Farmland irrigation water amount intelligent optimization method based on deep learning

Abstract

The invention discloses a deep learning-based intelligent optimization method for farmland irrigation water quantity. The method comprises the following steps of S1, obtaining a preprocessed farmland multi-mode perception data set; s2, constructing and training a multi-scale gated SIREN water content continuous prediction model by taking the preprocessed farmland multi-modal perception data set as input; s3, dividing a planned irrigation period into a plurality of time slices, defining a water volume decision vector space and constructing a differential evolution multi-target water volume optimization model; s4, outputting a Pareto optimal water quantity decision vector set meeting a multi-target constraint condition, and selecting an optimal farmland irrigation water quantity from the Pareto optimal water quantity decision vector set according to a user weight or a preset rule; s5, the optimal farmland irrigation water amount is issued to an intelligent valve control system, and an electromagnetic valve is driven to execute irrigation according to time slices. According to the invention, an intelligent irrigation prediction-decision-execution-feedback full-link closed-loop mechanism taking SIREN as a core is realized, and the method has remarkable water-saving and yield-increasing benefits and engineering deployability.

Description

Technical Field

[0001] The present invention relates to the technical field of agriculture, and in particular to an intelligent optimization method for farmland irrigation water volume based on deep learning. Background Art

[0002] With the rapid development of smart agriculture technology, farmland irrigation is gradually evolving from traditional experience-driven to data-driven and model-driven precise regulation. As the core link of agricultural water use, the quality of irrigation strategies directly affects crop yields, water resource utilization efficiency, and energy consumption levels. However, in the existing irrigation regulation system, there are generally technical bottlenecks such as insufficient characterization of soil moisture spatial distribution, limited irrigation decision-making optimization ability, and lack of a multi-objective dynamic regulation framework, which severely restricts the construction of an efficient water-saving irrigation system.

[0003] Currently, mainstream farmland irrigation strategies usually rely on tensiometer single-point sensor devices or irrigation threshold models based on empirical rules to judge the moisture status of the crop root zone. However, these methods are difficult to comprehensively obtain the high-frequency moisture content gradient changes in the 0–60 cm deep soil layer, and there are often coexisting phenomena of "waterlogging in low-lying areas and water shortage on ridge tops" in the same area in heterogeneous fields affected by factors such as terrain undulation, permeability differences, or shaded areas.

[0004] In recent years, some deep learning-based soil moisture modeling methods have been gradually applied to agricultural scenarios. However, most models still rely on regular grid partitioning and static data input, making it difficult to reconstruct the soil water content state at high resolution in a continuous space, and even more unable to provide a differentiable continuous prediction function interface for optimization algorithms. At the same time, some optimization strategies such as traditional genetic algorithms and particle swarm algorithms are prone to falling into local optima when dealing with high-dimensional non-linear irrigation water volume search spaces, lacking the ability to deeply express the multi-objective coupling relationship.

[0005] Therefore, there is an urgent need for a new method for complex farmland scenarios to solve the core problems of uneven irrigation space, weak optimization ability, and lack of a closed-loop learning mechanism in the existing technology. Summary of the Invention

[0006] An object of the present invention is to propose an intelligent optimization method for farmland irrigation water volume based on deep learning. The present invention realizes an intelligent irrigation prediction-decision-execution-feedback full-link closed-loop mechanism with SIREN as the core, and has significant water-saving and yield-increasing benefits and engineering deployability.

[0007] An intelligent optimization method for farmland irrigation water volume based on deep learning according to an embodiment of the present invention includes the following steps:

[0008] S1. Periodically collect a multi-modal perception dataset of farmland and perform preprocessing to obtain a preprocessed multi-modal perception dataset of farmland;

[0009] S2. With the preprocessed multi-modal perception dataset of farmland as the input, construct and train a multi-scale gated SIREN continuous moisture content prediction model to obtain the SIREN model parameters that can output the three-dimensional spatial-temporal continuous moisture content distribution and spatial gradient;

[0010] S3. Divide the planned irrigation period into several time slices, define the water volume decision vector space and construct a differential evolution multi-objective water volume optimization model. Each component of the water volume decision vector corresponds to the farmland irrigation water volume of a time slice-valve combination. In the differential evolution multi-objective water volume optimization model, construct a multi-objective fitness function based on the multi-scale gated SIREN continuous moisture content prediction model;

[0011] S4. Perform population initialization, differential mutation, crossover recombination and fitness selection on the differential evolution multi-objective water volume optimization model. In each fitness evaluation process, inject the candidate water volume decision vector as the conditional vector into the SIREN continuous moisture content prediction service to obtain the continuous moisture content distribution and spatial gradient and calculate the multi-objective fitness. Iteratively update until the termination condition is met, and output the Pareto optimal water volume decision vector set that meets the multi-objective constraint conditions, and select the optimal farmland irrigation water volume from it according to the user weight or preset rules;

[0012] S5. Send the optimal farmland irrigation water volume to the intelligent valve control system to drive the solenoid valve to perform irrigation according to the time slice.

[0013] Optionally, the S1 includes the following steps:

[0014] S11. Deploy sensing devices for periodically collecting field environmental factors in the irrigation area. The sensing devices include tensiometers for collecting soil water potential, frequency domain reflectometry moisture sensors for collecting soil moisture content, near-infrared imaging devices for collecting canopy temperature, and micro-meteorological stations for collecting meteorological factors. Sample the environmental information of the irrigation area at a unified time interval to construct a raw farmland multi-modal perception dataset D containing N perception data samples d i The raw farmland multi-modal perception dataset includes the sampling timestamp t raw , the two-dimensional spatial coordinates x i , y i of the sensing point, the soil water potential ψ i , the soil volumetric moisture content θ i , the canopy temperature T i , and the meteorological factor vector ξ i composed of air temperature, humidity, wind speed and solar radiation; i ;

[0015] S12. Synchronize the original multi-modal farmland perception dataset according to a preset time window, align asynchronous sampling data of various types to a unified time benchmark, smoothly complete the data at missing time points using time interpolation methods, and map the sampling data to regular grid coordinates according to the spatial interpolation grid resolution. Estimate the perception values at missing spatial positions using spatial interpolation algorithms to generate a multi-modal farmland spatial completion dataset;

[0016] S13. Perform anomaly elimination processing on the multi-modal farmland spatial completion dataset, calculate the deviation ratio of each data component, determine and eliminate abnormal data when the deviation ratio is greater than the preset deviation threshold, and convert all data items of the multi-modal farmland anomaly elimination dataset to a unified physical dimension to complete the unit unification process, obtaining the preprocessed multi-modal farmland perception dataset D final 。

[0017] Optionally, the S2 includes the following steps:

[0018] S21. Append the synchronous irrigation depth final in each sample of the preprocessed multi-modal farmland perception dataset D the future 24-hour rainfall forecast amount r j and the reference evapotranspiration amount e j to construct a training sample set S train ;

[0019] S22. Perform Fourier feature mapping on each group of spatial-temporal coordinates (x j , y j and depth coordinate z j ) consisting of to generate a high-frequency vector φ j , y j , z j , t j ): j :

[0020] φ j = [sin(2πB[x j , y j , z j , t j )] T ), cos(2πB[x j , y j , z j , t j )] T ;

[0021] where B is a fixed random frequency matrix;

[0022] S23. Based on the high-frequency vector φ j, construct a multi-scale gated SIREN continuous water content prediction model:

[0023]

[0024] Among them, f ΘSIREN (φ, p cum , r, e) represents the multi-scale gated SIREN continuous water content prediction model for predicting the volumetric water content in the crop root zone. α l (p cum , r, e) = σ(a l p cum + b l r + c l e + d l ) is the gating factor, and σ(·) is the Sigmoid function. is the set of model parameters to be trained. represents the sub-network constructed for the l-th soil layer, which is used to model the response characteristics of water content in different soil layers of 0–20 cm, 20–40 cm, and 40–60 cm.

