A greenhouse crop water and fertilizer integrated precision decision system based on deep learning

Abstract

The application discloses a greenhouse crop water and fertilizer integrated precision decision system based on deep learning, and relates to the technical field of precision irrigation and intelligent water and fertilizer management of facilities agriculture.The system core is composed of a water and fertilizer decoupling prediction module, a digital twin pre-simulation module, a conflict perception fusion module and an execution compensation module.The four modules are sequentially connected to form a complete water and fertilizer decision and execution closed loop.The system completes decoupling prediction of crop water and fertilizer demand through a double-branch deep neural network, establishes an ion migration rate matrix based on soil physical and chemical properties, calculates water and fertilizer spatio-temporal matching degree and optimizes fertilization timing, completes water and fertilizer scheme pre-simulation and multi-objective optimization through a lightweight crop digital twin, realizes conflict perception and intelligent fusion of data-driven and knowledge-driven decisions, and performs real-time compensation and correction on execution equipment deviation.The application significantly improves water and fertilizer utilization efficiency and adapts to the dynamic water and fertilizer management needs of greenhouse crops throughout the growth period.

Description

Technical Field

[0001] This invention relates to the field of precision irrigation and intelligent water and fertilizer management technology in facility agriculture, specifically a precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning. Background Technology

[0002] Integrated water and fertilizer management technology is a new agricultural technology that combines irrigation and fertilization. Utilizing a pressure system or natural terrain gradient, a fertilizer solution made from soluble solid or liquid fertilizer is mixed with irrigation water and delivered evenly, regularly, and quantitatively to the crop root zone through a controllable pipeline system. This technology maintains suitable moisture and nutrient levels in the soil around the crop roots by controlling irrigation and fertilization volumes, making it a crucial technique for achieving water and fertilizer conservation, and improving quality and efficiency in modern facility agriculture. With the rapid development of the Internet of Things, big data, and artificial intelligence technologies, intelligent integrated water and fertilizer management decision-making systems have become a research hotspot in precision agriculture. Their core lies in using sensors to collect environmental and crop growth data in real time and employing intelligent algorithms to generate optimal water and fertilizer application plans, replacing traditional experience-based management methods.

[0003] In the prior art, invention patent CN121638371A proposes a "training method, system, computer equipment, and medium for an integrated water and fertilizer decision-making model." This scheme acquires environmental state information and corresponding decision-making actions, inputs the environmental state information into a crop growth model to obtain predicted growth indicators, constructs a training dataset based on the environmental state information, predicted growth indicators, and decision-making actions, and trains the model using a deep reinforcement learning agent to obtain an integrated water and fertilizer decision-making model. This method introduces deep reinforcement learning technology, overcoming the short-sightedness of traditional decision-making methods and giving water and fertilizer management a certain degree of foresight. However, this scheme inputs water and fertilizer as coupling parameters into the same model, without considering the physical differences in the transport rates of water and fertilizer in the soil. In practical applications, this can easily lead to situations where water has reached the root zone before fertilizer has arrived, or fertilizer has arrived first but lacks water dissolution and absorption, resulting in asynchronous water and fertilizer application and affecting crop absorption efficiency.

[0004] The invention patent with publication number CN117744966A proposes a "method and device for intelligent water and fertilizer management of greenhouse tomatoes integrating agricultural experience knowledge." This scheme inputs the target parameters to be identified into a crop coefficient prediction model, obtaining the crop coefficient output by the model. The crop coefficient prediction model is obtained by training an extreme learning machine model based on a training set and a knowledge function. The knowledge function is determined based on a loss function and a conflict function, which in turn is determined based on the output results and domain experience knowledge. This method introduces agricultural experience knowledge as a constraint during the model training phase, improving irrigation accuracy. However, this scheme only constrains the model through the conflict function during the training phase. When data-driven results conflict with experience knowledge during actual system operation, it cannot provide real-time dynamic response and adjustment, making it difficult to cope with sudden diseases, extreme weather, and other emergencies.

[0005] The aforementioned existing technologies have not addressed the physical-level technical problem of the difference in the mechanics of water and fertilizer transport during integrated water and fertilizer management decisions. Water transport in soil is primarily influenced by factors such as soil texture, porosity, and initial moisture content, while the transport rates of fertilizers in different ionic forms are determined by ionic radius, ionic valence state, soil cation exchange capacity, and soil pH, exhibiting significant differences. Ignoring this physical difference and treating water and fertilizer as uniform coupling parameters during decision-making inevitably leads to a mismatch in the timing of water and fertilizer reaching the crop root zone during actual application, resulting in deep seepage or ion antagonism and reduced water and fertilizer use efficiency. Therefore, how to decouple and model the mechanics of water and fertilizer transport based on soil physicochemical properties, and generate time-matched water and fertilizer decision-making schemes accordingly, is a pressing technical problem that needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a precision decision-making system for integrated water and fertilizer management in greenhouse crops based on deep learning. By constructing a water and fertilizer decoupled neural network, it predicts water demand and fertilizer ion demand separately. Based on soil cation exchange capacity, it establishes an ion transport rate matrix, calculates the spatiotemporal matching degree between water and each ion, and adjusts the fertilization sequence until the target is met. This solves the problem of low synchronous absorption efficiency caused by ignoring the differences in the dynamics of water and fertilizer transport in existing technologies, and realizes precise synergy of water and fertilizer at the physical level.

[0007] To solve the above-mentioned technical problems, this invention provides the following technical solution: a deep learning-based precision decision-making system for integrated water and fertilizer management in greenhouse crops, comprising:

[0008] The water and fertilizer decoupling prediction module includes a dual-branch deep neural network. The first branch predicts water demand based on crop transpiration rate and rhizosphere matrix potential, while the second branch predicts fertilizer ion demand based on crop photosynthetic product distribution and rhizosphere ion concentration. The water and fertilizer decoupling prediction module also establishes an ion transport rate matrix based on soil cation exchange capacity, calculates the spatiotemporal matching degree between water and each ion according to the ion transport rate matrix, and adjusts the fertilization sequence when the spatiotemporal matching degree is lower than a preset matching degree threshold until the spatiotemporal matching degree reaches the preset matching degree threshold, and then outputs a preliminary water and fertilizer decision scheme.

[0009] The digital twin pre-simulation module includes a lightweight crop digital twin, which integrates a radiative transfer model and a photosynthetic model. The digital twin pre-simulation module inputs the preliminary water and fertilizer decision scheme into the lightweight crop digital twin, pre-simulates the photosynthetic rate and transpiration rate within a preset time period after the preliminary water and fertilizer decision scheme is executed, calculates the net photosynthetic gain index, and performs multi-objective optimization of the preliminary water and fertilizer decision scheme based on the net photosynthetic gain index before outputting an optimized water and fertilizer decision scheme.

[0010] The conflict perception and fusion module includes an agronomic knowledge graph. It acquires a data-driven decision-making scheme and a knowledge-driven decision-making scheme. The data-driven decision-making scheme originates from the optimized water and fertilizer decision-making scheme, and the knowledge-driven decision-making scheme originates from the agronomic knowledge graph. The conflict perception and fusion module calculates the conflict coefficient between the data-driven and knowledge-driven decision-making schemes. When the conflict coefficient exceeds a preset conflict threshold, it calls a third-party verification source to perform Bayesian confidence assessment. Based on the posterior confidence, it performs a weighted fusion of the data-driven and knowledge-driven decision-making schemes and outputs the final water and fertilizer decision-making scheme.

