An irrigation water requirement prediction method based on meteorological soil crop multi-source feature fusion

Abstract

This invention proposes a method for predicting irrigation water demand based on the fusion of multi-source features from meteorology, soil, and crops. Relating to the field of data prediction, the specific steps are as follows: First, multi-source feature data from farmland areas are collected to construct an original dataset. Then, a high-order perturbation encoder module is used to extract the nonlinear perturbation relationships between farmland features, and a structure energy tensor is introduced to describe the high-order structural correlation. Next, a dynamic gating feature generator module is introduced, combining a perturbation gain factor and a time coupling factor to generate a gating signal, achieving dynamic filtering of multi-source features. Subsequently, a cross-variable interaction modeling module is constructed to generate feature embedding vectors to model the complex coupling relationship between meteorology, soil, and crops. Finally, a multilayer sensing network combined with channel mapping and cross-memory mechanisms is used to output the predicted irrigation water demand, achieving efficient dynamic prediction of farmland water demand and improving water resource utilization efficiency and precise regulation capabilities.

Description

Technical Field

[0001] This invention belongs to the field of data prediction, specifically relating to a method for predicting irrigation water demand based on the fusion of multi-source characteristics of meteorology, soil, and crops. Background Technology

[0002] With the acceleration of agricultural modernization and the increasingly prominent problem of water supply and demand imbalance caused by climate change, the precision and intelligence of farmland irrigation have become the key to ensuring food security and sustainable agricultural development. Traditional methods for assessing irrigation water demand mainly rely on historical empirical formulas, soil moisture measurements, and crop growth stage judgments. These methods suffer from problems such as single data sources, delayed predictions, and poor timeliness. Especially in areas with frequent droughts and drastic fluctuations in weather conditions, it is difficult to achieve a precise response to the dynamic water demand of crops, which seriously restricts the efficiency of water resource allocation and the improvement of crop yield.

[0003] In recent years, with the advancement of remote sensing monitoring, agricultural IoT, and data fusion technologies, researchers have begun to use multi-source information from meteorological, hydrological, soil, and crop physiological parameters for water demand prediction. By introducing sensor networks, meteorological station data, soil moisture monitoring systems, and crop canopy index characteristics, combined with deep learning algorithms, the spatial resolution and temporal sensitivity of irrigation prediction have been significantly improved, gradually moving towards dynamic and refined regulation, and providing a new approach for the efficient utilization of agricultural water resources.

[0004] However, existing methods for predicting irrigation water demand still face many challenges: First, the inconsistent distribution of multi-source heterogeneous data across temporal and spatial scales limits feature fusion, making it difficult to comprehensively depict the actual water demand of crops. Second, the models are insufficient in modeling the response mechanisms of water demand dynamics under different crop types, soil structures, and meteorological backgrounds. Furthermore, current methods generally lack the ability to adaptively regulate nonlinear changes within the irrigation cycle, making it difficult to achieve high-frequency, high-precision dynamic predictions. Therefore, there is an urgent need for an intelligent irrigation water demand prediction method with multi-source feature fusion capabilities, support for dynamic response adjustment, and adaptability to regional differences, providing core support for the optimal allocation of agricultural water resources and the construction of intelligent irrigation systems. Summary of the Invention

[0005] This invention provides a method for predicting irrigation water demand based on the fusion of multi-source features of meteorology, soil and crops. It proposes a prediction model for complex multi-source feature data of meteorology, soil and crops, which consists of a high-order perturbation encoder module, a dynamic gating feature generator module, a cross-variable interaction modeling module and a prediction module.

[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps:

[0007] S1. Collect raw data from farmland areas and construct a raw dataset, which includes multi-source features of meteorology, soil, and crops;

[0008] S2. Design a nonlinear perturbation phase mapping matrix to represent the dynamic perturbation relationship of multi-source features. Determine the asymmetric perturbation characteristics of the original dataset at different time points based on the dynamic perturbation relationship. Introduce a structure energy tensor to describe the higher-order structural correlation between multi-source features to obtain the first dataset.

[0009] S3. Based on the characteristics of perturbation amplitude and time dependence, the first dataset is dynamically adjusted and gated to extract the response capability of multi-source features under perturbation driving, thus obtaining the second dataset.

[0010] S4. Based on the perturbation difference tensor and the asymmetric flux field, capture the complex gravitational direction relationship between variable pairs. The perturbation difference tensor characterizes the correlation fluctuation trend between multi-source features in the second dataset. The perturbation acceleration and the asymmetric flux field are weighted and fused to extract the instability interaction features and obtain the third dataset. The third dataset is divided into a training set and a prediction set.

[0011] S5. The training set input prediction module extracts key features through multi-scale channel mapping, uses a multilayer perceptron to fuse cross-memory features and dynamic weighting coefficients to output water demand prediction values, and uses a hybrid loss function to train the model.

[0012] S6. The prediction set is input into the trained water demand prediction model, and the final output is the predicted water demand value.

[0013] Preferably, in step S1, farmland-related data is collected by fusing multi-dimensional feature information from meteorological sensor data, soil monitoring data, crop growth status data, and historical irrigation records. This mainly includes meteorological data, farmland soil data, and crop data. Meteorological data includes rainfall, temperature, humidity, wind speed, and solar radiation; soil data includes soil texture, water content, and hydraulic conductivity; and crop data includes crop planting structure, crop planting area, crop growth stage, and crop root depth, thus obtaining the original dataset.

