CN113158565B

CN113158565B - Artificial intelligence method, system and electronic equipment for acquiring water-nitrogen strategy

Info

Publication number: CN113158565B
Application number: CN202110430308.XA
Authority: CN
Inventors: 吴宗翰; 胡睿琦; 潘平波; 于叶露; 娄帅
Original assignee: Inner Mongolia Hengyuan Water Engineering Co ltd; Lanzhou Lifeng Zhengwei Intelligent Technology Co ltd
Current assignee: Inner Mongolia Hengyuan Water Engineering Co ltd; Lanzhou Lifeng Zhengwei Intelligent Technology Co ltd
Priority date: 2021-04-21
Filing date: 2021-04-21
Publication date: 2023-05-23
Anticipated expiration: 2041-04-21
Also published as: CN113158565A

Abstract

The invention relates to the technical field of agriculture and provides an artificial intelligence method, an artificial intelligence system and electronic equipment for acquiring a water nitrogen strategy, wherein the method comprises the following steps: acquiring a target value of each parameter related to soil fertility at a preset time point; inputting a preset time point, a preset depth and all target values into a preference neural network with a preference mechanism to obtain a water-nitrogen strategy comprising target water filling quantity and target nitrogen application quantity, wherein the preference mechanism is used for representing the relation between each parameter and the water filling quantity and the nitrogen application quantity respectively, and the preference mechanism used for representing the relation between each parameter and the water filling quantity and the nitrogen application quantity respectively is firstly provided, so that the obtained preference neural network with the preference mechanism can naturally integrate the neural network structure with the priori rule of data in the nature, the convergence of the model, namely the preference neural network, can be greatly improved, the prediction precision can be effectively improved, and the optimal water-nitrogen strategy can be obtained.

Description

Artificial intelligence method, system and electronic equipment for acquiring water-nitrogen strategy

Technical Field

The invention relates to the technical field of agriculture, in particular to an artificial intelligence method, an artificial intelligence system and electronic equipment for acquiring a water nitrogen strategy.

Background

The method aims at improving the soil fertility of the root zone of the crop, and accurately provides a water nitrogen optimizing strategy under deep burying of the straw, so that the method has great significance in improving the land productivity of the river set irrigation zone, relieving the non-point source pollution and promoting the precise agricultural development. However, in the current research method, no matter field test, physical model or artificial intelligence means are adopted, most of the methods only can reveal the influence trend on soil and plant indexes through limited water nitrogen treatment, and then the optimized water nitrogen rule is obtained, so that the obtained water nitrogen strategy is generally and inaccurately.

Disclosure of Invention

The invention aims to solve the technical problem of providing an artificial intelligence method, an artificial intelligence system and electronic equipment for acquiring a water nitrogen strategy aiming at the defects of the prior art.

The technical scheme of the artificial intelligence method for acquiring the water-nitrogen strategy is as follows:

acquiring a target value of each parameter related to soil fertility at a preset time point;

inputting a preset time point and all target values into a preference neural network with a preference mechanism to obtain a water-nitrogen strategy comprising target water filling quantity and target nitrogen application quantity, wherein the preference mechanism is used for carrying out focused training on the neural network learning parameters according to objective relations between each parameter and the water filling quantity and the nitrogen application quantity respectively.

The strategy for acquiring the water nitrogen has the following beneficial effects:

the preference mechanism for representing the relation between each parameter and the water filling amount and the nitrogen application amount is provided for the first time, the obtained preference neural network with the preference mechanism can naturally integrate the neural network structure with the priori rule of the data in the nature, the convergence of the model, namely the preference neural network, is greatly improved, the prediction precision is effectively improved, and the optimal water-nitrogen strategy is obtained.

The technical scheme of the artificial intelligence system for acquiring the water-nitrogen strategy is as follows:

the device comprises a first acquisition module and a second acquisition module;

the first acquisition module is used for acquiring a target value of each parameter associated with soil fertility at a preset time point;

the second acquisition module is used for inputting a preset time point and all target values into a preference neural network with a preference mechanism to obtain a water-nitrogen strategy comprising target water filling quantity and target nitrogen application quantity, wherein the preference mechanism is used for representing the relation between each parameter and the water filling quantity and the nitrogen application quantity respectively.

