Crop efficient water irrigation opportunity prediction method based on machine learning

Abstract

The invention discloses a crop efficient water irrigation opportunity prediction method based on machine learning, relates to the technical field of intelligent agriculture and precise irrigation, and aims to solve the problem that a static irrigation decision cannot respond to spatial-temporal dynamic change of soil moisture. According to the method, rapid response to weather and crop water demand disturbance is realized, multi-depth water content and meteorological and phenological data are fused, a water stress index is calculated, and time-space correlation characteristics are constructed, so that a water deficit area and a key water demand time period are accurately identified. On the basis, in combination with phenological period sensing feature scaling and Pareto optimization, an efficient strategy for optimizing irrigation water quantity, energy consumption and yield risks is synchronously generated, so that resource waste caused by decision lag and global rearrangement is avoided, and finally, irrigation equipment is linked to form a closed-loop execution process, so that the irrigation efficiency is improved. Therefore, the irrigation accuracy, the self-adaptive capability and the resource utilization efficiency are obviously improved in a changeable farmland environment.

Description

Machine Learning-Based Method for Predicting the Timing of Efficient Irrigation for Crops Technical Field

[0001] This invention relates to the field of smart agriculture and precision irrigation technology, and more specifically, to a method for predicting the timing of efficient irrigation for crops based on machine learning. Background Technology

[0002] In intelligent irrigation decision-making scenarios based on soil volumetric moisture content, traditional methods usually rely on static threshold judgment of surface or single-point moisture content, or simply perform average and independent tomography processing on multi-depth moisture content data. However, these methods fail to effectively analyze the spatiotemporal dynamic transport patterns of soil moisture in the vertical and horizontal directions, and cannot accurately capture the synergistic evolution mechanism between it and meteorological elements and crop physiological water consumption.

[0003] In existing technologies, the timing of irrigation prediction fails to upgrade discrete, multi-level water content time-series data into a unified, continuous spatiotemporal field. At the same time, it lacks the ability to quantitatively model key water transport physical processes. This directly leads to the difficulty for prediction models to extract deep spatiotemporal correlation features from water content data that can truly and comprehensively reflect the crop's water stress state. They can only passively respond to the lagging changes in surface soil moisture and cannot make a forward-looking judgment on the dynamic surplus and deficit trend of effective water in the crop root zone. This inevitably causes lag in irrigation timing and deviation in irrigation water estimation, which not only leads to low water resource utilization efficiency but may also trigger the risk of stage-specific water stress in crops, affecting the efficiency of farmland production. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a machine learning-based method for predicting the timing of efficient irrigation for crops to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based method for predicting the timing of efficient irrigation for crops, comprising the following steps: In a preferred embodiment, time-series data of soil volumetric water content, meteorological data, crop phenological stage data, and weather forecast data for a preset future period are collected from different depths in farmland. After spatiotemporal alignment and cleaning, a fused dataset is generated and the current crop phenological stage is identified. Specific steps include: periodically collecting soil volumetric water content time-series data with spatial location information at different depths using soil moisture sensors, meteorological monitoring equipment, and crop growth recording equipment deployed in farmland. Data, meteorological data, and crop phenological data are collected, and weather forecast data for a preset future period is obtained through a weather forecast data service interface. Soil volumetric moisture content time-series data, meteorological data, crop phenological data, and weather forecast data are timestamped using a unified time base, and spatial location information is also uniformly processed for spatial alignment, thereby generating an aligned dataset. The aligned dataset is then cleaned to remove outliers and fill in missing data, and fused according to the unified time base to generate a fused dataset. Based on the phenological time stamps of the fused dataset and the crop phenological data, the current crop phenological stage is identified.

[0006] In a preferred embodiment, a soil moisture spatiotemporal map is constructed based on a fused dataset. Spatiotemporal correlation analysis is performed on the soil moisture spatiotemporal map to extract dynamic spatiotemporal feature vectors encompassing the vertical movement and lateral migration patterns of water. Specific steps include: based on the fused dataset, mapping time-series data of soil volumetric water content at different depths to three-dimensional grid nodes according to spatial location information to construct a soil moisture spatiotemporal map; performing spatiotemporal correlation analysis on the soil moisture spatiotemporal map, quantifying the vertical gradient of water content between adjacent three-dimensional grid nodes in the vertical direction as vertical water diffusion flux, and quantifying the water potential gradient between adjacent three-dimensional grid nodes in the horizontal direction as lateral water potential energy gradient; based on the vertical water diffusion flux and lateral water potential energy gradient, simultaneously generating vertical transport features representing the vertical movement pattern of water and lateral transport features representing the lateral migration pattern of water at each time slice of the soil moisture spatiotemporal map; stacking and integrating the vertical transport features and lateral transport features corresponding to each time slice to construct a three-dimensional tensor, and using the three-dimensional tensor as a dynamic spatiotemporal feature vector encompassing the vertical movement and lateral migration patterns of water.

[0007] In a preferred embodiment, dynamic spatiotemporal feature vectors are fused with meteorological data, crop phenological data, and weather forecast data at the feature level to construct an original feature set. The original feature set is then adaptively scaled based on the current phenological period to generate a phenological period-sensing feature set. Specific steps include: determining the types of water transport features to be extracted based on the current crop phenological period, and extracting corresponding vertical transport or lateral migration features from the dynamic spatiotemporal feature vector as phenological period-sensitive water features; fusing the phenological period-sensitive water features, meteorological data, crop phenological data, and weather forecast data according to a unified time benchmark to construct an original feature set containing different feature dimensions; and determining the adaptive scaling ratio of each feature dimension in the original feature set based on the physiological stage characteristics of the current crop phenological period, and adaptively scaling the original feature set to generate the phenological period-sensing feature set.

[0008] In a preferred embodiment, based on the physiological stage characteristics of the current crop phenological stage, an adaptive scaling ratio for each feature dimension in the original feature set is determined, and the original feature set is adaptively scaled to generate a phenological stage perception feature set. Specific steps include: determining the sensitivity level of the current crop phenological stage to water stress based on its physiological stage characteristics; determining the initial scaling coefficient for each feature dimension in the original feature set based on the deviation between the sensitivity level and a preset water demand baseline value; generating a dynamic meteorological correction factor by combining weather forecast data, meteorological data, and the dynamic changing trend of phenological stage sensitive water characteristics; inputting the dynamic meteorological correction factor into a preset scaling ratio mapping relationship and outputting an adjustment amount to the initial scaling coefficient; correcting the initial scaling coefficient based on the adjustment amount to generate an adaptive scaling ratio for each feature dimension; and performing adaptive feature scaling calculations on all feature dimensions of the original feature set based on the adaptive scaling ratio to generate the phenological stage perception feature set.

