Method for predicting stem xylem water potential of crops in saline-alkali soil and related equipment

Abstract

The application discloses a kind of saline-alkali soil crops stem xylem water potential prediction method and related equipment.The method comprises: using cloud migration optimization algorithm to optimize nonlinear Arp algorithm parameters, and cloud migration optimization algorithm and nonlinear Arp algorithm are constructed into integral cloud migration optimization-nonlinear Arp algorithm;Based on the stem xylem water potential data set collected, the crop stem xylem water potential model constructed by cloud migration optimization-nonlinear Arp algorithm is trained, and SHAP interpreter is used to optimize input factor;Based on the trained crop stem xylem water potential model, the saline-alkali soil crop stem xylem water potential is predicted.It can solve the problem that the traditional soil moisture or soil water potential determination method can only determine soil matric potential, it is difficult to determine the superposition water potential of osmotic potential and matric potential, and it is more difficult to accurately determine the time when crop physiological drought occurs.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the computer field, and more particularly, the present application relates to a salt-alkali soil crop stem xylem water potential prediction method and related equipment. BACKGROUND

[0002] In the salt-alkali soil, the crop root water absorption will not only be affected by the soil matric potential, but also be affected by the osmotic potential due to the accumulation of salt ions in the crop root. The traditional soil moisture or soil water potential measurement method can only measure the soil matric potential, and it is difficult to measure the superimposed water potential of the osmotic potential and the matric potential, and it is more difficult to accurately measure the time when the crop occurs physiological drought. The stem xylem water potential is recognized as one of the core indexes for evaluating the water condition of plants and judging the drought stress threshold. When the stem xylem water potential is observed to decrease, the stem xylem may have occurred embolism, and it is difficult to restore to the original state even if the water supply is restored. Therefore, accurately predicting the crop stem xylem water potential and irrigating in advance is an effective way to reduce the influence of drought stress on crops in the salt-alkali soil, and it is an important problem to be solved in the process of water management in the salt-alkali soil farmland. SUMMARY

[0003] A series of simplified concepts are introduced in the summary section, which will be further described in detail in the specific embodiment section. The summary section of the present application does not mean to try to limit the key features and necessary technical features of the claimed technical solution, and it does not mean to try to determine the protection scope of the claimed technical solution.

[0004] In order to solve the problem that the traditional soil moisture or soil water potential measurement method can only measure the soil matric potential, it is difficult to measure the superimposed water potential of the osmotic potential and the matric potential, and it is more difficult to accurately measure the time when the crop occurs physiological drought, in the first aspect, the present application provides a salt-alkali soil crop stem xylem water potential prediction method, the method comprises:

[0005] The cloud migration optimization algorithm is used to optimize the nonlinear Arp algorithm parameters, and the cloud migration optimization algorithm and the nonlinear Arp algorithm are constructed into an overall cloud migration optimization-nonlinear Arp algorithm;

[0006] Based on the collected stem xylem water potential data set, the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm is trained, and the SHAP interpreter is used to optimize the input factors;

[0007] Based on the trained crop stem xylem water potential model, the salt-alkali soil crop stem xylem water potential is predicted.

[0008] Optionally, the crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm is trained based on the collected stem xylem water potential data set, and the input factors are optimized using the SHAP interpreter, including:

[0009] The crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm is trained based on the collected stem xylem water potential data set, and the input factors of the model are eliminated using the SHAP interpreter until there is no factor to be eliminated, and the model training is ended.

[0010] Optionally, the crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm is trained based on the collected stem xylem water potential data set, including:

[0011] The collected stem xylem water potential data set is processed by using a sliding window and a feature transformation method to obtain new variant features;

[0012] The crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm is trained by using the new variant features.

[0013] Optionally, the stem xylem water potential data set includes stem xylem water potential value, stem xylem cell temperature, stem xylem voltage difference and stem xylem wet bulb temperature difference, and the collected stem xylem water potential data set is processed by using a sliding window and a feature transformation method to obtain new variant features, including:

[0014] The stem xylem temperature and stem xylem wet bulb temperature change value are slid to obtain the stem xylem cell temperature, stem xylem voltage difference and stem xylem wet bulb temperature difference of several time periods before the observation of the stem xylem water potential value;

[0015] The new variant features are obtained by feature transformation through the operation of the stem xylem water potential data set of the same period.

[0016] Optionally, the crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm is trained by using the new variant features, including:

[0017] In the training of the crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm by using the new variant features, the objective function is:

[0018]

[0019] Wherein, E max10％ is the value corresponding to the 10% quantile from large to small of the absolute error of all observed values and simulated values, and E median is the value corresponding to the median of the absolute error of all observed values and simulated values.

[0020] Optionally, the SHAP interpreter is used to optimize the input factors, including:

[0021] determining a target input factor with the largest absolute SHAP value;

[0022] eliminating input factors with absolute values less than 1% of the absolute value of the target input factor.

[0023] Optionally, it further includes:

[0024] Before model training, at least one of the following is used as a trigger condition for sliding alignment: near-surface temperature jump, wind speed drop, and decoupling of stem voltage difference in wet-bulb temperature difference; the target period is marked to obtain a salt crust candidate window;

[0025] The salt crust strength index is evaluated in combination with the conditions of wet-bulb difference sudden increase amplitude, wet-bulb difference sudden increase maintenance time, and wet-bulb difference sudden increase nighttime recovery, and the salt crust strength index is used to gate the salt crust candidate window during model training and inference;

[0026] During training of the model, a special sub-model is enabled for samples marked as salt crust candidate windows, which is used to adjust the saturation segment slope of the nonlinear Arp and reduce the weight of the wet-bulb temperature difference feature, and impose an online constraint on the positive effect of the wet-bulb temperature difference;

[0027] The marginal contribution direction and size of the wet-bulb temperature difference are reviewed within the salt crust candidate window, and if the monotonic / inhibitory relationship consistent with the salt crust mechanism is not met, the corresponding factor is eliminated or weighted until convergence;

[0028] When predicting based on the trained model, the inference branch of the special sub-model is used for periods gated as salt crust candidate windows, and the remaining periods use the normal model branch, to reduce the distortion effect of the wet-bulb temperature difference on the salt crust instantaneous formation.

[0029] In a second aspect, the present application further provides a device for predicting the stem xylem water potential of a crop in a saline-alkali soil, comprising:

[0030] A modeling unit is configured to optimize the nonlinear Arp algorithm parameters using a cloud migration optimization algorithm, and to construct the cloud migration optimization algorithm and the nonlinear Arp algorithm into an overall cloud migration optimization-nonlinear Arp algorithm.

[0031] A training unit is configured to train a crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on a collected stem xylem water potential dataset, and to optimize input factors using a SHAP interpreter.

[0032] A prediction unit is configured to predict the stem xylem water potential of the crop in the saline-alkali soil based on the trained crop stem xylem water potential model.

[0033] In a third aspect, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor is configured to implement the steps of the method for predicting the stem xylem water potential of the crop in the saline-alkali soil according to any one of the first aspect.

[0034] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, and the computer program is executable by a processor to implement the method for predicting the stem xylem water potential of the crop in the saline-alkali soil according to any one of the first aspect.

