Agricultural water intelligent decision-making method and device based on big data driving

Abstract

The invention relates to the technical field of artificial intelligence and data processing, in particular to an agricultural water intelligent decision-making method and device based on big data driving, and the method specifically comprises the following steps: collecting soil data and meteorological monitoring data, recording crop growth cycle data, then forming a data set, and carrying out the manual marking of a field; performing normalization processing on the data collected in the data set; the method comprises the following steps: constructing an agricultural water intelligent decision classification model, constructing a cost-sensitive feature cross mapping network, calculating a loss function of an extreme learning machine, carrying out parameter optimization of the extreme learning machine based on evapotranspiration constraint, dynamically adjusting a threshold value, and carrying out iterative training on the model for many times to obtain a trained model; newly collected data is preprocessed and then input into the trained agricultural water intelligent decision classification model for processing, and finally a decision result is output. According to the invention, the problems of precision, robustness and interpretability in agricultural water management can be solved.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and data processing, and particularly relates to an intelligent decision-making method and device for agricultural water use driven by big data. Background Art

[0002] Focusing on the ecological protection and high-quality development of the Yellow River Basin is a major national strategy. To implement this major national strategy, it can be specifically implemented in science and technology for water conservation, digital water governance, intelligent water management, and scientific water use. The integration of water conservancy development and information technology can be used for comprehensive management of water conservancy. Among them, water source management in agricultural production is a complex and key issue in intelligent water management and scientific water supply. A reasonable irrigation decision can not only increase crop yields but also effectively save water resources. However, the decision-making for agricultural water use faces various challenges. In particular, there are significant differences in water requirements at different growth stages of crops, and the spatio-temporal heterogeneity and multimodal characteristics of soil and meteorological data make irrigation demand prediction more complex.

[0003] Most traditional agricultural irrigation decision-making methods rely on fixed rules or simple statistical models, ignoring the non-linear coupling relationship between crop growth stages, soil properties, and meteorological conditions. In addition, traditional data processing methods such as global normalization often cannot effectively retain the water demand patterns of different growth periods and are difficult to adapt to problems such as uneven sensor distribution and data noise. There are significant deficiencies in the accuracy and adaptability of existing technologies when processing agricultural data, which affects the scientific nature and effectiveness of irrigation decisions.

[0004] The problems existing in the prior art are as follows: Most of the existing technologies adopt global normalization or simple time series alignment methods, which cannot effectively process the spatio-temporal heterogeneity of agricultural data and the differences in different crop growth stages, resulting in the inability of traditional methods to retain key farmland water characteristics and affecting the accuracy and adaptability of decisions. Traditional agricultural water use decision-making models are insensitive to the misclassification costs of different categories, especially the misjudgment cost of the "water shortage" category is relatively high, and the existing methods cannot accurately balance the cost differences between different categories. Many existing agricultural water use decision-making models rely on data-driven and do not consider the physical constraints in the agricultural field, resulting in the possibility that the model may output results that do not conform to agricultural common sense, making the model lack interpretability and scientific basis in practical applications. Existing agricultural decision-making models have poor adaptability to the time-varying characteristics of the agricultural environment and cannot flexibly respond to the influence of factors such as seasonal changes and meteorological mutations.

[0005] Therefore, the present invention proposes an intelligent decision-making method and device for agricultural water use driven by big data to solve the above problems. Summary of the Invention

[0006] In view of the deficiencies of the prior art, the present invention develops an intelligent decision-making method and device for agricultural water use driven by big data. Through the constructed intelligent decision-making classification model for agricultural water use, the present invention can solve the problems of accuracy, robustness, and interpretability in agricultural water use management.

[0007] On the one hand, the technical solution for the present invention to solve the technical problem is an intelligent decision-making method for agricultural water use driven by big data, including the following steps: S1. Data collection: Collect soil data through soil moisture sensors, collect meteorological monitoring data through meteorological detection devices, record crop growth cycle data, form a dataset for training from the collected data, and manually label irrigation decisions for the corresponding fields according to the collected data. S2. Data preprocessing: Normalize the data collected in the dataset through a piecewise normalization strategy based on dynamic spatio-temporal windows and spatial classes. S3. Train the constructed intelligent decision-making classification model for agricultural water use: Construct an intelligent decision-making classification model for agricultural water use, by constructing a cost-sensitive feature cross-mapping network, calculating the loss function of the extreme learning machine, optimizing the parameters of the extreme learning machine based on evapotranspiration constraints, and dynamically adjusting the threshold. Input the preprocessed data into the intelligent decision-making classification model for agricultural water use for training until the preset stop iteration condition is met, that is, complete the training of the intelligent decision-making classification model for agricultural water use. S4. Intelligent decision-making for agricultural water use: Input the newly collected data after preprocessing into the trained intelligent decision-making classification model for agricultural water use for processing, and finally output the decision result.

[0008] S1 is specifically as follows: The collected soil data includes, but is not limited to, volumetric water content, electrical conductivity, and temperature. The meteorological monitoring data includes, but is not limited to, precipitation, evaporation, air temperature, and wind speed. The crop growth cycle data includes monitoring data at different growth stages, and the growth stages include the germination stage, the heading stage, and the maturity stage. During the agricultural production process, several soil moisture sensors and meteorological detection devices are set in the fields. The soil moisture sensors and meteorological detection devices collect data on different fields and record the soil and meteorological data in real time. The time granularity of the collection is set to hours or days according to specific requirements. The collected data is transmitted to the central database for storage in real time through a wireless network. The data stored in the central database forms a dataset for training. Then, read the data in the central database, and according to the data collected for each field, experts manually label irrigation decisions for each field. The classification labels of the irrigation decisions include water shortage, suitability, and supersaturation.

[0009] S2 is as follows: The piecewise normalization strategy specifically divides the crop growth stages, retains the water requirement characteristics of each stage, and introduces spatial correlation correction using a spatial correction factor. The calculation formula of the piecewise normalization strategy is as follows: , where represents the normalized value of the data of the th type collected from the th field plot at the th growth stage of the crop, represents the original observed value of the data of the th type collected from the th field plot at the th growth stage of the crop, represents the spatial correction factor of the th field plot at the th growth stage of the crop, represents the division of the th growth period through a dynamic time window, represents the data of the th type collected from the th time point within the th field plot at the th growth stage of the crop.

[0010] S3 is as follows: Input the preprocessed data into the intelligent decision-making classification model for agricultural water use to train the model. The model includes an extreme learning machine model. The input data first passes through the extreme learning machine model. Through the calculation of the hidden layer of the extreme learning machine, the preprocessed data is mapped to the decision space to generate a preliminary decision; Then, improve the accuracy of the decision through the following operations: Construct a cost-sensitive feature cross-mapping network and dynamically adjust the weights of the parameters through the cost-sensitive learning mechanism; Adopt a double-weight adjustment mechanism based on the cost-sensitive matrix and sample density to calculate the loss function of the extreme learning machine; Optimize the model parameters of the extreme learning machine based on the evapotranspiration constraint; Use a dynamic threshold adjustment mechanism to automatically adjust the decision threshold according to the historical data fluctuations.

[0011] The cost-sensitive learning mechanism is as follows: The activation values of soil data and meteorological data at different hidden layer nodes are calculated using a dual-channel cross-mapping mechanism. Specifically, through cross-term design, the intelligent decision-making classification model for agricultural water use can explicitly learn the synergistic effect of soil-meteorological data, and the contribution degree of different data combinations is adjusted through a gating mechanism using dynamic coupling coefficients; The calculation formula for the activation value is as follows: , , where, represents the set of preprocessed soil data, represents the set of preprocessed meteorological data, represents the specific parameter index in is the specific parameter index in represents the activation value of the th hidden layer node in the soil data channel, represents the activation value of the th hidden layer node in the meteorological data channel, represents the rectified linear unit activation function ReLU, represents the soil data to the th hidden layer node connection weight, represents the soil data to the th hidden layer node connection weight, represents the input value of the preprocessed soil data , represents the meteorological data dynamic coupling weight in the soil channel, represents the meteorological data dynamic coupling weight in the meteorological channel, represents the input value of the preprocessed meteorological data , represents the Hadamard product used to capture the local interaction of soil-meteorological data, represents the Kronecker product used to capture the global coupling effect, represents the dynamic coupling coefficient for learning the non-linear synergistic effect of soil-meteorological data.

[0012] The loss function of the extreme learning machine is specifically as follows: The loss function of the extreme learning machine is calculated using a dual-weight adjustment mechanism based on the cost-sensitive matrix and sample density. Specifically, the sample density weight is used to alleviate the imbalance in the number of categories, and the cost difference defined by business knowledge is introduced using the cost matrix; The calculation formula of the loss function of the extreme learning machine is as follows: , where represents the loss function of the extreme learning machine, represents the true class label of the th sample, represents 's density weight, , represents the number of samples with label , represents the total number of classes of irrigation decisions, is the total number of training samples, represents the preset cost matrix, represents misjudging the true class label of the th sample as the th class's penalty coefficient, represents the transpose of the output vector of the th sample in the hidden layer of the extreme learning machine, represents the weight vector corresponding to the th class in the output layer of the extreme learning machine, represents the weight vector corresponding to the th class in the output layer, represents the focusing factor, is set to 2.

[0013] The process of optimizing the model parameters of the extreme learning machine based on the evapotranspiration constraint is as follows: Incorporate the FAO Penman-Monteith equation for calculating potential evapotranspiration in high-latitude regions as a soft constraint into the optimization process, and ensure that the model prediction conforms to the basic law of water transfer through the physical constraint term; The calculation formula is as follows: , , where represents the evapotranspiration of the th sample predicted by the model, represents the theoretical evapotranspiration of the th sample calculated by the FAO Penman-Monteith formula, represents the loss function of the extreme learning machine when the parameter is , represents the parameters of the extreme learning machine, represents the weight coefficient of the physical constraint term, is set to 0.3, represents the total number of training samples, represents the total number of time steps within the crop growth cycle, represents the vapor pressure deficit influence coefficient, which is obtained by optimizing the fitting of historical evapotranspiration data through the gradient descent method after random initialization, represents the net radiation influence coefficient, which is obtained by optimizing the fitting of historical evapotranspiration data through the gradient descent method after random initialization, represents the th sample's hidden layer output vector at the th time period; represents the th sample's vapor pressure deficit at the th time period; represents the th sample's net radiation amount at the th time period; Then, based on the loss function of the extreme learning machine, an improved Nesterov accelerated gradient method is used to update the parameters of the extreme learning machine.

[0014] The dynamic threshold adjustment is as follows: The sliding time window threshold calibration method is used to calibrate the output of the extreme learning machine, and the decision function is defined as: , where represents the irrigation decision prediction result at the th time period, 1 indicates irrigation is required, corresponding to water shortage, and 0 indicates no irrigation is required, corresponding to suitability and supersaturation; represents the prediction confidence of the extreme learning machine model at the th time period, represents the multi-source input feature vector at the th time period, represents the dynamically adjusted decision threshold at the th time period.

[0015] S4 is as follows: Adjust the irrigation amount within each time period of the field block according to the decision result, and the decision results include water shortage, suitability, and supersaturation; If the output is water shortage, increase the irrigation amount, If the output is suitability, keep it unchanged; If the output is supersaturation, drain the water in time.

[0016] On the other hand, the present invention also provides a computer storage medium, on which instructions are stored. When the instructions run on a computer, the computer is enabled to execute the intelligent decision-making method for agricultural water use based on big data drive.

[0017] The effects provided in the invention content are only the effects of the embodiments, rather than all the effects of the invention. The above technical solutions have the following advantages or beneficial effects: The present invention adopts a segmented normalization method based on dynamic time windows and spatial clustering. Through the segmented normalization of dynamic time windows and spatial clustering, the present invention can dynamically divide time periods according to the crop growth cycle, retain the water demand characteristics of different growth periods, and use a spatial correction factor to solve the local deviation caused by uneven sensor density, so that the water demand law of each stage is accurately retained, avoiding the blurring of key features by global normalization, and improving the accuracy of farmland water management; The present invention adopts a dual-channel cross-mapping mechanism, explicitly learns the synergistic effect of soil and meteorological parameters through cross terms, and uses a cost-sensitive method to adjust the model weights, enhancing the sensitivity to important decision categories and improving the accuracy and convergence speed of the model; Adopt a dual-weight adjustment mechanism based on the cost-sensitive matrix and sample density to adjust the weights of different categories. Especially when dealing with categories with large cost differences such as "water shortage" and "oversaturation", it improves the problem of class imbalance and enhances the recall rate of water shortage categories; The present invention integrates physical constraints into the model training process to ensure that the predictions of the model conform to the basic laws of water transfer, avoiding prediction results that may violate agronomic common sense by pure data-driven models, and improving the interpretability and physical rationality of the model; The present invention adopts a dynamic threshold adjustment mechanism. In view of the time-varying characteristics of agricultural water use decisions, it uses a sliding time window to adjust the decision threshold, automatically senses environmental changes, and adjusts the conservativeness of decisions according to historical data fluctuations, enhancing the model's adaptability to meteorological changes and seasonal fluctuations, and reducing the decision oscillation problem caused by meteorological mutations; In summary, it systematically solves the problems of accuracy, robustness, and interpretability in agricultural water management, and provides a set of innovative and practical technical solutions for precise agricultural water use. Brief Description of the Drawings

[0018] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention.

[0019] Figure 1 It is a schematic flowchart of the method of the present invention.

[0020] Figure 2 It is a feature discrimination comparison diagram between dynamic segmented normalization and global normalization.

[0021] Figure 3 It is a line chart of the classification accuracy of the method of the present invention and the existing method under different sample sizes.

[0022] Figure 4 This is a performance comparison chart of the confusion matrix of the present invention and the traditional confusion matrix.

[0023] Figure 5 This is a line chart of the evapotranspiration constraint effect in the present invention.

[0024] Figure 6 This is a schematic diagram of the evapotranspiration prediction of the unconstrained model.

[0025] Figure 7 This is a schematic diagram of the prediction with physical constraints in the present invention. Detailed implementation manners

[0026] In order to clearly illustrate the technical features of the present solution, the present invention will be elaborated in detail below through specific implementation manners and in combination with its accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and settings of specific examples are described below.

[0027] Embodiment 1 An intelligent decision-making method for agricultural water use based on big data driving includes the following steps: S1. Data collection: Collect soil data through soil moisture sensors, collect meteorological monitoring data through meteorological detection equipment, record crop growth cycle data, form a dataset for training from the collected data, and manually label irrigation decisions for the corresponding fields according to the collected data; S2. Data preprocessing: Normalize the data collected in the dataset through a segmentation normalization strategy based on a dynamic space-time window and space classes; S3. Train the constructed intelligent decision-making classification model for agricultural water use: Construct an intelligent decision-making classification model for agricultural water use, and optimize the parameters of the extreme learning machine and dynamically adjust the threshold by constructing a cost-sensitive feature cross-mapping network, calculating the loss function of the extreme learning machine, and performing evapotranspiration constraint; Input the preprocessed data into the intelligent decision-making classification model for agricultural water use for training until the preset stop iteration condition is met, that is, complete the training of the intelligent decision-making classification model for agricultural water use; S4. Intelligent decision-making for agricultural water use: Input the newly collected data after preprocessing into the trained intelligent decision-making classification model for agricultural water use for processing, and finally output the decision result.

[0028] In the specific implementation manner, S1 is as follows: The collected soil data includes, but is not limited to, volumetric water content, conductivity, and temperature, which reflect the soil moisture state, conductivity, and temperature changes; Meteorological monitoring data includes, but is not limited to, precipitation, evaporation, temperature, and wind speed, which reflect the impact of the external environment on soil moisture; Crop growth cycle data includes monitoring data for different growth stages, and the growth stages include the germination stage, the heading stage, and the maturity stage. The monitoring data for different growth stages reflects the change in crop water demand with the growth stage; During the agricultural production process, several soil moisture sensors and meteorological detection devices are set up in the fields. The soil moisture sensors and meteorological detection devices collect data on different fields and record the soil and meteorological data in real time. The time granularity of the collection is set to hours or days according to specific requirements; The collected data is transmitted to the central database for storage in real time through a wireless network. The data stored in the central database constitutes a dataset for training. Then, the data in the central database is read, and according to the data collected for each field, experts manually label the irrigation decisions for each field. The classification labels of the irrigation decisions include water shortage, suitability, and supersaturation.

[0029] In the specific implementation manner, S2 is as follows: The segmented normalization strategy specifically divides the crop growth stages, retains the water demand characteristics of each stage, and introduces spatial correlation correction using a spatial correction factor. The calculation formula of the segmented normalization strategy is as follows: , Among them, represents the normalized value of the data of the th type collected in the th field at the th growth stage of the crop, represents the original observed value of the data of the th type collected in the th field at the th growth stage of the crop, represents the spatial correction factor of the th field at the th growth stage of the crop, represents the th growth period divided by a dynamic time window. By dynamically dividing time periods, the problem of different water demand laws in different growth periods is solved. For example, the water demand during the heading stage is significantly higher than that during the germination stage. The segmented normalization can retain the characteristics of each stage and avoid the blurring of key water demand characteristics by global normalization; represents the data of the th type collected at the th time point within the th field at the th growth stage of the crop; Furthermore, the calculation formula of the spatial correction factor is as follows: , Among them, represents the spatial correlation coefficient. By adjusting the spatial correlation coefficient to balance local features and global distribution, the heterogeneity of different farmland areas can be adapted. is set to 0.2. represents the -dimensional feature mean vector of the th field block. represents the feature mean vector of the spatial clustering center to which the th field block belongs.

[0030] In the specific implementation manner, S3 is as follows: Input the preprocessed data into the intelligent decision-making classification model for agricultural water use to train the model. The model includes an extreme learning machine model. The input data first passes through the extreme learning machine model. Through the calculation of the hidden layer of the extreme learning machine, the preprocessed data is mapped to the decision space to generate a preliminary decision. Then, the following operations are performed to improve the accuracy of the decision: Construct a cost-sensitive feature cross-mapping network, and dynamically adjust the weights of the parameters through the cost-sensitive learning mechanism. Adopt a double-weight adjustment mechanism based on the cost-sensitive matrix and sample density to calculate the loss function of the extreme learning machine. Optimize the model parameters of the extreme learning machine based on the evapotranspiration constraint. Use a dynamic threshold adjustment mechanism to automatically adjust the decision threshold according to the historical data fluctuation.

[0031] Furthermore, the cost-sensitive learning mechanism is as follows: Adopt a dual-channel cross-mapping mechanism to calculate the activation values of soil data and meteorological data at different hidden layer nodes. Specifically, through the cross-term design, the intelligent decision-making classification model for agricultural water use can explicitly learn the synergistic effect of soil-meteorological data, and use the dynamic coupling coefficient to adjust the contribution degree of different data combinations through the gating mechanism. The calculation formula of the activation value is as follows: , , Among them, represents the set of preprocessed soil data. represents the set of preprocessed meteorological data. represents the specific parameter index in is The specific parameter index in represents the activation value of the th hidden layer node of the soil data channel, represents the activation value of the th hidden layer node of the meteorological data channel, represents the rectified linear unit activation function ReLU, represents the soil data to the th connection weight of the hidden layer nodes, represents the soil data to the th connection weight of the hidden layer nodes, represents the input value of the preprocessed soil data , represents the meteorological data in the dynamic coupling weight in the soil channel, represents the meteorological data in the dynamic coupling weight in the meteorological channel, represents the input value of the preprocessed meteorological data , represents the Hadamard product used to capture the local interaction of soil - meteorological data, represents the Kronecker product used to capture the global coupling effect, represents the dynamic coupling coefficient for learning the non - linear synergistic effect of soil - meteorological data.

[0032] Calculated by the cost - sensitive method and , and the calculation formula is as follows: , , where, represents the first basic coupling coefficient, with a default value of 0.5, represents the second basic coupling coefficient, with a default value of 0.3, represents the misclassification cost of the true class , represents the cost of the predicted class , represents dynamically adjusting the weight according to the misclassification cost. By adjusting the feature weight according to the cost ratio, the sensitivity to high - cost classes can be enhanced; For example, if the weight of misjudging water shortage as supersaturation (cost 5) is higher than the reverse misjudgment (cost 1), the sensitivity to high - cost classes is enhanced; Regarding the dynamic coupling coefficient The calculation formula is as follows: , , Among them, represents the Sigmoid activation function, and respectively represent the original soil data and meteorological data without preprocessing, represents the parameter interaction function; represents the first embedding matrix, which is obtained by random initialization and updated during model training. The update method can be gradient descent; represents the second embedding matrix, which is obtained by random initialization and updated during model training. The update method can be gradient descent; represents the bias term, which is updated during model training. The update method can be gradient descent.

[0033] Furthermore, the loss function of the extreme learning machine is specifically as follows: A dual-weight adjustment mechanism based on the cost-sensitive matrix and sample density is used to calculate the loss function of the extreme learning machine. Specifically, the sample density weight is used to alleviate the imbalance of the number of categories, and the cost matrix is used to introduce the cost difference defined by business knowledge; The calculation formula of the loss function of the extreme learning machine is as follows: , Among them, represents the loss function of the extreme learning machine, represents the true class label of the th sample, represents 's density weight, , represents the number of samples with label , represents the total number of categories of irrigation decisions, is the total number of training samples, represents the transpose of the output vector of the th sample in the hidden layer of the extreme learning machine, represents the weight vector corresponding to the th category in the output layer of the extreme learning machine, Dynamic sparsification is achieved through the L1 regularization term to suppress redundant features and solve the problem of redundant features in high-dimensional agricultural data; represents the weight vector corresponding to the th category in the output layer, represents the focusing factor, is set to 2; denotes a preset cost matrix, denotes the misjudgment of the true class label of the -th sample as the -th class, and the preset cost matrix is predefined by domain experts according to the actual irrigation cost differences. For example, for the classes water shortage (0), suitable (1), and over-saturation (2), the constructed cost matrix is as follows: , where the rows of the cost matrix represent the true classes, the columns represent the predicted classes, the cost of the diagonal (correct classification) is 0, the cost of misjudging water shortage (0) as over-saturation (2) is 5 (the highest penalty), the cost of misjudging water shortage (0) as suitable (1) is 3, the cost of misjudging over-saturation (2) as water shortage (0) is 1 (the lowest penalty), and the cost of misjudging suitable (1) as over-saturation (2) is 2; The update rule of is as follows: where denotes the updated -th class weight vector, denotes the learning rate of the weight vector, is set to 0.01, denotes the L1 regularization strength coefficient, is set to 0.001, denotes the sign function, outputting the sign (+1 / -1) of the input parameter, is the L1 norm.

[0034] Furthermore, the process of optimizing the model parameters of the extreme learning machine based on the evapotranspiration constraint is specifically as follows: Incorporate the high-latitude region potential evapotranspiration calculation equation FAO Penman-Monteith as a soft constraint into the optimization process, and ensure that the model prediction conforms to the basic law of water transfer through the physical constraint term; The calculation formula is as follows: , , where denotes the evapotranspiration of the -th sample predicted by the model, denotes the theoretical evapotranspiration of the -th sample calculated by the FAO Penman-Monteith formula, denotes the loss function of the extreme learning machine when the parameter is , denotes the parameter of the extreme learning machine, The weight coefficient representing the physical constraint term is set to 0.3 represents the total number of training samples represents the total number of time steps during the crop growth cycle represents the vapor pressure deficit influence coefficient, which is obtained by optimizing the fitting of historical evapotranspiration data through the gradient descent method after random initialization represents the net radiation influence coefficient, which is obtained by optimizing the fitting of historical evapotranspiration data through the gradient descent method after random initialization represents the -th sample's hidden layer output vector at the -th time period represents the -th sample's vapor pressure deficit at the -th time period represents the -th sample's net radiation amount at the -th time period Then, based on the loss function of the extreme learning machine, the improved Nesterov accelerated gradient method is used to update the parameters of the extreme learning machine, and the calculation formula is as follows , , where represents the update amount of the Nesterov momentum at the -th iteration represents the update amount of the Nesterov momentum at the -th iteration represents the momentum decay coefficient is set to 0.9 represents the base learning rate of the extreme learning machine parameters is set to 0.005 represents the joint gradient of the loss function of the extreme learning machine and the physical constraint term with respect to the extreme learning machine parameters represents the mean square loss of the evapotranspiration prediction error ; represents the model parameters of the extreme learning machine at the -th iteration represents the model parameters of the extreme learning machine at the -th iteration represents the momentum correction factor ; represents the L2 norm

[0035] Furthermore, the dynamic threshold adjustment is specifically as follows The output of the extreme learning machine is calibrated using the sliding time window threshold calibration method, and the decision function is defined as: , where represents the irrigation decision prediction result for the -th time period, 1 indicates irrigation required, corresponding to water shortage, and 0 indicates no irrigation required, corresponding to suitability and supersaturation; represents the prediction confidence of the extreme learning machine model in the -th time period, represents the multi-source input feature vector for the -th time period, represents the dynamically adjusted decision threshold for the -th time period; The update rule of is as follows: where represents the input feature vector at the -th historical moment within the sliding time window; represents the threshold smoothing coefficient, is set to 0.7; represents the dynamically adjusted decision threshold for the -th time period; represents the set of historical moment indices included within the sliding time window for the -th time period; represents the sensitivity coefficient, is set to 0.1; represents the standard deviation of the model output within the time window for the -th time period, which automatically senses the severity of environmental changes. When the data fluctuates greatly, increases, that is, the threshold conservatism is enhanced, thereby solving the decision oscillation caused by meteorological mutations; The calculation formula of is as follows: , where represents the input feature vector at the -th historical moment within the sliding time window.

[0036] Furthermore, S4 is specifically as follows: Adjust the irrigation amount in each time period of the field according to the decision result. The decision results include water shortage, suitability, and supersaturation; If water shortage is output, increase the irrigation amount, If suitability is output, keep it unchanged; If supersaturation is output, drain the water in a timely manner.

[0037] Example 2 A computer storage medium stores instructions thereon. When the instructions run on a computer, the computer is caused to execute an intelligent decision-making method for agricultural water use driven by big data.

[0038] Example 3 As Figure 2 shown, it is a feature discrimination comparison chart of dynamic segment normalization and global normalization. As can be seen from Figure 2 this, by comparing the effects of dynamic segment normalization, the effectiveness of segment normalization for feature retention is verified. In terms of key features such as volumetric water content and conductivity, the method of the present invention improves the discrimination by about 15% - 20% compared with global normalization. The experimental results show that the dynamic time window division effectively retains the feature differences in each growth stage of crops.

[0039] Example 4 As Figure 3 shown, it is a line chart of the classification accuracy of the method of the present invention and the existing method under different sample sizes. As can be seen from Figure 3 this, by comparing the performance of the feature cross network, the effectiveness of the dual-channel cross mapping is verified. When the sample size > 2000, the accuracy of the present invention is improved by 8% - 12% compared with the traditional extreme learning machine, and the convergence speed is faster. The test results show that the feature cross mechanism significantly improves the model's ability to model the parameter coupling relationship.

[0040] Example 5 As Figure 4 shown, it is a performance comparison chart of the confusion matrix of the present invention and the traditional confusion matrix. As can be seen from Figure 4 this, by analyzing the cost-sensitive loss effect, the effectiveness of the misclassification cost adjustment mechanism is verified. In terms of the recall rate of the water shortage category, the present invention is improved by 9%, and the over-saturation misjudgment rate is reduced by 30%. The test results show that the cost-sensitive mechanism effectively balances the business requirements of different misjudgment costs.

[0041] Example 6 As Figure 5 shown, Figure 5 it is a line chart of the evapotranspiration constraint effect in the present invention. As Figure 5 shown, to analyze the optimization effect of the evapotranspiration constraint and verify the guiding role of physical constraints on model optimization, when the constraint is introduced, the evapotranspiration prediction error is reduced by about 42%, and the training loss converges more stably. The experimental results show that the physical constraints effectively improve the interpretability and prediction rationality of the model.

[0042] Furthermore, as Figure 6 and Figure 7 shown, Figure 6 it is a schematic diagram of the evapotranspiration prediction of the unconstrained modelFigure 7 This is a schematic diagram for predicting physical constraints in the present invention. It can be seen from Figure 6 and Figure 7 that by comparing contour lines, the improvement effect of physical constraints on the rationality of model prediction can be verified. In the key parameter range of steam pressure difference and net radiation, the unconstrained model shows a phenomenon that violates the law of water transfer, and there is a systematic deviation between its predicted value and the theoretical calculated value. However, in this technology, by considering the evapotranspiration strategy as an optimization constraint, the predicted surface is highly consistent with the theoretical contour line. Especially in the extreme parameter combination regions such as high temperature and low humidity, physical consistency is still maintained. This optimization strategy that combines data-driven and physical laws not only improves the prediction accuracy, but also ensures that the model output conforms to the basic principles of agricultural hydrology, making the irrigation decision-making both data-intelligent and scientifically interpretable.

[0043] Although the specific implementation manners of the invention are described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present invention. Based on the technical solution of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A big data-driven intelligent decision-making method for agricultural water use, characterized by: The following steps are involved: S1. Data collection: soil data is collected through soil moisture sensors, meteorological monitoring data is collected through meteorological detection equipment, and crop growth cycle data is recorded. The collected data constitutes a data set for training, and irrigation decisions are manually marked for the corresponding fields based on the collected data; S2, data preprocessing: normalize the data collected in the dataset through a segmented normalization strategy based on dynamic space-time windows and spatial classes; S3. Training the constructed agricultural water use intelligent decision-making classification model: constructing an agricultural water use intelligent decision-making classification model by constructing a cost-sensitive feature cross-mapping network, calculating the loss function of the extreme learning machine, optimizing the parameters of the extreme learning machine based on the evapotranspiration constraint, and dynamically adjusting the threshold; The preprocessed data is input into the agricultural water use intelligent decision-making classification model for training until the preset stop iteration condition is met, that is, the training of the agricultural water use intelligent decision-making classification model is completed; S4. Intelligent decision-making on agricultural water use: The newly collected data is pre-processed and input into the trained intelligent decision-making classification model for agricultural water use for processing, and the decision result is finally output.

2. The intelligent decision-making method for agricultural water use based on big data drive according to claim 1 is characterized in that: S1 is as follows: The soil data collected include, but are not limited to, volumetric moisture content, electrical conductivity, and temperature; Meteorological monitoring data include but are not limited to precipitation, evaporation, temperature and wind speed; Crop growth cycle data includes monitoring data at different growth stages, including germination, heading, and maturity; During the agricultural production process, several soil moisture sensors and meteorological detection equipment are set up in the fields. The soil moisture sensors and meteorological detection equipment collect data on different fields and record soil and meteorological data in real time. The time granularity of the collection is set to hours or days according to specific needs; The collected data is transmitted to the central database in real time via the wireless network for storage. The data stored in the central database constitutes the data set for training. The data in the central database is then read, and experts manually mark the irrigation decisions for each field based on the data collected from each field. The classification labels for irrigation decisions include water shortage, suitability, and oversaturation.

3. The intelligent decision-making method for agricultural water use based on big data drive according to claim 2 is characterized in that: S2 is as follows: The segmented normalization strategy specifically divides the crop growth stages, retains the water demand characteristics of each stage, and uses the spatial correction factor to introduce spatial correlation correction. The calculation formula of the segmented normalization strategy is as follows: ， in, Indicates The first The data collected in this category are The normalized value of the growth stage, Indicates The first The data collected in this category are The original observations collected during the growth phase, Indicates The plots are in the first The spatial correction factor for each growth stage, Indicates that the dynamic time window is used to divide the A growing period, Indicates that in the crop The first growth stage Time point The first Class collected data.

4. The intelligent decision-making method for agricultural water use based on big data drive according to claim 3 is characterized in that: S3 is as follows: The preprocessed data is input into the agricultural water intelligent decision-making classification model to train the model. The model includes an extreme learning machine model. The input data first passes through the extreme learning machine model. Through the calculation of the hidden layer of the extreme learning machine, the preprocessed data is mapped to the decision space to generate a preliminary decision. Improve the accuracy of decision making by: Construct a cost-sensitive feature cross-mapping network and dynamically adjust the parameter weights through a cost-sensitive learning mechanism; A dual weight adjustment mechanism based on cost-sensitive matrix and sample density is used to calculate the loss function of the extreme learning machine; Optimize the model parameters of the extreme learning machine based on evapotranspiration constraints; Use a dynamic threshold adjustment mechanism to automatically adjust the decision threshold based on historical data fluctuations.

5. The intelligent decision-making method for agricultural water use based on big data drive according to claim 4 is characterized in that the cost The sensitive learning mechanism is as follows: A dual-channel cross-mapping mechanism is used to calculate the activation values ​​of soil data and meteorological data at different hidden layer nodes. Specifically, the cross-term design enables the agricultural water use intelligent decision-making classification model to explicitly learn the synergistic effect of soil-meteorological data, and uses the dynamic coupling coefficient to adjust the contribution of different data combinations through the gating mechanism. The activation value is calculated as follows: ， ， in, represents the collection of preprocessed soil data, represents the set of preprocessed meteorological data, express A specific parameter index in for A specific parameter index in Indicates soil data channel The activation value of the hidden layer nodes, Indicates the meteorological data channel The activation value of the hidden layer nodes, ReLU represents the rectified linear unit activation function. Represents soil data To The connection weights of hidden layer nodes, Represents soil data To The connection weights of hidden layer nodes, Represents the preprocessed soil data The input value of Displays weather data Dynamic coupling weights in soil channels, Displays weather data Dynamic coupling weights in the meteorological channel, Represents the preprocessed meteorological data The input value of represents the Hadamard product used to capture the local interaction of soil-meteorological data, represents the Kronecker product used to capture the global coupling effect, Represents the dynamic coupling coefficient for learning the nonlinear synergistic effect of soil-meteorological data.

6. The intelligent decision-making method for agricultural water use based on big data drive according to claim 5 is characterized in that: The loss function of the extreme learning machine is as follows: A dual weight adjustment mechanism based on cost-sensitive matrix and sample density is used to calculate the loss function of the extreme learning machine. Specifically, the sample density weight is used to alleviate the imbalance in the number of categories, and the cost matrix is ​​used to introduce the cost difference defined by business knowledge. The calculation formula of the extreme learning machine loss function is as follows: ， in, represents the loss function of the extreme learning machine, Indicates The true category labels of samples, express The density weight of , Indicates that the label is The number of samples, represents the total number of categories for irrigation decisions, is the total number of training samples, represents the preset cost matrix, Indicates that the The true category label of the sample is misclassified as The penalty coefficient for each category, Indicates The output vector of the samples in the hidden layer of the extreme learning machine is transposed. Indicates that the output layer of the extreme learning machine corresponds to The weight vector of each class, The output layer corresponds to The weight vector of each class, represents the focusing factor, Set to 2.

7. The intelligent decision-making method for agricultural water use based on big data drive according to claim 6 is characterized in that: The model parameter optimization process of the extreme learning machine based on evapotranspiration constraint is as follows: The FAO Penman-Monteith equation for calculating potential evapotranspiration in high latitudes is incorporated into the optimization process as a soft constraint, and the physical constraint terms are used to ensure that the model prediction conforms to the basic laws of water transport. The calculation formula is as follows: ， ， in, The model predicts the The evapotranspiration of the samples, represents the first The theoretical evapotranspiration of a sample is Indicates that the parameter is The loss function of the extreme learning machine when represents the parameters of the extreme learning machine, represents the weight coefficient of the physical constraint term, Set to 0.3, represents the total number of training samples, represents the total number of time steps in the crop growth cycle, represents the vapor pressure difference influence coefficient, which is obtained by optimizing the historical evapotranspiration data fitting through the gradient descent method after random initialization. Represents the net radiation influence coefficient, which is obtained by optimizing the historical evapotranspiration data fitting through the gradient descent method after random initialization. Indicates The sample in The hidden layer output vector of time periods; Indicates The sample in Vapor pressure difference over time period; Indicates The sample in The net radiation amount in a time period; Then, based on the loss function of the extreme learning machine, the improved Nesterov accelerated gradient method is used to update the parameters of the extreme learning machine.

8. The intelligent decision-making method for agricultural water use based on big data drive according to claim 7 is characterized in that: The dynamic threshold adjustment is as follows: The sliding time window threshold calibration method is used to calibrate the output of the extreme learning machine, and the decision function is defined as: ， in, Indicates The irrigation decision prediction results for each time period, 1 means irrigation is needed, corresponding to water shortage, and 0 means no irrigation is needed, corresponding to suitable and oversaturated; Indicates that the extreme learning machine model is The forecast confidence for a time period, Indicates The multi-source input feature vector for each time period, Indicates Dynamically adjusted decision threshold for each time period.

9. The intelligent decision-making method for agricultural water use based on big data drive according to claim 8 is characterized in that: S4 is as follows: Adjust the irrigation amount of the field in each time period according to the decision results, which include water shortage, suitable and oversaturation; If the output is short of water, increase the irrigation amount. If the output is suitable, it remains unchanged; If the output is oversaturated, drain the water in time.

10. A computer storage medium, characterized in that: Instructions are stored on a computer storage medium. When the instructions are executed on a computer, the computer executes the big data driven intelligent decision-making method for agricultural water use as described in any one of claims 1 to 9.

Images