[0009] In a preferred embodiment, the phenological period sensing feature set is input into a pre-trained irrigation decision model, and the model outputs preliminary irrigation demand probabilities and irrigation volume reference values ​​for different time nodes within a future preset period. Specific steps include: inputting the phenological period sensing feature set into the pre-trained irrigation decision model; performing forward inference calculations of the irrigation decision model to generate initial demand probability values ​​and initial irrigation volume values ​​for each time node within the corresponding future preset period; extracting feature dimensions from meteorological data and weather forecast data from the phenological period sensing feature set; and calculating the feature dimensions for consecutive preset periods before the current time node, using the current time node as a baseline. The changes in gradient and fluctuation frequency within the meteorological time series are used to generate meteorological time series evolution characteristics. Based on these characteristics, the initial demand probability value is smoothed using a sliding window to generate preliminary irrigation demand probabilities for each time node within a future preset period. Based on the current crop phenological stage and meteorological time series evolution characteristics, a decision threshold range is dynamically determined. The preliminary irrigation demand probability is compared with the decision threshold range for the corresponding time node. Based on the decision threshold range into which the preliminary irrigation demand probability falls, a preset nonlinear mapping function is selected to perform segmented correction on the initial irrigation amount value, outputting reference values ​​for irrigation amount at different time nodes within a future preset period.

[0010] In a preferred embodiment, the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk are calculated based on the initial irrigation demand probability, irrigation volume reference value, and preset candidate time step. Specific steps include: generating different candidate irrigation time nodes within a preset future time period based on the initial irrigation demand probability and preset candidate time step; calculating the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk for each irrigation time node based on the irrigation volume reference value; and in a preferred embodiment, constructing and solving a multi-objective function to generate a Pareto optimal irrigation strategy sequence, and scheduling irrigation equipment according to the Pareto optimal irrigation strategy sequence. Specific steps include: using the expected value of water use efficiency... With the optimization goals of maximizing, minimizing expected irrigation energy consumption, and minimizing expected crop yield risk, a multi-objective optimization function is constructed, constrained by a preset upper limit for total irrigation volume and a feasible interval for irrigation timing, forming a multi-objective optimization problem. A multi-objective optimization algorithm is used to iteratively solve the multi-objective optimization problem, generating a Pareto optimal solution set. Each solution in the Pareto optimal solution set corresponds to an irrigation time node and a corresponding reference value for irrigation volume. Each solution in the Pareto optimal solution set is defined as an irrigation strategy, forming a Pareto optimal irrigation strategy sequence. Target irrigation strategies are selected from the Pareto optimal irrigation strategy sequence according to preset decision preference rules. The irrigation time nodes and reference values ​​for irrigation volume in the target irrigation strategies are converted into irrigation control commands, and the irrigation equipment is scheduled to execute them.

[0011] In a preferred embodiment, a target irrigation strategy is selected from the Pareto optimal irrigation strategy sequence according to a preset decision preference rule. The specific steps include: obtaining the preference weights of the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk based on the preset decision preference rule, and constructing a decision function based on the preference weights; for each irrigation strategy in the Pareto optimal irrigation strategy sequence, inputting the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk corresponding to each irrigation strategy into the decision function, calculating the comprehensive evaluation value of each irrigation strategy, comparing the magnitude of the comprehensive evaluation values ​​of all irrigation strategies, identifying the largest comprehensive evaluation value, and selecting the corresponding irrigation strategy as the target irrigation strategy.

[0012] Compared with existing technologies, this invention has the following beneficial effects: 1. By constructing a spatiotemporal map of soil moisture and performing spatiotemporal correlation analysis, it achieves dynamic analysis and collaborative perception of the vertical movement and lateral migration patterns of farmland water. Based on adaptive feature scaling of phenological period perception, it dynamically highlights the water stress signals most relevant to the current crop growth stage, and combines meteorological time series evolution characteristics to smoothly correct the probability of irrigation demand, thereby forming a precise irrigation demand assessment model based on multi-source information fusion. Specifically, compared with existing methods using static thresholds or single soil moisture content criteria, this invention can accurately capture the spatiotemporal heterogeneity of soil moisture during the crop water-sensitive period. Through the synergistic effect of dynamic decision threshold intervals and nonlinear mapping functions, it significantly improves the prediction accuracy of irrigation timing and irrigation amount, and effectively suppresses decision misjudgments caused by environmental fluctuations or feature distortion.

[0013] 2. By introducing a multi-objective optimization framework centered on water use efficiency, irrigation energy consumption, and crop yield risk, this invention resolves the inherent contradiction in traditional single-objective irrigation decision-making where resource conservation and stable crop yield are difficult to coordinate. Secondly, based on the Pareto optimal irrigation strategy sequence generated from preliminary demand probability and irrigation volume reference values, it achieves dynamic trade-offs and global optimization of multi-dimensional objectives at different candidate time steps. Then, strategy selection is performed through preset decision preference rules to ensure that the final irrigation scheme, while meeting total volume constraints and timing feasibility, always matches the user's actual management objectives. This avoids resource waste or crop stress risks caused by a single objective. Therefore, while maintaining crop water demand satisfaction, this invention significantly improves the comprehensive benefits and resilience of the irrigation system, ensuring sustainable and refined management of farmland under the background of climate change and water resource constraints. Attached Figure Description

[0014] Figure 1 is a schematic diagram of the structure of the machine learning-based method for predicting the timing of efficient irrigation for crops according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Example 1: Figure 1 shows a schematic diagram of the structure of the efficient irrigation timing prediction method for crops based on machine learning according to the present invention, which includes the following steps: S1, collecting time-series data of soil volumetric water content, meteorological data, crop phenological data, and weather forecast data for a preset period of time deployed at different depths in farmland, and generating a fused dataset after spatiotemporal alignment and cleaning, and identifying the current crop phenological stage; S2, constructing a soil moisture spatiotemporal map based on the fused dataset, performing spatiotemporal correlation analysis on the soil moisture spatiotemporal map, and extracting dynamic spatiotemporal feature vectors covering the vertical movement and lateral migration patterns of water; S3, integrating the dynamic spatiotemporal feature vectors with meteorological data, crop phenological data, and weather forecast data. The data is fused at the feature level to construct an original feature set. Based on the current phenological period, the original feature set is adaptively scaled to generate a phenological period sensing feature set. S4. The phenological period sensing feature set is input into a pre-trained irrigation decision model to output the preliminary irrigation demand probability and irrigation volume reference value at different time nodes within the future preset time period. S5. Based on the preliminary irrigation demand probability, irrigation volume reference value and preset candidate time step, the expected value of water use efficiency, irrigation energy consumption and crop yield risk are calculated. A multi-objective function is constructed and solved to generate a Pareto optimal irrigation strategy sequence. Irrigation equipment is scheduled according to the Pareto optimal irrigation strategy sequence.