[0011] The execution compensation module monitors the status of the execution equipment and calculates the execution efficiency coefficient. Based on the execution efficiency coefficient, it compensates and corrects the final water and fertilizer decision scheme and generates an execution command to be sent to the execution equipment.

[0012] By constructing a water and fertilizer decoupling prediction module containing a dual-branch deep neural network, it is possible to independently predict water demand and fertilizer ion demand based on crop transpiration and photosynthetic product distribution, respectively, thus achieving precise decoupling of water and fertilizer at the demand level.

[0013] Furthermore, the dual-branch deep neural network includes a water demand prediction subnetwork and a fertilizer demand prediction subnetwork;

[0014] The input layer of the water demand prediction subnetwork receives crop transpiration rate, root zone matrix potential, leaf area index and canopy temperature data, and the output layer outputs water demand.

[0015] The fertilizer demand prediction subnetwork's input layer receives data on photosynthetically active radiation, carbon dioxide assimilation rate, crop growth stage, and rhizosphere ion concentration, while its output layer outputs the ionic demand for nitrogen, phosphorus, potassium, and trace elements.

[0016] By establishing an ion transport rate matrix based on soil cation exchange capacity, we can quantify the physical differences in the transport rates of fertilizers and water with different ion forms in the soil, providing a scientific basis for subsequent time-series adjustments.

[0017] Furthermore, the ion transport rate matrix is ​​expressed as:

[0018]

[0019] in, This represents the ion transport rate matrix, composed of water transport rate and the transport rates of each ion. Each ion rate is determined based on the soil cation exchange capacity. This indicates the rate of water transport in the soil. It is determined by soil texture, soil porosity, and initial soil moisture content;

[0020] , , , These represent the transport rates of nitrogen, phosphorus, potassium, and calcium ions in the soil, respectively. The transport rate of each ion is determined by its ionic radius, ionic valence state, soil cation exchange capacity, and soil pH.

[0021] By calculating the spatiotemporal matching degree between water and various ions and adjusting the fertilization sequence when it is below the threshold, water and fertilizer can reach the crop root zone synchronously, avoiding deep leakage or ion antagonism caused by differences in transport rates.

[0022] Furthermore, the spatiotemporal matching degree is expressed as:

[0023]

[0024] in, This indicates the spatiotemporal matching degree between water and various ions. Indicates the time it takes for water to reach the crop root zone. , , , These represent the time it takes for each ion to reach the crop root zone. This indicates taking the minimum value among all ratios.

[0025] By integrating a lightweight crop digital twin with a radiative transfer model and a photosynthetic model, the impact of water and fertilizer programs on crop photosynthetic and transpiration rates can be pre-simulated before decision-making and execution, thus quantifying short-term physiological responses.

[0026] Furthermore, the water-fertilizer decoupling prediction module also includes a fertilization timing adjustment unit. When the spatiotemporal matching degree is lower than the preset matching degree threshold, the fertilization timing adjustment unit applies ions with a migration rate lower than that of water in advance and applies ions with a migration rate higher than that of water in a later period, and recalculates the adjusted spatiotemporal matching degree until the spatiotemporal matching degree reaches the preset matching degree threshold.

[0027] By calculating the net photosynthetic gain index and performing multi-objective optimization, it is possible to avoid short-term photosynthetic inhibition that may be caused by irrigation decisions while pursuing long-term yield, so that crops are always in their physiological comfort zone.

[0028] Furthermore, the lightweight crop digital twin includes the PROSAIL radiative transport model and the Farquhar photosynthesis model;

[0029] The PROSAIL radiative transfer model outputs canopy reflectance and photosynthetically active radiation absorption ratio based on leaf structure parameters, chlorophyll content, and leaf tilt angle distribution data.

[0030] The Farquhar photosynthetic model outputs the photosynthetic rate based on ribulose-1,5-bisphosphate carboxylase activity, electron transport rate, carbon dioxide concentration, and light intensity data.

[0031] By monitoring the differences between data-driven and knowledge-driven decision-making in real time through the conflict perception module, contradictions between data and experience can be identified in a timely manner in case of emergencies, avoiding misjudgments caused by the failure of a single decision source.

[0032] Furthermore, the net photosynthetic gain index is expressed as:

[0033]

[0034] in, Indicates the net photosynthetic gain index. This represents the crop photosynthetic rate before implementing the preliminary water and fertilizer decision-making scheme. This represents the simulated predicted value of crop photosynthetic rate within a preset time period after implementing the preliminary water and fertilizer decision-making scheme. This represents the crop transpiration rate before implementing the preliminary water and fertilizer decision-making scheme. This represents the simulated predicted value of crop transpiration rate within a preset time period after implementing the preliminary water and fertilizer decision-making scheme. This represents the opportunity cost weighting coefficient for water consumed through evaporation.

[0035] By calling third-party verification sources to conduct Bayesian confidence assessments, the reliability of different decision sources can be dynamically calculated based on objective measurement data, enabling intelligent optimization and fusion of data and experience.

[0036] Furthermore, the conflict coefficient is expressed as:

[0037]

[0038] in, This represents the conflict coefficient between data-driven and knowledge-driven decision-making approaches. The feature vector representing the data-driven decision-making scheme. The feature vector representing the knowledge-driven decision-making scheme. The norm of a vector.

[0039] By using a digital twin unit for actuators to monitor equipment status in real time and identify execution efficiency coefficients, it can automatically compensate for execution deviations caused by equipment aging or blockages, ensuring consistency between theoretical decisions and actual execution.

[0040] Furthermore, the conflict-aware fusion module includes a hyperspectral image analysis unit and a Bayesian inference unit;

[0041] The hyperspectral image analysis unit acquires crop hyperspectral image data in real time and analyzes the actual nutrient element content of the crop.

[0042] The Bayesian inference unit uses the actual nutrient element content of the crop as observational evidence to calculate the first posterior confidence of the data-driven decision-making scheme and the second posterior confidence of the knowledge-driven decision-making scheme, and uses the first posterior confidence and the second posterior confidence as the weights for weighted fusion.

[0043] By dividing the final decision by the execution efficiency coefficient to generate the execution command, the control quantities issued to the pump station and solenoid valve can be accurately matched with the actual needs of the crop, overcoming the accuracy loss caused by the non-ideal state of the equipment.

[0044] Furthermore, the execution compensation module includes an actuator digital twin unit;

[0045] The actuator digital twin unit analyzes the pump room motor current data, pipeline pressure fluctuation data, and dripper flow data in real time, and identifies the execution efficiency coefficient of the current execution system online. The execution efficiency coefficient is the ratio of the actual execution flow to the theoretical decision flow.

[0046] The execution compensation module divides the final water and fertilizer decision scheme by the execution efficiency coefficient to generate an execution instruction.