[0014] Preferably, in actual agricultural planting scenarios, farmland irrigation water demand is significantly affected by spatiotemporal dynamic changes driven by multiple factors. Traditional methods often rely on empirical methods and static indicators for rough estimation, which is difficult to adapt to the nonlinear water demand fluctuations caused by the dynamic changes of multidimensional features in complex planting environments. Especially at critical nodes such as drought stress, seasonal transitions, and frequent agricultural operations, static models are unable to capture the abrupt trends and phased response mechanisms of water demand changes. Existing research is still insufficient in dealing with the asynchronous coupling relationship between crop physiological characteristics, soil moisture content, and meteorological elements, and cannot achieve fine dynamic coupling modeling. Therefore, this invention proposes an irrigation water demand prediction method based on the fusion of multi-source features of meteorology, soil, and crops. By constructing a high-order perturbation coding module and combining a perturbation-driven phase mapping strategy and a structure energy tensor mechanism, dynamic quantitative modeling of the intensity of farmland water demand response is achieved, and finally, a time-series embedded vector set is formed as the first dataset.

[0015] Preferably, the implementation process of the high-order disturbance encoder module includes:

[0016] S21. First, based on the feature representations of any two moments in the multidimensional feature time series, calculate the Euclidean norm. As a disturbance input variable, the disturbance frequency factor is synchronously fused through logarithmic scaling. Adjusting the rhythmic characteristics of the perturbation nodes at different time periods, and introducing feature center offset values. The overall offset of the feature center at each time step is measured, and the sensitivity to local dynamic changes is enhanced by using a periodic response function. This generates a nonlinear perturbation phase mapping matrix with time-varying control capabilities. The mathematical model is as follows:

[0017] .

[0018] Preferably, using The function is used as a whole to wrap the mapping, thereby enhancing the periodic perturbation. Used to characterize the difference between two moments, the difference value This is used to characterize non-stationary disturbances caused by changes in meteorological factors and soil moisture content under different crop growth conditions at different times, constituting a disturbance direction modulation factor. This enables the model to perceive the trend signal of evolution from normal offset to abnormal deviation. At the same time, a disturbance frequency factor is introduced. Adjust the frequency and rhythm consistency between different time points.

[0019] S22. Calculate the global mean of each feature dimension over the time series. For each time point, the local covariance of any two feature dimensions is calculated, and then mapped using a nonlinear perturbation phase mapping matrix. and stabilizing factors After normalization, the structure energy tensor is obtained. The mathematical model is:

[0020] ;

[0021] In the formula, For time points hour and Characteristic structure-energy tensor, molecule For a local co-offset structure, the denominator is adjusted by introducing a perturbation-sensitive term. This enables the model to strengthen its coordinated response when sudden drought stress occurs during crop growth, thus preventing crops from misjudging meteorological disturbances.

[0022] S23. In actual farmland water demand prediction tasks, the irrigation response changes of different crops are often jointly influenced by meteorological factors, soil moisture, and crop growth stages. To better capture these complex changes, a phase energy superposition device is proposed. The perturbation phase matrix is ​​ordered coupled with the structure energy tensor, and the structure energy tensor is... Phase mapping matrix with nonlinear perturbation Coupling is achieved by determining the coupling direction between feature dimensions through a sign function, and weighted aggregation is used to generate temporal embedding vectors with spatiotemporal correlation. The mathematical model is:

[0023] ;

[0024] In the formula, the first layer The second layer is used to weight the impact of perturbations between time steps. Taking into account the perturbation intensity at each time point in the time series, periods with high perturbations are given greater weight, enhancing the model's ability to respond to sudden events. Used to weight the coupling strength between feature dimensions, capturing the interactions between different features. By controlling the coupling direction between dimensions, the dynamic coupling relationship between feature dimensions is reflected, and the first dataset is finally obtained.

[0025] Preferably, the high-order perturbation encoder module first applies a logarithmic transformation to the distance between features through a perturbation-driven phase mapping matrix, and introduces a perturbation frequency factor and a center offset to enhance sensitivity to minor perturbations and unstable weather conditions. Then, by adjusting the response of the perturbation phase, the model can effectively perceive the interaction between typical characteristics of high load, pre-rainfall, and abrupt temperature changes, thereby improving the accuracy of predicting potential water demand changes. Finally, it effectively integrates multidimensional dynamic perturbations, the coupling relationship between variables, and the interaction between features to provide a high-dimensional feature representation. Through this innovative design, this module can accurately model the dynamic quantitative intensity of farmland water demand response, providing strong support for the precise formulation and dynamic adjustment of agricultural irrigation plans.

[0026] Preferably, the regulation of irrigation water demand during crop growth is affected by the coupling of disturbances from multiple sources of meteorological, soil moisture, and crop growth indicators. The nonlinear and non-stationary dynamic disturbance transmission mechanism cannot effectively extract the synergistic features between these fine-grained disturbances. Therefore, in response to the changing characteristics of farmland evapotranspiration demand and the nonlinear changes of multi-source feature coupling, this invention proposes a dynamic gating feature generator module. By introducing a disturbance gain factor and a time coupling factor, the module precisely adjusts the transmission of information flow. Based on the feature differences and disturbance intensity at each time step, the module dynamically adjusts the strength of the gating signal and the transmission mode of information flow, thereby enabling more flexible perception of the potential evapotranspiration change characteristics in the meteorological, soil, and crop systems. Ultimately, a more accurate and dynamic feature representation is obtained, which serves as a second dataset and provides efficient input for subsequent water demand prediction.

[0027] Furthermore, a dynamic gating feature generator module is proposed, the implementation process of which is as follows:

[0028] S31. First, in order to accurately adjust the transmission of multi-source disturbance characteristics in the environment, a disturbance gain factor was designed. The relative perturbation levels of multi-source features at different times in the first dataset are weighted and controlled. By calculating the global baseline value of the feature within the overall time series and combining it with the change magnitude of the feature at the current time, the contribution of the feature to the information flow is dynamically adjusted. The sensitivity of the information flow is controlled based on the difference between the change magnitude of the multi-source feature at the current time and the global mean. Through weighted processing of feature bias, larger perturbations will cause a stronger response, while smaller perturbations will cause a weaker response. The mathematical model is as follows:

[0029] .