[0035] In summary, the method for predicting the stem xylem water potential of the crop in the saline-alkali soil according to the present application optimizes the nonlinear Arp algorithm parameters by using the cloud migration optimization algorithm, and constructs the cloud migration optimization-nonlinear Arp algorithm as a whole; trains the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on the collected stem xylem water potential data set, and optimizes the input factors using the SHAP interpreter; and predicts the stem xylem water potential of the crop in the saline-alkali soil based on the trained crop stem xylem water potential model. Thus, the change in tension of the continuous water passage in the crop body is taken as the core predicted quantity, and the nonlinear autoregressive structure is used to depict the time inertia and threshold / saturation effect of the extreme interval of the water potential. The cloud migration optimization is used to realize the constrained global optimization in the parameter space containing noise and non-convexity, which naturally adapts to the randomness and ambiguity of the saline-alkali soil signal. With the help of the SHAP interpreter, the factor contribution, interaction, and monotonicity are reviewed and the structure is slimmed down, so that the learning result is consistent with the plant physiological mechanism, thereby maintaining reliable forward prediction ability under different salinity, different weather, and different device states. Compared with the trigger mode that only depends on the soil water content or tension threshold, the above scheme can simultaneously absorb the in vivo response of the combined action of matric potential and osmotic potential, is more sensitive to salt-induced latent drought, can give a forward warning before the arrival of osmotic potential-dominated latent drought, significantly reduces the occurrence rate of dangerous events, and moves the irrigation trigger to before the irreversible risk of the root-stem passage, significantly reducing the probability of xylem embolism. In the extreme afternoon with superimposed high temperature, low humidity, and high salinity, the model can maintain physically consistent monotonicity and saturation characteristics, and does not appear directional error or overfitting oscillation, and has extreme scene stability. The SHAP output clearly shows the key factors and interaction relationships, which is convenient for agronomic decision-making and threshold adjustment. Through the small sample retraining mechanism of cloud migration, the cross-season migration is fast and the downtime is short. While ensuring the avoidance of embolism risk, the uncertainty band control can be used to reduce excessive irrigation, reduce water consumption, and stabilize yield, thereby improving the economic and ecological performance of water management in saline-alkali soil.

[0036] The saline-alkali soil crop stem xylem water potential prediction method of the present application, other advantages, objects and features of the present application will be embodied in part by the following description, and part will be understood by those skilled in the art through research and practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0037] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are intended to further aid the understanding of the preferred embodiments, and are not intended to limit the present description. Moreover, like reference numerals denote like parts throughout the several views in the drawings. In the drawings:

[0038] Figure 1 A saline-alkali soil crop stem xylem water potential prediction method flow chart is provided for the embodiments of the present application;

[0039] Figure 2 The SHAP value of the key factor finally screened out in the saline-alkali soil crop stem xylem water potential prediction method provided for the embodiments of the present application;

[0040] Figure 3 A saline-alkali soil crop stem xylem water potential prediction method flow chart is provided for the embodiments of the present application;

[0041] Figure 4 A saline-alkali soil crop stem xylem water potential prediction device structure schematic diagram is provided for the embodiments of the present application;

[0042] Figure 5 A saline-alkali soil crop stem xylem water potential prediction device structure schematic diagram is provided for the embodiments of the present application; DETAILED DESCRIPTION

[0043] The terms "first", "second", "third", "fourth" and the like used in the description and claims of the present application, if any, are used for distinguishing between similar objects talking about the embodiments and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so termed is meant to embrace different subject matters and items that could be described in a different way and are not to be construed as prefacing the terminology used to create a specifically ordered sequence. Moreover, the terms "comprising", "having", "including", and "containing" are to be construed open-ended, i.e., meaning "including, but not limited to", unless otherwise noted. The terms "sub-portion" and "sub-portion" are intended to cover a part of the whole, and are not limited to a part of the whole. The terms "comprising", "having", "including" and "containing" are intended to cover not only the case where the stated elements are present, but also the case where the stated elements are not present or additional elements are present. The terms "first", "second", "third", "fourth" and the like used in the description and claims of the present application, if any, are used for distinguishing between similar objects talking about the embodiments and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so termed is meant to embrace different subject matters and items that could be described in a different way and are not to be construed as prefacing the terminology used to create a specifically ordered sequence. Moreover, the terms "comprising", "having", "including", and "containing" are to be construed open-ended, i.e., meaning "including, but not limited to", unless otherwise noted. The terms "sub-portion" and "sub-portion" are intended to cover a part of the whole, and are not limited to a part of the whole. The terms "comprising", "having", "including" and "containing" are intended to cover not only the case where the stated elements are present, but also the case where the stated elements are not present or additional elements are present. The terms "first", "second", "third", "fourth" and the like used in the description and claims of the present application, if any, are used for distinguishing between similar objects talking about the embodiments and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so termed is meant to embrace different subject matters and items that could be described in a different way and are not to be construed as prefacing the terminology used to create a specifically ordered sequence. Moreover, the terms "comprising", "having", "including", and "containing" are to be construed open-ended, i.e., meaning "including, but not limited to", unless otherwise noted. The terms "sub-portion" and "sub-portion" are intended to cover a part of the whole, and are not limited to a part of the whole. The terms "comprising", "having", "including" and "containing" are intended to cover not only the case where the stated elements are present, but also the case where the stated elements are not present or additional elements are present.

[0044] In order to solve the problem that the traditional soil moisture or soil water potential measurement method can only measure the soil matric potential, it is difficult to measure the superimposed water potential of osmotic potential and matric potential, and it is more difficult to accurately measure the time when physiological drought occurs in crops, please refer to Figure 1 A saline-alkali soil crop stem xylem water potential prediction method provided in the embodiment of the application, which can specifically include steps S110 to S130.

[0045] S110, the non-linear Arp algorithm parameters are optimized by using the cloud migration optimization algorithm, and the cloud migration optimization algorithm and the non-linear Arp algorithm are constructed into an overall cloud migration optimization-non-linear Arp algorithm.

[0046] S120, the crop stem xylem water potential model constructed by the cloud migration optimization-non-linear Arp algorithm is trained based on the collected stem xylem water potential data set, and the input factors are optimized by using the SHAP interpreter.

[0047] S130, the saline-alkali soil crop stem xylem water potential is predicted based on the trained crop stem xylem water potential model.

[0048] Exemplarily, the cloud migration optimization algorithm and the non-linear Arp algorithm can be constructed into an overall algorithm, wherein the cloud migration optimization algorithm is used to optimize the non-linear Arp algorithm parameters. The optimized parameters are the action radius of the radial basis kernel function of the Arp algorithm and the regularization coefficient. In addition, in order to improve the performance of the cloud migration optimization algorithm, the number of cloud populations, the number of iterations and the values of vb and vc for controlling the dynamic speed in the cloud migration optimization algorithm can also be adjusted, for example, the number of cloud populations is 50, the number of iterations is 300, vb=U(-0.1a, 0.1a), vc=U(-0.15b, 0.15b), and the values of a and b can be the same as those of the original model.

[0049] Exemplarily, the xylem psychrometer or equivalent sensors can be arranged in a typical saline-alkali land test plot in a representative manner of crop population, and the xylem water potential, xylem air chamber temperature, xylem voltage difference and xylem wet bulb temperature difference can be collected synchronously, with a sampling interval of 5-15 minutes, covering at least 20-30 diurnal cycles continuously; to eliminate device thermal drift and zero point offset, two-point baseline calibration can be performed in the morning low evapotranspiration period and the noon high evapotranspiration period respectively, and the daily weather conditions and irrigation / rainfall events can be recorded; quality control is performed on the original sequence: abnormal detection using a sliding window removes electrical signal spikes and breakpoints, time stamp alignment strategy is used to resample non-equidistant records to a uniform time step, and missing points are marked rather than hard filled; a separate calibration file is formed for each sensor, including sensor number, installation site, contact pressure and binding state, to ensure subsequent cross-season reproducibility. It can be understood that the xylem psychrometer-thermal state parameters are highly sensitive to changes in body water potential, and in saline-alkali land environment, the body response can comprehensively reflect the superimposed effect of soil matric potential and osmotic potential; through multi-period calibration and quality control, the input signal has a stable baseline and consistent time scale, which improves the reliability and comparability of subsequent modeling from the source, significantly reduces the system error caused by sensor drift and asynchronous sampling, and provides a clean data basis for the model to learn the true physiological signal.

[0050] Exemplarily, based on the obtained clean data, a feature set with time memory and physically interpretable can be constructed. The autoregressive lag term of the target quantity can be selected, preferably the xylem water potential of the last to the first three time steps, to reflect the inertia and diurnal rhythm of the conduit water filling / cavitation. Three original driving quantities can be retained, i.e. the xylem air chamber temperature, xylem voltage difference and xylem wet bulb temperature difference at the current time, to express the microenvironment vapor pressure deficit, heat conduction and electrical bridge response. A small number of nonlinear and interactive features can be introduced according to crop physiology and device mechanism, such as the interaction of temperature and wet bulb difference, threshold segmented indicators of wet bulb difference, to capture the response saturation under extreme hot and dry conditions. Physical consistency priors can be injected to key channels, for example, limiting the monotonicity of the predicted water potential rise with the increase of the wet bulb temperature difference, introducing saturation or inflection point priors in the high temperature section. Finally, robust standardization methods are used to scale each type of feature, and the diurnal phase such as the day cycle sine and cosine is used as an optional weak feature to enhance the model's perception of periodicity. In this way, the in vivo water tension driving, time memory and instrument response are fused with the least but high information density features, and the physiological common sense is fixed into the learning space through priors. Without significantly increasing the parameters, the model's expression power for extreme scenarios is improved, and the fitting direction that is contrary to physiology is prevented, thereby enhancing the generalization stability.

