the four stress states is 1, and perform Softmax normalization.

[0153] Meteorological forecast data: Meteorological elements such as temperature, humidity, and precipitation are smoothed using a sliding window and standardized using Z-score based on historical statistics;

[0154] Soil moisture dynamics: Soil moisture content is converted into relative moisture content (percentage of field capacity) and normalized to [0,1].

[0155] Risk factors: Risk factors such as drought, waterlogging, salinization, and frost damage are weighted and normalized, with the weights determined based on the local historical disaster frequency.

[0156] Based on the above processing steps, the estimated water and fertilizer demand intervals are output. , Risk assessment index Based on the comprehensive risk assessment results, the value range is 0-1; and the confidence interval of the demand forecast.

[0157] Step 5: Perform multi-objective robust optimization on the crop water and fertilizer demand prediction scheme to generate the optimal water and fertilizer allocation strategy.

[0158] The specific inputs for this step include: water and fertilizer requirement zones. , Demand forecast range from step 4; risk assessment The comprehensive risk index from step 4; resource constraints These include resource constraints such as water supply capacity, fertilizer inventory, and equipment power; and environmental constraints. Environmental constraints include environmental protection requirements, soil carrying capacity, and climate suitability.

[0159] The processing procedure specifically includes:

[0160] The core of this step is to establish a robust multi-objective optimization model that considers uncertainty. Traditional optimization methods typically solve under ideal conditions, while the model of this invention aims to find the optimal solution under worst-case conditions. It takes maximizing output, minimizing water resource consumption, and minimizing total cost as its three core optimization objectives:

[0161] ;

[0162] in Minimize operator; For the maximize operator; For decision variable vectors; For the decision variable space; For the uncertain parameter vector; It is a set of uncertainties; The objective function vector; Let the output objective function be... The objective function is the water resource consumption function. The objective function is the total cost. It belongs to the symbol.

[0163] The uncertainty of weather forecasts, the potential failure of equipment, and other risk factors are defined as a "set of uncertainties". In the solution process, it is ensured that the final solution has good performance (i.e., robustness) for all possible situations in this set.

[0164] The output function employs an improved AquaCrop model:

[0165] ;

[0166] in It is a production function; For decision variable vectors; For the uncertain parameter vector; For the maximum potential output; The symbol for multiplication; This represents the total number of growth stages. For the first Yield response coefficients for each growth stage; For the first Actual evapotranspiration at each growth stage; For the first Maximum evapotranspiration at each growth stage.

[0167] Uncertainty set It encompasses two main sources of uncertainty: weather forecast uncertainty and equipment performance uncertainty. For weather forecast uncertainty, the system sets a fluctuation amplitude coefficient. This is used to characterize the reliability and stability of weather forecast values. Similarly, the system sets a maximum deviation range to account for uncertainties in equipment performance. This ellipsoidal definition of uncertainty sets effectively characterizes various uncertainties encountered during system operation, providing a theoretical basis for subsequent robust optimization.

[0168] The constraints are enhanced to probabilistic constraints, reflecting robustness requirements. Specifically, these include the following aspects: Water supply adequacy constraint: The probability that the water supply is greater than or equal to the water demand is no less than 95% to ensure the reliability of irrigation water; Crop water stress control constraint: The probability that the crop water stress index does not exceed 0.6 is no less than 90% to avoid severe water stress on crops; Equipment availability constraint: The probability that irrigation equipment is available is no less than 95% to ensure stable system operation; Cost control constraint: The expected cost of system operation does not exceed 110% of the budget, controlling costs while ensuring system performance.

[0169] The introduction of these probabilistic constraints improves the robustness of the system, enabling it to better cope with various uncertainties.

[0170] These constraints guarantee that: the probability of water supply meeting demand is not less than 95%; the probability of water stress index being controlled below 0.6 is not less than 90%; equipment availability is not less than 95%; and expected cost does not exceed 110% of the budget.

