Agricultural intelligent irrigation scheduling method and system based on deep learning

Abstract

The invention discloses an agricultural intelligent irrigation scheduling method and system based on deep learning, and relates to the technical field of intelligent agriculture, and the method comprises the steps: collecting field crop data in unit time, the field crop data comprising soil parameters, meteorological parameters and crop parameters; constructing an irrigation decision model, and analyzing a field water and fertilizer loss value according to the field crop data; setting a water and fertilizer dynamic coupling mechanism for irrigation scheduling, and adopting a dynamic compensation strategy for the field water and fertilizer loss value; based on the growth trend of field crops, constructing a farmland digital twinborn model, simulating the dynamic compensation strategy and generating a feature contribution value; and visualizing the feature contribution value to generate a farmland decision reason analysis report, thereby effectively reducing the pollution risk of agricultural activities to underground water and drainage basin water bodies, realizing an accurate irrigation strategy of agricultural soil, and avoiding resource waste.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of smart agriculture technology, and in particular to an agricultural intelligent irrigation scheduling method and system based on deep learning. BACKGROUND

[0002] Intelligent irrigation and fertilization is a key technology for high-quality agricultural development, and is therefore being promoted on a large scale by the country. With the intelligentization of agriculture in China, traditional manual irrigation and fertilization methods cannot meet the actual needs, so it is necessary to achieve intelligent management of water and fertilizer integrated irrigation and fertilization.

[0003] At present, the Chinese invention patent with application number 202411136773.2 discloses an individualized precision irrigation method based on the Internet of Things and deep learning. The invention specifically includes: collecting and preprocessing data of crops; constructing a crop growth prediction model, learning the crop growth pattern through historical data, updating the model parameters, dynamically adjusting the irrigation strategy through the irrigation strategy optimization model, optimizing water resource allocation, and performing online learning and model updating, then making intelligent decisions and execution, combining the crop growth prediction and irrigation strategy optimization model to generate irrigation decisions, accurately controlling the irrigation time and water quantity through intelligent valves, specifically including irrigation decision generation methods and intelligent irrigation execution methods, and then collecting feedback and iterating strategies, continuously collecting crop growth conditions and soil moisture data through feedback data collection and evaluation methods and strategy iteration and optimization methods, and evaluating the actual effect of the irrigation strategy using preset evaluation indicators; the model is continuously optimized and dynamically deployed to enable real-time monitoring.

[0004] The above technology does not integrate physical mechanisms and digital twin simulation models, fails to mine the influence contribution of various parameters of soil and crops, and lacks quantitative decision basis for water and fertilizer irrigation. SUMMARY

[0005] The technical problem solved by the present application is that physical mechanisms and digital twin simulation models are not integrated, the influence contribution of various parameters of soil and crops is not mined, and there is a lack of quantitative decision basis for water and fertilizer irrigation.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, an agricultural intelligent irrigation scheduling method based on deep learning includes the following steps: Step S1, collecting crop data in a unit of time in a field, the crop data in the field including soil parameters, meteorological parameters and crop parameters; Step S2, constructing an irrigation decision model, and analyzing water and fertilizer loss values in the field according to the crop data in the field; Step S3, set up the water and fertilizer dynamic coupling mechanism of irrigation scheduling, and take a dynamic compensation strategy for the field water and fertilizer loss value; Step S4, based on the crop growth trend in the field, build a digital twin model of farmland, simulate the dynamic compensation strategy and generate a feature contribution value; Step S5, visualize the feature contribution value to generate a farmland decision reason analysis report.

[0007] Preferably, the step S1 includes the following sub-steps: Step S11, deploy soil parameter sensors, meteorological parameter sensors and crop physiological monitors in the field, and collect soil moisture, conductivity, pH value, soil nutrient content, soil depth and soil temperature in unit time as soil parameters; Step S12, collect air temperature, air pressure, humidity, light radiation, wind speed and rainfall in unit time as meteorological parameters; Step S13, collect crop growth status, crop composition, root water content and crop surface temperature in unit time as crop parameters; Step S14, save the soil parameters, meteorological parameters and crop parameters as field crop data using an edge gateway and upload them to the system cloud.

[0008] Preferably, the step S2 includes the following sub-steps: Step S21, synchronize the time stamps of the soil parameters, meteorological parameters and crop parameters of the field crop data, generate crop sample set data, calculate the mutual information value of a single feature in the crop sample set data with the conductivity and root water content respectively using the mutual information criterion, if the mutual information value of the current feature is less than the set correlation information threshold, discard the current feature, if the mutual information value of the current feature is greater than the set correlation information threshold, retain the current feature, and generate first relationship data according to the time stamp sorting; Step S22, calculate the soil heat flux in unit time according to the first relationship data, and the calculation expression of the soil heat flux is: ; Wherein, is the soil heat flux, is the soil heat capacity, is the air temperature at the i-th moment, is the air temperature at the i-1-th moment, is the time step, is the soil depth; Step S23, calculate the soil water loss in unit time according to the first relationship data and the soil heat flux, and the calculation expression of the soil water loss is: ; ; wherein, is the soil water loss amount, is the crop coefficient, is the reference soil water loss amount, is the net solar radiation on the crop surface, is the soil heat flux, is the soil temperature, is the daily average air temperature, is the saturated water vapor pressure, is the actual water vapor pressure, is the average air temperature saturated water vapor pressure and temperature curve slope, is the wind speed; The soil water loss amount calculated in a unit of time is sorted according to time in chronological order, and a soil water loss curve graph is drawn.

