Environment data-oriented remote control method for nitrogen and phosphorus loss amount of paddy field

Abstract

The present application relates to the field of agricultural environmental information technology, and discloses a remote control method for nitrogen and phosphorus loss amount of paddy field facing environmental data. The method comprises the following steps: obtaining historical nitrogen and phosphorus loss data, real-time meteorological data and water and fertilizer operation information of the paddy field; constructing a mechanism constraint sub-model combining nitrogen migration and transformation dynamics and phosphorus adsorption and desorption balance; performing multi-source feature extraction on the meteorological data to generate a standardized environmental vector; inputting the light residual convolutional neural network to output an initial prediction value; combining the mechanism model to generate a corrected prediction value conforming to the conservation of mass through the Lagrange multiplier method; when the prediction value exceeds the threshold value, generating a drainage gate opening degree and a fertilizer plan adjustment instruction, and delivering the instruction to an edge terminal for execution through a narrowband Internet of Things.

Description

Remote control method for nitrogen and phosphorus loss from paddy fields based on environmental data Technical Field

[0001] This invention belongs to the field of agricultural environmental information technology, specifically relating to a method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data. Background Technology

[0002] With the increasing urgency of agricultural non-point source pollution control, accurate prediction and dynamic regulation of nitrogen and phosphorus loss from paddy fields have become key technologies for ensuring water environment safety and achieving green agricultural development. Traditional research often relies on hydro-biogeochemical mechanism models. These models, by coupling physicochemical processes such as crop growth, water transport, and nutrient cycling, can relatively realistically reflect the intrinsic mechanisms of nitrogen and phosphorus migration and transformation in farmland ecosystems. However, these models typically contain a large number of empirical parameters, have high computational complexity, and require stringent spatiotemporal resolution of input data, making it difficult to meet the timeliness requirements of remote, real-time regulation in practical applications.

[0003] Data-driven intelligent prediction methods for environmental applications have received widespread attention in recent years. Deep learning models, such as LSTM and Transformer, with their powerful nonlinear fitting capabilities, can learn the statistical patterns of nitrogen and phosphorus loss directly from multi-source environmental data, including historical meteorological data, soil moisture, and rainfall intensity, and achieve rapid prediction without explicitly modeling the physical processes. However, these purely data-driven methods have significant drawbacks: on the one hand, they heavily rely on large-scale, high-quality labeled data, which is costly and results in sparse samples in agricultural scenarios; on the other hand, the internal decision-making logic of the models is opaque, lacking interpretability to support the physical processes, making them difficult for agronomic experts to trust and use for critical regulatory decisions.

[0004] Existing technologies exhibit a clear disconnect between mechanistic modeling and data-driven approaches: while mechanistic models offer good interpretability and process fidelity, they are limited by computational efficiency and parameter sensitivity, making them unable to support high-frequency, short-cycle dynamic regulation; whereas data-driven models, while responsive, suffer from poor generalization ability under small sample sizes or abnormal conditions due to their "black box" nature and data dependence, easily leading to inaccurate regulation strategies. This contradiction is particularly pronounced in scenarios requiring forward-looking predictions of nitrogen and phosphorus loss risks within the next 12 to 48 hours to trigger remote irrigation, drainage, or fertilization interventions. Therefore, a novel integrated modeling and regulation method that maintains fidelity to the physical process while possessing real-time predictive capabilities is urgently needed. Summary of the Invention

[0005] This invention provides a remote control method for nitrogen and phosphorus loss in paddy fields based on environmental data. It aims to address the technical challenges of existing technologies where purely data-driven AI models suffer from poor interpretability and reliance on large amounts of labeled data, while purely mechanistic models suffer from complex parameters, low computational efficiency, and difficulty in supporting real-time remote control. This invention constructs a hybrid modeling architecture that integrates physical mechanism constraints with data-driven learning capabilities, achieving high-precision, interpretable, and low-latency dynamic prediction of nitrogen and phosphorus loss processes in paddy fields. Based on this prediction, executable remote control commands are generated to optimize farmland water and fertilizer management strategies and reduce the risk of non-point source pollution.

[0006] This invention provides a method for remotely controlling nitrogen and phosphorus loss in paddy fields based on environmental data, which includes: acquiring historical nitrogen and phosphorus loss monitoring data, real-time meteorological and environmental sensing data, and current water and fertilizer management operation information of a target paddy field area; and constructing a mechanism-constrained sub-model that includes a nitrogen migration and transformation kinetic equation and a phosphorus adsorption and desorption balance relationship based on the historical nitrogen and phosphorus loss monitoring data and the current water and fertilizer management operation information.

[0007] Multi-source heterogeneous feature extraction is performed on the real-time meteorological and environmental sensing data to generate a standardized environmental state vector. The standardized environmental state vector is input into a pre-trained lightweight residual convolutional neural network to output an initial nitrogen and phosphorus loss trend prediction value. The initial nitrogen and phosphorus loss trend prediction value is coupled and verified with the output of the mechanism constraint sub-model. A loss function with physical conservation constraints is constructed using the Lagrange multiplier method, and the weight parameters of the lightweight residual convolutional neural network are fine-tuned online to generate a corrected nitrogen and phosphorus loss prediction value that conforms to the law of conservation of matter. The corrected nitrogen and phosphorus loss prediction value is compared with a preset environmental safety threshold. If the corrected nitrogen and phosphorus loss prediction value exceeds the environmental safety threshold, a drainage gate opening adjustment command and a fertilization plan adjustment command are generated. The commands are then sent to an edge control terminal deployed in the target paddy field area via a narrowband IoT communication module to execute remote control operations.

[0008] The acquisition of historical nitrogen and phosphorus loss monitoring data, real-time meteorological environment sensing data, and current water and fertilizer management operation information for the target paddy field area as described in this invention specifically includes: acquiring time series data of total nitrogen concentration, total phosphorus concentration, nitrate nitrogen concentration, ammonium nitrogen concentration, and dissolved phosphorus concentration over a historical period through online water quality monitoring stations deployed at paddy field drainage outlets; acquiring real-time data of rainfall, rainfall intensity, temperature, relative humidity, wind speed, and solar radiation intensity through micro-weather stations deployed in the field; and acquiring operation log data of current irrigation water volume, fertilizer type, fertilizer amount, fertilization time, and tillage method through a farmland operation record system.

[0009] The mechanistic constraint sub-model constructed in this invention, which includes nitrogen migration and transformation kinetic equations and phosphorus adsorption-desorption equilibrium relationships, specifically includes: establishing a soil moisture movement equation based on Darcy's law, and coupling a convection-dispersion equation to describe the migration process of nitrogen in the soil profile; introducing first-order kinetic reaction terms to characterize the rates of organic nitrogen mineralization, nitrification, and denitrification; using the Langmuir isotherm adsorption equation to describe the adsorption capacity and desorption potential of soil particles for phosphate ions; discretizing the above equation set into a finite difference scheme, and setting initial and boundary conditions to form a numerically solvable mechanistic constraint sub-model.

[0010] The present invention describes the extraction of multi-source heterogeneous features from the real-time meteorological environment sensing data to generate a standardized environmental state vector, specifically including: combining rainfall amount and rainfall intensity into a cumulative rainfall feature; fusing temperature, relative humidity, and solar radiation intensity into an evapotranspiration potential index; jointly calculating wind speed and rainfall intensity into a surface runoff scour intensity factor; and normalizing the above features to map their numerical range to the interval between 0 and 1, thereby forming a standardized environmental state vector with a dimension of 7.

[0011] The pre-trained lightweight residual convolutional neural network described in this invention includes three convolutional layers, two residual blocks, and a fully connected output layer. Each convolutional layer has a kernel size of 3 and the number of channels is 16, 32, and 64 respectively. Each residual block contains two convolutional layers with a kernel size of 3 and introduces skip connections. The fully connected output layer outputs two floating-point values, corresponding to the predicted values ​​of nitrogen loss and phosphorus loss, respectively. The network is trained end-to-end using historical monitoring datasets in the offline stage, and the loss function is a weighted sum of mean square error and relative error.

[0012] The present invention describes coupling and verifying the initial nitrogen and phosphorus loss trend prediction with the output of the mechanism-constrained sub-model, constructing a loss function with physical conservation constraints using the Lagrange multiplier method, and fine-tuning the weight parameters of the lightweight residual convolutional neural network online. Specifically, this includes: defining a nitrogen mass conservation constraint as the input nitrogen amount minus the output nitrogen amount per unit time equals the change in the soil nitrogen pool; defining a phosphorus adsorption equilibrium constraint as the phosphorus concentration in the solution phase and the adsorption amount in the solid phase satisfying the Langmuir equation; converting the above constraints into a quadratic penalty term, which, together with the original prediction error term, constitutes a composite loss function; and using the stochastic gradient descent algorithm, performing a single-step online update of the network weights after each new environmental data reception, with a fixed update step size of 0.001.

[0013] The preset environmental safety thresholds of this invention include a nitrogen loss threshold of 5 kg per hectare per day and a phosphorus loss threshold of 0.5 kg per hectare per day. When the corrected predicted nitrogen and phosphorus loss exceeds either threshold, a control mechanism is triggered. The drainage gate opening adjustment command is generated in stages according to the degree of exceedance: the gate is closed if the exceedance is less than 10%, opened to 1 / 3 of its opening if the exceedance is 10% to 30%, and fully opened if the exceedance is more than 30%. The fertilization plan adjustment command includes postponing the next fertilization time by no less than 48 hours and reducing the original fertilization amount by 20% to 50%, with the reduction ratio being linearly positively correlated with the degree of exceedance.

[0014] In this invention, the narrowband IoT communication module adopts the 800 MHz frequency band approved by the State Radio Regulatory Commission, the communication protocol follows the third version of the narrowband IoT standard, the data packets are verified for integrity using a 16-bit cyclic redundancy check code, and the instruction transmission delay does not exceed 5 seconds, ensuring the real-time and reliable issuance of control instructions.

[0015] This invention provides a remote control system for nitrogen and phosphorus loss in paddy fields based on environmental data. The system includes: a multi-source data acquisition unit for acquiring historical nitrogen and phosphorus loss monitoring data, real-time meteorological and environmental sensing data, and current water and fertilizer management operation information for a target paddy field area; a mechanism constraint modeling unit for constructing a mechanism constraint sub-model containing a nitrogen migration and transformation kinetic equation and a phosphorus adsorption-desorption balance relationship based on the historical nitrogen and phosphorus loss monitoring data and the current water and fertilizer management operation information; and an environmental feature encoding unit for extracting multi-source heterogeneous features from the real-time meteorological and environmental sensing data to generate a standardized environmental state vector.

[0016] The data-driven prediction unit is used to input the standardized environmental state vector into a pre-trained lightweight residual convolutional neural network and output an initial nitrogen and phosphorus loss trend prediction value. The physical constraint fusion unit is used to couple and verify the initial nitrogen and phosphorus loss trend prediction value with the output of the mechanism constraint sub-model, construct a loss function with physical conservation constraints through the Lagrange multiplier method, fine-tune the weight parameters of the lightweight residual convolutional neural network online, and generate a corrected nitrogen and phosphorus loss prediction value that conforms to the law of conservation of matter.

[0017] The remote control decision unit is used to compare the corrected nitrogen and phosphorus loss prediction value with the preset environmental safety threshold. If the corrected nitrogen and phosphorus loss prediction value exceeds the environmental safety threshold, it generates a drainage gate opening adjustment command and a fertilization plan adjustment command. The command issuance and execution unit is used to issue the command to the edge control terminal deployed in the target paddy field area through the narrowband Internet of Things communication module to execute the remote control operation.

[0018] The multi-source data acquisition unit of this invention includes an online water quality monitoring station, a micro weather station, and a farmland operation recording system; the online water quality monitoring station is located at the outlet of the main drainage ditch of the paddy field, and the sampling frequency is once per hour; the micro weather station is installed in the central area of ​​the paddy field, 1.5 meters above the ground, and the data update cycle is 5 minutes; the farmland operation recording system is synchronized with the farmer's mobile terminal, and the operation log is uploaded to the cloud database in real time.

[0019] The edge control terminal of the present invention includes an electric gate actuator and an intelligent fertilizer controller; after receiving the opening adjustment command, the electric gate actuator drives the gate to rise and fall through a stepper motor, with a position feedback accuracy of 1%; the intelligent fertilizer controller has a built-in fertilizer ratio algorithm, which can automatically modify the running time and flow rate of the liquid fertilizer injection pump according to the adjustment command, with a control error of less than 3%.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] This invention deeply integrates the conservation laws of the physical mechanism model with the generalization ability of the data-driven model, constructing a hybrid prediction architecture that combines high accuracy, strong interpretability, and low computational complexity. The mechanism-constrained sub-model ensures the physical rationality of the prediction results in the process of material migration and transformation, avoiding the non-physical interpretations that may occur in a purely data-driven model; the lightweight residual convolutional neural network significantly reduces the number of model parameters and inference latency while retaining sufficient expressive power, meeting the timeliness requirements of remote real-time control.

[0022] The online fine-tuning mechanism embeds physical constraints into the loss function using the Lagrange multiplier method, enabling adaptive correction of the model during operation and maintaining long-term predictive stability without relying on a large amount of labeled data. Furthermore, the control commands generated by this invention are directly linked to executable agricultural operations, such as adjusting gate opening and fertilizer application, forming a closed-loop control chain from environmental perception and intelligent prediction to precise execution. This effectively reduces the risk of nitrogen and phosphorus loss from paddy fields and improves the intelligence level and engineering practicality of agricultural non-point source pollution control. Attached Figure Description

[0023] Figure 1 is a schematic diagram of the overall technical solution architecture of the present invention;

[0024] Figure 2 is a schematic diagram of the core principle framework of hybrid modeling that integrates physical mechanism constraints and data-driven learning capabilities in this invention.

[0025] Figure 3 is a logical flowchart of the multi-source environmental data acquisition and standardized feature vector construction in this invention;

[0026] Figure 4 is a flowchart of the logical process of coupling verification and online fine-tuning of the lightweight residual convolutional neural network and the mechanism constraint sub-model in this invention.

[0027] Figure 5 is a decision logic framework diagram of generating remote control instructions based on the corrected nitrogen and phosphorus loss prediction value in this invention.

[0028] Figure 6 is a schematic diagram of the multi-level interaction relationship and data flow between the cloud-based control system and the paddy field edge control terminal in this invention. Detailed Implementation

[0029] This invention provides a remote control method for nitrogen and phosphorus loss in paddy fields based on environmental data. Its core lies in constructing a hybrid modeling architecture that integrates physical mechanism constraints with data-driven learning capabilities. This architecture enables high-precision, interpretable, and low-latency dynamic prediction of nitrogen and phosphorus loss processes in paddy fields, and generates executable remote control commands based on the prediction results. The specific implementation methods of each component of the system will be described in detail below with reference to Figures 1 to 5.

[0030] The method strictly follows the sequence of the following steps:

[0031] S1, acquire historical nitrogen and phosphorus loss monitoring data, real-time meteorological and environmental perception data, and current water and fertilizer management operation information of the target paddy field area;

[0032] S2, Based on the historical nitrogen and phosphorus loss monitoring data and the current water and fertilizer management operation information, construct a mechanism-constrained sub-model that includes the nitrogen migration and transformation kinetic equation and the phosphorus adsorption and desorption balance relationship;

[0033] S3, perform multi-source heterogeneous feature extraction on the real-time meteorological environment sensing data to generate a standardized environmental state vector;

[0034] S4, input the standardized environmental state vector into a pre-trained lightweight residual convolutional neural network, and output the initial nitrogen and phosphorus loss trend prediction value;

[0035] S5, the initial nitrogen and phosphorus loss trend prediction value is coupled and verified with the output of the mechanism constraint sub-model. A loss function with physical conservation constraints is constructed by using the Lagrange multiplier method. The weight parameters of the lightweight residual convolutional neural network are fine-tuned online to generate a corrected nitrogen and phosphorus loss prediction value that conforms to the law of conservation of matter.

[0036] S6. Based on the comparison between the corrected nitrogen and phosphorus loss prediction value and the preset environmental safety threshold, if the corrected nitrogen and phosphorus loss prediction value exceeds the environmental safety threshold, a drainage gate opening adjustment command and a fertilization plan adjustment command are generated, and the commands are sent to the edge control terminal deployed in the target paddy field area through the narrowband Internet of Things communication module to perform remote control operations.

[0037] In step S1, the specific implementation method for acquiring historical nitrogen and phosphorus loss monitoring data, real-time meteorological environment sensing data, and current water and fertilizer management operation information of the target paddy field area includes: continuously collecting time series data of total nitrogen concentration, total phosphorus concentration, nitrate nitrogen concentration, ammonium nitrogen concentration, and dissolved phosphorus concentration over a historical period by using an online water quality monitoring station deployed at the outlet of the main drainage ditch of the paddy field, with a sampling frequency of once per hour; the online water quality monitoring station is equipped with a multi-parameter water quality sensor, whose measurement range covers total nitrogen from 0 to 100 mg / L and total phosphorus from 0 to 10 mg / L, with a measurement accuracy better than 5%;

[0038] Meanwhile, a miniature weather station deployed in the center of the paddy field at a height of 1.5 meters above the ground acquires real-time data on rainfall, rainfall intensity, air temperature, relative humidity, wind speed, and solar radiation intensity with a 5-minute update cycle. The miniature weather station integrates a tipping bucket rain gauge, an ultrasonic anemometer, a thermocouple temperature sensor, a capacitive humidity sensor, and a silicon photovoltaic radiometer, and all sensors have passed national metrological certification.

[0039] In addition, the farmland operation record system synchronously obtains operation log data such as current irrigation water volume, fertilizer type, fertilizer amount, fertilization time and tillage method from farmers' mobile terminals. The record system adopts a structured log format to ensure that each operation includes four types of fields: timestamp, operation type, operation amount and operation location, and uploads them to the cloud database in real time to form a complete water and fertilizer management operation information chain.

[0040] In step S2, the specific implementation method for constructing a mechanistic constraint sub-model that includes the nitrogen migration and transformation kinetic equation and the phosphorus adsorption-desorption equilibrium relationship includes: firstly, establishing a soil water movement equation based on Darcy's law, expressed as: ;

[0041] in This refers to the volumetric water content. For time, Unsaturated hydraulic conductivity As the matrix potential, The gravitational potential is used; based on this, the coupled convection-dispersion equation describes the migration process of nitrogen in the soil profile, and its form is: ;

[0042] in Nitrogen concentration, The dispersion coefficient is... The pore water flow velocity, The nitrogen reaction term is introduced; a first-order kinetic reaction term is introduced to characterize the rates of organic nitrogen mineralization, nitration, and denitrification, i.e. ,in , , These are the rate constants for mineralization, nitrification, and denitrification, respectively. , , These represent the concentrations of organic nitrogen, ammonium nitrogen, and nitrate nitrogen, respectively. For phosphorus migration, the Langmuir isotherm adsorption equation is used to describe the adsorption capacity and desorption potential of soil particles for phosphate ions, and its expression is: ;

[0043] in This represents the amount of phosphorus adsorbed per unit mass of soil. For maximum adsorption capacity, It is the Langmuir constant. Let be the phosphorus concentration in the solution phase. The above partial differential equations are discretized spatially into a finite difference scheme and advanced temporally using the implicit Euler method. The initial condition is the measured nitrogen and phosphorus distribution in the soil profile, and the boundary conditions are that the upper boundary flux is determined by rainfall and irrigation, and the lower boundary is free drainage. Finally, a mechanism constraint sub-model that can be solved numerically is formed. Before each execution of step S5, the model dynamically updates the initial conditions and upper boundary flux based on the fertilization amount and irrigation water volume in the current water and fertilizer management operation information to ensure that the mechanism output is consistent with the current field conditions.

[0044] In step S3, the specific implementation of multi-source heterogeneous feature extraction from the real-time meteorological environment sensing data to generate a standardized environmental state vector includes: combining rainfall amount and rainfall intensity into a cumulative rainfall feature, which is calculated as a weighted sum of rainfall amount and current rainfall intensity over the past 24 hours, with weighting coefficients of 0.7 and 0.3, respectively; and fusing air temperature, relative humidity, and solar radiation intensity into an evapotranspiration potential index, using a simplified form of the Penman-Monteith formula. ;

[0045] in The intensity of solar radiation. For temperature, Relative humidity, , , , The empirical coefficients are 0.0023, 0.0015, 17.8, and 50, respectively; the surface runoff erosion intensity factor is calculated by combining wind speed and rainfall intensity, and is defined as: ;

[0046] in For wind speed, The rainfall intensity is used as the basis for determination. Additionally, the original temperature, relative humidity, and solar radiation intensity are retained as independent features to preserve high-frequency fluctuation information. This results in an original feature vector comprising seven dimensions: cumulative rainfall characteristics, evapotranspiration potential index, surface runoff erosion intensity factor, temperature, relative humidity, solar radiation intensity, and wind speed. This vector is then normalized using a min-max scaling method, mapping each feature to the 0-1 interval, as shown in the formula: ;

[0047] in and These are the minimum and maximum values ​​of the feature in the historical three-year dataset, respectively; thus, a standardized environment state vector with a dimension of 7 is generated, which serves as the sole input to the subsequent data-driven model.

[0048] In step S4, the specific implementation of inputting the standardized environmental state vector into the pre-trained lightweight residual convolutional neural network and outputting the initial nitrogen and phosphorus loss trend prediction value includes: the lightweight residual convolutional neural network is trained end-to-end using historical monitoring datasets in the offline stage. Its network structure includes 3 convolutional layers, 2 residual blocks, and a fully connected output layer; the input layer receives a one-dimensional vector with a dimension of 7, which first passes through the first convolutional layer with a one-dimensional convolutional kernel size of 3 and 16 channels. The output feature map is batch normalized and ReLU activated before entering the first residual block.

[0049] Each residual block consists of two consecutive one-dimensional convolutional layers, each with a kernel size of 3 and the same number of channels. A skip connection is introduced at the end of the block to directly add the input to the output of the second convolutional layer. The first residual block has 32 channels, and the second residual block has 64 channels. The residual blocks are followed by a third convolutional layer with a kernel size of 3 and 64 channels, used for feature compression. Finally, the dimension is reduced to a 64-dimensional vector through a global average pooling layer, and then mapped to two-dimensional floating-point values ​​through a fully connected output layer, corresponding to the predicted values ​​of nitrogen loss and phosphorus loss, respectively.

[0050] During the offline training phase, the network uses a weighted sum of mean squared error and relative error as the loss function, with a weight ratio of 7:3. The optimizer is Adam, the learning rate is set to 0.001, and the training epochs are 200. After training, the model parameters are fixed and deployed on the cloud inference server. During runtime, whenever a new standardized environmental state vector is received, a forward propagation is performed to output the initial nitrogen and phosphorus loss trend prediction value. The entire inference process takes no more than 200 milliseconds.

[0051] In step S5, the initial predicted nitrogen and phosphorus loss trend value is coupled and verified with the output of the mechanism-constrained sub-model. A loss function with physical conservation constraints is constructed using the Lagrange multiplier method. The specific implementation method for online fine-tuning the weight parameters of the lightweight residual convolutional neural network includes: First, defining the nitrogen mass conservation constraint condition as the input nitrogen amount minus the output nitrogen amount equals the change in soil nitrogen pool per unit time, mathematically expressed as: ;

[0052] in Determined by the amount of fertilizer applied and the amount of atmospheric deposition. The amount of nitrogen loss is given by the model prediction. The calculation is performed by integral integration using a mechanism-constrained sub-model; secondly, the phosphorus adsorption equilibrium constraint condition is defined as the phosphorus concentration in the solution phase and the amount of phosphorus adsorbed in the solid phase satisfying the Langmuir equation, i.e. ,in To predict the adsorption amount, To predict the phosphorus concentration in the solution phase, the two constraints mentioned above are transformed into quadratic penalty terms, in the following forms: ;

[0053] in and The penalty coefficient is 10 and 100 respectively; the original prediction error term is the mean squared error. ,in These are the initial predicted values. The output of the mechanism-constrained sub-model under the same environmental conditions; composite loss function. Defined as Using the stochastic gradient descent algorithm, after each new environmental data is received and step S4 is completed, the network weights are updated step-by-step online. The update formula is as follows: ;

[0054] in The learning rate is fixed at 0.001. This online fine-tuning mechanism ensures that the model output always satisfies the physical laws of nitrogen mass conservation and phosphorus adsorption balance, avoiding non-physical solutions and maintaining the model's adaptability to new environmental conditions. The fine-tuning process only updates the parameters of the fully connected output layer and the last residual block, while freezing the remaining layers to control computational overhead.

[0055] In step S6, the corrected predicted nitrogen and phosphorus loss is compared with a preset environmental safety threshold. If the corrected predicted nitrogen and phosphorus loss exceeds the environmental safety threshold, a drainage gate opening adjustment command and a fertilization plan adjustment command are generated. These commands are then sent to edge control terminals deployed in the target paddy field area via a narrowband IoT communication module. Specific implementation methods include:

[0056] The preset environmental safety thresholds include a nitrogen loss threshold of 5 kg / ha / day and a phosphorus loss threshold of 0.5 kg / ha / day. The comparison logic is that control is triggered as soon as either nitrogen or phosphorus exceeds the limit. When the corrected predicted nitrogen loss exceeds 5 kg / ha / day but does not exceed 5.5 kg / ha, or the predicted phosphorus loss exceeds 0.5 kg / ha / day but does not exceed 0.55 kg / ha, it is determined to exceed the limit by less than 10%, and a drainage gate closure command is generated. When the predicted nitrogen loss is between 5.5 and 6.5 kg / ha, or the predicted phosphorus loss is between 0.55 and 0.65 kg / ha, it is determined to exceed the limit by 10% to 30%, and a command to open the drainage gate to 1 / 3 of its opening is generated.

[0057] When the predicted nitrogen loss exceeds 6.5 kg or the predicted phosphorus loss exceeds 0.65 kg, it is considered an over-limit of 30%, and a command to fully open the drainage gate is generated. The fertilization plan adjustment command uniformly includes postponing the next fertilization time by no less than 48 hours and reducing the original fertilization amount by 20% to 50%. The reduction ratio is calculated using the following formula: ;

[0058] in The ratio of the excess amplitude to the threshold is truncated to 0.5. After all instructions are generated, they are encapsulated into standard JSON format data packets and sent out through the narrowband IoT communication module. This module operates in the 800 MHz frequency band approved by the State Radio Regulatory Commission, and the communication protocol follows the third version of the narrowband IoT standard. A 16-bit cyclic redundancy check code is added to the data packet to ensure integrity. The end-to-end delay from instruction generation to receipt by the edge control terminal does not exceed 5 seconds. After the edge control terminal receives the instruction, the electric gate actuator drives the gate to rise and fall through the stepper motor, with a position feedback accuracy of 1%. The intelligent fertilizer controller automatically adjusts the running time and flow rate of the liquid fertilizer injection pump according to the reduction ratio, with a control error of less than 3%, thereby completing closed-loop control.

[0059] The remote control system for nitrogen and phosphorus loss in paddy fields, oriented towards environmental data, includes a multi-source data acquisition unit, a mechanism constraint modeling unit, an environmental feature coding unit, a data-driven prediction unit, a physical constraint fusion unit, a remote control decision-making unit, and an instruction issuance and execution unit. The multi-source data acquisition unit consists of an online water quality monitoring station, a micro-weather station, and a farmland operation recording system. The online water quality monitoring station is located at the outlet of the main drainage ditch in the paddy field, with a sampling frequency of once per hour. The micro-weather station is installed in the central area of ​​the paddy field, 1.5 meters above the ground, with a data update cycle of 5 minutes. The farmland operation recording system is synchronized with farmers' mobile terminals, and operation logs are uploaded to a cloud database in real time.

[0060] The mechanistic constraint modeling unit receives historical nitrogen and phosphorus loss monitoring data and current water and fertilizer management operation information, and constructs a mechanistic constraint sub-model that includes nitrogen migration and transformation kinetic equations and phosphorus adsorption-desorption equilibrium relationships. This model employs finite difference discretization and supports dynamic boundary condition updates. The environmental feature encoding unit performs multi-source heterogeneous feature extraction on real-time meteorological and environmental sensing data, generating a standardized environmental state vector with a dimension of 7. The data-driven prediction unit deploys a pre-trained lightweight residual convolutional neural network, receives the standardized environmental state vector, and outputs initial nitrogen and phosphorus loss trend prediction values.

[0061] The physical constraint fusion unit performs coupled verification and online fine-tuning, embedding physical conservation constraints into the loss function using the Lagrange multiplier method to generate corrected predicted values ​​for nitrogen and phosphorus loss. The remote control decision unit compares the corrected predicted values ​​with environmental safety thresholds, generating drainage gate opening adjustment commands and fertilization plan adjustment commands based on the degree of exceedance. The command issuance and execution unit transmits the commands to the edge control terminal via a narrowband IoT communication module. The latter includes an electric gate actuator and an intelligent fertilization controller, which respectively execute gate opening adjustment and fertilization amount adjustment operations, with a position feedback accuracy of 1% and a fertilization control error of less than 3%.

[0062] This embodiment, through the aforementioned methods and systems, realizes a complete technical chain from environmental data perception and hybrid modeling prediction to precise execution and control. The mechanism-constrained sub-model ensures the physical rationality of the prediction results, the lightweight residual convolutional neural network guarantees real-time inference efficiency, the online fine-tuning mechanism enables adaptive model correction, and the hierarchical control strategy is directly linked to executable agricultural operations, effectively reducing the risk of nitrogen and phosphorus loss from paddy fields and improving the intelligent level of agricultural non-point source pollution control.

Claims

1. A remote control method for nitrogen and phosphorus loss from paddy fields based on environmental data, characterized in that, include: Acquire historical nitrogen and phosphorus loss monitoring data, real-time meteorological and environmental sensing data, and current water and fertilizer management operation information for the target paddy field area; Based on the historical nitrogen and phosphorus loss monitoring data and the current water and fertilizer management operation information, a mechanism-constrained sub-model is constructed, including the nitrogen migration and transformation kinetic equation and the phosphorus adsorption-desorption equilibrium relationship. Multi-source heterogeneous feature extraction is performed on the real-time meteorological and environmental sensing data to generate a standardized environmental state vector. The standardized environmental state vector is input into a pre-trained lightweight residual convolutional neural network to output an initial nitrogen and phosphorus loss trend prediction value. The initial nitrogen and phosphorus loss trend prediction value is coupled and verified with the output of the mechanism-constrained sub-model. A loss function with physical conservation constraints is constructed using the Lagrange multiplier method, and the weight parameters of the lightweight residual convolutional neural network are fine-tuned online to generate a corrected nitrogen and phosphorus loss prediction value that conforms to the law of conservation of matter. The corrected predicted nitrogen and phosphorus loss is compared with a preset environmental safety threshold. If the corrected predicted nitrogen and phosphorus loss exceeds the environmental safety threshold, a drainage gate opening adjustment command and a fertilization plan adjustment command are generated. These commands are then sent to edge control terminals deployed in the target paddy field area via a narrowband IoT communication module to execute remote control operations. The initial predicted nitrogen and phosphorus loss trend is coupled and verified with the output of the mechanism constraint sub-model. A loss function with physical conservation constraints is constructed using the Lagrange multiplier method. The weight parameters of the lightweight residual convolutional neural network are fine-tuned online, including: defining the nitrogen mass conservation constraint as the change in soil nitrogen pool per unit time equal to the input nitrogen amount minus the output nitrogen amount, mathematically expressed as: ;in Determined by the amount of fertilizer applied and the amount of atmospheric deposition. The amount of nitrogen loss is given by the model prediction. The phosphorus adsorption equilibrium constraint is defined as the phosphorus concentration in the solution phase and the adsorption amount in the solid phase satisfying the Langmuir equation. The above constraint is transformed into a quadratic penalty term, which together with the original prediction error term constitutes a composite loss function. Using the stochastic gradient descent algorithm, the network weights are updated step-by-step online after each new environmental data is received, with the update step size fixed at 0.

001.

2. The method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data according to claim 1, characterized in that, The system acquires historical nitrogen and phosphorus loss monitoring data, real-time meteorological and environmental sensing data, and current water and fertilizer management operation information for the target paddy field area. This includes: acquiring time-series data on total nitrogen concentration, total phosphorus concentration, nitrate nitrogen concentration, ammonium nitrogen concentration, and dissolved phosphorus concentration over historical periods through online water quality monitoring stations deployed at paddy field drainage outlets; acquiring real-time data on rainfall, rainfall intensity, temperature, relative humidity, wind speed, and solar radiation intensity through micro-weather stations deployed in the fields; and acquiring operation log data on current irrigation water volume, fertilizer type, fertilizer amount, fertilization time, and tillage methods through a farmland operation record system.

3. The method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data according to claim 2, characterized in that, A mechanistic constraint sub-model was constructed, which includes nitrogen migration and transformation kinetic equations and phosphorus adsorption-desorption equilibrium relationships. This includes: establishing a soil moisture movement equation based on Darcy's law and coupling it with a convection-dispersion equation to describe the migration process of nitrogen in the soil profile; introducing first-order kinetic reaction terms to characterize the rates of organic nitrogen mineralization, nitrification, and denitrification; using the Langmuir isotherm adsorption equation to describe the adsorption capacity and desorption potential of soil particles for phosphate ions; discretizing the above equations into a finite difference scheme and setting initial and boundary conditions to form a numerically solvable mechanistic constraint sub-model.

4. The method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data according to claim 3, characterized in that, Multi-source heterogeneous feature extraction is performed on the real-time meteorological environment sensing data to generate a standardized environmental state vector, including: combining rainfall amount and rainfall intensity into cumulative rainfall feature; fusing temperature, relative humidity and solar radiation intensity into evapotranspiration potential index; jointly calculating wind speed and rainfall intensity into surface runoff scour intensity factor; normalizing the above features to map their numerical range to the interval between 0 and 1, thus forming a standardized environmental state vector with a dimension of 7.

5. The method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data according to claim 4, characterized in that, The pre-trained lightweight residual convolutional neural network includes three convolutional layers, two residual blocks, and a fully connected output layer. Each convolutional layer has a kernel size of 3 and the number of channels is 16, 32, and 64 respectively. Each residual block contains two convolutional layers with a kernel size of 3 and introduces skip connections. The fully connected output layer outputs two floating-point values, corresponding to the predicted values ​​of nitrogen loss and phosphorus loss, respectively. The network is trained end-to-end using historical monitoring datasets in the offline stage, and the loss function is a weighted sum of mean square error and relative error.

6. The method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data according to claim 1, characterized in that, The preset environmental safety thresholds include a nitrogen loss threshold of 5 kg per hectare per day and a phosphorus loss threshold of 0.5 kg per hectare per day. When the corrected predicted nitrogen and phosphorus loss exceeds either threshold, the control mechanism is triggered. The drainage gate opening adjustment command is generated in stages according to the degree of exceedance: the gate is closed if the exceedance is less than 10%, opened to 1 / 3 of the opening if the exceedance is 10% to 30%, and fully opened if the exceedance is more than 30%. The fertilization plan adjustment command includes postponing the next fertilization time by no less than 48 hours and reducing the original fertilization amount by 20% to 50%, with the reduction ratio being linearly positively correlated with the degree of exceedance.

7. The method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data as described in claim 6, characterized in that, The narrowband IoT communication module adopts the 800 MHz frequency band approved by the State Radio Regulatory Commission. The communication protocol follows the third version of the narrowband IoT standard. Data packets are verified for integrity using a 16-bit cyclic redundancy check code. The instruction transmission delay does not exceed 5 seconds, ensuring the real-time and reliable issuance of control instructions.

8. The method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data according to claim 1, characterized in that, The online fine-tuning only updates the parameters of the fully connected output layer and the last residual block of the lightweight residual convolutional neural network, while keeping the parameters of the remaining network layers frozen.

9. The method for remotely controlling nitrogen and phosphorus loss from paddy fields based on environmental data according to claim 4, characterized in that, The normalization process employs a minimum-maximum scaling method, mapping each feature to the interval between 0 and 1, as shown in the formula: ,in and These are the minimum and maximum values ​​of this feature in the historical 3-year dataset, respectively.

Images