The artificial intelligence system for acquiring the water-nitrogen strategy has the following beneficial effects:

The technical scheme of the electronic equipment is as follows:

comprising a memory, a processor and a program stored on the memory and running on the processor, the processor implementing the steps of an artificial intelligence method for obtaining a water nitrogen policy as described in any one of the preceding claims when the program is executed.

Drawings

FIG. 1 is a schematic flow chart of an artificial intelligence method for obtaining a water-nitrogen strategy according to an embodiment of the invention;

FIG. 2 is a schematic diagram of a structure of a preferred neural network;

FIG. 3 is a schematic diagram of a Sigmoid function;

FIG. 4 is a schematic diagram of the derivative of a Sigmoid function;

FIG. 5 is a schematic diagram of a ReLU function;

FIG. 6 is a schematic diagram of the derivative of the ReLU function;

FIG. 7 is a schematic diagram of a structure of a neural network embedded in a preference matrix;

FIG. 8 is a graph showing the result of predicting soil organic matter content;

FIG. 9 is a graph showing the result of predicting the total nitrogen content of soil;

FIG. 10 is a graph showing the results of predicting the salt content of soil;

FIG. 11 is a graph showing the predicted pH;

FIG. 12 is a graph of EC value versus salt content;

FIG. 13 is a curved surface of the total amount of soil organic matters at the 17 th day after sowing, i.e., at the seedling stage;

FIG. 14 is a curved surface of the total amount of soil organic matters at the time of the jointing period at the 59 th day after sowing;

FIG. 15 is a curved surface of the total amount of soil organic matters at the time of the grouting period, which is 87 days after sowing;

FIG. 16 is a curved surface of the total nitrogen content of soil at 17 days after sowing, i.e., at the seedling stage;

FIG. 17 is a curved surface of total nitrogen content of soil at day 59 after sowing, i.e., the jointing period;

FIG. 18 is a graph showing the total nitrogen content of soil at day 87 after sowing, i.e., the grouting period;

FIG. 19 is a curved surface of the salt content of soil at 17 days after sowing, i.e., at the seedling stage;

FIG. 20 is a curved surface of soil salinity at day 59 after sowing, i.e., at the jointing stage;

FIG. 21 is a curved surface of soil salinity at day 87 after seeding, i.e., the grouting period;

FIG. 22 is a graph showing the pH at 17 days after sowing, i.e., at seedling stage;

FIG. 23 is a graph showing the pH at day 59 after seeding, i.e., the jointing period;

FIG. 24 is a graph showing pH at day 87 after seeding, i.e., the grouting period;

FIG. 25 is a graph of upper and lower limits of target fill volume for an optimal water-nitrogen strategy;

FIG. 26 is a graph of upper and lower limits of target nitrogen delivery for an optimal water nitrogen strategy;

FIG. 27 is a schematic diagram of an artificial intelligence system for obtaining a water nitrogen strategy according to an embodiment of the present invention.

Detailed Description

As shown in fig. 1, an artificial intelligence method for obtaining a water nitrogen strategy according to an embodiment of the present invention includes the following steps:

s1, acquiring a target value of each parameter related to soil fertility at a preset time point;

s2, inputting a preset time point and all target values into a preference neural network with a preference mechanism to obtain a water-nitrogen strategy comprising target water filling quantity and target nitrogen application quantity, wherein the preference mechanism is used for representing the relation between each parameter and the water filling quantity and the nitrogen application quantity respectively.

Wherein the plurality of parameters associated with soil fertility includes: soil organic matter content (Soil organic matter content), soil total nitrogen content (Soil Total nitrogen content), soil salt content (Soil salt content), pH value and the like.

Preferably, in the above technical solution, the method further includes:

s020, building a neural network;

s021, obtaining and normalizing preference values between each parameter and the water filling amount and the nitrogen application amount respectively to obtain a preference matrix, introducing the preference matrix into the neural network, and training by combining with an Adam algorithm to obtain the preference neural network with a preference mechanism.

Preferably, in the above technical solution, the building a neural network further includes:

s0201, taking the ReLU function as an activation function of the neural network, and taking the ReLU function as the activation function of the neural network, solves the gradient disappearance problem of the neural network.

s0202, introducing a Dropout algorithm into the neural network;

s0203, introducing batch standardization into the neural network.

The Dropout algorithm and batch standardization are introduced into the neural network, so that the error risk and the operation cost are effectively reduced, and the operation efficiency is improved.