[0051] For example, a candidate set of orders of non-linear autoregressive models, such as p = 2-4, need to be determined, and a lightweight non-linear expression such as piecewise linear, hyperbolic saturation or small feedforward network is selected as the regression kernel, while setting feasible regions and engineering constraints for each to-be-estimated parameter: including the sign of the parameter value, the upper / lower bound of the parameter, the range of the strength of the interaction term, and the penalty coefficient for violating the physical monotonicity. With the goal of increasing the weight of the sample in the drought risk interval, configure a robust loss, such as a target function that is less sensitive to large residuals and finely describes small residuals, and set the early stopping tolerance and minimum improvement threshold according to the training set and validation set division strategy. The combination of the above structure, constraints and loss can be handed over to the cloud migration optimizer for global optimization, and the cloud droplet population size, iteration number and annealing / shrinkage strategy are set to initial value range according to the training data volume and noise level, such as population 40-80, iterations 200-400, and after each iteration, the elite solution is executed to balance exploration and development. Thus, using the expectation, entropy and hyperentropy mechanisms of the cloud model, robust parameter estimation is achieved in a non-convex, noisy search space, and physiological / physical constraints are converted into soft constraints or feasible region clipping. While ensuring computational efficiency, the optimization stability is improved, the risk of falling into a local minimum is significantly reduced, and the final parameters meet the engineering interpretability and deployability.

[0052] For example, under the given order and structure setting, start cloud migration optimization, evaluate the comprehensive indicators of the candidate parameter set on the training set and validation set, such as weighted absolute error, median absolute error and risk segment hit rate, and monitor the validation set curve in real time to avoid overfitting. Once the early stopping condition is met or the maximum iteration number is reached, fix the current elite solution and perform K-fold cross-validation to evaluate its robustness, while recording the error distribution under different environmental intervals such as morning, noon, evening and night, as well as different salinity / hydration backgrounds; if it is found that the error of some intervals is systematically high, the aforementioned non-linear or interaction features will be adjusted, or the feasible region of the corresponding parameter will be tightened during the structure setting and hyperparameter feasible region determination process of the cloud migration optimization-non-linear Arp model, and the training-validation cycle will be entered again until the performance threshold is met in each interval. Finally, select the model with the best comprehensive indicators and that meets the physical monotonicity test from the candidate models as the finalized model. In this way, hierarchical validation and cross-validation can be used to cover day and night and salinity diversity, ensuring that the model is not a casual fit for a certain period or specific environment. The model maintains stable prediction accuracy on a full-day scale and multi-salinity hierarchy, avoiding performance collapse in critical intervals for early warning.

[0053] Exemplary, input the calibrated model with validation data into SHAP explainer, calculate global and local feature contributions, interaction strength and dependence curves, focus on checking three aspects: one is whether the marginal contribution of the latest lag of the target quantity in different environments is stable and consistent with water potential inertia; two is whether the marginal of wet bulb temperature difference and temperature in the high temperature and low humidity interval presents reasonable aggravating effect and saturation inflection point; three is whether the virtual high contribution of voltage difference signal appears under low temperature and high humidity conditions, which prompts possible instrument artifacts. Remove features with long-term unstable contributions or contradictory to physiological common sense from the structure, or apply more stringent monotonic / saturation constraints and retrain slightly. Significant interactions such as high temperature x large wet bulb difference can be solidified into explicit interaction terms to reduce model degrees of freedom. After that, repeat the fast training-validation process once, until the interpretation curves are consistent with the physiological mechanism and the overall error is not worse than the initial model. In this way, the explainable learning closed loop improves the credibility and robustness of the model, reduces variable redundancy and overfitting risk. While ensuring or improving accuracy, reduce the number of parameters and input dimensions, make the model more easily run on edge devices, and facilitate agronomists / water managers to understand and adopt.

[0054] Exemplary, the output model is run online in a sliding window manner, giving point predictions and uncertainty intervals of stem xylem water potential for future short periods such as 1 hour, 3 hours and 6 hours, which can be obtained by multiple perturbation sampling or using heuristic integration within the parameter neighborhood. Combined with the safety threshold and alarm threshold set by the variety or growth period, calculate the time and probability of the prediction curve reaching or crossing the threshold within the look-ahead window, generate irrigation recommendations when the probability of reaching exceeds the preset confidence level, which can include start time, recommended duration / water volume and end judgment rule. To deal with the slow recovery time lag dominated by osmotic potential in saline-alkali land, the system reserves a safe advance before the critical prediction time, and tracks the measured water potential trajectory after irrigation is executed in real time. If the recovery rate is lower than the empirical lower limit, automatically extend irrigation or suggest split irrigation; all alarm and disposal records are written back to the data warehouse for the next season to automatically calibrate the threshold and advance. In this way, the in vivo water potential is directly used as the control variable, bypassing the lagging judgment based only on soil matric potential. Under extreme heat and dryness and high salinity stress, irrigation can be triggered before the occlusion risk appears, significantly reducing the probability of irreversible damage, while avoiding over-irrigation through uncertainty band management.

[0055] Exemplarily, at the time of site replacement or seasonal change, one can first conduct short-term parallel observations along with the existing model and feature system to estimate the drift magnitude of signal distribution under the new environment. Based on this, only the initial value range of the expectation and entropy of cloud movement optimization is adjusted, and the physical constraints such as monotonicity / saturation are kept unchanged, and a small number of rounds of fast retraining are performed. For new instruments or replacement sensors, use the same two-point calibration method to refresh the baseline, and perform hierarchical audit on the prediction-observation residual in the first week after going online. If systematic bias is found in a certain period, then after the feasible region of the related parameters is directed to be widened or tightened, a fine tuning is performed again. All migration and fine tuning processes keep the old model as a fallback version, and in the case of on-site failure or extreme weather, the system falls back to the robust but slightly conservative version for operation. By decoupling the knowledge of structural constraints and data-driven parameters, it is ensured that only a small sample and limited iterations are needed to reset to a stable state when the environment changes. This can significantly reduce the deployment cost and downtime across sites or seasons, and ensure the reliability and economy of the system in long-term operation.

[0056] In summary, the salt-alkali soil crop stem xylem water potential prediction method provided by the embodiments of the present application optimizes the nonlinear Arp algorithm parameters by using the cloud migration optimization algorithm, constructs the cloud migration optimization-nonlinear Arp algorithm as a whole, trains the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on the collected stem xylem water potential data set, and optimizes the input factors by using the SHAP interpreter. The crop stem xylem water potential is predicted based on the trained crop stem xylem water potential model. Thus, the change in tension of the continuous water passage in the crop body is taken as the core predicted quantity, and the nonlinear autoregressive structure is used to depict the time inertia and threshold / saturation effect of the extreme interval of the water potential. The cloud migration optimization is used to realize the constrained global optimization in the parameter space containing noise and non-convexity, and is naturally adapted to the randomness and ambiguity of the salt-alkali soil signal. With the help of the SHAP interpreter, the factor contribution, interaction and monotonicity are reviewed and the structure is slimmed down, so that the learning result is consistent with the plant physiological mechanism, thereby maintaining reliable forward prediction ability under different salinity, different weather and different device states. Compared with the trigger mode relying only on the soil water content or the tension threshold, the above scheme can simultaneously absorb the in-vivo response of the combined action of the matric potential and the osmotic potential, is more sensitive to salt-induced latent drought, can give a forward warning before the latent drought dominated by the osmotic potential, significantly reduces the occurrence rate of the dangerous event of falling below, and moves the irrigation trigger to before the irreversible risk of the root-stem passage, thereby significantly reducing the probability of xylem embolism. In the extreme afternoon with superposition of high temperature, low humidity and high salinity, the model can maintain the monotonicity and saturation characteristics that are physically consistent, and does not appear directional error or overfitting oscillation, and has extreme scene stability. The SHAP output clearly shows the key factors and interaction relationships, which is convenient for agronomic decision-making and threshold adjustment. Through the small sample retraining mechanism of cloud migration, the seasonal migration is fast and the downtime is short. Moreover, while ensuring to avoid the risk of embolism, the uncertainty band control can be used to reduce excessive irrigation, reduce water consumption and stabilize yield, and improve the economy and ecology of water management in salt-alkali soil.

