CN116383589A - Spatial interpolation prediction method for soil nutrients - Google Patents

Spatial interpolation prediction method for soil nutrients Download PDF

Info

Publication number
CN116383589A
CN116383589A CN202310386123.2A CN202310386123A CN116383589A CN 116383589 A CN116383589 A CN 116383589A CN 202310386123 A CN202310386123 A CN 202310386123A CN 116383589 A CN116383589 A CN 116383589A
Authority
CN
China
Prior art keywords
residual
value
group
values
grid
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310386123.2A
Other languages
Chinese (zh)
Inventor
孟英
朱亮
魏海洋
宋鑫
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Heilongjiang Carbon Valley Industry And Trade Co ltd
Original Assignee
Heilongjiang Carbon Valley Industry And Trade Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Heilongjiang Carbon Valley Industry And Trade Co ltd filed Critical Heilongjiang Carbon Valley Industry And Trade Co ltd
Priority to CN202310386123.2A priority Critical patent/CN116383589A/en
Publication of CN116383589A publication Critical patent/CN116383589A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/17—Function evaluation by approximation methods, e.g. inter- or extrapolation, smoothing, least mean square method
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0499—Feedforward networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Biology (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computing Systems (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Algebra (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention relates to a spatial interpolation prediction method of soil nutrients, which specifically comprises the following steps: s1, estimating a soil nutrient space systematic effect by using a BP neural network, and obtaining a neural network fitting residual error; s2, carrying out gridding treatment on residual data to construct a residual grid; and S3, performing iterative updating on the full-value scaling residual error grid by using a Belman equation, and adding the predicted value of the BP neural network to the soil nutrient attribute of the sampling point and the predicted value of the Belman equation to the residual error of the sampling point to obtain a final spatial interpolation predicted result of the soil nutrient. The invention requires less expert knowledge: statistical assumptions about the distribution and stationarity of the target variables are not required, nor is modeling about the theoretical semi-variance function. According to the invention, as the neural network method is adopted, more auxiliary variables can be used for assisting interpolation, and the final prediction is obtained by adding the predicted value of the BP neural network and the residual value fitted by the Belman equation, so that the accuracy of interpolation is improved.