[0009] Preferably, the step S2 further comprises: The proportion of the soil nutrient content in the total soil content, the crop soil nutrient background concentration C of the first relationship data are counted, the soil enrichment factor EF is obtained by comparing the crop component of the crop parameter with the corresponding nutrient component in the soil, and the soil loss amount A is obtained by using the general soil loss equation, so as to improve the calculation of the soil nutrient loss value. The calculation expression of the improved calculation of the soil nutrient loss value is: ; The soil nutrient loss value in a unit of time is sorted according to time in chronological order, and a soil nutrient loss curve graph is drawn, and the soil water loss amount is one-to-one corresponding according to the time stamp, to obtain second relationship data; An irrigation decision-making model is constructed, a mechanism model based on HYDRUS is established, the second relationship data is time step verified, and the meteorological parameters and crop parameters are taken as input boundary conditions to predict the next soil nutrient loss value and soil water loss amount as the field water and fertilizer loss value.

[0010] Preferably, the step S3 specifically comprises: A water-fertilizer coupling rule of irrigation scheduling is set, the trigger condition of the water-fertilizer coupling rule includes a first condition and a second condition, when the field water and fertilizer loss value is in the state of the first condition, a first irrigation strategy is triggered, when the field water and fertilizer loss value is in the state of the second condition, a second irrigation strategy is triggered, and the first irrigation strategy and the second irrigation strategy are fused to dynamically compensate for the missing components of the soil until the field crop data reaches a balanced state. The first condition is used to determine the size of the soil water loss amount and the target soil water loss threshold value, and when the soil water loss amount is greater than the target soil water loss threshold value, the first condition is executed; The first condition is used to determine the size of the soil nutrient loss value and the target soil nutrient loss threshold value, and when the soil water loss value is greater than the target soil nutrient loss threshold value, the second condition is executed; The first irrigation strategy is used to adjust the valve and start the soil water replenishment irrigation strategy; The second irrigation strategy is used to adjust the valve and expect the soil nutrient irrigation strategy.

[0011] Preferably, the step S4 includes the following sub-steps: Step S41, the cloud field crop data is constructed in HYDRUS-2D 3D farmland all-around model, and the upper boundary, lower boundary and solute boundary are set, the upper boundary includes the atmospheric boundary, the meteorological parameter is input to the upper boundary, and the condition of triggering the first irrigation strategy of step S3 is superimposed, the lower boundary is set as variable pressure drainage function, and the solute boundary is used to superimpose the second irrigation strategy; Step S42, for the prediction time point of the field water and fertilizer loss value, a simulation irrigation compensation strategy is adopted at the prediction time point, and is valued in the 3D farmland all-around model; Step S43, the soil wetting peak, depth and uniformity after simulation irrigation are counted, the optimal interval crop soil water and fertilizer simulation value of the current crop parameter is analyzed according to the water content and nutrient concentration of crop root system, the 3D farmland all-around model is quantified by using four-dimensional evaluation system, and the fitting goodness index decision coefficient is determined according to the soil water and fertilizer simulation value and the measured value of soil water and fertilizer When The closer the value is to 1, the stronger the linear relationship between the soil water and fertilizer simulation value and the measured value of soil water and fertilizer, the irrigation simulation value is repeatedly adjusted until The result of the value is close to 1, the farmland numerical twin model is generated, and the characteristic contribution value is obtained.

[0012] Preferably, the first irrigation strategy and the second irrigation strategy specifically include: When the soil water loss amount is greater than the target soil water loss threshold value, the first irrigation strategy is adopted, and positive irrigation is adopted for the current soil according to the soil water loss amount, and the irrigation water amount is equal to the soil water loss amount; When the soil nutrient loss value is greater than the target soil nutrient loss threshold value, the second irrigation strategy is adopted, and positive irrigation is adopted for the current soil according to the soil nutrient loss amount, and the irrigation nutrient is equal to the soil nutrient loss amount.

[0013] Preferably, the method for obtaining the characteristic contribution value comprises: Adopting global sensitivity analysis method, quantifying the contribution degree of each feature of the irrigation simulation value to the measured value of soil water and fertilizer, taking soil nutrient content and water content as target scalar, taking each feature of the irrigation simulation value as input feature, defining feature distribution threshold, sampling the input features using Sobol sequence, generating multiple groups of parameter combinations, calling simulation irrigation compensation strategy for each group of parameters, outputting target variable, calculating the variance of the target variable and the sensitivity index of each input feature, the sensitivity index as the feature contribution value of each feature.