[0057] According to some embodiments, the training of the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on the collected stem xylem water potential data set and the optimization of the input factors by using the SHAP interpreter include:

[0058] The training of the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on the collected stem xylem water potential data set, and the optimization of the input factors of the model by using the SHAP interpreter are performed until there is no factor to be eliminated, and the model training is ended.

[0059] According to some embodiments, the training of the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on the collected stem xylem water potential data set includes:

[0060] The collected stem xylem water potential dataset is processed using a sliding window and feature transformation method to obtain new variant features;

[0061] The crop stem xylem water potential model constructed based on cloud migration optimization-nonlinear Arp algorithm is trained through the new variant features.

[0062] According to some embodiments, the stem xylem water potential dataset includes stem xylem water potential value, stem xylem chamber temperature, stem xylem voltage difference and stem xylem wet bulb temperature difference value, the collected stem xylem water potential dataset is processed using a sliding window and feature transformation method to obtain new variant features, including:

[0063] The stem xylem temperature and stem xylem wet bulb temperature change value are slid to obtain the stem xylem chamber temperature, stem xylem voltage difference and stem xylem wet bulb temperature difference value of several time periods before the observation stem xylem water potential value time;

[0064] The new variant features are obtained by feature transformation through the operation of the same period of stem xylem water potential dataset.

[0065] For example, the PSY-1 plant stem water potential instrument can be used for continuous determination of stem xylem water potential, and the observation interval is half or 1 hour. Abnormal data needs to be removed, and the values of continuous 2 time periods temperature change 0, stem xylem water potential greater than 0, and stem xylem water potential value, stem xylem chamber temperature, stem xylem voltage difference, stem xylem wet bulb temperature difference value more than 3 standard deviations from the mean value are regarded as abnormal values. The training and test dataset should remove the original data after removing the missing and abnormal data, and the data can be arranged into hour scale or day scale, for example, using day scale data, of which 70% is used as training sample set, and the rest is used as test sample set.

[0066] For example, the stem xylem temperature and stem xylem wet bulb temperature change value can be slid to obtain the stem xylem chamber temperature, stem xylem voltage difference, and stem xylem wet bulb temperature difference value of 1 and 2 time periods before the observation stem xylem water potential value time. In addition, based on different time periods such as the stem xylem wet bulb temperature difference value of the previous 1 and 2 time periods, the same period of stem xylem chamber temperature and stem xylem voltage difference can be used to multiply and divide to obtain new variant features, wherein the divisor should not be 0.

[0067] Exemplarily, the cloud migration algorithm-nonlinear Arp model can be trained using the newly added mutation features obtained in the above examples as input, the factor with the largest SHAP absolute value can be determined and the factors less than 1% of the absolute value of the factor can be removed. The stem xylem wet bulb temperature change rate of the previous 1 and 2 periods and the product of the stem xylem temperature change rate of the previous 1 period and the stem xylem temperature change rate of the previous 1 period can be determined as key factors, and the stem xylem temperature of the previous 1 and 2 periods can be removed. The input factor combination determined by the process model training is: previous period air chamber temperature (T_1), previous period potential difference (dT_1), previous period wet bulb temperature difference (WB_1), previous two period temperature difference (WB_2), dT_1 / T_1, dT_1 x T_1, T_1 x WB_1, dT_1 x WB_1, T_1 x WB_2, dT_1 x WB_2, WB_1 / dT1, WB_1 / dT2, WB_1 / T1, WB_1 / T2. For example, as shown in Figure 2 It can be known that the above method can finally determine 6 key factors, and the contributions from large to small are dT_1 x WB_1, WB_2, WB_2 / dT_1, T_1 x WB_2, dT_1 / T1 and dT_1 x WB_2. As shown in Figure 3 , 343 independent samples are selected to constitute a third-party data set for performance evaluation, and the MMER of the observation value and the prediction value is 0.52, the R 2 2 is 0.967, and the RMSE is 0.032 MPa. It can be seen that the above method can accurately predict the time of physiological drought of crops in saline-alkali soil, reduce the probability of drought and xylem embolism of crops in saline-alkali soil, and bring more practical guiding significance of irrigation time determination, thereby providing technical support and decision basis for saline-alkali soil farmland irrigation management. Thus, the technical problems that the related art is difficult to measure the superimposed water potential of osmotic potential and matric potential, and it is more difficult to accurately measure the time of physiological drought of crops are solved.

[0068] Ex, entropy En, and hyper-entropy He to initialize the cloud model, and a number of cloud droplets are randomly generated, each carrying a set of Arp parameters to form an initial population. Then, the fitness of each individual is calculated using the nonlinear Arp model, the training data is input into the Arp model instantiated by each cloud droplet parameter, the predicted sequence is obtained, and the fitness is calculated according to the set robust objective. The objective can be a combination of robust loss and heavy tail suppression index to simultaneously consider overall accuracy and safety in critical sections. If any termination condition is met, the result is output; if not, enter the weight / position update adjustment, perform expected drift on elite cloud droplets according to fitness to strengthen development, and perform entropy contraction on ordinary cloud droplets to reduce search radius and combine hyper-entropy disturbance to maintain diversity, while individuals that violate physical priors are punished or removed, and the iteration continues after updating. When the termination condition is met, the current elite parameters and the finalized nonlinear Arp model are output as a candidate model to enter the explanation-screening stage. Calculate the SHAP value of the candidate model on the validation data to check the global importance, direction consistency, and physical reasonableness of each input factor. If there is an input with extremely low contribution, unstable direction, or inconsistent with prior mechanism, delete the factor and quickly retrain the model with the remaining factors, and repeat the SHAP calculation until all factors meet the screening conditions. When the screening conditions are met, the final result is output, and the final model is obtained. Input the real-time measured stem xylem cell temperature, stem xylem voltage difference, stem xylem wet bulb temperature difference, and recent stem xylem water potential history into the final model to obtain the stem xylem water potential prediction for several time periods in the future, and the safety threshold set according to the crop and growth period can be used for early warning and irrigation linkage.

