Artificial intelligence-guided fertilizer reduction data model based on crop planting area differentiation

Abstract

The present invention discloses an artificial intelligence-guided fertilizer reduction data model based on crop planting area differentiation. Based on the massive planting data accumulated by soil testing and formula fertilization, the model introduces an artificial intelligence platform to conduct data-based intelligent analysis and iterative optimization processing of crop zoning planting and / or fertilizer reduction, and constructs the existing simple tabular data format into a guided data model based on data algorithms; based on the data-based construction of crop planting zoning, it is compatible with the further introduction of different types of fertilizer reduction and / or crop yield prediction progressive models, thereby achieving global fit for the introduction of artificial intelligence data optimization platform at the data structure level. The present invention has carried out joint development of planting + data, laying a foundation for the subsequent artificial intelligence iterative optimization of wheat zoning fertilizer control and reduction problems.

Description

Technical Field

[0001] The present invention relates to the field of agricultural informatization and data intelligence technology, and in particular to an artificial intelligence-oriented data model for wheat weight loss analysis and optimization in different regions, as well as its construction and application. Background Art

[0002] The previous research of the research group of the present invention used the test data of the "Nitrogen, Phosphorus and Potassium 3414" project of soil testing and fertilization formula in Hebei Province to deeply analyze the appropriate amount of wheat nutrients in different regions of Hebei Province, explore the potential for saving fertilizer in different regions of wheat, and provide technical support for reducing fertilizer application in Hebei Province. However, the above research is still based on data analysis based on manual experience in the processing of data information, which is relatively traditional and extensive.

[0003] In recent years, artificial intelligence has developed rapidly and vigorously, and the topic of fertilizer control and reduction for wheat and other crops happens to involve a large amount of data analysis and optimization. Therefore, combining artificial intelligence data iterative optimization technology with the existing zoning planting data and the massive data accumulated from soil testing and formula fertilization has become a valuable and very interesting new research direction.

[0004] Based on this, we have carried out joint development of planting + data with sister colleges and universities, and conducted pioneering and basic research in this new technical direction. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome various deficiencies of the prior art and provide a method for optimizing the weight loss potential of wheat by region.

[0006] In order to solve the above technical problems, the technical solutions adopted by the present invention are as follows.

[0007] The artificial intelligence-guided fertilizer reduction data model based on crop planting area differentiation is based on the massive planting data accumulated from soil testing and formula fertilization. It is guided by the introduction of an artificial intelligence platform to conduct data-based intelligent analysis and iterative optimization processing of crop zoning planting and / or fertilizer reduction, and the existing simple tabular data format is constructed into a guided data model based on data algorithms. Based on the data-based construction of crop planting zoning, it is compatible with the further introduction of different types of fertilizer reduction and / or crop yield prediction progressive models, thereby achieving global fit with the introduction of the artificial intelligence data optimization platform at the data structure level.

[0008] As a preferred technical solution of the present invention, for the digitization of crop planting zoning, multiple groups of optional digitization models are constructed to adapt to a variety of specific numerical analysis and data processing tasks; the multiple groups of optional digitization models include at least the following two groups of specific optional models: a discrete numerical zoning model and a functional fitting zoning model.

[0009] As a preferred technical solution of the present invention, the main data architecture of the discretized numerical partitioning model is: constructing a land similarity matrix based on all measurement values ​​of soil data and its associated geographic or climate data, and deriving components of the similarity matrix using the Euclidean distance or other similarity metrics of the soil and its associated data to represent the similarity between different land areas; and performing effectiveness division of crop planting areas based on this data architecture.

[0010] As a preferred technical solution of the present invention, the main data architecture of the functional fitting partitioning model is: for the multidimensional attributes and multivariate attributes of soil data and its related data in space and time, multivariate spatial statistical equations are used to fit the spatial continuity and heterogeneity of crop-planting land data and its related data; and the effectiveness division of crop-planting areas is carried out on this data architecture.

[0011] As a preferred technical solution of the present invention, the construction of the discretized numerical partition model specifically includes:

[0012] To construct a land similarity matrix, we first create a characteristic data description for each piece of land that includes soil geographic and climate data. This description is specifically represented as a data set or data vector, which contains multiple indicators of each piece of land, including but not limited to soil pH, organic matter content, texture, macro-element levels, medium and trace element levels, water content, and temperature, precipitation, sunshine hours, wind speed, wind direction and relative humidity under climatic conditions, as well as other related data added subsequently. These indicators are key factors in evaluating land characteristics, and together they constitute the unique fingerprint of each piece of land.