[0171] The constraints have been strengthened from the traditional "must be met" to "high probability of being met," for example, requiring that "the probability of water supply meeting demand is not less than 95%." This makes the decision-making scheme more realistic and reliable. The system uses an improved NSGAIII genetic algorithm to solve this complex multi-objective problem and can adjust the optimization direction according to the user's risk preferences (e.g., whether they prefer high output or low cost). The fitness function uses a weighted summation method, comprehensively considering the three optimization objectives and robustness index, and balancing the importance of each objective through weight coefficients.

[0172] The optimal solution set obtained by the improved NSGAIII genetic algorithm constitutes a robust optimized irrigation and fertilization scheme. The data is stored in a structured format and includes operational parameters such as daily irrigation water volume, fertilizer type and quantity, and specific execution time for the next week. The system selects the final solution from the optimal solution set based on the user's risk preference weight (conservative, balanced, or aggressive) and converts it into an executable sequence of control instructions.

[0173] The robustness index is defined as a measure of the system's stability under uncertain conditions. Specifically, it is measured by minimizing the sum of squares of constraint violations across all possible uncertain scenarios. The degree of constraint violation represents the extent to which the system does not satisfy a particular constraint. The smaller the index, the better the system's robustness. Based on the robustness index, the system further calculates the decision confidence level. Decision confidence It is a comprehensive score, ranging from 0 to 1, providing decision-makers with a quantitative assessment of the reliability of a solution. It is derived through a weighted average of three key factors:

[0174] Constraint satisfaction probability is the probability that key constraints will be satisfied in an uncertain scenario.

[0175] Objective function stability measures the degree of fluctuation in the value of an objective function under conditions of uncertainty.

[0176] Model prediction error is a statistical measure of the deviation between historical model predictions and actual results.

[0177] In addition, the system is designed with tiered emergency response plans. Based on the level of risk index, the system will automatically trigger different levels of response measures, ranging from simple adjustments to irrigation parameters to activating backup water sources, and then to the highest level of emergency irrigation and manual intervention, ensuring that losses can be minimized in various emergencies.

[0178] The emergency response plan employs a tiered triggering mechanism, automatically activating the corresponding level of the plan based on the risk index:

[0179] Level 1 emergency plan ( ):

[0180] Automatic adjustment of irrigation parameters: The planned irrigation water volume is adjusted according to the risk index. When the risk index increases by 0.1, the irrigation water volume increases by 2% to cope with potential risks.

[0181] Increase monitoring frequency to once per hour;

[0182] Activate the early warning information push system.

[0183] Level II emergency response plan ( ):

[0184] Activate backup water sources and equipment, with a backup capacity of 20% of normal demand;

[0185] Send alert messages to administrators and relevant technical personnel;

[0186] Adjusting irrigation strategies to an active mode, increasing irrigation frequency and volume to address potential water stress risks, and increasing the frequency of mobile sensor inspections.

[0187] Level III emergency plan ( ):

[0188] Initiate emergency whole-field irrigation procedures;

[0189] Manual personnel take over system control, and automated systems switch to auxiliary mode;

[0190] Initiate agricultural insurance claims process;

[0191] Call in the emergency technical support team.

[0192] The above three-level contingency plans, along with their triggering conditions, response measures, and execution procedures, together constitute the system's emergency response plan set. The emergency response plan set is stored in a decision tree structure, which can automatically select the appropriate level of plan based on the real-time risk index and smoothly transition when the risk status changes. Emergency Response Plan Set Each contingency plan includes detailed execution steps, resource requirements, and division of responsibilities to ensure a rapid and effective response in emergency situations.

[0193] Before performing multi-objective optimization, the objective function and constraints need to be standardized:

[0194] Objective function normalization: The three objective functions of maximizing output, minimizing water consumption, and minimizing total cost are normalized according to their historical best values ​​to ensure that each objective is optimized on the same order of magnitude.

[0195] Constraint standardization: Capacity normalization is performed on resource constraints such as water supply capacity, fertilizer inventory, and equipment power.

[0196] Uncertainty parameters: Probability distribution fitting and standardization of uncertainty parameters such as weather forecast errors and equipment failure rates;

[0197] Risk preference coefficient: Converts users' risk preferences (conservative, balanced, aggressive) into numerical weights and normalizes them.