[0014] Preferably, step S5 specifically comprises: Using a bee colony diagram, the feature contribution value is generated into a visual large screen, the feature contribution value corresponding to each feature point as a data point, according to the data point distribution feature, analyzing the farmland decision-making reason, when the feature points are mainly distributed on the right side, indicating that the irrigation amount and nutrient loss present positive influence, when the feature points are distributed on both sides, indicating that the irrigation amount and nutrient loss present negative influence, when the feature points are distributed dispersedly, the irrigation amount and nutrient loss have smaller mutual influence.

[0015] In a second aspect, an agricultural intelligent irrigation scheduling system based on deep learning includes a data acquisition module, a data analysis module, a strategy compensation module, a digital simulation module, and a visualization analysis module. The data acquisition module is configured to acquire crop data in a unit of time and comprehensively obtain real-time state information of the farmland. The data analysis module is configured to construct an irrigation decision-making model and quantify the water and fertilizer loss value in the field according to the crop data. The strategy compensation module is configured to develop a precise irrigation and fertilization scheme, set a water and fertilizer dynamic coupling mechanism for irrigation scheduling, and adopt a first irrigation strategy and a second irrigation strategy for the water and fertilizer loss value in the field. The digital simulation module is configured to construct a digital twin model of the farmland based on the growth trend of the crops in the field, simulate the dynamic compensation strategy, and calculate the feature contribution value. The visualization analysis module is configured to visually display the influence of each decision-making feature using a chart and generate a farmland decision-making reason analysis report by visualizing the feature contribution value.

[0016] The application has the beneficial effects that: by deploying sensors, the state information of the farmland is comprehensively acquired to provide a data basis for decision-making, an irrigation decision-making model based on deep learning and physical mechanism is constructed for the collected farmland data, the farmland data is comprehensively analyzed, the current field water and fertilizer loss value is calculated, the loss trend in a future period of time is predicted, the linkage rules between irrigation and fertilization are established to ensure that the water and fertilizer amounts that need to be supplemented are calculated synchronously and optimized, then the dynamic compensation strategy is simulated in the digital simulation model, the soil wetting state is visualized, the advantages and disadvantages of different strategies are compared through hypothesis analysis to ensure the scientificity and reliability of the decision-making, and finally the visual analysis report is established, the application effectively reduces the pollution risk of agricultural activities on groundwater and watershed water, and realizes the coordinated development of agricultural production and environmental protection. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A step flow chart of an agricultural intelligent irrigation scheduling method based on deep learning provided by an embodiment of the application is shown in the figure. Figure 2 A basic flow schematic diagram of an agricultural intelligent irrigation scheduling system based on deep learning provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0018] In order to make the above-mentioned objects, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments.

[0019] Embodiment 1, with reference to Figure 1 provides an agricultural intelligent irrigation scheduling method based on deep learning, which comprises the following steps. Step S1, collecting field crop data in a unit time, the field crop data comprising soil parameters, meteorological parameters and crop parameters; Step S2, constructing an irrigation decision-making model, and analyzing the field water and fertilizer loss value according to the field crop data; Step S3, setting a water and fertilizer dynamic coupling mechanism of irrigation scheduling, and taking a dynamic compensation strategy for the field water and fertilizer loss value; Step S4, constructing a farmland digital twin model based on the field crop growth trend, simulating the dynamic compensation strategy and generating a feature contribution value; Step S5, visualizing the feature contribution value to generate a farmland decision-making reason analysis report.

[0020] In this embodiment, the state information of the farmland is comprehensively acquired by deploying sensors, providing a data basis for decision-making, and constructing an irrigation decision-making model based on deep learning and physical mechanism for the collected farmland data. The farmland data is comprehensively analyzed, the current field water and fertilizer loss value is calculated, the loss trend in the future period is predicted, the linkage rules between irrigation and fertilization are established to ensure the synchronous optimization of the water and fertilizer amount that needs to be supplemented, and then the dynamic compensation strategy is simulated in the digital simulation model. The soil wetting state is visualized, the advantages and disadvantages of different strategies are compared through hypothesis analysis, the scientificity and reliability of the decision-making are ensured, and finally the visual analysis report is established. The characteristic contribution value and simulation results are visualized to generate a report, and the characteristic influence value of each decision is intuitively displayed.

[0021] Step S1 includes the following sub-steps: Step S11, deploying soil parameter sensors, meteorological parameter sensors and crop physiological monitors in the field to collect soil moisture, electrical conductivity, pH value, soil nutrient content, soil depth and soil temperature per unit time as soil parameters; Step S12, collecting air temperature, air pressure, humidity, light radiation, wind speed and rainfall per unit time as meteorological parameters; Step S13, collecting crop growth state, crop composition, root water content and crop surface temperature per unit time as crop parameters; Step S14, saving the soil parameters, meteorological parameters and crop parameters as field crop data by using the edge gateway and uploading to the system cloud.