[0030] Preferably, the absolute difference It measures the magnitude of fluctuation of each feature relative to the global mean, reflecting the strength of feature changes. It is an adjustment factor that controls the sensitivity to changes, and determines the intensity of the influence of characteristic deviation. This is achieved by adjusting... The sensitivity of the model can be flexibly controlled under different fluctuation stages and environmental response states, thereby improving the accuracy of water demand estimation. This ensures that under normal circumstances, the perturbation gain factor does not excessively suppress the information flow, while guaranteeing a response to weak perturbations, so that the model has a certain degree of sensitivity and avoids overly ignoring small changes.

[0031] S32. To measure the difference in feature changes between adjacent time steps and to adjust the intensity of information flow transmission, this invention proposes a time coupling factor. The model dynamically adjusts time dependence by comparing the differences in characteristics between adjacent time points and combining the standard deviation information of each feature channel. The standard deviation and adjustment factors enhance the sensitivity to changes, enabling the model to capture nonlinear time-series dependencies in meteorological, soil, and crop systems. This results in a higher response capability to drastic disturbances and potential abrupt changes in water demand. The mathematical model is as follows:

[0032] ;

[0033] In the process of irrigation water demand regulation, the changes in characteristics between continuous time steps can represent the dynamic fluctuations of the environmental system and the potential risk of water shortage. Therefore, the following is introduced: The absolute variation difference better captures dynamic changes between consecutive time periods, helping the model identify potential risks to irrigation water demand.

[0034] S33. Finally, the fusion perturbation gain factor Coupling factor with time To control the dynamic transmission of multi-source feature information, two adjustable parameters are introduced. and These correspond to the degree of characteristic perturbation and the intensity of time dependence, respectively, and regulate and control the amplitude of information flow transmission to generate dynamic gating signals. The mathematical model is:

[0035] ;

[0036] In the formula, the molecular part This reflects the control of the perturbation gain factor on the information flow, and represents the strength of the influence of the current feature change on the model output. The denominator part... Used to control the effect of the time coupling factor on the information flow, by adjusting the parameters. and The model can adaptively control its sensitivity to different types of disturbances and time dependencies to adapt to the nonlinear coupling of irrigation demand with different crop growth states, soil moisture changes, and meteorological disturbance combinations, ultimately yielding... As the second dataset.

[0037] Preferably, the dynamic gating feature generator module adaptively adjusts the information flow transmission of input features based on the temporal fluctuations in multi-source dynamic disturbances of meteorology, soil, and crops. By designing a disturbance gain factor, the module can flexibly enhance its response to potential abnormal disturbances based on the changes between features and the global mean. The temporal coupling factor measures the feature differences between adjacent time steps, enabling the model to perceive the temporal dependence in multi-source disturbances of farmland. Finally, the gating signal combines these two types of factors to adjust the transmission intensity of the information flow, ensuring that the model can accurately capture key changes in the crop water demand process and provide more efficient input for accurate irrigation estimation.

[0038] Preferably, crop water requirements during agricultural planting typically exhibit nonlinear abrupt changes, significant local fluctuations, and complex coupling interference characteristics among multiple variables. Furthermore, these disturbances often do not change synchronously with meteorological factors and soil moisture characteristics. Therefore, this invention proposes a cross-variable interaction modeling module. The module first constructs a perturbation difference tensor to quantify the relative amplitude and asymmetry of perturbations between different data sources. Then, it introduces a perturbation energy density tensor and combines it with uniform difference tension and coupling direction weights to construct a global asymmetric water conduction flux field. Finally, the multi-source interaction embedding vector is aggregated, fully considering the key coupling impacts caused by weakly perturbed meteorological factors and soil parameters, the non-steady-state characteristics reflected by historical average deviations, and the regulatory effect of flux-oriented differences on the prediction weights of different variables. This ensures that the output features not only characterize the local drastic changes in crop water requirements dynamics but also reflect the moderating trend of slowly varying variables on the overall interaction pattern.

[0039] Preferably, the implementation process of the intervariate interaction modeling module is as follows:

[0040] S41. First, to measure the relative perturbation asymmetry between variables, we define the perturbation difference tensor. The mathematical model is:

[0041] .

[0042] Preferably, This is a standard proportional normalized form used to enhance interactions between variables across scales. This method measures the intensity of the shift in current crop water requirements between two source characteristics, enhances the sensitivity to cross-scale perturbations, strengthens the nonlinear amplification effect of historical mean deviation on perturbation responses, and incorporates historical moving averages. As a benchmark, it is used to extract current water demand response characteristics. The degree of deviation, combined with The function's smooth growth characteristic effectively suppresses the unstable effects of extreme outliers on the output while amplifying moderate deviations from the response. Furthermore, the added adjustment factor... This improves the model's robustness to numerical convergence under small offsets, preventing the model from converging too quickly or too slowly in the early stages of training.

[0043] S42. In actual agricultural irrigation, crop water deficit is not caused by a single variable, but is often the result of the combined effect of multiple variables. The magnitude of a single perturbation is insufficient to reveal the risk, and it is necessary to introduce a perturbation rate coupling term. Therefore, this invention defines the perturbation energy density tensor of the variable pair. The mathematical model is:

[0044] ;

[0045] In the formula, This is the disturbance magnitude term, used to measure the strength of disturbances between variables. It mainly captures the amplification effect of differences between variables, and enhances the contribution of moderate disturbances to energy tension through squaring, thus amplifying the disturbance response. This represents the coupling between the intensity of the disturbance and the rate of change. By smoothing the fusion, we can avoid overfitting the model of high-frequency data disturbances and improve the accuracy of identifying abrupt water deficit states.