[0017] The process involves collecting time-series data on soil volumetric moisture content, meteorological data, crop phenological stage data, and future weather forecast data at different depths in farmland. After spatiotemporal alignment and cleaning, a fused dataset is generated and the current crop phenological stage is identified. Specifically, multiple soil moisture sensors, meteorological monitoring equipment, and crop growth recording devices deployed in the farmland synchronously collect time-series data on soil volumetric moisture content, meteorological data, and crop phenological stage data at different depths, along with spatial location information of their installation points, at fixed intervals. For example, soil moisture sensors are buried at depths of 10 cm, 30 cm, and 50 cm below the surface; meteorological monitoring equipment includes, for example, small weather stations; and crop growth recording devices are installed at depths of 10 cm, 30 cm, and 50 cm below the surface. Long-term recording devices include, for example, fixed-point cameras with image recognition capabilities or manual recording terminals. The fixed collection period is, for example, every 30 minutes. Meteorological data includes temperature, humidity, wind speed, and solar radiation intensity. Crop phenological data is used to identify key growth stages of crops, such as sowing period, seedling stage, jointing stage, heading stage, and maturity stage. At the same time, by calling the weather forecast data service interface provided by the Central Meteorological Observatory or commercial meteorological data service providers, weather forecast data for a future preset time period is obtained at the same period or daily timed. The future preset time period is, for example, the next 24 hours. The weather forecast data includes the predicted future temperature, precipitation probability, wind speed, and sunshine duration corresponding to the location of the farmland.

[0018] The collected soil volumetric moisture content time-series data, meteorological data, crop phenological data, and weather forecast data are time-stamp aligned based on a unified time reference. This involves converting the soil volumetric moisture content time-series data, meteorological data, crop phenological data, and weather forecast data to a unified Coordinated Universal Time (UTC) timestamp. Then, the spatial location information carried by the soil volumetric moisture content time-series data, meteorological data, crop phenological data, and weather forecast data is unified by using the latitude and longitude coordinates of the center point of the farmland as a unified coordinate reference. All data carrying spatial location information are spatially mapped and aligned, thereby generating an aligned dataset with timestamp and spatial location alignment.

[0019] Statistical methods, such as the 3σ principle, are used to identify and remove outliers from the aligned dataset. For missing data caused by temporary equipment failures, time-series linear interpolation or spatial interpolation based on adjacent sensor data is used to fill in the missing data. After cleaning, the dataset is fused according to a unified time reference. Specifically, the aligned dataset is integrated into the same data table according to a unified Coordinated Universal Time (UTC) timestamp. Each row corresponds to a unified sampling time node defined by the UTC timestamp, and each column corresponds to time-series data of soil volumetric water content at different depths, meteorological data, crop phenological data, and weather forecast data, thereby generating a structured fused dataset.

[0020] The system reads crop phenological data from the fusion dataset, which is recorded by crop growth recording devices or manually input. This crop phenological data contains time markers for each phenological stage, such as the jointing stage. The system compares the current time or the time of the last data collection with these time markers to identify the current crop phenological stage. For example, if the current time is between the jointing stage time marker and the next phenological stage heading stage time marker, then the current crop phenological stage is identified as the jointing stage.

[0021] Based on the fused dataset, a soil moisture spatiotemporal map is constructed. Spatiotemporal correlation analysis is performed on the soil moisture spatiotemporal map to extract dynamic spatiotemporal feature vectors covering the vertical movement and lateral migration patterns of water. Specifically, the implementation is as follows: Based on unified spatial location information, a three-dimensional grid covering the target farmland area is defined. On the horizontal plane, the three-dimensional grid is divided into regular cells with the center point of the farmland as the origin and a preset resolution, such as 1 meter by 1 meter. In the vertical direction, based on the actual burial depth of the soil moisture sensors and combined with the characteristics of the soil profile, several depth layers representing different soil layers are defined. Then, the time series data of soil volumetric water content corresponding to different depths in the fused dataset are mapped to each three-dimensional grid node of the three-dimensional grid through spatial interpolation. Each three-dimensional grid node is uniquely determined by its horizontal coordinates and vertical depth. For any three-dimensional grid node, the time series data of soil volumetric water content of adjacent soil moisture sensors are obtained, and then the water content of the three-dimensional grid node is calculated using the inverse distance weighted interpolation method.

[0022] The core of the inverse distance weighted interpolation method is to perform weighted averaging based on the principle that the closer the distance, the greater the influence. Specifically, the spatial straight-line distance from each adjacent soil moisture sensor to the 3D grid node to be determined is calculated. Each soil moisture sensor collects time-series data of soil volumetric water content. The reciprocal of the square of each spatial straight-line distance is used as the basic weight corresponding to the measured value of the soil volumetric water content at the current moment in the time-series data. All basic weights are normalized to obtain the final weight coefficient of the measured value of the soil volumetric water content at the current moment in the time-series data. The measured value of the soil volumetric water content at the current moment is multiplied by the corresponding final weight coefficient, and all products are summed. The sum is the water content of the 3D grid node. For each unified timestamp, the water content of all 3D grid nodes in the entire 3D grid at that moment can be calculated, thereby generating a water content grid map reflecting the 3D spatial distribution of soil moisture at that moment. Finally, the water content grid maps generated by each timestamp are integrated to construct a soil moisture spatiotemporal map.

[0023] Spatiotemporal correlation analysis of soil moisture maps is performed as follows: At any 3D grid node in the 3D grid, the vertical gradient of water content between the two adjacent 3D grid nodes above and below it in the vertical direction (i.e., depth) is calculated. This vertical gradient of water content is quantified as the vertical water diffusion flux at the 3D grid node, which characterizes the intensity of water migration in the soil profile. At the same time, the water potential gradient between the 3D grid node and its four adjacent 3D grid nodes to the east, west, south, and north in the horizontal direction, based on water content conversion, is calculated. This water potential gradient is quantified as the lateral water potential energy gradient that drives the spatial redistribution of water in the horizontal direction.

[0024] Based on the vertical water diffusion flux and lateral water potential energy gradient of all three-dimensional grid nodes, the following operations are performed synchronously on each time slice defined by the unified and coordinated universal time timestamp corresponding to the soil moisture spatiotemporal map: the vertical water diffusion flux of all three-dimensional grid nodes is organized into a two-dimensional matrix corresponding to the horizontal and vertical structure of the three-dimensional grid. This two-dimensional matrix is ​​defined as the vertical transport feature characterizing the overall law of vertical movement of soil moisture under this time slice. At the same time, the lateral water potential energy gradient of all three-dimensional grid nodes is organized into another two-dimensional matrix with the same structure. This two-dimensional matrix is ​​defined as the lateral transport feature characterizing the law of lateral migration of soil moisture under this time slice.

[0025] The vertical water diffusion flux and lateral water potential energy gradient of each 3D grid node in the 3D spatial grid are divided into continuous time slices in chronological order. For each time slice, its vertical and lateral transport features are extracted. Then, the vertical and lateral transport features corresponding to each time slice arranged in chronological order are combined and sequentially stacked and integrated along the third dimension of the time order to finally construct a 3D tensor. The 3D tensor is used as a dynamic spatiotemporal feature vector that covers the laws of vertical water movement and lateral migration.