[0022] The soil sensor network is used to continuously monitor soil moisture, electrical conductivity (EC), pH value, main nutrient content (nitrogen, phosphorus, potassium), temperature of different soil layers and depth information. The field meteorological station is used to collect environmental data such as air temperature, air pressure, humidity, light radiation, wind speed and rainfall. The crop physiological monitor (hyperspectral sensor) is used to obtain physiological indicators such as crop growth state (leaf area index), biochemical composition, root water content and crop canopy surface temperature. All sensor data is preliminarily processed, protocol converted, aggregated and standardized by the edge gateway, and finally stably uploaded to the cloud data center to form a structured and time-synchronized field crop data set. This embodiment simultaneously captures environmental driving factors (meteorology), soil response states (water, fertilizer, heat) and crop physiological feedback (growth, stress), places high-frequency and real-time data collection and processing tasks in the field site (edge side) with uncertain network conditions, responsible for caching, preprocessing and ensuring data continuity; and places energy-consuming storage and complex analysis tasks in the cloud, realizing the optimal allocation of computing resources.

[0023] Step S2 includes the following sub-steps: Step S21, synchronizing the time stamps of the soil parameters, meteorological parameters and crop parameters of the field crop data, generating crop sample set data, calculating the mutual information values of a single feature in the crop sample set data with the electrical conductivity and root water content respectively using the mutual information criterion, if the mutual information value of the current feature is less than the set correlation information threshold, discarding the current feature, if the mutual information value of the current feature is greater than the set correlation information threshold, retaining the current feature, and generating first relationship data in chronological order; Step S22, calculating the soil heat flux per unit time according to the first relationship data, and the calculation expression of the soil heat flux is: ; wherein, is the soil heat flux, is the soil heat capacity, is the air temperature at the i-th moment, is the air temperature at the (i-1)-th moment, is the time step, is the soil depth; Step S23, calculating the soil water loss per unit time according to the first relationship data and the soil heat flux, and the calculation expression of the soil water loss is: ; ; wherein, is the soil water loss, is the crop coefficient, is the reference soil water loss, is the net solar radiation on the crop surface, is the soil heat flux, is the soil temperature, is the daily average air temperature, is the saturated water vapor pressure, is the actual water vapor pressure, is the average air temperature saturated water vapor pressure and temperature curve slope, is the wind speed; sequentially sorting the soil water loss per unit time calculated, and drawing a soil water loss curve graph.

[0024] The crop coefficient takes wheat and rice as an example, the wheat jointing stage: 0.7-0.85, the rice tillering stage: 1.1-1.2, the saturated water vapor pressure curve slope (kPa / ℃), the soil density empirical value is divided into sandy soil, clay and loam, sandy soil: ~1500 kg / m³; clay: ~1200 kg / m³; loam: ~1300 kg / m³, the mutual information correlation threshold is usually the median, average or determined by observation of all feature mutual information values, and the median is set in this embodiment 10%, as a mutual information correlation threshold; In this embodiment, the multi-source heterogeneous field crop data is time-synchronized to generate a unified crop sample set, then the mutual information non-parametric standard is used to quantify the correlation of each feature with the key target (electrical conductivity EC, root water content), and by setting a threshold, the most relevant feature subset to the target is automatically selected to generate high-quality first relationship data, effectively removing redundancy and noise, and realizing real-time or near real-time decision support.

[0025] Step S2 further includes: The proportion of soil nutrient content in total soil content in the first relationship data is counted, the crop component of the crop parameter is compared with the corresponding nutrient component in the soil to obtain a soil enrichment factor EF, and the general soil loss equation is used to obtain a soil loss amount A, so as to improve the calculation of the soil nutrient loss value. The calculation expression for improving the calculation of the soil nutrient loss value is: ; The soil nutrient loss values in a unit time are sorted according to time, a soil nutrient loss curve graph is drawn, and the soil water loss amount is one-to-one corresponding to the time stamp to obtain second relationship data; An irrigation decision-making model is constructed, a mechanism model based on HYDRUS is established, the second relationship data is time-step verified, and meteorological parameters and crop parameters are taken as input boundary conditions to predict and generate the next step soil nutrient loss value and soil water loss amount as the field water and fertilizer loss value.

[0026] First, the background concentration C (i.e. the nutrient amount in unit mass of soil, unit: mg / kg) of a specific nutrient (such as nitrogen, phosphorus) in the soil is calculated, the soil enrichment factor EF (which refers to the enrichment degree of the nutrient concentration in the lost sediment relative to the original soil concentration, usually >1) is calculated by analyzing the crop component and the soil component, and finally, C and EF are taken as key parameters to improve the classic general soil loss equation RUSLE, so that it is upgraded from predicting the sediment loss amount to directly predicting the soil nutrient loss amount L; In this embodiment, the improved RUSLE is used, which recognizes that nutrients mainly migrate by being adsorbed on the lost sediment (especially phosphorus and ammonium nitrogen). This method can quickly quantify the nutrient loss amount and provide a data benchmark for subsequent steps. Then, water and fertilizer coupling correlation is adopted to correlate and compare the water loss curve and the nutrient loss curve in the time dimension, which can intuitively analyze the time sequence relationship change and whether there is a lag effect between water loss and nutrient loss. Based on the physical law, the water and fertilizer transport equation is numerically solved to effectively convert the abstract data relationship into a visual and understandable causal chain.