An artificial intelligence method for obtaining a water nitrogen strategy of the present application is described in detail by one embodiment:

s10, selecting a test field, specifically:

the test is carried out in an inner Mongolia river set irrigation area in 4 months to 10 months in 2017, and is carried out in a double river town agriculture comprehensive water saving demonstration area in the inner Mongolia river set irrigation area, wherein the test area belongs to a Yongji irrigation area (40 DEG 42 'in North latitude and 107 DEG 24' in east longitude), and belongs to a semi-arid continental climate in a medium temperature zone, the average precipitation amount for years is 138mm, the average evaporation amount for many years is 2332mm, the rainfall is mostly concentrated in summer and autumn, the surface salt returning in spring and winter is serious, and the rainfall in the corn growth period in 2017 is 75.3mm respectively. The soil with 0-40cm of the test area belongs to silt loam, and the average volume weight is 1.42-1.53 g.cm ^-3 The soil physical properties of the moderate saline-alkali soil are shown in the following table 1:

TABLE 1

S11, test design. Specifically:

the corn of the main crop planted in the inner Mongolia river set irrigation area is used as a test material, and a local conventional variety (Simon 168) is adopted to carry out field test. The test is carried out under moderately saline soil (EC value is between 0.7 and 1.24 ms/cm), a corn straw layer is buried in a depth of 40cm soil layer, the thickness is 5cm, then the soil is leveled, the atrazine herbicide is selected as the pesticide, and the application amount is conventional in the local area. In addition to autumn irrigation and spring irrigation, yellow river water is irrigated for 3 times in the whole growth period of corn, the irrigation mode is furrow irrigation, 3 irrigation levels are set, and irrigation rates, namely irrigation amounts are 60mm, 90mm and 120mm respectively; applying fertilisers Urea (46% N) and diammonium phosphate (18% N,46% P2O 5) were used at 4 levels of nitrogen application, 135kg/hm each ² 、180kg/hm ² 、225kg/hm ² 、280kg/hm ² . In the test, 50% of nitrogenous fertilizer and all phosphorus and potassium fertilizers are taken as a base Shi Shi before sowing, and the residual nitrogenous fertilizer is respectively applied before the secondary water and the tertiary water according to 25% of equal quantity each time; the test had 13 treatments, 3 replicates, 39 cells, 8×9=72m per cell area ² The periphery is separated by polyethylene plastic films with the burial depth of 1.2m, the top is left for 30cm, fertilizer and water channeling of each cell is prevented, and the field management is consistent with the local farmer management. Corn is covered with mulch film and sown in the last ten days of 4 months, and harvested in the middle and last ten days of 9 months. Corn film width 1.1m, one film 2 rows, plant spacing about 45cm, row spacing 35cm, corn planting density 6 ten thousand plants/hm ² . The specific design is shown in the following table 2:

numbering device	Treatment of	Irrigation quota (mm)	Nitrogen application amount (kg/hm) ² )
				1	W ₁ N ₁	60	135
2	W ₁ N ₂	60	180
				3	W ₁ N ₃	60	225
4	W ₁ N ₄	60	280
				5	W ₂ N ₁	90	135
6	W ₂ N ₂	90	180
				7	W ₂ N ₃	90	225
8	W ₂ N ₄	90	280
				9	W ₃ N ₁	120	135
10	W ₃ N ₂	120	180
				11	W ₃ N ₃	120	225
12	W ₃ N ₄	120	280
				13	Local area	135	325

TABLE 2

S13, observation and measurement, specifically:

comprehensively evaluating soil fertility by four indexes, wherein the four indexes comprise: the four parameters are soil organic matter content, soil total nitrogen content, soil salt content and pH value. The sampling and measuring method is as follows:

s130, before the test starts, 5 layers of soil samples of 0-20cm, 20cm-40cm, 40cm-60cm, 60cm-80cm and 80cm-100cm are drilled by soil in each treated district, the depth is 100cm, and each layer is repeated three times. Measuring the soil volume weight after taking the soil, and analyzing the particle size of soil particles;

S131, taking 5 layers every 20cm in the whole growth period of crops, wherein the depth is 100cm. Soil samples of soil layers of 0-20cm, 20cm-40cm, 40cm-60cm, 60cm-80cm and 80cm-40cm are respectively collected by an earth auger in a sampling area by adopting a three-point method, and are refrigerated and stored at the temperature of 4 ℃ to measure the organic matter content, the total nitrogen content, the salt content and the pH value of the soil, specifically:

the total nitrogen content of soil is measured by a KDN-AA double-tube nitrogen determination instrument, the organic matter content of soil is measured by a MWD-2 microwave universal digestion device, the salt content of soil is measured by a TU1810PC type ultraviolet-visible spectrophotometer, and the pH value is measured by a TU18950 double-beam ultraviolet-visible spectrophotometer.