[0026] The dynamic spatiotemporal feature vector is fused with meteorological data, crop phenological data, and weather forecast data at the feature level to construct an original feature set. Based on the current phenological period, the original feature set is adaptively scaled to generate a phenological period sensing feature set. Specifically, based on the identified current crop phenological period, the type of water transport feature to be extracted is determined, and the corresponding vertical transport feature or lateral migration feature is extracted from the dynamic spatiotemporal feature vector as phenological period sensitive water features. Specifically, if the current crop phenological stage is the seedling stage or jointing stage dominated by root development, then the type of water transport feature to be extracted is determined to be the vertical transport feature, which characterizes the intensity of water migration in the soil profile. If the current crop phenological stage is the heading stage or grain-filling stage, which is key to canopy transpiration and lateral water allocation, then the type of water transport feature to be extracted is determined to be the lateral transport feature, which drives the spatial redistribution of water in the horizontal direction. Subsequently, from the three-dimensional tensor representing the dynamic spatiotemporal feature vector, a two-dimensional matrix of vertical transport features composed of the vertical water diffusion flux of all three-dimensional grid nodes or a two-dimensional matrix of lateral transport features composed of the lateral water potential energy gradient of all three-dimensional grid nodes is extracted from the corresponding time slice. The extracted two-dimensional matrix is ​​flattened and converted into a one-dimensional feature vector. This one-dimensional feature vector is defined as the phenological stage sensitive water feature.

[0027] Phenological water-sensitive characteristics, meteorological data, crop phenological data, and weather forecast data are fused using a unified time base to construct an original feature set containing different feature dimensions. Specifically, the Coordinated Universal Time (UTC) timestamp is used as the unified time base and unique association key for feature set fusion. This involves matching and connecting phenological water-sensitive characteristics with multi-source data, which includes: observed meteorological data (meteorological element values, including temperature, humidity, wind speed, and solar radiation intensity); discrete representations of crop phenological data (phenological period codes); and predicted results from weather forecast data (meteorological forecast values, including future temperature, precipitation probability, wind speed, and sunshine duration).

[0028] Once the phenological period-sensitive moisture characteristics and multi-source data are matched and connected under a unified time reference, a multi-column data structure is formed. This data structure is the final original feature set containing different feature dimensions. Each row in the original feature set corresponds to a unified sampling time node, and each column corresponds to a specific feature dimension. These feature dimensions specifically cover phenological period-sensitive moisture characteristics, meteorological element values, phenological period codes, and meteorological forecast values.

[0029] Based on the physiological stage characteristics of the current crop phenological stage, an adaptive scaling ratio for each feature dimension in the original feature set is determined, and the original feature set is adaptively scaled to generate a phenological stage perception feature set. Specifically, according to the physiological stage characteristics of the current crop phenological stage, a pre-defined phenological stage-water stress sensitivity mapping table is queried to determine the sensitivity level of the current crop phenological stage to water stress. The phenological stage-water stress sensitivity mapping table is pre-defined based on agronomic knowledge regarding the sensitivity of different physiological stages to water stress. It associates different crop phenological stages with a sensitivity level representing water stress, where the sensitivity level is represented by a positive integer; a larger value indicates greater sensitivity to water stress. For example, for maize, the physiological stage characteristics corresponding to each crop phenological stage and the preset mapping relationship are as follows: the seedling stage, which is in the early stage of growth, has the physiological stage characteristics of initial root development and relatively small water requirement, so it corresponds to sensitivity level 1; the jointing stage, which is in the critical period of vegetative growth, has the physiological stage characteristics of rapid stem elongation and greatly increased water requirement, so it corresponds to sensitivity level 3; the tasseling stage, which is in the critical period of reproductive growth, has the physiological stage characteristics of pollen activity and pollination process being extremely sensitive to water, so it corresponds to sensitivity level 5; the grain-filling stage, which is in the critical period of yield formation, has the physiological stage characteristics of grain filling requiring water but with slightly increased tolerance, so it corresponds to sensitivity level 4; and the maturity stage, which is in the late stage of growth, has the physiological stage characteristics of weakened physiological activity and reduced water requirement, so it corresponds to sensitivity level 2.

[0030] A preset water requirement baseline value corresponding to the current crop phenological stage is obtained. This preset water requirement baseline value is a scalar value representing the theoretically desired water conditions for the current crop phenological stage. The absolute difference between the sensitivity level and this preset water requirement baseline value is calculated to obtain the deviation. Then, based on this deviation, an initial scaling factor is assigned to each feature dimension in the original feature set through a preset linear function. Specifically, the assignment rules of the preset linear function are as follows: For meteorological element values ​​representing the current environmental state and phenological stage-sensitive water features that directly reflect water conditions in the original feature set, their initial scaling factor is determined to be the product of 1 plus the deviation multiplied by 0.1. For crop phenological stage codes used as category identifiers, their initial scaling factor is always 1. For meteorological forecast values ​​from weather forecast data, their initial scaling factor is determined to be the sum of the products of 1 divided by the number in parentheses plus the deviation multiplied by 0.05.

[0031] For example, in an application scenario of a winter wheat field in the North China Plain, the current crop phenological stage is identified as the critical jointing stage requiring water. The preset water demand baseline value for the current crop phenological stage is 30, while the current water stress sensitivity level is 3. The absolute difference between the two is calculated, and the deviation is 27. Based on this deviation, initial scaling factors are assigned to each feature dimension in the original feature set: for meteorological element values ​​that represent the current environmental state and phenological stage-sensitive water features that directly reflect water conditions, the initial scaling factor is amplified to 3.7; for crop phenological stage codes that serve as category identifiers, the initial scaling factor remains unchanged at 1; and for meteorological forecast values ​​from weather forecast data, the initial scaling factor is reduced to 0.425.

[0032] Extract weather forecast data, such as the weather forecast data for the next 24 hours. The weather forecast data includes predicted values ​​for future temperature, precipitation probability, wind speed, and sunshine duration. Meteorological data includes temperature, humidity, wind speed, and solar radiation intensity. Simultaneously, obtain phenological sensitive moisture characteristics for the current moment and several consecutive past time points. Calculate the slope of the changing trend formed by these phenological sensitive moisture characteristics obtained in chronological order to characterize the dynamic changing trend of the phenological sensitive moisture characteristics. The method for generating the dynamic meteorological correction factor is performed according to the following steps: Calculate the arithmetic mean of the predicted precipitation probability values ​​in the weather forecast data for the next 24 hours, divide it by the humidity observed at the current moment to obtain the humidity correction term, then calculate the ratio of the predicted cumulative sunshine duration for the next 24 hours to the solar radiation intensity observed at the current moment, and record this ratio as the radiation correction term. Finally, add the absolute value of the changing trend slope, the humidity correction term, and the radiation correction term, and calculate the arithmetic mean of the sum to obtain the average value. The average value is defined as the dynamic meteorological correction factor.

[0033] The dynamic weather correction factor is input into a preset scaling ratio mapping relationship, and the output is the adjustment amount of the initial scaling coefficient. This preset scaling ratio mapping relationship is defined as a piecewise linear function. Specifically, the dynamic weather correction factor is used as the only input of this piecewise linear function. By querying the scaling ratio mapping relationship, an adjustment amount of the initial scaling coefficient is output. The mapping rules of the piecewise linear function are clearly defined as follows: First, when the input dynamic weather correction factor is less than or equal to 1.0, the piecewise linear function outputs a fixed adjustment amount of 0; when the dynamic weather correction factor is greater than 1.0 but does not exceed 2.0, its output adjustment amount is the difference obtained by subtracting 1.0 from the dynamic weather correction factor, and then multiplying the difference by 0.2 to obtain the adjustment amount; when the dynamic weather correction factor is greater than 2.0, the piecewise linear function outputs a constant adjustment amount of 0.2.