[0027] Step S3 specifically includes: The water-fertilizer coupling rule for irrigation scheduling is set, and the trigger conditions of the water-fertilizer coupling rule include a first condition and a second condition. When the field water-fertilizer loss value is in the state of the first condition, the first irrigation strategy is triggered. When the field water-fertilizer loss value is in the state of the second condition, the second irrigation strategy is triggered. The first irrigation strategy and the second irrigation strategy are fused to dynamically compensate for the missing components of the soil until the field crop data reaches a balanced state. The first condition is used to judge the size of the soil water loss amount and the target soil water loss threshold value. When the soil water loss amount is greater than the target soil water loss threshold value, the first condition is executed. The first condition is used to judge the size of the soil water loss amount and the target soil water loss threshold value. When the soil water loss amount is greater than the target soil water loss threshold value, the first condition is executed. The first irrigation strategy is used to mobilize the valve and start the soil water compensation irrigation strategy. The second irrigation strategy is used to adjust the valve and expect the soil nutrient irrigation strategy.

[0028] The water-fertilizer coupling rule is set in advance. The rule includes two independent trigger conditions (the first condition and the second condition) and two corresponding execution strategies (the first irrigation strategy and the second irrigation strategy). When only the first condition is established (i.e., only water is lacking), the system triggers the first strategy. The strategy executes a single soil water compensation irrigation by controlling the valve and water pump of the irrigation system to supplement water without supplementing fertilizer. When only the second condition is established (i.e., only fertilizer is lacking), the system triggers the second strategy. The strategy executes soil nutrient irrigation (such as fertilizer irrigation through a drip irrigation system) by controlling the fertilizer injection pump and valve to supplement fertilizer without additionally increasing a large amount of water. When both conditions are established (i.e., water and fertilizer are both lacking), the system fuses the first and second strategies to start integrated water-fertilizer irrigation, which simultaneously supplements water and nutrients. Irrigation will continue until the real-time monitored field crop data indicates that the soil water and fertilizer content returns to the preset balanced state (i.e., below the target threshold value). The system then stops irrigation. This embodiment maximally reduces the amount of water and fertilizer input by compensating on demand, directly reduces production costs, and reduces deep percolation and runoff loss caused by excessive irrigation and fertilization from the source, thereby reducing the risk of non-point source pollution.

[0029] Step S4 includes the following sub-steps: Step S41, the cloud field crop data is constructed in HYDRUS-2D to build a 3D farmland all-around model, and the upper boundary, the lower boundary and the solute boundary are set. The upper boundary includes the atmospheric boundary, the meteorological parameters are input to the upper boundary, and the condition of triggering the first irrigation strategy in step S3 is superimposed. The lower boundary is set as a variable pressure drainage function, and the solute boundary is used to superimpose the second irrigation strategy. Step S42, for the predicted time point of field water and fertilizer loss value, the simulation irrigation compensation strategy is adopted at the predicted time point, and is valued in the 3D farmland all-around model; Step S43, the soil wetting peak value, depth, and uniformity after simulation irrigation are counted, the optimal interval of crop soil water and fertilizer simulation value of crop root water content and nutrient concentration is analyzed according to the current crop parameters, the 3D farmland all-around model is quantified by using a four-dimensional evaluation system, and the fitting goodness index determines the coefficient of determination When The closer the value is to 1, the stronger the linear relationship between the soil water and fertilizer simulation value and the soil water and fertilizer measured value, the simulation value of irrigation is repeatedly adjusted until The result of the value is close to 1, the farmland numerical twin model is generated, and the characteristic contribution value is obtained.

[0030] The historical and real-time data (soil, weather, crop) of the cloud are imported into HYDRUS-2D, a parameterized 3D farmland all-around model is constructed, the upper boundary is set as an atmospheric boundary, meteorological parameters (rainfall, ) are input, and the decision logic of step S3 is superimposed on this boundary, that is, when the simulation conditions trigger the first or second condition, the corresponding irrigation or fertilization event is automatically applied, the lower boundary is set as a variable pressure water head or free drainage to simulate more realistic root zone bottom water exchange, and the solute boundary is used to define the fertilization event corresponding to the triggering of the second irrigation strategy. In the digital twin, different simulation irrigation compensation strategies (for example, changing the irrigation amount, duration, and fertilization concentration) are run for future time points that need to be predicted. After the simulation is completed, HYDRUS outputs high spatio-temporal resolution results, and the peak value, depth, and uniformity of the soil wetting peak are counted. According to the current crop root distribution and water and fertilizer demand law (from the crop parameters), the optimal interval of root layer water content and nutrient concentration is determined. The proportion of the simulation result in the optimal interval is taken as the soil water and fertilizer simulation value. A four-dimensional evaluation system is used to comprehensively quantify the performance of the digital twin model, including water balance accuracy, solute transport accuracy, root layer state accuracy, and spatio-temporal dynamic accuracy. The coefficient of determination R² is used as the goodness of fit index, and the soil water and fertilizer simulation value is linearly regressed with the soil water and fertilizer measured value at the same period. The key input parameters (such as soil hydraulic parameters, root water absorption parameters, and solute transport parameters) in the HYDRUS model are repeatedly adjusted and the simulation is re-run until the R² value is close to 1. At this time, the model is considered to be highly calibrated and can be used as a reliable farmland digital twin model. In the optimization process, through sensitivity analysis, the contribution of each adjusted parameter to the improvement of model accuracy can be calculated, that is, the characteristic contribution value is obtained; The embodiment perfectly combines digitalization, modeling and optimization control of farmland, constructs a digital twin model, continuously runs, receives real-time data, and rolls over to predict water and fertilizer conditions in the next few days, thereby providing the rule engine in step S3 with advanced and accurate decision basis.