Description

Spatial interpolation prediction method for soil nutrients
Technical Field
The invention relates to a soil assessment method, in particular to a spatial interpolation prediction method of soil nutrients.
Background
Agriculture is the source of people's clothing and food and the national survival cost, and is the basis of all production and living. Precision agriculture is an important development model of modern agriculture. The processing analysis of the agricultural data reduces the trial-and-error cost of agricultural development, reduces investment and improves production efficiency at the same time so as to obtain larger benefits. Limited by economic level, technical means, and topography conditions, the loss of soil nutrient data is a common problem. The spatial interpolation of the soil nutrient data can estimate the nutrient data value of an unknown point through adjacent data sampling points so as to make up for the defect of data missing, which is particularly important for aspects such as agricultural production, soil resource monitoring and protection. However, due to the limitations of economic level, technical means, and topography conditions, soil nutrient data in many places are difficult to obtain, in which case the spatial interpolation prediction method provides an evaluation method for estimating the data of the non-sampled places, which can estimate the approximation of a specific place based on the sampled values measured at other places.
Neural networks are widely used in interpolation algorithms based on functional methods because of their superior performance in terms of function approximation. The neural network can establish a nonlinear feedback system, and theoretically, the neural network with a three-layer structure can approach any complex function. Taking BP neural network trained by error back propagation as an example, the optimization method is a gradient descent method. The connection weight between the layers of networks is continuously adjusted by using back propagation so that the networks can be better predicted. In addition, the neural network can use other data besides the soil to-be-measured attribute to assist in interpolation, so that the accuracy of prediction can be improved to a certain extent. However, these studies still have the following problems:
(1) Although neural networks are widely used in interpolation, most spatial interpolation methods using neural networks do not take geographical properties into consideration, i.e. do not fully exploit spatial information contained in the residual.
(2) The neural network residual kriging method is a method for improving the neural network interpolation in residual in a small number. However, there are many disadvantages to using the kriging process for processing the residuals using the neural network residual kriging method: the path between two points is not unique, some paths are far more important in value propagation than straight-line paths between two points.
(3) Using the normal kri Jin Nige residual, a severe assumption is required: statistical assumptions about the distribution and stationarity of the target variables model about the theoretical half-variance function. It may not be easy to find a reliable geostatistical model that fits all residual data when actually used.
The existence of the above problems results in a large prediction result error of the neural network-based interpolation method, and cannot be used for guiding practice.
Disclosure of Invention
The invention aims to provide a spatial interpolation prediction method for soil nutrients, which aims to solve the problem of larger prediction result error in the existing neural network interpolation method.
The purpose of the invention is realized in the following way:
a spatial interpolation prediction method of soil nutrients comprises the following steps:
s1, estimating a soil nutrient space systematic effect by using a BP neural network, and obtaining a neural network fitting residual error; the specific operation mode is as follows:
s11, collecting data information of organic matter nutrient content of each sampling point, and carrying out normalization processing according to the following formula:
u i = (z i -z min ) / (z max -z min )
wherein,,z i is the firstiThe attribute values of the individual sample points,z max for the maximum value of the sample point attribute value,z min is the minimum value of the sample point attribute value.
S12, predicting soil attribute values of all sampling points by using BP neural network, wherein the number of the input layer nodes is determined by longitude and latitude dimensions and geographic attribute dimensions of the sampling points, and the number of hidden layer nodesHIs determined by the following empirical formula:
H=(I+O) 1/2 +a
wherein,,Hfor the number of hidden layer nodes,Ifor the number of nodes of the input layer,Ofor the number of output layer nodes, setO=1，aIs constant.
S13, using attribute value data of organic matter nutrient content of each sampling point, and acquiring earth surface attribute information of each sampling point according to longitude and latitude data information of the sampling point; splicing the soil attribute value information of each sampling point with the earth surface attribute information to construct a multi-source data set; dividing the multi-source data set into a training set and a testing set, putting the training set into a BP neural network to obtain predicted values for all sampling points in the testing set, and combining the true values of all the sampling points in the testing set to obtain residual errors of all the sampling points in the training set; and combining the residual errors of all sampling points of the training set with the longitude and latitude data of all sampling points of the training set and the longitude and latitude data of all sampling points of the test set to form a residual error data set.
S2, performing gridding treatment on the residual data set to construct a residual grid; the specific operation mode is as follows:
s21, calculating the minimum Euclidean distance between every two longitude and latitude coordinates of all sampling points in the residual data setd min 。
S22, minimum Euclidean distanced min Setting the initial diagonal length of the grid, constructing the grid, calculating the relative positions of all sampling points according to longitude and latitude, and placing the residual values of all sampling points of the training set calculated by the BP neural network in the constructed grid to form a residual grid.
