Intelligent decision-making method for smart agriculture based on deep learning

Abstract

The application discloses a kind of intelligent decision-making methods of wisdom agriculture based on deep learning, it is related to agricultural expert system and intelligent control technical field, including the following steps: S1, generate digital farmland panorama hyper voxel map feature vector;S2, output critical state environmental feature vector;S3, apply non-equilibrium thermodynamics entropy flow-entropy generation coupling feature evolution mechanism, output crop self-organization mechanism feature vector;S4, output multi-source mechanism collaborative feature vector;S5, update multi-source mechanism collaborative feature vector;S6, generate crop growth evolution state matrix;S7, construct crop growth dynamics system state equation, calculate optimal agricultural decision control trajectory.The application overcomes the limitations of traditional data-driven methods, such as ignoring physical mechanisms, weak model generalization ability, and lack of theoretical basis for decision-making, providing a precise and reliable physical-data dual-driven solution for intelligent decision-making in wisdom agriculture.

Description

Technical Field

[0001] This invention relates to the field of agricultural expert systems and intelligent control technology, and in particular to a smart agriculture intelligent decision-making method based on deep learning. Background Technology

[0002] As the dimensions of data collection in smart agriculture continue to expand, crop growth monitoring systems face the severe challenge of deep fusion of heterogeneous multi-source data and analysis of physical mechanisms. Existing deep learning-based decision-making methods mainly rely on data-driven feature extraction models, processing UAV point clouds and soil profile data through convolutional neural networks. While these methods can acquire surface visual geometric features, this purely statistical learning-based feature representation ignores the implicit energy dissipation laws and material cycle dynamics mechanisms within the crop growth system. As a typical open system, the growth and evolution of crops is essentially a non-equilibrium thermodynamic process involving the exchange of matter and energy with the environment. Existing data-driven models struggle to construct the mapping relationship between thermodynamic forces and thermodynamic flows, and cannot quantify the degree to which the system deviates from equilibrium, thus limiting the model's generalization ability under complex environmental changes. Furthermore, existing technologies lack effective characterization of crop self-organized critical states, losing the physical driving factors behind changes in crop growth states, and making it difficult to establish constraint boundaries that conform to physical laws in the feature space, thereby reducing the interpretability and scientific validity of intelligent decision-making results.

[0003] Therefore, how to provide a smart decision-making method for agriculture based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes a deep learning-based intelligent decision-making method for smart agriculture. It utilizes a non-equilibrium thermodynamic entropy flow-entropy generation coupling feature evolution mechanism. Based on the critical state environmental feature vector and the crop visual feature vector, a system state vector is constructed. The deviation of this vector from the thermodynamic equilibrium reference vector is calculated to determine the thermodynamic force vector. Combined with the transport coefficient matrix, the mass and energy fluxes are solved, thereby deriving the entropy generation flow and entropy exchange rate. The negative entropy flow is calculated as an indicator of crop growth orderliness. The negative entropy flow value is fused with the entropy generation flow and entropy exchange rate values ​​to generate a crop self-organization mechanism feature vector, which is then used to construct a closed convex set parameter space as a physical boundary constraint. This mechanism maps the crop growth process to the evolution trajectory of the thermodynamic state space. By quantifying the entropy generation process within the system and the entropy exchange process with the external environment, it achieves a deep integration of data-driven features and thermodynamic physical mechanisms. This ensures that the output of the decision model conforms to the laws of energy dissipation and mass conservation, significantly improving the physical interpretability and scientific rigor of crop growth state assessment in complex environments. This invention overcomes the limitations of traditional data-driven methods, such as neglecting physical mechanisms, weak model generalization ability, and lack of theoretical basis for decision-making, and provides a precise and reliable physical-data dual-driven solution for intelligent decision-making in smart agriculture.

[0005] According to an embodiment of the present invention, a smart agriculture intelligent decision-making method based on deep learning specifically includes:

[0006] S1. Collect canopy point cloud and soil profile data from UAVs to construct a three-dimensional voxelized mesh. Calculate the three-dimensional special Euclidean group transformation matrix based on Lie algebra to construct an invariant feature descriptor. Aggregate and generate feature vectors of the digital farmland panoramic supervoxel map.

[0007] S2. The feature vectors of the panoramic hypervoxel image of digital farmland are processed by three-dimensional convolution and fractal dimension to extract the visual feature vectors of crops. Based on soil profile data, the evolution trajectory of environmental order parameters is calculated by applying Ginzburg-Landau free energy functional and the critical state environmental feature vector is output.

[0008] S3. Based on the critical state environment feature vector and crop visual feature vector, the non-equilibrium thermodynamic entropy flow-entropy generation coupling feature evolution mechanism is applied to calculate the entropy generation flow, entropy exchange rate and negative entropy flow, and output the crop self-organization mechanism feature vector.

[0009] S4. Based on the Cauchy-Schwarz inequality, construct the divergence loss terms of the visual feature vector, the critical state environment feature vector, and the crop self-organization mechanism feature vector. Perform iterative optimization with the goal of minimizing the divergence loss term value to output the multi-source mechanism collaborative feature vector.

[0010] S5. Introduce the maximum entropy principle into the multi-source mechanism collaborative feature vector, couple the data fitting term, mechanism consistency constraint term and information entropy regularization term, construct the maximum entropy-mechanism consistency constraint functional, and perform gradient backpropagation iterative optimization to update the multi-source mechanism collaborative feature vector;

[0011] S6. Construct a closed convex set parameter space based on the multi-source mechanism collaborative feature vector and the crop self-organization mechanism feature vector. Use the Euclidean projection operator to project the gradient calculated based on the maximum entropy-mechanism consistency constraint functional to the feasible region to obtain the crop growth evolution state matrix.

[0012] S7. Input the real-time multi-source mechanism collaborative feature vector into the state equation of the crop growth dynamics system constructed from the crop growth evolution state matrix, apply the Pontryagin maximum principle to solve the extreme value of the Hamiltonian function, and calculate the optimal agricultural decision control trajectory.

[0013] Optionally, S1 specifically includes:

[0014] S11. Use airborne lidar to collect crop canopy point cloud, use soil profile sensor to collect soil profile data, preprocess and coordinate register crop canopy point cloud and soil profile data, and construct a three-dimensional voxel grid integrating crop and soil.

[0015] S12. Perform super-voxel over-segmentation on the three-dimensional voxelized mesh, calculate the geometric center and principal curvature direction of the voxel point cloud within the super-voxel, and generate the three-dimensional special Euclidean group transformation matrix of the corresponding Lie group based on Lie algebra.

[0016] S13. Construct invariant feature descriptors with rotation invariance and translation invariance based on the three-dimensional special Euclidean group transformation matrix. Perform manifold aggregation on the invariant feature descriptors of adjacent supervoxels in the three-dimensional voxelized mesh to generate feature vectors of digital farmland panoramic supervoxel map.

[0017] Optionally, S2 specifically includes:

[0018] S21. Perform sliding window convolution operation on the feature vector of the digital farmland panoramic hypervoxel image, calculate the weighted sum of pixel values ​​within the window to obtain the local geometric texture feature vector; count the number of cover boxes on the point cloud surface at different measurement scales, calculate the linear regression slope between the logarithm of the measurement scale and the logarithm of the number of cover boxes, and use the slope value as the fractal dimension; concatenate the local geometric texture feature vector with the fractal dimension value to output the crop visual feature vector.