[0047] Through the collaborative work of four modules—water and fertilizer decoupling prediction, digital twin pre-simulation, conflict perception fusion, and execution compensation—water and fertilizer decision-making can be comprehensively optimized from four dimensions: physical process decoupling, physiological response prediction, cognitive conflict handling, and execution deviation compensation, thereby achieving precise spatiotemporal coordination of water and fertilizer.

[0048] Compared with existing technologies, this deep learning-based precision decision-making system for integrated water and fertilizer management in greenhouse crops has the following advantages:

[0049] I. This invention constructs a water and fertilizer decoupling prediction module containing a dual-branch deep neural network to independently predict crop water requirements and fertilizer ion requirements, achieving precise decoupling of water and fertilizer at the demand level. By establishing an ion transport rate matrix based on soil cation exchange capacity, it quantifies the physical differences in the transport rates of water and different fertilizer ions in the soil, calculates the spatiotemporal matching degree of water and fertilizer, and dynamically adjusts the fertilization sequence. This enables water and fertilizer ions to reach the crop root zone synchronously, effectively avoiding deep leakage and ion antagonism problems caused by asynchronous water and fertilizer transport, and significantly improving water and fertilizer utilization efficiency and crop absorption efficiency.

[0050] Second, this invention constructs a lightweight digital twin of crops that integrates radiative transfer and photosynthetic models. This allows for pre-simulation of crop physiological responses before water and fertilizer decisions are implemented. Multi-objective optimization of the decision-making scheme is performed based on the net photosynthetic gain index, ensuring that the water and fertilizer scheme matches the crop's physiological growth needs and avoiding photosynthetic inhibition caused by improper water and fertilizer application. Furthermore, the invention achieves intelligent integration of data-driven and knowledge-driven decision-making through a conflict perception fusion module, and real-time correction of execution deviations through an execution compensation module. This constructs a complete closed loop for water and fertilizer decision-making, improving the reliability of the decision-making scheme and the accuracy of field execution, and adapting to the dynamic water and fertilizer management needs throughout the crop's entire growth period.

[0051] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0053] Figure 1 This is a diagram of the overall system architecture of the present invention;

[0054] Figure 2 This is a schematic diagram of the dual-branch deep neural network structure of the present invention;

[0055] Figure 3 This is a flowchart of the water and fertilizer decision-making timing matching and adjustment process of the present invention. Detailed Implementation

[0056] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0057] Example

[0058] This embodiment provides a deep learning-based precision decision-making system for integrated water and fertilizer management in greenhouse crops. Deployed on a greenhouse intelligent management platform, the platform is equipped with an edge computing terminal. It communicates with the greenhouse environmental sensor array, hyperspectral imaging equipment, and integrated water and fertilizer management equipment via wired and wireless dual-link communication. The communication protocols used are Modbus-RTU and MQTT to ensure real-time and reliable data transmission. This embodiment focuses on greenhouse overwintering tomato cultivation. The tomato growth cycle encompasses four core stages: seedling, flowering and fruit setting, peak fruiting, and harvest. The system operates throughout the entire tomato growth cycle, updating decisions every 30 minutes to ensure real-time matching of water and fertilizer decisions with crop growth dynamics and environmental changes.

[0059] like Figure 1 As shown in the figure, the overall architecture of the deep learning-based precision decision-making system for greenhouse crops with integrated water and fertilizer is as follows. The overall system architecture is divided into four core functional modules, which are, from top to bottom, the water and fertilizer decoupling prediction module, the digital twin pre-simulation module, the conflict perception and fusion module, and the execution compensation module. The four modules are connected in sequence to form a complete closed loop of water and fertilizer decision-making and execution.

[0060] In this embodiment, the water and fertilizer decoupling prediction module is implemented as follows:

[0061] like Figure 2 As shown, the water and fertilizer decoupling prediction module is the core decision-making unit of the system. Its core function is to independently predict crop water requirements and fertilizer ion requirements, as well as optimize the temporal matching of water and fertilizer spatiotemporal transport, and output a preliminary water and fertilizer decision scheme. Specifically, the water and fertilizer decoupling prediction module includes four core components: a dual-branch deep neural network, an ion transport rate matrix calculation unit, a spatiotemporal matching degree calculation unit, and a fertilization timing adjustment unit.

[0062] Specifically, the water demand prediction subnetwork adopts a hybrid network structure combining a one-dimensional convolutional neural network and a gated recurrent unit (GRU). The network has eight layers: two one-dimensional convolutional layers, two pooling layers, two GRU layers, one fully connected layer, and one output layer. The network's input layer receives four types of input data: crop transpiration rate, root zone matrix potential, leaf area index (LAI), and canopy temperature. Crop transpiration rate is calculated using the Penman Montes formula, with the calculation based on data from greenhouse environmental sensors: air temperature, relative humidity, wind speed, and photosynthetically active radiation (PAIR) data collected every minute. A 10-minute moving average preprocessing step is performed before inputting the data into the network. Root zone matrix potential is collected using tensiometer sensors buried 15 cm deep in the crop root zone, with one sensor deployed per crop. The arithmetic mean of sensor data within the same planting row is used as the input value. Leaf area index is obtained by inverting canopy images collected by a multispectral camera mounted on the greenhouse roof, with the inversion frequency being once daily at 12:00. Canopy temperature is collected using an infrared thermometer, which is positioned directly in the upper part of the crop canopy. The temperature is collected once per minute and pre-processed by a 10-minute moving average before being input into the network.

[0063] The output layer of the water demand prediction subnetwork outputs the water demand in cubic meters per hectare. The network is trained using supervised learning. The training dataset consists of 120,000 samples from three consecutive years of tomato cultivation data in the target greenhouse. The training set to validation set ratio is 8:2. The Adam optimizer is used, with an initial learning rate of 0.001 and a mean squared error loss function. The training iterations are set to 200. Training is terminated early when the validation set loss does not decrease for 10 consecutive iterations. The final network prediction accuracy is controlled to be below 5% average absolute percentage error.

[0064] Specifically, the fertilizer demand prediction subnetwork adopts a hybrid network structure derived from the water demand prediction subnetwork, also set to 8 layers, including 2 one-dimensional convolutional layers, 2 pooling layers, 2 gated recurrent unit layers, 1 fully connected layer, and 1 output layer. The network weights are trained and updated independently. The network's input layer receives four types of input data: photosynthetically active radiation (PARF), carbon dioxide assimilation rate, crop growth stage, and rhizosphere ion concentration. PARF is collected by PARF sensors deployed above the canopy at a frequency of once per minute, and preprocessed with a 10-minute moving average before being input into the network. Carbon dioxide assimilation rate is obtained through a combination of periodic calibration of a portable photosynthesis system and model inversion, with calibration occurring every 7 days. Routine inversion is based on PARF, canopy temperature, and air carbon dioxide concentration data. The crop growth stage is divided into four stages according to the tomato growth cycle: seedling stage, flowering and fruit setting stage, peak fruiting stage, and harvest stage. These stages are converted into numerical data using one-hot encoding and input into the network. Rhizosphere ion concentrations were collected using soil ion sensors buried in the root zone. The collected parameters included the concentrations of nitrate ions, phosphate ions, potassium ions, calcium ions, and magnesium ions. The collection frequency was once every 30 minutes, and the arithmetic mean of the sensor data within the same planting row was used as the input value.