[0034] The initial scaling factor is corrected based on the adjustment amount, and an adaptive scaling ratio for each feature dimension is generated. For meteorological element values ​​and phenological period-sensitive moisture features, the adaptive scaling ratio is calculated by adding the adjustment amount to the corresponding initial scaling factor. For meteorological forecast values, the adaptive scaling ratio is calculated by subtracting the adjustment amount from the corresponding initial scaling factor. For crop phenological period codes, the adaptive scaling ratio remains unchanged, which is the initial scaling factor 1. Through the correction process, an adaptive scaling ratio calibrated by dynamic meteorological factors is finally generated for each feature dimension in the original feature set.

[0035] Adaptive feature scaling is performed on all feature dimensions of the original feature set according to an adaptive scaling ratio to generate the phenological period sensing feature set. First, the data structure of the original feature set is defined: each row corresponds to a unified time sampling node, and each column corresponds to a specific feature dimension. In practice, adaptive feature scaling is performed on all feature dimensions. The process iterates through each column corresponding to each feature dimension in the original feature set, multiplying all time sampling nodes in each column by the adaptive scaling ratio of that feature dimension. This ultimately generates a dataset, which is the phenological period sensing feature set.

[0036] The phenological period sensing feature set is input into a pre-trained irrigation decision model, which outputs the preliminary irrigation demand probability and irrigation volume reference values ​​for different time nodes within a preset future period. Specifically, the phenological period sensing feature set is input into an irrigation decision model that has been pre-trained using historical data. The forward inference calculation of the irrigation decision model is performed to generate the initial demand probability value and initial irrigation volume value for each time node within the preset future period. For example, the initial demand probability value and initial irrigation volume value for each hour in the next 24 hours. The initial demand probability value is a value between 0 and 1, representing the probability that irrigation is needed at that future time node according to the irrigation decision model. The initial irrigation volume value is a value in millimeters, representing the irrigation water depth initially suggested by the irrigation decision model.

[0037] Feature dimensions of meteorological and weather forecast data are extracted from the phenological period perception feature set. The observation values ​​collected at fixed time intervals within a continuous preset period before the current time node are obtained. These observation values ​​constitute a time series. To calculate the linear regression slope of the time series, the order of each observation value in the time series is first defined as the time sequence number. The time sequence number is used as the independent variable, and the corresponding observation value is used as the dependent variable. The least squares method is used to fit a straight line. The least squares method determines the slope and intercept of the line by minimizing the sum of the squares of the vertical distances from all observation values ​​to the line. The slope of the determined line is used as the linear regression slope, which is also the gradient of change.

[0038] The difference between each observed value in the time series and the predicted value corresponding to the same time sequence number of the fitted straight line is calculated. These differences constitute a residual sequence. In order to quantify the volatility of the residual sequence, fluctuation events are identified by detecting the continuous change of the numerical sign of the residual sequence: when traversing this residual sequence in turn, when the sign of a difference is different from that of the previous difference, it is considered that a valid fluctuation event has occurred. The total number of fluctuation events in the continuous preset time period is counted and divided by the total duration of the continuous preset time period. The resulting value is the fluctuation frequency. The change gradient and fluctuation frequency in the continuous preset time period before the current time node calculated from all feature dimensions together constitute the meteorological time series evolution characteristics.

[0039] For each future time node within a preset future time period, with its corresponding initial demand probability value as the center, select the initial demand probability values ​​corresponding to several adjacent time nodes before and after it to form a sliding window on the time series. Calculate the arithmetic mean of all initial demand probability values ​​within this sliding window to obtain the window probability mean. Then, process the meteorological time series evolution characteristics: calculate the absolute value of the change gradient of all meteorological time series evolution characteristics, and calculate the arithmetic mean of these absolute values ​​of change gradients. This arithmetic mean is defined as the comprehensive trend strength. At the same time, calculate the arithmetic mean of the fluctuation frequency of all meteorological time series evolution characteristics, and define it as the comprehensive fluctuation level.

[0040] Based on the current crop phenological period query and the preset weight configuration table, the phenological period adjustment weights corresponding to the comprehensive trend intensity and comprehensive fluctuation level are obtained respectively. The smoothing correction process is completed through a weighted fusion calculation: the window probability mean, comprehensive trend intensity, and comprehensive fluctuation level are linearly weighted and summed. The weight of the window probability mean is fixed at 1. The comprehensive trend intensity and comprehensive fluctuation level are multiplied by their corresponding phenological period adjustment weights respectively, and then added to the window probability mean. The result of this weighted summation is determined as the preliminary irrigation demand probability of each time node in the future preset period.

[0041] Based on the current crop phenological stage and meteorological time series evolution characteristics, the decision threshold range is dynamically determined. First, a preset phenological stage-basic threshold mapping table is consulted according to the current crop phenological stage to directly obtain a baseline demand probability threshold. Then, the change gradients in all meteorological time series evolution characteristics are converted into absolute values ​​and the average value is calculated. This average value is multiplied by a preset sensitivity coefficient to obtain the threshold adjustment amount.

[0042] The preset sensitivity coefficient is determined by combining prior agronomic knowledge with historical data statistical patterns. This coefficient reflects the sensitivity and adjustment range of the impact of meteorological time-series evolution on the decision threshold range. Specifically, based on the known water response characteristics of the target crop at different phenological stages, domain experts set an initial empirical coefficient range. Then, using historical data from multiple complete growing seasons of the deployed farmland, backfitting and optimization are performed to calibrate the sensitivity coefficient that enables the decision threshold range to achieve the optimal balance between water conservation and yield when guiding irrigation. For example, for meteorologically stable areas or drought-resistant crops, this sensitivity coefficient can be set to a lower value, such as 0.05; while for areas with drastic meteorological fluctuations or water-sensitive key phenological stages, the sensitivity coefficient will be increased accordingly, such as to 0.15, thus allowing the decision threshold range to adapt more sensitively to environmental changes.

[0043] Subtracting the threshold adjustment amount from the demand probability threshold yields the lower limit of the decision threshold interval. Adding the threshold adjustment amount to the demand probability threshold yields the upper limit of the decision threshold interval. The lower and upper limits together constitute the decision threshold interval.

[0044] The initial irrigation demand probability is compared with the decision threshold interval for the corresponding time node. Based on the decision threshold interval into which the initial irrigation demand probability falls, a preset nonlinear mapping function is selected to perform piecewise correction on the initial irrigation amount value, outputting reference values ​​for irrigation amount at different time nodes within a preset future period. The construction and operation rules of this nonlinear mapping function are as follows: The nonlinear mapping function takes the initial irrigation demand probability, the lower limit of the decision threshold interval, the upper limit of the decision threshold interval, and the initial irrigation amount value as inputs. Its core rule is to dynamically select and apply one of three calculation scenarios based on the position of the initial irrigation demand probability relative to the decision threshold interval to complete the piecewise correction of the initial irrigation amount value, and finally calculate the reference value of irrigation amount.