[0198] Based on the above processing steps, a robust optimized irrigation and fertilization scheme is generated. This constitutes a set of emergency response plans ( This involves classifying contingency plans into levels based on risk index thresholds, designing response measures, and organizing them into structured documents; additionally, decision confidence levels... By evaluating the robustness index of the optimization results, calculating the probability of constraint satisfaction, and comprehensively considering the model prediction error, a weighted average confidence score is finally obtained.

[0199] This step enables the development of robust water and fertilizer allocation schemes that balance yield, cost, and resource consumption in uncertain environments.

[0200] Step 6: Input the optimal water and fertilizer allocation strategy into the adaptive PID controller to realize multi-source linkage closed-loop feedback control, and continuously optimize the system parameters through an online learning mechanism to achieve adaptive evolution;

[0201] Step 6.1, Multi-source linkage and closed-loop feedback control;

[0202] The specific inputs for this step include: the robust optimized irrigation and fertilization scheme generated in step 5. Emergency Response Plan (Includes the triggering conditions and execution strategies for the Level 3 emergency plan), field zoning information. (Field zoning data obtained through a geographic information system, including attributes such as area, soil type, and topographic features of each zone), Equipment status (Real-time monitoring of equipment operating parameters via sensor networks, including indicators such as availability, health, and response time of each irrigation device)

[0203] The processing procedure specifically includes:

[0204] This step establishes a distributed collaborative control architecture to enable coordinated execution of irrigation equipment in multiple regions.

[0205] First, an adaptive proportional-integral-derivative (PID) controller is constructed. The PID controller is a classic control algorithm widely used in industrial control, achieving precise regulation of the controlled object through a combination of proportional, integral, and derivative components. This invention improves the traditional PID controller, enabling its parameters to be dynamically adjusted according to changes in soil moisture.

[0206] Each control partition The formula for calculating the control output is:

[0207] ;

[0208] in For the first Each partition at time Control output; For the first Each partition at time Proportional control gain; For the first Each partition at time Integral control gain; For the first Each partition at time The differential control gain; For the first Each partition at time Control error; For integration variables; From 0 to The definite integral of .

[0209] The adaptive adjustment of PID parameters adopts an online optimization strategy based on gradient descent:

[0210] ;

[0211] ;

[0212] ;

[0213] in For the first Each partition at time Proportional control gain; For the first Each partition at time Proportional control gain; The learning rate is the proportional term. For the first Each partition at time Integral control gain; For the first Each partition at time Integral control gain; The learning rate is the integral term. For the first Each partition at time The differential control gain; The differential gain at the current moment; The learning rate is the differential term. The performance index function is calculated using the following formula:

[0214] ;

[0215] in For performance metrics functions; From 0 to The definite integral; To control the cycle; To control the precision weight; Energy consumption weighting; For stability weights; For the first Each partition at time Control error; For the first Each partition at time Control output; For the first Each partition at time The amount of change in the control output.

[0216] After optimizing the PID control parameters for a single zone, considering the interrelationships between zones in the farmland irrigation system, such as water resource sharing and equipment linkage, it is necessary to further address the inter-zone coordination control problem. Therefore, this invention proposes the following inter-zone coordination control optimization problem:

[0217] ;

[0218] in Minimize operator; To optimize variables, and The control outputs for the first and last partitions are respectively. Total number of partitions; The summation symbol; For the first Performance metrics for each partition; Weights for the coordination items; For neighbor partition sets; For the first The first partition and the first Coordination weights between partitions; and The first The and the first Control outputs for each partition.

[0219] In actual operation, due to abnormal situations such as equipment failure and communication interruption, it is necessary to adjust the control strategy in a timely manner to ensure the reliable operation of the system. To address this issue, this invention proposes a fault-tolerant task reallocation algorithm based on cost minimization:

[0220] ;

[0221] ;

[0222] in This is the result of the reallocation; The parameters that minimize the objective function; For the allocation scheme; The summation symbol; It is the cost function; For the penalty function; To sum over available devices; To sum the results for the faulty equipment; This represents the original task quantity; Mark the constraints; This represents the total number of partitions.