[0065] The output layer of the fertilizer demand prediction subnetwork outputs the ionic demand for nitrogen, phosphorus, potassium, and trace elements, in kilograms per hectare, specifically including the individual element demand for nitrate, phosphate, potassium, calcium, and magnesium ions. The network training process also employs supervised learning. The training dataset consists of 120,000 samples from three consecutive years of historical nutrient management data for tomato cultivation in the target greenhouse. The training set to validation set ratio is 8:2. The Adam optimizer is used, with an initial learning rate of 0.0005 and a mean squared error loss function. The training iterations are set to 200. Training is terminated early when the validation set loss does not decrease for 10 consecutive iterations. The final network prediction accuracy is controlled to be below 6% average absolute percentage error.

[0066] Specifically, the ion transport rate matrix calculation unit establishes the ion transport rate matrix based on the soil cation exchange capacity. The expression for the ion transport rate matrix is ​​as follows:

[0067]

[0068] The matrix represents the ion transport rate. The row dimension of the matrix is ​​1, and the column dimension is determined by the number of water and fertilizer transport media involved in the decision. In this embodiment, the column dimension is 6, corresponding to water, nitrogen ions, phosphorus ions, potassium ions, calcium ions, and magnesium ions in sequence.

[0069] This indicates the rate of water transport in the soil, measured in centimeters per hour. The values ​​are determined by three core parameters: soil texture, soil porosity, and initial soil moisture content. Soil texture is determined according to the international soil texture classification standard; in this embodiment, the soil in the root zone of the target greenhouse is loam. Soil porosity is obtained through field measurement using the ring sampler method, with a measurement frequency of once per planting season. Initial soil moisture content is collected using a time-domain reflectometry sensor, with a collection frequency of once every 30 minutes. The calculation was performed using Darcy's law, and the effects of soil saturated hydraulic conductivity and matrix potential gradient were considered simultaneously during the calculation process to ensure that the accuracy of the water transport rate calculation was consistent with the actual field conditions.

[0070] The value represents the rate of nitrogen ion transport in the soil, expressed in centimeters per hour. In this embodiment, nitrogen ions are mainly in the form of nitrate ions. The values ​​are determined by four core parameters: ionic radius, ionic valence state, soil cation exchange capacity, and soil pH. The ionic radius of the nitrate ion is 0.179 nm, and its valence state is -1. Soil cation exchange capacity is determined in the field using the ammonium acetate method, with a measurement frequency of once per planting season. Soil pH is collected using a pH sensor buried in the root zone, with a sampling frequency of once every 30 minutes.

[0071] The value represents the rate of phosphorus ion transport in the soil, expressed in centimeters per hour. In this embodiment, phosphorus ions are mainly in the form of dihydrogen phosphate ions. The values ​​are determined by four core parameters: ionic radius, ionic valence state, soil cation exchange capacity, and soil pH. The ionic radius of the dihydrogen phosphate ion is 0.238 nanometers, and its valence state is -1. The values ​​of soil cation exchange capacity and soil pH are derived from the aforementioned parameters, ensuring parameter consistency in the calculation process.

[0072] This indicates the rate of potassium ion transport in the soil, expressed in centimeters per hour. The values ​​are determined by four core parameters: ionic radius, ionic valence state, soil cation exchange capacity, and soil pH. The ionic radius of potassium ions is 0.138 nanometers, and their valence state is +1. The values ​​of soil cation exchange capacity and soil pH are derived from the aforementioned parameters.

[0073] This indicates the rate of calcium ion transport in the soil, expressed in centimeters per hour. The values ​​are determined by four core parameters: ionic radius, ionic valence state, soil cation exchange capacity, and soil pH. The ionic radius of calcium ions is 0.100 nanometers, and the ionic valence state is +2. The values ​​of soil cation exchange capacity and soil pH are derived from the aforementioned parameters.

[0074] This indicates the rate of magnesium ion transport in the soil, expressed in centimeters per hour. The values ​​are determined by four core parameters: ionic radius, ionic valence state, soil cation exchange capacity, and soil pH. The ionic radius of magnesium ions is 0.072 nanometers, and the ionic valence state is +2. The values ​​of soil cation exchange capacity and soil pH are derived from the aforementioned parameters.

[0075] The ion transport rate matrix is ​​updated at the same frequency as the water and fertilizer decision-making frequency, which is once every 30 minutes, to ensure that the matrix values ​​can respond in real time to the dynamic changes in soil environmental parameters.

[0076] Specifically, the spatiotemporal matching degree calculation unit calculates the spatiotemporal matching degree between water and each ion based on the ion transport rate matrix. The expression for the spatiotemporal matching degree is:

[0077]

[0078] This represents the spatiotemporal matching degree between water and various ions; it is dimensionless and ranges from 0 to 1. The closer the value is to 1, the better the time synchronization of water and various ions to the crop root zone, and the higher the degree of time and space matching of water and fertilizer.

[0079] This indicates the time it takes for water to reach the crop root zone, expressed in hours. The value is determined by the rate of water transport. Based on the vertical distance calculated from the water and fertilizer application point to the crop root zone, in this embodiment, the vertical distance between the drip irrigation tape placement point and the root zone is 15 cm. The calculation method is to divide the vertical distance by the water transport rate. .

[0080] This indicates the time it takes for nitrogen ions to reach the crop root zone, expressed in hours. The value is determined by the nitrogen ion transport rate. The distance from the water and fertilizer application point to the crop root zone is calculated by dividing the vertical distance by the nitrogen ion transport rate. .

[0081] This indicates the time it takes for phosphorus ions to reach the crop root zone, expressed in hours. The value is determined by the phosphorus ion transport rate. The value is calculated by dividing the vertical distance from the water and fertilizer application point to the crop root zone by the phosphorus ion transport rate. .

[0082] This indicates the time it takes for potassium ions to reach the crop root zone, expressed in hours. The value is determined by the potassium ion transport rate. The value is calculated by dividing the vertical distance from the water and fertilizer application point to the crop root zone by the potassium ion transport rate. .

[0083] This indicates the time it takes for calcium ions to reach the crop root zone, expressed in hours. The value is determined by the calcium ion transport rate. The value is calculated by dividing the vertical distance from the water and fertilizer application point to the crop root zone by the calcium ion transport rate. .

[0084] This indicates the time it takes for magnesium ions to reach the crop root zone, expressed in hours. The value is determined by the magnesium ion transport rate. The value is calculated by dividing the vertical distance from the water and fertilizer application point to the crop root zone by the magnesium ion transport rate. .

[0085] This means taking the minimum value among all ratios, i.e., from... The smallest value among the ratios to the arrival times of each ion is selected as the spatiotemporal matching degree. The final value is determined to ensure that the matching degree between all ions and water meets the preset requirements.

[0086] like Figure 3 As shown, the core function of the fertilization timing adjustment unit is to adjust the fertilization timing of different ions when the spatiotemporal matching degree is lower than a preset matching degree threshold, until the spatiotemporal matching degree reaches the preset matching degree threshold. Specifically, in this embodiment, the preset matching degree threshold is set to 0.85, when the spatiotemporal matching degree... When the calculated value is less than 0.85, the fertilization timing adjustment unit initiates the timing adjustment process. The core rule of timing adjustment is to apply ions with a transport rate lower than that of water earlier, and ions with a transport rate higher than that of water later. The duration of advance or delay is determined by the difference between the time it takes for ions to reach the root zone and the time it takes for water to reach the root zone. Specifically, the advance application duration equals the ion arrival time minus the water arrival time, and the delay application duration equals the water arrival time minus the ion arrival time. After the timing adjustment is completed, the spatiotemporal matching degree calculation unit recalculates the adjusted spatiotemporal matching degree. If the adjusted spatiotemporal matching degree is greater than or equal to the preset matching degree threshold, the adjustment process stops. If the adjusted spatiotemporal matching degree is still less than the preset matching degree threshold, the timing adjustment process is repeated until the spatiotemporal matching degree reaches the preset requirement.

[0087] After the water and fertilizer decoupling prediction module completes the timing adjustment, it outputs a preliminary water and fertilizer decision scheme. The preliminary water and fertilizer decision scheme includes the total water volume of a single irrigation cycle, the total amount of each nutrient ion applied in a single cycle, the fertilization timing nodes of each nutrient ion, and the irrigation duration and irrigation flow parameters.

[0088] In this embodiment, the digital twin pre-simulation module is implemented as follows:

[0089] The core function of the digital twin pre-simulation module is to perform pre-simulation verification of the preliminary water and fertilizer decision-making scheme before its execution. Based on the simulation results of crop physiological responses, the module optimizes the decision-making scheme and outputs an optimized water and fertilizer decision-making scheme. The digital twin pre-simulation module consists of three core components: a lightweight crop digital twin, a simulation calculation unit, and a multi-objective optimization unit.

[0090] Specifically, the lightweight crop digital twin integrates a radiative transfer model and a photosynthetic model. In this embodiment, the radiative transfer model adopts the PROSAIL radiative transfer model, and the photosynthetic model adopts the Farquhar photosynthetic model. The model parameters of the lightweight crop digital twin are calibrated using field measurement data at a frequency of once every 15 days to ensure that the model output is consistent with the actual growth status of the crop.

[0091] The core function of the PROSAIL radiative transfer model is to output canopy reflectance and photosynthetically active radiation absorption ratio based on crop leaf and canopy parameters. The input parameters of the PROSAIL radiative transfer model include leaf structure parameters, chlorophyll content, and leaf tilt angle distribution data. Leaf structure parameters represent the internal tissue structure characteristics of the leaf and are dimensionless. In this example, the leaf structure parameters for tomato leaves range from 1.0 to 2.0, determined based on field sampling and measurement data. Chlorophyll content was obtained through acetone extraction and measured in the field, in micrograms per square centimeter, with a measurement frequency of once every 7 days. Leaf tilt angle distribution data was obtained by collecting canopy structure data using a three-dimensional laser scanning device, with a collection frequency of once every 15 days. The output parameters of the PROSAIL radiative transfer model include canopy reflectance and photosynthetically active radiation absorption ratio. Canopy reflectance is a dimensionless parameter with a value range of 0 to 1. The photosynthetically active radiation absorption ratio is a dimensionless parameter with a value range of 0 to 1, representing the proportion of incident photosynthetically active radiation absorbed by the canopy, providing the basic input for subsequent photosynthetic rate calculations.

[0092] The core function of the Farquhar photosynthetic model is to output the crop's photosynthetic rate based on crop biochemical and environmental parameters. The input parameters of the Farquhar photosynthetic model include ribulose-1,5-bisphosphate carboxylase activity, electron transport rate, carbon dioxide concentration, and light intensity data. Ribuluose-1,5-bisphosphate carboxylase activity was obtained through leaf enzyme activity assays, measured in micromoles per square meter per second, with a measurement frequency of once every 15 days. The electron transport rate was measured using a chlorophyll fluorometer, also in micromoles per square meter per second, with a measurement frequency of once every 7 days. Carbon dioxide concentration was collected by carbon dioxide sensors deployed inside the greenhouse, with a collection frequency of once per minute. The light intensity data is the photosynthetically active radiation data, which is from the same source as the input parameters of the aforementioned fertilizer demand prediction subnetwork. The output parameter of the Farquhar photosynthetic model is the crop's net photosynthetic rate, measured in micromoles per square meter per second, characterizing the crop's photosynthetic productivity.

[0093] Specifically, the simulation calculation unit inputs the preliminary water and fertilizer decision scheme into the lightweight crop digital twin and pre-simulates the photosynthetic rate and transpiration rate within a preset time period after the execution of the preliminary water and fertilizer decision scheme. In this embodiment, the preset time period is set to 24 hours, that is, simulating the dynamic changes of the crop's photosynthetic rate and transpiration rate within 24 hours after the execution of the preliminary water and fertilizer decision scheme, and outputting the time-weighted average of the photosynthetic rate and the time-weighted average of the transpiration rate within the simulation period, providing a data basis for the subsequent calculation of the net photosynthetic gain index.

[0094] Specifically, the simulation calculation unit calculates the net photosynthetic gain index based on the simulation output results. The expression for the net photosynthetic gain index is as follows:

[0095]

[0096] The value represents the net photosynthetic gain index, which is dimensionless. The larger the value, the better the photosynthetic gain effect of the initial water and fertilizer decision-making scheme on crops, while the lower the cost of transpiration water consumption and the better the overall performance of the scheme.

[0097] This indicates the crop photosynthetic rate before the implementation of the initial water and fertilizer decision-making plan, expressed in micromoles per square meter per second. The value is the time-weighted average of the measured net photosynthetic rate of crops in the 24 hours before the decision is implemented. The data comes from the real-time inversion results of the Farquhar photosynthetic model.

[0098] This represents the simulated predicted value of crop photosynthetic rate within a preset time period after the implementation of the initial water and fertilizer decision-making scheme, expressed in micromoles per square meter per second. The value is the time-weighted average of the net photosynthetic rate of the crop output by the lightweight crop digital twin within 24 hours after the decision is executed.

[0099] This indicates the crop transpiration rate before the implementation of the initial water and fertilizer decision-making plan, expressed in millimeters per hour. The value is the time-weighted average of the crop transpiration rate measured within 24 hours before the decision is implemented. The data comes from the real-time calculation results of the Penmanmont formula.

[0100] This represents the simulated predicted crop transpiration rate within a preset time period after the implementation of the initial water and fertilizer decision-making scheme, expressed in millimeters per hour. The value is the time-weighted average of the crop transpiration rate output by the lightweight crop digital twin, within 24 hours after the decision is executed.

[0101] The opportunity cost weighting coefficient for water loss through evaporation is dimensionless and ranges from 0 to 1. The value of is determined based on the water resource endowment and irrigation cost of the planting area. In this embodiment, The value of is set to 0.3, which means that in the decision optimization process, the weight of evapotranspiration water cost is set to 30% of the photosynthetic gain weight.

[0102] Specifically, the multi-objective optimization unit performs multi-objective optimization on the initial water and fertilizer decision scheme based on the photosynthetic net gain index. The objective function of the optimization is to maximize the photosynthetic net gain index. The optimization constraints include irrigation water volume not exceeding the maximum irrigation quota for the crop's growth stage, application of each nutrient ion not exceeding the maximum fertilization quota for the crop's growth stage, and the spatiotemporal matching degree of water and fertilizer not being lower than a preset matching degree threshold. The optimization algorithm uses a non-dominated sorting genetic algorithm with a population size of 50, 100 iterations, a crossover probability of 0.8, and a mutation probability of 0.05. After optimization, the optimized water and fertilizer decision scheme with the largest photosynthetic net gain index is output. Based on the initial water and fertilizer decision scheme, the optimized scheme refines the irrigation water volume, fertilization amount, and fertilization timing, ensuring that the scheme achieves water and fertilizer conservation while improving crop photosynthetic capacity.

[0103] In this embodiment, the conflict-aware fusion module is implemented as follows:

[0104] The core function of the conflict perception and fusion module is to identify and intelligently integrate conflicts between data-driven and knowledge-driven decision-making, avoiding decision biases caused by the failure of a single decision source, and outputting the final water and fertilizer decision scheme. The conflict perception and fusion module comprises five core components: an agronomic knowledge graph, a conflict coefficient calculation unit, a hyperspectral image analysis unit, a Bayesian inference unit, and a decision fusion unit.

[0105] Specifically, the agronomic knowledge graph is constructed based on authoritative agronomic knowledge in the field of greenhouse tomato cultivation. The entity types in the knowledge graph include crop varieties, growth stages, environmental parameter thresholds, water and fertilizer management indicators, and pest and disease control requirements. The relationship types between entities include hierarchical relationships, causal relationships, constraint relationships, and collaborative relationships. In this embodiment, the knowledge sources for the agronomic knowledge graph include Chinese agricultural industry standards, academic monographs on tomato cultivation, research results on greenhouse tomato water and fertilizer management published in core journals, and planting guidance programs issued by local agricultural technology extension departments, ensuring the authority and applicability of knowledge-driven decision-making. The knowledge-driven decision-making scheme is generated by the agronomic knowledge graph based on real-time collected data on crop growth stages, environmental parameters, and crop growth status, including recommended irrigation water volume, recommended fertilizer application rate, fertilization frequency, and water-fertilizer ratio parameters for the corresponding growth stage.

[0106] Specifically, the conflict coefficient calculation unit acquires both data-driven and knowledge-driven decision-making schemes and calculates the conflict coefficient between them. The data-driven decision-making scheme originates from the optimized water and fertilizer decision-making scheme output by the digital twin pre-simulation module, while the knowledge-driven decision-making scheme originates from the reasoning results of the agronomic knowledge graph. The expression for the conflict coefficient is:

[0107]

[0108] This represents the conflict coefficient between data-driven and knowledge-driven decision-making approaches. It is dimensionless and ranges from 0 to 1. The closer the value is to 1, the greater the difference between the two decision options and the higher the degree of conflict. The closer the value is to 0, the better the consistency between the two decision options and the lower the degree of conflict.

[0109] The feature vector represents the data-driven decision-making scheme. The vector has a dimension of 6 and corresponds to the amount of irrigation water, nitrogen ion application, phosphorus ion application, potassium ion application, calcium ion application, and magnesium ion application in a single round, respectively. All elements in the vector are normalized, with a normalization range of 0 to 1, to ensure the consistency of calculation for parameters with different dimensions.

[0110] The feature vector representing the knowledge-driven decision-making scheme, with the vector's dimensions and element types as... Completely consistent, corresponding sequentially to the recommended irrigation water volume, nitrogen ion application rate, phosphorus ion application rate, potassium ion application rate, calcium ion application rate, and magnesium ion application rate for a single round. Each element in the vector is also normalized to the range of 0 to 1.

[0111] The norm of a vector is used in this embodiment. The L2 norm, or Euclidean norm, is used to quantify the magnitude of a vector and the distance between vectors. This represents the Euclidean distance between the feature vectors of data-driven decision-making and knowledge-driven decision-making. The Euclidean modulus of the feature vector representing data-driven decision-making is given. The Euclidean modulus of the feature vector representing knowledge-driven decision-making is given.

[0112] Specifically, in this embodiment, the preset conflict threshold is set to 0.3, when the conflict coefficient... When the calculated value is less than or equal to 0.3, it is determined that the two decision schemes have no significant conflict, and the decision fusion unit directly performs weighted fusion of the two decision schemes in an equal weight manner. When the conflict coefficient is less than or equal to 0.3, the two decision schemes are considered to have no significant conflict, and the decision fusion unit directly performs weighted fusion of the two decision schemes in an equal weight manner. When the calculated value is greater than 0.3, it is determined that there is a significant conflict between the two decision schemes, and the system calls a third-party verification source to perform a Bayesian confidence assessment. In this embodiment, the third-party verification source is the actual nutrient element content data of the crop, and the data comes from the real-time analysis results of the hyperspectral image analysis unit.

[0113] Specifically, the hyperspectral image analysis unit acquires real-time hyperspectral image data of the crop canopy using hyperspectral imaging equipment installed inside the greenhouse. Acquisition occurs once daily between 11:00 AM and 1:00 PM. The hyperspectral images cover a spectral range of 400 nm to 1000 nm, with a spectral resolution of 3 nm and a spatial resolution of 1 cm. After preprocessing the acquired image data through radiometric calibration, atmospheric correction, and geometric correction, the hyperspectral image analysis unit analyzes the actual content of nitrogen, phosphorus, potassium, calcium, and magnesium in crop leaves based on a pre-trained nutrient element inversion model, expressed in milligrams per gram of dry weight, providing observational evidence for Bayesian inference.

[0114] Specifically, the Bayesian inference unit uses the actual nutrient element content of crops as observational evidence to calculate the first posterior confidence of the data-driven decision-making scheme and the second posterior confidence of the knowledge-driven decision-making scheme. The core process of Bayesian inference is as follows: First, the prior confidence of the data-driven and knowledge-driven decision-making schemes is set. In this embodiment, the prior confidence of both schemes is set to 0.5, representing that the initial reliability of the two schemes is consistent in the absence of observational evidence. Then, a likelihood function is constructed based on historical data. The likelihood function represents the probability of observing the current nutrient element content of the crops under the condition that the decision scheme is valid. Finally, the posterior confidence of the two schemes is calculated based on Bayes' theorem. The posterior confidence ranges from 0 to 1; the larger the value, the higher the reliability of the corresponding decision scheme.

[0115] Specifically, the decision fusion unit uses the first and second posterior confidence levels as weights for weighted fusion to perform weighted fusion of the data-driven and knowledge-driven decision schemes, generating the final water and fertilizer decision scheme. The weighted fusion calculation method is as follows: each parameter value of the final water and fertilizer decision scheme equals the corresponding parameter value of the data-driven decision scheme multiplied by the first posterior confidence level, plus the corresponding parameter value of the knowledge-driven decision scheme multiplied by the second posterior confidence level. The final water and fertilizer decision scheme includes parameters such as the total irrigation water volume in a single cycle, the total application amount of each nutrient ion in a single cycle, the fertilization timing nodes of each nutrient ion, irrigation duration, irrigation flow rate, and fertilizer pump operating frequency, providing a basis for subsequent compensation execution.

[0116] In this embodiment, the compensation module is implemented as follows:

[0117] The core function of the execution compensation module is to compensate for execution deviations in the final water and fertilizer decision-making scheme, ensuring that the actual output of the water and fertilizer execution equipment is completely consistent with the decision requirements, and generating execution instructions to be sent to the execution equipment. The execution compensation module consists of three core components: an actuator digital twin unit, an execution efficiency calculation unit, and an instruction generation unit.

[0118] Specifically, the actuator digital twin unit interacts with the integrated water and fertilizer management equipment in real time, collecting pump motor current data, pipeline pressure fluctuation data, and dripper flow data. Pump motor current data is collected via a built-in current transformer in the motor, at a frequency of once per second. Pipeline pressure fluctuation data is collected via pressure sensors installed in the main and branch irrigation pipelines, at a frequency of once per second. Dripper flow data is collected via flow sensors at the inlet of each drip irrigation tape row, at a frequency of once per second. Based on the collected real-time data, the actuator digital twin unit constructs a dynamic simulation model of the execution system, identifies the dynamic characteristics of the execution system online, and provides a data foundation for calculating execution efficiency.

[0119] Specifically, the execution efficiency calculation unit identifies the execution efficiency coefficient of the current execution system online based on the output data of the actuator digital twin unit. The expression for the execution efficiency coefficient is:

[0120]

[0121] This represents the execution efficiency coefficient of the execution system; it is dimensionless and ranges from 0 to 1. The closer the value is to 1, the better the consistency between the actual output of the execution system and the theoretical requirements, and the higher the execution efficiency. When the value is less than 1, it indicates that there is a loss of traffic in the execution system, which needs to be compensated and corrected.

[0122] This indicates the actual execution flow of the system, expressed in cubic meters per hour. The value is the sum of the real-time flow rates collected by all drip irrigation tape flow sensors, and the 10-second moving average is taken as the final calculated value.

[0123] This represents the theoretical decision flow rate of the execution system, expressed in cubic meters per hour. The value of is derived from the irrigation flow parameters set in the final water and fertilizer decision scheme, that is, the theoretical flow value that the execution system should output according to the decision requirements.

[0124] Specifically, the instruction generation unit compensates and corrects the final water and fertilizer decision scheme based on the execution efficiency coefficient. The calculation method for the compensation correction is as follows: the corrected execution flow rate value equals the theoretical decision flow rate value in the final water and fertilizer decision scheme divided by the execution efficiency coefficient; the corrected irrigation duration equals the theoretical irrigation duration in the final water and fertilizer decision scheme multiplied by the execution efficiency coefficient, ensuring that the corrected total irrigation water volume and total fertilizer application amount are completely consistent with the requirements of the final water and fertilizer decision scheme. After the compensation correction is completed, the instruction generation unit converts the corrected control parameters into execution instructions in the standard Modbus-RTU protocol and sends them to the integrated water and fertilizer execution equipment, including the irrigation pump station, solenoid valve, fertilizer pump, and flow regulating valve, through a wired communication link, controlling the execution equipment to complete the water and fertilizer application operation according to the decision requirements.

[0125] In this embodiment, the complete operation of the system is implemented as follows:

[0126] This embodiment uses a water and fertilizer decision-making process during the peak fruiting period of greenhouse tomatoes as an example to fully illustrate the system's operation. The environmental parameters during the peak fruiting period of tomatoes are: air temperature 28 degrees Celsius, relative humidity 60%, photosynthetically active radiation 800 micromoles per square meter per second, root zone matrix potential -15 kPa, initial soil moisture content 22%, soil pH 6.8, and soil cation exchange capacity 15 centimoles per kilogram.

[0127] First, the dual-branch deep neural network of the water and fertilizer decoupling prediction module performs demand prediction. The water demand prediction sub-network outputs a water demand of 25 cubic meters per hectare, and the fertilizer demand prediction sub-network outputs ion demand of 30 kg per hectare for nitrogen ions, 10 kg per hectare for phosphorus ions, 35 kg per hectare for potassium ions, 5 kg per hectare for calcium ions, and 3 kg per hectare for magnesium ions.

[0128] Then, the ion transport rate matrix calculation unit calculates the ion transport rate matrix. It is 1.2 centimeters per hour. It is 1.0 cm per hour. It is 0.3 cm per hour. It is 0.8 cm per hour. It is 0.6 cm per hour. It is 0.7 centimeters per hour.

[0129] The spatiotemporal matching degree calculation unit calculates the arrival time of each ion in the root region. It takes 1.25 hours. It takes 1.5 hours. It lasts for 5 hours. It takes 1.875 hours. It takes 2.5 hours. The time is 2.14 hours. The spatiotemporal matching degree is calculated. The value is 0.25, which is lower than the preset matching threshold of 0.85, so the fertilization timing adjustment is initiated.

[0130] The fertilization timing adjustment unit adjusts the fertilization timing for each ion. Phosphorus ions, with the lowest transport rate, are applied 3.75 hours earlier. Calcium ions are applied 1.25 hours earlier. Magnesium ions are applied 0.89 hours earlier. Potassium ions are applied 0.625 hours earlier. Nitrogen ions are applied 0.25 hours earlier. The spatiotemporal matching degree is recalculated after the adjustment. The value is 0.92, reaching the preset matching threshold, and a preliminary water and fertilizer decision scheme is output.

[0131] Next, the digital twin pre-simulation module inputs the preliminary water and fertilizer decision-making scheme into the lightweight crop digital twin to pre-simulate the crop's physiological response over 24 hours. The simulation yields... It is 22 micromoles per square meter per second. 28 micromoles per square meter per second It is 0.8 millimeters per hour. It is 0.9 millimeters per hour. Taking a value of 0.3, the calculation yields... The value is 0.9355. The multi-objective optimization unit is based on... After optimization, the irrigation water volume is adjusted to 24 cubic meters per hectare, the nitrogen ion application rate to 29 kg per hectare, the phosphorus ion application rate to 10 kg per hectare, the potassium ion application rate to 34 kg per hectare, the calcium ion application rate to 5 kg per hectare, and the magnesium ion application rate to 3 kg per hectare. The value is 0.95, and the optimized water and fertilizer decision scheme is output.

[0132] Then, the conflict perception fusion module obtains the optimized water and fertilizer decision scheme as a data-driven decision scheme, and the agronomic knowledge graph reasoning obtains the knowledge-driven decision scheme, with irrigation water volume of 26 cubic meters per hectare, nitrogen ion of 32 kg per hectare, phosphorus ion of 11 kg per hectare, potassium ion of 36 kg per hectare, calcium ion of 5 kg per hectare, and magnesium ion of 3 kg per hectare. The conflict coefficient is calculated. The value is 0.22, which is lower than the preset conflict threshold of 0.3. Equal weight fusion is adopted to obtain the final water and fertilizer decision scheme: irrigation water volume of 25 cubic meters per hectare, nitrogen ion of 30.5 kg per hectare, phosphorus ion of 10.5 kg per hectare, potassium ion of 35 kg per hectare, calcium ion of 5 kg per hectare, and magnesium ion of 3 kg per hectare.

[0133] Finally, the compensation module collects data from the execution device and calculates the execution efficiency coefficient. The value is set to 0.95. The final water and fertilizer decision scheme is compensated and corrected. The corrected irrigation flow rate is the theoretical flow rate divided by 0.95, and the irrigation duration is the theoretical duration multiplied by 0.95. An execution command is generated and sent to the execution equipment to complete this water and fertilizer decision and application operation.

[0134] This embodiment achieves decoupled prediction of water and fertilizer demand, spatiotemporal matching optimization, physiological response pre-simulation, decision conflict fusion and execution deviation compensation through the coordinated operation of four modules, forming a complete closed loop for precise water and fertilizer decision-making. This can effectively improve water and fertilizer utilization efficiency and adapt to the growth needs of greenhouse crops throughout their entire growth period.

[0135] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A precision decision-making system for integrated water and fertilizer management in greenhouse crops based on deep learning, characterized in that, The system consists of: The water and fertilizer decoupling prediction module includes a dual-branch deep neural network. The first branch predicts water demand based on crop transpiration rate and rhizosphere matrix potential, while the second branch predicts fertilizer ion demand based on crop photosynthetic product distribution and rhizosphere ion concentration. The water and fertilizer decoupling prediction module also establishes an ion transport rate matrix based on soil cation exchange capacity, calculates the spatiotemporal matching degree between water and each ion according to the ion transport rate matrix, and adjusts the fertilization sequence when the spatiotemporal matching degree is lower than a preset matching degree threshold until the spatiotemporal matching degree reaches the preset matching degree threshold, and then outputs a preliminary water and fertilizer decision scheme. The digital twin pre-simulation module includes a lightweight crop digital twin, which integrates a radiative transfer model and a photosynthetic model. The digital twin pre-simulation module inputs the preliminary water and fertilizer decision scheme into the lightweight crop digital twin, pre-simulates the photosynthetic rate and transpiration rate within a preset time period after the preliminary water and fertilizer decision scheme is executed, calculates the net photosynthetic gain index, and performs multi-objective optimization of the preliminary water and fertilizer decision scheme based on the net photosynthetic gain index before outputting an optimized water and fertilizer decision scheme. The conflict perception and fusion module includes an agronomic knowledge graph. It acquires a data-driven decision-making scheme and a knowledge-driven decision-making scheme. The data-driven decision-making scheme originates from the optimized water and fertilizer decision-making scheme, and the knowledge-driven decision-making scheme originates from the agronomic knowledge graph. The conflict perception and fusion module calculates the conflict coefficient between the data-driven and knowledge-driven decision-making schemes. When the conflict coefficient exceeds a preset conflict threshold, it calls a third-party verification source to perform Bayesian confidence assessment. Based on the posterior confidence, it performs a weighted fusion of the data-driven and knowledge-driven decision-making schemes and outputs the final water and fertilizer decision-making scheme. The execution compensation module monitors the status of the execution equipment and calculates the execution efficiency coefficient. Based on the execution efficiency coefficient, it compensates and corrects the final water and fertilizer decision scheme and generates an execution command to be sent to the execution equipment.

2. The precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The dual-branch deep neural network includes a water demand prediction subnetwork and a fertilizer demand prediction subnetwork; The input layer of the water demand prediction subnetwork receives crop transpiration rate, root zone matrix potential, leaf area index and canopy temperature data, and the output layer outputs water demand. The fertilizer demand prediction subnetwork's input layer receives data on photosynthetically active radiation, carbon dioxide assimilation rate, crop growth stage, and rhizosphere ion concentration, while its output layer outputs the ionic demand for nitrogen, phosphorus, potassium, and trace elements.

3. The precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The ion transport rate matrix is ​​represented as follows: in, This represents the ion transport rate matrix, composed of water transport rate and the transport rates of each ion. Each ion rate is determined based on the soil cation exchange capacity. This indicates the rate of water transport in the soil. It is determined by soil texture, soil porosity, and initial soil moisture content; , , , These represent the transport rates of nitrogen, phosphorus, potassium, and calcium ions in the soil, respectively. The transport rate of each ion is determined by its ionic radius, ionic valence state, soil cation exchange capacity, and soil pH.

4. The precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The spatiotemporal matching degree is expressed as: in, This indicates the spatiotemporal matching degree between water and various ions. Indicates the time it takes for water to reach the crop root zone. , , , These represent the time it takes for each ion to reach the crop root zone. This indicates taking the minimum value among all ratios.

5. The precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The water and fertilizer decoupling prediction module also includes a fertilization timing adjustment unit. When the spatiotemporal matching degree is lower than the preset matching degree threshold, the fertilization timing adjustment unit applies ions with a migration rate lower than that of water in advance and applies ions with a migration rate higher than that of water in a later manner, and recalculates the adjusted spatiotemporal matching degree until the spatiotemporal matching degree reaches the preset matching degree threshold.

6. The precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The lightweight crop digital twin includes the PROSAIL radiative transfer model and the Farquhar photosynthesis model; The PROSAIL radiative transfer model outputs canopy reflectance and photosynthetically active radiation absorption ratio based on leaf structure parameters, chlorophyll content, and leaf tilt angle distribution data. The Farquhar photosynthetic model outputs the photosynthetic rate based on ribulose-1,5-bisphosphate carboxylase activity, electron transport rate, carbon dioxide concentration, and light intensity data.

7. The precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The net photosynthetic gain index is expressed as: in, Indicates the net photosynthetic gain index. This represents the crop photosynthetic rate before implementing the preliminary water and fertilizer decision-making scheme. This represents the simulated predicted value of crop photosynthetic rate within a preset time period after implementing the preliminary water and fertilizer decision-making scheme. This represents the crop transpiration rate before implementing the preliminary water and fertilizer decision-making scheme. This represents the simulated predicted value of crop transpiration rate within a preset time period after implementing the preliminary water and fertilizer decision-making scheme. This represents the opportunity cost weighting coefficient for water consumed through evaporation.

8. The precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The conflict coefficient is expressed as: in, This represents the conflict coefficient between data-driven and knowledge-driven decision-making approaches. The feature vector representing the data-driven decision-making scheme. The feature vector representing the knowledge-driven decision-making scheme. The norm of a vector.

9. The precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The conflict perception fusion module includes a hyperspectral image analysis unit and a Bayesian inference unit. The hyperspectral image analysis unit acquires crop hyperspectral image data in real time and analyzes the actual nutrient element content of the crop. The Bayesian inference unit uses the actual nutrient element content of the crop as observational evidence to calculate the first posterior confidence of the data-driven decision-making scheme and the second posterior confidence of the knowledge-driven decision-making scheme, and uses the first posterior confidence and the second posterior confidence as the weights for weighted fusion.

10. A precision decision-making system for integrated water and fertilizer management of greenhouse crops based on deep learning according to claim 1, characterized in that, The execution compensation module includes an actuator digital twin unit; The actuator digital twin unit analyzes the pump room motor current data, pipeline pressure fluctuation data, and dripper flow data in real time, and identifies the execution efficiency coefficient of the current execution system online. The execution efficiency coefficient is the ratio of the actual execution flow to the theoretical decision flow. The execution compensation module divides the final water and fertilizer decision scheme by the execution efficiency coefficient to generate an execution instruction.

Images