S23, scaling the size of the residual grid to form a scaling grid, and placing residual values of sampling points into the scaling grid, wherein the relative positions of the residual values in the scaling grid are kept unchanged; when each residual value is placed in the scaling grid, if the position of the placement point conflicts, placing the residual value which is needed to be placed at the back on the position of the adjacent cell of the conflict position; and after all residual values of the sampling points are put in, forming a scaled residual grid.
S3, assigning the scaled residual grids into full-value scaled residual grids, performing sub-sampling treatment, and performing iterative updating on the full-value scaled residual grids by using a Bellman equation until convergence to obtain predicted values of the Bellman equation on sampling point residuals; and adding the predicted value of the BP neural network to the soil nutrient attribute of the sampling point and the predicted value of the Belman equation to the residual error of the sampling point to obtain a final spatial interpolation predicted result of the soil nutrient.
Further, the activation function used by the hidden layer in the BP neural network is a ReLU function:
f(x)=max(0,x)
wherein,,f(x)to activate the value of the function, max (0,x) To take inputxAnd a maximum value between 0.
Further, the loss function used in the BP neural network is a square loss function:
L(Y,f(X))=(Y-f(X)) 2
wherein,,Lin order to be lost, the process is carried out,Yas a result of the actual output value,f(X) Input variables for model pairsXIs a predicted value of (a).
Further, the specific operation manner of step S13 is: the residual data set is divided into 5 parts, namely a group A, a group B, a group C, a group D and a group E, and the five groups are mixed and recombined into a BCD group, an ACD group, an ABD group and an ABC group. Training BP neural network by using BCD group to obtain predicted value of A groupy A And first prediction of group Ey E1 The method comprises the steps of carrying out a first treatment on the surface of the Similarly, the predicted values of group B are obtained successivelyy B And second prediction value of group Ey E2 Predicted value of group Cy C And third prediction value of group Ey E3 Predicted value of group Dy D And fourth predictive value of group Ey E4 The method comprises the steps of carrying out a first treatment on the surface of the Then, through the true values of group A, group B, group C and group DA、B、C、DObtaining residuals of four groups of A, B, C and Dε A 、ε B 、ε C 、ε D The method comprises the steps of carrying out a first treatment on the surface of the Residual error of group Eε E Temporarily setting the sampling point as 0, integrating longitude and latitude data and corresponding residual errors of sampling points of five groups of A, B, C, D and E into a residual error data set in the mode of [ + ]A 1 ,A 2 ,ε) The method comprises the steps of carrying out a first treatment on the surface of the Wherein, the method comprises the following steps ofA 1 ,A 2 ) Is the longitude and latitude data of the sampling point,εis the residual of the sample point.
Further, the number of sample point residuals with position conflicts transferred in step S23 is controlled to be within 2% of the total number of cells in the grid.
Further, the specific operation manner of step S3 is as follows:
and S31, assigning values to the cells without residual values in the scaling residual grids to form a full-value scaling residual grid, wherein the assigned values are obtained by adding all the residual values which are already placed in the scaling residual grid and dividing the total number of the cells in the scaling residual grid.
S32, sub-sampling is carried out on the full-value scaling residual grids to obtain space weight parameters, and the sub-sampling proportion of the full-value scaling residual grids is randomly determined in the residual grids constructed in the step S22; assigning a random number between 0 and 1 to each cell in the full-value scaling residual grid; if the random number is smaller than the sub-sampling proportion number, reserving a filling value in the cell; if the random number is greater than the sub-sampled proportional number, replacing the fill value in the cell with a null value; after filling values in all cells of the full-value scaling residual grid are processed in this way, generating a sub-grid with a specified sub-sampling proportion; assigning values to the vacant cells in the sub-grids, wherein the assigned values are obtained by adding filling values reserved in the sub-grids and dividing the filling values by the total number of the cells in the sub-grids; and determining the space weight factor gamma of the sub-grid by means of random search parameters, and taking the space weight factor gamma as the space weight factor gamma of the full-value scaling residual grid.
S33, in the Markov process, the state transition probability that the state S at a certain moment t reaches another state S' at the next moment t+1 is as follows:
P SS' =P[S t+1 =S'|S t =S]
wherein, P is the probability,SandS t the state is that the time is t,S'andS t+1 are alltState at +1.
S34, iteratively updating the full-value scaling residual error grid by using a Belman equation until convergence, and obtaining a predicted value of the sampling point residual error after iteratively updating by using the Belman equation:
v(s) = E[R t+1 +γv(s t+1 )|s t =s]
wherein R is t+1 For immediate rewards to reach a certain state, gamma is a spatial weighting factor,v(s t+1 ) For a subsequent state cost function to which the spatial weight factor gamma is added.
And S35, adding the predicted value of the BP neural network on the soil nutrient attribute at the sampling point with the predicted value of the Belman equation on the residual error of the sampling point, and taking the added result as a final spatial interpolation predicted result of the soil nutrient.
The beneficial effects of the invention are as follows:
1. and combining the BP neural network with a Belman equation, fitting the trend existing in the space by using the BP neural network, and processing the residual error by using the Belman equation. Compared with common kriging and neural network residual kriging, the prediction method of the invention requires much less expert knowledge, and does not need statistical assumptions about the distribution and stationarity of the target variable, nor modeling about the theoretical half variance function.