[0045] In the first calculation scenario, when the probability of initial irrigation demand is less than the lower limit of the decision threshold interval, the output value of the nonlinear mapping function is always 0, meaning that irrigation is not required and the reference value for irrigation amount is 0. In the second calculation scenario, when the probability of initial irrigation demand is within the decision threshold interval, i.e., when the probability of initial irrigation demand is greater than or equal to the lower limit and less than or equal to the upper limit, the nonlinear mapping function performs a first-type piecewise correction on the initial irrigation amount value. Specifically, the output value of the nonlinear mapping function is the initial irrigation amount value multiplied by a proportional coefficient. This proportional coefficient is designed to be a nonlinear function that is positively correlated with the probability of initial irrigation demand. Specifically, the proportional coefficient is obtained by dividing the probability of initial irrigation demand by the upper limit of the decision threshold interval and then squaring the quotient. The calculated proportional coefficient is then multiplied by the initial irrigation amount value, and the product is used as the reference value for irrigation amount in this scenario.

[0046] In the third calculation scenario, when the initial irrigation demand probability exceeds the upper limit of the decision threshold interval, the nonlinear mapping function performs a second type of piecewise correction on the initial irrigation amount. The output value of the nonlinear mapping function is the initial irrigation amount plus a dynamic increment. This dynamic increment is calculated by subtracting the upper limit of the decision threshold interval from the initial irrigation demand probability, multiplying this excess by a preset unit over-threshold irrigation coefficient, and obtaining the result as the dynamic increment. The result of adding the dynamic increment to the initial irrigation amount is used as the irrigation reference value in this scenario. By integrating the three calculation scenarios into a unified nonlinear mapping function for piecewise correction, the final output is the irrigation reference value for each time node within the preset future time period.

[0047] For example, in a winter wheat field on the North China Plain, the crop is currently in the heading stage, which is extremely sensitive to water conditions. Preliminary irrigation demand probabilities and initial irrigation amounts are generated for different predicted future time points. For instance, for the first time point, the calculated preliminary irrigation demand probability is 0.70. Combined with a decision threshold range (lower limit 0.65, upper limit 0.75), the initial irrigation amount for this time point is 10. According to a preset segmented correction rule, since the probability value of 0.70 falls within the range of 0.65 to 0.75, a second calculation method is used. The preliminary irrigation demand probability is divided by the upper limit of the decision threshold range, and the quotient is squared to obtain a proportionality coefficient of approximately 0.87. The initial irrigation amount is then multiplied by this proportionality coefficient to obtain the corrected irrigation amount for this time point. The reference value is approximately 8.7. When the initial irrigation demand probability at another time point is 0.60, since the initial irrigation demand probability is lower than the lower limit of the decision threshold interval (0.65), the first calculation scenario is automatically applied, directly determining that no irrigation is needed at that time, and the irrigation reference value is 0. However, for the third time point where the initial irrigation demand probability reaches 0.80, the initial irrigation demand probability has exceeded the upper limit of the decision threshold interval (0.75), so the third calculation scenario is activated. The excess amplitude obtained by subtracting the upper limit of the decision threshold interval from the initial irrigation demand probability is 0.05, equivalent to exceeding 5 percentage points. This excess amplitude is multiplied by the preset unit excess irrigation coefficient of 1.0 mm per percentage point to obtain a dynamic increment of 5 mm. Finally, this dynamic increment is added to the initial irrigation amount value, outputting the irrigation reference value of 15 mm for that time point. By integrating a unified nonlinear mapping function of the three calculation scenarios for piecewise correction, the irrigation reference value for each time point within the preset future time period is finally output.

[0048] Based on the initial irrigation demand probability, irrigation volume reference value, and preset candidate time steps, the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk are calculated. Specifically, based on the initial irrigation demand probability at each time node in the future preset period, several candidate irrigation execution time intervals, i.e., preset candidate time steps, are preset. These candidate time steps can be set to 4 hours, 6 hours, or 8 hours. For each selected candidate time step, starting from the beginning of the future preset period, a series of equally spaced candidate irrigation time nodes are automatically divided within the future preset period at fixed intervals.

[0049] Subsequently, the irrigation time nodes are screened: the probability of the initial irrigation demand corresponding to each irrigation time node is checked in turn. If the probability of the initial irrigation demand is greater than a preset activation threshold, the irrigation time node is retained as a valid irrigation time node. If the probability of the initial irrigation demand is less than or equal to the preset activation threshold, the irrigation time node is discarded, and the next irrigation time node is automatically included in the check process. In this way, different candidate irrigation time nodes are generated in the future preset time period.

[0050] For each candidate irrigation time node, obtain the reference value of irrigation amount corresponding to the irrigation time node. First, calculate the expected value of water use efficiency: the expected value of water use efficiency is obtained by multiplying the reference value of irrigation amount by a water use efficiency coefficient. The water use efficiency coefficient is a value between 0 and 1. The water use efficiency coefficient is obtained by querying the preset correspondence table based on the current soil texture and crop phenology.

[0051] Calculate the expected value of irrigation energy consumption: The expected value of irrigation energy consumption is calculated by multiplying the reference value of irrigation volume by a unit water volume energy consumption factor, and then multiplying it by a terrain resistance coefficient related to the slope of the farmland. The unit water volume energy consumption factor is pre-determined based on the rated power and efficiency of the irrigation system pumps, while the terrain resistance coefficient is obtained by referring to a preset slope and resistance comparison table based on the average slope of the farmland.

[0052] Finally, the expected value of crop yield risk is calculated as follows: The expected value of crop yield risk is calculated as follows: the time difference between the irrigation time node and the center point of the ideal irrigation time window of the crop is calculated, the absolute value of the time difference is taken to obtain the absolute time difference, and then the absolute time difference is multiplied by a phenological period time sensitivity coefficient. The ideal irrigation time window is predefined based on the agronomic knowledge of the current crop phenological period. The phenological period time sensitivity coefficient is obtained by querying a mapping table pre-set based on the current crop phenological period. Finally, the expected value of water use efficiency, the expected value of irrigation energy consumption, and the expected value of crop yield risk are generated for each irrigation time node.

[0053] A multi-objective function is constructed and solved to generate a Pareto optimal irrigation strategy sequence. Irrigation equipment is then scheduled according to this sequence. Specifically, the multi-objective optimization function is constructed with the simultaneous optimization directions of maximizing the expected value of water use efficiency, minimizing the expected value of irrigation energy consumption, and minimizing the expected value of crop yield risk. Specifically, the multi-objective optimization function aims to simultaneously optimize three objectives: the first objective is to maximize the sum of the expected values ​​of water use efficiency at all irrigation time points; the second objective is to minimize the sum of the expected values ​​of irrigation energy consumption at all irrigation time points; and the third objective is to minimize the sum of the expected values ​​of crop yield risk at all irrigation time points.