[0013] In particular, after the above-mentioned characteristic data description data set or data vector is created, the kernel function mapping data structure is further introduced as an adaptive mathematical structure on this basis, thereby converting the original data set or data vector characteristic description into a new space, and setting the upper limit of the possible dimension of this space as an open upper limit, so that it has a sufficiently high data analysis capacity; at this time, based on the kernel function mapping mathematical tool, a set of mathematical expressions are used to characterize the data conversion corresponding to the kernel function mapping, and a parameter variable is implicitly constructed in the above-mentioned mathematical expression, and based on this parameter variable, their similarity measurement is adjusted according to the difference between the characteristic descriptions of the two pieces of land. The above-mentioned implicitly constructed parameter variable is constructed as the rate at which the similarity of the corresponding land area changes with the characteristic difference; after completing the above-mentioned mapping conversion, the data structure is transformed into a new data space, and the similarity between the lands reveals a nonlinear pattern that can be captured, thereby expanding based on the kernel function mapping to obtain more comprehensive validity data, and at this time according to the following data connotation: each component of the similarity matrix is ​​constructed to characterize the degree of similarity between the two pieces of land in the new space, and data output is performed to obtain a similarity matrix that characterizes the degree of association between the lands;

[0014] Further data optimization: A. In order to ensure the stability and reliability of the land similarity matrix, the above data output is normalized. This process involves adjusting the values ​​of the land similarity matrix so that the sum of the values ​​of each row and column in the matrix is ​​1, so as to eliminate the possible scale differences between the sample points and make the matrix more balanced and standardized; the mathematical operation method corresponding to this data processing is: normalization based on the diagonal matrix so that the elements on each diagonal of the diagonal matrix are the sum of the values ​​of the corresponding rows in the similarity matrix; B. Simultaneously simplify and optimize the data structure and reduce the calculation process, and further transform the normalized similarity matrix. Since only relatively large similarity values ​​are important in actual data processing and those smaller values ​​should be ignored, this conversion is completed by setting a threshold. Only when the value in the matrix is ​​greater than this threshold will the value be retained, otherwise it will be set to 0; thereby reducing unnecessary calculations and retaining valuable key data information; because in actual applications, only relatively large similarity values ​​are important, and those smaller values ​​can usually be ignored. In this way, we can preserve key similarity information while drastically reducing computational complexity and resource consumption.

[0015] Finally, a clustering algorithm is constructed based on the above-mentioned optimized land similarity matrix. The optimized similarity matrix is ​​used as input to divide the land into different planting areas with real differences in global geography, climate and other related data. Specifically, the following data process is introduced to handle complex data structures and reveal subtle clustering patterns in the data: First, we construct and create a graphical representation according to the following rules to show the similarity between lands: In the construction of this graph, each piece of land is regarded as a node, and the similarity between lands is represented by the edges connecting these nodes; then, these edges are assigned a weight or correlation data strength, which is given by the above-mentioned land similarity matrix; further, a numerical representation of the correlation of each node is constructed based on the sum of the weights or correlation data strengths of all edges connected to each node, thereby obtaining a torsion matrix, which records The data resonance of each node in the graph is recorded; at this time, consider introducing the Laplace method, and obtain a new Laplace representation matrix by subtracting the similarity matrix from the twisted matrix, which linearly represents the structural characteristics of the initially constructed graph and its clustering characteristics. At this time, the eigenvalues ​​and eigenvectors of the obtained Laplace representation matrix can be extracted; among them, the eigenvalue represents the information of how the associated graph is divided into clusters, and the corresponding eigenvector represents the shape of each associated cluster; the next step is to extract the eigenvectors corresponding to the smallest several non-zero eigenvalues, and use these eigenvectors to create a new data representation for each piece of land. Based on this data representation, the similarity between the lands is reflected, so that the existing standardized clustering algorithms such as K-means or other standard clustering algorithm tools can be directly quoted or linked to divide the land into several areas, and the land in each area has similar characteristics.

[0016] As a preferred technical solution of the present invention, the last section of the data model maps the clustering results back to the original land to obtain standardized partitions, that is, dividing the land into different planting areas with authentic differences in global geography, climate and other related data. It not only includes the planting area differences visible in the original feature data set or data vector, but the above data processing process also expands and displays the inherent similarities in the original feature data.

[0017] As a preferred technical solution of the present invention, after the discretized numerical partitioning model is initially constructed according to the above steps, an artificial intelligence platform is introduced as needed based on the existing accumulated data volume for supervised iterative optimization; it has the following data optimization potential in terms of data process: it not only targets the multidimensional attributes and multivariate attributes of soil data and its related data in space and time, but also is compatible with the spatial distribution of land characteristics and the continuity of land characteristics; the above data optimization potential is derived through the subsequent introduction and iterative optimization of the artificial intelligence platform.