2. Compared to the inverse distance weighted interpolation, the prediction method of the present invention can use more auxiliary information about the point of the value to be measured.
3. Because of the neural network approach, more auxiliary variables can be used to assist in interpolation. Since there is some correlation between interpolation points, these correlations may help interpolation. Thus, using the BP neural network in combination with the Belman processing residual, the trend present in the system is processed using the BP neural network first, and then the residual field is processed using the Belman equation. And adding the predicted value of the BP neural network and the residual value fitted by the Belman equation together to obtain final prediction, thereby improving the accuracy of interpolation.
Drawings
FIG. 1 is a flow chart of a method for spatial interpolation prediction of soil nutrients according to the present invention.
FIG. 2 is a grid scaling schematic; wherein a is an initial residual grid map; b is a scaled mesh map; c is what diagram the training set and the test set divide the grid; d is what diagram of the final mesh.
FIG. 3 is a state set and state transition matrix diagram; wherein a is a state set diagram; b is the state transition matrix.
Fig. 4 is a neighbor distribution and value calculation graph.
FIG. 5 is a schematic diagram of value transfer in an iterative process; wherein a is the initialization process of the residual grid; b-f are the results of each iterative process.
FIG. 6 is a distribution diagram of 450 soil nutrient sampling points within the area of Tangshan city.
Fig. 7 is a root mean square error comparison graph of prediction results of five interpolation prediction methods.
Detailed Description
Examples:
the soil nutrient sampling points of the embodiment are derived from 450 soil nutrient sampling points in the area of Tangshan city, and the content of the sampling points comprises longitude, latitude, surface attribute, organic nutrient content of soil and the like. The longitude and latitude information of the sampling point is the position coordinate information of the sampling point, the earth surface attribute of the sampling point comprises information describing earth surface conditions such as gradient, earth surface fluctuation and the like, and the organic matter nutrient content of the sampling point is attribute value information of soil. Fig. 6 gives a specific distribution of these 450 sample points. The method is to conduct interpolation prediction on the organic matter nutrient content in the soil of the tested area. In this embodiment, 90 sampling points are randomly selected as test set data, and the remaining 360 sampling points are selected as training set data, so as to perform interpolation prediction.
As shown in FIG. 1, the spatial interpolation prediction method of the soil nutrients comprises the following steps:
s1, estimating a soil nutrient space systematic effect by using a BP neural network, and obtaining a neural network fitting residual error. The specific operation mode is as follows:
s11, collecting data information of organic matter nutrient content of each sampling point, and carrying out normalization processing according to the following formula:
u i = (z i -z min ) / (z max -z min )
wherein,,z i is the firstiThe attribute values of the individual sample points,z max for the maximum value of the sample point attribute value,z min is the minimum value of the sample point attribute value.
S12, predicting soil attribute values of the sampling points by using the BP neural network, wherein the number of the input layer nodes is determined by longitude and latitude dimensions and geographic attribute dimensions of the sampling points, and the number of hidden layer nodesHIs determined by the following empirical formula:
H=(I+O) 1/2 +a
wherein,,Hfor the number of hidden layer nodes,Ifor the number of nodes of the input layer,Ofor the number of output layer nodes, setO=1，aIs constant.
The activation function used by the hidden layer in the BP neural network is the ReLU function:
f(x)=max(0,x)
wherein,,f(x)to activate the value of the function, max (0,x) To take inputxAnd a maximum value between 0.
The loss function used in the BP neural network is a square loss function:
L(Y,f(X))=(Y-f(X)) 2
wherein,,Lin order to be lost, the process is carried out,Yas a result of the actual output value,f(X) Is a model predictive value for the input variable X.
Obtaining interpolation residuals using a BP neural network, wherein:
when predicting the attribute value of the organic matter nutrient content of the soil: the number of the input layer nodes is 8, the number of the hidden layer nodes is 1, the number of the hidden layer nodes is 11, and the number of the output layer nodes is 1; the neural network training frequency was 150.
When predicting nitrogen properties: the number of the input layer nodes is 8, the number of the hidden layer nodes is 1, the number of the hidden layer nodes is 13, and the number of the output layer nodes is 1; the neural network training number is 220.
When predicting available phosphorus properties: the number of the input layer nodes is 8, the number of the hidden layer nodes is 1, the number of the hidden layer nodes is 9, and the number of the output layer nodes is 1; the neural network training frequency was 200.
Predicting the slow potassium property: the number of the input layer nodes is 8, the number of the hidden layer nodes is 1, the number of the hidden layer nodes is 14, and the number of the output layer nodes is 1; the neural network training frequency was 200.
When predicting the quick-acting potassium attribute: the number of the input layer nodes is 8, the number of the hidden layer nodes is 1, the number of the hidden layer nodes is 16, and the number of the output layer nodes is 1. The neural network training number was 180.