[0223] After completing the above control algorithm design and optimization, the system needs to transform the theoretical model into practically executable operation commands and establish a feedback evaluation mechanism to form a complete closed-loop control system. This process includes three key steps: command generation, feedback data acquisition, and performance evaluation. First, based on the PID controller output... Based on the optimization results of inter-regional coordination and control, the system generates partition execution instructions. Partition execution command It is a structured dataset containing specific operational parameters for each irrigation zone, such as start-up time, runtime, flow rate settings, and fertilizer ratio. These instructions are organized in JSON format for easy identification and execution by various irrigation equipment, ensuring that theoretical control strategies are accurately translated into equipment actions. Secondly, to ensure the effective implementation of control strategies and timely response to abnormal situations, the system establishes a real-time execution feedback mechanism. Execution feedback mechanism A sensor network distributed throughout the farmland periodically (every 1-5 minutes) collects operational status data from irrigation equipment, including key indicators such as actual flow rate, pressure, on / off status, and execution delay. This feedback data is stored in a time-series structure, with each record containing a timestamp, equipment ID, parameter type, and parameter value. This provides the control system with necessary observational information, enabling the control loop to form a closed loop. Finally, to objectively evaluate the control effect and provide a basis for subsequent optimization, the system constructs a control effect evaluation system. This assessment quantifies control performance across three dimensions: response timeliness, control accuracy, and system stability, by comparing the ratio of the root mean square error between the actual control value and the target value to the predetermined baseline error. These dimensional indicators are weighted and averaged into a comprehensive score ranging from 0 to 1, intuitively reflecting the overall operational quality of the control system. Simultaneously, the scores for each dimension provide direction for targeted optimization. Through these three stages, the system achieves end-to-end management from theoretical control algorithms to actual execution results, ensuring that water and fertilizer allocation strategies can be efficiently and accurately implemented in farmland practices.

[0224] Based on the above processing, output parameter partitioning execution instructions are generated. Implementation feedback mechanism and control effect evaluation .

[0225] Step 6.2, Online Learning and System Evolution Optimization;

[0226] The specific inputs for this step include: execution feedback mechanism. (Real-time execution status data generated in step 6.1), actual output data (Wheat yield data for each region obtained through harvest period measurements, in kilograms / hectare), environmental change data. (Includes long-term monitoring data such as climate change trends, changes in soil fertility, and the occurrence of pests and diseases), user feedback (Farmers' subjective evaluations and suggestions for improvement of system performance, using a 1-5 point rating system), control effect evaluation (Control effect evaluation data generated from step 6.1).

[0227] The processing procedure specifically includes:

[0228] This step establishes an online incremental learning mechanism to enable continuous optimization and adaptive evolution of the system.

[0229] Online incremental learning is an important branch of machine learning, allowing models to be continuously updated as they receive new data without retraining. This invention employs an improved stochastic gradient descent algorithm to achieve online updates of model parameters:

[0230] ;

[0231] For a moment The parameter vector; For a moment The parameter vector; The learning rate; For parameters Find the gradient; The loss function; For a moment The true value; For parameters The model function; For a moment Input; Historical information weighting coefficients; The summation symbol; For the first Weights of historical data; For the first The true value of historical data; For the first Input of historical data;

[0232] The model parameters are continuously updated through the above online incremental learning, ultimately resulting in the updated parameter set. It contains knowledge and patterns extracted from historical data.

[0233] Learning rate The learning rate is dynamically adjusted over time, specifically calculated by dividing the initial learning rate by the square root of the time step plus 1. This method allows the learning rate to gradually decrease during training, which helps the model converge.

[0234] loss function The mean squared error (MSE) method is used: the square of the difference between the model's predicted value and the true value is multiplied by 1 / 2. This loss function can effectively measure the accuracy of the model's predictions, and the coefficient 1 / 2 is used to eliminate the exponent when differentiating.

[0235] Historical data weights employ an exponential decay mechanism to ensure that new data has a higher weight:

[0236] ;

[0237] in For the first The weight of each data point; is the base of the natural logarithm; Let be the time decay coefficient, and ; This is the current time step; For the first The importance weight of each data point is determined based on the degree to which the data deviates from the historical mean, and the absolute value of the standardized deviation is used to measure the importance of the data.

[0238] The model performance evaluation adopts a multi-dimensional indicator system:

[0239] Water use efficiency (WUE) is calculated by dividing actual yield by total water consumption. Actual yield is expressed as yield per hectare (kg / ha), and total water consumption is expressed in millimeters. This indicator can be used to assess the efficiency of crop water use; a higher value indicates greater crop yield per unit of water consumption.

[0240] Prediction accuracy is measured using Mean Absolute Percentage Error (MAPE):

[0241] ;

[0242] in Mean absolute percentage error; The number of samples; The summation symbol; For the first Predicted yield for each sample; For the first The actual output of each sample.

[0243] System reliability is comprehensively evaluated by considering both runtime and task success rate, using the product of runtime ratio and task success rate. Based on these multi-dimensional evaluation metrics, a structured system performance report is generated. It includes time-series data and visualizations of key indicators such as water use efficiency, forecast accuracy, system reliability, and return on investment, providing decision-makers with a comprehensive view of the system's operational status.

[0244] Economic benefits are evaluated using the return on investment (ROI) metric. ROI is calculated by subtracting the total system cost from the increased revenue generated by the system, and then dividing by the total system cost. The increased revenue includes economic benefits such as increased output, improved quality, and water and fertilizer savings; the total system cost includes expenditures on hardware, software development, installation, commissioning, operation, and maintenance. Calculating the ROI percentage provides a direct assessment of the system's economic feasibility.

[0245] During continuous optimization, the system faces the problem of selecting various models and strategies. To strike a balance between exploring new strategies and utilizing known effective strategies, this invention introduces the Multi-Armed Bandit algorithm as the core mechanism for model selection, making the optimal choice under uncertain conditions. Specifically, it employs the Upper Confidence Bound (UCB) algorithm:

[0246] ;

[0247] in The upper confidence bound algorithm; This is the current time step; For the model At any moment Average performance; For the model The number of times it was selected; It is the natural logarithm; This is the square root operator.

[0248] Building upon short-term optimization, the system also needs to consider maximizing long-term benefits. The seasonal nature of agricultural production requires control strategies to adapt to long-term environmental changes. Therefore, the system employs the Q-learning method within the reinforcement learning framework to construct a long-term optimal strategy. Unlike multi-armed slot machines that focus on short-term choices, Q-learning learns the long-term optimal decision sequence through the state-action value function, thus forming the optimal control strategy. Its update formula is:

[0249] ;

[0250] in The state-action value function; This is the current state; For the current action; The learning rate; For instant rewards; Discount factor; For all possible actions Take the maximum value; The next state; For the next action.

[0251] In Q-learning, states and actions are specifically represented by state space and action space, where state space... Specific definition:

[0252] Environmental conditions: standardized values ​​of temperature, humidity, precipitation, and light intensity;

[0253] Crop status: growth stage, leaf area index, and standardized value of biomass;

[0254] System status: Standardized values ​​for device health, data quality, and control accuracy;

[0255] Action space include:

[0256] Model parameter adjustment: PID controller parameters, fusion weights, threshold settings;

[0257] Strategy switching: conservative strategy, aggressive strategy, balanced strategy;

[0258] Sampling frequency adjustment: sensor sampling interval, data update frequency;

[0259] Based on the above processing steps, the system continuously optimizes the model parameters through an online incremental learning algorithm. Simultaneously, a system performance report is generated by combining multiple dimensions of indicators such as water use efficiency, forecast accuracy, system reliability, and return on investment. And the optimal control strategy is formed under different environmental conditions using the Q-learning framework. These output parameters together constitute the system's adaptive optimization capability.

[0260] A smart agricultural crop demand-driven intelligent water and fertilizer allocation system is used to execute the aforementioned smart agricultural crop demand-driven multi-source linkage method, including:

[0261] The data acquisition module is used to acquire multi-source heterogeneous data through an adaptive sparse sensor network;

[0262] The data fusion module is used to perform high-precision fusion and uncertainty quantification of the multi-source heterogeneous data based on a Bayesian framework and data credibility weights.

[0263] The stress identification module is used to accurately identify the stress status of different crop varieties by using a variety-adaptive crop water stress index model combined with health status correction.

[0264] The demand forecasting module is used to generate crop water and fertilizer demand forecasting schemes that take risks into account, based on the value at risk model and combined with stress conditions and weather forecasts.

[0265] The optimization decision module is used to perform multi-objective robust optimization using the NSGAIII genetic algorithm to generate the optimal water and fertilizer allocation strategy.

[0266] The linkage control and learning module is used to execute closed-loop feedback control according to the optimal strategy through an adaptive PID controller, and continuously optimize the model parameters and control strategy in the system through an online incremental learning mechanism based on the actual execution effect.

[0267] Here, the present invention provides an application example:

[0268] The greenhouses in Area A utilize high-standard, intelligent facilities, covering a total planting area of ​​5 hectares. The main variety planted is the winter wheat, Jimai 22. This variety exhibits strong cold and disease resistance and is currently in the jointing stage (120 days after sowing), a crucial stage for wheat growth and development with high water and fertilizer requirements. The greenhouses are equipped with an intelligent environmental control system, precisely controlling the temperature within the range of 15-25°C (maintaining a diurnal temperature range of 10°C) and relative humidity within a suitable range of 60%-80%. An LED supplemental lighting system ensures stable light intensity of 400-600 μmol / (m²). 2 The greenhouse utilizes a CO2 generator to maintain the CO2 concentration at an ideal level of 350-400 ppm. It is also equipped with an intelligent integrated water and fertilizer system that adjusts irrigation and fertilization plans in real time according to crop growth needs, ensuring optimal environmental conditions for crop growth.

[0269] Area B's open-field farmland is located in a typical northern agricultural region, with a total cultivated area of ​​50 hectares. The main crop is the adaptable spring wheat variety Longmai 26. Currently, the crop is in the grain-filling stage (150 days after sowing), a stage where water supply requirements are particularly stringent and directly impact the final yield. The region's climate is characterized by moderate temperatures, with daily average temperatures fluctuating between 18-28°C and monthly average precipitation around 50 mm, classifying it as a semi-arid climate. The farmland is consistently affected by moderate winds of 2-4 m / s, which promotes crop transpiration. The soil is predominantly loam, with a pH value between 6.5 and 7.2, good soil structure, and strong water and fertilizer retention capacity. The farmland is equipped with an intelligent irrigation system, including a soil moisture monitoring network, a meteorological monitoring station, and intelligent sprinkler irrigation equipment, enabling precision irrigation and integrated water and fertilizer management. Simultaneously, satellite remote sensing and drone patrols are used to monitor and assess crop growth in real time, providing data support for water and fertilizer management decisions.

[0270] To comprehensively monitor the growth environment and physiological state of crops, this system deploys multiple high-precision sensors. The following table shows the real-time monitoring data and reliability indicators of each sensor. The real-time monitoring data of the multi-source sensors are shown in the table below:

[0271] Table 1: Real-time Monitoring Data from Multi-Source Sensors

[0272]

[0273] The data shows that the data quality of each sensor is generally high, with confidence levels all above 0.88, providing a reliable data foundation for the system's accurate decision-making.

[0274] To verify the technical advantages of this system, we conducted a comprehensive comparison with traditional agricultural water and fertilizer management methods. The system performance comparison and verification results are shown in the table below:

[0275] Table 2: System Performance Comparison and Verification Table

[0276]

[0277] The comparison results show that the system has achieved improvements in all key indicators, especially in system response time and resource utilization efficiency.

[0278] To assess the system's adaptability under different climatic conditions, we conducted multi-year follow-up verification. The results of the environmental adaptability verification are shown in the table below:

[0279] Table 3: Environmental Adaptability Verification Table

[0280]

[0281] Data shows that the system maintains high stability and reliability under various climatic conditions, and exhibits stronger resource-saving effects, especially in drought years.

[0282] To understand the system's ability to continuously improve, we conducted a three-year follow-up evaluation. The statistical results of the long-term operating performance are shown in the table below:

[0283] Table 4: Statistical Table of Long-Term Operational Results

[0284]

[0285] Long-term operational data shows that through continuous learning and optimization, the system's various performance indicators have shown a steady upward trend, demonstrating good development potential.

[0286] Variety-adaptive water stress identification: A variety-specific CWSI calculation method is proposed for the first time, with an identification accuracy of over 85%.

[0287] Multi-source data fusion technology: Innovatively combining Bayesian data fusion with variational kriging interpolation, improving fusion accuracy by 30%.

[0288] Robust multi-objective optimization: An optimization framework considering uncertainties was established, reducing yield losses by 15%–20% under extreme weather conditions.

[0289] Closed-loop control and online learning: This enables adaptive control with multi-source linkage, reducing system response time to less than 5 minutes.

[0290] Application results in greenhouses in region A: After 3 months of application, the system achieved a 28% increase in water use efficiency, a 15% increase in crop yield, and maintained a 97.5% system stability.

[0291] Application results in open-field fields in Region B: Over a full growing season, wheat yield increased by 12%, water savings were achieved by 25%, and the system failure rate was only 3% (lower than the industry average of 8%).

[0292] Economic benefits: After the system was implemented, the average income per hectare increased by 440 yuan, achieving a return on investment of 32.1%, while operating costs were reduced by 21.6%.

[0293] Social benefits: By increasing water resource utilization efficiency by 26.5% and fertilizer utilization efficiency by 38.8%, it effectively promotes sustainable agricultural development.

[0294] Environmental benefits: While reducing agricultural non-point source pollution, it improves soil health and reduces carbon emission intensity.

[0295] This intelligent water and fertilizer allocation system, driven by crop demand, achieves precise, efficient, and intelligent agricultural water and fertilizer management through key technologies such as multi-source data fusion, variety adaptive identification, robust optimization decision-making, and closed-loop control. The system demonstrates good adaptability and stability in various application scenarios, providing crucial technical support for the digital transformation of modern agriculture.

[0296] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multi-source linkage method driven by crop demand in smart agriculture, characterized in that, The method comprises the following steps: A layout optimization model and a dynamic programming algorithm are used to construct an adaptive sparse sensor network and plan a mobile inspection path; Based on a Bayesian framework and by introducing a data credibility weight mechanism, multi-source heterogeneous data collected by the adaptive sparse sensor network is fused, and a variational Kriging interpolation method is used to realize the unification of spatial and temporal resolutions; A variety-specific crop water stress index model is constructed using the unified data, and a health state correction factor and a multi-state classifier are used to realize variety-adaptive crop stress state recognition; The data credibility weight mechanism is expressed as: ； wherein is the fused parameter posterior distribution, , , denote the observation data of the 1st, 2nd, 3rd data source, respectively; is the likelihood function of the 1st data source; is the parameter prior distribution; is the data reliability weight of the 1st data source; is proportional to the sign; is the multiplication symbol; is the number of data sources; The variety-specific crop water stress index is specifically expressed as: ； wherein is a variety-specific water stress index; is a leaf temperature; is an air temperature; is a variety-specific lower baseline; is a variety-specific upper baseline; is an effective accumulated temperature; is an atmospheric vapor pressure deficit; is a health status correction factor; Introducing breed-specific lower baseline and breed-specific upper baseline , dynamically adjusting the stress determination criteria by considering the physiological characteristics of different breeds; at the same time, adding a health state correction factor , correcting non-water stress factors; According to the recognized stress state probability, a crop water and fertilizer demand prediction scheme considering yield loss risk is generated in combination with a risk value model and future weather; specifically, a fuzzy water stress coefficient is calculated based on the stress state probability, a variety-specific crop coefficient and an uncertainty coefficient are introduced, a water and fertilizer demand interval is predicted based on the fuzzy water stress coefficient, the variety-specific crop coefficient and the uncertainty coefficient, and a crop water and fertilizer demand prediction scheme considering yield loss risk is generated based on the water and fertilizer demand interval, the risk value model and future weather; A multi-objective robust optimization is performed on the crop water and fertilizer demand prediction scheme to generate an optimal water and fertilizer allocation strategy; The optimal water and fertilizer allocation strategy is input into an adaptive PID controller to realize multi-source linkage closed-loop feedback control, and an online learning mechanism is used to continuously optimize system parameters to realize adaptive evolution. 2.The smart agriculture crop demand-driven multi-source linkage method according to claim 1, characterized in that, The objective function of the sensor layout optimization model of the adaptive sparse sensor network is to minimize the weighted sum of deployment cost, monitoring error and maintenance difficulty, wherein the cost weight is 40%, the monitoring error weight is 40%, and the maintenance difficulty weight is 20%, and the constraints are that the monitoring coverage rate is not less than 90% and the redundancy is not less than 2. 3.The smart agriculture crop demand-driven multi-source linkage method according to claim 1, characterized in that, The constraint conditions of the multi-objective robust optimization include that the hydraulic uniformity coefficient of the irrigation system is not less than 0.85, the total amount of single fertilization is not more than 110% of the maximum absorption threshold of the crop, and the total operating cost is within the preset budget range. 4.The smart agriculture crop demand-driven multi-source linkage method according to claim 1, characterized in that, In the multi-source linkage closed-loop feedback control, an adaptive proportional, integral and differential controller is constructed to realize accurate water and fertilizer regulation of each control partition, and the control output calculation formula is: ； wherein is the control output of the first zone at time ; is the proportional control gain of the first zone at time ; is the integral control gain of the first zone at time ; is the derivative control gain of the first zone at time ; is the control error of the first zone at time ; is the integral variable; is the definite integral from 0 to . 5.The smart agriculture crop demand-driven multi-source linkage method according to claim 1, characterized in that, The adaptive evolution by continuously optimizing system parameters through an online learning mechanism includes: An online incremental learning mechanism is established, an improved stochastic gradient descent algorithm is used to realize online updating of model parameters, and the model is allowed to continuously update without retraining when new data is obtained; An exponential decay mechanism is used to ensure that new data has a higher weight, and the weight of historical data decays over time; A multi-dimensional index system is established to evaluate the performance of the model, including water use efficiency, prediction accuracy, system reliability and economic benefit; A multi-armed bandit algorithm is used for model selection, and a confidence upper bound algorithm is used to make the optimal selection in an uncertain environment; A long-term optimization strategy based on a Q-learning reinforcement learning framework is established, the optimal strategy is realized by learning the state-action value function, and the continuous optimization and adaptive evolution of the system are realized. 6.The smart agriculture crop demand-driven multi-source linkage method according to claim 1, characterized in that, The stress state is divided by using a Softmax classifier, and after division, there are four states: When , it is a non-stress state; when , it is a light stress state; when , it is a moderate stress state; and when , it is a severe stress state. 7.The smart agriculture crop demand-driven multi-source linkage method according to claim 1, characterized in that, The analysis of the stress state uses a hidden Markov model to model the time series of crop stress states, and predict the evolution trend of the stress state in the next 24 hours.

8. The smart agricultural crop demand-driven water and fertilizer intelligent allocation system, characterized in that, The intelligent agriculture crop demand-driven multi-source linkage method comprises the following steps: A data acquisition module is configured to acquire multi-source heterogeneous data through an adaptive sparse sensor network. A data fusion module is configured to perform high-precision fusion and uncertainty quantification on the multi-source heterogeneous data based on a Bayesian framework and data credibility weight. A stress identification module is configured to accurately identify the stress state of different crop varieties through a crop water stress index model adaptive to varieties and combined with health state correction. A demand prediction module is configured to generate a crop water and fertilizer demand prediction scheme considering risks based on a risk value model, combined with stress state and weather forecast. An optimization decision module is configured to perform multi-objective robust optimization by using an NSGAIII genetic algorithm to generate an optimal water and fertilizer allocation strategy. A linkage control and learning module is configured to perform closed-loop feedback control according to the optimal strategy through an adaptive PID controller, and continuously optimize the model parameters and control strategy in the system through an online incremental learning mechanism according to the actual execution effect.

Images