[0018] As a preferred technical solution of the present invention, the construction of the functional fitting partition model is mainly based on the model construction direction dominated by the introduction of spatial statistics and differential equation mathematical tools, specifically including:

[0019] Firstly, considering the optimal distribution prediction of spatial data and the spatial correlation between data points, the collected soil and climate data were spatially interpolated using interpolation methods to estimate the continuous distribution of these variables in the entire region.

[0020] Next, a land property function is constructed that comprehensively covers all relevant soil, climate, geography and other related variables. Simple linear combination sub-terms are set in this function, while sub-terms containing complex nonlinear terms are set to reflect the interaction between different variables. Each component of this function is constructed based on Kriging interpolation. In particular, a weight is assigned to each variable to correspond to the size of its measurable influence on land properties.

[0021] Furthermore, in order to ensure that the zoning model can reflect the spatial continuity of land characteristics, a partial differential equation is constructed to further process the land characteristic function; specifically, given that the zoning is guided by fertilizer reduction and the diffusion properties of fertilizer and soil, the Allen-Cahn equation is considered to simulate the separation and pattern formation between cultivated lands, and the parameters of the equation are superimposed and adjusted to characterize the changes in land characteristics over time and space and the formation of boundaries between different land characteristics; at this time, the numerical method can be directly used to solve the partial differential equation, and its stable solution can characterize the final pattern of land characteristics; on this basis, the regional boundaries are delineated. Specifically, two optional The data paradigm is as follows: ① By setting a specific threshold, the boundaries of different regions are determined; ② The level set data paradigm is introduced to determine the boundaries of different regions; finally, on the basis of the above regional boundaries, clustering algorithms are further introduced to cluster these land characteristic patterns, thereby constructing a partition model defined or represented by precise quantitative data. From the perspective of data process, it has the following data optimization potential: not only for the multidimensional and multivariate attributes of soil data and its related data in space and time, but also for the spatial distribution and continuity of land characteristics; the above data optimization potential is derived through the introduction and iterative optimization of the subsequent artificial intelligence platform.

[0022] As a preferred technical solution of the present invention, regarding the subsequent different types of fertilizer reduction and / or crop yield prediction progressive models, the fertilizer reduction related model is constructed using a model based on survey data or guidance data, or is newly characterized and constructed using physical parameters based on fertilizer soil science; the crop yield prediction progressive model focuses on adaptive construction based on the iterative nature of artificial intelligence.

[0023] As a preferred technical solution of the present invention, the crop is wheat or other conventionally planted crops.

[0024] The beneficial effect of adopting the above technical scheme is that: the present invention carries out joint development of planting + data, and is guided by the introduction of an artificial intelligence platform to perform data-based intelligent analysis and iterative optimization processing on crop zoning planting and / or fertilizer reduction. The traditional gridded data format is constructed into a directional data model based on data algorithms, and an attempt is made to achieve global fit for the introduction of the artificial intelligence data optimization platform at the data structure level, thereby laying the foundation for the subsequent artificial intelligence iterative optimization of the wheat zoning fertilizer control and reduction problems, and also providing valuable basic scientific research references for this new technical subject direction. DETAILED DESCRIPTION

[0025] Example 1: Data architecture and technical route of discretized numerical partitioning model

[0026] The main data architecture of the discretized numerical zoning model is: construct a land similarity matrix based on all measurements of soil data and its associated geographic or climate data, and derive the components of the similarity matrix using the Euclidean distance or other similarity metrics of the soil and its associated data to represent the similarity between different land areas; and perform effective division of crop planting areas based on this data architecture.

[0027] First consider the construction of the land similarity matrix.

[0028] Generally speaking, the initial classification of existing planting data is as follows: a. Soil data: pH value, organic matter content, soil texture (clay, loam, sand, etc.), macronutrient content (such as N, P, K, etc.) and trace element content, soil moisture content, etc.; b. Climate and geographical data, temperature, precipitation, sunshine duration, wind speed and direction, relative humidity, etc.; c. Other related data: c-1. Wheat variety data, growth cycle, disease and pest resistance, yield, quality (such as protein content), etc.; c-2. Fertilizer data, fertilizer type (organic fertilizer, chemical fertilizer, etc.), fertilizer composition, fertilization time, fertilizer amount, etc.; c-3. Historical yield data, wheat yield and its related data, etc.; d. Other data or subsequent data added as needed.

[0029] These data arrangements, including classification rules, classification standards, data types and data volumes, can be adjusted and added or deleted as needed.