S13, using attribute value data of organic matter nutrient contents of 450 sampling points in an area of Tangshan city, and simultaneously using Arcgis software to perform grid turning point operation on a corresponding area in a national high-resolution national soil information grid basic attribute data set (2010-2018) of a national earth system science data center, and acquiring surface attribute information of the 450 sampling points according to longitude and latitude data information of the sampling points. And splicing the soil attribute value information of each sampling point with the earth surface attribute data to construct a multi-source data set. Dividing the multi-source data set into a training set and a testing set, putting the training set into a BP neural network to obtain predicted values for all sampling points in the testing set, and combining the true values of all the sampling points in the testing set to obtain residual errors of all the sampling points in the training set. And then combining and sorting the residual errors of all sampling points of the training set, the longitude and latitude data of all sampling points of the training set and the longitude and latitude data of all sampling points of the test set into a residual error data set.
The specific operation mode is as follows: the residual data set is divided into 5 parts, namely a group A, a group B, a group C, a group D and a group E, and the five groups are mixed and recombined into a BCD group, an ACD group, an ABD group and an ABC group. Training BP neural network by using BCD group to obtain predicted value of A groupy A And first prediction of group Ey E1 . Similarly, the predicted values of group B are obtained successivelyy B And second prediction value of group Ey E2 Predicted value of group Cy C And third prediction value of group Ey E3 Predicted value of group Dy D And fourth predictive value of group Ey E4 . Then, through the true values of group A, group B, group C and group DA、B、C、DObtaining residuals of four groups of A, B, C and Dε A 、ε B 、ε C 、ε D . Residual error of group Eε E Temporarily setting the sampling point as 0, integrating longitude and latitude data and corresponding residual errors of sampling points of five groups of A, B, C, D and E into a residual error data set in the mode of [ + ]A 1 ,A 2 ,ε). Wherein, the method comprises the following steps ofA 1 ,A 2 ) Is the longitude and latitude data of the sampling point,εis the residual of the sample point.
S2, performing gridding processing on the residual data set to construct a residual grid.
As shown in fig. 2, the specific operation of step S2 is as follows:
s21, calculating the minimum Euclidean distance between every two longitude and latitude coordinates of all sampling points in the residual data setd min ；
S22, minimum Euclidean distanced min Initial set to gridAnd (3) the width is used for constructing a grid, the relative positions of all sampling points are calculated according to the longitude and latitude, and the residual values of all sampling points of the training set calculated by the BP neural network are placed in the constructed grid to form a residual grid.
S23, scaling the size of the residual grid, namely using cells with gray scale points as blank cells without residual values, and representing the cells with residual values by using cells with transverse lines (fig. 2 a), wherein each cell is ensured to be provided with at most one residual value. The residual grid formed at this time is very sparse. In order to improve the processing efficiency of the residual field, the size of the cells in the residual grid is scaled. The processed mesh forms an initial mesh. And counting the side length of the cell and the point position condition for placing the residual value. And placing residual values of the sampling points into an initial grid, wherein the relative positions of the residual values in the initial grid are kept unchanged. In the process of placing each residual value into the initial grid, if the cells into which the residual values are placed have position conflicts, placing the residual values to be placed at the back on the positions of adjacent cells at the conflict positions, namely one of the adjacent positions including the upper position, the lower position, the left position and the right position of the conflict positions, and the placement result is shown in fig. 2 b. This neighboring location is referred to as a "neighbor location". In order to ensure that the overall relative position of each cell into which the residual value is placed is unchanged, the number of sample point residual values of the position conflicts is shifted to be controlled to be 1.2% of the total number of cells in the grid. Within the transferred control proportion, the maximum cell side length is the final scaling grid. And after all residual values of the sampling points are put into the cells, forming a scaled residual grid. The cell in which the E group data is located in the grid is set as the test set (the cell with the forward diagonal line drawn in fig. 2 c). And for blank cells, wherein the cells with horizontal alternate lines are cells after iterative processing, and the grid data of the residual values are put in advance and kept unchanged. Cells with reverse diagonal lines are updated cells of the test set. The result after interpolation is shown in fig. 2 d. The root mean square error is calculated using the residual values in the forward-slashed cells and the residual values in the reverse-slashed cells.
S3, assigning the scaled residual grids into full-value scaled residual grids, performing sub-sampling treatment, and performing iterative updating on the full-value scaled residual grids by using a Bellman equation until convergence to obtain predicted values of the Bellman equation on sampling point residuals; and adding the predicted value of the BP neural network to the soil nutrient attribute of the sampling point and the predicted value of the Belman equation to the residual error of the sampling point to obtain a final spatial interpolation predicted result of the soil nutrient. The specific operation mode is as follows:
and S31, assigning values to the cells without residual values in the scaling residual grids to form a full-value scaling residual grid, wherein the assigned values are obtained by adding all the residual values which are already placed in the scaling residual grid and dividing the total number of the cells in the scaling residual grid.
