Soil moisture prediction method based on agricultural disturbance and feature contribution dynamic redistribution
By dynamically adjusting the feature contribution weights of the soil moisture prediction model when the disturbance intensity exceeds a threshold, the problems of high computational complexity and insufficient stability of existing models are solved, and efficient and stable soil moisture prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-01
- Publication Date
- 2026-07-28
AI Technical Summary
Existing deep learning-based soil moisture prediction models suffer from high computational complexity during parameter updates and are sensitive to noise, resulting in insufficient prediction stability and response speed.
By using a dynamic redistribution method based on agricultural disturbances and feature contributions, feature contribution weights are redistributed only when the disturbance intensity exceeds a threshold. By combining the instantaneous response intensity of features, the type of agricultural disturbance, and the error feedback correction term, the contribution weights of input features are dynamically adjusted, reducing computational complexity and improving prediction stability.
It significantly reduces the computational complexity of online calculations, improves the response speed and prediction stability of soil moisture prediction in non-stationary agricultural environments, reduces unnecessary computational overhead, and enhances the interpretability of results.
Smart Images
Figure CN122470962A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of humidity prediction technology, and more specifically to a soil humidity prediction method based on the dynamic redistribution of agricultural disturbances and characteristic contributions. Background Technology
[0002] With the development of precision agriculture and smart irrigation, soil moisture has become a core state variable determining irrigation scheduling, crop water diagnosis, and water-fertilizer synergistic management strategies. The dynamic evolution of field soil moisture is a complex physical and biochemical process, constrained not only by intrinsic factors such as soil texture and crop root growth, but also by external environmental factors such as rainfall, artificial irrigation, evapotranspiration, solar radiation, and temperature changes. Due to the strong nonlinear coupling and significant multi-timescale characteristics among these factors, achieving high-precision and high-stability soil moisture prediction has always been a technical challenge in the field of precision agriculture.
[0003] In current soil moisture prediction technologies, time-series prediction models based on deep learning have been widely used. These models typically follow the standard gradient descent training paradigm, which involves iteratively updating all parameters within the neural network at each sampling time using backpropagation of the loss function. However, this full-parameter update mechanism has significant drawbacks: on the one hand, the number of parameters within the model is enormous, requiring the calculation of numerous gradients and multiple rounds of iterative optimization for each update, resulting in extremely high computational complexity, especially in complex network structures involving multi-scale fusion; on the other hand, forcibly updating parameters at each sampling point can easily lead to the model becoming overly sensitive to instantaneous noise, causing weight oscillations and ultimately disrupting the model's stable capture of the long-term evolution of soil moisture. Summary of the Invention
[0004] The purpose of this invention is to provide a soil moisture prediction method based on the dynamic redistribution of agricultural disturbances and feature contributions, which solves the problem of excessive computational burden caused by real-time gradient updates of all parameters when performing moisture prediction in neural networks.
[0005] To achieve the above objectives, embodiments of the present invention provide a soil moisture prediction method based on the dynamic redistribution of agricultural disturbances and characteristic contributions, the prediction method comprising: Acquire multi-source sensor data of farmland environment; Multi-scale sequences are constructed based on the multi-source sensor data to obtain standardized input features; The disturbance intensity index is calculated based on the input characteristics; Determine whether the disturbance intensity index is greater than the disturbance threshold; If the disturbance intensity index is determined to be greater than the disturbance threshold, the type of agricultural disturbance is determined based on the amount of environmental change. Based on the input features, obtain the instantaneous response strength, historical contribution level, and prediction error feedback correction term of the features; The comprehensive contribution adjustment factor is determined based on the instantaneous response intensity of the features, the type of agricultural disturbance, the degree of historical contribution, and the prediction error feedback correction term. The contribution weights of the input features are redistributed according to the comprehensive contribution adjustment factor to obtain the predicted soil moisture values for each scale sub-network. The fusion weights of subnetworks at each scale are determined based on the type of agricultural disturbance. The predicted soil moisture values of each scale sub-network are fused according to the fusion weights to obtain a fused moisture prediction value; A dynamically changing boundary is constructed to correct the fused humidity prediction value to obtain the final humidity prediction value.
[0006] Optionally, calculating the disturbance intensity index based on the input features includes: The basic environmental change indicators are determined according to formula (1). (1) The disturbance intensity index is determined according to formula (2). (2) in, For scale Down Indicators of basic environmental changes at any given time. For the first A time scale The total dimension of the input features. For scale Down Time of the first Each input feature For scale Down Time of the first Each input feature For scale Down The intensity index of the disturbance at any given time. The length of the sliding window. For scale Down Indicators of basic environmental changes at any given time. This is the time offset within the sliding window, and , For scale The following The length of the end at time is The average value of basic environmental change indicators within the window.
[0007] Optionally, the type of agricultural disturbance can be determined based on the amount of environmental change, including: The mechanism response coefficients are determined according to formula (3). (3) in, For agricultural disturbance types Next Mechanistic response coefficients of each input feature to changes in soil moisture. For scale Down The type of agricultural disturbance corresponding to each moment. For agricultural disturbance types Next to the The response direction of each input feature. For agricultural disturbance types Next The response strength corresponding to each input feature For agricultural disturbance types Next The response lag step size corresponding to each input feature For agricultural disturbance types Next The response decay coefficient corresponding to each input feature.
[0008] Optionally, the instantaneous response strength of the features, the historical contribution level, and the prediction error feedback correction term are obtained based on the input features, including: The characteristic instantaneous response intensity is determined according to formula (4). (4) The degree of correlation between input features and changes in soil moisture is determined according to formula (5). (5) The error feedback correction term is determined according to formula (6). (6) in, For scale Down Time of the first The instantaneous response strength of each input feature For scale Down Time of the first A standardized input feature, For scale Down Time of the first A standardized input feature, The summation index is used to iterate through all input features, and , It is the first stability constant. For scale Down Time of the first The degree of correlation between each input feature and changes in soil moisture The length of the evaluation window for historical contributions. The time offset within the historical contribution evaluation window, and , For scale Down Historical moment A standardized input feature, For The time is the end, The first in the historical contribution evaluation window for length The mean of each input feature, for Historical soil moisture values The average of the actual soil moisture values within the historical contribution evaluation window. It is the second stability constant. For scale Down Time of the first The prediction error feedback correction term for each input feature. For symbolic functions, For scale Down The predicted soil moisture value output by the time subnetwork. for Real-time soil moisture values.
[0009] Optionally, a comprehensive contribution adjustment factor is determined based on the instantaneous response intensity of the features, the type of agricultural disturbance, the historical contribution level, and the prediction error feedback correction term, including: The comprehensive contribution adjustment factor for each input feature is determined according to formula (7). (7) in, For scale Down Time of the first The comprehensive contribution adjustment factor of each input feature. These are the first adjustment coefficient, the second adjustment coefficient, the third adjustment coefficient, and the fourth adjustment coefficient, respectively.
[0010] Optionally, the contribution weights of the input features are reallocated according to the comprehensive contribution adjustment factor to obtain the predicted soil moisture values for each scale sub-network, including: The disturbance preservation coefficient is determined according to formula (8). (8) The contribution weights of the input features are redistributed according to formula (9). (9) The constraints are determined according to formula (10). (10) The contribution weights are normalized according to formula (11). (11) in, For scale Down The disturbance preservation coefficient at any given time. For scale Down The disturbance preservation coefficient at any given time. This is the cooling attenuation coefficient. scale after redistribution Down Time of the first The contribution weights of each input feature, For scale Down Time of the first The contribution weights of each input feature, The scale after upper and lower limit constraints Down Time of the first The contribution weights of each input feature, For the first Each input feature at scale The minimum allowable contribution weight. For the first Each input feature at scale The maximum allowable contribution weight, For scale Down Time of the first The final contribution weight of each input feature It is the third stability constant. The scale after upper and lower limit constraints Down Time of the first The contribution weights of each input feature. The summation index is used to iterate through all input features, and .
[0011] Optionally, the contribution weights of the input features are reallocated according to the comprehensive contribution adjustment factor to obtain the predicted soil moisture values for each scale sub-network, including: The weighted input features are determined according to formula (12). (12) The predicted soil moisture values for each scale sub-network are determined according to formula (13). (13) in, For scale Down Time of the first A weighted input feature, For scale Down Time of the first A standardized input feature, For scale Down The predicted soil moisture value output by the time-scale subnetwork. For scale subnetworks, For scale Down Time of the first A weighted input feature, For scale Down Time of the first A weighted input feature, For scale The corresponding historical window length.
[0012] Optionally, the fusion weights of sub-networks at each scale are determined based on the type of agricultural disturbance, including: The historical average absolute percentage error is determined according to formula (14). (14) According to formula (15), the scale bias factor under the agricultural disturbance type (15) The unnormalized fusion score is determined according to formula (16). (16) The fusion weights are determined according to formula (17). (17) in, For scale exist The historical mean absolute percentage error corresponding to that time point. To integrate the evaluation window length, For scale Down The predicted soil moisture value output by the time subnetwork. To integrate historical time offsets within the evaluation window, and , To integrate the length of the evaluation window, for Real-time soil moisture values It is the fourth stability constant. For scale Down The scale bias factor at time. This is a mapping function between the type of disturbance and the time scale. For scale Down The type of agricultural disturbance corresponding to each moment. For scale Below The non-normalized fusion score at any given moment It is the fifth stability constant. This is the scale bias adjustment coefficient. scale set The first in A time scale For scale Below Moment-wise fusion weights It is the sixth stability constant.
[0013] Optionally, a dynamically changing boundary is constructed to correct the fused humidity prediction value to obtain the final humidity prediction value, including: Construct the dynamic boundary according to formula (18). (18) The predicted fusion humidity value is corrected by directional limiting according to formula (19). (19) According to formula (20), the predicted fusion humidity value is truncated at the boundary. (20) in, for Types of agricultural disturbances The maximum permissible humidity variation under constraints Calculate the function for the changing boundary. for The type of agricultural disturbance at any given moment. For parameters related to soil type, These are parameters related to irrigation method, irrigation volume, or irrigation status. For meteorological and environmental parameters, Parameters related to crop growth status or cover conditions. After being corrected by physical constraints of the perturbation type Real-time fusion of humidity prediction values, for Real-time soil moisture values For correction functions, for Real-time fusion humidity prediction value and The difference between the actual soil moisture values at any given time. for Types of agricultural disturbances The maximum allowable increase in humidity, for Types of agricultural disturbances The maximum allowable decrease in humidity, The constraint boundary is in the direction of decreasing humidity. The boundary is the constraint for the direction of increasing humidity. After the boundary is truncated Final humidity forecast value at any time The maximum permissible soil moisture value, This is the minimum permissible soil moisture value.
[0014] Optionally, the prediction method further includes: If the disturbance intensity index is determined to be less than the disturbance threshold, the contribution weight of each feature at the current moment is used as the contribution weight of each feature at the next moment. The predicted soil moisture value for the next time step is obtained based on the contribution weights of each feature at the next time step.
[0015] Through the above technical solution, this invention provides a soil moisture prediction method based on dynamic redistribution of agricultural disturbances and feature contributions. When the disturbance intensity index exceeds the disturbance threshold, the fixed feature weights are no longer used. Instead, a comprehensive contribution adjustment factor is dynamically calculated by combining the instantaneous response intensity of features, the type of agricultural disturbance, and error feedback. Based on this, the contribution weights of the input features of each scale sub-network are redistributed in real time to obtain the moisture prediction value. This significantly reduces the online computational complexity and improves the response speed, prediction stability, and interpretability of soil moisture prediction in non-stationary agricultural environments. When the disturbance intensity index does not exceed the disturbance threshold, the current contribution weights are kept unchanged, and no redistribution calculation is performed, avoiding weight oscillations caused by high-frequency updates and reducing unnecessary computational overhead.
[0016] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description
[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a prediction method according to an embodiment of the present invention; Figure 2 This is a flowchart for determining a disturbance intensity index according to one embodiment of the present invention; Figure 3 This is a flowchart for determining the comprehensive contribution adjustment factor according to one embodiment of the present invention; Figure 4 This is a flowchart illustrating the redistribution of contribution weights of input features according to an embodiment of the present invention. Figure 5 This is a flowchart of the predicted soil moisture output by a scale subnetwork according to an embodiment of the present invention; Figure 6 This is a flowchart illustrating the determination of fusion weights according to one embodiment of the present invention; Figure 7 This is a flowchart of obtaining the final humidity prediction value according to one embodiment of the present invention. Detailed Implementation
[0018] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.
[0019] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.
[0020] Figure 1 This is a flowchart of a prediction method according to an embodiment of the present invention, in which the prediction method includes: In step S1, multi-source sensor data of the farmland environment are acquired.
[0021] In step S2, a multi-scale sequence is constructed based on multi-source sensor data to obtain standardized input features.
[0022] In step S3, the disturbance intensity index is calculated based on the input characteristics.
[0023] In step S4, it is determined whether the disturbance intensity index is greater than the disturbance threshold. If the disturbance intensity index is greater than the disturbance threshold, step S5 is executed; otherwise, step S5 is executed, where the contribution weights of each feature at the current time are used as the contribution weights of each feature at the next time, and the predicted soil moisture value for the next time is obtained based on the contribution weights of each feature at the next time. Specifically, when the disturbance intensity index is greater than the disturbance threshold, the scale is determined. The current environment is in a state of significant disturbance; otherwise, the judgment scale is determined. The current environment is in a relatively stable state. The disturbance threshold can be preset or adaptively determined based on different soil types, seasonal conditions, farmland regional characteristics, or historical statistical results.
[0024] In step S5, the type of agricultural disturbance is determined based on the amount of environmental change.
[0025] In step S6, the instantaneous response strength of the feature, the historical contribution level, and the prediction error feedback correction term are obtained based on the input features.
[0026] In step S7, the comprehensive contribution adjustment factor is determined based on the characteristic instantaneous response intensity, agricultural disturbance type, historical contribution level, and prediction error feedback correction term.
[0027] In step S8, the contribution weights of the input features are redistributed according to the comprehensive contribution adjustment factor to obtain the predicted soil moisture values for each scale sub-network.
[0028] In step S9, the fusion weights of sub-networks at each scale are determined according to the type of agricultural disturbance.
[0029] In step S10, the predicted soil moisture values of each scale sub-network are fused according to the fusion weights to obtain the fused moisture prediction value.
[0030] In step S11, a dynamically changing boundary is constructed to correct the fused humidity prediction value to obtain the final humidity prediction value.
[0031] In steps S1 to S11, this invention differs from conventional model parameter iterative update methods based on backpropagation of loss functions. It does not update all parameters within the neural network at each sampling time using gradients; instead, it triggers the redistribution of input feature contribution weights only when the perturbation intensity exceeds a perturbation threshold. From the perspective of the update object, conventional models based on backpropagation of loss functions require updating a large number of network parameters, such as weights, biases, and gating parameters. This invention updates only the input feature contribution weights online, with the update dimension being the number of input features. From the perspective of computational complexity, the complexity of existing error-minimization-based model parameter updates is related to the number of internal model parameters and the number of iterations. Since model parameters typically require multiple iterations at each sampling time, the computational complexity is extremely high, especially in complex network structures with multi-scale fusion. In contrast, the computational complexity of this invention is mainly related to the number of input features and the number of scales. It directly calculates the comprehensive contribution adjustment factor only under perturbation trigger conditions and completes the feature contribution weight redistribution based on this factor, significantly reducing online computational complexity and improving the response speed, prediction stability, and interpretability of soil moisture prediction in non-stationary agricultural environments. Furthermore, when the agricultural environment is stable, this invention directly maintains the current contribution weight without performing redistribution calculations, thus avoiding weight oscillations caused by high-frequency updates and reducing unnecessary computational overhead.
[0032] In this embodiment, the methods for acquiring multi-source sensor data of the farmland environment can be various and known to those skilled in the art. In one example of the present invention, it is achieved by relying on a multi-source sensor acquisition system deployed in the farmland environment. This multi-source sensor acquisition system includes at least a soil moisture sensor, and may further include a soil temperature sensor, an air temperature and humidity sensor, a rainfall sensor, a solar radiation sensor, an evaporation sensor, and an irrigation status sensor, to collect data on soil moisture, soil temperature, air temperature, relative humidity, rainfall, evaporation, or solar radiation intensity and irrigation status of the farmland environment. Each sensor acquires data synchronously according to a uniform sampling period, preferably 15 minutes, but can also be set to other values within the range of 5 to 30 minutes according to actual application needs.
[0033] The collected raw data undergoes preprocessing at edge computing nodes or agricultural IoT gateways. Specifically, this includes missing value imputation, outlier removal, timestamp alignment, data format standardization, and chronological storage as multidimensional time-series data. After preprocessing, at time... A unified multidimensional observation vector is formed at this point: ,in, for The original multidimensional observation vector at time t, For the first observed variable in The observed value at time, For the second observed variable in The observed value at time, For the first Each observed variable in The observed value at time, The total dimension of the observed variables, These are discrete sampling times. Therefore, the original multidimensional time series data set can be obtained: ,in, The original multidimensional time series data set, This is the original multidimensional observation vector at the first discrete sampling time. This is the original multidimensional observation vector at the second discrete sampling time. For the first The original multidimensional observation vector at each discrete sampling time. This represents the total length of all historical observations.
[0034] In this implementation, a multi-scale sequence is constructed based on the original multi-dimensional time-series data set. Let the scale set be: ,in, It is a set consisting of all time scales. For the first time scale, For the second time scale, For the first A time scale This represents the total number of time scales. For each scale... Preset a corresponding history window length ,in The unit is the number of sampling steps. If the sampling period is 15 minutes, a preferred setting for different scales is shown in Table 1: Table 1. Historical window lengths at different scales
[0035] exist At any given moment, the scale The corresponding original historical sequence is defined as: ,in, for Time, Scale The corresponding original multidimensional observation vector, for Time, Scale The corresponding original multidimensional observation vector, for Time, Scale The corresponding original multidimensional observation vector, for Time, Scale The corresponding original multidimensional observation vector, For scale The corresponding historical window length.
[0036] To improve the comparability of features with different dimensions and enhance the stability of subsequent perturbation identification and prediction modeling, in this embodiment, the present invention standardizes the input features at each scale. For the scale... Next Each feature at time The original value at the location Its standardized result is defined as: ,(twenty one) in, For scale Down Time of the first A standardized input feature, For scale Next Each input feature in The observation at time t, For scale Next Each input feature corresponds to the mean within a historical window. For scale Next Each input feature corresponds to the standard deviation within a historical window. It is the seventh stability constant. On the scale... The standardized multi-scale input sequence is then: , For scale The down-standardized multi-scale input sequence, For scale Down Standardized input features at each time step. For scale Down Time-standardized input features For scale The corresponding historical window length.
[0037] In agricultural environments, soil moisture changes are often significantly disturbed by factors such as sudden rainfall, concentrated irrigation, and sustained high-temperature evaporation. If a fixed feature contribution relationship is used for prediction, moisture prediction is prone to response lag or error accumulation. Therefore, this invention constructs an agricultural disturbance intensity index based on multi-scale input sequences to characterize the degree of change in the current environmental state and further identify agricultural disturbance types, providing a basis for subsequent dynamic redistribution of feature contributions. Specifically, such as... Figure 2As shown, it includes the following steps: For scale Next Each input feature is defined according to formula (22) to represent the standardized variation magnitude between two adjacent time points. ,(twenty two) in, For scale Next One feature in The instantaneous change in time, vertical line This represents absolute value operations, used to eliminate the influence of the direction of change and retain only the intensity of the change.
[0038] Furthermore, the magnitudes of change of each characteristic are aggregated. In step S31, the basic environmental change index is determined according to formula (1). (1) To reduce the impact of single-measurement noise on the disturbance determination result, the present invention further specifies a length of... Within a sliding window, fluctuation enhancement processing is applied to the basic environmental change index to obtain a smoothed disturbance intensity index. In step S32, the disturbance intensity index is determined according to formula (2). (2) in, For scale Down Indicators of basic environmental changes at any given time. For the first A time scale The total dimension of the input features. For scale Down Time of the first Each input feature For scale Down Time of the first Each input feature For scale Down The intensity index of the disturbance at any given time. The length of the sliding window. For scale Down Indicators of basic environmental changes at any given time. This is the time offset within the sliding window, and , For scale The following The length of the end at time is The average value of basic environmental change indicators within the window.
[0039] In this embodiment, after determining that the current environment is in a state of significant disturbance, the present invention further identifies the disturbance type based on the changing characteristics of different agricultural environmental variables. Let the set of agricultural disturbance types be: ,in, This is a set of agricultural disturbance types. This indicates a state of no significant disturbance or stability. This indicates a disturbance caused by rainfall infiltration. This indicates an irrigation-related disturbance. This indicates a high-temperature evaporation-type disturbance. This indicates a complex disturbance. Specifically, when rainfall or the magnitude of rainfall variation exceeds a preset threshold, it is classified as a rainfall infiltration disturbance; when the irrigation state variable changes from an unirrigated state to an irrigated state, or the irrigation flow exceeds a preset threshold, it is classified as an irrigation replenishment disturbance; when air temperature, solar radiation, or evaporation continuously increases while air humidity decreases, it is classified as a high-temperature evaporation disturbance; when two or more disturbance conditions are met simultaneously, it is classified as a complex disturbance; when the significant disturbance condition is not met, it is classified as a state with no significant disturbance or a stable state. For ease of subsequent calculations, the scale is [not specified]. Next moment The disturbance type determination result is recorded as: ,in, For scale Down The type of agricultural disturbance corresponding to a given moment can also be understood as the scale. Down Moment Agricultural-like disturbance types. In this invention, not only can disturbance intensity indices at different time scales be obtained, but also... It can also obtain the corresponding agricultural disturbance type. The perturbation intensity is used to determine whether to initiate the subsequent dynamic redistribution process of feature contributions, while the perturbation type is used to determine the direction of contribution adjustment for different input features in the current agricultural scenario, thus providing a direct basis for the subsequent dynamic redistribution mechanism of feature contributions.
[0040] To reflect the differentiated impacts of various agricultural disturbance events on soil moisture changes, this invention constructs an agricultural disturbance type-time scale-input feature response tensor: ,in, The agricultural disturbance type-timescale-input feature response tensor. For the first A time scale For the first Each input feature For the first Agricultural disturbance types of input features The corresponding mechanistic response coefficients. By constructing the above response tensor, this invention can determine the contribution adjustment direction of different input features according to the current agricultural scenario, thereby avoiding blind updates based solely on error gradients. Furthermore, for different perturbation types, the feature response relationship can be set according to the agricultural moisture change mechanism. Specifically, as shown in Table 2: Table 2. Relationship between agricultural disturbance type and the enhancement or reduction of contribution characteristics.
[0041] Unlike methods that only set a single weight for input features, this response tensor considers three factors simultaneously: agricultural disturbance type, time scale, and input features. For example, under rainfall infiltration disturbances, rainfall amount and soil moisture change rate show strong positive responses at short scales; under high-temperature evaporation disturbances, air temperature, solar radiation, and evaporation show strong responses at medium scales; and under steady-state conditions, long-scale historical soil moisture and soil temperature have higher reference value. In a preferred embodiment, the mechanism response coefficient... It is determined by the response direction, response intensity, response hysteresis step size, and response duration decay coefficient. Specifically, it includes: The mechanism response coefficients are determined according to formula (3). (3) in, For agricultural disturbance types Next Mechanistic response coefficients of each input feature to changes in soil moisture. For scale Down The type of agricultural disturbance corresponding to each moment. For agricultural disturbance types Next to the The response direction of each input feature, when Time represents the contribution to enhancing the input feature, when When this indicates a reduction in the contribution of the input feature, The time represents the contribution of preserving the input feature. For agricultural disturbance types Next The response strength corresponding to each input feature For agricultural disturbance types Next The response lag step size corresponding to each input feature is used to characterize the lag time after a disturbance occurs, which affects the contribution relationship of that input feature. For agricultural disturbance types Next The response decay coefficient corresponding to each input feature is used to characterize the degree to which the impact of the disturbance gradually weakens over time. Therefore, this feature response tensor is not only used to provide the mechanistic response coefficients for adjusting the input feature contribution, but also to describe the response direction, response intensity, response hysteresis, and degree of persistent impact of different agricultural disturbance types on each input feature at different time scales, thereby improving the interpretability and scenario adaptability of the dynamic redistribution process of feature contribution.
[0042] In determining the current scale When the environment is under significant disturbance, according to the type of disturbance The corresponding feature response relationship is invoked, and combined with the instantaneous change intensity of the input features, their historical contribution, and prediction error feedback, the comprehensive contribution adjustment factor of each input feature is calculated. Specifically, such as... Figure 3 As shown, it includes the following steps: In step S61, the characteristic instantaneous response intensity is determined according to formula (4). (4) When a certain feature changes significantly under the current disturbance state, its A larger value indicates that this feature is more responsive to current environmental changes.
[0043] To avoid misjudgment caused by a single instantaneous change, this invention further introduces the historical contribution level. In step S62, the correlation contribution level between the input feature and the soil moisture change is determined according to formula (5). (5) To ensure that the feature contribution redistribution process reflects recent prediction bias, this invention introduces a prediction error feedback correction term. The scale is determined according to formula (23). Down Prediction error at time, ,(twenty three) in, For scale Down Prediction error at time, For scale Down The predicted soil moisture value output by the time subnetwork. for Real-time soil moisture values.
[0044] Further, in step S63, the error feedback correction term is determined according to formula (6). (6) This error feedback correction term is only used as a correction factor for feature contribution redistribution and is not used to directly perform gradient descent updates of the model's internal parameters.
[0045] in, For scale Down Time of the first The instantaneous response strength of each input feature For scale Down Time of the first A standardized input feature, For scale Down Time of the first A standardized input feature, The summation index is used to iterate through all input features, and , This is the first stability constant, used to prevent small positive numbers with a denominator of zero. For scale Down Time of the first The degree of correlation contribution of each input feature to soil moisture changes is used to characterize the explanatory power of a certain input feature for soil moisture changes within a recent historical window. The length of the evaluation window for historical contributions. The time offset within the historical contribution evaluation window, and , For scale Down Historical moment A standardized input feature, For The time is the end, The first in the historical contribution evaluation window for length The mean of each input feature, for Historical soil moisture values The average of the actual soil moisture values within the historical contribution evaluation window. It is the second stability constant. For scale Down Time of the first The prediction error feedback correction term for each input feature. The sign function indicates that the feature contribution is reversed based on whether the prediction is too high or too low. For scale Down The predicted soil moisture value output by the time subnetwork. for Real-time soil moisture values For scale No. The standardized variation of each feature.
[0046] Taking into account the instantaneous response strength of the features, the guidance of the disturbance type mechanism, the degree of historical contribution, and the prediction error feedback, the comprehensive contribution adjustment factor is determined. The comprehensive contribution adjustment factor for each input feature is determined according to formula (7). (7) in, For scale Down Time of the first The comprehensive contribution adjustment factor of each input feature. They are the first adjustment coefficient, the second adjustment coefficient, the third adjustment coefficient, and the fourth adjustment coefficient, respectively, and satisfy the following conditions: , The absolute value is used to represent the explanatory power of the relevant contribution strength. The comprehensive contribution adjustment factor determined in this invention is not a single error gradient, but is jointly determined by agricultural disturbance mechanisms, feature change responses, historical contributions, and error feedback. It is used to characterize the adjustment direction and magnitude of the input feature contribution weights under the current agricultural scenario. By introducing an adjustment coefficient, the adjustment of feature weights can both quickly respond to sudden disturbances and maintain stable tracking of long-term trends. Combined with subsequent disturbance preservation coefficients and weight constraint mechanisms, high accuracy, high stability, and strong interpretability of soil moisture prediction are ultimately achieved.
[0047] Considering that agricultural disturbances often have a certain degree of persistence—for example, changes in soil moisture do not immediately cease after rainfall infiltration, and the impact of high-temperature evaporation on soil moisture may continue for multiple sampling periods—this invention establishes a disturbance retention coefficient. This is used to control the persistent impact of perturbation events on the feature contribution weights. After obtaining the perturbation coefficients, the contribution weights of the input features are redistributed based on the comprehensive contribution adjustment factor and the perturbation preservation coefficient. Specifically, such as... Figure 4 As shown, it includes the following steps: In step S81, the disturbance retention coefficient is determined according to formula (8). (8) Among them, the current is time.
[0048] In this invention, the feature contribution weights corresponding to each scale sub-network have a recursive update relationship. During the initialization phase, the contribution weights of each input feature can be set to the same weight, or the initial contribution weights can be preset based on historical data, expert experience, and crop type. In subsequent prediction processes, the feature contribution weights obtained at the previous or current time are used as the update starting point, and are redistributed in conjunction with the perturbation preservation coefficient and the comprehensive contribution adjustment factor. Specifically, in step S82, the contribution weights of the input features are redistributed according to formula (9). (9) To ensure that the feature contribution weights have physical meaning and interpretability, this invention further sets upper and lower limit constraints. In step S83, the constraint conditions are determined according to formula (10). (10) In step S84, the contribution weights are normalized according to formula (11). (11) in, For scale Down The disturbance preservation coefficient at any given time. For scale Down The disturbance preservation coefficient at any given time. The cooling attenuation coefficient and The disturbance occurrence period indicates that the conditions for a significant disturbance are still met at the current moment, while the disturbance cooling period indicates the stage where the significant disturbance has just ended but its effects have not yet completely disappeared. scale after redistribution Down Time of the first The contribution weights of each input feature. For scale Down Time of the first The contribution weights of each input feature. The scale after upper and lower limit constraints Down Time of the first The contribution weights of each input feature. For the first Each input feature at scale The minimum allowable contribution weight. For the first Each input feature at scale The maximum allowable contribution weight, For scale Down Time of the first The final contribution weight of each input feature It is the third stability constant. The scale after upper and lower limit constraints Down Time of the first The contribution weights of each input feature. The summation index is used to iterate through all input features, and , To find the minimum value function, This is the function for finding the maximum value.
[0049] In steps S81 to S84, the perturbation hold-and-cool-down adjustment prevents the feature contribution weights from abruptly changing at the moment the perturbation ends, thus improving the stability of the prediction process. Upper and lower limit constraints confine the weights within a physically reasonable range, preventing individual features from dominating the prediction due to instantaneous noise or extreme adjustments, thereby improving the accuracy of soil moisture prediction.
[0050] In this embodiment, the methods for obtaining the predicted soil moisture value can be various and known to those skilled in the art. In a preferred example of the present invention, in order to learn the soil moisture change patterns at different time scales, the present invention performs calculations for each scale. A corresponding scale sub-network is constructed, and multiple scale sub-networks are arranged in parallel to form a multi-scale, multi-head sequence prediction structure. Each scale sub-network receives a normalized historical sequence at its corresponding scale and, combined with the feature contribution weights at that scale, outputs the predicted soil moisture for the next time step. Specifically, it can be as follows: Figure 5 The method shown. Figure 5 The prediction method also includes: In step S85, the weighted input features are determined according to formula (12). (12) In step S86, the predicted soil moisture values for each scale sub-network are determined according to formula (13). (13) in, For scale Down Time of the first A weighted input feature, For scale Down Time of the first A standardized input feature, For scale Down The predicted soil moisture value output by the time-scale subnetwork. For scale subnetworks, For scale Down Time of the first A weighted input feature, For scale Down Time of the first A weighted input feature, For scale The corresponding historical window length.
[0051] The above This can be implemented using gated recurrent unit networks, one-dimensional temporal convolutional networks, temporal perceptron networks, or other lightweight network structures capable of processing temporal data. In a preferred example of this invention, the scale sub-network... It is implemented using a gated cyclic unit network. For scale... At this scale The time is the end, and the length is The weighted input sequence is input into the gated recurrent unit network, and the weighted input sequence is: ,in, Representing scale The following The weighted input sequence at time 1 is the end. Representing scale The next historical moment The weighted multidimensional input feature vector, , Representing scale The corresponding historical window length. The gated recurrent unit network extracts the time-dependent features of soil moisture changes at this scale through update and reset gates, obtaining the terminal hidden state. Then, the output layer at that scale is obtained. Predicted soil moisture at any given time: ,(twenty four) in, For scale Subnet output Predicted soil moisture values at any time For scale Lower gated recurrent unit network in The hidden state at the end of the output at any time. For output layer weights, The output layer is biased. At different scales, each scale sub-network has a different input window length. The short-scale sub-network is used to learn rapid humidity changes caused by rainfall or irrigation, the medium-scale sub-network is used to learn periodic humidity changes caused by intraday evaporation, temperature and air humidity changes, and the long-scale sub-network is used to learn soil water storage capacity and long-term humidity evolution trends caused by continuous rainfall or continuous drought.
[0052] In steps S85 to S86, the present invention can enable short-scale models to have higher weight in sudden disturbance scenarios such as rainfall and irrigation, enable mesoscale models to play a role in scenarios with continuous changes in evaporation and temperature and humidity, and enable long-scale models to provide trend constraints in stable scenarios, thereby improving the adaptability of soil moisture prediction in non-stable agricultural environments.
[0053] Obtaining the output of subnetworks at various scales Subsequently, this invention does not directly select a result at a specific scale as the final predicted value. Instead, it dynamically allocates scale fusion weights based on the prediction performance of each scale sub-network within a recent historical window, and considers the differences in the impact of the current agricultural disturbance type on different time scales, thereby forming a unified fusion prediction result. Specifically, it can be as follows: Figure 6 The method shown. Figure 6 The prediction method also includes: To quantitatively characterize the prediction reliability of each scale sub-network within the recent historical window, in step S91, the historical mean absolute percentage error is determined according to formula (14). (14) In step S92, according to formula (15), the scale bias factor under the agricultural disturbance type, (15) In a preferred embodiment of the present invention, the perturbation type scale bias factor can be determined according to Table 3. Specifically, it includes: Table 3 Correspondence between Agricultural Disturbance Types and Priority Enhancement Scales
[0054] Among them, short scale, medium scale, and long scale correspond to scale sets, respectively. The specific length of the historical window can be determined based on the sampling period, crop type, soil type, and regional climate conditions. For sandy soils, due to their rapid infiltration and weak water retention capacity, soil moisture changes quickly after rainfall or irrigation. Short-scale windows can be set to 0.5 to 1 hour, medium-scale windows to 12 to 24 hours, and long-scale windows to 3 to 7 days. For loam, due to its moderate infiltration and water retention capacity, short-scale windows can be set to 1 to 2 hours, medium-scale windows to 24 hours, and long-scale windows to approximately 7 days. For clay soils, due to their slower infiltration and stronger water retention capacity, soil moisture changes exhibit a stronger lag. Short-scale windows can be set to 2 to 3 hours, medium-scale windows to 24 to 48 hours, and long-scale windows to 7 to 14 days. For leafy vegetables and shallow-rooted crops, the short- and medium-scale windows can be appropriately shortened because root zone water changes are relatively rapid. For deep-rooted crops such as fruit trees and corn, the medium- and long-scale windows can be appropriately lengthened because root zone water changes are relatively slow. In hot and arid regions, the weight of the medium-scale window can be appropriately increased to characterize intraday evaporation and continuous water consumption processes. In regions with frequent rainfall, the short-scale window can be appropriately shortened to improve the response speed to rainfall infiltration disturbances.
[0055] Taking into account historical prediction performance and perturbation type scale bias, in step S93, the unnormalized fusion score is determined according to formula (16). (16) When a certain scale has a small error in recent historical predictions and the scale is more consistent with the current type of agricultural disturbance, its unnormalized fusion score increases accordingly.
[0056] In step S94, the fusion weights are determined according to formula (17). (17) in, For scale exist The historical mean absolute percentage error corresponding to that time point. To integrate the evaluation window length, For scale Down The predicted soil moisture value output by the time subnetwork. To integrate historical time offsets within the evaluation window, and , To integrate the length of the evaluation window, for Real-time soil moisture values It is the fourth stability constant. For scale Down The scale bias factor at time. This is a mapping function between the type of disturbance and the time scale. For scale Down The type of agricultural disturbance corresponding to each moment. For scale Below The non-normalized fusion score at any given moment It is the fifth stability constant. Rate historical performance. This is the scale bias adjustment coefficient, used to control the degree of influence of the perturbation type on the fusion weights. scale set The first in A time scale For scale Below The fusion weights at all times are non-negative and sum to 1. It is the sixth stability constant. This represents absolute value operations.
[0057] In steps S91 to S94, the present invention achieves the synergistic fusion of historical prediction performance and agricultural disturbance types, enabling prediction results at different time scales to be adaptively combined according to the dynamic changes in the current agricultural environment, thereby achieving a balance between rapid disturbance response and long-term trend stability.
[0058] After obtaining the fusion weights, the outputs of each sub-network are weighted and summed based on the fusion weights at each scale. Specifically, the fused humidity prediction value is obtained according to formula (25). (25) in, for Real-time humidity prediction values are integrated.
[0059] Although soil moisture prediction results can be obtained after agricultural disturbance perception, dynamic redistribution of feature contributions, and multi-scale dynamic fusion, soil moisture, as an agricultural state variable with clear physical meaning, is constrained by factors such as soil type, rainfall infiltration, irrigation replenishment, evaporation loss, and crop cover status. If the fused output exhibits instantaneous jumps that do not conform to the actual water evolution pattern, it may affect the stability of subsequent precision irrigation control. Therefore, this invention further introduces a physical correction mechanism for soil moisture under disturbance type constraints based on the fused prediction results. Specifically, such as... Figure 7 As shown, it includes the following steps: Considering the varying ability of soil moisture to change under different types of agricultural disturbance, this invention does not use a fixed limiting constant, but instead constructs dynamically changing boundary parameters based on the type of agricultural disturbance. In step S111, the dynamic boundary is constructed according to formula (18). (18) Since the permissible ranges for soil moisture increase and decrease are not entirely the same under different disturbance types, this invention further sets moisture increase and decrease boundaries and performs directional limiting correction on the original predicted change. In a preferred embodiment of this invention, the dynamic change boundary can be set in segments according to the disturbance type. When the disturbance is of the rainfall infiltration type, a large increase in soil moisture is permissible in a short period of time; when For irrigation-related disturbances, increased humidity is permissible but limited by irrigation volume and soil infiltration capacity; when When the disturbance is of the high-temperature evaporative type, it mainly limits the rate of humidity decrease; when When the humidity is in a stable state, sudden and large changes in humidity are strictly limited; when When the disturbance is a complex disturbance, the boundary of change is determined based on multiple disturbance factors. Specifically, this includes: (27) (28) in, This is the uncorrected maximum humidity rise boundary. This represents the uncorrected maximum humidity drop boundary. This serves as the basic boundary for the direction of humidity increase. This forms the basic boundary for the direction of humidity decrease. and These represent the maximum allowable increase and decrease in humidity under steady-state conditions, respectively, and are typically less than the corresponding baseline change boundaries. for Rainfall amount or intensity at any given time, for The amount of irrigation, the flow rate, or the irrigation status at any given time. for Evaporation rate at any given time for The intensity of solar radiation at any given time. for air temperature at any given moment for The relative humidity of the air at any given time. , , , , and These are the boundary adjustment coefficients corresponding to rainfall, irrigation, evaporation, solar radiation, air temperature, and air humidity, respectively.
[0060] To further consider the impact of soil type on its ability to change moisture, a soil adjustment coefficient can be introduced. The above boundaries are then corrected: (29) (30) The corrected dynamic boundary can be expressed as: , The corrected maximum humidity rise threshold, This is the corrected maximum humidity decrease boundary. This is the adjustment coefficient of soil infiltration capacity to the moisture rise boundary. This is the adjustment coefficient of soil water retention capacity to the boundary of humidity decrease. That is to For sandy soils, where moisture changes rapidly, the boundary of variation can be appropriately increased. For clay soils, where moisture changes slowly, the boundary of variation can be appropriately decreased. For loam, a moderate adjustment coefficient can be used.
[0061] In step S112, the directional limiting correction is performed on the fused humidity prediction value according to formula (19). (19) The predicted change is determined according to formula (26). (26) To ensure that the predicted values always remain within a reasonable physical range of soil moisture, this invention further performs boundary truncation. In step S113, boundary truncation is performed on the fused moisture prediction values according to formula (20). (20) in, for Types of agricultural disturbances The maximum permissible humidity variation under constraints, i.e., the corrected maximum permissible humidity variation. Calculate the function for the changing boundary. for The type of agricultural disturbance at any given moment. For parameters related to soil type, These are parameters related to irrigation method, irrigation volume, or irrigation status. For meteorological and environmental parameters, Parameters related to crop growth status or cover conditions. After being corrected by physical constraints of the perturbation type Real-time fusion of humidity prediction values, for Real-time soil moisture values For correction functions, Indicates the variable Limited to the range Inside, for Real-time fusion humidity prediction value and The difference between the actual soil moisture values at any given time, if This indicates a predicted increase in humidity. This indicates a predicted decrease in humidity. for Real-time fusion of humidity prediction values, for Types of agricultural disturbances The maximum allowable increase in humidity, for Types of agricultural disturbances The maximum allowable decrease in humidity, i.e., the corrected maximum decrease in humidity. That is to , The constraint boundary is in the direction of decreasing humidity. The boundary is the constraint for the direction of increasing humidity. After the boundary is truncated Final humidity forecast value at any time The maximum permissible soil moisture value, The minimum permissible soil moisture value, and The settings can be configured based on sensor range, soil properties, suitable moisture range for crops, or engineering experience. To find the minimum value function, This is the function for finding the maximum value.
[0062] In steps S111 to S113, the final humidity prediction value obtained not only retains the ability of the multi-scale prediction model to express complex agricultural environmental changes, but also meets the physical boundary constraints that soil moisture changes should follow under different types of agricultural disturbances. Thus, it can more safely and stably serve precision irrigation control, agricultural environmental monitoring, and intelligent regulation and control systems for facility agriculture.
[0063] Furthermore, when the environment is stable, the feature contribution weights are kept relatively stable to avoid weight oscillations caused by frequent adjustments. Specifically, this includes: using the contribution weights of each feature at the current moment as the contribution weights of each feature at the next moment, and obtaining the predicted soil moisture value for the next moment based on the contribution weights of each feature at the next moment.
[0064] Through the above technical solution, this invention provides a soil moisture prediction method based on dynamic redistribution of agricultural disturbances and feature contributions. When the disturbance intensity index exceeds the disturbance threshold, the fixed feature weights are no longer used. Instead, a comprehensive contribution adjustment factor is dynamically calculated by combining the instantaneous response intensity of features, the type of agricultural disturbance, and error feedback. Based on this, the contribution weights of the input features of each scale sub-network are redistributed in real time to obtain the moisture prediction value. This significantly reduces the online computational complexity and improves the response speed, prediction stability, and interpretability of soil moisture prediction in non-stationary agricultural environments. When the disturbance intensity index does not exceed the disturbance threshold, the current contribution weights are kept unchanged, and no redistribution calculation is performed, avoiding weight oscillations caused by high-frequency updates and reducing unnecessary computational overhead.
[0065] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0066] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.
Claims
1. A method for predicting soil moisture based on dynamic redistribution of agricultural disturbances and characteristic contributions, characterized in that, The prediction method includes: Acquire multi-source sensor data of farmland environment; Multi-scale sequences are constructed based on the multi-source sensor data to obtain standardized input features; The disturbance intensity index is calculated based on the input characteristics; Determine whether the disturbance intensity index is greater than the disturbance threshold; If the disturbance intensity index is determined to be greater than the disturbance threshold, the type of agricultural disturbance is determined based on the amount of environmental change. Based on the input features, obtain the instantaneous response strength, historical contribution level, and prediction error feedback correction term of the features; The comprehensive contribution adjustment factor is determined based on the instantaneous response intensity of the features, the type of agricultural disturbance, the degree of historical contribution, and the prediction error feedback correction term. The contribution weights of the input features are redistributed according to the comprehensive contribution adjustment factor to obtain the predicted soil moisture values for each scale sub-network. The fusion weights of subnetworks at each scale are determined based on the type of agricultural disturbance. The predicted soil moisture values of each scale sub-network are fused according to the fusion weights to obtain a fused moisture prediction value; A dynamically changing boundary is constructed to correct the fused humidity prediction value to obtain the final humidity prediction value.
2. The prediction method according to claim 1, characterized in that, The disturbance intensity index is calculated based on the input features, including: The basic environmental change indicators are determined according to formula (1). ，（1） The disturbance intensity index is determined according to formula (2). ，（2） in, For scale Down Indicators of basic environmental changes at any given time. For the first A time scale The total dimension of the input features. For scale Down Time of the first Each input feature For scale Down Time of the first Each input feature For scale Down The intensity index of the disturbance at any given time. The length of the sliding window. For scale Down Indicators of basic environmental changes at any given time. This is the time offset within the sliding window, and , For scale The following The length of the end at time is The average value of basic environmental change indicators within the window.
3. The prediction method according to claim 2, characterized in that, Agricultural disturbance types are determined based on environmental changes, including: The mechanism response coefficients are determined according to formula (3). ，（3） in, For agricultural disturbance types Next Mechanistic response coefficients of each input feature to changes in soil moisture. For scale Down The type of agricultural disturbance corresponding to each moment. For agricultural disturbance types Next to the The response direction of each input feature. For agricultural disturbance types Next The response strength corresponding to each input feature For agricultural disturbance types Next The response lag step size corresponding to each input feature For agricultural disturbance types Next The response decay coefficient corresponding to each input feature.
4. The prediction method according to claim 3, characterized in that, Based on the input features, the instantaneous response strength, historical contribution level, and prediction error feedback correction term are obtained, including: The characteristic instantaneous response intensity is determined according to formula (4). ，（4） The degree of correlation between input features and changes in soil moisture is determined according to formula (5). ，（5） The error feedback correction term is determined according to formula (6). ，（6） in, For scale Down Time of the first The instantaneous response strength of each input feature For scale Down Time of the first A standardized input feature, For scale Down Time of the first A standardized input feature, The summation index is used to iterate through all input features, and , It is the first stability constant. For scale Down Time of the first The degree of correlation between each input feature and changes in soil moisture The length of the evaluation window for historical contributions. The time offset within the historical contribution evaluation window, and , For scale Down Historical moment A standardized input feature, For The time is the end, The first in the historical contribution evaluation window for length The mean of each input feature, for Historical soil moisture values The average of the actual soil moisture values within the historical contribution evaluation window. It is the second stability constant. For scale Down Time of the first The prediction error feedback correction term for each input feature. For symbolic functions, For scale Down The predicted soil moisture value output by the time subnetwork. for Real-time soil moisture values.
5. The prediction method according to claim 4, characterized in that, The comprehensive contribution adjustment factor is determined based on the instantaneous response intensity of the features, the type of agricultural disturbance, the historical contribution level, and the prediction error feedback correction term, including: The comprehensive contribution adjustment factor for each input feature is determined according to formula (7). ，（7） in, For scale Down Time of the first The comprehensive contribution adjustment factor of each input feature. These are the first adjustment coefficient, the second adjustment coefficient, the third adjustment coefficient, and the fourth adjustment coefficient, respectively.
6. The prediction method according to claim 5, characterized in that, The contribution weights of the input features are redistributed according to the comprehensive contribution adjustment factor to obtain the predicted soil moisture values for each scale sub-network, including: The disturbance preservation coefficient is determined according to formula (8). ，（8） The contribution weights of the input features are redistributed according to formula (9). ，（9） The constraints are determined according to formula (10). ，（10） The contribution weights are normalized according to formula (11). ，（11） in, For scale Down The disturbance preservation coefficient at any given time. For scale Down The disturbance preservation coefficient at any given time. This is the cooling attenuation coefficient. scale after redistribution Down Time of the first The contribution weights of each input feature. For scale Down Time of the first The contribution weights of each input feature. The scale after upper and lower limit constraints Down Time of the first The contribution weights of each input feature. For the first Each input feature at scale The minimum allowable contribution weight. For the first Each input feature at scale The maximum allowable contribution weight, For scale Down Time of the first The final contribution weight of each input feature It is the third stability constant. The scale after upper and lower limit constraints Down Time of the first The contribution weights of each input feature. The summation index is used to iterate through all input features, and .
7. The prediction method according to claim 6, characterized in that, The contribution weights of the input features are redistributed according to the comprehensive contribution adjustment factor to obtain the predicted soil moisture values for each scale sub-network, including: The weighted input features are determined according to formula (12). ，（12） The predicted soil moisture values for each scale sub-network are determined according to formula (13). ，（13） in, For scale Down Time of the first A weighted input feature, For scale Down Time of the first A standardized input feature, For scale Down The predicted soil moisture value output by the time-scale subnetwork. For scale subnetworks, For scale Down Time of the first A weighted input feature, For scale Down Time of the first A weighted input feature, For scale The corresponding historical window length.
8. The prediction method according to claim 7, characterized in that, The fusion weights of subnetworks at each scale are determined based on the type of agricultural disturbance, including: The historical average absolute percentage error is determined according to formula (14). ，（14） According to formula (15), the scale bias factor under the agricultural disturbance type ，（15） The unnormalized fusion score is determined according to formula (16). ，（16） The fusion weights are determined according to formula (17). ，（17） in, For scale exist The historical mean absolute percentage error corresponding to that time point. To integrate the evaluation window length, For scale Down The predicted soil moisture value output by the time subnetwork. To integrate historical time offsets within the evaluation window, and , To integrate the length of the evaluation window, for Real-time soil moisture values It is the fourth stability constant. For scale Down The scale bias factor at time. This is a mapping function between the type of disturbance and the time scale. For scale Down The type of agricultural disturbance corresponding to each moment. For scale Below The non-normalized fusion score at any given moment It is the fifth stability constant. This is the scale bias adjustment coefficient. scale set The first in A time scale For scale Below Moment-wise fusion weights It is the sixth stability constant.
9. The prediction method according to claim 8, characterized in that, The process of constructing a dynamically changing boundary to correct the fused humidity prediction value to obtain the final humidity prediction value includes: Construct the dynamic boundary according to formula (18). ，（18） The predicted fusion humidity value is corrected by directional limiting according to formula (19). ，（19） According to formula (20), the predicted fusion humidity value is truncated at the boundary. ，（20） in, for Types of agricultural disturbances The maximum permissible humidity variation under constraints Calculate the function for the changing boundary. for The type of agricultural disturbance at any given moment. For parameters related to soil type, These are parameters related to irrigation method, irrigation volume, or irrigation status. For meteorological and environmental parameters, Parameters related to crop growth status or cover conditions. After being corrected by physical constraints of the perturbation type Real-time fusion of humidity prediction values, for Real-time soil moisture values For correction functions, for Real-time fusion humidity prediction value and The difference between the actual soil moisture values at any given time. for Types of agricultural disturbances The maximum allowable increase in humidity, for Types of agricultural disturbances The maximum allowable decrease in humidity, The constraint boundary is in the direction of decreasing humidity. The boundary is the constraint for the direction of increasing humidity. After the boundary is truncated Final humidity forecast value at any time The maximum permissible soil moisture value, This is the minimum permissible soil moisture value.
10. The prediction method according to claim 2, characterized in that, The prediction method further includes: If the disturbance intensity index is determined to be less than the disturbance threshold, the contribution weight of each feature at the current moment is used as the contribution weight of each feature at the next moment. The predicted soil moisture value for the next time step is obtained based on the contribution weights of each feature at the next time step.