[0030] Based on the existing planting data and its initial classification, a characteristic data description containing soil geography and climate data is first created for each piece of land, which is specifically represented as a data set or data vector, which contains multiple indicators of each piece of land, including but not limited to soil pH, organic matter content, texture, macro-element level, medium and trace element level, water content, and temperature, precipitation, sunshine hours, wind speed, wind direction and relative humidity under climatic conditions, and other related data added later. The indicators of these constructed data sets or data vectors are digitized factors for evaluating land characteristics and together constitute the unique fingerprint of each piece of land.

[0031] In particular, after the above-mentioned feature data description data set or data vector is created, the kernel function mapping data structure is further introduced as an adaptive mathematical structure on this basis, thereby converting the original data set or data vector feature description into a new space, and setting the upper limit of the possible dimension of this space as an open upper limit, so that it has a sufficiently high data analysis capacity; at this time, based on the kernel function mapping mathematical tool, a set of mathematical expressions are used to characterize the data conversion corresponding to the kernel function mapping, and a parameter variable is implicitly constructed in the above mathematical expression, and based on this parameter variable, their similarity measurement is adjusted according to the difference between the feature descriptions of the two pieces of land, and the above-mentioned implicitly constructed parameter variable is constructed as the rate at which the similarity of the corresponding land area changes with the feature difference; after completing the above-mentioned mapping conversion, the data structure is transformed into a new data space, and the similarity between the lands reveals a nonlinear pattern that can be captured, thereby expanding based on the kernel function mapping to obtain more comprehensive validity data, and at this time according to the following data connotation: each component of the similarity matrix is ​​constructed to characterize the degree of similarity between the two pieces of land in the new space, and data output is performed to obtain a similarity matrix characterizing the degree of association between the lands.

[0032] Further data optimization:

[0033] A. In order to ensure the stability and reliability of the land similarity matrix, the above data output is normalized. This process involves adjusting the values ​​of the land similarity matrix so that the sum of the values ​​of each row and column in the matrix is ​​1, so as to eliminate the possible scale differences between the sample points and make the matrix more balanced and standardized; the mathematical operation method corresponding to this data processing is: normalization based on the diagonal matrix so that the elements on each diagonal of the diagonal matrix are the sum of the values ​​of the corresponding rows in the similarity matrix.

[0034] B. Simultaneously simplify and optimize the data structure and reduce the calculation process, and further transform the normalized similarity matrix. Since only relatively large similarity values ​​are important in actual data processing and those smaller values ​​should be ignored, this transformation is completed by setting a threshold. Only when the value in the matrix is ​​greater than this threshold will the value be retained, otherwise it will be set to 0; thereby reducing unnecessary calculations and retaining valuable key data information.

[0035] Because in the actual processing of wheat zoning planting and fertilizer control problems, only relatively large similarity values ​​are important, while those small values ​​can usually be ignored. In this way, we can retain the key similarity information while greatly reducing the computational complexity and resource consumption.

[0036] Secondly, a clustering algorithm is constructed based on the above optimized land similarity matrix.

[0037] Taking the optimized similarity matrix as input, the land is divided into different planting areas with real differences in global geography, climate and other related data; specifically, the following data process is introduced to handle complex data structures and reveal subtle clustering patterns in the data: First, we construct and create a graphical representation according to the following rules to show the similarities between lands: In the construction of this graph, each piece of land is regarded as a node, and the similarities between lands are represented by the edges connecting these nodes; then, these edges are assigned a weight or correlation data strength, which is given by the above-mentioned land similarity matrix; further, based on the sum of the weights or correlation data strengths of all edges connected to each node, a numerical representation of the correlation of each node is constructed, thereby obtaining a torsion matrix, which records the data commonality of each node in the graph. vibration; at this time, consider introducing the Laplace method, and obtain a new Laplace characterization matrix by subtracting the similarity matrix from the twisted matrix, which linearly characterizes the structural characteristics of the initially constructed graph and its clustering characteristics. At this time, the eigenvalues ​​and eigenvectors of the obtained Laplace characterization matrix can be extracted; among which, the eigenvalue represents the information of how the associated graph is divided into clusters, and the corresponding eigenvector represents the shape of each associated cluster; the next step is to extract the eigenvectors corresponding to the smallest several non-zero eigenvalues, and use these eigenvectors to create a new data representation for each piece of land. Based on this data representation, the similarity between the lands is reflected, thereby directly citing or linking existing standardized clustering algorithms such as K-means or other standard clustering algorithm tools to divide the land into several areas, and the land in each area has similar characteristics.

[0038] Embodiment 2: Example of constructing a discretized numerical partitioning model

[0039] Under the guidance of the above dominant data architecture and technical route, a specific discretized numerical partitioning model can be constructed by using high-dimensional feature mapping and distance measurement based on kernel methods.

[0040] At this time, the mathematical representation of the land similarity matrix is ​​a matrix S, whose element Sij represents the similarity between the i-th piece of land and the j-th piece of land (characterized by the distance measurement based on soil data and climate data combined with high-dimensional feature mapping). The steps for constructing the land similarity matrix S include:

[0041] Feature vector construction: For each piece of land i, first construct a feature vector xi, which contains all the measured values ​​of soil data and climate data. This vector may include pH value, organic matter content, soil texture, macro-element content, medium and trace element content, soil moisture content, temperature, precipitation, sunshine duration, wind speed and direction, relative humidity, etc.

[0042] Kernel function mapping: Use the kernel function (k(xi, xj)) to map the original feature space to a high-dimensional feature space (possibly infinite-dimensional) in order to better capture complex nonlinear similarities in this space.

[0043] As an initial modeling, the commonly used Gaussian radial basis function can be used as the initial kernel function:

[0044] k(xi, xj) = exp(-|xi - xj|^2 / (2σ^2));

[0045] Among them, σ is the bandwidth parameter of the kernel function, which controls the speed at which the similarity changes with distance.

[0046] Similarity matrix calculation: Calculate the similarity matrix S between the lands, where each element Sij is given by the kernel function (k(xi, xj)), representing the similarity between land i and land j in the high-dimensional feature space.

[0047] First-level normalization of land similarity matrix: In order to make the similarity matrix S more robust, it can be normalized. As the initialization model before the initial data iteration, the Laplace method is introduced here, which is specifically characterized as: Sreg = D^(-1 / 2) SD^(-1 / 2) Where D is a diagonal matrix, and its diagonal element Dii is the sum of the i-th row of the matrix S; this normalization can make the sum of the rows and columns of the similarity matrix equal to 1, thereby eliminating the scale differences of different land sample points.

[0048] Secondary simplification of the land similarity matrix: There are multiple options at this point. For example, you can consider further converting the similarity matrix Sreg into a sparse matrix to reduce the computational complexity.

[0049] Based on the characteristics of the land similarity matrix S, it can be further used as input to cluster the planting areas. Considering the many dimensions and complexity of land data, we have constructed a clustering technology based on graph theory and linear algebra. The data process includes:

[0050] Graph construction: The land similarity matrix S is considered as the adjacency matrix of a weighted undirected graph G, where each node of the graph represents a piece of land and the weight of the edge is given by the similarity value in S. Twist matrix calculation: Calculate the degree matrix D of the graph G, which is represented as a diagonal matrix, each of whose diagonal elements Dii is equal to the sum of the weights of all edges of node i. Laplacian matrix construction: Calculate the scale-free Laplacian matrix L of the graph G, defined as L = D - S; the use of scale-free Laplacian matrix can reveal the spectral properties of the graph. Eigenvalue and eigenvector calculation: Calculate the eigenvalues ​​and corresponding eigenvectors of the Laplacian matrix L. Since L is a symmetric semi-positive definite matrix, all eigenvalues ​​are non-negative. Eigenvector selection: Select the eigenvectors u1, u2, ..., uk corresponding to the smallest k non-zero eigenvalues. These eigenvectors form a new feature space, and each land i is represented in this space as yi = (u1(i), u2(i), ...,uk(i)). Clustering in the new feature space: In the new feature space, the vector yi is clustered using traditional clustering algorithms such as K-means to divide the land into k regions, each of which contains land with similar characteristics. The last section of the data model maps the clustering results back to the original land to obtain a stereotyped partition, that is, the land is divided into different planting areas with real differences in global geography, climate and other related data. It not only contains the planting area differences visible in the original feature data set or data vector, but the above data processing process expands and reveals the inherent similarities in the original feature data.

[0051] Example 3: Functionalized fitting partition model

[0052] The basic data organization of the functional fitting zoning model is consistent with Example 1; its main data architecture is: for the multidimensional attributes and multivariate attributes of soil data and its related data in space and time, multivariate spatial statistical equations are used to fit the spatial continuity and heterogeneity of crop-planting land data and its related data; the effectiveness division of crop-planting areas is carried out on this data architecture.