[0046] S43. Using asymmetry as a criterion for dynamic flow, based on the perturbation energy density tensor of variable pairs. By utilizing the synergistic fusion of two sets of weighting factors to reflect the complex coupling relationship between variables, and by normalizing the difference tension and introducing coupling direction weights, an asymmetric flux field is proposed. To assess the absorption intensity of external stimuli by each variable at its location within the overall water demand system, the mathematical model is as follows:

[0047] ;

[0048] In the formula, This is the normalization term for asymmetric perturbation flux, which accurately measures the asymmetric strength of perturbations between variables. This is a weighting term used to determine the direction of the gravitational force between each pair of variables.

[0049] Preferably, the asymmetric flux field Multiplying the acceleration response of the current variable captures the deep-seated unstable interaction trends between variables, thereby forming a feature embedding vector. As the third dataset, the mathematical model is:

[0050] ;

[0051] In the formula, The term represents the acceleration of the current feature sequence, used to characterize the potential water demand risk caused by drastic changes in farmland conditions; the term represents the offset. By introducing the concept of non-stationarity in time series data, this approach is used to reflect the chronic water demand risk regulation caused by long-term deviations from steady state in variables. By weighted fusion of the two unstable components, acceleration and mean deviation, a joint modeling of the perturbation between variables and the unstable state of the independent variables themselves is achieved, resulting in a third dataset. The third dataset is then divided into a training set and a prediction set in a 7:3 ratio.

[0052] Preferably, the cross-variable interactive modeling module can adapt to the evolution of water disturbance under asymmetric coupling relationships in different agricultural ecosystems. The module first constructs a disturbance difference tensor to measure the disturbance intensity and asymmetry between variable pairs, then introduces a disturbance energy density tensor to fuse the disturbance amplitude and velocity change trends, and finally models the dominant directionality of the current disturbance by constructing an asymmetric flux tensor. Ultimately, it forms a fusion response vector of water change under dynamic disturbance, providing high-timeliness and high-accuracy irrigation prediction results for intelligent irrigation systems.

[0053] Preferably, feature representations at different scales in the current time step are integrated, including the original interaction vector, average features across channels, and extreme features. These are then mapped to a unified representation space using a learnable projection matrix to achieve comprehensive monitoring of crop water demand status. The core design lies in simultaneously capturing instantaneous anomalies, long-term load trends, and extreme events, and simulating the accumulation and compensation process of crop response through nonlinear interaction. The mathematical model is as follows:

[0054] ;

[0055] In the formula, Some features and details are directly preserved. Some indicators, reflecting cumulative risk, are extracted from the time dimension using global average pooling. Some sensitive channels capture extreme values, corresponding to high-risk events, using the ReLU activation function. A nonlinear threshold response is introduced to simulate the critical mechanism of crop water demand.

[0056] Preferably, a cross-memory factor is introduced. The mathematical model for simulating asymmetric stress response and time-series coordinated processes is as follows:

[0057] ;

[0058] In the formula, To guide the cross-modulation matrix, which is used to learn the stress transfer relationship between features. It is a nonlinear compression function that makes the model produce a saturated response to abnormal signals, while maintaining low sensitivity to normal fluctuations.

[0059] Preferably, based on the temporal dependence and high-dimensional data imbalance of crop water requirement prediction, this invention introduces a multilayer perceptron structure to output the final predicted crop water requirement value and fuses a cross-memory factor. With dynamic weighting coefficients A single-layer perceptron connected to a ReLU activation function is used to output a normalized irrigation water demand prediction. The mathematical model is as follows:

[0060] ;

[0061] In the formula, For crops in time The predicted irrigation water demand, regression direction quantity Ensure the model converges stably.

[0062] Preferably, to further improve the model prediction accuracy, this invention combines anomaly focusing and prediction bias suppression to design a hybrid loss function mechanism. The mathematical model is as follows: This model measures and optimizes the difference between predicted values ​​and true labels.

[0063] ;

[0064] In the formula, This represents the actual irrigation water requirement for crops. This is a loss balancing factor used to control the weighting between quadratic bias and aberration focusing. To focus on the modulation coefficients and enhance the model's attention to error-prone samples, a well-trained crop irrigation water storage prediction model is obtained. By making predictions on the prediction set, the predicted irrigation water demand can be obtained. .

[0065] In summary, this invention proposes an irrigation water demand prediction method based on the fusion of multi-source features from meteorology, soil, and crops. The method comprises the following modules: a high-order perturbation encoder module, a dynamic gated feature generator module, a cross-variable interaction modeling module, and a prediction module. First, the high-order perturbation encoder module constructs perturbation difference tensors and perturbation density tensors to dynamically characterize the evolution trends and shifts of meteorological factors, soil parameters, and crop status at the temporal and scale levels, effectively enhancing the response sensitivity to sudden droughts and rainfall perturbations. Next, the dynamic gated feature generator introduces perturbation gain factors and time weight parameters to autonomously adjust the information throughput and temporal dependence strength of the input features, achieving adaptive dynamic screening of the importance among multi-source features. The process involves selecting a variable interaction modeling module, which integrates asymmetric gravitational field, acceleration response term, and offset energy field to construct a multi-factor coupling path, characterizing the deep unstable interaction mechanism between meteorology, soil, and crops. Finally, the prediction module introduces joint modeling of gating features and interaction features, utilizing a weighted memory mechanism and a multilayer sensing network to accurately output the predicted irrigation water demand. Simultaneously, it combines a heterofocusing loss function to enhance the prediction and identification capabilities for extreme drought and excessive wet periods. Compared with traditional single-factor and shallow regression methods, this invention can deeply integrate multi-source nonlinear features, dynamically adjust the input importance and modeling path, and comprehensively improve the accuracy, stability, and adaptability of irrigation water demand prediction, providing scientific support for agricultural water-saving regulation and precision irrigation. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the steps of a method for predicting irrigation water demand based on the fusion of multi-source characteristics of meteorology, soil, and crops.

[0067] Figure 2 A structural diagram of a crop irrigation water demand prediction model.

[0068] Figure 3 This is a structural diagram of a high-order disturbance encoder module.

[0069] Figure 4 Structure diagram of the module for modeling cross-variable interactions.

[0070] Figure 5 This is a diagram of the model training process.

[0071] Figure 6 The fitting effect diagram for the prediction model to assess irrigation water demand. Detailed Implementation

[0072] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] Please see Figures 1-6 This invention provides a technical solution: a method for predicting irrigation water demand based on the fusion of multi-source features of meteorology, soil, and crops, including a high-order perturbation encoder module, a dynamic gating feature generator module, a cross-variable interaction modeling module, and a prediction module. The specific steps are as follows: Figure 1 As shown.

[0074] A crop irrigation water demand prediction model is constructed, the structure of which is as follows: Figure 2 As shown, the specific steps are as follows.

[0075] S1. Collect raw data from farmland areas and construct a raw dataset, which includes meteorological, soil, and crop multi-source features.

[0076] Furthermore, the dataset of this invention contains 1,000 farmland-related data points, and the dataset is divided into a training set and a test set in a 7:3 ratio.

[0077] S2. Design a nonlinear perturbation phase mapping matrix to represent the dynamic perturbation relationship of multi-source features. Determine the asymmetric perturbation characteristics of the original dataset at different time points based on the dynamic perturbation relationship. Introduce a structure energy tensor to describe the higher-order structural correlation between multi-source features to obtain the first dataset.

[0078] Furthermore, the structure of the high-order disturbance encoder module is as follows: Figure 3 As shown, firstly, a perturbation-driven phase mapping matrix is ​​proposed based on the radial basis function widely used in traditional machine learning. The original mathematical model is as follows: By applying exponential decay to feature differences, the distance between samples is mapped to similarity weights. However, when facing water demand prediction tasks in agricultural irrigation scenarios, radial basis functions lack the ability to perceive the collaborative changes between dimensions in the feature space. This invention, based on the above kernel function idea, performs structural deformation and nonlinear enhancement, and calculates the Euclidean norm based on the feature representations of any two moments in a multidimensional feature time series. As a disturbance input variable, the disturbance frequency factor is synchronously fused through logarithmic scaling. Adjusting the rhythmic characteristics of the perturbation nodes at different time periods, and introducing feature center offset values. The overall offset of the feature center at each time step is measured, and the sensitivity to local dynamic changes is enhanced by using a periodic response function, thereby generating a nonlinear perturbation phase mapping matrix with time-varying control capabilities. The mathematical model is:

[0079] ;

[0080] In the formula, The function introduces a periodic perturbation response to enhance the ability to detect minor perturbation anomalies. For the first Feature vectors at each time point Let Euclidean norm be the feature at time t. With time The difference between them is expressed using a logarithmic function. By employing gradual compression, the model is made to no longer drastically decay in response to large distance differences between characteristic time pairs when dealing with long-period dynamic changes caused by crop water absorption rhythms, soil disturbance diffusion processes, and meteorological driving forces. This enhances the model's ability to respond to long-range disturbances across periods while maintaining its sensitivity to subtle perturbation signals. This is the perturbation frequency factor, used to control the perturbation rhythm. It is initialized based on the crop water absorption characteristics and meteorological fluctuation features, with an initial value of 0.8, and subsequently adjusted within the interval... Random value is selected from within. For a moment The feature center offset value reflects the time. The overall meteorological and soil stress trends are modeled as follows:

[0081] ;

[0082] In the formula, The number of feature dimensions at each time point, with a value of 12.

[0083] Furthermore, to capture the dynamic disturbances of different crop types and growth stages in farmland, a structural energy tensor is proposed based on the traditional covariance matrix model. In traditional statistics, the covariance matrix is ​​used to measure the linear relationship between variables. However, in farmland irrigation data, the relationship between water demand response features is nonlinear, stage-dependent, and time-varying. Therefore, the covariance matrix cannot directly meet the modeling requirements. This invention designs a structural energy tensor based on covariance to construct a high-dimensional structure for dynamically sensing the relationship between water demand variables, and calculates the global mean of each feature dimension over the time series. For each time point, the local covariance of any two feature dimensions is calculated, and then mapped using a nonlinear perturbation phase mapping matrix. and stabilizing factors After normalization, the structure energy tensor is obtained. The mathematical model is:

[0084] ;

[0085] In the formula, For time points hour and The characteristic structure-energy tensor, with the molecule having a local covariance structure, represents the first... With the The degree of synchronization offset of each feature This is a stabilizing factor used to prevent division by zero in the denominator; its value is set to 0.001. For the first The global mean of the dimensional feature across the entire time series is mathematically modeled as follows:

[0086] ;

[0087] In the formula, This represents the total number of time steps, with a value of 1000.

[0088] Furthermore, the perturbation phase matrix is ​​coupled with the structure energy tensor, and the structure energy tensor is... Phase mapping matrix with nonlinear perturbation Coupling is achieved by determining the coupling direction between feature dimensions using a sign function, thus generating the final temporal embedding vector. The mathematical model is:

[0089] ;

[0090] In the formula, For the first At the time step, the first With the The coupling direction factor between the features is mathematically modeled as follows:

[0091] ;

[0092] In the formula, For symbolic functions, For the first The standard deviation of the first feature is used to measure the first feature. The mathematical model for the fluctuation range of each feature dimension over the entire time period is as follows:

[0093] ;

[0094] when When it is greater than 0.55, it indicates that the first These characteristics fluctuate greatly and are highly unstable under the spatiotemporal variations of meteorology, soil, and crops. When it is less than 0.05, it indicates that the first... These characteristics are relatively stable, with small fluctuations, and high reliability. It is a joint fluctuation term, along with the mean difference. Together they determine the direction of the asymmetric coupling between these two dimensions, if It is a positive number, indicating that the first... With the The directional offset between features and the degree of instability increase in the same direction, indicating positive coupling. A negative value indicates that the two features are in opposite directions, indicating negative coupling. If the value is zero, it indicates no coupling directionality. The final output is... As the first dataset.

[0095] S3. Based on the characteristics of perturbation amplitude and time dependence, the first dataset is dynamically adjusted and gating, and the response capability of multi-source features under perturbation is extracted to obtain the second dataset.

[0096] Furthermore, to effectively capture the coupling and non-stationary changes between meteorological disturbances and multi-source features of soil and crops during changes in farmland irrigation demand, this invention proposes a dynamic gated feature generator module based on the classical kernel function method and disturbance-driven model, as shown in the diagram below. Figure 4 As shown, in the specific implementation, the input features come from the first dataset output by the higher-order perturbation encoder module. First, the perturbation gain factor was designed. The relative perturbation levels of the multi-source features in the first dataset are weighted and controlled. By calculating the global baseline value of the feature within the overall time series and combining it with the change magnitude of the feature at the current moment, the contribution of the feature to the information flow is dynamically adjusted. The mathematical model is as follows:

[0097] ;

[0098] In the formula, The global mean of the feature over the entire time series is used to measure the relative change of the feature. The adjustment factor, used to control sensitivity to disturbances, is set to 0.8. This is the bias term, used for fine-tuning, with a value set to 0.1; it is also the perturbation gain factor. By measuring the deviation of each feature from the global mean of the overall time series at the current moment, adjustable information flow weights are dynamically assigned to different features, thereby achieving selective control during feature propagation. Greater than When, it indicates that the feature has a strong perturbation trend. Less than When this occurs, it indicates that the characteristics tend to stabilize.

[0099] Furthermore, considering the interactions between features, this invention designs a time coupling factor. The time dependence is dynamically controlled by comparing the feature differences between adjacent time points and combining the standard deviation information of each feature channel to capture the nonlinear temporal dependence of multi-source features. The effect of the gating mechanism is dynamically adjusted, taking into account the time dependence in the meteorological change process, that is, the perturbation coupling between the current time and the previous time. The mathematical model is as follows:

[0100] ;

[0101] In the formula, For the first The standard deviation of a feature is used to quantify its volatility. The parameter used to control the nonlinear adjustment and the sensitivity of the disturbance response is set to 2.

[0102] Furthermore, the fusion perturbation gain factor Coupling factor with time This method controls the dynamic transmission of multi-source feature information by introducing two adjustable control coefficients, corresponding to the degree of feature perturbation and the intensity of time dependence, respectively, to regulate and control the amplitude of information flow transmission, ultimately generating a dynamic gating signal. It is used to regulate the transmission of information flow, and its mathematical model is as follows:

[0103] ;

[0104] In the formula, the parameter The adjustment factor is the perturbation gain factor, which is mainly related to the magnitude of change in the soil-crop system during short-term meteorological perturbations. The initial value is set to 1, and the subsequent adjustment range is... , A value less than 1.5 results in a smaller impact of the perturbation gain factor on the information flow, making it suitable for crops in a stable moisture state. A value greater than 2.5 enhances the impact of disturbances on information flow, making it suitable for situations where water demand is significantly affected by disturbances and climate shocks. It is an adjustment of the time coupling factor The impact on information flow transmission controls the coupling strength of feature changes between adjacent time steps, affecting the model's order dependency perception of the crop's current water state. The initial value is set to 1, and subsequent values ​​are... The values ​​are dynamically selected between them. A value less than 1 indicates a weaker time dependence, suitable for time-series characteristics of crops under continuous and stable irrigation conditions. Values ​​greater than 1.5 enhance the coupling effect between temporal features, making it suitable for abrupt changes in crop water demand in complex hydrological environments. Then, the gating signal... Combined with the first dataset, a gated feature representation is generated, and the mathematical model is as follows:

[0105] ;

[0106] Finally, output As the second dataset.

[0107] S4. Based on the perturbation difference tensor and the asymmetric flux field, capture the complex gravitational direction relationship between variable pairs. The perturbation difference tensor characterizes the correlation fluctuation trend between multi-source features in the second dataset. The perturbation acceleration and the asymmetric flux field are weighted and fused to extract the instability interaction features and obtain the third dataset. The third dataset is divided into a training set and a prediction set.

[0108] Furthermore, using the second dataset as input to the intervariate interaction modeling module, a set of asymmetric perturbation mapping processes driven by meteorological, soil, and crop factors is constructed. Firstly, in practical agricultural irrigation applications, crop water demand is often coupled with multiple dynamic factors. These influencing variables often exhibit scale differences and response delays, making their potential cross-driving characteristics easily overlooked. Therefore, based on the theory of nonlinear material strain response and deformation stability, this invention designs a perturbation difference tensor model with a tension response enhancement mechanism to strengthen the tension changes that induce large responses from small perturbations, reflecting the high sensitivity of crop water status. The original mathematical model is as follows: Based on the original mathematical model, a perturbation difference tensor is defined. The mathematical model is:

[0109] ;

[0110] In the formula, As a stability factor, the value is set to 0.01. This is an exponential parameter used to enhance nonlinear sensitivity; its value is set to 3. To control the adjustment factor of the sensitivity of the perturbation tensor to the degree of historical offset, the control variable is... The effect of the current time-to-historical mean shift on the gain of the overall perturbation tensor is initialized to 1 and can be adjusted according to the system's sensitivity to abnormal shifts, with a range of [value missing]. , For variables In recent The historical moving average over several time steps can effectively suppress the amplification of short-term mutations on the system, thereby constructing a more robust response mechanism. The mathematical model is as follows:

[0111] ;

[0112] In the formula, The length of the history window is 20.

[0113] Furthermore, we define the perturbation energy density tensor of the variable pair. A perturbation density function is introduced to reflect the coupling between the degree of difference between variables and the rate of change. The mathematical model is as follows:

[0114] ;

[0115] In the formula, and This is the fusion weighting factor for the disturbance amplitude term and the disturbance rate term. Set the value to 0.6. The value is set to 0.4. Representing variables The rate of change over time, if the variable If the rate of change within an adjacent time step exceeds 0.35, it indicates a sudden change trend, which will trigger a higher energy response.

[0116] Furthermore, crop root water absorption exhibits significant directionality and local asymmetry. Soil layer distribution, water potential gradient, and meteorological disturbances often asymmetrically drive changes in water demand. Therefore, this invention designs an asymmetric flux field based on the "flux difference normalized absorption model" in fluid mechanics. First, based on the perturbation energy density tensor of variable pairs. The complex coupling relationship between variables is reflected by the synergistic fusion of two sets of weighting factors. Then, by normalizing the difference tension and introducing coupling direction weights, the mathematical model is as follows:

[0117] ;

[0118] In the formula, As a stabilization factor, its value is set to 1 to prevent the denominator from being zero and to avoid gradient explosion. The weighting term is used to determine the direction of the gravitational force between each pair of variables. The mathematical model is as follows:

[0119] .

[0120] Furthermore, based on the principles of physical tension and unsteady behavior modeling, the asymmetric flux field is multiplied by the acceleration response of the current variable to capture the deep-seated unstable interaction trends between variables, thereby forming a feature embedding vector. As the third dataset, the mathematical model is:

[0121] ;

[0122] In the formula, Gating features The second-order time derivative is used to characterize the trend of the disturbance. To adjust the weighting coefficients for the two instability components, acceleration and mean deviation, the initial value is set to 1, and subsequently... The values ​​are selected within the range to obtain the third dataset, which is then divided into a training set and a prediction set in a 7:3 ratio.

[0123] S5. The training set input prediction module extracts key features through multi-scale channel mapping, uses a multilayer perceptron to fuse cross-memory features and dynamic weighting coefficients to output water demand prediction values, and uses a hybrid loss function to train the model.

[0124] Furthermore, the training set was input into the irrigation water demand prediction model. The model adopted the PyTorch deep learning framework, ran on a Linux operating system, and was accelerated using an NVIDIA V100 32GB GPU. During training, the batch size was set to 128. In the training phase, a multi-scale channel mapping was first proposed based on the principle of biomechanical temporal feature coupling. This fused feature representations at different scales in the current time step, including the original interaction vector, cross-channel average features, and extreme features. These were then mapped to a unified representation space through a learnable projection matrix, achieving comprehensive monitoring of crop water demand status. The core design lies in simultaneously capturing instantaneous anomalies, long-term load trends, and extreme events, and simulating the crop response accumulation and compensation process through nonlinear interaction. The mathematical model is as follows:

[0125] ;

[0126] In the formula, , , There are three sets of learnable projection matrices, applied to the original vector, global average pooling features, and max pooling features, respectively. This is the bias term, with a value set to 0.01. The ReLU activation function is used to improve nonlinear modeling capabilities, and then a cross-memory factor is introduced. The mathematical model for simulating asymmetric stress response and time-series coordinated processes is as follows:

[0127] ;

[0128] In the formula, To guide the cross-modulation matrix, This is element-wise multiplication, used to fuse mappings and modulation factors. This is a non-linear compression function used to adjust the offset strength of the cross-memory factor. This is the bias term, with a value set to 0.01.

[0129] Furthermore, by introducing a multilayer sensor structure to output the current irrigation water demand prediction value, and fusing the cross-memory factor... With dynamic weighting coefficients A single-layer perceptron connected to a ReLU activation function is used to output normalized predicted crop irrigation water requirements. The mathematical model is as follows:

[0130] ;

[0131] In the formula, To output the current irrigation water demand prediction value by introducing a multilayer sensor structure, The final regression direction vector, whose values ​​follow a normal distribution. , These are dynamic weighting coefficients, initially set to 0.3, and subsequently... The values ​​are dynamically selected between them. This is the bias term, with a value set to 0.01.

[0132] Furthermore, to further improve the model's prediction accuracy, this invention combines anomaly focusing with prediction bias suppression to design a hybrid loss function mechanism. The mathematical model is as follows: This model measures and optimizes the difference between predicted values ​​and true labels.

[0133] ;

[0134] In the formula, To reflect the actual irrigation water demand, This is the loss balancing factor, used to control the weight between secondary bias and aberration focusing. Its initial value is 0.5, and it is subsequently adjusted... The values ​​are dynamically selected between them. To focus on the modulation coefficients and enhance the model's attention to error-prone samples, a value of 2 is set. The training process is as follows: Figure 5 As shown, the trained irrigation water demand prediction model is finally obtained.

[0135] S6. The prediction set is input into the trained water demand prediction model, and the final output is the predicted water demand value.

[0136] Furthermore, the crop irrigation water demand prediction model achieves the following water demand prediction fitting effect diagram: Figure 6 As shown in the figure, the horizontal axis represents time, the vertical axis represents irrigation water demand, the solid dotted line represents the actual irrigation water demand value, and the dashed cross line represents the model prediction value. It can be seen from the figure that the trend of the predicted value and the actual irrigation water demand value are roughly similar. The experimental results show that the crop irrigation water demand prediction model can effectively capture the trend of farmland data and can predict the crop irrigation water demand value well.

Claims

1. A method for predicting irrigation water demand based on the fusion of multi-source characteristics of meteorology, soil, and crops, characterized in that: Collect raw data from farmland areas to construct a raw dataset, which includes multi-source features of meteorology, soil, and crops. A nonlinear perturbation phase mapping matrix is ​​designed to represent the dynamic perturbation relationship of multi-source features. Based on the dynamic perturbation relationship, the asymmetric perturbation characteristics of the original dataset at different time points are determined. A structure energy tensor is introduced to describe the higher-order structural correlation between multi-source features, thus obtaining the first dataset. The specific steps are as follows: Based on the feature representations at any two moments in the multi-dimensional feature time series, the Euclidean norm is calculated. As a disturbance input variable, the disturbance frequency factor is synchronously fused through logarithmic scaling. By adjusting the rhythmic characteristics of the perturbation nodes at different time periods, a feature mean metric is introduced to measure the overall state of the characteristics at different times. A periodic response function is used to enhance the sensitivity to local dynamic changes, generating a nonlinear perturbation phase mapping matrix with time-varying control capabilities. Calculate the global mean of each feature dimension over the time series. For each time point, the local covariance of any two feature dimensions is calculated, and then the nonlinear perturbation phase mapping matrix and stability factor are used. Normalization is performed on the structure energy tensor. Coupled with the nonlinear perturbation phase mapping matrix, the coupling direction between feature dimensions is determined by a sign function, and a weighted aggregation is performed to generate a temporal embedding vector with spatiotemporal correlation. The mathematical model is: ； In the formula, For the first At the time step, the first With the The coupling direction factor between features, ultimately output As the first dataset; Based on the perturbation amplitude and time dependence characteristics, the first dataset is dynamically adjusted and gated to extract the response capability of multi-source features under perturbation-driven conditions, thus obtaining the second dataset. The specific steps are as follows: designing a perturbation gain factor. The relative perturbation levels of multi-source features at different times are weighted and controlled. By calculating the global baseline value of the feature within the overall time series and combining it with the change amplitude of the feature at the current time, a nonlinear enhancement function is introduced to dynamically adjust the contribution of the feature to the information flow. The mathematical model is as follows: ； In the formula, The feature is the global mean of the entire time series. and To adjust the factor, a time coupling factor is designed. The time dependence is dynamically controlled by comparing the feature differences between adjacent time points and combining the standard deviation information of each feature channel to capture the nonlinear temporal dependence of multi-source features. The mathematical model is as follows: ； In the formula, For the first Standard deviation of features To control the parameters of nonlinear adjustment, a fusion perturbation gain factor is used. Coupling factor with time To control the dynamic transmission of multi-source feature information, two adjustable control coefficients are introduced, corresponding to the degree of feature perturbation and the intensity of time dependence, respectively, to regulate and control the amplitude of information flow transmission. The mathematical model is as follows: ； In the formula, and To adjust the parameters, the gating signal By performing element-wise weighting with the first dataset, the gating control feature representation is obtained. This serves as the second dataset for use as input to subsequent modules; Based on the perturbation difference tensor and the asymmetric flux field, the complex gravitational directional relationship between variable pairs is captured. The perturbation difference tensor characterizes the correlation fluctuation trend between multi-source features in the second dataset. The perturbation acceleration and the asymmetric flux field are weighted and fused to extract the unstable interaction features and obtain the third dataset, which is divided into a training set and a prediction set. The training set is input to the prediction module, which extracts key features through multi-scale channel mapping, uses a multilayer perceptron to fuse cross-memory factors and dynamic weighting coefficients to output water demand prediction values, and uses a hybrid loss function to train the model. The prediction set is input into the trained water demand prediction model, and the final output is the predicted water demand value.

2. The irrigation water demand prediction method based on the fusion of multi-source characteristics of meteorology, soil, and crops according to claim 1, characterized in that, The original dataset includes meteorological data, farmland soil data, and crop data. The meteorological data includes rainfall, temperature, humidity, wind speed, and solar radiation; the soil data includes soil texture, water content, and hydraulic conductivity; and the crop data includes crop planting structure, crop planting area, crop growth stage, and crop root depth, thus obtaining the original dataset.

3. The irrigation water demand prediction method based on the fusion of multi-source characteristics of meteorology, soil, and crops according to claim 2, characterized in that, Using the second dataset as input to the intervariate interaction modeling module, an asymmetric perturbation mapping process between variables is constructed, and the perturbation difference tensor is defined. The mathematical model is: ； In the formula, As a stabilizing factor, For exponential parameters, To control the adjustment factor for the sensitivity of the perturbation tensor to the degree of historical offset, For variables In recent Historical moving average over a time step.

4. The irrigation water demand prediction method based on the fusion of multi-source characteristics of meteorology, soil, and crops according to claim 3, characterized in that, Perturbation energy density tensor based on variable pairs By utilizing the synergistic fusion of two fusion weighting factors to reflect the complex coupling relationship between variables, and by normalizing the difference tension and introducing coupling direction weights, the asymmetric flux field is proposed. To assess the absorption intensity of external stimuli by each variable at its location within the overall water demand system, the mathematical model is as follows: ； In the formula, As a stabilizing factor, As a weighting term, the asymmetric flux field is multiplied by the acceleration response of the current variable to capture the deep-seated unstable interaction trends between variables, forming a feature embedding vector. The mathematical model is: ； In the formula, Gating features The second time derivative, The weighting coefficients are used to obtain the third dataset, which is then divided into a training set and a prediction set in a 7:3 ratio.

5. The irrigation water demand prediction method based on the fusion of multi-source characteristics of meteorology, soil, and crops according to claim 4, characterized in that, The multi-scale channel mapping integrates feature representations at different scales in the current time step, including the original interaction vector, average features across channels, and extreme features. These are then mapped to a unified representation space using a learnable projection matrix, and the cross-memory factor is introduced. Simulating asymmetric stress response and time-series coordination processes, a multilayer perceptron structure is used to fuse the cross-memory factor and dynamic weighting coefficients. Output the predicted value of irrigation water demand A hybrid loss function mechanism combining anomaly focusing and prediction bias suppression is designed. The mathematical model is as follows: This model measures and optimizes the difference between predicted values ​​and true labels. ； In the formula, To reflect the actual irrigation water demand, As a loss balance factor, To focus on the modulation coefficients, the prediction set is ultimately input into the trained irrigation water demand prediction model to obtain the predicted value of irrigation water demand. .

Images