A self-adaptive dynamic grid agricultural precision irrigation method based on micro-potential flow induction

Abstract

This invention discloses an adaptive dynamic grid precision irrigation method based on micro-advection induction. It involves real-time acquisition of multi-source heterogeneous farmland data within a farmland area, followed by preprocessing and spatiotemporal matching. The processed farmland data is input into a deep fusion model based on extended Kalman filtering and convolutional long short-term memory networks to generate a multi-dimensional visualized farmland state map. An improved quadtree adaptive subdivision and merging algorithm dynamically divides the farmland area into several dynamic grid units. Based on the real-time environmental state of each dynamic grid unit, a corresponding micro-advection induction mode is selected within the micro-advection induction mechanism, and the required physical parameters are calculated. A precision irrigation decision model based on the crop water stress index is established to obtain the target irrigation amount for each dynamic grid unit. Irrigation execution instructions for each dynamic grid unit are generated to implement precise irrigation of the farmland area, and the irrigation effect is evaluated.

Description

Technical Field

[0001] This invention relates to the fields of agricultural engineering and irrigation technology, and in particular to an adaptive dynamic grid agricultural precision irrigation method based on micro-advection induced by micro-advection. Background Technology

[0002] Current agricultural irrigation practices mainly exhibit two extreme models: one is "full irrigation," which pursues complete soil moisture coverage. While this method guarantees yield, it results in significant water waste and ignores the heat exchange potential between the crop canopy and the atmosphere. The other is "deficit irrigation," which actively reduces irrigation volume. Although this method saves water, it often relies on experience to set the deficit amount, which can easily cause crops to close their stomata due to water stress during high-temperature periods, leading to a surge in canopy temperature—a risk of "heat-induced death"—ultimately resulting in reduced yield. Existing research typically treats farmland as a homogeneous thermodynamic whole, neglecting the microclimate interactions between "wet patches" and "dry patches" within the farmland. In fact, when wind blows across the surface of moist crops, transpiration produces cool, moist air. This cool air, driven by the wind, flows downwind, forming "advection cooling." However, existing irrigation systems cannot fully quantify this effect, leading to unnecessary repeated irrigation in areas that could have been cooled by advection, or failure to replenish water in a timely manner when the advection effect fails. Summary of the Invention

[0003] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides an adaptive dynamic grid agricultural precision irrigation method based on micro-advection induced by micro-advection. Based on the collected multi-source heterogeneous farmland data, the farmland area is dynamically divided into several dynamic grid units, and precision irrigation is implemented for each dynamic grid unit.

[0004] Technical solution: To achieve the above objectives, the present invention provides an adaptive dynamic grid agricultural precision irrigation method based on micro-advection induction, comprising the following steps:

[0005] S1. Deploy several distributed sensor nodes within the farmland area to collect multi-source heterogeneous farmland data, including soil data, crop data, and micro-meteorological data, in real time, and perform preprocessing and spatiotemporal matching on the multi-source heterogeneous farmland data.

[0006] S2. Construct a deep fusion model based on extended Kalman filter and convolutional long short-term memory network. Input the preprocessed farmland data into the deep fusion model based on extended Kalman filter and convolutional long short-term memory network to generate a multi-dimensional visualized farmland status map.

[0007] S3. Based on the multi-dimensional visualization of farmland status map, an improved quadtree adaptive subdivision and merging algorithm is used to construct a dynamic irrigation grid within the farmland area, dynamically dividing the farmland area into several dynamic grid units.

[0008] S4. Based on the real-time environmental state of each dynamic grid cell, select the corresponding micro-advection induction mode in the micro-advection induction mechanism, and calculate the physical parameters required for micro-advection induction.

[0009] S5. Establish a precision irrigation decision model based on the crop water stress index, and obtain the target irrigation amount for each dynamic grid unit based on the crop water stress index calculated for each dynamic grid unit.

[0010] S6. Based on the calculation results of steps S4 and S5, generate irrigation execution instructions for each dynamic grid cell, irrigate the farmland area based on the irrigation execution instructions for each dynamic grid cell, and evaluate the irrigation effect of the farmland area.

[0011] Furthermore, in step S1, the multi-source heterogeneous farmland data is transmitted to an edge computing gateway via a wireless network for preprocessing and spatiotemporal matching. For spatiotemporal matching, using the least common multiple time step as a benchmark, a cubic spline interpolation algorithm is employed to unify soil data, crop data, and micro-meteorological data onto the same time axis. In the preprocessing operation, a sliding window statistical method is used to remove outliers from the multi-source heterogeneous farmland data. To eliminate the order-of-magnitude differences between different environmental parameters, a Min-Max normalization method is used to normalize the dimensions of the multi-source heterogeneous farmland data. The specific normalization formula is as follows:

[0012]

[0013] In the formula, X' i，t X represents the normalized value of the i-th type of environmental parameter at time t; i，t Represented as the original collected values; X max and X min These represent the maximum and minimum values ​​of this type of parameter within the historical statistical period, respectively.

[0014] Furthermore, in step S2, a deep fusion model based on Extended Kalman Filter (EKF) and Convolutional Long Short-Term Memory (LSM) network is constructed. This deep fusion model includes a data assimilation module, a spatiotemporal feature extraction module, and a decoding module. In the data assimilation module, the EKF algorithm is used to fuse the physical model and the observation data from the sensor nodes. The soil moisture transport physical equation is used as the state transition equation, and the observation data from the sensor nodes is used as the measurement equation to optimally estimate the state of discrete points within the farmland area. In the spatiotemporal feature extraction module, a LSM network is constructed. The discrete point data obtained after processing by the EKF algorithm is converted into gridded image data through Kriging interpolation. The gridded image data is then input into the LSM network for spatiotemporal feature extraction to obtain the spatiotemporal feature map H. t The decoding module is constructed from a multi-branch convolutional decoder, which converts the spatiotemporal feature map H... t The data is input into a multi-branch convolutional decoder for nonlinear mapping and channel separation operations, and then fused to generate a multi-dimensional visualized farmland state map. The formulas for state update and Kalman gain calculation in the extended Kalman filter algorithm are shown below:

[0015]

[0016]

[0017] In the formula, This is represented as the soil moisture state vector corrected at time k; Represented as prior predictions based on a physical model; z k Represented as the actual observation vector of the sensor node; K k H represents the Kalman gain; H represents the observation matrix; H T R is represented as the transpose of the observation matrix; R is represented as the measurement noise covariance matrix. It is represented as the prior error covariance matrix.