[0054] Solving the multi-objective optimization function requires satisfying the following constraints: the sum of the reference values ​​of irrigation amount corresponding to all selected irrigation time nodes must not exceed a preset upper limit of total irrigation amount, which is preset based on the water requirement of the crop at the current phenological stage; any selected irrigation time node must be located within a preset feasible interval of irrigation timing, which is preset based on the operable time window of the farmland, such as the daytime working hours of each day. The upper limit of total irrigation amount and the feasible interval of irrigation timing together constitute the multi-objective optimization problem.

[0055] A multi-objective optimization algorithm is used to iteratively solve the multi-objective optimization problem, generating a Pareto optimal solution set. Each solution in the Pareto optimal solution set corresponds to an irrigation time node and a corresponding reference value for irrigation volume. In specific implementation, a non-dominated sorting genetic algorithm based on an elite strategy is selected as the solver, i.e., a specific multi-objective optimization algorithm is used for solving the problem. During the initialization of this multi-objective optimization algorithm, a certain number of solutions are randomly generated as an initial population, where each solution represents a specific irrigation scheme. This irrigation scheme specifies a series of candidate irrigation time nodes and corresponding reference values ​​for irrigation volume. In each iteration of the multi-objective optimization algorithm, for each solution in the current population, the three objective function values ​​are calculated by calling the multi-objective optimization function, namely, the expected value of water use efficiency, the expected value of irrigation energy consumption, and the expected value of crop yield risk that may be caused by the irrigation scheme planned based on the solution.

[0056] Subsequently, the multi-objective optimization algorithm determines the Pareto dominance relationship among all solutions based on their tri-objective function values, and performs non-dominated sorting and crowding calculation accordingly to comprehensively evaluate the quality of the solutions. On this basis, new offspring populations are generated through genetic operations such as selection, crossover, and mutation to achieve continuous evolution of the population. When the iteration process reaches the preset termination condition, the multi-objective optimization algorithm outputs a Pareto optimal solution set. Each solution in this Pareto optimal solution set is a Pareto optimal solution, corresponding to a set of feasible irrigation time nodes and corresponding irrigation amount reference values. Finally, each solution in the Pareto optimal solution set is defined as a complete irrigation strategy, and all solutions together constitute a Pareto optimal irrigation strategy sequence.

[0057] The target irrigation strategy is selected from the Pareto optimal irrigation strategy sequence according to a preset decision preference rule. Specifically, the preset decision preference rule is a rule base containing multiple decision scenarios, each corresponding to a set of pre-defined preference weights. For example, in the water-saving priority decision scenario, the preference weight for the expected value of water use efficiency is set to the highest value of 0.6, the preference weight for the expected value of irrigation energy consumption is set to 0.3, and the preference weight for the expected value of crop yield risk is set to 0.1; in the balanced production decision scenario, the preference weights for the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk are all set to 0.333; in the stable yield and guaranteed income decision scenario, the preference weight for the expected value of crop yield risk is set to the highest value of 0.5, while the preference weights for the expected values ​​of water use efficiency and irrigation energy consumption are 0.3 and 0.2, respectively.

[0058] The specific selection of decision-making scenarios adopts an automatic judgment logic based on the current crop phenological stage. For example, based on the current crop phenological stage, if the current crop phenological stage is the heading stage or grain-filling stage, which is critical for water demand and sensitive to stress, then the stable yield and guaranteed income decision-making scenario is automatically selected. If the current crop phenological stage is the seedling stage or maturity stage, which has moderate water demand, then the water-saving priority decision-making scenario is automatically selected. In other cases, the balanced production decision-making scenario is selected as the default setting. Based on this judgment logic, the corresponding preference weights of expected water use efficiency, expected irrigation energy consumption, and expected crop yield risk are retrieved from the rule base.

[0059] The specific construction process of the decision function is described as follows: For any irrigation strategy in the Pareto optimal irrigation strategy sequence, normalization is performed. The expected value of its water use efficiency is divided by the maximum value of the expected water use efficiency of all irrigation strategies in the Pareto optimal irrigation strategy sequence to obtain the normalized water use efficiency value. The expected value of its irrigation energy consumption is divided by the maximum value of the expected irrigation energy consumption of all irrigation strategies in the Pareto optimal irrigation strategy sequence to obtain the normalized irrigation energy consumption value. The expected value of its crop yield risk is divided by the maximum value of the expected crop yield risk of all irrigation strategies in the Pareto optimal irrigation strategy sequence to obtain the normalized yield risk value.

[0060] After normalization, the water use efficiency component is obtained by multiplying the normalized water use efficiency value by the preference weight of the expected water use efficiency value. The irrigation energy consumption component is obtained by multiplying the normalized irrigation energy consumption value by the preference weight of the expected irrigation energy consumption value, and then subtracting the result from 1. The crop yield risk component is obtained by multiplying the normalized yield risk value by the preference weight of the expected crop yield risk value, and then subtracting the result from 1. Finally, the calculated water use efficiency component, irrigation energy consumption component, and crop yield risk component are added together, and the sum is the comprehensive evaluation value of the irrigation strategy. The function defined through this process is the decision function.

[0061] Subsequently, the Pareto optimal irrigation strategy sequence is traversed and calculated for each irrigation strategy. The expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk corresponding to each irrigation strategy are input into the decision function, and the comprehensive evaluation value of each irrigation strategy is calculated one by one. After the calculation is completed, the comprehensive evaluation values ​​of all irrigation strategies are compared, the largest comprehensive evaluation value is identified, and the corresponding irrigation strategy is selected as the target irrigation strategy.

[0062] The irrigation time nodes of the target irrigation strategy are converted into specific execution times in the irrigation control system schedule, and the irrigation volume reference values ​​of the target irrigation strategy are converted into specific operating parameters of the designated irrigation equipment. The specific execution times and specific operating parameters are integrated into structured irrigation control instructions, which are then scheduled and sent to the corresponding irrigation equipment through the communication network to drive the irrigation equipment to complete precise irrigation operations.

[0063] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0064] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0065] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0067] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for predicting the timing of efficient irrigation for crops, characterized by: S1. Collect time-series data on soil volumetric moisture content, meteorological data, crop phenological data, and weather forecast data for a preset future period from different depths in farmland. After spatiotemporal alignment and cleaning, generate a fused dataset and identify the current crop phenological stage. S2. Based on the fused dataset, construct a soil moisture spatiotemporal map. Perform spatiotemporal correlation analysis on the soil moisture spatiotemporal map to extract dynamic spatiotemporal feature vectors covering the vertical movement and lateral migration patterns of water. S3. Perform feature-level fusion of the dynamic spatiotemporal feature vectors with meteorological data, crop phenological data, and weather forecast data to construct an original feature set. Adaptively scale the original feature set based on the current phenological stage to generate a phenological stage perception feature set. S4. Input the phenological stage perception feature set into a pre-trained irrigation decision model to output the preliminary irrigation demand probability and irrigation volume reference values ​​for different time nodes within a preset future period. S5. Based on the initial irrigation demand probability, irrigation volume reference value and preset candidate time step, calculate the expected value of water use efficiency, irrigation energy consumption and crop yield risk, construct and solve the multi-objective function, generate the Pareto optimal irrigation strategy sequence, and schedule irrigation equipment according to the Pareto optimal irrigation strategy sequence.

2. The method for predicting the timing of efficient irrigation for crops based on machine learning according to claim 1, characterized in that: The specific steps of S1 include: periodically collecting soil volumetric water content time-series data, meteorological data, and crop phenological data at different depths with spatial location information using soil moisture sensors, meteorological monitoring equipment, and crop growth recording equipment deployed in farmland; obtaining weather forecast data for a preset period through a weather forecast data service interface; aligning the soil volumetric water content time-series data, meteorological data, crop phenological data, and weather forecast data with timestamps using a unified time base, and uniformly aligning the spatial location information to generate an aligned dataset; cleaning the aligned dataset to remove outliers and fill in missing data, and fusing it according to the unified time base to generate a fused dataset; identifying the current crop phenological stage based on the phenological time stamps of the fused dataset and the crop phenological data.

3. The method for predicting the timing of efficient irrigation for crops based on machine learning according to claim 2, characterized in that: The specific steps of S2 include: based on the fused dataset, mapping the time series data of soil volumetric water content at different depths to three-dimensional grid nodes according to spatial location information to construct a soil moisture spatiotemporal map; performing spatiotemporal correlation analysis on the soil moisture spatiotemporal map, quantifying the vertical gradient of water content between adjacent three-dimensional grid nodes in the vertical direction into vertical water diffusion flux, and quantifying the water potential gradient between adjacent three-dimensional grid nodes in the horizontal direction into horizontal water potential energy gradient; based on the vertical water diffusion flux and horizontal water potential energy gradient, simultaneously generating vertical transport features representing the vertical movement law of water and horizontal transport features representing the horizontal migration law of water at each time slice of the soil moisture spatiotemporal map; stacking and integrating the vertical transport features and horizontal transport features corresponding to each time slice to construct a three-dimensional tensor, and using the three-dimensional tensor as a dynamic spatiotemporal feature vector covering the vertical movement and horizontal migration law of water.

4. The method for predicting the timing of efficient irrigation for crops based on machine learning according to claim 3, characterized in that: The specific steps of S3 include: determining the type of water transport feature to be extracted based on the current crop phenological stage, and extracting the corresponding vertical transport feature or lateral migration feature from the dynamic spatiotemporal feature vector as phenological stage sensitive water features; fusing the phenological stage sensitive water features, meteorological data, crop phenological stage data, and weather forecast data according to a unified time benchmark to construct an original feature set containing different feature dimensions; determining the adaptive scaling ratio of each feature dimension in the original feature set based on the physiological stage characteristics of the current crop phenological stage, and performing adaptive feature scaling on the original feature set to generate a phenological stage sensing feature set.

5. The method for predicting the timing of efficient irrigation for crops based on machine learning according to claim 4, characterized in that: Based on the physiological stage characteristics of the current crop phenological stage, the adaptive scaling ratio of each feature dimension in the original feature set is determined, and the original feature set is adaptively scaled to generate a phenological stage perception feature set. Specific steps include: determining the sensitivity level of the current crop phenological stage to water stress based on its physiological stage characteristics; determining the initial scaling coefficient of each feature dimension in the original feature set based on the deviation between the sensitivity level and the preset water demand baseline value; generating a dynamic meteorological correction factor by combining weather forecast data, meteorological data, and the dynamic changing trend of phenological stage sensitive water characteristics; inputting the dynamic meteorological correction factor into a preset scaling ratio mapping relationship and outputting the adjustment amount to the initial scaling coefficient; correcting the initial scaling coefficient based on the adjustment amount to generate the adaptive scaling ratio of each feature dimension; and performing adaptive feature scaling calculations on all feature dimensions of the original feature set based on the adaptive scaling ratio to generate the phenological stage perception feature set.

6. The method for predicting the timing of efficient irrigation for crops based on machine learning according to claim 5, characterized in that: The specific steps of S4 include: inputting the phenological period sensing feature set into the pre-trained irrigation decision model, performing forward inference calculations of the irrigation decision model, and generating initial demand probability values ​​and initial irrigation amounts for each time node within the corresponding future preset time period; extracting feature dimensions from meteorological data and weather forecast data from the phenological period sensing feature set, and calculating the gradient and fluctuation frequency of the feature dimensions over the consecutive preset time periods before the current time node, based on the current time node, to generate meteorological time-series evolution features; performing sliding window smoothing correction on the initial demand probability values ​​according to the meteorological time-series evolution features, and generating preliminary irrigation demand probabilities for each time node within the corresponding future preset time period; dynamically determining the decision threshold range based on the current crop phenological period and meteorological time-series evolution features; comparing the preliminary irrigation demand probability with the decision threshold range for the corresponding time node, and selecting a preset nonlinear mapping function to perform segmented correction on the initial irrigation amount value according to the decision threshold range into which the preliminary irrigation demand probability falls, and outputting reference values ​​for irrigation amounts at different time nodes within the future preset time period.

7. The method for predicting the timing of efficient irrigation for crops based on machine learning according to claim 6, characterized in that: Based on the initial irrigation demand probability, irrigation volume reference value, and preset candidate time step, the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk are calculated. The specific steps include: generating different candidate irrigation time nodes within a preset future period based on the initial irrigation demand probability and preset candidate time step; and calculating the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk corresponding to each irrigation time node based on the irrigation volume reference value.

8. The method for predicting the timing of efficient irrigation for crops based on machine learning according to claim 7, characterized in that: The process involves constructing and solving a multi-objective function to generate a Pareto optimal irrigation strategy sequence, and then scheduling irrigation equipment based on this sequence. The specific steps include: constructing a multi-objective optimization function with the optimization directions of maximizing expected water use efficiency, minimizing expected irrigation energy consumption, and minimizing expected crop yield risk, and forming a multi-objective optimization problem with preset upper limits for total irrigation volume and feasible intervals for irrigation timing as constraints; iteratively solving the multi-objective optimization problem using a multi-objective optimization algorithm to generate a Pareto optimal solution set, where each solution corresponds to an irrigation time node and a corresponding irrigation volume reference value; defining each solution in the Pareto optimal solution set as an irrigation strategy, forming a Pareto optimal irrigation strategy sequence; selecting target irrigation strategies from the Pareto optimal irrigation strategy sequence according to preset decision preference rules; and converting the irrigation time nodes and irrigation volume reference values ​​in the target irrigation strategies into irrigation control commands and scheduling irrigation equipment to execute them.

9. The method for predicting the timing of efficient irrigation for crops based on machine learning according to claim 8, characterized in that: The target irrigation strategy is selected from the Pareto optimal irrigation strategy sequence according to a preset decision preference rule. The specific steps include: obtaining the preference weights of the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk based on the preset decision preference rule, and constructing a decision function based on the preference weights; for each irrigation strategy in the Pareto optimal irrigation strategy sequence, inputting the expected values ​​of water use efficiency, irrigation energy consumption, and crop yield risk corresponding to each irrigation strategy into the decision function, calculating the comprehensive evaluation value of each irrigation strategy, comparing the magnitude of the comprehensive evaluation values ​​of all irrigation strategies, identifying the largest comprehensive evaluation value, and selecting the corresponding irrigation strategy as the target irrigation strategy.

Images