[0019] S22. Perform Kriging interpolation on soil profile data to construct a numerical matrix of soil physical properties distributed in three-dimensional space as a soil physical field model; set soil moisture content as an order parameter and construct an energy density function containing the order parameter gradient as a Ginzburg-Landau free energy functional; calculate the variational derivative of the Ginzburg-Landau free energy functional with respect to the order parameter, construct a time-dependent evolution equation containing a time derivative term using the variational derivative, solve the evolution equation using the finite difference method, and obtain the numerical sequence of the order parameter changing with time as the evolution trajectory.

[0020] S23. Calculate the first derivative value of the order parameter evolution trajectory, identify the points where the first derivative value is zero as critical points, and extract the order parameter value corresponding to the critical point as the critical point feature; calculate the second derivative value of the order parameter evolution trajectory, and take the sign of the second derivative value as the phase transition trend feature; concatenate the critical point feature and the phase transition trend feature to construct the critical state environment feature vector, and concatenate the crop visual feature vector and the critical state environment feature vector as the multimodal feature output.

[0021] Optionally, the evolution mechanism of the non-equilibrium thermodynamic entropy flow-entropy generation coupling characteristics specifically includes:

[0022] The elements of the critical state environment feature vector and the elements of the crop visual feature vector are concatenated sequentially to construct the system state vector; the difference between each element value in the system state vector and the corresponding element value in the preset thermodynamic equilibrium reference vector is calculated, and the vector composed of the differences is used as the thermodynamic force vector.

[0023] A diagonal matrix containing the light energy conversion coefficient and the mass transport coefficient is constructed as the transport coefficient matrix. The thermodynamic force vector is multiplied by the transport coefficient matrix to obtain the mass and energy flux vectors. The dot product of the mass and energy flux vectors and the thermodynamic force vector is calculated to obtain the scalar entropy generation flux value. The difference between the input flux vector and the output flux vector at the system boundary is calculated, and the dot product of the difference vector and the environmental gradient vector is performed to obtain the entropy exchange rate value.

[0024] The difference between the entropy exchange rate and the entropy generation flow is used to obtain the negative entropy flow value. The negative entropy flow value is used as an indicator to measure the orderliness of crop growth. This indicator value is concatenated with the entropy generation flow value and the entropy exchange rate value to generate a feature vector of crop self-organization mechanism.

[0025] Optionally, S4 specifically includes:

[0026] S41. Calculate the squared Euclidean distances between each pair of the visual feature vector, the critical state environment feature vector, and the crop self-organization mechanism feature vector, and construct an upper bound constraint function for the distance based on the Cauchy-Schwarz inequality.

[0027] S42. Construct a divergence loss term based on the distance upper bound constraint function, define a measure of the distribution difference of multi-source features in the manifold space, and set the minimization of the divergence loss term value as the optimization objective.

[0028] S43. Use the gradient descent algorithm to iteratively optimize the visual feature vector, critical state environment feature vector and crop self-organization mechanism feature vector until they converge to a preset threshold, and then output the multi-source mechanism collaborative feature vector.

[0029] Optionally, S5 specifically includes:

[0030] S51. Calculate the sum of squares of the differences between each element of the multi-source mechanism collaborative feature vector and the corresponding vector of the real observation data, and then divide it by the total number of elements to obtain the mean square error value as the data fitting term; calculate the sum of the product of the difference between the probability distribution histogram of the multi-source mechanism collaborative feature vector and the histogram of the preset prior mechanism distribution and the logarithmic product to obtain the KL divergence value as the mechanism consistency constraint term.

[0031] S52. Calculate the probability value of each element in the multi-source mechanism collaborative feature vector, take the logarithm of the probability value and multiply it with the probability value, sum all the product results and take the negative value to obtain the information entropy regularization term.

[0032] S53. Using the preset Lagrange multiplier values ​​as weight coefficients, the values ​​of the data fitting term, the mechanism consistency constraint term, and the information entropy regularization term are multiplied by their respective weight coefficients and then summed to obtain the total loss function value as the maximum entropy-mechanism consistency constraint functional. The partial derivative of the total loss function value with respect to each element in the multi-source mechanism collaborative feature vector is calculated to obtain the gradient vector. The gradient vector is multiplied by the preset learning rate and the element values ​​of the multi-source mechanism collaborative feature vector are updated. This process is iterated until the total loss function value is less than the preset threshold.

[0033] Optionally, S6 specifically includes:

[0034] S61. Extract the negative entropy flow value from the crop self-organization mechanism feature vector as the physical boundary threshold, set the upper and lower limits of the parameter values, and construct a set of linear inequality constraints for the range of values ​​of each element in the multi-source mechanism collaborative feature vector; define the polyhedral boundary of the parameter space based on the set of linear inequality constraints, cut off the spatial region that satisfies all constraints, and construct a closed convex set parameter space restricted by the physical boundary threshold.

[0035] S62. Calculate the partial derivative of the maximum entropy-mechanism consistency constraint functional with respect to the multi-source mechanism cooperative feature vector to obtain the gradient descent vector; determine whether each element of the gradient descent vector is within the closed convex set parameter space. If it is outside the range, calculate the vertical distance from the element to the boundary of the closed convex set parameter space, correct the element value to the boundary value to realize the Euclidean projection operation, and generate the projected feasible gradient vector.

[0036] S63. Using the opposite direction of the projected feasible gradient vector as the search direction, calculate the product of the preset step size and the feasible gradient vector, move the current parameter point along the search direction by the product distance to obtain a new parameter point; determine whether the new parameter point satisfies the iteration convergence condition. If it does, stop the iteration and take the current new parameter point as the optimal parameter point. Reconstruct the optimal parameter point into a two-dimensional matrix by arranging rows and columns to generate a crop growth evolution state matrix that represents the crop growth evolution law.

[0037] Optionally, S7 specifically includes:

[0038] S71. Acquire real-time canopy point cloud and soil profile data from UAVs, and generate real-time multi-source mechanism collaborative feature vectors through three-dimensional voxelized mesh construction and invariant feature descriptor aggregation.

[0039] S72. The real-time multi-source mechanism synergistic feature vector is used as an input term and matrix multiplication is performed with the crop growth evolution state matrix to construct the state equation of the crop growth dynamics system. The state parameters in the crop growth evolution state matrix are set as system state variables, the external input environmental regulation parameters are set as control variables, and the auxiliary variables describing the rate of change of state variables are set as adjoint variables.

[0040] S73. Construct a Hamiltonian function expression by linearly weighting and summing the system state variables, control variables, and adjoint variables. Calculate the partial derivatives of the control variables in the Hamiltonian function expression and set the partial derivatives to zero to obtain the expression for the values ​​of the control variables. Set the derivative of the adjoint variables with respect to time as the negative of the partial derivative of the Hamiltonian function with respect to the state variables, construct the adjoint equation, and solve it numerically to obtain the evolution trajectory of the adjoint variables. Substitute the evolution trajectory of the adjoint variables into the expression for the values ​​of the control variables to calculate the numerical sequence of the control variables that optimizes the system performance index, which is taken as the optimal control trajectory.

[0041] The beneficial effects of this invention are:

[0042] (1) This invention achieves deep coupling and mathematical representation of crop phenotypic geometric features and environmental dynamics and physiological features by constructing feature vectors of digital farmland panoramic hypervoxel images and the Ginzburg-Landau phase transition evolution process. A feature descriptor with rotation and translation invariance is constructed using Lie algebra to generate a three-dimensional special Euclidean group transformation matrix. Combined with fractal dimension quantification of point cloud surface complexity, crop visual features are accurately extracted. The Ginzburg-Landau free energy functional is applied to soil profile data to solve for the order parameter evolution trajectory. A time-dependent evolution equation is constructed by calculating variational derivatives, accurately capturing environmental critical point features and phase transition trend features. This mechanism elevates discrete observation data to a continuous dynamic system state description, effectively filtering out measurement noise and revealing the physical phase transition laws hidden behind the data, providing a solid physical foundation for multimodal feature fusion.

[0043] (2) This invention establishes a physical mechanism-driven feature constraint optimization and parameter space optimization system by constructing a non-equilibrium thermodynamic entropy flow-entropy generation coupled feature evolution mechanism. Based on the critical state environment feature vector and the crop visual feature vector, a system state vector is constructed. Its deviation from the thermodynamic equilibrium reference vector is calculated to determine the thermodynamic force vector. The transport coefficient matrix is ​​used to solve for the mass and energy fluxes, thereby deriving the entropy generation flow and entropy exchange rate. The difference between the two is used to construct the negative entropy flow as a core indicator for measuring the orderliness of crop growth. This mechanism strictly follows the fundamental laws of non-equilibrium thermodynamics, transforming the abstract growth state into calculable entropy change dynamics, achieving a deep integration of data-driven features and thermodynamic mechanisms, and providing a quantitative basis that conforms to the physical essence for subsequent decision-making.

[0044] (3) This invention constructs a crop growth dynamics system and applies the Pontryagin maximum principle to achieve a mathematical closed-loop solution from state perception to optimal control decision-making. The real-time multi-source mechanism synergistic feature vector is mapped to the state space constructed from the crop growth evolution state matrix. System state variables, control variables, and adjoint variables are defined, and the adjoint equation and control equation are derived using the Hamiltonian function extremum condition. This mechanism abandons the traditional data-fitting prediction model and instead uses variational methods to solve for the optimal control variable sequence with optimal performance indicators, directly outputting the optimal control trajectory that conforms to the dynamic evolution law. This achieves precise quantitative regulation of the crop growth process and provides rigorous mathematical and physical theoretical support for smart agriculture decision-making. Attached Figure Description

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0046] Figure 1This is an overall flowchart of a deep learning-based intelligent decision-making method for smart agriculture proposed in this invention.

[0047] Figure 2 This is a flowchart illustrating the working principle of the non-equilibrium thermodynamic entropy flow-entropy generation coupling feature evolution mechanism of a deep learning-based intelligent agricultural decision-making method proposed in this invention. Detailed Implementation

[0048] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0049] refer to Figure 1 and Figure 2 A smart decision-making method for smart agriculture based on deep learning, specifically including:

[0050] S1. Collect canopy point cloud and soil profile data from UAVs to construct a three-dimensional voxelized mesh. Calculate the three-dimensional special Euclidean group transformation matrix based on Lie algebra to construct an invariant feature descriptor. Aggregate and generate feature vectors of the digital farmland panoramic supervoxel map.

[0051] S2. The feature vectors of the panoramic hypervoxel image of digital farmland are processed by three-dimensional convolution and fractal dimension to extract the visual feature vectors of crops. Based on soil profile data, the evolution trajectory of environmental order parameters is calculated by applying Ginzburg-Landau free energy functional and the critical state environmental feature vector is output.

[0052] S3. Based on the critical state environment feature vector and crop visual feature vector, the non-equilibrium thermodynamic entropy flow-entropy generation coupling feature evolution mechanism is applied to calculate the entropy generation flow, entropy exchange rate and negative entropy flow, and output the crop self-organization mechanism feature vector.

[0053] S4. Based on the Cauchy-Schwarz inequality, construct the divergence loss terms of the visual feature vector, the critical state environment feature vector, and the crop self-organization mechanism feature vector. Perform iterative optimization with the goal of minimizing the divergence loss term value to output the multi-source mechanism collaborative feature vector.

[0054] S5. Introduce the maximum entropy principle into the multi-source mechanism collaborative feature vector, couple the data fitting term, mechanism consistency constraint term and information entropy regularization term, construct the maximum entropy-mechanism consistency constraint functional, and perform gradient backpropagation iterative optimization to update the multi-source mechanism collaborative feature vector;

[0055] S6. Construct a closed convex set parameter space based on the multi-source mechanism collaborative feature vector and the crop self-organization mechanism feature vector. Use the Euclidean projection operator to project the gradient calculated based on the maximum entropy-mechanism consistency constraint functional to the feasible region to obtain the crop growth evolution state matrix.

[0056] S7. Input the real-time multi-source mechanism collaborative feature vector into the state equation of the crop growth dynamics system constructed from the crop growth evolution state matrix, apply the Pontryagin maximum principle to solve the extreme value of the Hamiltonian function, and calculate the optimal agricultural decision control trajectory.

[0057] In this embodiment, S1 specifically includes:

[0058] S11. Crop canopy point cloud is collected using airborne lidar, and soil profile data is collected using soil profile sensors. The crop canopy point cloud and soil profile data are preprocessed and coordinate registered to construct a three-dimensional voxel grid integrating crop and soil. Statistical filtering is performed on the crop canopy point cloud to calculate the average distance from each point to its 10 neighboring points and remove outlier noise with an average distance greater than 3 times the standard deviation. Smoothing interpolation is performed on the soil profile data to fill in missing values. Common feature points of the two types of data are extracted, rotation and translation transformation parameters are calculated and converted to a unified geodetic coordinate system. The voxel side length is set to 5 cm. The data points are traversed to calculate the voxel index. The geometric center of all points within a voxel is used as the voxel center coordinate. The average normal vector of points within a voxel is calculated as the voxel normal vector. The two are combined to generate a three-dimensional voxel grid integrating crop and soil.

[0059] S12. Perform super-voxel oversegmentation on the 3D voxelized mesh, calculate the geometric center and principal curvature direction of the voxel point cloud within the super-voxel, and generate the 3D special Euclidean group transformation matrix of the corresponding Lie group based on Lie algebra. Select seed points as initial cluster centers for every 20 voxels, calculate the difference in the normal vector angle between the voxel and the seed point and the Euclidean distance, set the weight of the normal vector angle difference to 0.6 and the weight of the Euclidean distance to 0.4, and use the weighted sum as the distance metric. Iteratively search for neighboring voxels and merge voxels with a distance metric less than 0.2 into the cluster, update the cluster centers and repeat the search until all voxels are assigned, completing the super-voxel oversegmentation. Calculate the average spatial coordinates of the voxel point cloud within the super-voxel as the geometric center, traverse all points within the super-voxel, calculate the deviation value of each point's coordinates from the geometric center, multiply the deviation value by its transpose vector to obtain a 3x3 covariance matrix, calculate the characteristic equation of the covariance matrix, solve for three eigenvalues ​​and sort them by value, and select the smallest value. The eigenvectors corresponding to the eigenvalues ​​are used as the principal curvature directions. A local coordinate system is constructed with the geometric center as the origin. The cosine of the angle between the principal curvature direction vector and the three coordinate axes of the local coordinate system is calculated. The cosine of the angle is multiplied by a set scaling factor to obtain the first three components of the rotation vector. The three coordinate values ​​of the geometric center are used as translation components. A 6-dimensional rotation vector is generated by combining these components. The magnitude of the principal curvature direction vector is calculated as the rotation angle. The principal curvature direction vector is divided by the rotation angle to obtain the unit rotation axis vector. A 3x3 antisymmetric matrix is ​​constructed based on the three components of the unit rotation axis vector. The sine of the rotation angle is calculated as the first weighting coefficient. The result of 1 minus the cosine of the rotation angle is calculated as the second weighting coefficient. The sum of the identity matrix, the antisymmetric matrix multiplied by the first weighting coefficient, and the square of the antisymmetric matrix multiplied by the second weighting coefficient is calculated to obtain a 3x3 rotation matrix. The rotation matrix and the translation vector are combined to construct a 4x4 three-dimensional special Euclidean group transformation matrix.

[0060] S13. Construct invariant feature descriptors with rotation and translation invariance based on the 3D special Euclidean group transformation matrix. Perform manifold aggregation on the invariant feature descriptors of adjacent supervoxels in the 3D voxelized mesh to generate feature vectors of the digital farmland panoramic supervoxel map. Specifically, extract the rotation and translation parts from the 3D special Euclidean group transformation matrix. Calculate the trace value of the rotation part matrix, add 1, and divide the result by 4 to obtain the real part of the quaternion. Calculate the difference combination of the diagonal elements of the rotation part matrix and divide it by 4 times the real part of the quaternion to obtain the imaginary part of the quaternion. The process involves combining the real and imaginary parts of a quaternion to form a quaternion representation. Each component of the quaternion is then concatenated with the three values ​​of the translation vector to construct an initial feature descriptor. The trace of the rotation matrix is ​​subtracted by 1, and the result is divided by 2 to obtain the cosine of the rotation angle. The inverse cosine function is used to calculate the rotation angle. Finally, each element of the rotation matrix is ​​subtracted from its transpose, and the result is divided by twice the sine of the rotation angle to obtain the rotation axis vector. Multiplying the rotation axis vector by the rotation angle value yields the three-dimensional rotation variable, which is then combined with the translation vector to obtain a six-dimensional Lie algebra vector as a geometric constraint term. The initial feature descriptor is concatenated with the geometric constraint term to generate an invariant feature descriptor. Super-voxels in the 3D voxelized mesh are traversed, and super-voxels sharing faces, edges, or vertices are identified as adjacent, and an adjacency graph is constructed. The arithmetic mean of the Lie algebra vectors in the invariant feature descriptors of adjacent super-voxels is calculated as the tangent space reference point. The difference vector between each super-voxel Lie algebra vector and the tangent space reference point is calculated, and the weight values ​​of the difference vectors are calculated using a Gaussian kernel function. The difference vectors are multiplied by the corresponding weight values ​​and summed in the neighborhood. The summation result is added to the tangent space reference point. The aggregated Lie algebra vector is obtained. The magnitude of the three-dimensional rotation variable is extracted as the rotation angle. The three-dimensional rotation variable is divided by the rotation angle to obtain the unit rotation axis vector. The antisymmetric matrix of the unit rotation axis vector is calculated. The rotation matrix is ​​calculated by adding the unit matrix, the antisymmetric matrix multiplied by the sine of the rotation angle, and the square of the antisymmetric matrix multiplied by 1 minus the cosine of the rotation angle. This results in the rotation matrix, which is combined with the aggregated translation vector to generate the aggregated three-dimensional special Euclidean group transformation matrix. The feature values ​​of this matrix are extracted to generate the feature vector of the digital farmland panoramic hypervoxel map.

[0061] In this embodiment, S2 specifically includes:

[0062] S21. Perform sliding window convolution operation on the feature vector of the digital farmland panoramic hypervoxel image. Set the convolution kernel size to 3x3. Traverse the feature vector matrix and calculate the sum of squared differences between the central element and the eight neighboring elements around the central element. Use this sum of squared differences as the local geometric texture feature value of the central element. Count the number of cover boxes on the point cloud surface at different measurement scales. Set the step size of the measurement scale to an integer power of 2. Calculate the number of cover boxes in the scale range from 1 to 64. Calculate the linear regression slope between the logarithm of the measurement scale and the logarithm of the number of cover boxes. Use the slope value as the fractal dimension. Concatenate the local geometric texture feature vector with the fractal dimension value to output the crop visual feature vector.

[0063] S22. Perform Kriging interpolation on the soil profile data. Search for sampling points within a 5-meter radius of the point to be interpolated as neighboring points. Calculate the Euclidean distance between each pair of neighboring points as the lag distance. Calculate half the square of the difference in attribute values ​​between neighboring points as the semivariogram value. Divide the lag distance into multiple distance groups at 1-meter intervals. Calculate the average lag distance and average semivariogram value for all point pairs within each distance group. Plot a scatter plot with the average lag distance as the x-axis and the average semivariogram value as the y-axis. Fit the scatter plot using the spherical model formula. The semivariogram value corresponding to a lag distance of 0 is taken as the nugget value. The value at which the lag distance increases to the point where the semivariogram value no longer increases and begins to plateau is taken as the sill value. The lag distance value corresponding to the semivariogram value reaching the sill value is taken as the range. Construct a numerical matrix of soil physical properties distributed in three-dimensional space as a soil physical field model. Set the soil moisture content value as the order parameter. Calculate the rate of change of the order parameter in the three coordinate axes of space and sum the squares. Calculate the summation result. Multiplying by the gradient coefficient 0.5 yields the gradient energy term. The sum of the square of the order parameter value multiplied by -1 and the fourth power of the order parameter value is calculated. Multiplying this sum by a coefficient of 0.25 yields the potential energy term. Adding the gradient energy term and the potential energy term constructs an energy density function containing the order parameter gradient, serving as the Ginzburg-Landau free energy functional. The derivative term is obtained by multiplying the cubic value of the order parameter by 4 and then subtracting the value multiplied by 2. The sum of the second derivatives of the order parameter in the three spatial directions is then multiplied. The diffusion term result is obtained by setting the gradient term coefficient to 0.5. The derivative term result is then subtracted from the diffusion term result to obtain the rate of change of the order parameter over time. A time-dependent evolution equation containing the time derivative term is constructed. The spatial grid spacing is set to 0.1 meters and the time step is 0.01 seconds. The evolution equation is solved using the finite difference method. The result of adding the current order parameter value to the rate of change value and multiplying by the time step of 0.01 seconds is used as the predicted value for the next time step. After 500 iterations, the numerical sequence of the order parameter change over time is obtained as the evolution trajectory.

[0064] S23. Calculate the first derivative value of the order parameter evolution trajectory. Subtract the previous value from the current value and divide the result by the time step of 0.01 seconds. Identify the points where the first derivative value is zero as critical points. Extract the order parameter value corresponding to the critical point as the critical point feature. Calculate the second derivative value of the order parameter evolution trajectory. Subtract the previous value from the first derivative value from the current value and divide the result by the time step of 0.01 seconds. Use the sign of the second derivative value as the phase transition trend feature. When the sign is positive, it is determined to be an upward trend; when the sign is negative, it is determined to be a downward trend. Concatenate the critical point feature and the phase transition trend feature to construct the critical state environment feature vector. Concatenate the crop visual feature vector and the critical state environment feature vector as the multimodal feature output.

[0065] In this embodiment, the evolution mechanism of the non-equilibrium thermodynamic entropy flow-entropy generation coupling characteristics specifically includes:

[0066] The critical state environment feature vector and the crop visual feature vector are concatenated end to end. The value of each element in the thermodynamic equilibrium reference vector is set to 0.5. The thermodynamic force vector is obtained by subtracting 0.5 from each element of the system state vector.

[0067] Set the light energy conversion coefficient to 0.8 and the mass transfer coefficient to 0.6. Construct a square matrix whose number of rows and columns are equal to the dimension of the thermodynamic force vector. Set the value of the element in the first row and first column of the square matrix to 0.8, the value of the element in the second row and second column to 0.6, and the value of the remaining elements to 0 to obtain the transfer coefficient matrix. Calculate the product of the thermodynamic force vector and the transfer coefficient matrix to obtain the mass and energy flux vector. Calculate the inner product of the mass and energy flux vector and the thermodynamic force vector as the entropy generation flow. Calculate the difference vector between the input flux vector and the output flux vector. Calculate the inner product of the difference vector and the environmental gradient vector as the entropy exchange rate.

[0068] The entropy exchange rate is calculated by subtracting the entropy generation flow from the entropy exchange rate to obtain the negative entropy flow. The three values ​​of negative entropy flow, entropy generation flow, and entropy exchange rate are concatenated and spliced ​​to generate the crop self-organization mechanism feature vector.

[0069] The non-equilibrium thermodynamic entropy flow-entropy generation coupling feature evolution mechanism proposed in this step is similar to the traditional multi-source feature fusion method in that both are based on feature vector concatenation operations to construct system state descriptions, and use vector operations and matrix multiplication to realize the transformation and mapping of feature space, and both output numerical feature vectors for subsequent analysis.

[0070] The difference lies in that this invention breaks away from the limitations of traditional methods that rely solely on statistical data characteristics for feature stacking or weighted fusion. Instead of directly outputting a concatenated vector, this invention adds a non-equilibrium thermodynamic mechanism modeling step, introducing a pre-defined thermodynamic equilibrium reference vector as the system's steady-state benchmark. In the feature evolution calculation step, a diagonal transport coefficient matrix containing light energy conversion coefficients and mass transport coefficients is used to transform the thermodynamic force vector representing system state deviations into a mass and energy flux vector, rather than a simple feature projection. Finally, in the feature generation step, the entropy generation flow is calculated based on the dot product of flux and force. The entropy exchange rate is obtained by combining the boundary flux difference and environmental gradient calculations, and then the difference between the two is calculated to obtain the negative entropy flow. The negative entropy flow, entropy generation flow, and entropy exchange rate are jointly output as the crop self-organization mechanism feature vector, rather than a single state vector.

[0071] The beneficial effects of this improvement are that by introducing thermodynamic equilibrium reference and non-equilibrium dissipative structure theory, the material and energy transfer mechanism in the crop growth process can be integrated into the feature representation, breaking the limitation of traditional methods that ignore the physical and biological processes behind the features. This achieves a deep physical mapping from visual and environmental appearance features to the orderliness and self-organization ability of crop growth. This design significantly enhances the model's ability to analyze the intrinsic driving force of crop growth state and can more accurately quantify the system evolution trend of crops in environmental interaction. The feature vectors constructed based on physical indicators such as negative entropy flow effectively improve the physical interpretability and discrimination accuracy of features in crop growth monitoring and stress resistance assessment.

[0072] In this embodiment, S4 specifically includes:

[0073] S41. Calculate the sum of squares of the differences between corresponding elements of the visual feature vector and the critical state environment feature vector to obtain the first distance value. Calculate the sum of squares of the differences between corresponding elements of the visual feature vector and the crop self-organization mechanism feature vector to obtain the second distance value. Calculate the sum of squares of the differences between corresponding elements of the critical state environment feature vector and the crop self-organization mechanism feature vector to obtain the third distance value. Calculate the sum of squares of each element of the three feature vectors and take the square root to obtain the magnitude value of each vector. Calculate the product of the magnitude values ​​of each pair of vectors as the upper bound threshold of the distance between the pair of vectors.

[0074] S42. Calculate the difference between the first distance value, the second distance value, and the third distance value minus the sum of the corresponding upper boundary threshold and 0.01. Set the difference less than 0 to 0. Summate the three processed values ​​to obtain the divergence loss term value.

[0075] S43. Set the learning rate to 0.01, calculate the partial derivative of the divergence loss term with respect to each feature vector element as the gradient value, and subtract 0.01 times the gradient value from the feature vector element value to complete one iteration update. When the change in the divergence loss term value is less than 0.001 for 5 consecutive iterations, concatenate the three converged feature vectors end to end to output the multi-source mechanism collaborative feature vector.

[0076] In this embodiment, S5 specifically includes:

[0077] S51. Calculate the square of the difference between each element of the multi-source mechanism collaborative feature vector and the corresponding vector of the real observation data. Summate all the squared values ​​to obtain the sum of squares. Divide the sum of squares by the total number of elements to obtain the mean square error value as the data fitting term. Construct a probability distribution histogram by statistically analyzing the frequency distribution of each numerical interval in the multi-source mechanism collaborative feature vector. Read the preset prior mechanism distribution histogram and calculate the probability difference between the two histograms in the same numerical interval. Calculate the prior probability logarithm. Multiply the probability difference by the prior probability logarithm. Summate the product results for all intervals to obtain the KL divergence value as the mechanism consistency constraint term.

[0078] S52. The frequency of occurrence of each element in the multi-source mechanism collaborative feature vector is used as the probability value. The logarithm of each probability value is calculated with the natural constant e as the base. The probability value is multiplied by the corresponding logarithm to obtain the product result. The product results of all elements are summed and the negative value is taken to obtain the information entropy regularization term value.

[0079] S53. Set the Lagrange multiplier values ​​to 0.5, 0.3, and 0.2 respectively. Multiply the data fitting term value by 0.5, the mechanism consistency constraint term value by 0.3, and the information entropy regularization term value by 0.2. Summate the three product results to obtain the total loss function value. Calculate the partial derivative of the total loss function value with respect to each element in the multi-source mechanism collaborative feature vector to obtain the gradient vector. Set the learning rate to 0.01. Multiply the gradient vector by 0.01 to obtain the update step size. Subtract the update step size from the element values ​​of the multi-source mechanism collaborative feature vector to update the parameters. Repeat the calculation of the total loss function value and parameter updates. Stop the iteration when the total loss function value is less than 0.001.

[0080] In this embodiment, S6 specifically includes:

[0081] S61. Extract the negative entropy flow value from the crop self-organization mechanism feature vector as the physical boundary threshold. Set the upper limit of the parameter value to 1.0 and the lower limit of the parameter value to -1.0. For each element in the multi-source mechanism collaborative feature vector, construct three linear inequalities: element value less than or equal to the physical boundary threshold, element value less than or equal to 1.0, and element value greater than or equal to -1.0. Summarize the linear inequalities of all elements to form a linear inequality constraint group. Select the region in the multidimensional space that simultaneously satisfies the above three inequality conditions, remove the spatial region outside the boundary, and retain all inequality intersection regions to construct a closed convex set parameter space restricted by the physical boundary threshold.

[0082] S62. Calculate the partial derivative values ​​of the maximum entropy-mechanism consistency constraint functional with respect to the multi-source mechanism collaborative feature vector. Combine the partial derivative values ​​of each element to obtain the gradient descent vector. Determine whether the value of each element of the gradient descent vector is within the closed convex set parameter space. If the element value is greater than 1.0, correct the element value to 1.0. If the element value is less than -1.0, correct the element value to -1.0. If the absolute value of the element value is greater than the physical boundary threshold and is between the upper and lower limits, correct it to the physical boundary threshold. Generate the projected feasible gradient vector through Euclidean projection operation.

[0083] S63. Invert each element of the projected feasible gradient vector to obtain the search direction vector. Set the preset step size to 0.1. Calculate the product of the search direction vector and 0.1 to obtain the displacement vector. Add the element values ​​of the current parameter point to the corresponding element values ​​of the displacement vector to obtain the new parameter point. Calculate the Euclidean distance between the new parameter point and the current parameter point. When the Euclidean distance is less than 0.001, it is determined that the iteration convergence condition is met and the iteration stops. Take the current new parameter point as the optimal parameter point. Count the total number of elements of the optimal parameter point. Calculate the square root of the total number of elements and round it up to obtain the matrix dimension value. Arrange the elements of the optimal parameter point in row and column order to reconstruct a two-dimensional matrix and generate a crop growth evolution state matrix that represents the crop growth evolution law.