S32, sub-sampling is carried out on the full-value scaling residual grids to obtain space weight parameters, and the sub-sampling proportion of the full-value scaling residual grids is randomly determined in the residual grids constructed in the step S22; assigning a random number between 0 and 1 to each cell in the full-value scaling residual grid; if the random number is smaller than the sub-sampling proportion number, reserving a filling value in the cell; if the random number is greater than the sub-sampled proportional number, replacing the fill value in the cell with a null value; after filling values in all cells of the full-value scaling residual grid are processed in this way, generating a sub-grid with a specified sub-sampling proportion; assigning values to the vacant cells in the sub-grids, wherein the assigned values are obtained by adding filling values reserved in the sub-grids and dividing the filling values by the total number of the cells in the sub-grids; and determining the space weight factor gamma of the sub-grid by means of random search parameters, and taking the space weight factor gamma as the space weight factor gamma of the full-value scaling residual grid.
Determining the space weight factor gamma of the organic matter attribute of the sub-grid as 0.5418, the space weight factor gamma of the total nitrogen attribute as 0.4926, the space weight factor gamma of the effective phosphorus attribute as 0.4127, the space weight factor gamma of the slow potassium attribute as 0.2998 and the space weight factor gamma of the quick-acting potassium attribute as 0.3633 by means of random search parameters; taking the space weight factor gamma of the sub-grid of each attribute as the space weight factor gamma of the whole residual grid of the attribute.
S33, in the Markov process, the state transition probability that the state S at a certain moment t reaches another state S' at the next moment t+1 is as follows:
P SS' =P[S t+1 =S'|S t =S]
wherein, P is the probability,SandS t the state is that the time is t,S'andS t+1 are alltState at +1.
The present invention uses state transition probabilities to measure the impact of each neighbor on the current cell. For simplicity of processing, the neighbor's impact on the current cell is considered the same here. In the state set shown in fig. 3a, different positions of the cells in the grid are given. In the state transition matrix presented in fig. 3b, the weight of each neighbor's impact on the current cell is shown. For example, a cell has three neighbors, and the three neighbors have the same influence on the current grid, and are all 1/3.
S34, iteratively updating the full-value scaling residual error grid by using a Belman equation until convergence, and obtaining a predicted value of the sampling point residual error after iteratively updating by using the Belman equation:
v(s) = E[R t+1 +γv(s t+1 )|s t =s]
wherein R is t+1 For immediate rewards to reach a certain state, gamma is a spatial weighting factor,v(s t+1 ) For a subsequent state cost function to which the spatial weight factor gamma is added.
FIG. 4 depicts the positional relationship of the current cell and its neighbors in the grid and gives the residual values in the neighbor positionsv 1 、v 2 、v 3 ) Residual error value in current cellv) Functional relation between the residual values of the current cellsv) And multiplying the residual values in the neighbor grids by a space weight factor lambda, and then taking an average value. As can be seen in one particular example shown in FIG. 5, the grid cell valuesThe iterative process is performed under the constraints of the bellman equation. Wherein fig. 5a is an initialization process of the residual grid; fig. 5b to 5f are five iterative processes. In this iteration process, the numerical gap between each grid is gradually smaller, and fig. 5f is the final iteration result. The arithmetic expressions listed on the right side of fig. 5 illustrate in detail the update calculation process of each cell from a to b of fig. 5 in this embodiment.
And S35, adding the predicted value of the BP neural network to the soil nutrient attribute at a certain sampling point and the predicted value of the Belman equation to the residual error of the sampling point to serve as a prediction model of the invention, namely a neural network residual error Belman method (Neural Network Residual Bellman, NNRB), and taking the added result as a final spatial interpolation prediction result of the soil nutrient.
In order to prevent over fitting in the network training process, training is terminated when training errors continuously decrease and test errors start to increase. And adding the predicted value of the BP neural network for the soil nutrient attribute with the residual predicted value obtained by iteratively updating the residual through a Belman equation, and obtaining a final spatial interpolation predicted result.
Fig. 7 shows the root mean square error comparison of the predicted results of the neural network residual bellman method (Neural Network Residual Bellman, NNRB) employed in the present invention with the predicted results of other interpolation prediction methods including the normal Kriging (OK), the inverse distance weighted interpolation (Inverse Distance Weight, IDW), the BP neural network interpolation (BP Neural Network Interpolation Algorithm, BPNNIA), and the neural network residual Kriging (Neural Network Residual Kriging, NNRK). As can be seen from the comparison result of fig. 7, among the comprehensive evaluation of various interpolation predictions of organic matter property, nitrogen property, phosphorus property, quick-acting potassium property and slow-acting potassium property of soil in the area to which the tangshan belongs, the best interpolation effect is the neural network residual bellman method of the present invention, and the worst is the BP neural network interpolation BPNNIA. The interpolation prediction of the quick-acting potassium attribute has the best interpolation effect, namely the neural network residual Belman method NNRB and the worst neural network residual Kriging method NNRK; and only on this property, the effect of the neural network residual kriging NNRK is not as good as the BP neural network interpolation BPNNIA. In interpolation prediction of other attributes, the neural network residual kriging method NNRK is superior to the BP neural network interpolation method BPNNIA. For the attribute value prediction of five soils in the area of Tangshan city, the prediction effect of the inverse distance weighted interpolation method IDW on the nitrogen attribute is superior to that of the common Kriging interpolation method OK, and the accuracy of the interpolation of the other attributes is inferior to that of the common Kriging interpolation method OK.