[0053] Specifically, we first consider the optimal distribution prediction of spatial data and the spatial correlation between data points, and introduce the Kriging interpolation method to perform spatial interpolation processing on the collected soil and climate data, thereby estimating the continuous distribution of these variables in the entire region;

[0054] Basically, a geostatistical method based on optimal unbiased prediction can be constructed directly using Kriging interpolation, where for each soil and climate variable v a random field Zv(x) is constructed, where x is the spatial position; this interpolation takes into account the spatial correlation and describes it through its variogram (semivariogram) γv(h), where h is the spatial lag;

[0055] Specifically, here we construct the basic Kriging difference as:

[0056] Z(x) is set as a random variable of soil or climate properties at location x, and the interpolation model is formalized as:

[0057] Z(x) = μ(x) + ε(x);

[0058] Among them, μ(x) is a trend function representing large-scale spatial variation (for simplicity, it is generally set to a polynomial structure), and ε(x) is a random error term representing small-scale variability; in the zoning problem of cultivated land here, they are usually spatially correlated;

[0059] Next, the trend function and semivariogram are estimated by solving the following optimization problem:

[0060] minimize over μ(x), σ², φ the sum from i=1 to n of (Z(xi) - μ(xi))²

[0061] subject to the constraints σ² ≥ 0 and φ ≥ 0

[0062] where σ² is the variance and φ is the spatial correlation parameter.

[0063] In addition, for more complex and common data analysis modes, we can also consider building a multivariate interpolation model based on the Kriging interpolation method to adapt to the many data dimensions in each wheat partition and the many related spatial variables such as land geography and climate. To this end, we tried to build the following multivariate interpolation model.

[0064] Adapting to the many data dimensions in each wheat zone and the many related spatial variables such as land geography and climate, suppose we have m spatially related variables Z1(x), Z2(x), ..., Zm(x), and each random variable Zj(x) represents the jth soil or climate attribute at position x. The multivariate model can be expressed as:

[0065] Zj(x) = μj(x) + εj(x), where j = 1,2,...,m;

[0066] in:

[0067] μj(x) is the trend function of the jth variable (polynomial in position x).

[0068] εj(x) is the random error term of the jth variable, representing small-scale variability.

[0069] It can be seen that the multivariate model constructed here further takes into account the cross-semivariogram between variables, that is, the spatial correlation between different variables.

[0070] In the subsequent optimization, the generalization only requires estimating the parameters between the trend function μj(x), the variance σj² and the cross-semivariogram γjk(h), where h is the spatial distance and γjk(h) represents the spatial association between the jth and kth variables at distance h.

[0071] For example, according to the current planting data processing task, the optimization objective may be set to minimize the weighted squared difference of all variables:

[0072] minimize over μj(x), σj², γjk(h) the sum from j=1 to m of the sumfrom i=1 to n of wj(Zj(xi) - μj(xi))²; where wj is the weight of the jth variable.

[0073] Next, a land property function is constructed that comprehensively covers all relevant soil and climate geography and other related variables. Simple linear combination sub-terms are set in this function, while sub-terms containing complex nonlinear terms are set to reflect the interactions between different variables. Each component of this function is constructed based on Kriging interpolation. In particular, a weight is assigned to each variable to correspond to the size of its measurable influence on land properties.

[0074] Here, based on the above interpolation method, a land property function F(x) can be specifically defined, which should be constructed as a multivariate function to combine all relevant soil and climate variables. This function can be linear or nonlinear, depending on the interaction between the variables. For example, F can be defined in data structure as:

[0075] F(x) = w1 ZpH(x) + w2 ZOM(x) + ... + wn ZClimate_n(x) + Φ(x)

[0076] Among them, wi is the weight parameter, ZpH, ZOM, ..., ZClimate_n are the random fields of each variable obtained by Kriging interpolation, and Φ(x) is a function that may contain nonlinear terms to describe the interaction between variables.

[0077] Furthermore, in order to ensure that the zoning model can reflect the spatial continuity of land characteristics, a partial differential equation is constructed to further process the land characteristic function; specifically, given that the orientation of zoning is fertilizer reduction and the diffusion properties of fertilizer and soil, the Allen-Cahn equation is considered to simulate the separation and pattern formation between cultivated lands, and the parameters of the equation are superimposed and adjusted to characterize the changes of land characteristics over time and space and the formation of boundaries between different land characteristics; we conducted a simplified partial differential equation modeling, that is, using the Allen-Cahn equation based on the partial differential model to further process the land characteristic function F:

[0078] ; where Δ is the Laplace operator, ε is the diffusion coefficient, and W'(F) is the derivative of a bistable state that models the boundaries between different land properties.

[0079] At this point, numerical methods such as the finite element method or the finite difference method can be directly used to solve the partial differential equation, and its stable solution can characterize the final pattern of land characteristics; on this basis, the regional boundaries are delineated. Specifically, two optional data paradigms are constructed in this data step: ① Determine the boundaries of different regions by setting a specific threshold; ② Introduce the level set data paradigm to determine the boundaries of different regions.

[0080] Finally, based on the above-mentioned regional boundary demarcation, a clustering algorithm is further introduced to cluster these land characteristic patterns, thereby constructing a zoning model defined or represented by precise quantitative data. In terms of data process, it has the following data optimization potential: it not only targets the multidimensional and multivariate attributes of soil data and its related data in space and time, but also is compatible with the spatial distribution of land characteristics and the continuity of land characteristics; the above-mentioned data optimization potential is derived through the introduction and iterative optimization of the subsequent artificial intelligence platform.

[0081] Example 4: Subsequent Progression

[0082] Subsequently, it will be necessary to develop different types of fertilizer reduction and / or crop yield prediction progressive models based on the effectiveness division of planting zones.

[0083] Among them, the fertilizer reduction related model is constructed based on survey data or guidance data, or is completely characterized and constructed based on physical parameters of fertilizer soil science. For the latter, a possible effective data route is to build a model based on partial differential equations to simulate the diffusion of fertilizers in the soil and the absorption of fertilizers by crops. For example, if the diffusion process of fertilizers in the soil follows Fick's second law, the corresponding partial differential equation representation can be constructed by combining fertilizer concentration, diffusion coefficient, crop absorption rate of fertilizers (depending on fertilizer concentration and time), etc., and then the above partial differential equations are solved through boundary conditions and initial conditions and the introduction of finite difference methods, while fitting historical yield data and fertilizer usage to estimate the parameters of the function; under this model, artificial intelligence algorithms such as genetic algorithms can be directly introduced for optimization iteration.

[0084] Among them, the crop yield prediction progressive model focuses on adaptive construction based on the iteration of artificial intelligence. Generally, it needs to combine the output of the planting zoning model and the fertilizer reduction model, and then build a comprehensive yield prediction model based on partial differential equations or statistical learning. Considering the access convenience and adaptability of the artificial intelligence platform, the statistical learning route may be more effective.

[0085] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for constructing an artificial intelligence-guided fertilizer reduction data model based on crop planting area differentiation, characterized in that: The method is based on the data construction of crop planting partitions. According to the data construction of crop planting partitions, multiple groups of data models are constructed to adapt to various specific numerical analysis and data processing tasks; the multiple groups of data models include at least the following two groups of specific models: a discrete numerical partition model and a functional fitting partition model; The main data structure of the discretized numerical partitioning model is as follows: constructing a land similarity matrix based on all measured values ​​of soil data and its associated geographic or climate data and deriving components of the similarity matrix using the Euclidean distance of the soil to represent the similarity between different land areas; and performing validity division of crop planting areas based on this data structure; Specifically, the method for constructing the discretized numerical partitioning model includes: To construct the land similarity matrix, firstly, a characteristic data description containing soil geographic and climate data is created for each piece of land, which is specifically represented as a data set or data vector, which contains multiple indicators of each piece of land, including soil pH, organic matter content, texture, macro-element level, medium and trace element level, water content, and temperature, precipitation, sunshine hours, wind speed, wind direction and relative humidity under climatic conditions; After the above-mentioned characteristic data description data set or data vector is created, the kernel function mapping data structure is further introduced as an adaptive mathematical structure on this basis, thereby converting the original data set or data vector characteristic description into a new space, and setting the upper limit of the possible dimension of this space as an open upper limit, so that it has a sufficiently high data analysis capacity; at this time, based on the kernel function mapping mathematical tool, a set of mathematical expressions are used to characterize the data conversion corresponding to the kernel function mapping, and a parameter variable is constructed in the above-mentioned mathematical expression, and based on this parameter variable, their similarity measurement is adjusted according to the difference between the characteristic descriptions of the two pieces of land. The above-mentioned constructed parameter variable is constructed as the rate at which the similarity of the corresponding land area changes with the characteristic difference; after the mapping conversion is completed, the data structure is transformed into a new data space, and the similarity between the lands reveals a nonlinear pattern that can be captured, thereby expanding based on the kernel function mapping to obtain more comprehensive validity data, and at this time according to the following data connotation: each component of the similarity matrix is ​​constructed to characterize the degree of similarity between the two pieces of land in the new space, and the data output is performed to obtain a similarity matrix that characterizes the degree of association between the lands; Further data optimization: A. In order to ensure the stability and reliability of the land similarity matrix, the above data output is normalized. This process involves adjusting the values ​​of the land similarity matrix so that the sum of the values ​​of each row and column in the matrix is ​​1, so as to eliminate the possible scale differences between the sample points and make the matrix more balanced and standardized; the mathematical operation method corresponding to this data processing is: normalization based on the diagonal matrix so that the elements on each diagonal of the diagonal matrix are the sum of the values ​​of the corresponding rows in the similarity matrix; B. Simultaneously simplify and optimize the data structure and reduce the calculation process, and further transform the normalized similarity matrix. Since only relatively large similarity values ​​are important in actual data processing and those smaller values ​​should be ignored, this transformation is completed by setting a threshold. Only when the value in the matrix is ​​greater than this threshold will this value be retained, otherwise it will be set to 0; thereby reducing unnecessary calculations and retaining valuable key data information; Finally, a clustering algorithm is constructed based on the optimized land similarity matrix. The optimized similarity matrix is ​​used as input to divide the land into different planting areas with real differences in global geography and climate. Specifically, the following data process is introduced to handle complex data structures and reveal subtle clustering patterns in the data: First, a graphical representation is constructed and created according to the following rules to show the similarities between lands: In the construction of this graph, each piece of land is regarded as a node, and the similarities between lands are represented by the edges connecting these nodes; then, these edges are given a correlation data strength, which is given by the above-mentioned land similarity matrix; further, a numerical representation of the correlation of each node is constructed based on the sum of the weights or correlation data strengths of all edges connected to each node, thereby obtaining a twisted connection. Matrix, which records the data resonance of each node in the graph; at this time, consider introducing the Laplace method, and obtain a new Laplace representation matrix by subtracting the similarity matrix from the twisted matrix, which linearly represents the structural characteristics of the initially constructed graph and its clustering characteristics. At this time, the eigenvalues ​​and eigenvectors of the obtained Laplace representation matrix can be extracted; among them, the eigenvalue represents the information of how the associated graph is divided into clusters, and the corresponding eigenvector represents the shape of each associated cluster; the next step is to extract the eigenvectors corresponding to the smallest several non-zero eigenvalues, and use these eigenvectors to create a new data representation for each piece of land. Based on this data representation, the similarity between the lands is reflected, so that the existing standardized clustering algorithm tools can be directly referenced to divide the land into several areas, and the land in each area has similar characteristics.