[0018] Furthermore, the multidimensional visualized farmland state map includes a micro-advection induced gradient layer, a potential energy layer of soil moisture energy state, and a CWSI crop water deficit index layer; the micro-advection induced gradient layer aims to quantify the inducing effect of environmental potential energy on soil moisture movement at the microscale, calculating the micro-advection induced vector at any grid point (x, y) in space. The calculated micro-advection induced vectors are combined and mapped in space to construct a global micro-advection induced vector field, which is then used as a micro-advection induced gradient layer. The calculation formula is as follows:

[0019]

[0020] In the formula, Represented as the temperature gradient vector at grid point (x, y), Let α represent the soil moisture content gradient vector at grid point (x, y); α represents the thermal induction coefficient, and β represents the wet induction coefficient.

[0021] Furthermore, in step S3, based on the multi-dimensional visualized farmland state map, an improved quadtree adaptive subdivision and merging algorithm is used to construct a dynamic irrigation grid within the farmland area, dynamically dividing the farmland area into several dynamic grid units; the farmland area is used as the first initial grid unit, and the heterogeneity evaluation index of the current first initial grid unit is calculated in real time; when the heterogeneity evaluation index of the first initial grid unit exceeds a preset splitting threshold T... split When the time limit is reached, an improved quadtree adaptive subdivision and merging algorithm is used to divide the current first initial grid cell into four first sub-grid cells; all four first sub-grid cells are used as second initial grid cells, and the heterogeneity evaluation index of any second initial grid cell is calculated in real time; when the heterogeneity evaluation index of any second initial grid cell exceeds the preset splitting threshold T, the algorithm is called to divide the first initial grid cell into four sub-grid cells. split If the heterogeneity evaluation index of all initial grid cells does not exceed the preset splitting threshold T, then an improved quadtree adaptive subdivision and merging algorithm is used to divide the current second initial grid cell into four second sub-grid cells; this process continues until the heterogeneity evaluation index of all initial grid cells does not exceed the preset splitting threshold T. split This yields several dynamic mesh elements;

[0022] In any initial grid cell, the heterogeneity evaluation index of the four sub-grid cells is combined with a preset merging threshold T. merge For comparison, when there are two or more sub-grid cells with heterogeneity evaluation indices below the merging threshold T... merge If the two or more sub-grid cells are then merged into a new dynamic grid cell, the result is a new dynamic grid cell.

[0023] Furthermore, the heterogeneity evaluation index of the initial grid cell is a weighted comprehensive index of the variance of soil moisture content, the variance of crop water requirement, and the mean of micro-advection induction intensity within the initial grid cell. The calculation process is as follows:

[0024]

[0025] In the formula, H i Let be the heterogeneity evaluation index of the i-th grid cell. This is expressed as the spatial variance of soil moisture content within a grid cell. This is expressed as the spatial variance of crop evapotranspiration demand within a grid cell. It is expressed as the average magnitude of the micro-adaptation induced vector within the grid cell, where w1, w2, and w3 are the weight coefficients of each factor, and w1+w2+w3=1.

[0026] Furthermore, at regular time intervals, the farmland area is dynamically divided into several dynamic grid cells. A time step adjustment model based on the rate of environmental change is constructed to adaptively update the execution time step, thus obtaining the time interval Δt for the next division of dynamic grid cells. k+1 The calculation process is as follows:

[0027]

[0028]

[0029] In the formula, Δt k+1 T represents the time interval for the next dynamic mesh generation. base R represents the preset reference time step, λ represents the time sensitivity coefficient; change Expressed as the rate of change of environmental state, H i (k) and H i (k-1) represent the heterogeneity evaluation indices at the current time k and the previous time k-1, respectively, Δt k This represents the time interval at the current moment.

[0030] Furthermore, in step S5, a precision irrigation decision model based on the crop water stress index is established. The target irrigation amount for each dynamic grid cell is calculated based on the crop water stress index calculated for each dynamic grid cell. The formula for calculating the crop water stress index is as follows:

[0031]

[0032] In the formula, CWSI represents the crop water stress index, and T c T represents the measured canopy temperature. wet T represents the canopy temperature under conditions of adequate water supply. dry Represented as canopy temperature under complete water stress;

[0033] The target irrigation amount for each dynamic grid cell is calculated based on the crop water stress index of each dynamic grid cell. The calculation process is as follows:

[0034]

[0035] In the formula, I i Let K represent the target irrigation amount for the i-th dynamic grid cell. cRepresented as crop coefficient, ET0 represents reference evapotranspiration, A i Let f(CWSI) represent the area of ​​the i-th dynamic mesh cell. i ) represents the water stress correction function. P represents the duration of the current irrigation decision cycle. eff Expressed as effective precipitation, ΔS i This is represented by changes in soil water storage.

[0036] Beneficial Effects: This invention provides an adaptive dynamic grid precision irrigation method for agriculture based on micro-advection induction. By deploying several distributed sensor nodes within a farmland area, it collects multi-source heterogeneous farmland data in real time, including soil, crop, and micro-meteorological data, thereby improving the monitoring accuracy of large-scale farmland areas. A deep fusion model based on extended Kalman filtering and convolutional long short-term memory networks is used to process the multi-source heterogeneous data, obtaining a multi-dimensional visualized farmland state map and constructing a dynamic irrigation grid within the farmland area. The ratio and arrangement of "wet spots" and "dry spots" are dynamically adjusted to avoid localized drought and waterlogging caused by evenly distributed irrigation. A suitable micro-advection induction mechanism is selected, and the physical parameters of micro-advection and the target irrigation amount for each dynamic grid unit are calculated, achieving precise irrigation for each dynamic grid unit and maximizing water conservation while avoiding thermal stress. Through dynamic and refined management of the farmland irrigation area, precise control of the irrigation process is achieved, significantly improving water resource utilization efficiency. Attached Figure Description

[0037] Figure 1 The flowchart shows an adaptive dynamic grid agricultural precision irrigation method based on micro-adaptive flow induction.

[0038] Figure 2 A schematic diagram of the dynamic irrigation grid division within a farmland area;

[0039] Figure 3 This is a schematic diagram of the architecture of an adaptive dynamic grid agricultural precision irrigation model based on micro-advection. Detailed Implementation

[0040] The invention will now be further described with reference to the accompanying drawings.

[0041] like Figure 1 As shown, an adaptive dynamic grid agricultural precision irrigation method based on micro-advection-induced precision irrigation includes the following steps:

[0042] S1. Deploy several distributed sensor nodes in the farmland area to collect multi-source heterogeneous farmland data, including soil data, crop data and micro-meteorological data, in real time. Transmit the multi-source heterogeneous farmland data to the edge computing gateway via wireless network for preprocessing and spatiotemporal matching.

[0043] S2. Construct a deep fusion model based on EKF extended Kalman filter and ConvLSTM convolutional long short-term memory network. Input the preprocessed farmland data into the deep fusion model based on extended Kalman filter and convolutional long short-term memory network to generate a multi-dimensional visualized farmland state map, which is high-precision farmland state data.

[0044] S3. Based on the multi-dimensional visualization of farmland status map, an improved quadtree adaptive subdivision and merging algorithm is used to construct a dynamic irrigation grid within the farmland area, dynamically dividing the farmland area into several dynamic grid units.

[0045] S4. For each dynamic grid cell, select the corresponding micro-advection induction mode in the micro-advection induction mechanism according to the real-time environmental state of each dynamic grid cell, and calculate the physical parameters required for micro-advection induction.

[0046] S5. Establish a precision irrigation decision model based on the crop water stress index, and obtain the target irrigation amount for each dynamic grid unit based on the crop water stress index calculated for each dynamic grid unit.

[0047] S6. Based on the calculation results of steps S4 and S5, generate irrigation execution instructions for each dynamic grid cell, irrigate the farmland area based on the irrigation execution instructions for each dynamic grid cell, and evaluate the irrigation effect of the farmland area.

[0048] In step S1, several distributed sensor nodes are deployed within the farmland area to construct a three-dimensional sensing network encompassing roots, ground, and air to acquire basic environmental information. Multi-parameter soil probes are buried at different depths (e.g., 10cm, 30cm, and 60cm) in the rhizosphere to collect soil data in real time. The density of these probes is dynamically adjusted based on the spatial variability of soil texture, with increased density in areas with abrupt changes in texture. In the canopy layer, non-contact infrared temperature sensors and multispectral cameras are used to collect crop data. Micro-meteorological stations are set up in the near-surface micro-meteorological layer to collect micro-meteorological data. The soil data includes soil volumetric water content, soil matrix potential, soil temperature, and soil electrical conductivity. The crop data includes crop canopy temperature, normalized difference vegetation index (NDVI), and leaf area index (LAI). The micro-meteorological data includes air temperature and humidity, photosynthetically active radiation, micro-wind speed, and wind direction. All sensor nodes transmit the collected multi-source heterogeneous farmland data to an edge computing gateway via a wireless network.

[0049] Several distributed sensor nodes are deployed within the farmland area, including fault diagnosis and emergency response mechanisms. When a sensor node fails or an actuator malfunctions, the system automatically switches to a backup sensor node or uses interpolation estimation based on historical data, and activates a pre-defined emergency irrigation plan. The multi-source heterogeneous farmland data also includes UAV remote sensing data. UAVs are deployed at high altitudes, and their multispectral cameras collect crop normalized vegetation index and canopy cover information as UAV remote sensing data, improving the monitoring accuracy of large-scale farmland.

[0050] In step S1, multi-source heterogeneous farmland data is transmitted to an edge computing gateway via a wireless network for preprocessing and spatiotemporal matching. For spatiotemporal matching, to address the issue of inconsistent sampling frequencies in the multi-source heterogeneous farmland data, a cubic spline interpolation algorithm is used, based on the least common multiple time step, to unify soil data, crop data, and micro-meteorological data onto the same time axis. In the preprocessing stage, to remove noise from the multi-source heterogeneous farmland data, a sliding window statistical method is used to eliminate outliers. To eliminate the order-of-magnitude differences between different environmental parameters, the Min-Max normalization method is used to normalize the dimensions of the multi-source heterogeneous farmland data. The specific normalization formula is as follows:

[0051]

[0052] In the formula, X' i，t X represents the normalized value of the i-th type of environmental parameter at time t; i，t Represented as the original collected values; X max and X minThese represent the maximum and minimum values ​​of this type of parameter within the historical statistical period, respectively.

[0053] In step S2, a deep fusion model based on Extended Kalman Filter (EKF) and Convolutional Long Short-Term Memory (LSTM) networks is constructed. This deep fusion model includes a data assimilation module, a spatiotemporal feature extraction module, and a decoding module. The data assimilation module uses the EKF EKF algorithm to fuse the physical model and the observation data from sensor nodes. The soil moisture transport physical equation is used as the state transition equation, and the observation data from the sensor nodes is used as the measurement equation to optimally estimate the state of discrete points within the farmland area. The spatiotemporal feature extraction module uses a ConvLSTM network. The discrete point data obtained after processing by the EKF EKF algorithm is converted into continuous gridded image data through Kriging interpolation. The gridded image data is then input into the ConvLSTM network for spatiotemporal feature extraction to obtain the spatiotemporal feature map H. t The decoding module is constructed from a multi-branch convolutional decoder, which converts the spatiotemporal feature map H... t The data is input to a multi-branch convolutional decoder for nonlinear mapping and channel separation operations, and then fused to generate a multi-dimensional visualized farmland state map. The spatiotemporal feature extraction module and the decoding module are used to dynamically predict the evolution of the farmland and output the multi-dimensional map.

[0054] The formulas for calculating the state update and Kalman gain in the EKF extended Kalman filter algorithm are as follows:

[0055]

[0056]

[0057] In the formula, This is represented as the soil moisture state vector corrected at time k; Represented as prior predictions based on a physical model; z k Represented as the actual observation vector of the sensor node; K k H represents the Kalman gain; H represents the observation matrix; H T R is represented as the transpose of the observation matrix; R is represented as the measurement noise covariance matrix. This is represented as the prior error covariance matrix. In the above processing, the Kalman gain K... kIt plays a role in dynamic weight adjustment, used to measure the reliability of the physical model predictions and the actual observations of the sensor nodes; the state update operation feeds back the residuals of the actual observations of the sensor nodes to the prior predictions through Kalman gain, thereby effectively filtering out measurement noise in the observation data of the sensor nodes and correcting the systematic bias of the physical model; after processing by the EKF extended Kalman filter algorithm, the output data format is still discrete point data corresponding to the physical location of each sensor node.

[0058] Discrete point data is transformed into continuous gridded image data through Kriging interpolation. This gridded image data is then input into a ConvLSTM (Convolutional Long Short-Term Memory) network for spatiotemporal feature extraction. The ConvLSTM network incorporates convolution operations into its gating mechanism to simultaneously capture the spatial topological correlations and temporal dependencies of farmland environmental parameters. The ConvLSTM network extracts the spatiotemporal feature map H. t Spatiotemporal feature map H t Includes spatial distribution characteristics of humidity, temperature, etc. in farmland areas and their dynamic evolution over time; the spatiotemporal feature map H... t The data is input to a multi-branch convolutional decoder for nonlinear mapping and channel separation operations, fusing spatial distribution characteristics of farmland humidity and temperature, as well as their dynamic evolution over time, to generate a multi-dimensional, visualized farmland state map. The gating state calculation formula within the ConvLSTM convolutional long short-term memory network is shown below:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] In the formula, X t H represents the gridded image data input at time t; t and H t-1 Let C represent the hidden layer states at the current and previous time steps, respectively; t and C t-1 These represent the cell states at the current and previous time steps, respectively; i t f t o t These are represented as the input gate, forget gate, and output gate, respectively; * indicates a convolution operation. It is represented as the Hadamard product, σ represents the Sigmoid activation function, and W and b represent the corresponding convolution kernel weight matrix and bias term, respectively.

[0065] The multidimensional visualized farmland state map includes a micro-advection induced gradient layer, a potential energy layer of soil moisture and energy state, and a CWSI crop water deficit index layer. The micro-advection induced gradient layer is obtained by a decoding module that extracts temperature and humidity feature channels, and calculates the micro-advection induced vector at any grid point (x, y) in space. The following layers are generated: a micro-advection induced gradient layer; a soil moisture energy state potential energy layer, where the decoding module extracts soil matrix potential characteristic channels and generates a potential energy distribution layer reflecting the strength of soil moisture adsorption and capillary action through normalization mapping; and a CWSI crop water deficit index layer, where the decoding module extracts canopy temperature and micro-meteorological characteristic channels and, combined with a water stress calculation model, outputs the crop water deficit index grid-by-grid to form the CWSI crop water deficit index layer. The aforementioned grids and grid points divide the farmland area into several uniform parameter calculation grids and several parameter calculation grid points. These divided parameter calculation grids and grid points are used to calculate and generate the micro-advection induced gradient layer, the soil moisture energy state potential energy layer, and the CWSI crop water deficit index layer.

[0066] The micro-advection induced gradient layer aims to quantify the inducing effect of environmental potential energy on soil moisture movement at the microscale, calculating the micro-advection induced vector at any grid point (x, y) in the computational space. The calculated micro-advection induced vectors are combined and mapped in space to construct a global micro-advection induced vector field, which is then used as a micro-advection induced gradient layer. The calculation formula is as follows:

[0067]

[0068] In the formula, Represented as the temperature gradient vector at grid point (x, y), Let α represent the soil moisture content gradient vector at grid point (x, y); α represents the thermal induction coefficient, used to characterize the induction strength of the temperature gradient on water migration; β represents the wet induction coefficient, used to characterize the induction strength of the soil moisture content gradient on water migration. The micro-adaptation induction vectors at all grid points are also considered. The combined global vector field intuitively represents the natural migration trend of fluids driven by thermodynamics and concentration differences at the microscale.

[0069] Because several distributed sensor nodes are deployed within the farmland area to collect multi-source heterogeneous farmland data, including soil data, crop data, and micro-meteorological data, the multi-dimensional visualized farmland state map generated by the deep fusion model based on extended Kalman filter and convolutional long short-term memory network should also be dynamic. That is, every time interval, a new multi-dimensional visualized farmland state map is generated using the deep fusion model based on extended Kalman filter and convolutional long short-term memory network to update the multi-dimensional visualized farmland state map. This time interval is the same as the time interval for dynamically dividing the farmland area into several dynamic grid units.

[0070] like Figure 1 As shown, in step S3, based on the multi-dimensional visualized farmland state map, an improved quadtree adaptive subdivision and merging algorithm is used to construct a dynamic irrigation grid within the farmland area. The farmland area is dynamically divided into several dynamic grid units according to the calculation time interval. The farmland area is used as the first initial grid unit, and the heterogeneity evaluation index of the current first initial grid unit is calculated in real time. When the heterogeneity evaluation index of the first initial grid unit exceeds a preset splitting threshold T... split When the time limit is reached, an improved quadtree adaptive subdivision and merging algorithm is used to divide the current first initial grid cell into four first sub-grid cells; all four first sub-grid cells are used as second initial grid cells, and the heterogeneity evaluation index of any second initial grid cell is calculated in real time; when the heterogeneity evaluation index of any second initial grid cell exceeds the preset splitting threshold T, the algorithm is called to divide the first initial grid cell into four sub-grid cells. split If the heterogeneity evaluation index of all initial grid cells does not exceed the preset splitting threshold T, then an improved quadtree adaptive subdivision and merging algorithm is used to divide the current second initial grid cell into four second sub-grid cells; this process continues until the heterogeneity evaluation index of all initial grid cells does not exceed the preset splitting threshold T. split This yields several dynamic mesh elements;

[0071] In any initial grid cell, the heterogeneity evaluation index of the four sub-grid cells is combined with a preset merging threshold T. merge For comparison, when there are two or more sub-grid cells with heterogeneity evaluation indices below the merging threshold T... merge If necessary, the two or more sub-grid cells are merged into a new dynamic grid cell to reduce computational redundancy. The obtained dynamic grid cells undergo grid boundary smoothing, and a spline interpolation algorithm is used to smoothly transition the grid boundaries of adjacent dynamic grid cells, avoiding abrupt changes in irrigation water volume at the boundaries.

[0072] The heterogeneity evaluation index of the initial grid cell is a weighted composite index of the variance of soil moisture content, the variance of crop water requirement, and the mean of micro-advection induction intensity within the initial grid cell. The calculation process is shown below:

[0073]

[0074] In the formula, H i Let be the heterogeneity evaluation index of the i-th initial grid cell. This is expressed as the spatial variance of soil moisture content within the initial grid cell. This represents the spatial variance of crop evapotranspiration demand within the initial grid cell. The mean modulus of the micro-advection induced vector within the initial grid cell represents the average micro-advection induced intensity, characterizing the water migration potential. w1, w2, and w3 are the weighting coefficients for each factor, namely the spatial variance of soil moisture content, the spatial variance of crop evapotranspiration demand, and the mean modulus of the micro-advection induced vector, respectively, with w1 + w2 + w3 = 1. The aforementioned variances of soil moisture content, crop water demand, and the mean micro-advection induced intensity within the initial grid cell are all obtained through spatial statistical analysis of data points in the corresponding layers of the multidimensional visualized farmland condition map within the current initial grid cell.

[0075] At regular intervals, the farmland area is dynamically divided into several dynamic grid cells. A time step adjustment model based on the rate of environmental change is constructed to adaptively update the execution time step, thus obtaining the time interval for the next division of dynamic grid cells. The calculation process is as follows:

[0076]

[0077]

[0078] In the formula, T represents the time interval for the next dynamic mesh generation. base R represents the preset reference time step, λ represents the time sensitivity coefficient; change Expressed as the rate of change of environmental state, H i (k) and H i (k-1) represent the heterogeneity evaluation indices for the current time k and the previous time k-1, respectively. This represents the time interval at the current moment. The mechanism of dynamically dividing farmland areas into several dynamic grid units based on the calculated time interval introduces a rate of change to ensure that the update interval is automatically shortened when the environment changes drastically, i.e., when the heterogeneity evaluation index changes significantly per unit time, and the interval is extended when the environment is stable.

[0079] In step S4, a corresponding micro-advection induction mode is selected from the micro-advection induction mechanism based on the real-time environmental state of each dynamic grid cell, and the physical parameters required for micro-advection induction are calculated. The micro-advection induction mechanism includes three modes: thermocapillary effect induction mode, concentration gradient induction mode, and surfactant gradient induction mode. Based on the real-time environmental state within each dynamic grid cell, the optimal micro-advection induction mode is adaptively selected by comprehensively considering soil texture, current soil moisture content, and crop growth period.

[0080] When the soil texture of the dynamic grid cell is detected as heavy clay soil and the current soil moisture content is lower than the preset infiltration threshold, the surfactant gradient induction mode is selected; when the crop is detected to be in a critical water-demanding growth stage and the multi-dimensional visualized farmland status map shows a rhizosphere water deficit, the concentration gradient induction mode is selected; when the ambient temperature is detected to be lower than the preset temperature threshold or there is localized heat stress, the thermocapillary effect induction mode is selected. Considering the limitations of energy consumption and physical scale, the thermocapillary effect induction mode is mainly suitable for micro-precision irrigation scenarios such as high-value-added greenhouses, hydroponics, or miniature potted plants; while in large-scale farmland areas, when the system evaluates that the expected energy consumption of the thermocapillary effect induction mode exceeds the preset economic threshold, the system automatically triggers a degradation strategy, prioritizing the use of the concentration gradient induction mode or the surfactant gradient induction mode to control overall energy consumption. When two or more of the above single triggering conditions are met simultaneously, a multi-mode collaborative control strategy is adopted. Normalized weights are allocated according to the degree of deviation of each environmental factor, and multiple modes are selected to combine together to achieve micro-advection induction. For example, the thermocapillary effect induction mode and the surfactant gradient induction mode are used together to achieve micro-advection induction.

[0081] The thermocapillary effect-induced mode establishes a local temperature gradient by placing micro-heating elements at specific locations in the irrigation network, inducing irrigation water to flow directionally along the temperature gradient direction. The thermocapillary velocity is calculated using the following formula:

[0082]

[0083] In the formula, Expressed as thermocapillary velocity,

[0084] It is expressed as the surface tension temperature coefficient. Let μ represent the temperature difference, μ represent the fluid dynamic viscosity, and L represent the characteristic length.

[0085] The concentration gradient induced flow model establishes an osmotic pressure gradient between the rhizosphere and the surrounding soil by adding a biodegradable osmotic regulator to the irrigation water, thereby promoting the directional migration of water towards the roots. The formula for calculating the concentration gradient induced flow rate is shown below:

[0086]

[0087] In the formula, K represents the concentration gradient-induced flow rate. c It is expressed as the soil permeability and conductivity coefficient, C root Represented as the concentration of regulatory substances in the rhizosphere region, C soil The concentration of regulating substances in the surrounding soil is represented by d, and the effective interaction distance between the rhizosphere and the surrounding soil is represented by d.

[0088] The surfactant gradient-induced flow mode employs biodegradable surfactants to establish a surface tension gradient field in soil pores, enhancing the transport efficiency of water in unsaturated soil. The formula for calculating the surfactant gradient-induced flow rate is shown below:

[0089]

[0090] In the formula, The value is expressed as the surfactant gradient-induced flow rate, k represents the inherent soil permeability, and μ represents the fluid dynamic viscosity. Expressed as the rate of change of surface tension with surfactant concentration, It is represented as the surfactant concentration gradient vector in space.

[0091] When the thermocapillary effect induction mode is selected as the micro-advection induction mode, the physical parameter required for micro-advection induction is the thermocapillary velocity; when the concentration gradient induction mode is selected, the physical parameter required for micro-advection induction is the concentration gradient induced velocity; when the surfactant gradient induction mode is selected, the physical parameter required for micro-advection induction is the surfactant gradient induced velocity. The thermocapillary velocity, concentration gradient induced velocity, and surfactant gradient induced velocity represent the water flow velocities during irrigation under different micro-advection induction modes.

[0092] In step S5, a precision irrigation decision model based on the crop water stress index is established. This precision irrigation decision model is a quantitative relationship model that calculates the target irrigation amount for each dynamic grid unit using the input crop water stress index. Based on the crop water stress index calculated for each dynamic grid unit, the target irrigation amount for each dynamic grid unit is calculated. The formula for calculating the crop water stress index is as follows:

[0093]

[0094] In the formula, CWSI represents the crop water stress index, and T c T represents the measured canopy temperature. wet T represents the canopy temperature under conditions of adequate water supply. dryRepresented as canopy temperature under complete water stress;

[0095] The target irrigation amount for each dynamic grid cell is calculated based on the crop water stress index of each dynamic grid cell. The calculation process is as follows:

[0096]

[0097] In the formula, I i Let K represent the target irrigation amount for the i-th dynamic grid cell. c Represented as crop coefficient, ET0 represents reference evapotranspiration, A i Let f(CWSI) represent the area of ​​the i-th dynamic mesh cell. i ) represents the water stress correction function. P represents the duration of the current irrigation decision cycle. eff Expressed as effective precipitation, ΔS i This is represented by changes in soil water storage.

[0098] In step S6, irrigation execution instructions for each dynamic grid unit are generated based on the calculation results of steps S4 and S5, and irrigation is carried out on the farmland area based on the irrigation execution instructions of each dynamic grid unit; the selected micro-advection induction mode and the physical parameters required for micro-advection induction are obtained from step S4, and the target irrigation amount for each dynamic grid unit is obtained from step S5; the selected micro-advection induction mode and the physical parameters required for micro-advection induction are encoded and fused with the target irrigation amount for each dynamic grid unit to generate irrigation execution instructions containing water volume control parameters and micro-advection induction control parameters.

[0099] The irrigation execution command specifically includes water volume control parameters and micro-adjustment flow induction control parameters. The water volume control parameters include irrigation start and stop times for controlling the target irrigation volume, the opening degree of the branch pipe solenoid valve, and the operating frequency of the variable frequency water pump. The micro-adjustment flow induction control parameters are determined according to the selected micro-adjustment flow induction mode. When the thermocapillary effect induction mode is selected, it includes the heating power and heating duration of the micro heating element. When the concentration gradient induction mode is selected, it includes the injection concentration and injection rate of the osmotic conditioning substance. When the surfactant gradient induction mode is selected, it includes the injection ratio and mixing flow rate of the biodegradable surfactant.

[0100] The execution process of the irrigation execution command involves sending the generated irrigation execution command to the irrigation execution mechanism deployed in the farmland area through an edge computing gateway for hardware execution. The irrigation execution mechanism includes a variable frequency water pump, a branch solenoid valve, a micro heating element array, and an intelligent proportional dosing pump. The specific execution process is as follows: First, based on the water volume control parameters, the operating frequency of the variable frequency water pump is adjusted to provide basic pipeline pressure, and the opening and closing sequence and opening degree of the branch solenoid valve of the corresponding dynamic grid unit are controlled to achieve precise water distribution. Secondly, based on the micro-advection induction control parameters, the corresponding induction execution hardware is activated. If the thermocapillary effect induction mode is executed, the micro-heating element array is activated to establish a local temperature gradient according to the calculated heating power, inducing irrigation water to flow along a predetermined path. If the concentration gradient induction mode is executed, the intelligent proportional dosing pump is controlled to mix osmotic conditioning substances into the irrigation main pipe according to the set injection concentration and rate, establishing an osmotic pressure gradient. If the surfactant gradient induction mode is executed, the intelligent proportional dosing pump is controlled to mix surfactants according to the set injection ratio, optimizing soil moisture infiltration characteristics. If the multi-mode collaborative control strategy is executed, the operating status of the micro-heating element array and the intelligent proportional dosing pump is synchronously coordinated. During the irrigation of farmland areas, the soil moisture content changes of each dynamic grid unit are monitored in real time, and the parameters in the irrigation execution command are dynamically adjusted using a model predictive control algorithm to ensure that the deviation between the actual irrigation amount and the target irrigation amount is within the allowable range.

[0101] like Figure 3 As shown, after irrigation is completed, a comprehensive evaluation of the irrigation effect in the farmland area is conducted. The evaluation indicators include irrigation uniformity, water use efficiency, energy consumption indicators, and crop growth response. The Christensen uniformity coefficient is used to calculate the irrigation uniformity of the farmland area, and the calculation process is shown below:

[0102]

[0103] Where CU represents irrigation uniformity, x i This is represented as the irrigation amount at each sampling point. This represents the average irrigation amount, and n is the number of sampling points.

[0104] The water use efficiency is used to assess the efficiency of converting irrigation water into crop yield, and the calculation formula is as follows:

[0105]

[0106] In the formula, WUE represents water use efficiency, Y represents the actual crop yield per unit area, and I... total P represents the total irrigation amount per unit area during the growing season. eff This is expressed as effective rainfall.

[0107] The energy consumption index is used to quantify the comprehensive energy consumption during irrigation and micro-advection induction processes, and the calculation formula is as follows:

[0108]

[0109] In the formula, ECI represents the comprehensive energy consumption index per unit of irrigation water, and E pump E represents the operating energy consumption of a variable frequency water pump. heat E is expressed as the thermally induced energy consumption of a micro heating element array. mix This is represented by the operating energy consumption of the intelligent proportional dosing pump. The irrigation system has a built-in automatic trade-off mechanism based on energy consumption and benefits. In actual operation, when the calculated value of the ECI comprehensive energy consumption index exceeds the set economic benefit threshold, the system will automatically adjust the micro-adjustment induced control parameters, such as reducing the heating power of the micro heating element or shortening the heating time; or directly triggering the above-mentioned mode degradation strategy, thereby ensuring the economic practicality and scalability of the entire irrigation system while protecting crops from water stress.

[0110] The crop growth response is assessed based on remote sensing multispectral data to evaluate changes in crop growth before and after irrigation, and the calculation formula is as follows:

[0111]

[0112] In the formula, CGR represents the crop growth comprehensive response index, which is the crop growth response; NDVI pre and NDVI post The normalized difference in vegetation index (LAI) before and after the irrigation cycle are represented respectively. pre and LAI post W1 and W2 represent the leaf area index before and after the irrigation cycle, respectively. W1 and W2 are the set weighting coefficients, and W1 + W2 = 1.

[0113] The method for precision irrigation of agriculture based on micro-advection-induced adaptive dynamic grid is described, and its corresponding model is the precision irrigation model of agriculture based on micro-advection-induced adaptive dynamic grid. Based on the above evaluation results, a reinforcement learning algorithm based on deep deterministic policy gradient is used to optimize the precision irrigation model of agriculture based on micro-advection-induced adaptive dynamic grid online, continuously improving the model parameters. The steps include:

[0114] S1-1. Define the state space, taking the multi-dimensional visualized farmland state map features, current micro-meteorological data, and irrigation uniformity, water use efficiency, energy consumption indicators, and crop growth response from the previous irrigation cycle as the environmental state input S for the reinforcement learning agent. t ;

[0115] S1-2, Define the action space, and define the action A output by the reinforcement learning algorithm based on the deep deterministic policy gradient.t Action At represents the continuous adjustment of key parameters in the adaptive dynamic grid agricultural precision irrigation model based on micro-adaptation; specifically, it includes the fine-tuning of the weight coefficients w1, w2, and w3 of each factor in the heterogeneity evaluation index, and the preset splitting threshold T. split The adjustment amount, the compensation value of thermal induction coefficient α and wet induction coefficient β in the micro-adaptation induced gradient layer;

[0116] S1-3, Constructing the reward function R t To maximize water use efficiency and crop growth response while minimizing energy consumption, a comprehensive reward function Rt is constructed; the calculation process is shown below:

[0117]

[0118] In the formula, λ1, λ2, λ3, and λ4 represent the reward weights of each evaluation indicator; WUE t Let CGR be the water use efficiency for the t-th irrigation cycle. t Let ECI represent the crop growth response during the t-th irrigation cycle. t Let CU represent the energy consumption index for the t-th irrigation cycle. t This represents the irrigation uniformity during the t-th irrigation cycle.

[0119] S1-4, Online interaction and parameter updates: After each irrigation decision cycle, the reinforcement learning algorithm updates the parameters based on the current state S. t and perform action A t Update the parameters of the adaptive dynamic grid agricultural precision irrigation model based on micro-advection induced by micro-advection.

[0120] After obtaining the actual irrigation assessment results in the next cycle, the immediate reward function R is calculated. t And observed a new environmental state S t+1 ; the empirical tuple (S t A t R t S t+1 The samples are stored in the experience replay pool; the reinforcement learning algorithm based on deep deterministic policy gradient periodically randomly extracts small batches of samples from the experience replay pool, and uses time difference error to update the neural network weights of the policy network and the value network respectively; through continuous online interaction with the farmland environment, the reinforcement learning algorithm continuously tries and iterates, and finally enables the adaptive dynamic grid agricultural precision irrigation model based on micro-advection to converge to the optimal parameter configuration, thereby maximizing the comprehensive benefits of water saving, energy saving and increased production.

[0121] The above description is merely a preferred embodiment of the present invention. Those skilled in the art can make several modifications and optimizations based on the above disclosure without departing from the basic principles described above. These modifications and optimizations should be considered within the scope of protection as understood by the present invention.

Claims

1. A method for adaptive dynamic grid agricultural precision irrigation based on micro-advection induction, characterized in that: Includes the following steps: S1. Deploy several distributed sensor nodes within the farmland area to collect multi-source heterogeneous farmland data, including soil data, crop data, and micro-meteorological data, in real time, and perform preprocessing and spatiotemporal matching on the multi-source heterogeneous farmland data. S2. Construct a deep fusion model based on extended Kalman filter and convolutional long short-term memory network. Input the preprocessed farmland data into the deep fusion model based on extended Kalman filter and convolutional long short-term memory network to generate a multi-dimensional visualized farmland status map. S3. Based on the multi-dimensional visualization of farmland status map, an improved quadtree adaptive subdivision and merging algorithm is used to construct a dynamic irrigation grid within the farmland area, dynamically dividing the farmland area into several dynamic grid units. S4. Based on the real-time environmental state of each dynamic grid cell, select the appropriate micro-advection induction mode in the micro-advection induction mechanism, and calculate the physical parameters required for micro-advection induction. S5. Establish a precision irrigation decision model based on the crop water stress index, and obtain the target irrigation amount for each dynamic grid unit according to the crop water stress index calculated for each dynamic grid unit. S6. Based on the calculation results of steps S4 and S5, generate irrigation execution instructions for each dynamic grid cell, irrigate the farmland area based on the irrigation execution instructions for each dynamic grid cell, and evaluate the irrigation effect of the farmland area.

2. The adaptive dynamic grid agricultural precision irrigation method based on micro-adaptive flow induction according to claim 1, characterized in that: In step S1, multi-source heterogeneous farmland data is transmitted to an edge computing gateway via a wireless network for preprocessing and spatiotemporal matching. For spatiotemporal matching, using the least common multiple time step as a benchmark, a cubic spline interpolation algorithm is employed to unify soil data, crop data, and micro-meteorological data onto the same time axis. In preprocessing, a sliding window statistical method is used to remove outliers from the multi-source heterogeneous farmland data. To eliminate the order-of-magnitude differences between different environmental parameters, the Min-Max normalization method is used to normalize the dimensions of the multi-source heterogeneous farmland data. The specific normalization formula is as follows: In the formula, X' i，t X represents the normalized value of the i-th type of environmental parameter at time t; i，t Represented as the original collected values; X max and X min These represent the maximum and minimum values ​​of this type of parameter within the historical statistical period, respectively.

3. The adaptive dynamic grid agricultural precision irrigation method based on micro-adaptive flow induction according to claim 1, characterized in that: In step S2, a deep fusion model based on Extended Kalman Filter (EKF) and Convolutional Long Short-Term Memory (LSM) network is constructed. This deep fusion model includes a data assimilation module, a spatiotemporal feature extraction module, and a decoding module. The data assimilation module uses the EKF algorithm to fuse the physical model and the observation data from sensor nodes, using the soil moisture transport physical equation as the state transition equation and the sensor node observation data as the measurement equation to optimally estimate the state of discrete points within the farmland area. The spatiotemporal feature extraction module uses a LSM network to convert the discrete point data obtained after EKF processing into gridded image data through Kriging interpolation. The gridded image data is then input into the LSM network for spatiotemporal feature extraction to obtain the spatiotemporal feature map H. t The decoding module is constructed from a multi-branch convolutional decoder, which converts the spatiotemporal feature map H... t The data is input into a multi-branch convolutional decoder for nonlinear mapping and channel separation operations, and then fused to generate a multi-dimensional visualized farmland state map. The formulas for state update and Kalman gain calculation in the extended Kalman filter algorithm are shown below: In the formula, This is represented as the soil moisture state vector corrected at time k; Represented as prior predictions based on a physical model; z k Represented as the actual observation vector of the sensor node; K k H represents the Kalman gain; H represents the observation matrix; H T R is represented as the transpose of the observation matrix; R is represented as the measurement noise covariance matrix. It is represented as the prior error covariance matrix.

4. The adaptive dynamic grid agricultural precision irrigation method based on micro-adaptive flow induction according to claim 3, characterized in that: The multidimensional visualized farmland state map includes a micro-advection induced gradient layer, a potential energy layer of soil moisture energy state, and a CWSI crop water deficit index layer. The micro-advection induced gradient layer aims to quantify the inducing effect of environmental potential energy on soil moisture movement at the microscale, calculating the micro-advection induced vector at any grid point (x, y) in space. The calculated micro-advection induced vectors are combined and mapped in space to construct a global micro-advection induced vector field, which is then used as a micro-advection induced gradient layer. The calculation formula is as follows: In the formula, Represented as the temperature gradient vector at grid point (x, y), Let α represent the soil moisture content gradient vector at grid point (x, y); α represents the thermal induction coefficient, and β represents the wet induction coefficient.

5. The adaptive dynamic grid agricultural precision irrigation method based on micro-adaptive flow induction according to claim 1, characterized in that: In step S3, based on the multi-dimensional visualized farmland state map, an improved quadtree adaptive subdivision and merging algorithm is used to construct a dynamic irrigation grid within the farmland area, dynamically dividing the farmland area into several dynamic grid units; the farmland area is used as the first initial grid unit, and the heterogeneity evaluation index of the current first initial grid unit is calculated in real time; when the heterogeneity evaluation index of the first initial grid unit exceeds a preset splitting threshold T... split When the time limit is reached, an improved quadtree adaptive subdivision and merging algorithm is used to divide the current first initial grid cell into four first sub-grid cells; all four first sub-grid cells are used as second initial grid cells, and the heterogeneity evaluation index of any second initial grid cell is calculated in real time; when the heterogeneity evaluation index of any second initial grid cell exceeds the preset splitting threshold T, the algorithm is called to divide the first initial grid cell into four sub-grid cells. split If the heterogeneity evaluation index of all initial grid cells does not exceed the preset splitting threshold T, then an improved quadtree adaptive subdivision and merging algorithm is used to divide the current second initial grid cell into four second sub-grid cells; this process continues until the heterogeneity evaluation index of all initial grid cells does not exceed the preset splitting threshold T. split This yields several dynamic mesh elements; In any initial grid cell, the heterogeneity evaluation index of the four sub-grid cells is combined with a preset merging threshold T. merge For comparison, when there are two or more sub-grid cells with heterogeneity evaluation indices below the merging threshold T... merge If the two or more sub-grid cells are then merged into a new dynamic grid cell, the result is a new dynamic grid cell.

6. The adaptive dynamic grid agricultural precision irrigation method based on micro-adaptive flow induction according to claim 5, characterized in that: The heterogeneity evaluation index of the initial grid cell is a weighted composite index of the variance of soil moisture content, the variance of crop water requirement, and the mean of micro-advection induction intensity within the initial grid cell. The calculation process is shown below: In the formula, H i Let be the heterogeneity evaluation index of the i-th grid cell. This is expressed as the spatial variance of soil moisture content within a grid cell. This is expressed as the spatial variance of crop evapotranspiration demand within a grid cell. It is expressed as the average magnitude of the micro-adaptation induced vector within the grid cell, where w1, w2, and w3 are the weight coefficients of each factor, and w1+w2+w3=1.

7. The adaptive dynamic grid agricultural precision irrigation method based on micro-adaptive flow induction according to claim 5, characterized in that: At regular intervals, the farmland area is dynamically divided into several dynamic grid cells. A time step adjustment model based on the rate of environmental change is constructed to adaptively update the execution time step, thus obtaining the time interval for the next division of dynamic grid cells. The calculation process is as follows: In the formula, T represents the time interval for the next dynamic mesh generation. base R represents the preset reference time step, λ represents the time sensitivity coefficient; change Expressed as the rate of change of environmental state, H i (k) and H i (k-1) represent the heterogeneity evaluation indices at the current time k and the previous time k-1, respectively, Δt k This represents the time interval at the current moment.

8. The adaptive dynamic grid agricultural precision irrigation method based on micro-adaptive flow induction according to claim 1, characterized in that: In step S5, a precision irrigation decision model based on the crop water stress index is established. The target irrigation amount for each dynamic grid cell is calculated based on the crop water stress index calculated for each dynamic grid cell. The formula for calculating the crop water stress index is as follows: In the formula, CWSI represents the crop water stress index, and T c T represents the measured canopy temperature. wet T represents the canopy temperature under conditions of adequate water supply. dry Represented as canopy temperature under complete water stress; The target irrigation amount for each dynamic grid cell is calculated based on the crop water stress index of each dynamic grid cell. The calculation process is as follows: In the formula, I i Let K represent the target irrigation amount for the i-th dynamic grid cell. c Represented as crop coefficient, ET0 represents reference evapotranspiration, A i Let f(CWSI) represent the area of ​​the i-th dynamic mesh cell. i ) represents the water stress correction function. P represents the duration of the current irrigation decision cycle. eff This is expressed as effective precipitation. This is represented by changes in soil water storage.

Images