Claims (6)

1. A spatial interpolation prediction method of soil nutrients is characterized by comprising the following steps:
s1, estimating a soil nutrient space systematic effect by using a BP neural network, and obtaining a neural network fitting residual error; the specific operation mode is as follows:
s11, collecting data information of organic matter nutrient content of each sampling point, and carrying out normalization processing according to the following formula:
u i = (z i - z min ) / ( z max - z min )
wherein,,z i is the firstiThe attribute values of the individual sample points,z max for the maximum value of the sample point attribute value,z min the minimum value of the attribute value of the sampling point;
s12, predicting soil attribute values of all sampling points by using BP neural network, wherein the number of the input layer nodes is determined by longitude and latitude dimensions and geographic attribute dimensions of the sampling points, and the number of hidden layer nodesHIs determined by the following empirical formula:
H=(I+O) 1/2 +a
wherein,,Hfor the number of hidden layer nodes,Ifor the number of nodes of the input layer,Ofor the number of output layer nodes, setO=1，aIs a constant;
s13, using attribute value data of organic matter nutrient content of each sampling point, and acquiring earth surface attribute information of each sampling point according to longitude and latitude data information of the sampling point; splicing the soil attribute value information of each sampling point with the earth surface attribute information to construct a multi-source data set; dividing the multi-source data set into a training set and a testing set, putting the training set into a BP neural network to obtain predicted values for all sampling points in the testing set, and combining the true values of all the sampling points in the testing set to obtain residual errors of all the sampling points in the training set; combining and sorting the residual errors of all sampling points of the training set, the longitude and latitude data of all sampling points of the training set and the longitude and latitude data of all sampling points of the test set into a residual error data set;
s2, performing gridding treatment on the residual data set to construct a residual grid; the specific operation mode is as follows:
s21, calculating the minimum Euclidean distance between every two longitude and latitude coordinates of all sampling points in the residual data setd min ；
S22, minimum Euclidean distanced min Setting the initial diagonal length of the cells in the grid, constructing the grid, calculating the relative positions of all sampling points according to longitude and latitude, and placing the residual values of all sampling points of the training set calculated by the BP neural network in the constructed grid to form a residual grid;
s23, scaling the size of the residual grids to form initial grids, and placing residual values of sampling points into the initial grids, wherein the relative positions of the residual values in the initial grids are kept unchanged; in the process of putting each residual value into the initial grid, if the cells in which the residual values are put have position conflicts, the residual values which are needed to be put in the back are placed at the positions of the adjacent cells in the conflict positions; the residual values of the points to be sampled are all put into the scaling residual grid;
s3, assigning the scaled residual grids into full-value scaled residual grids, performing sub-sampling treatment, and performing iterative updating on the full-value scaled residual grids by using a Bellman equation until convergence to obtain predicted values of the Bellman equation on sampling point residuals; and adding the predicted value of the BP neural network to the soil nutrient attribute of the sampling point and the predicted value of the Belman equation to the residual error of the sampling point to obtain a final spatial interpolation predicted result of the soil nutrient.
2. The method of claim 1, wherein the activation function used by the hidden layer in the BP neural network is a ReLU function:
f(x)=max(0,x)
wherein,,f(x)to activate the value of the function, max (0,x) To take inputxAnd a maximum value between 0.
3. The method of claim 1, wherein the loss function used in the BP neural network is a square loss function:
L(Y,f(X))=(Y-f(X)) 2
wherein,,Lin order to be lost, the process is carried out,Yas a result of the actual output value,f(X) Is a model predictive value for the input variable X.
4. The method for predicting the spatial interpolation of soil nutrients according to claim 1, wherein the specific operation mode of step S13 is as follows: equally dividing the residual data set into 5 parts, namely a group A, a group B, a group C, a group D and a group E, and mixing and recombining the five groups into a BCD group, an ACD group, an ABD group and an ABC group; training BP neural network by using BCD group to obtain predicted value of A groupy A And first prediction of group Ey E1 The method comprises the steps of carrying out a first treatment on the surface of the Similarly, the predicted values of group B are obtained successivelyy B And second prediction value of group Ey E2 Predicted value of group Cy C And third prediction value of group Ey E3 Predicted value of group Dy D And fourth predictive value of group Ey E4 The method comprises the steps of carrying out a first treatment on the surface of the Then, through the true values of group A, group B, group C and group DA、B、C、DObtaining residuals of four groups of A, B, C and Dε A 、ε B 、ε C 、ε D The method comprises the steps of carrying out a first treatment on the surface of the Residual error of group Eε E Temporarily setting the sampling point as 0, integrating longitude and latitude data and corresponding residual errors of sampling points of five groups of A, B, C, D and E into a residual error data set in the mode of [ + ]A 1 ,A 2 ,ε) The method comprises the steps of carrying out a first treatment on the surface of the Wherein, the method comprises the following steps ofA 1 ,A 2 ) Is the longitude and latitude data of the sampling point,εis the residual of the sample point.
5. The method for spatial interpolation prediction of soil nutrients according to claim 1, wherein the number of samples at which residual values of position conflicts are shifted in step S23 is controlled to be within 2% of the total number of cells in the grid.
6. The method for predicting the spatial interpolation of soil nutrients according to claim 1, wherein the specific operation mode of the step S3 is as follows:
s31, assigning values to the cells without residual values in the scaling residual grids to form a full-value scaling residual grid, wherein the assigned values are obtained by adding all residual values which are already placed in the scaling residual grid and dividing the added values by the total number of the cells in the scaling residual grid;
s32, sub-sampling is carried out on the full-value scaling residual grids to obtain space weight parameters, and the sub-sampling proportion of the full-value scaling residual grids is randomly determined in the residual grids constructed in the step S22; assigning a random number between 0 and 1 to each cell in the full-value scaling residual grid; if the random number is smaller than the sub-sampling proportion number, reserving a filling value in the cell; if the random number is greater than the sub-sampled proportional number, replacing the fill value in the cell with a null value; after filling values in all cells of the full-value scaling residual grid are processed in this way, generating a sub-grid with a specified sub-sampling proportion; assigning values to the vacant cells in the sub-grids, wherein the assigned values are obtained by adding filling values reserved in the sub-grids and dividing the filling values by the total number of the cells in the sub-grids; determining a space weight factor gamma of the sub-grid in a random search parameter mode, and taking the space weight factor gamma as a space weight factor gamma of a full-value scaling residual grid;
s33, in the Markov process, the state transition probability that the state S at a certain moment t reaches another state S' at the next moment t+1 is as follows:
P SS' =P[S t+1 = S'|S t =S]
wherein, P is the probability,SandS t the state is that the time is t,S'andS t+1 are alltA state at time +1;
s34, iteratively updating the full-value scaling residual error grid by using a Belman equation until convergence, and obtaining a predicted value of the sampling point residual error after iteratively updating by using the Belman equation:
v(s) = E[R t+1 +γv(s t+1 )|s t =s]
wherein R is t+1 For immediate rewards to reach a certain state, gamma is a spatial weighting factor,v(s t+1 ) A subsequent state cost function added with a space weight factor gamma;
and S35, adding the predicted value of the BP neural network to the soil nutrient attribute at a certain sampling point and the predicted value of the Belman equation to the residual error of the sampling point, and taking the added result as a final spatial interpolation predicted result of the soil nutrient.
CN202310386123.2A 2023-04-12 2023-04-12 Spatial interpolation prediction method for soil nutrients Pending CN116383589A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310386123.2A CN116383589A (en) 2023-04-12 2023-04-12 Spatial interpolation prediction method for soil nutrients