[0069] For example, based on the collected stem xylem water potential dataset, first, a number of historical time period windows can be set based on the target observation time, for example, with a sampling step of 10 minutes, 6-18 time periods within 60-180 minutes before observation are selected, and the stem xylem air chamber temperature, stem xylem voltage difference and stem xylem wet bulb temperature difference are synchronously sampled and time-aligned. Specifically, a fixed number of sampling steps can be traced back at each target time to construct the corresponding multivariate sequence slice. In order to eliminate the field clock drift and accidental missing, uniform time grid resampling is used, adjacent interpolation is used for short-time missing and missing markers are retained, and sliding window anomaly detection is used to remove or replace sequences with obvious spikes / breaks. At the same time, in order to avoid systematic bias caused by sensor thermal drift, baseline calibration can be performed once a day during low evapotranspiration and high evapotranspiration periods and written into the calibration file; finally, a sliding window feature matrix corresponding to each target water potential observation and quality control label are obtained. The above window feature matrix can include temperature, voltage difference and wet bulb temperature difference of each historical period. Thus, the instantaneous driving force is converted into time series evidence with memory, and the causal time sequence relationship between the proximal microenvironment, device response and in vivo water potential is explicitly exposed. The missing information or mismatch caused by time asynchronization and noise is significantly reduced, ensuring the time sequence consistency and data availability of downstream feature transformation and model learning. After completing the sliding slice, feature operations without formulas are carried out on the stem xylem water potential sequence of the same period to obtain new variation features for enhancing separability: including but not limited to short-term change intensity and direction markers, such as water potential change value sequence of adjacent sampling steps, cumulative change amplitude and rebound / slide duration of recent steps; stability and volatility characterization, such as range, quantile interval width, proportion of abnormal points and quality label summary within the sliding window; phase and rhythm information, such as period indication aligned with circadian rhythm, relative position in the same agricultural period; synchronous coupling features with driving force, such as water potential change direction consistency marker synchronized with wet bulb temperature difference, average water potential level and recovery rate in high temperature sub-interval; and threshold proximity features, indicating the buffer size of the current water potential distance from the preset safety threshold and the speed of approaching the threshold in the last few steps. The above operations are completed on the same time grid as sliding sampling and time alignment, and all new features are scaled and missing proportion screened to remove stable missing or information redundant candidates; the output is a set of variation features parallel to the original sliding window, which is directly used for subsequent training. Thus, only relying on the original instantaneous quantity and lag quantity is not enough to indicate the dynamic process of approaching risk, and the trend, fluctuation, phase and threshold proximity are explicitly exposed through the same window operation on the water potential itself. Without introducing complex formulas, the model is provided with more discriminative evidence that is closer to the physiological process, improving the recognition of critical situations such as critical downward exploration and failure to recover.The sliding window features obtained by sliding sampling and time alignment are spliced with the variation features obtained by water potential operation in the same period as column splicing, which are used as input factor set, and the target amount of stem xylem water potential is used for training cloud migration optimization-nonlinear Arp model. Specifically, the order of the nonlinear Arp candidate and the lightweight nonlinear structure such as piecewise linear, saturation unit or small feedforward network can be determined first, and the engineering feasible region and the physical consistency soft constraint are set, for example, the increase of the wet bulb temperature difference should not cause the predicted water potential to rise, and the response saturation is allowed in the high temperature section. Subsequently, the robust loss and risk section weighting are used as the training target, such as giving higher weight to the data close to the safety threshold, and the cloud migration optimizer is introduced for global parameter search, and the cloud droplet population size, iteration round, expected translation and entropy contraction strategy can be set according to the data size and noise level, and the penalty is applied to the candidate that violates the monotonicity / saturation priori in each round. Through training or validation hierarchical evaluation, the daytime, night, different salinity / weather sub-interval are distinguished, and early stopping and cross validation are used to avoid overfitting. The output is the calibrated model parameter and structure that meets the constraints and meets the performance standard in hierarchical validation. Thus, the time memory Arp and nonlinear response are combined with the global robust parameter search of cloud migration to adapt to the random, fuzzy and non-convex characteristics of saline-alkali land signals, and the physiological common sense is injected into the learning space. In the extreme hot and dry or salinity mutation conditions, it can still make stable prediction, and the error in the critical section is smaller and the advance amount is more reliable.

[0070] According to some embodiments, the crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm is trained by the new variation features, comprising:

[0071] In the training of the crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm by the new variation features, the objective function is:

[0072]

[0073] Wherein, E max10％ is the value corresponding to the 10% quantile from large to small of the absolute error of all observed values and simulated values, E median is the value corresponding to the median of the absolute error of all observed values and simulated values, and n is the sample size.

[0074] It should be noted that strong evaporation in the afternoon or drastic temperature change may cause salt shell to form around the instrument, local relative humidity and leaf boundary layer to be capped, sensor wet bulb difference to be abnormally enlarged, and model to be misjudged as extreme drought. In order to solve the above problems, according to some embodiments, the method further comprises:

[0075] Before model training, at least one of the sudden change of near-surface temperature, the sudden drop of wind speed, and the decoupling of stem voltage difference from wet-bulb temperature difference after sliding alignment is used as a trigger condition to mark the target period to obtain a salt crust candidate window;

[0076] The salt crust intensity index is evaluated in combination with the conditions of the amplitude of sudden increase in wet-bulb difference, the maintenance time of sudden increase in wet-bulb difference, and the recovery of sudden increase in wet-bulb difference at night, which is used to gate the salt crust candidate window in the model training and inference process;

[0077] When training the model, a special sub-model is enabled for samples marked as salt crust candidate windows, which is used to adjust the saturation segment slope of the nonlinear Arp and reduce the weight of the wet-bulb temperature difference feature, and impose an upper limit constraint on the positive effect of the wet-bulb temperature difference;

[0078] The marginal contribution direction and size of the wet-bulb temperature difference are reviewed within the salt crust candidate window. If it does not meet the monotonic / inhibitory relationship consistent with the salt crust mechanism, the corresponding factor is removed or weighted until convergence;

[0079] When predicting based on the trained model, the inference branch of the special sub-model is used for the period gated as the salt crust candidate window, and the remaining period uses the normal model branch, to reduce the distortion effect of the salt crust instantaneous formation on the wet-bulb temperature difference.

[0080] It can be understood that, for the physical phenomenon that the rapid surface salt crust under the strong evaporation condition of saline-alkali soil locally caps the near-surface wet heat exchange, which leads to the abnormal amplification of the stem wet-bulb temperature difference and the measurement distortion of the abnormal amplification of the stem wet-bulb temperature difference, which is not a direct reflection of the real decrease of the water potential in the body, so that the observation bias caused by the salt crust can be separated from the change of the plant water potential at the model level. Specifically, before the data enters the modeling process, the salt crust candidate identification and labeling link can be set. Firstly, the main triggering conditions are the sudden change of the near-surface temperature, the sudden drop of the wind speed from moderate to weak, the rapid rise of the wet-bulb temperature difference without the corresponding voltage difference and the stem water potential, and the non-falling of the wet-bulb temperature difference before and after sunset. The sliding window is used to calculate the three indicators of the sudden increase of the wet-bulb difference, the abnormal duration and the recovery at night to form a salt crust intensity index. The time slice that meets the threshold combination is marked as a salt crust candidate window. In the feature engineering stage, the salt crust intensity index and the candidate label are added to the attribute column of the training sample. In the nonlinear Arp structure, a special branch is enabled for the labeled sample. The parameterization strategy includes setting a monotone upper limit constraint for the wet-bulb temperature difference channel to prevent the model from learning the abnormally increased wet-bulb difference as positive evidence of the continued decrease of the water potential. In the segmented saturation unit, the marginal weight of the wet-bulb difference is reduced and the weight of the autoregressive term is increased to strengthen the dominant role of the body inertia on short-term prediction. At the same time, the weight range of the voltage difference and the chamber temperature remains unchanged and the sudden jump across the segments in the candidate window is prohibited. During training, cloud migration optimization is used to conduct global parameter searching in the feasible region that meets the above constraints. The SHAP interpreter is used to output the local and global marginal contributions and dependence relationships on the validation set. The contribution direction and amplitude of the wet-bulb difference in the salt crust window are reviewed. When the positive amplification appears, which is inconsistent with the physical mechanism, the related interactive term is automatically down-weighted or removed, and then the model is quickly retrained until the interpretation result is consistent with the field mechanism. In the online inference stage, if a time slice is identified as a salt crust candidate window, the special branch is scheduled for reasoning and the upper bound of the output uncertainty is appropriately widened. At the same time, several sampling steps after the end of the candidate window are recorded to monitor whether the normal exchange is restored at night. Once the recovery is confirmed, the normal model branch is switched back. Therefore, since the wet heat boundary layer anomaly caused by the salt crust is identified as an observable state and is suppressed in the model by constraints and branch weight adjustment, the system can significantly reduce false positives caused by the virtual high of the wet-bulb difference in the strong evaporation afternoon, while maintaining the response sensitivity to the real physiological drought intensification. Thus, the warning near the critical water potential is more stable, the irrigation trigger time is closer to the actual water demand of the body, and the prediction consistency across days and weather types is improved without increasing the overall complexity of the model.