[0084] The crop growth evolution state matrix construction process proposed in this step is similar to that of traditional gradient optimization algorithms in that both determine the search direction for parameter updates based on the gradient information of the objective function and use iterative updates to find the optimal parameter solution. Both also involve the determination of convergence conditions and the reconstruction of the parameter matrix.

[0085] The difference lies in that this invention breaks through the limitations of traditional methods in gradient descent within unconstrained or simply boundary-constrained spaces. Building upon the traditional model's direct use of gradient direction to update parameters, this invention adds a physical boundary constraint construction step. It introduces the negative entropy flow value from the crop self-organization mechanism feature vector as a physical boundary threshold, combining it with preset upper and lower limits to construct a set of linear inequality constraints to define the closed convex set parameter space. In the parameter update step, Euclidean projection is used to project gradient vectors exceeding the closed convex set's range to the boundary, correcting them into feasible gradient vectors, rather than directly using the original gradient. Finally, in the state matrix generation step, new parameter points are calculated based on the opposite direction of the feasible gradient vector and a preset step size. The converged optimal parameter points are reconstructed into a two-dimensional matrix form, rather than a single feature vector.

[0086] The beneficial effects of this improvement are that by introducing a negative entropy flow physical boundary threshold and Euclidean projection operation, the thermodynamic physical constraints of crop growth can be integrated into the parameter optimization process. This breaks the limitation of traditional methods that ignore physical mechanism constraints, resulting in parameter solutions violating crop growth laws, and achieves a leap from pure mathematical optimization to physical constraint optimization. This design significantly enhances the rationality and feasibility of the model parameters in a physical sense, and can more accurately ensure that the parameters are always within the effective growth state space through closed convex set constraints. The crop growth evolution state matrix constructed based on physical constraints effectively improves the physical interpretability of feature representation and enhances the robustness and accuracy of the system in complex growth environment evolution modeling.

[0087] In this embodiment, S7 specifically includes:

[0088] S71. Acquire real-time canopy point cloud data and soil profile data from UAVs. Set the voxel grid side length to 0.05 meters. Divide the point cloud data into a three-dimensional voxel grid based on the grid side length. Calculate the average coordinate of all points in each voxel grid as the feature point coordinate of that voxel. Calculate the variance of the angle between the normal vectors of all points in the grid as the geometric feature descriptor value. Concatenate and splice the water content and nutrient concentration values ​​in the soil profile data with the geometric feature descriptor value to generate a real-time multi-source mechanism collaborative feature vector.

[0089] S72. Arrange the values ​​of each element of the real-time multi-source mechanism synergistic feature vector into a state column vector. Read the crop growth evolution state matrix. Multiply each element of the state column vector with the element of the corresponding row in the crop growth evolution state matrix and sum them to calculate the state prediction vector for the next moment. Add the state prediction vector to the externally input environmental control parameter vector to obtain the calculation result of the state equation of the crop growth dynamics system. Set the state parameters in the crop growth evolution state matrix as system state variables. Set the externally input temperature, light intensity and irrigation values ​​as control variables. Set the auxiliary variables describing the rate of change of state variables as companion variables.

[0090] S73. Construct a Hamiltonian function expression that includes system performance indicators and dynamic constraints. Calculate the weighted sum of the system state variables and control variables as the performance indicator value. Calculate the product of the rate of change of the adjoint variable and the state variable as the dynamic constraint value. Add the performance indicator value and the dynamic constraint value to obtain the Hamiltonian function value. Calculate the partial derivative of the control variable in the Hamiltonian function value, set the partial derivative value to zero, and solve for the value expression of the control variable. Calculate the partial derivative of the Hamiltonian function value with respect to the state variable, take the negative value of the partial derivative value to obtain the rate of change of the adjoint variable, construct the adjoint equation, set the integration step size to 0.01 seconds, and calculate the rate of change of the adjoint variable at the current time multiplied by 0.01 seconds to obtain the slope K. 1. Calculate the rate of change of the accompanying variable at the current time plus 0.005 seconds, multiply by 0.01 seconds to obtain the slope K2. Calculate the rate of change of the accompanying variable at the current time plus 0.005 seconds, multiply by 0.01 seconds to obtain the slope K3. Calculate the rate of change of the accompanying variable at the current time plus 0.01 seconds, multiply by 0.01 seconds to obtain the slope K4. Add K1, twice K2, twice K3 and K4 and divide by 6 to obtain the weighted slope value. Subtract the weighted slope value from the accompanying variable value to obtain the accompanying variable value at the next time step. Iterate to obtain the evolution trajectory of the accompanying variable. Substitute the evolution trajectory value of the accompanying variable into the value expression of the control variable to calculate the sequence of control variable values ​​that optimizes the system performance index, which is taken as the optimal control trajectory.

[0091] The optimal control trajectory solution process proposed in this step is similar to that of traditional model predictive control methods in that both are based on describing the dynamic evolution process using the system state equation, using the principle of minimizing the objective function to find the optimal value, and relying on gradient information or variational principles to iteratively calculate the optimal control sequence.

[0092] The difference lies in that this invention breaks away from the limitations of traditional methods that rely solely on rolling time-domain optimization to obtain numerical solutions while neglecting the system's intrinsic physical mechanisms. Building upon traditional models that only utilize state error weighting to construct the objective function, this invention adds a step to construct the dynamic system's state equation. It performs matrix multiplication on the real-time multi-source mechanism synergistic feature vector and the crop growth evolution state matrix to construct a state equation that incorporates physical mechanism constraints, rather than relying solely on a data-driven black-box model. In the optimal control solution step, the adjoint equation is constructed using the negative relationship between the time derivative of the adjoint variable and the partial derivative of the Hamiltonian function. The Runge-Kutta method is used for numerical integration to solve for the evolution trajectory of the adjoint variable, rather than directly solving the discrete algebraic equation. Finally, in the control trajectory output step, the evolution trajectory of the adjoint variable is substituted into the expression for the control variable's value to calculate the numerical sequence of control variables that optimizes the system's performance indicators, rather than a single static control quantity.

[0093] The beneficial effects of this improvement lie in the fact that by constructing a state equation that incorporates the synergistic characteristics of multiple mechanisms and introducing an adjoint equation for solution, the physical evolution mechanism of crop growth and energy constraints can be integrated into the control decision-making process. This breaks through the limitations of traditional methods that ignore the deep dynamic characteristics of the system, leading to short-sighted control strategies, and achieves a leap from data-driven optimization to optimal control based on mechanistic models. This design significantly enhances the decision model's ability to perceive dynamic changes in crop growth and can more accurately capture the optimal change path of control variables through the extreme conditions of the Hamiltonian function. The optimal control trajectory calculated based on the evolution trajectory of the adjoint equation effectively improves the scientificity and accuracy of intelligent decision-making, and enhances the system's adaptability and decision-making efficiency in complex agricultural environment regulation tasks.

[0094] Example 1: To verify the feasibility of this invention in intelligent decision-making for smart agriculture, the method of this invention was applied to the intelligent greenhouse tomato planting management system of a modern agricultural demonstration zone of a provincial state-owned agricultural group (hereinafter referred to as "Farm P"). In traditional greenhouse environmental control systems, simple control logic based on threshold judgment or black-box model prediction based on historical data are usually used. These methods not only fail to accurately analyze the deep physical mechanisms of crop growth, but also cannot effectively integrate heterogeneous data of canopy three-dimensional morphology and soil profile, easily leading to problems such as delayed water and fertilizer regulation, resource waste, and low crop growth environment matching. To solve the above problems, Farm P decided to adopt the intelligent decision-making method for smart agriculture based on deep learning proposed in this invention.

[0095] During implementation, Farm P first used a lidar scanning system deployed on a drone to acquire 3D point cloud data of the tomato canopy. Simultaneously, it used a pluggable soil profile sensor to obtain soil moisture and nutrient concentration data. Through 3D voxelized mesh construction, 3D special Euclidean group transformation matrix calculation, and invariant feature descriptor aggregation, an aligned feature input containing both spatial geometric and deep environmental dimensions was constructed. Meanwhile, agronomic experts at Farm P labeled the collected multi-source data with crop growth status and calibrated optimal environmental parameters, serving as a benchmark for model training and decision evaluation.

[0096] Farm P extracted visual feature vectors containing canopy morphology and structure, and critical-state environmental feature vectors containing soil environmental conditions, using a Lie group hypervoxel image feature extraction network. It then calculated the self-organizing mechanism feature vector characterizing crop vitality using a non-equilibrium thermodynamic entropy flow-entropy generation coupling mechanism. Next, it constructed a multi-source heterogeneous data collaborative divergence loss term using the Cauchy-Schwarz inequality, and output a multi-source collaborative feature vector through iterative optimization. Subsequently, it introduced the maximum entropy principle to construct an intelligent decision-making maximum entropy-mechanism consistency constraint functional, updated the feature vectors through gradient backpropagation, and reconstructed the crop growth intelligent evolution state matrix from the calculation results of the closed convex set parameter space projection based on physical boundary threshold constraints.

[0097] In the core decision-making and control stage, this invention constructs the state equation of the crop growth dynamics system, processes multi-source collaborative feature vectors, applies the Pontryagin maximum principle to solve the extreme value of the Hamiltonian function, constructs the adjoint equation and performs numerical integration, calculates the numerical sequence of control variables that optimizes the system performance indicators, and generates the optimal agricultural decision-making and control trajectory that includes irrigation amount, fertilizer amount and environmental regulation parameters, thus realizing a closed-loop leap from perception to decision-making.

[0098] During implementation, the technical team at Farm P discovered that, compared to traditional threshold control and conventional data-driven models, the method of this invention significantly improves the accuracy and scientific rigor of crop growth regulation. Traditional methods cannot quantify the physical limits and energy dissipation of crop growth and have weak capabilities for fusing multi-source heterogeneous data. In contrast, the method of this invention, through thermodynamic mechanism coupling, physical constraint projection, and maximum principle solving, effectively achieves precise evolutionary deduction of crop growth status and formulation of optimal control strategies.

[0099] To further verify the actual performance of the method of the present invention, farm P conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1:

[0100] Table 1. Performance Comparison of Smart Decision-Making Methods in Farm P's Smart Agriculture

[0101] Accuracy rate of crop growth status prediction (%) 81.2 95.6 +14.4% Water and fertilizer resource utilization rate (%) 68.5 92.3 +23.8% Average yield per mu (kg) per crop 4200 4850 +15.5% Environmental control response lag time (minutes) 30 2 -93.3% Disease incidence rate (%) 8.5 1.2 -85.9% Average irrigation water consumption per mu (cubic meters) 185 125 -32.4% Average fertilizer application rate per mu (kg) 45 28 -37.8% Cost of manual inspection and decision-making (yuan / mu) 350 120 -65.7% Rate of high-quality fruit (%) 75.0 93.0 +18.0% Overall satisfaction with system decision-making (%) 79.0 96.0 +17.0%

[0102] As shown in Table 1, the performance of the intelligent agricultural decision-making system has been comprehensively improved after applying the method of this invention. The accuracy of crop growth status prediction has increased from 81.2% using traditional methods to 95.6%, and the water and fertilizer resource utilization rate has increased from 68.5% to 92.3%, significantly improving the level of refined management in agricultural production and providing a reliable basis for resource conservation. The response lag time for environmental control has been significantly shortened from 30 minutes to 2 minutes, significantly enhancing the system's timeliness. In addition, the disease incidence rate has decreased from 8.5% to 1.2%, and the average irrigation water consumption per mu has decreased from 185 cubic meters to 125 cubic meters, significantly reducing production costs and environmental impact. The overall satisfaction with the system's decision-making has also significantly improved, from 79.0% to 96.0%.

[0103] Through the method of this invention, farm P has successfully achieved precise perception and intelligent decision-making control of crop growth process, effectively improving the efficiency of water and fertilizer resource utilization and crop yield and quality, ensuring the efficient and safe operation of facility agriculture, greatly improving the level of intelligence and digitalization of agricultural production, significantly reducing the decision-making burden of managers, enhancing the stability and robustness of the production system, and providing strong technical support for the construction of smart agriculture.

[0104] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart decision-making method for smart agriculture based on deep learning, characterized in that, Includes the following steps: S1. Collect canopy point cloud and soil profile data from UAVs to construct a three-dimensional voxelized mesh. Calculate the three-dimensional special Euclidean group transformation matrix based on Lie algebra to construct an invariant feature descriptor. Aggregate and generate feature vectors of the digital farmland panoramic supervoxel map. S2. The feature vectors of the panoramic hypervoxel image of digital farmland are processed by three-dimensional convolution and fractal dimension to extract the visual feature vectors of crops. Based on soil profile data, the evolution trajectory of environmental order parameters is calculated by applying Ginzburg-Landau free energy functional and the critical state environmental feature vector is output. S3. Based on the critical state environment feature vector and crop visual feature vector, the non-equilibrium thermodynamic entropy flow-entropy generation coupling feature evolution mechanism is applied to calculate the entropy generation flow, entropy exchange rate and negative entropy flow, and output the crop self-organization mechanism feature vector. S4. Based on the Cauchy-Schwarz inequality, construct the divergence loss terms of the visual feature vector, the critical state environment feature vector, and the crop self-organization mechanism feature vector. Perform iterative optimization with the goal of minimizing the divergence loss term value to output the multi-source mechanism collaborative feature vector. S5. Introduce the maximum entropy principle into the multi-source mechanism collaborative feature vector, couple the data fitting term, mechanism consistency constraint term and information entropy regularization term, construct the maximum entropy-mechanism consistency constraint functional, and perform gradient backpropagation iterative optimization to update the multi-source mechanism collaborative feature vector; S6. Construct a closed convex set parameter space based on the multi-source mechanism collaborative feature vector and the crop self-organization mechanism feature vector. Use the Euclidean projection operator to project the gradient calculated based on the maximum entropy-mechanism consistency constraint functional to the feasible region to obtain the crop growth evolution state matrix. S7. Input the real-time multi-source mechanism collaborative feature vector into the state equation of the crop growth dynamics system constructed from the crop growth evolution state matrix, apply the Pontryagin maximum principle to solve the extreme value of the Hamiltonian function, and calculate the optimal agricultural decision control trajectory.

2. The intelligent decision-making method for smart agriculture based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Use airborne lidar to collect crop canopy point cloud, use soil profile sensor to collect soil profile data, preprocess and coordinate register crop canopy point cloud and soil profile data, and construct a three-dimensional voxel grid integrating crop and soil. S12. Perform super-voxel over-segmentation on the three-dimensional voxelized mesh, calculate the geometric center and principal curvature direction of the voxel point cloud within the super-voxel, and generate the three-dimensional special Euclidean group transformation matrix of the corresponding Lie group based on Lie algebra. S13. Construct invariant feature descriptors with rotation invariance and translation invariance based on the three-dimensional special Euclidean group transformation matrix. Perform manifold aggregation on the invariant feature descriptors of adjacent supervoxels in the three-dimensional voxelized mesh to generate feature vectors of digital farmland panoramic supervoxel map.

3. The intelligent decision-making method for smart agriculture based on deep learning according to claim 1, characterized in that, S2 specifically includes: S21. Perform sliding window convolution operation on the feature vector of the digital farmland panoramic hypervoxel image, calculate the weighted sum of pixel values ​​within the window to obtain the local geometric texture feature vector; count the number of cover boxes on the point cloud surface at different measurement scales, calculate the linear regression slope between the logarithm of the measurement scale and the logarithm of the number of cover boxes, and use the slope value as the fractal dimension; concatenate the local geometric texture feature vector with the fractal dimension value to output the crop visual feature vector. S22. Perform Kriging interpolation on soil profile data to construct a numerical matrix of soil physical properties distributed in three-dimensional space as a soil physical field model; set soil moisture content as an order parameter and construct an energy density function containing the order parameter gradient as a Ginzburg-Landau free energy functional; calculate the variational derivative of the Ginzburg-Landau free energy functional with respect to the order parameter, construct a time-dependent evolution equation containing a time derivative term using the variational derivative, solve the evolution equation using the finite difference method, and obtain the numerical sequence of the order parameter changing with time as the evolution trajectory. S23. Calculate the first derivative value of the order parameter evolution trajectory, identify the points where the first derivative value is zero as critical points, and extract the order parameter value corresponding to the critical point as the critical point feature; calculate the second derivative value of the order parameter evolution trajectory, and take the sign of the second derivative value as the phase transition trend feature; concatenate the critical point feature and the phase transition trend feature to construct the critical state environment feature vector, and concatenate the crop visual feature vector and the critical state environment feature vector as the multimodal feature output.

4. The intelligent decision-making method for smart agriculture based on deep learning according to claim 1, characterized in that, The specific evolution mechanism of the non-equilibrium thermodynamic entropy flow-entropy generation coupling characteristics includes: The elements of the critical state environment feature vector and the elements of the crop visual feature vector are concatenated sequentially to construct the system state vector; the difference between each element value in the system state vector and the corresponding element value in the preset thermodynamic equilibrium reference vector is calculated, and the vector composed of the differences is used as the thermodynamic force vector. A diagonal matrix containing the light energy conversion coefficient and the mass transport coefficient is constructed as the transport coefficient matrix. The thermodynamic force vector is multiplied by the transport coefficient matrix to obtain the mass and energy flux vectors. The dot product of the mass and energy flux vectors and the thermodynamic force vector is calculated to obtain the scalar entropy generation flux value. The difference between the input flux vector and the output flux vector at the system boundary is calculated, and the dot product of the difference vector and the environmental gradient vector is performed to obtain the entropy exchange rate value. The difference between the entropy exchange rate and the entropy generation flow is used to obtain the negative entropy flow value. The negative entropy flow value is used as an indicator to measure the orderliness of crop growth. This indicator value is concatenated with the entropy generation flow value and the entropy exchange rate value to generate a feature vector of crop self-organization mechanism.

5. The intelligent decision-making method for smart agriculture based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. Calculate the squared Euclidean distances between each pair of the visual feature vector, the critical state environment feature vector, and the crop self-organization mechanism feature vector, and construct an upper bound constraint function for the distance based on the Cauchy-Schwarz inequality. S42. Construct a divergence loss term based on the distance upper bound constraint function, define a measure of the distribution difference of multi-source features in the manifold space, and set the minimization of the divergence loss term value as the optimization objective. S43. Use the gradient descent algorithm to iteratively optimize the visual feature vector, critical state environment feature vector and crop self-organization mechanism feature vector until they converge to a preset threshold, and then output the multi-source mechanism collaborative feature vector.

6. The intelligent decision-making method for smart agriculture based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Calculate the sum of squares of the differences between each element of the multi-source mechanism collaborative feature vector and the corresponding vector of the real observation data, and then divide it by the total number of elements to obtain the mean square error value as the data fitting term; calculate the sum of the product of the difference between the probability distribution histogram of the multi-source mechanism collaborative feature vector and the histogram of the preset prior mechanism distribution and the logarithmic product to obtain the KL divergence value as the mechanism consistency constraint term. S52. Calculate the probability value of each element in the multi-source mechanism collaborative feature vector, take the logarithm of the probability value and multiply it with the probability value, sum all the product results and take the negative value to obtain the information entropy regularization term. S53. Using the preset Lagrange multiplier values ​​as weight coefficients, the values ​​of the data fitting term, the mechanism consistency constraint term, and the information entropy regularization term are multiplied by their respective weight coefficients and then summed to obtain the total loss function value as the maximum entropy-mechanism consistency constraint functional. The partial derivative of the total loss function value with respect to each element in the multi-source mechanism collaborative feature vector is calculated to obtain the gradient vector. The gradient vector is multiplied by the preset learning rate and the element values ​​of the multi-source mechanism collaborative feature vector are updated. This process is iterated until the total loss function value is less than the preset threshold.

7. The intelligent decision-making method for smart agriculture based on deep learning according to claim 1, characterized in that, S6 specifically includes: S61. Extract the negative entropy flow value from the crop self-organization mechanism feature vector as the physical boundary threshold, set the upper and lower limits of the parameter values, and construct a set of linear inequality constraints for the range of values ​​of each element in the multi-source mechanism collaborative feature vector; define the polyhedral boundary of the parameter space based on the set of linear inequality constraints, cut off the spatial region that satisfies all constraints, and construct a closed convex set parameter space restricted by the physical boundary threshold. S62. Calculate the partial derivative of the maximum entropy-mechanism consistency constraint functional with respect to the multi-source mechanism cooperative feature vector to obtain the gradient descent vector; determine whether each element of the gradient descent vector is within the closed convex set parameter space. If it is outside the range, calculate the vertical distance from the element to the boundary of the closed convex set parameter space, correct the element value to the boundary value to realize the Euclidean projection operation, and generate the projected feasible gradient vector. S63. Using the opposite direction of the projected feasible gradient vector as the search direction, calculate the product of the preset step size and the feasible gradient vector, move the current parameter point along the search direction by the product distance to obtain a new parameter point; determine whether the new parameter point satisfies the iteration convergence condition. If it does, stop the iteration and take the current new parameter point as the optimal parameter point. Reconstruct the optimal parameter point into a two-dimensional matrix by arranging rows and columns to generate a crop growth evolution state matrix that represents the crop growth evolution law.

8. The intelligent decision-making method for smart agriculture based on deep learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Acquire real-time canopy point cloud and soil profile data from UAVs, and generate real-time multi-source mechanism collaborative feature vectors through three-dimensional voxelized mesh construction and invariant feature descriptor aggregation. S72. The real-time multi-source mechanism synergistic feature vector is used as an input term and matrix multiplication is performed with the crop growth evolution state matrix to construct the state equation of the crop growth dynamics system. The state parameters in the crop growth evolution state matrix are set as system state variables, the external input environmental regulation parameters are set as control variables, and the auxiliary variables describing the rate of change of state variables are set as adjoint variables. S73. Construct a Hamiltonian function expression by linearly weighting and summing the system state variables, control variables, and adjoint variables. Calculate the partial derivatives of the control variables in the Hamiltonian function expression and set the partial derivatives to zero to obtain the expression for the values ​​of the control variables. Set the derivative of the adjoint variables with respect to time as the negative of the partial derivative of the Hamiltonian function with respect to the state variables, construct the adjoint equation, and solve it numerically to obtain the evolution trajectory of the adjoint variables. Substitute the evolution trajectory of the adjoint variables into the expression for the values ​​of the control variables to calculate the numerical sequence of the control variables that optimizes the system performance indicators, which serves as the optimal agricultural decision control trajectory.

Images