[0031] The first irrigation strategy and the second irrigation strategy specifically include: When the soil water loss amount is greater than the target soil water loss threshold, the first irrigation strategy is adopted, and positive irrigation is performed on the current soil according to the soil water loss amount, and the irrigation water amount is equal to the soil water loss amount. When the soil nutrient loss value is greater than the target soil nutrient loss threshold, the second irrigation strategy is adopted, and positive irrigation is performed on the current soil according to the soil nutrient loss amount, and the irrigation nutrient amount is equal to the soil nutrient loss amount.

[0032] The system continuously monitors the real-time water and fertilizer loss values calculated by step S2, executes the first irrigation strategy (water replenishment): the system controls the irrigation valve and the water pump to execute the irrigation operation. The water amount of irrigation is directly equal to the monitored water loss amount, executes the second irrigation strategy (fertilizer replenishment): the system controls the fertilizer injection pump and valve to inject the fertilizer mother liquor into the irrigation system according to the proportion. The nutrient amount of irrigation is directly equal to the monitored nutrient loss amount, the two strategies can be independently triggered and executed, or can be triggered at the same time, realizing the water and fertilizer integrated synchronous compensation. The biggest advantage of this strategy is that it is extremely simple, easy to understand and realize.

[0033] The feature contribution value obtaining method includes: The global sensitivity analysis method is adopted, the contribution degree of each feature of the irrigation simulation value to the soil water and fertilizer measured value is quantified, the soil nutrient content and water content are taken as target scalars, each feature of the irrigation simulation value is taken as an input feature, a feature distribution threshold is defined, the input features are sampled by using Sobol sequence, a plurality of groups of parameter combinations are generated, the simulation irrigation compensation strategy is called for each group of parameters, the target variable is output, the variance of the target variable and the sensitivity index of each input feature are calculated, and the sensitivity index is taken as the feature contribution value of each feature.

[0034] In this embodiment, Sobol sequence is used to perform quasi-Monte Carlo sampling on all input features. Sobol sequence can generate uniformly distributed sample points in multi-dimensional space, and higher computational efficiency can be achieved with fewer sampling times. Each sampling generates a complete set of parameter combinations (e.g., [irrigation amount = 15 mm, fertilizer concentration = 0.3%,...]), for each set of parameter combinations, a calibrated digital twin model (HYDRUS) is called to run the simulation irrigation compensation strategy and obtain the corresponding target variables. Based on all sampling points and their corresponding outputs, the total variance of the output results is analyzed and decomposed into each input feature and its interaction. Through rigorous mathematical framework and a large number of physical model simulations, the most in-depth mechanism insight and the most reliable decision support are provided for complex agricultural systems.

[0035] Step S5 specifically includes: Using the swarm plot, the feature contribution value is generated on a visual large screen. Each feature point corresponds to a data point, and the feature contribution value is analyzed according to the data point distribution characteristics. When the feature points are mainly distributed on the right side, it indicates that the irrigation amount and nutrient loss present a positive influence. When the feature points are distributed on both sides, it indicates that the irrigation amount and nutrient loss present a negative influence. When the feature points are scattered, the influence of irrigation amount and nutrient loss on each other is small.

[0036] In this embodiment, the abstract sensitivity index is converted into a large number of individual cases of the specific influence of each feature by using the swarm plot. The positive correlation, negative correlation or complex nonlinear relationship between each feature and the target is clearly displayed. In addition, the width and density of the point distribution provide additional dimensional information, which may be strongly dependent on the cooperation of other conditions. The visual large screen can intuitively understand and directly guide the visual insight and action guide of agricultural practice.

[0037] Embodiment 2, refer to Figure 2 provides a deep learning-based agricultural intelligent irrigation scheduling system, which includes a data acquisition module, a data analysis module, a strategy compensation module, a digital simulation module and a visualization analysis module. The data acquisition module is used to acquire crop data in a unit of time and comprehensively acquire real-time state information of the farmland. The data analysis module is used to construct an irrigation decision-making model and quantize the water and fertilizer loss value in the field according to the crop data. The strategy compensation module is used to formulate a precise irrigation and fertilization scheme, set a water and fertilizer dynamic coupling mechanism for irrigation scheduling, and adopt a first irrigation strategy and a second irrigation strategy according to the water and fertilizer loss value in the field. The digital simulation module is used to construct a digital twin model of the farmland based on the growth trend of the crops in the field, simulate the dynamic compensation strategy and calculate the feature contribution value. The visualization analysis module is used for visually displaying the influence of each decision feature by using a chart and visualizing the feature contribution value to generate a farmland decision reason analysis report.