[0081] It should be noted that the sudden irrigation in the high temperature period brings cold water and wet heat exchange, the chamber temperature of the stem sensor drops rapidly, the bridge output is temporarily reversed, and the model mistakenly believes that the water potential rises rapidly. In order to solve the above problems, according to some embodiments, further comprising:

[0082] The cloud migration optimization algorithm is used to optimize the parameters of the nonlinear Arp algorithm, and the two are constructed as a cloud migration optimization-nonlinear Arp algorithm; the model constructed by the algorithm is trained based on the collected stem xylem water potential data set, wherein an asymmetric historical sliding window is established according to the irrigation start time before training, and a thermal shock indicator and a recovery lag time feature are generated, and in the training, the sudden drop of the chamber temperature is not equivalent to the rise of the water potential, and the SHAP interpreter is used to remove and reduce the weight of the input factors; based on the trained model, prediction is carried out, and the lag correction branch is enabled within the hot shock window, and only when the recovery lasts more than the preset time and amplitude threshold, the correction is removed to output the final predicted value.

[0083] It can be understood that the sudden irrigation during the high temperature period will cause instantaneous heat exchange and sudden drop of the chamber temperature at the stem measurement site, and the short-time reverse signal caused by the sensor bridge and the heat and moisture exchange process is not equivalent to the effective water supply in the plant body. If such short-time cooling effect is directly explained as the real recovery of the stem xylem water potential, it is easy to make an early judgment of the irrigation stop time. An asymmetric historical window is constructed with the irrigation start time as the anchor point, a longer warming evapotranspiration background section is retained forwardly, and a shorter but high-resolution thermal shock section is set backwardly, features such as chamber temperature drop, temperature drop duration, voltage difference phase inversion amplitude, and apparent water potential recovery speed upper limit are extracted, and the monotonic constraint that the short-time drop of the chamber temperature cannot be learned as positive evidence of the rise of the water potential is explicitly added when training the cloud migration optimization-nonlinear Arp model, and in the prediction stage, the data branch identified as the thermal shock window adopts the recovery delay judgment rule. Only when the apparent recovery of the water potential is within the preset minimum duration and minimum amplitude threshold at the same time, the delay is removed and the normal model inference is restored. In this way, the thermodynamic heat dissipation process and the hydrodynamic water supply process are distinguished in time sequence, so that the system will not prematurely determine that the water potential has safely recovered due to the short-time temperature effect of the cold water, significantly reducing the risk of early pump stop after irrigation leading to a second drop in the morning, and improving the identification ability of the recovery failure scenario while keeping the overall error not worse.

[0084] It should be noted that once the catheter embolism occurs, even if the external conditions improve, the water potential recovery has memory and plateau, and it is difficult for linear or ordinary NAR to recover. In order to solve the above problems, according to some embodiments, further comprising:

[0085] The cloud migration optimization algorithm is used to optimize the parameters of the nonlinear Arp algorithm; based on the stem wood water potential data set, the model is trained, the memory strength features of the near period minimum point, the platform duration and the number of failed recoveries are generated before training, and the memory gate is set in the model structure to increase the autoregressive weight, reduce the environmental driving weight and limit the maximum recovery slope when the memory is strong, and the SHAP interpreter is used to verify the dominant contribution of the memory features in the recovery stage and accordingly to slim down; based on the trained model, the prediction is carried out, and the conservative forward-looking result after gating is output for the samples in the memory strong state to reduce the risk of overestimation in recovery.

[0086] It can be understood that once the xylem vessel experiences a significant negative water potential interval, air bubbles will form in the vessel and the hydraulic structure will degenerate, even if the external evapotranspiration load is reduced or the soil water content is improved, the recovery of the internal water potential will show a slow process of platform, and the traditional model relying only on external driving quantity is easy to give an overly optimistic recovery estimate in this stage. The memory strength features such as the minimum water potential value in the near period, the duration of the low water potential platform, and the count of previous recovery attempts that failed to reach the set recovery amplitude can be calculated from the historical sequence before training, and the gating coefficient based on the above strength is introduced in the nonlinear Arp structure to dynamically adjust the relative weight of the autoregressive term and the external driving term, while limiting the maximum recovery slope of the water potential of the model in the memory strong state, and the SHAP interpreter is used to verify that the memory features have main explanatory power in the recovery stage and are consistent with the experience and physiological knowledge, and the structure is fixed. Thus, the slow recovery rule after a serious stress is incorporated into the internal mechanism of the model, so that the prediction in the memory strong stage no longer gives an overestimated recovery value due to the improvement of the environmental variables, thereby reducing the misjudgment probability near the critical water potential and optimizing the decision basis for the irrigation duration time.

[0087] It should be noted that when the nearby water pump and inverter work introduce electromagnetic noise, the voltage difference appears periodic peaks, and simple denoising will also wipe out the real fast drop. In order to solve the above problems, according to some embodiments, the cloud migration optimization algorithm is used to optimize the parameters of the nonlinear Arp algorithm; when training the model, the shielded reference signal of the voltage difference is collected in parallel first and the difference spectrum robustness and phase locking degree are calculated, the frequency band penalty and weight suppression are applied to the voltage difference features according to the frequency band that is phase-locked and independent of water potential / temperature, while the fast drop fidelity window is reserved, and the SHAP interpreter is used to review that the voltage difference still has a positive contribution in the real drop situation; in the prediction stage, the input detected as the disturbed frequency band is used to adopt the penalized feature weight and output the corresponding robust prediction.

[0088] It can be understood that the periodic electromagnetic noise generated by the field motor, inverter and other equipment will form narrow-band peaks in the stem clamp bridge channel that are unrelated to plant physiological processes. If such peaks are included in the model learning without discrimination, the model will produce false high sensitivity in the interference frequency band, thereby masking the real rapid decline signal or creating false sharp changes. The shielding reference or empty channel can be recorded in parallel during the data acquisition stage, and the differential power spectrum robustness and phase lock degree of the voltage difference are calculated before training to identify frequency bands that are unrelated to water potential and temperature changes but have high phase stability. These frequency bands are weighted and suppressed or explicitly penalized for voltage-related features, while preserving the fidelity sub-window in the time domain to reflect physiological rapid changes to avoid over-denoising. The SHAP interpreter is used to review the voltage difference on the real water potential rapid decline sample to ensure reasonable contribution. Thus, the misleading of the model by specific electromagnetic frequency bands can be significantly reduced, enabling the model to identify electrical signal information related to plant water status in an environment containing operating disturbances, suppressing false positives caused by false peaks, and retaining the ability to respond to real rapid changes.

[0089] It should be noted that stem growth or changes in the tightness of the binding belt can cause baseline drift, which can easily lead to long-term drought. To solve the above problems, according to some embodiments, a cloud migration optimization algorithm is used to optimize the parameters of the nonlinear Arp algorithm; the stem clamp pressure and deformation are recorded before training to generate interdiurnal baseline drift rate and baseline stability under the same temperature and humidity conditions, and a bias branch is used to fit the mechanical drift and apply a daily scale smoothing constraint during training, while the SHAP interpreter is used to ensure that the main physiological feature direction is unchanged and to remove redundant factors that are strongly correlated with drift; the mechanical drift is estimated and deducted by the bias branch before outputting the stem xylem water potential prediction value in the prediction stage.

[0090] It can be understood that as crops grow and the tightness of the binding belt changes, the mechanical conditions of the stem clamp and the stem contact interface will slowly change, causing long-term baseline drift in the measurement system. This drift is not equivalent to physiological water potential changes, and if used directly for training, the model will mistakenly identify mechanical drift as a drought trend, resulting in systematic errors. The clamping pressure, displacement, or structural calibration value can be recorded simultaneously during the acquisition stage, the baseline drift rate and baseline consistency index under the same temperature and humidity conditions are calculated before training, and an additive bias branch is set in the model to specifically fit the slowly changing channel. A daily scale smoothing constraint is applied to the bias branch to prevent the absorption of short-term real changes, and the main branch continues to learn the time series features related to the physiological process. After training, the SHAP verifies that the contribution direction of the main physiological factor has not been changed by the bias, and the final water potential prediction is output after estimating and deducting the bias during inference. Thus, the device and structural factors are separated from the physiological signal at the output layer of the model, the interdiurnal consistency of the prediction is significantly improved during long-term operation, and the false trigger caused by the change in the state of the device and the threshold drift are reduced on the seasonal scale.