[0025] S24. Using the soil volumetric water content θ j as the supervision signal, train the multi-scale gated SIREN continuous water content prediction model Minimize the model loss function

[0026]

[0027] Among them, λ is the physical constraint regularization weight, K(θ j ) is the hydraulic conductivity function corresponding to the soil volumetric water content, M represents the total number of samples used to train the multi-scale gated SIREN continuous water content prediction model. represents the time change rate of the soil volumetric water content at the j-th sample. represents the divergence term of the flow controlled by the soil water potential. The second term is the physical consistency residual term, which is used to evaluate whether the predicted value conforms to the Richards soil water migration control equation.

[0028] S25. After the training of the multi-scale gated SIREN continuous water content prediction model converges, rank based on the mutual information value I(θ, E) between the soil volumetric water content prediction result and the unit yield energy consumption E, retain the top K frequency components in the first-layer frequency matrix B, and obtain the pruned frequency matrix B K , and update the model parameters to the compressed model parameter set

[0029] S26. For the model parameter set Perform 8-bit weight quantization processing and insert the MC-Dropout structure to generate the final multi-scale gated SIREN water content continuous prediction model G. The call interface is defined as:

[0030]

[0031] where p opt is an element of a certain optimized water volume decision vector, is the predicted soil volumetric water content, is its gradient vector in the spatial coordinates, is the predicted standard deviation estimated based on MC-Dropout.

[0032] Optionally, the S3 includes the following steps:

[0033] S31. Divide the planned farmland irrigation period T irrig into T non-overlapping time slice sets {τ1, τ2, …, τ T}, and the duration of each time slice is Δτ t , satisfying that each time slice τ t corresponds to an electromagnetic valve control operation for performing phased farmland irrigation tasks;

[0034] S32. Set the number of independent electromagnetic valves deployed in the farmland irrigation area as V. Each time slice can control V independent farmland irrigation channels, and construct the farmland irrigation water volume decision vector where, represents the farmland irrigation water volume of the v-th electromagnetic valve within the time slice τ t ;

[0035] S33. Define the upper and lower boundary limits for each electromagnetic valve-time slice combination, satisfying that the farmland irrigation water volume of the v-th electromagnetic valve within the time slice τ t is between the minimum farmland irrigation water volume t and the maximum farmland irrigation water volume of the v-th electromagnetic valve within the time slice τ . The upper and lower boundary limits are set according to the pump station water supply capacity, crop waterlogging tolerance, and soil infiltration threshold;

[0036] S34. Construct a differential evolution multi-objective water volume optimization model, and use the farmland irrigation water volume decision vector p opt as the coding structure of each individual in the differential evolution multi-objective water volume optimization model, where the farmland irrigation water volume decision vector p optA combination including T time slices and V valves, where each dimension represents the farmland irrigation water volume of a certain time slice-valve combination. In the differential evolution population, all individuals are initialized within the water volume boundary constraint interval to form the initial population. The number of individuals in the initial population is N, and the search space is jointly defined by the upper and lower limits of all T×V groups of valve-controlled water volumes;

[0037] S35. In the differential evolution multi-objective water volume optimization model, based on the individuals in each generation of the population, a mutant vector is generated through the differential mutation operation. The differential mutation operation is constructed depending on the water volume difference between three different individuals. After mutation, a crossover operation based on the set crossover probability C is performed on each individual. r In the crossover operation, each dimension in the candidate water volume decision vector will randomly select its source between the original individual and the mutant vector and at least one dimension comes from the mutant vector. After the crossover operation, a child water volume decision vector is formed and enters the next round of fitness evaluation;

[0038] S36. In the fitness evaluation session, the multi-scale gated SIREN water content continuous prediction model G is called. Taking the farmland irrigation water volume corresponding to each time slice-valve combination in the water volume decision vector p opt as one of the model inputs, and at the same time combining the spatial-temporal position, cumulative irrigation depth, reference evapotranspiration, and rainfall forecast volume to obtain the predicted volumetric water content under this water volume configuration The spatial gradient of the predicted volumetric water content and the standard deviation of the predicted volumetric water content The predicted volumetric water content The spatial gradient of the predicted volumetric water content and the standard deviation of the predicted volumetric water content are coupled with four optimization objectives to construct a multi-objective fitness function F(p opt ):

[0039] F(p opt ) = ω1f1 + ω2f2 + ω3f3 + ω4f4;

[0040] Among them, the multi-objective preference weight vector ω = [ω1, ω2, ω3, ω4], f1 is the sum of squared differences in water content term, f2 is the surface runoff risk control term, f3 is the term to maximize transpiration satisfaction, and f4 is the irrigation energy consumption and total water volume control term.

[0041] Optionally, the sum of squared differences in water content term is expressed as the sum of squared differences between the target volumetric water content and the predicted volumetric water content;

[0042] The surface runoff risk control term is expressed as the degree of deviation between the spatial gradient of the predicted volumetric water content and the target water uniformity;

[0043] The transpiration satisfaction maximization term is expressed as the water content prediction uncertainty corresponding to the standard deviation of the predicted volumetric water content, which is used to guide the optimization to search in areas with high uncertainty;

[0044] The irrigation energy consumption and total water volume control term is expressed as a comprehensive index of the irrigation water volume per unit area of farmland and the energy consumption per unit yield, which is used to balance crop yield, water-saving efficiency and energy consumption.

[0045] Optionally, S4 includes the following steps:

[0046] S41. Based on the differential evolution multi-objective water volume optimization model and the multi-objective fitness function, initialize the water volume population and execute the generational evolution process. In each generation, use differential mutation and crossover recombination operations to generate candidate water volume individuals, and use the multi-scale gated SIREN water content continuous prediction model as the prediction service to obtain the water content prediction value, spatial gradient and uncertainty estimation, and accordingly complete the calculation of the multi-objective fitness function and fitness optimization;

[0047] S42. In each generation of iteration process, update the non-dominated solution set in real time according to the non-dominance judgment principle, and construct the Pareto optimal water volume decision vector set P Pareto , and record the change trend of each objective function for convergence monitoring;

[0048] S43. When the number of iterations reaches the maximum generation G or the average improvement value of the objective function in δ consecutive generations is lower than the convergence threshold ∈, it is determined that the differential evolution multi-objective water volume optimization model has met the convergence condition, terminate the iteration and output the final Pareto solution set P Pareto ;

[0049] S44. According to the multi-objective preference weight vector ω = [ω1, ω2, ω3, ω4] set by the user, or combined with the agricultural expert rules of the crop growth stage, select the final optimal farmland irrigation water volume vector p Pareto from the Pareto optimal water volume decision vector set P opt* .

[0050] Optionally, S6 includes the following steps:

[0051] S61. Send the optimal farmland irrigation water volume decision vector to the intelligent valve control system through the communication gateway. The intelligent valve control system contains V electromagnetic valve nodes, and each valve node performs quantitative irrigation control according to the water volume t within its own time slice τ ;

[0052] S62. Within each time slice τ t , the intelligent valve control system collects the actual irrigation flow and combined with the time slice length Δτ t calculate the actual farmland irrigation water volume Compare it with the optimal farmland irrigation water volume If there is a deviation then trigger the flow feedback correction logic to update the current valve control parameters to offset the deviation, where ∈ q is the threshold value;

[0053] S63. After the irrigation cycle ends, re-collect the farmland multi-modal perception data set D new using the latest farmland multi-modal perception data set D new the measured soil volumetric water content at each spatio-temporal point in Compare with the predicted water content output by calling the multi-scale gated SIREN water content continuous prediction model G to calculate the residual sequence ε j and calculate the mean square error of all residual samples. If the residual mean square error ε MSE > δ model i.e., exceeding the model accuracy threshold δ model then trigger the high-frequency spectral increment retraining process of the multi-scale gated SIREN water content continuous prediction model;

[0054] S64. During the high-frequency spectral increment retraining process, only fine-tune the parameter set of the high-frequency activation channels in the multi-scale gated SIREN water content continuous prediction model and retain the top K frequency components based on the current mutual information sorting structure. The updated frequency matrix is denoted as At the same time, perform a pruning operation to remove low-contribution channels and execute 8-bit weight quantization to generate the updated model parameter set

[0055] S65. Redeploy the updated multi-scale gated SIREN water content continuous prediction model to the field edge computing gateway for the water volume fitness evaluation process in the next round of differential evolution multi-objective water volume optimization model, realizing a continuous closed-loop learning and updating mechanism for farmland irrigation water volume regulation.

[0056] Optionally, step S5 further includes extracting the output classification criteria for the optimal farmland irrigation water volume and defining the farmland irrigation water volume recommendation rules under different regulation modes according to the distribution characteristics of the Pareto solution set, the model prediction error distribution, and the crop growth stage in historical iterations:

[0057] Drought repair type: If the average residual in the past three days is negatively biased and the evaporation amount ≥ 120% of the daily historical value, then increase the upper limit of the water volume component in each time slice by 10%;

[0058] Water-saving and stable type: If the absolute values of the residuals are all less than the threshold and the prediction uncertainty then, on the premise of maintaining the satisfaction constraint of the target transpiration, the total farmland irrigation water volume is compressed by no more than 15%;

[0059] Energy consumption constraint type: If the energy consumption index per unit output is more than 20% higher than the historical average, then some water uniformity indexes are sacrificed according to the target priority, and the optimal solution priority sorting strategy is switched from ω2>ω3 to ω4>ω1.

[0060] The beneficial effects of the present invention are as follows:

[0061] (1) Based on the traditional SIREN network, the present invention introduces a multi-scale hierarchical structure and an irrigation-meteorological gating mechanism to construct a continuous soil moisture prediction model that can hierarchically perceive the response characteristics of different soil depths. By fusing the control residuals of the Richards equation as a physical consistency term during the model training process, the model can not only perform high-frequency reconstruction in the space-time coordinate system, but also have physical interpretability, effectively avoiding prediction drift. Compared with the traditional neural network without prior knowledge, the mean square error of soil moisture prediction of the proposed model decreases in heterogeneous soil scenarios, and the spatial gradient continuity is significantly improved, which can be used as a stable fitness evaluator for optimization algorithms.

[0062] (2) The present invention designs a differential evolution optimization structure with the prediction result as the feedback closed loop, injects the irrigation water volume decision vector into the SIREN model, dynamically obtains the soil moisture prediction, spatial gradient and prediction variance corresponding to each individual, and constructs a multi-objective fitness function that integrates four objectives. By means of adjustable weights, the multiple optimization requirements under different crop and climate conditions are met, significantly improving the scientificity and flexibility of irrigation decision-making.

[0063] (3) The present invention designs a closed-loop feedback control structure with an intelligent valve control system as the execution body, collects the water volume execution results of the solenoid valve during the actual irrigation cycle, and combines the latest field perception data to calculate the prediction residuals of the SIREN model in real time. When the mean square of the residuals exceeds the threshold, an incremental retraining mechanism with a high-frequency channel as the core is automatically triggered to ensure that the model still maintains stable performance in the face of extreme climate changes or sudden changes in soil conditions. Combined with pruning and quantization compression strategies, the updated model can be redeployed in the MCU-level edge gateway to ensure the continuous online nature and lightweight efficiency of the field prediction and control process. Description of the Drawings

[0064] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0065] Figure 1Flowchart of an intelligent optimization method for farmland irrigation water volume based on deep learning proposed by the present invention. Specific implementation mode

[0066] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0067] Reference Figure 1 , an intelligent optimization method for farmland irrigation water volume based on deep learning, includes the following steps:

[0068] S1. Periodically collect the farmland multi-modal perception data set and perform preprocessing to obtain the preprocessed farmland multi-modal perception data set;

[0069] S2. Using the preprocessed farmland multi-modal perception data set as the input, construct and train a multi-scale gated SIREN moisture content continuous prediction model to obtain the SIREN model parameters that can output the three-dimensional space-time continuous moisture content distribution and spatial gradient;

[0070] S3. Divide the planned irrigation period into several time slices, define the water volume decision vector space and construct a differential evolution multi-objective water volume optimization model. Each component of the water volume decision vector corresponds to the farmland irrigation water volume of a time slice-valve combination. In the differential evolution multi-objective water volume optimization model, construct a multi-objective fitness function based on the multi-scale gated SIREN moisture content continuous prediction model;

[0071] S4. Perform population initialization, differential mutation, crossover recombination and fitness selection on the differential evolution multi-objective water volume optimization model. In each fitness evaluation process, inject the candidate water volume decision vector as the conditional vector into the SIREN continuous moisture content prediction service to obtain the continuous moisture content distribution and spatial gradient and calculate the multi-objective fitness. Iteratively update until the termination condition is met, and output the Pareto optimal water volume decision vector set that meets the multi-objective constraint conditions, and select the optimal farmland irrigation water volume from it according to the user weight or preset rules;

[0072] S5. Send the optimal farmland irrigation water volume to the intelligent valve control system to drive the solenoid valve to perform irrigation according to the time slice.

[0073] In this implementation mode, S1 includes the following steps:

[0074] S11. Deploy sensing devices for periodically collecting field environmental factors in the irrigation area. The sensing devices include a tensiometer for collecting soil water potential, a frequency domain reflectometry moisture sensor for collecting soil moisture content, a near-infrared imaging device for collecting canopy temperature, and a micro meteorological station for collecting meteorological factors. Sample environmental information in the irrigation area at a unified time interval to construct an original farmland multi-modal perception data set D containing N perception data samples d i ; raw The original farmland multi-modal perception data set includes the sampling timestamp t i , the two-dimensional spatial coordinates x i , y i of the sensing point, the soil water potential ψ i , the soil volumetric water content θ i , the canopy temperature T i , and the meteorological factor vector ξ i composed of air temperature, humidity, wind speed and solar radiation;

[0075] S12. Perform time synchronization processing on the original farmland multi-modal perception data set according to a preset time window, align asynchronous sampling data of various types to a unified time reference, smoothly complete the data at missing time points by using the time interpolation method, and map the sampling data to regular grid coordinates according to the spatial interpolation grid resolution. Use the spatial interpolation algorithm to estimate the perception values at missing spatial positions to generate a spatially completed farmland multi-modal data set;

[0076] S13. Perform abnormal data elimination processing on the spatially completed farmland multi-modal data set, calculate the deviation ratio of each data component, determine the abnormal data and eliminate it when the deviation ratio is greater than the preset deviation threshold, and convert all data items of the farmland multi-modal abnormal data elimination data set to a unified physical dimension to complete the unit unification process to obtain the preprocessed farmland multi-modal perception data set D final .

[0077] In this embodiment, S2 includes the following steps:

[0078] S21. Append the synchronous irrigation depth final to each sample of the preprocessed farmland multi-modal perception data set D , the future 24-hour rainfall forecast volume r j and the reference evapotranspiration e j to construct a training sample set S train ;

[0079] S22. For each group with the spatial-temporal coordinates (x j , y j and the depth coordinate z j constituting the spatial-temporal coordinates (x j , y j,z j ,t j ) Perform Fourier feature mapping to generate the high-frequency vector φ j :

[0080] φ j = [sin(2πB[x j ,y j ,z j ,t j )] T ), cos(2πB[x j ,y j ,z j ,t j )] T )];

[0081] Among them, B is a fixed random frequency matrix;

[0082] S23. Based on the high-frequency vector φ j , construct a multi-scale gated SIREN moisture content continuous prediction model:

[0083]

[0084] Among them, represents the multi-scale gated SIREN moisture content continuous prediction model for predicting the volumetric moisture content in the crop root zone. α l (p cum , r, e) = σ(a l p cum + b l r + c l e + d l ) is the gating factor, σ(·) is the Sigmoid function, is the set of model parameters to be trained, represents the sub-network constructed for the l-th soil layer, which is used to model the response characteristics of the moisture content in different soil layers of 0–20 cm, 20–40 cm, and 40–60 cm;

[0085] S24. Using the soil volumetric moisture content θ j as the supervision signal, train the multi-scale gated SIREN moisture content continuous prediction model Minimize the model loss function

[0086]

[0087] Among them, λ is the physical constraint regularization weight, K(θ j ) is the hydraulic conductivity function corresponding to the soil volumetric moisture content, and M represents the total number of samples used to train the multi-scale gated SIREN moisture content continuous prediction model. represents the time variation rate of soil volumetric water content at the j-th sample, represents the flow divergence term controlled by soil water potential, and the second term is the physical consistency residual term, which is used to evaluate whether the predicted value conforms to the Richards soil moisture migration control equation;

[0088] S25. After the training of the multi-scale gated SIREN water content continuous prediction model converges, rank based on the mutual information value I(θ, E) between the predicted result of soil volumetric water content and the unit yield energy consumption E, and retain the top K frequency components in the first-layer frequency matrix B to obtain the pruned frequency matrix B K , and update the model parameters to the compressed model parameter set

[0089] S26. Perform 8-bit weight quantization processing on the model parameter set , and insert the MC-Dropout structure to generate the final multi-scale gated SIREN water content continuous prediction model G. The calling interface is defined as:

[0090]

[0091] where p opt is an element of a certain optimized water volume decision vector, is the predicted soil volumetric water content, is its gradient vector in the spatial coordinates, is the predicted standard deviation estimated based on MC-Dropout.

[0092] In this embodiment, S3 includes the following steps:

[0093] S31. Divide the planned farmland irrigation period T irrig into a set of T non-overlapping time slices {τ1, τ2, …, τ T}, and the duration of each time slice is Δτ t , satisfying that each time slice τ t corresponds to one electromagnetic valve control operation for performing phased farmland irrigation tasks;

[0094] S32. Set the number of independent electromagnetic valves deployed in the farmland irrigation area to V. Each time slice can control V independent farmland irrigation channels, and construct a farmland irrigation water volume decision vector where represents the farmland irrigation water volume of the v-th electromagnetic valve in the time slice τ t ;

[0095] S33. Define the upper and lower boundary limits for each electromagnetic valve-time slice combination, satisfying the time slice τ tThe farmland irrigation water volume of the v-th electromagnetic valve within the time slice τ t The minimum farmland irrigation water volume of the v-th electromagnetic valve and the maximum farmland irrigation water volume are set with upper and lower boundaries based on the pumping station's water supply capacity, crop waterlogging tolerance, and soil infiltration threshold;

[0096] S34. Construct a differential evolution multi-objective water volume optimization model. Take the farmland irrigation water volume decision vector p opt as the coding structure of each individual in the differential evolution multi-objective water volume optimization model. The farmland irrigation water volume decision vector p opt contains the combination of T time slices and V valves. Each dimension represents the farmland irrigation water volume of a certain time slice-valve combination. In the differential evolution population, all individuals are initialized within the water volume boundary constraint interval to form an initial population. The number of individuals in the initial population is N, and the search space is jointly defined by the upper and lower limits of all T×V groups of valve-controlled water volumes;

[0097] S35. In the differential evolution multi-objective water volume optimization model, taking the individuals in each generation of the population as the basis, generate mutant vectors through differential mutation operations. The differential mutation operations are constructed based on the water volume differences between three different individuals. After mutation, perform a crossover operation based on the set crossover probability C r for each individual. In the crossover operation, each dimension in the candidate water volume decision vector will randomly select its source between the original individual and the mutant vector and at least contain one dimension from the mutant vector. After completing the crossover operation, form the offspring water volume decision vector and enter the next round of fitness evaluation;

[0098] S36. In the fitness evaluation, call the multi-scale gated SIREN water content continuous prediction model G. Take the farmland irrigation water volume corresponding to each time slice-valve combination in the water volume decision vector p opt as one of the model inputs. At the same time, combine the spatial-temporal position, cumulative irrigation depth, reference evapotranspiration, and rainfall forecast volume to obtain the predicted volumetric water content under this water volume configuration The spatial gradient of the predicted volumetric water content and the standard deviation of the predicted volumetric water content Couple the predicted volumetric water content The spatial gradient of the predicted volumetric water content and the standard deviation of the predicted volumetric water content with four optimization objectives to construct a multi-objective fitness function F(p opt ):

[0099] F(p opt ) = ω1f1 + ω2f2 + ω3f3 + ω4f4;

[0100] Among them, the multi-objective preference weight vector ω = [ω1, ω2, ω3, ω4], f1 is the sum of squares of moisture content deviation terms, f2 is the surface runoff risk control term, f3 is the term to maximize transpiration satisfaction, and f4 is the irrigation energy consumption and total water volume control term.

[0101] In this embodiment, the sum of squares of moisture content deviation terms is expressed as the sum of squares of the deviation between the target volume moisture content and the predicted volume moisture content;

[0102] The surface runoff risk control term is expressed as the degree of deviation between the spatial gradient of the predicted volume moisture content and the target moisture uniformity;

[0103] The term to maximize transpiration satisfaction is expressed as the moisture content prediction uncertainty corresponding to the standard deviation of the predicted volume moisture content, which is used to guide the optimization to search in areas with high uncertainty;

[0104] The irrigation energy consumption and total water volume control term is expressed as a comprehensive index of the irrigation water volume per unit area of farmland and the energy consumption per unit yield, which is used to balance crop yield, water-saving efficiency and energy consumption.

[0105] In this embodiment, the sum of squares of moisture content deviation terms:

[0106]

[0107] Among them, θ target is the target volume moisture content, representing the optimal moisture level required at the current growth stage of the crop, and f1 represents the overall error accumulation of the moisture content deviation from the target value in each valve control area, measuring the irrigation accuracy;

[0108] In this embodiment, the surface runoff risk control term:

[0109]

[0110] Among them, is the component of the spatial gradient in the vertical direction, representing the degree of moisture increase per unit depth; γ runoff is the surface runoff determination threshold, and exceeding this value will trigger the runoff risk, and f2 measures the potential waterlogging or moisture overload risk;

[0111] In this embodiment, the term to maximize transpiration satisfaction:

[0112]

[0113] Among them, e t represents the reference evapotranspiration, representing the theoretical water requirement intensity of the crop; ∈ is a very small positive number used to prevent division by zero; f3 represents the water supply level per unit reference transpiration, and minimizing its negative value is equivalent to the most sufficient water supply;

[0114] In this embodiment, the irrigation energy consumption and total water volume control item:

[0115]

[0116] Among them, α v is the energy consumption coefficient required for irrigating a unit water volume of the v-th valve channel, which is obtained by measuring the field energy consumption; f4 measures the energy consumption intensity during the irrigation process of the total pumping station.

[0117] In this embodiment, S4 includes the following steps:

[0118] S41. Based on the differential evolution multi-objective water volume optimization model and the multi-objective fitness function, initialize the water volume population and execute the generational evolution process. In each generation, use differential mutation and crossover recombination operations to generate candidate water volume individuals, and use the multi-scale gated SIREN water content continuous prediction model as the prediction service to obtain the water content prediction value, spatial gradient, and uncertainty estimation, and accordingly complete the calculation of the multi-objective fitness function and fitness optimization;

[0119] S42. In each generation of the iterative process, update the non-dominated solution set in real time according to the non-dominance judgment principle, and construct the Pareto optimal water volume decision vector set P Pareto , and at the same time record the change trend of each objective function for convergence monitoring;

[0120] S43. When the number of iterations reaches the maximum number of generations G or the average improvement value of the objective function for δ consecutive generations is lower than the convergence threshold ∈, it is determined that the differential evolution multi-objective water volume optimization model has met the convergence conditions, terminate the iteration, and output the final Pareto solution set P Pareto ;

[0121] S44. According to the multi-objective preference weight vector ω = [ω1, ω2, ω3, ω4] set by the user, or combined with the agricultural expert rules of the crop growth stage, select the final optimal farmland irrigation water volume vector p Pareto from the Pareto optimal water volume decision vector set P opt* .

[0122] In this embodiment, S6 includes the following steps:

[0123] S61. Send the optimal farmland irrigation water volume decision vector to the intelligent valve control system through the communication gateway. The intelligent valve control system includes V electromagnetic valve nodes, and each valve node performs quantitative irrigation control according to the water volume t within its own time slice τ ;

[0124] S62. Within each time slice τ t , the intelligent valve control system collects the actual irrigation flow Combined with the time slice length Δτ t Calculate the actual farmland irrigation water volume Compare it with the optimal farmland irrigation water volume If there is a deviation Then trigger the flow feedback correction logic, update the current valve control parameters to offset the deviation, where ∈ q Is the threshold;

[0125] S63. After the irrigation cycle ends, re-collect the farmland multi-modal perception data set D new With the latest farmland multi-modal perception data set D new The measured soil volumetric water content at each spatio-temporal point in Compare with the predicted water content output by calling the multi-scale gated SIREN water content continuous prediction model G Calculate the residual sequence ε j And calculate the mean square error of all residual samples. If the residual mean square error ε MSE > δ model That is, exceeding the model accuracy threshold δ model Then trigger the high-frequency spectral increment retraining process of the multi-scale gated SIREN water content continuous prediction model;

[0126] S64. During the high-frequency spectral increment retraining process, only fine-tune the parameter set of the high-frequency activation channels in the multi-scale gated SIREN water content continuous prediction model And retain the top K frequency components based on the current mutual information sorting structure. The updated frequency matrix is expressed as At the same time, perform a pruning operation to remove low-contribution channels and perform 8-bit weight quantization to generate the updated model parameter set

[0127] S65. Redeploy the updated multi-scale gated SIREN water content continuous prediction model to the field edge computing gateway for the water volume fitness evaluation process in the next round of differential evolution multi-objective water volume optimization model, and implement the continuous closed-loop learning update mechanism for farmland irrigation water volume regulation.

[0128] In this embodiment, S5 also includes extracting the optimal farmland irrigation water volume output classification standard based on the distribution characteristics of the Pareto solution set, the model prediction error distribution, and the crop growth stage in historical iterations, and defining the farmland irrigation water volume recommendation rules under different regulation modes:

[0129] Drought repair type: If the average residual in the past three days is negatively biased and the evaporation amount ≥ 120% of the daily average historical value, then increase the upper limit of the water volume component in each time slice by 10%;

[0130] Water-saving and stable type: If the absolute values of the residuals are all less than the threshold and the prediction uncertainty then, while maintaining the satisfaction constraint of the target transpiration, the total farmland irrigation water volume is compressed by no more than 15%;

[0131] Energy consumption constraint type: If the energy consumption index per unit output is more than 20% higher than the historical average, then some water uniformity indicators are sacrificed according to the target priority, and the optimal solution priority ranking strategy is switched from ω2 > ω3 to ω4 > ω1.

[0132] Example 1:

[0133] At 06:00 am on August 6, 2024, a 9.6-hectare cornfield in Area A entered the third-round irrigation window period. The irrigation target was summer corn from the end of jointing to the initial stage of tasseling. This area was divided into 16 solenoid valve control units, each covering approximately 0.6 hectares. To perform the precise irrigation task, the farm manager enabled the system of the present invention deployed on the LoRa edge gateway in the field.

[0134] At 06:05 am, the system automatically collected the latest data from 64 FDR moisture sensors, 32 tensiometers, 16 infrared canopy temperature monitoring nodes, and 1 meteorological micro-station. In a certain block, the sensor "node-fdr-03" returned the following records: Timestamp: 2024-08-06 06:00; Spatial coordinates: (X = 112.1371, Y = 36.3758); Soil volumetric water content: 0.177 cm 3 / cm 3 ; Soil water potential: -38 kPa; Canopy temperature: 31.2 °C; Surface wind speed: 2.3 m / s; Air humidity: 61%; Solar radiation intensity: 673 W / m 2 .

[0135] After the system completed the data collection of all nodes, anomaly rejection (rejecting the abnormal drift data from "node-tension-11"), unified time alignment, and performed spatial interpolation, it automatically constructed the latest round of perception dataset D at 06:12 final .

[0136] Subsequently, the SIREN moisture content continuous prediction model G started to work, performing high-frequency prediction on the target prediction window (from 06:00 on August 6 to 06:00 on August 7, a total of 24 hours). With the spatial coordinates (112.1371, 36.3758, 0.35), cumulative irrigation depth of 31.4 mm, predicted rainfall of 2.1 mm, and reference evapotranspiration of 6.3 mm as inputs, the predicted value of the model at this point was:

[0137] Moisture content Spatial gradient Prediction uncertainty

[0138] Based on the spatial average results of 179 prediction points in the whole field, the system evaluated that the proportion of the area where the moisture content distribution was below the lower limit of maize water requirement reached 38.2%. At the same time, spatial gradient mutations occurred in multiple regions, indicating an obvious heterogeneous distribution phenomenon of "drought on ridge tops - waterlogging in depressions" inside the field.

[0139] The differential evolution multi-objective water volume optimization module was started. The system set the optimization dimension to 192 (12 time slices × 16 valves), and the objective function weight vector was set as:

[0140] The weight of moisture content deviation ω1 = 0.35; the weight of moisture content spatial gradient ω2 = 0.25; the weight of uncertainty-induced search ω3 = 0.15; the weight of energy consumption per unit yield ω4 = 0.25;

[0141] The system started to execute DE optimization. The initial population size was set to 50, the crossover probability was 0.9, and the maximum number of iteration generations was 80. Each generation of candidate solutions was injected into model G for a complete 24-hour prediction, and the obtained fitness results were used for population evolution. It converged in the 67th generation. The system output 16 groups of Pareto optimal solutions and automatically called the "water-saving priority at the late jointing stage" strategy preset by agronomy experts to select the optimal solution.

[0142] Finally, the system output the optimal water volume vector p opt* , and issued and executed as follows (excerpting the first four valves part):

[0143] Table 1 Valves execute time slices according to the optimal water volume vector

[0144] Valve number Time slice 1 Time slice 2 Time slice 3 ... Time slice 12 V-01 22.0mm 20.0mm 19.5mm ... 0.0mm V-02 14.0mm 18.0mm 18.5mm ... 0.0mm V-03 0.0mm 0.0mm 6.5mm ... 0.0mm V-04 16.5mm 20.0mm 21.0mm ... 1.0mm

[0145] At 08:00, irrigation was officially started, and the system monitored the irrigation execution situation in real time according to the valve status and the flowmeter reading. At 08:27, the system found that the flow deviation of "V-03" in time slice 2 exceeded ±12% (the target was 0, and the actual measurement was 2.4 mm), and automatically executed feedback control to force the water volume in its subsequent time slices to zero.

[0146] The irrigation cycle was completed at 20:00. At 21:00, post-irrigation data was collected. Combining with the model prediction value, the residual mean square error of the whole field was calculated to be 0.026 cm 3 / cm 3 , exceeding the threshold of 0.020. The system automatically triggered the incremental retraining of the SIREN high-frequency channel, updated the model parameters and compressed them into an edge deployment version, which took 11 minutes.

[0147] Table 2 shows the comparison results between the method of the present invention and the traditional tensiometer threshold irrigation method

[0148] Index Method of the present invention Traditional tensiometer threshold irrigation method Total water volume for single irrigation <![CDATA[14160m 3 > <![CDATA[17720m 3 > Average corn yield 661 kg / mu 612 kg / mu Spatial standard deviation of water content 0.0098 0.0175 Average prediction error of water content 0.021 No modeling Average deviation of valve-controlled flow execution ±4.1％ Unmonitorable Model update period Automatically after each irrigation No update mechanism Optimized calculation time 18 minutes No optimization module

[0149] Example 1 proves that in the context of significant spatial heterogeneity of fields and complex water distribution, the present invention can effectively avoid the problem of "partial over-irrigation and partial under-irrigation", achieve stable crop yields, save water, and reduce the energy consumption per unit of output, and has significant practicality and engineering promotion value.

[0150] Based on the traditional SIREN network, the present invention introduces a multi-scale hierarchical structure and an irrigation-meteorological gating mechanism to construct a continuous water content prediction model that can hierarchically perceive the response characteristics of different soil depths. By integrating the control residuals of the Richards equation as a physical consistency term during the model training process, the model can not only perform high-frequency reconstruction in the spatio-temporal coordinate system but also has physical interpretability, effectively avoiding prediction drift. Compared with traditional neural networks without prior knowledge, the proposed model has a decrease in the mean square error of water content prediction in heterogeneous soil scenarios, and at the same time, the spatial gradient continuity is significantly improved, and it can be used as a stable fitness evaluator for optimization algorithms.

[0151] The present invention designs a differential evolution optimization structure with the prediction result as the feedback closed-loop, injects the irrigation water volume decision vector into the SIREN model, dynamically obtains the soil water content prediction, spatial gradient, and prediction variance corresponding to each individual, and constructs a multi-objective fitness function that integrates four objectives. By means of adjustable weights, it meets the multiple optimization requirements under different crop and climate conditions, significantly improving the scientificity and flexibility of irrigation decision-making.

[0152] The present invention designs a closed-loop feedback control structure with an intelligent valve control system as the execution body. During the actual irrigation cycle, it collects the water volume execution results of the solenoid valve and combines the latest field perception data to calculate the prediction residuals of the SIREN model in real time. When the mean square of the residuals exceeds the threshold, it automatically triggers an incremental retraining mechanism centered on the high-frequency channel to ensure that the model still maintains stable performance in the face of extreme climate changes or sudden changes in soil conditions. Combined with pruning and quantization compression strategies, the updated model can be redeployed in the MCU-level edge gateway to ensure the continuous online nature and lightweight efficiency of the field prediction and control process.

[0153] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. An intelligent optimization method for farmland irrigation water volume based on deep learning, characterized in that, It includes the following steps: S1. Periodically collect the multi-modal perception dataset of farmland and perform preprocessing to obtain the preprocessed multi-modal perception dataset of farmland; S2. Taking the preprocessed multi-modal perception dataset of farmland as the input, construct and train a multi-scale gated SIREN continuous moisture content prediction model to obtain the SIREN model parameters that can output the three-dimensional space-time continuous moisture content distribution and spatial gradient; S3. Divide the planned irrigation period into several time slices, define the water volume decision vector space and construct a differential evolution multi-objective water volume optimization model. Each component of the water volume decision vector corresponds to the farmland irrigation water volume of a time slice-valve combination. In the differential evolution multi-objective water volume optimization model, construct a multi-objective fitness function based on the multi-scale gated SIREN continuous moisture content prediction model; S4. Perform population initialization, differential mutation, crossover recombination and fitness selection on the differential evolution multi-objective water volume optimization model. In each fitness evaluation process, inject the candidate water volume decision vector as the condition vector into the SIREN continuous moisture content prediction service to obtain the continuous moisture content distribution and spatial gradient and calculate the multi-objective fitness. Iteratively update until the termination condition is met, and output the Pareto optimal water volume decision vector set that meets the multi-objective constraint conditions, and select the optimal farmland irrigation water volume from it according to the user weight or preset rules; S5. Send the optimal farmland irrigation water volume to the intelligent valve control system to drive the solenoid valve to perform irrigation according to the time slice.

2. The intelligent optimization method for farmland irrigation water volume based on deep learning according to claim 1, wherein The S1 includes the following steps: S11. Deploy sensing devices for periodically collecting field environmental factors in the irrigation area. The sensing devices include tensiometers for collecting soil water potential, frequency domain reflectometry moisture sensors for collecting soil moisture content, near-infrared imaging devices for collecting canopy temperature, and micro-meteorological stations for collecting meteorological factors. Sample environmental information in the irrigation area at a unified time interval to construct a raw dataset D of multi-modal perception of farmland containing N perception data samples d i ; raw The raw dataset of multi-modal perception of farmland includes the sampling timestamp t i , the two-dimensional spatial coordinates x i , y i of the sensing point, the soil water potential ψ i , the soil volumetric water content θ i , the canopy temperature T i and the meteorological factor vector ξ composed of air temperature, humidity, wind speed and solar radiation i ; S12. Perform time synchronization processing on the original multi-modal perception dataset of farmland according to the preset time window, align the asynchronous sampling data of each type to the unified time benchmark, smoothly complete the data at the missing time points by using the time interpolation method, and map the sampling data to the regular grid coordinates according to the spatial interpolation grid resolution. Use the spatial interpolation algorithm to estimate the perception values at the missing spatial positions to generate the multi-modal spatial completion dataset of farmland; S13. Perform anomaly rejection processing on the farmland multi-modal spatial completion dataset, calculate the deviation ratio of each data component, determine the data as abnormal and reject it when the deviation ratio is greater than the preset deviation threshold, and convert all data items in the farmland multi-modal anomaly rejection dataset into a unified physical dimension to complete the unit unification process, obtaining the preprocessed farmland multi-modal perception dataset D final 。 3. The intelligent optimization method for farmland irrigation water volume based on deep learning according to claim 2, characterized in that, The S2 includes the following steps: S21. Append the synchronous irrigation depth, the future 24-hour rainfall forecast volume r and the reference evapotranspiration e j to each sample of the preprocessed farmland multi-modal perception dataset D final , and construct the training sample set S train ; final j j train ​​​​​ S22. For each group of plane coordinates x j ,y j and depth coordinate z j The space-time coordinate (x j ,y j ,z j ,t j ) performs Fourier feature mapping to generate a high-frequency vector φ j ; S23. Based on the high-frequency vector φ j , construct a multi-scale gated SIREN continuous water content prediction model: Among them, represents a multi-scale gated SIREN moisture content continuous prediction model for predicting the volumetric water content in the root zone of crops, and α l (p cum , r, e) = σ(a l p cum + b l r + c l e + d l ) is the gating factor, and σ(·) is the Sigmoid function. is the set of model parameters to be trained. represents the sub-network constructed for the l-th soil layer, which is used to model the response characteristics of the moisture content in different soil layers of 0–20 cm, 20–40 cm, and 40–60 cm. S24. Using the soil volumetric water content θ j as the supervision signal, train the multi-scale gated SIREN water content continuous prediction model to minimize the model loss function where λ is the physical constraint regularization weight, K(θ j ) is the hydraulic conductivity function corresponding to the volumetric soil water content, M represents the total number of samples used to train the multi-scale gated SIREN water content continuous prediction model, represents the time change rate of the volumetric soil water content at the j-th sample, represents the divergence term of the flow controlled by soil water potential, and the second term is the physical consistency residual term, which is used to evaluate whether the predicted value conforms to the Richards soil water migration control equation; After the training of the multi-scale gated SIREN moisture content continuous prediction model converges, sort based on the mutual information value I(θ, E) between the predicted result of soil volumetric water content and the energy consumption per unit yield E, and retain the top K frequency components in the first-layer frequency matrix B to obtain the pruned frequency matrix B K , and update the model parameters to the compressed model parameter set S26. Quantize the model parameter set with 8-bit weight quantization and insert the MC-Dropout structure to generate the final multi-scale gated SIREN water content continuous prediction model G.

4. The intelligent optimization method for farmland irrigation water volume based on deep learning according to claim 1, characterized in that, The S3 includes the following steps: S31. Divide the planned farmland irrigation period T irrig into T non-overlapping time slice sets {τ1, τ2, …, τ T}, and the duration of each time slice is Δτ t , such that each time slice τ t corresponds to an electromagnetic valve control operation for performing phased farmland irrigation tasks; S32. Set the number of independent electromagnetic valves deployed within the farmland irrigation area to V. Each time slice can control V independent farmland irrigation channels, and construct a decision vector for the farmland irrigation water volume wherein, represents the farmland irrigation water volume of the v-th electromagnetic valve within the time slice τ t ; S33. Define upper and lower boundary limits for each electromagnetic valve - time slice combination to satisfy the farmland irrigation water volume of the v-th electromagnetic valve within the time slice τ t and the minimum farmland irrigation water volume of the v-th electromagnetic valve within the time slice τ t is between the minimum farmland irrigation water volume and the maximum farmland irrigation water volume . The upper and lower boundary limits are set according to the pumping station water supply capacity, crop waterlogging tolerance, and soil infiltration threshold; S34. Construct a differential evolution multi-objective water volume optimization model, and use the farmland irrigation water volume decision vector p opt as the coding structure of each individual in the differential evolution multi-objective water volume optimization model, where the farmland irrigation water volume decision vector p opt contains combinations of T time slices and V valves. Each dimension represents the farmland irrigation water volume of a certain time slice-valve combination. In the differential evolution population, all individuals are initialized within the water volume boundary constraint interval to form an initial population. The number of individuals in the initial population is N, and the search space is jointly defined by the upper and lower limits of all T×V groups of valve-controlled water volumes; S35. In the differential evolution multi-objective water volume optimization model, based on the individuals in each generation of the population, a mutant vector is generated through differential mutation operation. The differential mutation operation is constructed depending on the water volume differences among three different individuals. After mutation, a crossover operation based on the set crossover probability C is performed on each individual. r In the crossover operation, each dimension of the candidate water volume decision vector randomly selects its source between the original individual and the mutant vector, and at least one dimension comes from the mutant vector. After completing the crossover operation, a child water volume decision vector is formed and enters the next round of fitness evaluation. S36. In the fitness evaluation step, call the multi-scale gated SIREN continuous soil moisture prediction model G, using the farmland irrigation water volume corresponding to each time slice-valve combination in the water volume decision vector p opt as one of the model inputs, and at the same time, combine the spatial-temporal position, cumulative irrigation depth, reference evapotranspiration, and rainfall forecast volume to obtain the predicted volumetric soil moisture content under this water volume configuration The spatial gradient of the predicted volumetric soil moisture content and the standard deviation of the predicted volumetric soil moisture content Couple the predicted volumetric soil moisture content The spatial gradient of the predicted volumetric soil moisture content and the standard deviation of the predicted volumetric soil moisture content with four optimization objectives to construct a multi-objective fitness function F(p opt ): F(p opt ) = ω1f1 + ω2f2 + ω3f3 + ω4f4; Among them, the multi-objective preference weight vector ω = [ω1, ω2, ω3, ω4], f1 is the sum of squares of moisture content deviation terms, f2 is the surface runoff risk control term, f3 is the term for maximizing transpiration satisfaction, and f4 is the irrigation energy consumption and total water volume control term.

5. The intelligent optimization method for farmland irrigation water volume based on deep learning according to claim 4, characterized in that, The sum of squares of moisture content deviation terms is expressed as the sum of squares of the deviation between the target volume moisture content and the predicted volume moisture content; The surface runoff risk control term is expressed as the degree of deviation between the spatial gradient of the predicted volume moisture content and the target moisture uniformity; The term for maximizing transpiration satisfaction is expressed as the moisture content prediction uncertainty corresponding to the standard deviation of the predicted volume moisture content, which is used to guide the optimization to search in the high-uncertainty area; The irrigation energy consumption and total water volume control term is expressed as a comprehensive index of the farmland irrigation water volume per unit area and the energy consumption per unit yield, which is used to balance crop yield, water saving efficiency and energy consumption.

6. The intelligent optimization method for farmland irrigation water volume based on deep learning according to claim 4, characterized in that The S4 includes the following steps: S41. Initialize the water volume population and execute the generational evolution process based on the differential evolution multi-objective water volume optimization model and the multi-objective fitness function. In each generation, generate candidate water volume individuals using differential mutation and crossover recombination operations, and use the multi-scale gated SIREN water content continuous prediction model as a prediction service to obtain water content prediction values, spatial gradients, and uncertainty estimates, and accordingly complete the calculation of the multi-objective fitness function and fitness optimization; S42. In each iteration, the non-dominated solution set is updated in real time according to the non-dominance judgment principle, and the Pareto optimal water volume decision vector set P is constructed. Pareto Meanwhile, the changing trends of each objective function are recorded for convergence monitoring. S43. When the number of iterations reaches the maximum generation G or the average improvement value of the objective function for consecutive δ generations is lower than the convergence threshold ∈, it is determined that the differential evolution multi-objective water volume optimization model has met the convergence condition, the iteration is terminated, and the final Pareto solution set P is output Pareto ; S44. Select the final optimal farmland irrigation water volume vector p from the Pareto optimal water volume decision vector set P according to the multi-objective preference weight vector ω = [ω1, ω2, ω3, ω4] set by the user, or in combination with the agricultural expert rules of the crop growth stage Pareto opt* .​ 7. An intelligent optimization method for farmland irrigation water volume based on deep learning according to claim 6, characterized in that, The said S6 includes the following steps: S61. Transmit the optimal farmland irrigation water volume decision vector to the intelligent valve control system through the communication gateway. The intelligent valve control system includes V electromagnetic valve nodes, and each valve node performs quantitative irrigation control according to the water volume t within its respective time slice τ ; S62. During each time slice τ t the intelligent valve control system collects the actual irrigation flow rate and combines it with the time slice length Δτ t to calculate the actual farmland irrigation water volume Compare it with the optimal farmland irrigation water volume If there is a deviation then trigger the flow feedback correction logic to update the current valve control parameters to offset the deviation, where ∈ q is the threshold value; After the irrigation cycle ends, re-collect the multi-modal perception dataset D of the farmland new , and use the latest multi-modal perception dataset D of the farmland new to compare the measured soil volumetric water content at each spatio-temporal point with the predicted water content output by calling the multi-scale gated SIREN water content continuous prediction model G to calculate the residual sequence ε j , and calculate the mean square error of all residual samples. If the residual mean square error ε MSE >δ model , that is, exceeding the model accuracy threshold δ model , then trigger the high-frequency spectral increment retraining process of the multi-scale gated SIREN water content continuous prediction model; S64. During the high-frequency spectral incremental retraining process, only the parameter set of the high-frequency activation channels in the multi-scale gated SIREN water content continuous prediction model is fine-tuned, and the top K frequency components are retained based on the current mutual information sorting structure. The updated frequency matrix is expressed as At the same time, a pruning operation is performed to remove low-contribution channels, and 8-bit weight quantization is executed to generate an updated model parameter set S65. Redeploy the updated multi-scale gated SIREN water content continuous prediction model to the field edge computing gateway for the water volume fitness evaluation process in the next round of the differential evolution multi-objective water volume optimization model, and implement a continuous closed-loop learning and updating mechanism for farmland irrigation water volume regulation.

8. The intelligent optimization method for farmland irrigation water volume based on deep learning according to claim 7, characterized in that The said S5 also includes extracting the optimal farmland irrigation water volume output classification criteria based on the distribution characteristics of the Pareto solution set, the model prediction error distribution, and the crop growth stage in historical iterations, and defining the farmland irrigation water volume recommendation rules under different regulation modes: Drought repair type: If the average residual in the past three days is negatively biased and the evaporation is ≥ 120% of the daily historical value, then increase the upper limit of the water volume component in each time slice by 10%; Water-saving and stable type: If the absolute values of the residuals are all less than the threshold and the prediction uncertainty then, on the premise of maintaining the satisfaction constraint of the target transpiration, the total farmland irrigation water volume is compressed by no more than 15%; Energy consumption constraint type: If the energy consumption index per unit output is more than 20% higher than the historical average, then sacrifice some water uniformity indicators according to the target priority, and switch the optimal solution priority sorting strategy from ω2>ω3 to ω4>ω1.

Images