2. The method for constructing an artificial intelligence-guided fertilizer reduction data model based on crop planting area differentiation according to claim 1 is characterized in that: After dividing the land into several areas, the clustering results are mapped back to the original land to obtain stereotyped partitions, that is, the land is divided into different planting areas with real differences in global geography and climate. It not only includes the planting area differences visible in the original feature data set or data vector, but also expands and reveals the inherent similarities in the original feature data.

3. The method for constructing an artificial intelligence-guided fertilizer reduction data model based on crop planting area differentiation according to claim 2 is characterized in that: The construction of the functional fitting partition model is mainly based on the model construction direction dominated by the introduction of spatial statistics and differential equation mathematical tools, specifically including: Firstly, considering the optimal distribution prediction of spatial data and the spatial correlation between data points, the interpolation method was introduced to perform spatial interpolation processing on the collected soil and climate data, thereby estimating the continuous distribution of these variables in the entire region; Next, a land property function that comprehensively covers all relevant soil and climate geography is constructed. Simple linear combination sub-terms are set in this function, while sub-terms containing complex nonlinear terms are set to reflect the interaction between different variables. Each component of this function is constructed based on Kriging interpolation. A weight is assigned to each variable to correspond to the size of its measurable influence on land properties. Furthermore, in order to ensure that the zoning model can reflect the spatial continuity of land characteristics, a partial differential equation is constructed to further process the land characteristic function; specifically, given that the orientation of zoning is fertilizer reduction and the diffusion properties of fertilizer and soil, the Allen-Cahn equation is used to simulate the separation and pattern formation between cultivated lands, and the parameters of the equation are superimposed and adjusted to characterize the changes in land characteristics over time and space and the formation of boundaries between different land characteristics; at this time, numerical methods can be directly used to solve this partial differential equation, and its stable solution characterizes the final pattern of land characteristics; on this basis, regional boundaries are delineated, specifically, two data paradigms are constructed: ① by setting a specific threshold to determine the boundaries of different regions; ② by introducing the level set data paradigm to determine the boundaries of different regions; finally, on the basis of the above regional boundary delineation, clustering algorithms are further introduced to cluster these land characteristic patterns as needed, thereby constructing a zoning model defined or characterized by precise quantitative data.

4. The method for constructing an artificial intelligence-guided fertilizer reduction data model based on crop planting area differentiation according to claim 1 is characterized in that: Regarding the subsequent different types of fertilizer reduction and / or crop yield prediction progressive models, the fertilizer reduction related models are constructed based on survey data or guidance data, or based on physical parameters of fertilizer soil science; the crop yield prediction progressive model is adaptively constructed based on the iterative nature of artificial intelligence.

5. The method for constructing an artificial intelligence-guided fertilizer reduction data model based on crop planting area differentiation according to any one of claims 1 to 4, characterized in that: The crop is wheat.