[0091] It should be noted that in the case of row spacing, plant height leading to wind corridor, the sensor microclimate difference in the same plot is large, the unified model generalization is poor, in order to solve the above problems, according to some embodiments, further comprising: using cloud migration optimization algorithm to optimize nonlinear Arp algorithm parameters; In the training, the point micro location information and the relative angle of wind direction are introduced to generate the wind corridor index, and the mixed expert structure is used to gate the participation proportion of the common model branch and the wind corridor special branch, and the SHAP interpreter is used to verify the dominance of the wind corridor index on the gating and to cut unnecessary inputs; In the prediction stage, the branches are automatically switched or mixed according to the real-time wind corridor index to obtain the water potential prediction adapted to the local microclimate.

[0092] It can be understood that the wind corridor phenomenon formed by the same plot due to different row spacing, canopy height and dominant wind direction will cause the microclimate at the sensor to be continuously different from the regional average weather. If the model only explains all points with a unified structure, there will be system deviation at points with significant wind corridor. Then in the layout stage, the position of each point in the row, the angle with the main wind direction and the surrounding shielding condition can be recorded, and a wind corridor index is constructed to represent the difference in air exchange intensity and evapotranspiration environment. In training, a mixed expert structure is used to gate the participation proportion of the common branch and the wind corridor special branch, so that the special branch is mainly relied on when the wind corridor is significant, and the common model is returned when the wind corridor is not significant. At the same time, SHAP is used to check the dominance of the wind corridor index in branch allocation and to eliminate unnecessary collinear inputs. Thus, under the condition of keeping the same overall framework and controllable parameter amount, the most suitable inference path is automatically selected for local microclimate difference, reducing the system error of different points in the same plot and improving the generalization stability across points.

[0093] It should be noted that the transition of soil infiltration potential after salt removal or sudden irrigation is easy to cause the instantaneous misalignment of training distribution and online distribution, in order to solve the above problems, according to some embodiments, further comprising: using cloud migration optimization algorithm to optimize nonlinear Arp algorithm parameters; In the training and online operation, the quantile drift of input features and residuals is continuously monitored, the initial value range of cloud migration optimization expectation and entropy is adjusted when the drift event is triggered, and small batch rapid retraining is performed, at the same time, the uncertainty band is widened and calibrated, and the stability of the rule is reviewed and removed by SHAP interpreter; In the prediction stage, the forward-looking result under the conservative calibration of uncertainty coverage is output before the drift recovery to maintain the reliability of early warning.

[0094] It can be understood that after salt removal, water replenishment or sudden heavy rainfall, the soil solution osmotic potential of saline-alkali soil will produce a cascade transition, so that the input characteristics and the target distribution are significantly deviated from the training stage. If the model does not have an adaptive mechanism for distribution mutation, the prediction and threshold determination will not be reliable. The quantile structure of the input characteristics and the residual can be continuously calculated in online operation and compared with the reference quantile of the training period. When the deviation exceeds the set threshold, a small batch rapid retraining process is triggered. By adjusting the cloud migration optimization, the expected value and the entropy initial value are expanded to expand the search space and retain the physical monotony and saturation constraints, so that the model completes parameter adaptation within a limited iteration, and the uncertainty output is calibrated to expand the prediction confidence band. The SHAP is used to review the direction and contribution stability of important features under the new distribution, and the online model is replaced after confirmation. Thus, when the environmental distribution mutates, the prediction can still be continuously available without stopping large-scale retraining, the coverage of the warning remains controllable during the transition period, and the false alarm and false alarm during the mutation period are significantly reduced.

[0095] It should be noted that the wet-bulb difference is low due to night leaf surface dew, which seems to be an environmental improvement, but the root-stem water supply is insufficient, and it drops in the morning. In order to solve the above problems, according to some embodiments, the non-linear Arp algorithm parameters are optimized using a cloud migration optimization algorithm; before training, the dew point is approached and the wet-bulb difference is low to mark the dew window, generate the post-dew exploration risk feature, and apply post-dew penalty to the night window in training to suppress trust in short-term recovery, while using SHAP interpreter to confirm the positive contribution of the post-dew risk feature in the morning stage; in the prediction stage, the warning sensitivity is increased for the forward-looking period after the dew window, and the water potential prediction with penalty correction is output to reduce the morning exploration false alarm.

[0096] It can be understood that condensation often occurs when the temperature of the night leaf surface and the near stem surface is close to the dew point, and the wet-bulb temperature difference decreases and appears as an environmental improvement. However, the body water guide has not been effectively compensated by the root supply, and the evapotranspiration will be enhanced again after the morning light rises, which may cause a second drop. If the model regards the night appearance as real recovery, the sensitivity of the morning warning will be reduced. The dew window can be identified before training according to the combination of dew point approximation, weak wind speed and rapid decrease of wet-bulb difference, and the empirical distribution of water potential change within two hours after dew is counted to construct the post-dew exploration risk feature. In model training, the positive recovery evidence of the night window is suppressed, the warning sensitivity is increased in the morning period after dew, and the SHAP interpreter is used to confirm the positive contribution of the risk feature in the morning judgment and not to cause excessive penalty for normal night recovery. Thus, without changing the response of the main model to real recovery, the pseudo-recovery caused by dew is kept cautious enough to improve the early identification of potential morning drop and reduce the water potential drop below the threshold event caused by false optimism.

[0097] Please refer toFigure 4 An embodiment of the salt-alkali soil crop stem xylem water potential prediction device in the embodiment of the application can include:

[0098] The modeling unit 21 is configured to optimize the nonlinear Arp algorithm parameters by using the cloud migration optimization algorithm, and construct the cloud migration optimization algorithm and the nonlinear Arp algorithm into an overall cloud migration optimization-nonlinear Arp algorithm.

[0099] The training unit 22 is configured to train the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on the collected stem xylem water potential data set, and optimize the input factors by using the SHAP interpreter.

[0100] The prediction unit 23 is configured to perform salt-alkali soil crop stem xylem water potential prediction based on the trained crop stem xylem water potential model.

[0101] In summary, the salt-alkali soil crop stem xylem water potential prediction device provided by the embodiment of the application optimizes the nonlinear Arp algorithm parameters by using the cloud migration optimization algorithm, constructs the cloud migration optimization algorithm and the nonlinear Arp algorithm into an overall cloud migration optimization-nonlinear Arp algorithm, trains the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on the collected stem xylem water potential data set, and optimizes the input factors by using the SHAP interpreter. The salt-alkali soil crop stem xylem water potential prediction is performed based on the trained crop stem xylem water potential model. Thus, compared with the triggering mode that only depends on the soil water content or tension threshold, the above scheme can simultaneously absorb the in vivo response of the combined action of matric potential and osmotic potential, is more sensitive to salt-induced latent drought, can give a forward warning before the latent drought dominated by osmotic potential, significantly reduces the occurrence rate of a dangerous event, and moves the irrigation trigger to before the irreversible risk of the root-stem pathway, thereby significantly reducing the probability of xylem embolism. In an extreme post-noon period with superimposed high temperature, low humidity, and high salinity, the model can maintain physically consistent monotonicity and saturation characteristics, and does not appear directional errors or overfitting oscillations, and has extreme scenario stability. The SHAP output clearly shows the key factors and interaction relationships, which facilitates agronomic decision-making and threshold adjustment. Through the small sample retraining mechanism of cloud migration, the model can be quickly migrated across seasons with short downtime. While ensuring the avoidance of embolism risk, the model can also control excessive irrigation by using uncertainty, reduce water consumption, and stabilize yield, thereby improving the economic and ecological performance of water management in salt-alkali soil.

[0102] As shown in Figure 5 The embodiment of the application also provides an electronic device 300, which includes a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, the steps of any method for predicting the salt-alkali soil crop stem xylem water potential are implemented.

[0103] The cloud migration optimization algorithm is used to optimize the parameters of the nonlinear Arp algorithm, and the cloud migration optimization algorithm and the nonlinear Arp algorithm are constructed into an overall cloud migration optimization-nonlinear Arp algorithm.

[0104] The crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm is trained based on the collected stem xylem water potential data set, and the input factors are optimized using the SHAP interpreter.

[0105] The salt-alkali land crop stem xylem water potential is predicted based on the trained crop stem xylem water potential model.

[0106] Since the electronic device described in the embodiment is the device used to implement the salt-alkali land crop stem xylem water potential prediction device in the embodiment, based on the method described in the embodiment, those skilled in the art can understand the specific implementation of the electronic device in the embodiment and its various forms, so the implementation of the electronic device in the method of the embodiment is not described in detail, as long as the device used to implement the method in the embodiment belongs to the scope of protection of the present application.

[0107] In the specific implementation process, the computer program 311 can realize Figure 1 Any implementation of the corresponding embodiment:

[0108] The cloud migration optimization algorithm is used to optimize the parameters of the nonlinear Arp algorithm, and the cloud migration optimization algorithm and the nonlinear Arp algorithm are constructed into an overall cloud migration optimization-nonlinear Arp algorithm.

[0109] The crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm is trained based on the collected stem xylem water potential data set, and the input factors are optimized using the SHAP interpreter.

[0110] The salt-alkali land crop stem xylem water potential is predicted based on the trained crop stem xylem water potential model.

[0111] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0112] Those skilled in the art will appreciate that embodiments of the application can be readily used as software, hardware, or a combination of software and hardware. In one embodiment, the application can be implemented in software and / or firmware. In this embodiment, the software implementation can include computer readable code stored in a computer readable storage medium (alternatively referred to as a computer readable medium, a processor readable storage medium, or a processor readable code) that, when taken in whole or in part, can program one or more computer processors (alternatively referred to as computer processors, computer processor(s), or processors) to perform various embodiments disclosed herein. The computer readable storage medium or media can be tangible and non-transitory (alternatively referred to as a tangible and / or a non-transitory computer readable medium or media). The term "non-transitory" can also be used in some embodiments including the following embodiments. The term "non-transitory" is to be taken to mean that the computer readable storage medium is a tangible and / or physical computer readable storage medium, as opposed to modulated data signals or carrier waves. The term "non-transitory" specifically disavows the mere bunination of a computer readable medium that is transitory in nature, such as a modulated data signal or a carrier wave. The term "non-transitory" is not intended to encompass transitory signals per se.

[0113] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart blocks and / or diagrams can represent code, programs, or subprograms that can be executed on computing devices. Figure 1 The flowchart blocks and / or diagrams can represent code, programs, or subprograms that can be executed on computing devices.

[0114] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart blocks and / or diagrams can represent code, programs, or subprograms that can be executed on computing devices. Figure 1 The flowchart blocks and / or diagrams can represent code, programs, or subprograms that can be executed on computing devices.

[0115] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart blocks and / or diagrams can represent code, programs, or subprograms that can be executed on computing devices. Figure 1 The flowchart blocks and / or diagrams can represent code, programs, or subprograms that can be executed on computing devices.

[0116] Embodiments of the application also provide a computer program product including computer software instructions that, when executed on a processing device, cause the processing device to perform the steps of any of the embodiments described herein. ​ A flowchart of the saline-alkali soil crop stem xylem water potential prediction in the corresponding embodiments.

[0117] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that the computer can store or be integrated into a data storage device such as a server, data center, etc. containing one or more available media. The available media can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid state disk (SSD)), etc.

[0118] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0119] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or components shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, and can be electrical, mechanical or other forms.

[0120] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0121] In addition, each of the functional units in the various embodiments of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0122] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in part, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the various embodiments of the method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and various other media that can store program codes.

[0123] The above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting stem xylem water potential of a crop in saline-alkali soil, characterized in that, The method comprises the following steps: An Arp algorithm parameter is optimized by using a cloud migration optimization algorithm, and the cloud migration optimization algorithm and the nonlinear Arp algorithm are combined to form an overall cloud migration optimization-nonlinear Arp algorithm; A crop stem xylem water potential model constructed by using the cloud migration optimization-nonlinear Arp algorithm is trained based on a collected stem xylem water potential data set, and an input factor is optimized by using a SHAP interpreter. A salt-alkali soil crop stem xylem water potential is predicted based on the trained crop stem xylem water potential model, and the stem xylem water potential data set comprises a stem xylem water potential value, a stem xylem air chamber temperature, a stem xylem voltage difference and a stem xylem wet bulb temperature difference. Before model training, at least one of the following is used as a trigger condition: a near-surface temperature jump, a wind speed drop, and a wet bulb temperature difference decoupling from a stem voltage difference, a target period is marked to obtain a salt crust candidate window. A salt crust intensity index is evaluated in combination with the following conditions: a wet bulb difference sudden increase amplitude, a wet bulb difference sudden increase maintenance time and whether the wet bulb difference sudden increase is recovered at night, and the salt crust intensity index is used for gating the salt crust candidate window during model training and inference. During the training of the model, a special sub-model is enabled for the samples marked as the salt crust candidate window, the special sub-model is used for adjusting the saturation segment slope of the nonlinear Arp and reducing the weight of the wet bulb temperature difference feature, and an upper limit constraint is applied to the positive influence of the wet bulb temperature difference. The marginal contribution direction and size of the wet bulb temperature difference are reviewed within the salt crust candidate window, and if the monotonicity / inhibition relationship consistent with the salt crust mechanism is not met, the corresponding factor is removed or weighted until convergence. During prediction based on the trained model, the inference branch of the special sub-model is used for the period gated as the salt crust candidate window, and the general model branch is used for the remaining period, so as to reduce the distortion influence of the salt crust instantaneous formation on the wet bulb temperature difference.

2. The method of claim 1, wherein, The crop stem xylem water potential model constructed by using the cloud migration optimization-nonlinear Arp algorithm is trained based on the collected stem xylem water potential data set, and the input factor is optimized by using the SHAP interpreter, which comprises the following steps: The crop stem xylem water potential model constructed by using the cloud migration optimization-nonlinear Arp algorithm is trained based on the collected stem xylem water potential data set, and the input factor of the model is removed by using the SHAP interpreter until there is no removable factor, and then the model training is ended.

3. The method of claim 2, wherein, The crop stem xylem water potential model constructed by using the cloud migration optimization-nonlinear Arp algorithm is trained based on the collected stem xylem water potential data set, which comprises the following steps: The collected stem xylem water potential data set is processed by using a sliding window and a feature transformation method to obtain new variant features. The crop stem xylem water potential model constructed by using the cloud migration optimization-nonlinear Arp algorithm is trained by using the new variant features.

4. The method of claim 3, wherein, The collected stem xylem water potential data set is processed by using a sliding window and a feature transformation method to obtain new variant features, which comprises the following steps: The stem xylem temperature and the stem xylem wet bulb temperature change value are slid to obtain the stem xylem air chamber temperature, the stem xylem voltage difference and the stem xylem wet bulb temperature difference value of several time periods before the observation stem xylem water potential value moment; The characteristic transformation is obtained by the operation of the same period of stem xylem water potential data set.

5. The method of claim 3, wherein, The crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm is trained by the new variant feature, including: In the training of the crop stem xylem water potential model constructed based on the cloud migration optimization-nonlinear Arp algorithm by the new variant feature, the objective function is: wherein, is the value corresponding to the 10th percentile of the absolute error of all observed and simulated values from large to small, is the value corresponding to the median of the absolute error of all observed and simulated values, n is the number of samples.

6. The method of claim 1, wherein, The SHAP interpreter is used to optimize the input factor, including: Determine the target input factor with the largest SHAP absolute value; Remove the input factor less than 1% of the absolute value of the target input factor.

7. A device for predicting stem xylem water potential of a crop in saline-alkali soil, characterized in that, The method of any one of claims 1-6 can be implemented, and the device comprises: A modeling unit is configured to optimize the nonlinear Arp algorithm parameters using the cloud migration optimization algorithm, and to construct the cloud migration optimization algorithm and the nonlinear Arp algorithm into an overall cloud migration optimization-nonlinear Arp algorithm; A training unit is configured to train the crop stem xylem water potential model constructed by the cloud migration optimization-nonlinear Arp algorithm based on the collected stem xylem water potential data set, and to optimize the input factor using the SHAP interpreter; A prediction unit is configured to predict the saline-alkali soil crop stem xylem water potential based on the trained crop stem xylem water potential model.

8. An electronic device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to implement the steps of the saline-alkali soil crop stem xylem water potential prediction method according to any one of claims 1-6 when executing the computer program stored in the memory.

9. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the saline-alkali soil crop stem xylem water potential prediction method according to any one of claims 1-6.

Images