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310386123.2A CN116383589A (en) 2023-04-12 2023-04-12 Spatial interpolation prediction method for soil nutrients

Publications (1)

Publication Number Publication Date
CN116383589A true CN116383589A (en) 2023-07-04

Family

ID=86964153

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310386123.2A Pending CN116383589A (en) 2023-04-12 2023-04-12 Spatial interpolation prediction method for soil nutrients

Country Status (1)

Country Link
CN (1) CN116383589A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117540409A (en) * 2024-01-10 2024-02-09 中化现代农业有限公司 Soil sampling sample point encryption method and device, electronic equipment and storage medium

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117540409A (en) * 2024-01-10 2024-02-09 中化现代农业有限公司 Soil sampling sample point encryption method and device, electronic equipment and storage medium
CN117540409B (en) * 2024-01-10 2024-04-19 中化现代农业有限公司 Soil sampling sample point encryption method and device, electronic equipment and storage medium

Similar Documents

Publication Publication Date Title
CN109492822B (en) Air pollutant concentration time-space domain correlation prediction method
Matsumura et al. Maize yield forecasting by linear regression and artificial neural networks in Jilin, China
CN105243435B (en) A kind of soil moisture content prediction technique based on deep learning cellular Automation Model
CN109478045A (en) Goal systems is controlled using prediction
CN110689179A (en) Water bloom prediction method based on space-time sequence mixed model
CN112446559B (en) Large-range ground subsidence space-time prediction method and system based on deep learning
CN113379107B (en) Regional ionosphere TEC forecasting method based on LSTM and GCN
CN107423561A (en) A kind of evaluation method of soil attribute interpolation
CN107860889A (en) The Forecasting Methodology and equipment of the soil organism
CN113324923B (en) Remote sensing water quality inversion method combining space-time fusion and deep learning
Kişi Evolutionary fuzzy models for river suspended sediment concentration estimation
CN114091766B (en) CEEMDAN-LSTM-based space load prediction method
CN110909447B (en) High-precision short-term prediction method for ionization layer region
CN116383589A (en) Spatial interpolation prediction method for soil nutrients
CN114881323A (en) Foundation pit dewatering area underground water level prediction and updating method based on deep neural network
CN114723188A (en) Water quality prediction method, device, computer equipment and storage medium
Nguyen Multi-objective automatic calibration of hydrodynamic models-development of the concept and an application in the Mekong Delta
CN113901348A (en) Oncomelania snail distribution influence factor identification and prediction method based on mathematical model
CN112668606A (en) Step type landslide displacement prediction method based on gradient elevator and quadratic programming
Tsai et al. From parameter calibration to parameter learning: Revolutionizing large-scale geoscientific modeling with big data
CN117131991A (en) Urban rainfall prediction method and platform based on hybrid neural network
Yang et al. Predictor selection for CNN-based statistical downscaling of monthly precipitation
CN115182398B (en) Groundwater level and earth surface subsidence prediction method for earthquake early warning area
Sampson et al. Operational evaluation of air quality models
Masoomi et al. Modeling uncertainties in sodium spatial dispersion using a computational intelligence-based kriging method

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination