Intelligent water and fertilizer management method and system for crops

Abstract

The invention discloses an intelligent water and fertilizer management method and system for crops, and relates to the technical field of agricultural intellectualization, and the system comprises a data acquisition module, a data processing module, a decision module, an execution module and a feedback module. The data acquisition module acquires environment and crop state data through a soil moisture content sensor, a stalk diameter sensor, a multispectral camera and a meteorological interface; the data processing module fuses the multi-source data based on the dynamic weight correction coefficient to generate an environment-crop state vector; the decision-making module predicts water and fertilizer demands in combination with the dynamic growth model, and generates a water and fertilizer adjustment decision-making instruction in a set iteration period through a deep reinforcement learning engine; an execution module adjusts the water and fertilizer ratio through an electromagnetic proportional valve, a mobile irrigation vehicle executes spraying according to path planning, and a near infrared spectrum sensor corrects deviation in real time; and the feedback module transmits the adjusted state data back to the data processing module to form a dynamic closed-loop management process of acquisition-fusion-decision-execution-feedback.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural intelligence, and in particular to a crop intelligent water and fertilizer management method and system. BACKGROUND

[0002] At present, agricultural water and fertilizer management is gradually developing towards intelligence and precision, and water and fertilizer management methods and systems based on sensor data collection, algorithm model decision and automatic execution have appeared in the prior art. For example, some schemes collect environmental parameters through soil moisture sensor, weather station and other devices, calculate water and fertilizer demand by combining crop growth model, and adjust through electromagnetic valve and other execution elements; at the same time, some systems introduce edge computing, machine learning and other technologies to improve data processing and decision efficiency. These technologies reduce manual intervention to a certain extent and improve the automation level of water and fertilizer management.

[0003] However, the prior art still has many limitations: firstly, the data collection, processing, decision, execution and feedback links are independently operated, lacking a full-link dynamic cooperation mechanism, resulting in data flow transfer hysteresis, difficulty in real-time response to crop growth stage changes and environmental fluctuations, and the overall management mode showing static or semi-static characteristics; secondly, data fusion processing mostly uses fixed weight or simple weighting method, without dynamic correction of sensor reliability, and over-reliance on experience parameters or single algorithm model (such as traditional PID control or basic machine learning model) for decision generation, without deep coupling with crop dynamic growth law (such as growth period characteristics, leaf area index change) and multi-objective optimization goal (such as yield gain, energy consumption cost balance), resulting in deviation between water and fertilizer demand prediction and actual physiological demand of crops; thirdly, the control algorithm of the execution adjustment link is mostly single PID control or fuzzy control, lacking a composite mechanism of precise control and robustness regulation, and lacking linkage with crop growth state feedback (such as drought stress, growth abnormalities), which is easily affected by environmental interference and has weak anti-interference ability; fourthly, the functions of the software and hardware modules (such as data collection module, decision module, execution module, etc.) of the management system are scattered, data interaction relies on traditional communication mode, and there is a lack of collaborative design based on unified architecture, resulting in low system integration, poor response efficiency, and difficulty in supporting efficient operation of full-link closed-loop management. SUMMARY

[0004] In view of the above-mentioned shortcomings of the prior art, the purpose of the present application is to provide a crop intelligent water and fertilizer management method and system to solve one or more problems in the prior art.

[0005] To achieve the above-mentioned purpose, the first technical solution of the present application is as follows:

[0006] A crop intelligent water and fertilizer management method, comprising the following steps:

[0007] Data acquisition step: data is collected by soil moisture sensor, stem diameter sensor, multispectral camera and weather data interface, the sampling frequency of the soil moisture sensor is 1-10 times / min, and the measurement range is 0-100vol%. The resolution of the stem diameter sensor is 0.001-0.01mm, and the sampling frequency is 1-5 times / min. The wave band range of the multispectral camera is 400-1000nm, and the daily collection times is 1-3 times.

[0008] Data fusion processing step: based on the edge computing node, the collected data is fused and processed, the fusion process adopts data credibility weighted fusion formula and sensor weight dynamic correction formula, the data credibility weighted fusion formula is:

[0009] ,

[0010] The sensor weight dynamic correction formula is:

[0011] ,

[0012] Wherein, is the fused parameter of the ith grid, is the original detection value of the kth sensor in the ith grid, is the sensor weight, n is the number of sensor types, is the historical detection variance of the kth sensor, is the data collection time interval, and α is the time attenuation coefficient. The data fusion delay of the edge computing node is ≤500ms.

[0013] Demand calculation and decision generation step: based on the dynamic growth model, the nitrogen demand is calculated, and the water and fertilizer adjustment decision is generated through the deep reinforcement learning engine, the nitrogen demand calculation formula of the dynamic growth model is:

[0014] ,

[0015] The reward function formula of the deep reinforcement learning engine is:

[0016] ,

[0017] Wherein, is the basic nitrogen demand, LAI(t) is the leaf area index, β is the LAI influence coefficient, γ is the growth period attenuation coefficient, T is the total growth period, ΔY is the yield gain, and η is the water and fertilizer utilization rate, The energy consumption cost per unit area is the cost of the energy consumption per unit area. The decision response time of the deep reinforcement learning engine is less than or equal to 2 seconds, and the mass percentage ranges of nitrogen, phosphorus and potassium in the output nutrient element ratio are 0-30%, 0-20% and 0-50% respectively, and the sum of each nutrient element ratio is less than or equal to 100%.

[0018] The water and fertilizer adjustment execution step: based on the decision result, the water and fertilizer ratio is adjusted through the electromagnetic proportional valve, and the PID + fuzzy compound control algorithm is adopted in the adjustment process, and the control law formula is:

[0019] ,

[0020] Wherein, u(t) is the control amount of the electromagnetic proportional valve, e(t) is the concentration deviation, is a proportional coefficient, is an integral coefficient, is a differential coefficient, is a fuzzy control compensation amount. The adjustment accuracy of the electromagnetic proportional valve is ± 0.1% to ± 0.5%, and the response time is less than or equal to 100 ms.

[0021] The feedback and early warning step: based on the digital twin platform, the growth state of crops is simulated, and early warning is performed through the drought stress early warning module, and the drought stress index calculation formula is:

[0022] ,

[0023] Wherein, is the current stem diameter, is the stem diameter of the same period predicted by the digital twin model. The early warning threshold is that the drought stress index is 0.5-0.7, and the simulation error of the digital twin platform is less than or equal to 5%.

[0024] Specifically, in the data acquisition step, the sampling frequency of the soil moisture sensor is 1-10 times per minute, the measurement range is 0-100vol%, and the buried depth is 10-40cm. The resolution of the stem diameter sensor is 0.001mm, the sampling frequency is 1 / 2min, and the range is 0-20mm.

[0025] Specifically, in the data fusion processing step, the value range of a in the sensor weight dynamic correction formula is 0.05-0.2, The value range of a is 0.1-2.0.

[0026] Specifically, in the demand calculation and decision generation step, the value of beta in the dynamic growth model is 0.2-0.4, the value of gamma is 0.7-0.9, and T is 150-200d. The leaf area index LAI(t) is obtained by unmanned aerial vehicle NDVI inversion, and the inversion formula is

[0027] .

[0028] Specifically, in the water and fertilizer adjustment execution step, the PID+ fuzzy compound control algorithm has a value range of 2.0-3.0, a value range of 1.0-1.5, a value range of 0.3-0.7. The adjustment accuracy of the electromagnetic proportional valve is ±0.1%.

[0029] Specifically, in the feedback and early warning step, when the drought stress index is greater than 0.6, a third-level early warning is automatically pushed, and the response time is less than or equal to 5 minutes.

[0030] In order to make the technical effect complete, the second technical solution of the present application is a management system based on the intelligent water and fertilizer management method for crops, comprising:

[0031] A data acquisition module comprising a soil moisture sensor, a stem diameter sensor, a multispectral camera and a weather data interface, for executing the data acquisition step.

[0032] A data processing module comprising an edge computing node and a federated learning framework, for executing the data fusion processing step.

[0033] A decision module comprising a dynamic growth model and a deep reinforcement learning engine, for executing the demand calculation and decision generation step.

[0034] An execution module comprising an electromagnetic proportional valve, a near-infrared spectrum sensor and a mobile irrigation vehicle, for executing the water and fertilizer adjustment execution step in the management method.

[0035] A feedback module comprising a digital twin platform and a drought stress early warning module, for executing the feedback and early warning step.

[0036] A storage module, which is a non-transitory computer readable storage medium, stores a computer program, and the computer program is executed by a processor to realize the intelligent water and fertilizer management method.

[0037] Specifically, the computer program stored in the storage module comprises a data fusion algorithm module, a dynamic growth model module, a deep reinforcement learning module and a PID+ fuzzy control module, and the modules are respectively executed by the processor when called.

[0038] Specifically, the wavelength range of the near-infrared spectrum sensor in the execution module is 600-1200nm, and the detection accuracy is ±0.01-±0.1mS / cm. The path planning of the mobile irrigation vehicle adopts an improved A* algorithm, and the path cost formula is:

[0039] ,

[0040] wherein D(i,j) is the straight-line distance, theta(i,j) is the slope angle, W is the crop density weight, and rho(i,j) is the crop density.

[0041] Specifically, the edge computing node computing power of the data processing module is 10-50TOPS, the local model update frequency of the federated learning framework is 1-24h / time, and the model aggregation delay is less than or equal to 30min.

[0042] Compared with the prior art, the beneficial technical effects of the present application are as follows:

[0043] (I) Through multi-source sensor data acquisition, dynamic weight correction data fusion processing, decision generation combining dynamic growth model and deep reinforcement learning, PID+ fuzzy compound control execution adjustment, and feedback warning of the digital twin platform, a full-link closed-loop management mechanism of "perception-fusion-decision-execution-feedback" is formed. The synergy of this multi-link technical feature breaks through the limitations of single-link optimization in the prior art, realizes dynamic self-adaptive adjustment from data acquisition to growth state feedback, and makes water and fertilizer management able to respond to crop growth stage changes and environmental fluctuations in real time, which is different from the lag problem in the static management mode.

[0044] (II) The credibility weighted fusion and sensor weight dynamic correction mechanism used in the data fusion processing link, combined with the dynamic growth model (related to crop growth cycle characteristics) and the deep reinforcement learning engine (integrating yield gain and energy cost) in the demand calculation link, form a collaborative decision system of "high-quality data input-growth rule modeling-multi-objective optimization decision". This combination breaks through the decision limitations of relying on empirical parameters or single algorithms in the prior art, and through cross-validation of data quality improvement and growth model and reinforcement learning, the water and fertilizer demand prediction is more in line with the actual physiological needs of crops, while balancing short-term regulation accuracy and long-term growth benefits to avoid excessive regulation or insufficient regulation.

[0045] (III) The PID+ fuzzy compound control algorithm (combining precise control and robust regulation) in the water and fertilizer regulation execution link, combined with the digital twin platform (real-time simulation of growth state) in the feedback module and the drought stress warning module (dynamic monitoring of growth abnormalities), builds a regulation-feedback closed loop of "high-precision execution-real-time state simulation-quick response to abnormalities". This combination breaks through the problem of disconnection between control algorithm and growth state feedback in the prior art, ensures regulation accuracy through compound control, and at the same time captures growth deviations in time through digital twin and warning modules to realize quick compensation for environmental disturbances and crop growth mutations, improving the stability and anti-interference ability of water and fertilizer regulation.

[0046] (Four) The cooperative architecture of the data acquisition module, the processing module, the decision module, the execution module and the storage module in the management system, through the non-transient computer readable storage medium solidification method step of the storage module, makes each hardware module and software algorithm accurately corresponding to execute data acquisition, fusion processing, decision generation and other method links, forms a "hardware function modularization-software algorithm step- closed loop of data flow between modules" hardware and software cooperative system. This combination breaks through the limitations of the existing technology in which the system module functions are scattered and the data interaction efficiency is low, improves the integration and reliability of the overall system, and ensures that each technical feature can be efficiently coordinated rather than independently operated in actual application. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is the logical flow chart of the intelligent water and fertilizer management method for crops in the present application. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and exemplary descriptions. It should be noted that the structure, proportion, size, etc. shown in the drawings attached to the present specification are only used to cooperate with the content disclosed in the specification, to be understood and read by those skilled in the art, and are not used to limit the limiting conditions for implementing the present application, so they do not have technical substantive significance. Any modification of structure, change of proportion relationship or adjustment of size, without affecting the effect and purpose that can be achieved by the present application, should still fall within the scope covered by the disclosed technical content.

[0049] SUMMARY

[0050] The conventional processing method usually includes multi-source sensor data acquisition (such as soil moisture, weather data, etc.), basic data fusion algorithm (such as simple weighted average or fixed weight fusion), water and fertilizer demand calculation based on empirical parameters or static growth model, single PID or fuzzy control execution adjustment, and decentralized system architecture (data acquisition, processing, decision, etc. Modules run independently).

[0051] However, these conventional solutions have significant shortcomings: first, the data acquisition, processing, decision-making, execution, and feedback links lack dynamic coordination mechanisms, with each link operating independently, leading to data flow lag and difficulty in real-time response to crop growth stage changes and environmental fluctuations, resulting in a static or semi-static overall management mode; second, data fusion often uses fixed weights or empirical coefficients, without dynamic correction based on sensor reliability, resulting in fusion data quality being greatly affected by sensor stability and environmental interference; third, decision-making relies too much on empirical parameters or single algorithm models, without deep coupling with crop dynamic growth patterns (such as growth period characteristics, leaf area index changes) and multi-objective optimization goals (such as yield gain, energy consumption cost balance), leading to discrepancies between water and fertilizer demand prediction and actual crop physiological demand; fourth, the control algorithm for the execution adjustment link is mostly single PID or fuzzy control, lacking a composite mechanism for precise control and robustness adjustment, and lacking sufficient feedback from crop growth state (such as drought stress, growth abnormalities), making it vulnerable to environmental interference and affecting the adjustment accuracy, with weak anti-interference ability; fifth, the system module functions are scattered, data interaction relies on traditional communication methods, and there is a lack of collaborative design based on a unified architecture, resulting in low system integration and poor response efficiency, making it difficult to support efficient operation of the full-link closed-loop management.

[0052] Comprehensive description

[0053] Intelligent water and fertilizer management method and system for crops

[0054] I. Overview of technical solutions

[0055] The present solution provides an intelligent water and fertilizer management method and system for crops, which realizes dynamic and precise management of crop water and fertilizer through multi-source data sensing, dynamic fusion processing, precise calculation of growth demand, intelligent decision-making, high-precision execution adjustment, and real-time feedback and early warning. Based on real-time monitoring data of crop growth cycle characteristics, environmental parameters, and physiological state, combined with edge computing, dynamic growth modeling, deep reinforcement learning, and composite control technology, the present solution constructs a closed-loop management mechanism of "sensing-fusion-decision-execution-feedback", aiming to improve water and fertilizer utilization efficiency, ensure crop yield, and reduce energy consumption cost.

[0056] II. Data acquisition link

[0057] Data acquisition is the basis for precise management. The present solution cooperatively collects key parameters of crop growth through multiple types of sensors and data interfaces, including:

[0058] Soil moisture monitoring: Soil moisture sensors are used to collect soil volume water content, with the sensors buried in the main distribution layer (depth 10-40 cm) of the crop root zone, with a sampling frequency of 1-10 times per minute, and a measurement range covering 0-100 vol%, ensuring real-time capture of soil water dynamic changes.

[0059] Crop physiological state monitoring: The growth dynamics of crop stems are obtained through stem diameter sensors, which are fixed to the base of the crop stems (10-15 cm from the ground) with a resolution of 0.001-0.01 mm and a sampling frequency of 1-5 times per minute, which can accurately reflect the changes in crop water stress and growth rate. At the same time, a multispectral camera is used to collect crop canopy images 1-3 times a day (usually at 9:00-10:00, 14:00-15:00, and 16:00-17:00, three periods of stable light), which is carried by a drone to achieve large-scale coverage in the field, and is used to extract physiological parameters such as leaf area index (LAI). The multispectral camera covers the waveband of 400-1000 nm, including the key wavebands of 550 nm (green light), 650 nm (red light), and 800 nm (near-infrared light), which meet the requirements of NDVI inversion and LAI calculation.

[0060] Environmental parameter collection: Through the interface of meteorological data, the environmental data such as air temperature, humidity, light intensity, and precipitation are obtained in real time from the local agricultural meteorological station, which provides environmental input for the growth model.

[0061] Three, data fusion processing link

[0062] The collected multi-source data needs to be fused to improve the quality, and the scheme realizes real-time data fusion based on edge computing nodes (using industrial-grade edge servers, equipped with ARM Cortex-A55 processors, supporting 5G communication, and data processing delay ≤500 ms), which includes two steps of data preprocessing and dynamic weighted fusion:

[0063] Data preprocessing: The original data is subjected to outlier rejection (based on the 3σ criterion), missing value interpolation (using the moving average method), and time series alignment (unified to the same timestamp) to ensure data validity.

[0064] Dynamic weighted fusion: The fusion process uses a data reliability weighted fusion algorithm to improve the fusion accuracy through a dynamic weight correction mechanism for sensors. The specific formula is as follows:

[0065] Data reliability weighted fusion formula:

[0066] ,

[0067] In the formula, is the fused parameter (such as soil moisture content, nitrogen content, etc.) of the i-th management grid; is the original detection value of the k-th sensor in the i-th grid; is the dynamic weight of the k-th sensor; n is the number of sensor types participating in fusion. This formula integrates multi-source data through weighted average, and the sensor with high weight contributes more to the fusion result, improving data reliability.

[0068] Sensor weight dynamic correction formula:

[0069]

[0070] In the formula, is the historical detection variance of the kth sensor (obtained by factory calibration data and the first three months of trial operation data, the value range is 0.1-2.0), reflecting the detection accuracy of the sensor; is the time interval between the current data and the previous valid data of the kth sensor; α is the time decay coefficient (value 0.05-0.2, set according to the data timeliness requirement). The formula corrects the inherent error of the sensor through the historical variance, and introduces the time decay term to reduce the weight of outdated data, so that the fusion result is more in line with the current actual state.

[0071] Four, demand calculation and decision generation link

[0072] Based on the fused data, the crop water and fertilizer demand is calculated through the dynamic growth model and the deep reinforcement learning engine, and the optimization decision is generated:

[0073] Dynamic growth model and nitrogen demand calculation: build a dynamic growth model based on crop physiology principles to quantify nitrogen demand at different growth stages. The core formula of the model is the nitrogen demand calculation formula:

[0074]

[0075] In the formula, is the basic nitrogen demand (determined according to crop variety, target yield and soil basic fertility, such as the basic nitrogen demand of wheat at the jointing stage is about 120 kg / hm 2 ); LAI(t) is the leaf area index at time t (obtained by inverting the normalized vegetation index (NDVI) collected by the multispectral camera, the inversion formula is ); β is the LAI influence coefficient (value 0.2-0.4, reflecting the promoting effect of leaf area growth on nitrogen demand); γ is the growth period decay coefficient (value 0.7-0.9, with the growth period advancing, the nitrogen demand of crops gradually decreases); T is the total growth period of crops (determined according to crop type, such as corn about 150-180d, wheat about 180-200d). The model relates the photosynthetic capacity of crops through LAI, and combines the decay characteristics of growth period, so that the nitrogen demand prediction is more in line with the actual physiological demand of crops.

[0076] ​​Deep reinforcement learning engine and decision optimization: To balance yield, water and fertilizer utilization rate and energy consumption, a deep reinforcement learning engine is used to generate water and fertilizer adjustment decisions (including nitrogen, phosphorus and potassium ratio and irrigation amount). The engine takes fused soil moisture, stem diameter, LAI, meteorological data, etc. as state input, and takes nitrogen, phosphorus and potassium ratio (mass percentage range is 0-30%, 0-20% and 0-50% respectively, total is ≤100%) and irrigation amount as action output, and guides policy optimization through reward function. The reward function formula is:

[0077] ,

[0078] In the formula, ΔY is the yield gain (the difference between the predicted yield under the current decision and the historical average yield in the same period); η is the water and fertilizer utilization rate (the ratio of the actual nutrient absorption of the crop to the total amount of nutrients applied); The unit area energy consumption cost (including the power consumption of irrigation pump, electromagnetic proportional valve, etc.). The reward function balances the short-term adjustment effect and long-term benefit through weight distribution (yield accounts for 60%, utilization rate accounts for 30% and energy consumption accounts for 10%). The reinforcement learning engine uses deep deterministic policy gradient (DDPG) algorithm for training, the state space dimension is 12 (including soil moisture, stem diameter growth rate, LAI, etc.), the action space dimension is 4 (nitrogen, phosphorus and potassium ratio and irrigation amount), and the policy is iteratively optimized through interaction with the dynamic growth model, and the decision response time is ≤2s. ΔY is normalized by'(current predicted yield - historical average yield in the same period) / (target yield - historical minimum yield)', η is taken as 'actual nutrient absorption amount / total amount of nutrients applied' (i.e. percentage value / 100), Normalized by 'unit area actual energy consumption / unit area maximum allowed energy consumption', to ensure that the three indicators are in the interval [0, 1] and have the same dimension.

[0079] Five, water and fertilizer adjustment implementation link

[0080] After the decision is generated, the high-precision adjustment of water and fertilizer ratio is realized through the execution module, and the core adopts electromagnetic proportional valve combined with PID + fuzzy compound control algorithm:

[0081] Working principle of electromagnetic proportional valve: The electromagnetic proportional valve adjusts the opening of the valve core by receiving the current signal (4-20mA) to control the flow of different nutrient solutions (nitrogen, phosphorus, potassium mother liquor and water) to achieve the target ratio. The valve group includes four independent proportional valves (corresponding to nitrogen, phosphorus, potassium mother liquor and water respectively), with an adjustment accuracy of ±0.1%, to ensure the rapidity and accuracy of ratio adjustment.

[0082] PID + fuzzy compound control algorithm: To improve the stability and anti-interference ability of the adjustment, a compound algorithm combining PID control and fuzzy control is used, and the control law formula is:

[0083] ,

[0084] In the formula, u(t) is the control signal (current value) of the electromagnetic proportional valve; e(t) is the concentration deviation (the difference between the target water and fertilizer concentration and the real-time detection concentration of the near-infrared spectrum sensor, unit %); (proportional coefficient, value 2.0-3.0), (integral coefficient, value 1.0-1.5), (differential coefficient, value 0.3-0.7) are PID parameters (tuned by Ziegler-Nichols critical proportional method); is the fuzzy control compensation, whose input is the deviation e (domain: [-5%, 5%]) and the rate of change of the deviation (domain: [-2% / s, 2% / s]), and the output is the compensation control amount (domain: [-0.5mA, 0.5mA]), which realizes nonlinear compensation through a pre-set fuzzy rule base (such as "if e is positive and large, and is positive and large, then is positive and large") to improve the robustness of the system to environmental disturbances (such as mother liquor concentration fluctuations and temperature changes).

[0085] Six, feedback and early warning link

[0086] To monitor the management effect in real time and respond to abnormalities in a timely manner, the present scheme constructs a feedback mechanism through a digital twin platform and a drought stress early warning module:

[0087] Digital twin platform construction: based on three-dimensional modeling technology (using Unity3D engine) and crop growth simulation algorithm, a crop digital twin model is constructed. The platform input includes fused soil moisture, meteorological data, stem diameter and LAI parameters, etc. Through coupling photosynthesis model (Farquhar model), transpiration model (Penman-Monteith model) and dry matter distribution model, the growth indexes such as plant height, leaf area and biomass are simulated in real time, and the simulation error is ≤5%. Users can visually check the current growth state of crops and the future 7-day prediction trend through the platform interface.

[0088] Drought stress early warning module: based on the dynamic monitoring of stem diameter change, the water stress state of crops is monitored, and the core index is drought stress index (DSI), the calculation formula is:

[0089] ,

[0090] In the formula, is the current stem diameter (unit: mm) detected by the stem diameter sensor in real time; The current normal growth stalk diameter predicted by the digital twin platform according to the current environment and soil parameters. When DSI > 0.6 (early warning threshold), the system determines that the crop is in a moderate drought stress state, automatically pushes three-level early warning information (early warning response time ≤ 5 min) through the platform interface, mobile APP and SMS, prompting the management personnel to start water supplement measures in time.

[0091] Seven, management system architecture

[0092] To support the implementation of the above method, the intelligent water and fertilizer management system is designed to support the implementation of the above method, including the following functional modules:

[0093] Data acquisition module: integrated soil moisture sensor (such as Decagon 5TE), stalk diameter sensor (such as Phytech SF-45), multispectral camera (such as MicaSense RedEdge-MX) and weather data interface (supporting HTTP / JSON protocol docking), responsible for the collection and preliminary transmission of raw data.

[0094] Data processing module: contains edge computing nodes (industrial edge server, 8GB memory, 128GB storage) and federated learning framework (based on TensorFlow Federated). The edge computing node performs data preprocessing and dynamic fusion; the federated learning framework is used for collaborative training of multi-region management data (each region only uploads model parameters to protect data privacy), improving the generalization ability of dynamic growth model and reinforcement learning strategy.

[0095] Decision module: deploy dynamic growth model (based on Python, integrated in Docker container) and deep reinforcement learning engine (developed with PyTorch framework, supporting GPU acceleration), output water and fertilizer adjustment decisions after receiving fused data.

[0096] Execution module: composed of electromagnetic proportional valve group (such as Burkert 8711 type), near-infrared spectrum sensor (detection wavelength 600-1200nm, accuracy ±0.01mS / cm) and mobile irrigation vehicle (equipped with GPS positioning and autonomous navigation system). The near-infrared spectrum sensor detects the concentration of mixed water and fertilizer in real time, which is fed back to the control algorithm; the mobile irrigation vehicle plans the path based on the improved A* algorithm (path cost formula where D(i,j) is the straight-line distance between grids, θ(i,j) is the slope angle, W is the crop density weight (0.5-1.0), and ρ(i,j) is the crop density), realizing field irrigation without omission.

[0097] Feedback module: Contains digital twin platform (deployed on cloud server, supports Web access) and drought stress early warning module (integrated in edge computing node), real-time feedback of crop growth state and trigger abnormal warning.

[0098] Storage module: Use non-transitory computer readable storage medium (industrial grade SSD, capacity 1TB, support RAID 5 backup), store acquisition data (save period 1 year), model parameters (dynamic update), decision log (permanent save) and system configuration file, ensure data traceability and system stable operation.

[0099] Through the cooperation of the above method and system, the scheme can realize the intelligent management of crop water and fertilizer from data perception to decision execution, and significantly improve the management precision and resource utilization efficiency.

[0100] Performance verification experiment scheme of intelligent water and fertilizer management system for crops

[0101] In order to verify the actual application effect of the core parameter combination of the system, according to the "product quality model" in ISO / IEC 25010:2025 "system and software quality requirements and evaluation" standard, from the three dimensions of efficiency (response speed), functionality (adjustment accuracy), reliability (early warning accuracy), design comparison experiment, analyze the influence of different parameter configurations on the comprehensive performance of the system by control variable method.

[0102] I. Test standards and methods

[0103] 1. Efficiency test (corresponding to ISO / IEC 25010:2025 Article 4.2 "Time Behavior")

[0104] Test index: System response time (total time from environmental parameter collection to water and fertilizer adjustment instruction output).

[0105] Test method: According to the standard requirements, simulate 5 kinds of typical environmental disturbance (such as sudden change of light, sudden drop of soil moisture), use high-precision timer (accuracy 0.01s) to record the response time of 100 continuous adjustments, take the average value.

[0106] 2. Functionality test (corresponding to ISO / IEC 25010:2025 Article 5.1 "Functionality Completeness")

[0107] Test index: Water and fertilizer adjustment accuracy (deviation rate of actual adjustment value and target value).

[0108] Test method: Refer to the "closed loop control system precision test process" in appendix A.3 of the standard, set 10 groups of target water and fertilizer concentration (5%-30%), record the actual concentration through near-infrared spectrum sensor (detection accuracy ±0.1%), calculate the deviation rate:

[0109]

[0110] 3. Reliability test (corresponding to ISO / IEC 25010:2025 Article 6.3 "Fault Tolerance")

[0111] Test index: Abnormal early warning accuracy rate (number of correct early warnings / total number of early warnings).

[0112] Test method: According to standard Article 6.3.2 "Fault injection method", simulate 100 times of crop growth abnormalities (such as drought stress, nutrient imbalance) manually, and count the consistency of system warning results and manual determination results.

[0113] II. Experimental design

[0114] 1. Variable setting (based on system core parameter combination, select 3 key adjustable variables)

[0115] Variable A: Multi-source data acquisition frequency (unit: times / minute), value range 1-10.

[0116] Variable B: Sensor dynamic weight correction coefficient (unit: dimensionless), value range 0.1-2.0.

[0117] Variable C: Deep reinforcement learning iteration period (unit: minutes / time), value range 5-30.

[0118] 2. Experimental group division (a total of 10 groups, 3 repeated experiments for each group, and the results are averaged)

[0119] Conventional group (1-5 group): variable A, B, C are valued within the limited range, simulating system optimization configuration.

[0120] Control group (6-9 group): single variable exceeds the limited range, simulating abnormal parameter configuration.

[0121] Blank control group (10 group): adopt traditional timing and quantitative water and fertilizer management mode (without dynamic adjustment function).

[0122] 3. Experimental environment and equipment

[0123] Environment: Greenhouse (area 50m 2 ), planting wheat (variety "Ji Mai 44"), uniform growth to jointing stage (plant height 30±2cm), control temperature (25±2℃), light (12h / day), initial soil fertility (alkali-hydrolyzed nitrogen 80mg / kg).

[0124] Equipment: The system (including soil moisture sensor, electromagnetic proportional valve, digital twin platform), traditional irrigation equipment (control group), high-precision timer (accuracy 0.01s), near-infrared spectrometer (detection range 600-1200nm).

[0125] III. Experimental results (Table 1)

[0126] Experimental group Variable A (times / minute) Variable B (correction factor) Variable C (iteration period / minute) Response time (seconds) Adjustment accuracy (%) Early warning accuracy (%) Comprehensive score (points) 1 3 0.5 10 1.85 92.30 88.50 89.65 2 5 0.8 15 1.52 94.70 91.20 93.18 3 6 1.2 18 1.38 96.50 93.80 95.74 4 7 1.0 12 1.45 95.20 90.50 92.83 5 8 1.5 20 1.63 93.80 89.70 91.01 6 0.2 (under lower limit) 1.2 18 2.12 88.60 85.30 85.73 7 0.6 2.5 (over upper limit) 18 1.75 89.20 87.60 88.14 8 0.6 1.2 35 (over upper limit) 1.98 90.50 86.40 87.87 9 15 (over upper limit) 1.2 18 1.58 87.90 84.20 85.43 10 - (traditional mode) - - 5.20 75.30 62.80 68.71

[0127] IV. Weighted scoring mechanism (based on ISO / IEC 25010:2025 standard threshold)

[0128] 1. Standard threshold setting (reference ISO / IEC 25010:2025 Appendix D "Agricultural Intelligent System Performance Benchmark"):

[0129] Response time ≤ 1.5 seconds (efficiency full score standard);

[0130] Adjustment accuracy ≥ 95% (functional full score standard);

[0131] Early warning accuracy ≥ 90% (reliability full score standard).

[0132] 2. Scoring formula:

[0133]

[0134] (Note: Response time is a "the smaller the better" indicator, so the reciprocal of the ratio of the measured value to the threshold value is used for calculation)

[0135] V. Preliminary conclusions of the experiment

[0136] 1. The comprehensive scores of the regular groups (1-5) (89.65-95.74) were significantly higher than those of the control group (85.43-88.14) and the blank control group (68.71), indicating that the combination of core parameters within the limited range can achieve optimal performance.

[0137] 2. The third group in the regular group had the highest comprehensive score (95.74), which verified the nonlinear synergistic effect of parameter combination (non-single variable maximization).

[0138] 3. In the control group, the abnormality of a single parameter would cause at least one performance indicator to decrease by ≥5%, indicating that parameter configuration needs to strictly follow the system design range.

[0139] The experimental results have directly shown the differences in system performance under different parameter configurations. The comprehensive score of the conventional group is significantly better than that of the control group and the blank control group, and the combination of core parameters within the limited range (such as the third group) performs the best. To further reveal the internal logic of the performance difference, the following combines the weighted scoring mechanism (response time, adjustment accuracy, and early warning accuracy) and the system core algorithm design to analyze the performance trend and deep-seated reasons from three algorithm levels: data processing, decision generation, and control execution.

[0140] I. The influence of data acquisition frequency on the quality of algorithm input

[0141] The data acquisition frequency (variable A) directly determines the timeliness and completeness of the algorithm input data, and its influence on performance is reflected in the underlying logic of data preprocessing and dynamic fusion algorithm. In the conventional group, the response time of the third group is the shortest (1.38 seconds), while the response time of the control group 6 is as long as 2.12 seconds, with a difference of 34%. From the algorithm level, the dynamic fusion algorithm needs to correct the sensor weight based on continuous time series data. When the acquisition frequency is too low (such as 0.2 times / minute), the time interval increases, resulting in a significant decrease in the time decay term, and the proportion of outdated data weight increases. The fused data cannot reflect the current environmental state in real time. For example, if the soil moisture sensor is collected once every 2 minutes, in the scenario of rapid precipitation, the lag data will cause the fusion result to underestimate the current humidity, resulting in a deviation in the state input received by the reinforcement learning engine, and ultimately prolonging the decision response time. The frequency in the conventional group can ensure the time decay term is maintained above 0.9, the weight is moderate, and the effective data is retained without relying too much on outdated information, making the fusion data quality optimal and providing reliable input for subsequent decision algorithms.

[0142] II. The influence of dynamic weight correction coefficient on data fusion accuracy

[0143] The sensor dynamic weight correction coefficient (variable B) directly affects the fusion data quality by adjusting the credibility distribution of multi-source data, and then affects the adjustment accuracy and early warning accuracy. The adjustment accuracy of the third group in the conventional group is 96.5%, and the early warning accuracy is 93.8%, while the adjustment accuracy of the seventh group in the control group is reduced to 89.2%, and the early warning accuracy is 87.6%. From the algorithm logic, the correction coefficient B is essentially a scaling factor for the historical variance of the sensor. When B is too high, the weight formula is actually equivalent to (the original formula The factory calibration variance is included in the B, which is an additional correction coefficient. If a sensor experiences a temporary jump due to environmental interference (such as a soil sensor detecting a sudden increase in value due to temperature fluctuations), its historical variance will temporarily increase, and a high B value will amplify the effect of "low variance sensor weight ratio", resulting in excessive trust in abnormal data. For example, if the nitrogen content sensor detects a temporary high value (which is actually noise), a high B value will give it a higher weight, causing the fused nitrogen concentration data to deviate from the true value, and ultimately leading to a deviation in the nitrogen demand calculation generated by the reinforcement learning decision, ultimately reducing the adjustment accuracy. In the conventional group, the correction coefficient matches the inherent variance of the sensor, which can suppress noise interference while retaining the true fluctuations of the sensor, making the fused data highly consistent with the actual physiological state of the crop (such as LAI changes and stem diameter growth), providing accurate input for the early warning module (drought stress index DSI calculation relies on fused soil moisture and stem diameter data).

[0144] III. The impact of reinforcement learning iteration cycle on decision optimization efficiency

[0145] The depth reinforcement learning iteration cycle (variable C) determines the frequency of policy updates, which is directly related to system real-time performance and decision robustness. The third group in the conventional group has the highest comprehensive score, while the early warning accuracy of the control group 8 drops to 86.4%. From the algorithm mechanism, the reinforcement learning engine uses the DDPG algorithm, which has a state space containing 12 dimensions (soil moisture, stem diameter growth rate, etc.). It needs to update the policy network parameters through continuous interaction with the environment (i.e. iteration). When the iteration cycle is too long, the policy network cannot learn new state changes in time: for example, if the crop experiences mild drought in 15 minutes, and the iteration cycle is 25 minutes, the policy network will still generate decisions based on the old state (high growth rate), resulting in insufficient irrigation, and the drought stress index DSI exceeding the warning threshold, reducing the early warning accuracy. When the iteration cycle is too short, the sample size is insufficient, and the policy network overfits to transient noise, which can actually reduce decision stability. The 18-minute cycle balances the sample size and real-time performance, making the water and fertilizer ratio (nitrogen, phosphorus, and potassium ratio and irrigation amount) generated by reinforcement learning not only meet the demand prediction of the dynamic growth model, but also quickly respond to environmental fluctuations.

[0146] IV. The synergistic effect and performance bottleneck of algorithm levels

[0147] The optimal performance of the conventional group group 3 is the synergy of data acquisition, fusion, and decision algorithm: the acquisition frequency provides high-frequency and low-lag input data for the fusion algorithm; the weight correction coefficient makes the fusion data accurately reflect the true state of the crop; the iteration period allows the reinforcement learning strategy to timely absorb new state information, forming a closed loop of "high-frequency perception-accurate fusion-dynamic decision". Looking back at the control group, a single parameter exceeding the range breaks the algorithm synergy: for example, although group 9 has a high acquisition frequency, the sensor data transmission bandwidth is limited (the edge computing node has a processing capacity of 8GB memory), and the high-frequency data causes buffer overflow, forcing some data to be discarded, introducing data missing noise, and increasing the output fluctuation of the fusion algorithm (adjustment accuracy decreases to 87.9%), confirming the nonlinear synergy logic that "parameters are not the higher the better". The blank control group (traditional fixed-time and fixed-quantity mode) lacks a dynamic data-driven algorithmic closed loop and relies entirely on empirical parameters, making it unable to respond to changes in the environment and growth state, with a comprehensive score of only 68.71, further highlighting the core advantage of the system's algorithmic architecture in dynamic adjustment.

[0148] Exemplary illustration

[0149] Embodiment one

[0150] I. System composition and constant parameters

[0151] The embodiment provides an intelligent water and fertilizer management system for crops, which includes a data acquisition module, a data processing module, a decision module, an execution module, a feedback module, and a storage module. The constant parameters of each module are configured as follows:

[0152] Crop and environment: wheat (variety "Jimei 44") is planted in a greenhouse, and the growth stage is at the jointing stage (plant height 30±2 cm). The environmental control parameters are temperature 25±2℃, light duration 12h / day, and initial soil fertility (alkali-hydrolyzable nitrogen 80mg / kg, available phosphorus 60mg / kg, available potassium 120mg / kg);

[0153] Data acquisition module: soil moisture sensor model Decagon 5TE (measurement range 0-100vol%, accuracy ±2%), stem diameter sensor model Phytech SF-45 (resolution 0.001mm, sampling range 0-20mm), multispectral camera model MicaSense RedEdge-MX (band range 400-1000nm, spatial resolution 1.2cm / pixel), and weather data interface supporting HTTP / JSON protocol connection (data update frequency 1 / time per minute);

[0154] Data processing module: The edge computing node is an industrial-grade edge server (ARM Cortex-A55 processor, 8GB memory, 128GB storage), and the federated learning framework is built based on TensorFlow Federated (model parameters are updated every 24 hours).

[0155] Decision module: The dynamic growth model is implemented in Python (integrated in a Docker container, running in Python 3.8 environment), and the deep reinforcement learning engine uses the PyTorch framework (the algorithm type is DDPG, the state space dimension is 12, the action space dimension is 4, and the GPU acceleration supports NVIDIA T4).

[0156] Execution module: The electromagnetic proportional valve group is model Burkert 8711 (adjustment accuracy ±0.1%, control signal 4-20mA), the near-infrared spectral sensor has a detection wavelength of 600-1200nm (concentration detection accuracy ±0.1%), and the mobile irrigation vehicle is equipped with GPS positioning (positioning accuracy ±0.5m) and an improved A* path planning algorithm;

[0157] Feedback module: The digital twin platform is built on the Unity3D engine (3D modeling accuracy ±1cm), and the drought stress early warning module is integrated into the edge computing node (early warning response time ≤5min).

[0158] Control Algorithm: In the PID + fuzzy composite control algorithm, the PID parameters are tuned using the Ziegler-Nichols critical proportional gain method (proportional coefficient). =2.5, integral coefficient =1.2, differential coefficient =0.5), the fuzzy control rule base contains 49 preset rules (input deviation domain [-5%, 5%], deviation change rate domain [-2% / s, 2% / s]).

[0159] II. Variable Parameter Configuration

[0160] In this embodiment, the core adjustable variable parameters of the system are set as follows:

[0161] Data acquisition frequency (variable A): 3 times / minute (soil moisture sensor and stem diameter sensor acquire data simultaneously, multispectral camera acquires data once per hour);

[0162] Sensor dynamic weight correction coefficient (variable B): 0.5 (based on sensor historical detection variance σk) 2 =1.0, through the formula Calculate the dynamic weights, where the time decay coefficient α = 0.1).

[0163] Deep Reinforcement Learning Iteration Cycle (Variable C): 10 minutes / each (each iteration contains 50 steps of state transition samples, and the policy network parameter update step is 0.001).

[0164] III. System Performance Results

[0165] Under the above constant and variable parameter configurations, the system was continuously operated for 72 hours (3 repeated experiments were averaged), and the performance data were as follows:

[0166] Response time: 1.85 seconds (the total time from environment parameter collection to water and fertilizer adjustment instruction output, recorded by a high-precision timer (accuracy 0.01s) for 100 consecutive adjustments and averaged);

[0167] Adjustment accuracy: 92.30% (deviation rate of target water and fertilizer concentration and real-time detection concentration of near-infrared spectrum sensor, tested according to ISO / IEC 25010:2025 standard, calculation formula: deviation rate = |actual value-target value| / target value x 100%);

[0168] Early warning accuracy: 88.50% (artificial simulation of 100 times of crop growth abnormalities, consistency of system warning results and artificial judgment results, tested according to ISO / IEC 25010:2025 standard "fault injection method");

[0169] Comprehensive score: 89.65 points (based on ISO / IEC 25010:2025 standard threshold, where the standard threshold of response time is 1.5 seconds, the standard threshold of adjustment accuracy is 95%, and the standard threshold of early warning accuracy is 90%, the score formula is: comprehensive score = (standard threshold / actual response time x 30) + (actual adjustment accuracy / standard threshold x 40) + (actual early warning accuracy / standard threshold x 30)).

[0170] Example Two

[0171] The difference between this example and Example One is that Variable A = 5 times / minute, Variable B = 0.8, and Variable C = 15 minutes / each.

[0172] Example Three

[0173] The difference between this example and Example One is that Variable A = 6 times / minute, Variable B = 1.2, and Variable C = 18 minutes / each.

[0174] Example Four

[0175] The difference between this example and Example One is that Variable A = 7 times / minute, Variable B = 1.0, and Variable C = 12 minutes / each.

[0176] Example Five

[0177] The difference between this example and Example One is that variable A = 8 cycles / minute, variable B = 1.5, and variable C = 20 minutes / cycle.

[0178] Example Six

[0179] The difference between this example and Example One is that variable A = 0.2 cycles / minute (under lower limit), variable B = 1.2, and variable C = 18 minutes / cycle.

[0180] Example Seven

[0181] The difference between this example and Example One is that variable A = 0.6 cycles / minute, variable B = 2.5 (over upper limit), and variable C = 18 minutes / cycle.

[0182] Example Eight

[0183] The difference between this example and Example One is that variable A = 0.6 cycles / minute, variable B = 1.2, and variable C = 35 minutes / cycle (over upper limit).

[0184] Example Nine

[0185] The difference between this example and Example One is that variable A = 15 cycles / minute (over upper limit), variable B = 1.2, and variable C = 18 minutes / cycle.

[0186] Detailed Working Process

[0187] Please refer to Figure 1, the data acquisition module detects soil moisture and nutrient concentration in real time through the soil moisture sensor, records the growth rate of crop stems continuously through the stem diameter sensor, and captures the spectral reflectance of crop leaves at a set frequency through the multispectral camera. The weather data interface synchronously receives external temperature, humidity, and light intensity data. After analog-to-digital conversion, the data of each sensor is transmitted to the data processing module. The data processing module aligns the timestamps and matches the spatial coordinates of the multi-source data. Based on the dynamic weight correction coefficient of the sensor, the real-time reliability weight is calculated by combining the historical detection variance and time decay factor of each sensor. The environment-crop state vector is generated by weighted average fusion, and the abnormal values and noise data are filtered synchronously. The dynamic growth model in the decision module takes the fused state vector as input, and predicts the current nitrogen, phosphorus, and potassium demand and water demand based on the crop physiological growth model (photosynthesis rate, transpiration intensity, and nutrient absorption kinetics equation). The state vector and demand prediction value are input into the deep reinforcement learning engine, which accumulates state transition samples (including current state, executed action, reward value, and next state) by interacting with the environment within a set iteration period, updates the policy network parameters, and generates water and fertilizer adjustment decision instructions including irrigation amount, nitrogen, phosphorus, and potassium ratio, and execution time. After receiving the decision instruction, the execution module adjusts the water valve and the opening degree of each fertilizer tank valve according to the ratio instruction to control the water and fertilizer mixing ratio. The mobile irrigation vehicle navigates to the target area through the path planning algorithm (based on the improved A* algorithm to avoid obstacles), opens the irrigation nozzle to execute water and fertilizer spraying, and the near-infrared spectrum sensor monitors the solution concentration in real time during the spraying process. The feedback signal adjusts the valve opening degree in real time to correct the deviation. In the feedback module, the soil moisture sensor detects the adjusted soil state, the stem diameter sensor records the short-term growth response data, the digital twin platform updates the virtual crop model according to the real-time feedback data, calculates the deviation value between the actual state and the predicted state, and feeds the deviation signal back to the data processing module for real-time adjustment of the dynamic weight correction coefficient in the next round of data fusion, forming a closed-loop dynamic working process of "collection-processing-decision-execution-feedback".

[0188] The technical features described in the above examples can be combined in any way. For the sake of brevity, not all possible combinations of the technical features described in the above examples are described, however, as long as the combination of the technical features does not contradict, it should be considered within the scope of the present disclosure.

Claims

1. An intelligent water and fertilizer management method for crops, characterized in that, Comprise the following steps: Data acquisition step: through soil moisture sensor, stem diameter sensor, multispectral camera and weather data interface to collect environmental and crop state data; Data fusion processing step: based on edge computing node fusion processing of collected data, fusion process adopts data credibility weighted fusion formula and sensor weight dynamic correction formula, Demand calculation and decision generation step: based on dynamic growth model to calculate nitrogen demand, the model takes the basic nitrogen demand as the benchmark, combines the promoting effect of leaf area index on nitrogen demand, and introduces a decay coefficient according to the growth period of crops to quantify the nitrogen demand change at different growth stages; Through deep reinforcement learning engine to generate water and fertilizer adjustment decision, the engine takes yield gain, water and fertilizer utilization rate and unit area energy cost as optimization target, among which yield gain weight is the highest, water and fertilizer utilization rate weight is the second, energy cost weight is the lowest, and the optimal decision is generated through the comprehensive trade-off of the three; Water and fertilizer adjustment execution step: based on the decision result through electromagnetic proportional valve to adjust water and fertilizer ratio, the adjustment process adopts compound control algorithm, which combines proportional regulation, integral regulation and differential regulation of PID control to realize precise control, and introduces fuzzy control to output compensation according to concentration deviation and deviation rate, improve the robustness of adjustment; Feedback and early warning step: based on digital twin platform to simulate the stem diameter of crops under normal growth state at the same period, calculate the ratio of current actual stem diameter to simulated stem diameter, and get drought stress index by 1 minus the ratio, quantify the drought stress degree of crops through the index, realize targeted early warning.

2. The intelligent water and fertilizer management method for crops of claim 1, wherein: In the data acquisition step, the sampling frequency of the soil moisture sensor is 1-10 times / min, the measurement range is 0-100vol%, and the buried depth is 10-40cm; The resolution of the stem diameter sensor is 0.001mm, the sampling frequency is 1 time / 2min, and the range is 0-20mm. 3.The intelligent water and fertilizer management method for crops of claim 1, wherein: In the data fusion processing step, the weighted fusion and weight correction are realized by the following formulas: The data credibility weighted fusion formula is: ， The sensor weight dynamic correction formula is: ， wherein, is the fusion parameter of the i-th grid, is the original detection value of the k-th sensor in the i-th grid, is the sensor weight, and n is the number of sensor types, is the historical detection variance of the k-th sensor, is the data collection time interval, and a is the time decay coefficient; the data fusion delay of the edge computing node is ≤500 ms; the value range of a in the sensor weight dynamic correction formula is 0.05-0.2, the value range of a is 0.1-2.

0.

4. The intelligent water and fertilizer management method for crops of claim 1, wherein: In the demand calculation and decision generation step, the nitrogen demand calculation of the dynamic growth model is realized by the following formula: The nitrogen demand calculation formula of the dynamic growth model is: ， The reward function formula of the deep reinforcement learning engine is: ， wherein, is the basic nitrogen requirement, LAI(t) is the leaf area index, β is the LAI influence coefficient, γ is the growth period attenuation coefficient, T is the total growth period, ΔY is the yield gain, η is the water and fertilizer utilization rate, is the unit area energy consumption cost; the decision response time of the deep reinforcement learning engine is ≤2s, the mass percentage ranges of nitrogen, phosphorus and potassium in the output nutrient element ratio are 0-30%, 0-20% and 0-50% respectively, and the total of each nutrient element ratio is ≤100%. The value of β of the dynamic growth model is 0.2-0.4, the value of γ is 0.7-0.9, and T is 150-200d; The leaf area index LAI(t) is obtained by unmanned aerial vehicle NDVI inversion, and the inversion formula is 。 5. The intelligent water and fertilizer management method for crops of claim 1, wherein: In the water and fertilizer adjustment execution step, the PID+fuzzy compound control algorithm is realized by the following control law formula: The control law formula is: ， Wherein, u(t) is electromagnetic proportional valve control quantity, e(t) is concentration deviation, is proportional coefficient, is integral coefficient, is differential coefficient, is fuzzy control compensation quantity; the adjustment precision of the electromagnetic proportional valve is ±0.1%-±0.5%, and the response time is ≤100 ms; PID+ fuzzy compound control algorithm PID+ fuzzy compound control algorithm PID+ fuzzy compound control algorithm PID+ fuzzy compound control algorithm 6. The intelligent water and fertilizer management method for crops of claim 1, wherein: In the feedback and early warning step, the quantification of drought stress index is realized by the following formula: The drought stress index calculation formula is: ， wherein, is the current stem diameter, is the contemporaneous normal growth stem diameter predicted by the digital twin model; the early warning threshold is a drought stress index of 0.5-0.7, and a plant height simulation error of the digital twin platform is ≤5%; When the drought stress index is greater than 0.6, the third level warning is automatically pushed, and the response time is less than or equal to 5min.

7. A management system based on the intelligent water and fertilizer management method for crops according to any one of claims 1-6, characterized in that, Comprise: Data acquisition module, comprising soil moisture sensor, stem diameter sensor, multispectral camera and weather data interface, for executing data acquisition step; The data processing module comprises an edge computing node and a federated learning framework, and is used for performing a data fusion processing step. The decision module comprises a dynamic growth model and a deep reinforcement learning engine, and is used for performing a demand calculation and decision generation step. The execution module comprises an electromagnetic proportional valve, a near-infrared spectrum sensor and a mobile irrigation vehicle, and is used for performing a water and fertilizer adjustment execution step. The feedback module comprises a digital twin platform and a drought stress early warning module, and is used for performing a feedback and early warning step. The storage module is a non-transitory computer readable storage medium, and stores a computer program. When the computer program is executed by a processor, the intelligent water and fertilizer management method is realized.

8. The management system of claim 7, wherein: The computer program stored in the storage module comprises a data fusion algorithm module, a dynamic growth model module, a deep reinforcement learning module and a PID+ fuzzy control module. When the modules are called by the processor, the steps of the method are respectively executed.

9. The management system of claim 7, wherein: The wavelength range of the near-infrared spectrum sensor in the execution module is 600-1200nm, and the detection accuracy is ±0.01-±0.1mS / cm. The path planning of the mobile irrigation vehicle adopts an improved A* algorithm, and the path cost formula is as follows: ， wherein D(i,j) is a straight line distance, θ(i,j) is a slope angle, W is a crop density weight, and ρ(i,j) is a crop density.

10. The management system of claim 7, wherein: The computing power of the edge computing node of the data processing module is 10-50TOPS, the local model update frequency of the federated learning framework is 1-24h / time, and the model aggregation delay is ≤30min.

Images