[0038] In the embodiment, the embodiment is divided into five modules, which represent a highly integrated and systematic solution. The modules form a complete closed-loop intelligent system from perception analysis to decision verification. The data flow starts from collection and finally guides the decision through the visualization report, and the decision effect is verified through the collection module, forming an enhanced loop that can continuously learn and optimize, making the system highly intelligent in data processing.

[0039] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be realized by any type of volatile or non-volatile storage device or their combination, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk. These computer program instructions can also be stored in a computer readable storage medium that can guide the computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction means, which realize the functions specified in the flow Figure 1 one or more flows and / or blocks Figure 1 the functions specified in one or more flows or blocks.

[0040] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalent ones without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based intelligent irrigation scheduling method for agriculture, characterized in that, Includes the following steps: Step S1: Collect field crop data within a unit of time, including soil parameters, meteorological parameters, and crop parameters; Step S2: Construct an irrigation decision model and analyze the field water and fertilizer loss values ​​based on the field crop data; Step S3: Set up a dynamic water and fertilizer coupling mechanism for irrigation scheduling, and adopt a dynamic compensation strategy for the water and fertilizer loss values ​​in the field. Step S4: Based on the crop growth trend in the field, construct a digital twin model of farmland, simulate the dynamic compensation strategy, and generate feature contribution values; Step S5: Visualize the feature contribution values ​​to generate a farmland decision-making cause analysis report.

2. The intelligent agricultural irrigation scheduling method based on deep learning as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11: Deploy soil parameter sensors, meteorological parameter sensors, and crop physiological monitoring instruments in the field to collect soil moisture, electrical conductivity, pH value, soil nutrient content, soil depth, and soil temperature per unit time as soil parameters. Step S12: Collect temperature, air pressure, humidity, solar radiation, wind speed and rainfall per unit time as meteorological parameters; Step S13: Collect crop growth status, crop composition, root moisture content and crop surface temperature per unit time as crop parameters. Step S14: Use the edge gateway to save the soil parameters, meteorological parameters and crop parameters as field crop data and upload them to the system cloud.

3. The intelligent agricultural irrigation scheduling method based on deep learning as described in claim 2, characterized in that, Step S2 includes the following sub-steps: Step S21: Align the soil parameters, meteorological parameters, and crop parameters of the field crop data with the synchronous timestamps to generate crop sample set data. Calculate the mutual information value between a single feature in the crop sample set data and the electrical conductivity and root moisture content, respectively, using the mutual information criterion. If the mutual information value of the current feature is less than the set relevant information threshold, discard the current feature. If the mutual information value of the current feature is greater than the set relevant information threshold, retain the current feature and generate first relational data by sorting according to the timestamp. Step S22: Calculate the soil heat flux per unit time based on the first relationship data. The expression for calculating the soil heat flux is: ； in, For soil heat flux, For soil heat capacity, Let be the temperature at time i. Let be the temperature at time i-1. For time step, Soil depth; Step S23: Calculate the soil moisture loss per unit time based on the first relationship data and soil heat flux. The expression for calculating the soil moisture loss is as follows: ； ； in, This represents the amount of soil moisture lost. For crop coefficients, For reference purposes, soil moisture loss Net light radiation to the crop surface, For soil heat flux, For soil temperature, The average daily temperature The saturated water vapor pressure, This is the actual water vapor pressure. The slope of the curve representing saturated water vapor pressure versus temperature at average air temperature. Wind speed; Soil moisture loss calculated per unit time is sorted in chronological order, and a soil moisture loss curve is plotted.

4. The intelligent agricultural irrigation scheduling method based on deep learning as described in claim 3, characterized in that, Step S2 further includes: The proportion of soil nutrient content in the first relational data to the total soil content is statistically analyzed, and the background concentration of crop soil nutrients C is calculated. The crop components of the crop parameters are extracted and compared with the corresponding nutrient components in the soil to obtain the soil enrichment factor EF. Then, the soil loss amount A is obtained using the general soil loss equation. The soil nutrient loss value is calculated in an improved manner. The expression for the improved calculation of the soil nutrient loss value is as follows: ； Soil nutrient loss values ​​per unit time are sorted according to time sequence, and a soil nutrient loss curve is plotted. The curve is then matched one-to-one with the soil moisture loss according to the timestamp to obtain the second relationship data. An irrigation decision model is constructed, a HYDRUS-based mechanistic model is established, the second relationship data is validated by time step, and the meteorological parameters and crop parameters are used as input boundary conditions to predict and generate the next soil nutrient loss value and soil moisture loss value as field water and fertilizer loss value.

5. The intelligent agricultural irrigation scheduling method based on deep learning as described in claim 4, characterized in that, Step S3 specifically includes: Set up water and fertilizer coupling rules for irrigation scheduling. The triggering conditions of the water and fertilizer coupling rules include a first condition and a second condition. When the field water and fertilizer loss value is under the first condition, the first irrigation strategy is triggered. When the field water and fertilizer loss value is under the second condition, the second irrigation strategy is triggered. The first irrigation strategy and the second irrigation strategy are combined to dynamically compensate for the missing soil components until the field crop data reaches a balanced state. The first condition is used to determine the magnitude of soil moisture loss and the target soil moisture loss threshold. When the soil moisture loss is greater than the target soil moisture loss threshold, the first condition is executed. The first condition is used to determine the magnitude of the soil nutrient loss value and the target soil nutrient loss threshold. When the soil moisture loss value is greater than the target soil nutrient loss threshold, the second condition is executed. The first irrigation strategy is used to adjust the valve and initiate the soil replenishment irrigation strategy; The second irrigation strategy is used to adjust the valves and anticipate soil nutrient irrigation strategies.

6. The intelligent agricultural irrigation scheduling method based on deep learning as described in claim 5, characterized in that, Step S4 includes the following sub-steps: Step S41: Construct a 3D all-round farmland model in HYDRUS-2D using the field crop data in the cloud, and set an upper boundary, a lower boundary, and a solute boundary. The upper boundary includes an atmospheric boundary. Input the meteorological parameters into the upper boundary and overlay the conditions for triggering the first irrigation strategy in step S3. The lower boundary is set to variable pressure drainage function. The solute boundary is used to overlay the second irrigation strategy. Step S42: For the predicted time point of the field water and fertilizer loss value, a simulated irrigation compensation strategy is adopted at the predicted time point and assigned to the 3D farmland all-round model. Step S43: Statistically analyze the peak soil moisture content, depth, and uniformity after simulated irrigation. Based on current crop parameters, analyze the optimal range of crop root water content and nutrient concentration in the simulated soil water and fertilizer values. Quantify the 3D farmland all-round model using a four-dimensional evaluation system. Based on the simulated soil water and fertilizer values ​​and the measured soil water and fertilizer values, determine the goodness-of-fit index, the coefficient of determination. ,when The closer the value is to 1, the stronger the linear relationship between the simulated soil water and fertilizer values ​​and the measured soil water and fertilizer values. The irrigation simulation values ​​are repeatedly adjusted until... The result of the value is infinitely close to 1, generating a numerical twin model of farmland and obtaining the feature contribution value.

7. The intelligent agricultural irrigation scheduling method based on deep learning as described in claim 6, characterized in that, The first irrigation strategy and the second irrigation strategy specifically include: When the soil moisture loss exceeds the target soil moisture loss threshold, the first irrigation strategy is adopted, which involves positive irrigation of the current soil in response to the soil moisture loss, with the irrigation water volume equal to the soil moisture loss. When the soil nutrient loss value exceeds the target soil nutrient loss threshold, a second irrigation strategy is adopted, which involves positive irrigation of the current soil in response to the soil nutrient loss, with the irrigation nutrients equal to the soil nutrient loss.

8. The intelligent agricultural irrigation scheduling method based on deep learning as described in claim 7, characterized in that, The method for obtaining the feature contribution value includes: A global sensitivity analysis method is adopted to quantify the contribution of each feature of the irrigation simulation value to the measured soil water and fertilizer value. Soil nutrient content and water content are used as target scalars, and each feature of the irrigation simulation value is used as input feature. A feature distribution threshold is defined, and the input features are sampled using Sobol sequences to generate multiple sets of parameter combinations. For each set of parameters, a simulated irrigation compensation strategy is called to output the target variable. The variance of the target variable and the sensitivity index of each input feature are calculated, and the sensitivity index is used as the feature contribution value of each feature.

9. The intelligent agricultural irrigation scheduling method based on deep learning as described in claim 8, characterized in that, Step S5 specifically includes: Using a bee colony graph, the feature contribution values ​​are generated into a large visualization screen. The feature contribution value corresponding to each feature point is taken as a data point. Based on the distribution characteristics of the data points, the reasons for farmland decision-making are analyzed. When the feature points are mainly distributed on the right side, it indicates that irrigation amount and nutrient loss have a positive impact. When the feature points are distributed on both sides, it indicates that irrigation amount and nutrient loss have a negative impact. When the feature points are scattered, the mutual influence between irrigation amount and nutrient loss is relatively small.

10. A deep learning-based intelligent agricultural irrigation scheduling system, applied in any one of the deep learning-based intelligent agricultural irrigation scheduling methods as described in claims 1-9, characterized in that, It includes a data acquisition module, a data analysis module, a strategy compensation module, a digital simulation module, and a visualization analysis module; The data acquisition module is used to collect field crop data within a unit of time to comprehensively obtain real-time status information of farmland. The data analysis module is used to build an irrigation decision model and quantify the water and fertilizer loss value in the field based on field crop data. The strategy compensation module is used to formulate precise irrigation and fertilization plans, set up a dynamic water and fertilizer coupling mechanism for irrigation scheduling, and adopt a first irrigation strategy and a second irrigation strategy for the field water and fertilizer loss value. The digital simulation module is used to construct a farmland digital twin model based on the crop growth trend in the field, simulate the dynamic compensation strategy, and calculate the feature contribution value. The visualization analysis module is used to visually display the impact of each decision feature using charts and graphs, and to visualize the contribution values ​​of the features to generate a farmland decision cause analysis report.

Images