S14, defining preference values, specifically:

in the past, in the field of water conservancy, the pearson correlation coefficient r (X, Y) method is generally used to express the correlation between two variables, and the formula is:

where r (X, Y) represents the PEARSON correlation coefficient, which was originally designed by the collectist Called PEARSON design (PEARSON, E.S. et al, 1955), is a quantity used to study the degree of linear correlation between variables X, Y, var [ X ]]Variance of X, var [ Y ]]Variance of Y;

wherein Cov (X, Y) is the covariance of X and Y, representing the overall error of two variables, the formula is:

Assuming that both sets of data X and Y contain n elements, E represents the expected value of X, Y, it can be seen that Cov (X, Y) is positive if the variable X, Y trend is consistent (e.g., greater than the expected value of itself), whereas it is negative if X, Y distribution has no correlation (X) _i -E(X _i ))(Y _i -E(Y _i ) Positive and negative, resulting in Cov (X, Y) eventually approaching 0. If X, Y there is a significant correlation, then

The absolute value is large, so that Cov (X, Y) tends to be 1;

if all sample correlation coefficients are symmetrically distributed, the significance of r can be performed by t-test. The significance of the correlation between X, Y is commonly denoted by P<0.05 represents that X is significantly correlated with Y, P<0.01 represents that X is extremely significantly correlated with Y; p is required to be obtained according to a table look-up of t value under a specific sample capacity, and the formula of the t value is as follows:

where n is the number of samples and n is a positive integer, r is the pearson correlation coefficient, and the degree of freedom is n-2 due to the presence of two variables X, Y.

In order to improve the convergence of the redundant fully-connected neural network, the preference value is used as a statistic for calculating the dependency degree among different indexes in the farmland water conservancy, and is a centralized representation of the correlation and the significance. The preference value is based on the pearson correlation coefficient, and the inherent positive and negative correlation between the variables is further amplified or reduced by combining the significance, so that the interaction degree between the indexes is more remarkably expressed while the correlation is represented, thereby more comprehensively and effectively reflecting the inherent relation between test control factors such as water nitrogen treatment and the like and the indexes such as soil, crops and the like, and the calculation formula of the preference value (Preference Values) is as follows:

Wherein PV _ij As variable X _i And Y _j Preference value, X _i Representing the ith variable or parameter related to the test treatment, such as the level of water filling or the level of nitrogen application or the amount of nitrogen application, Y _j Representing the j-th index variable obtained by test monitoring under different treatments, specifically: soil organic matter content, total nitrogen content of soil, soil salt content, pH value, and soil sampling time point, depth, e is constant and e takes 0.001 for preventing formula denominator from being 0.

S15, normalization of preference values, specifically:

in order to make the preference values among the indexes in the same order of magnitude so as to be more stable and convenient to calculate, the preference values are normalized, and the formula is as follows:

wherein (1)>

Take absolute value, and PV _normal Is determined by the sign of the correlation r.

S16, building a neural network, specifically:

s160, defining an input terminal and an output terminal of the neural network, specifically:

1) Defining the input end of the neural network as a matrix

Where n is the sample size, e.g., n=420; d is the dimension of each input vector, such as d=6, wherein { xi } i=1, …, n e X represents a vectorized set of four indexes, namely parameters, namely time of nitrogen application and irrigation (expressed by days after sowing) of a target value of soil organic matter content, a target value of total nitrogen content, a target value of soil salt content and a target value of pH value for measuring soil fertility, namely a preset time point;

2) Defining the output end of the neural network as a matrix representing the irrigation quantity and the nitrogen application quantity

The goal of the neural network is to learn a fixed mapping +.>

And obtaining a preference neural network, wherein θ is a learned parameter which can be well optimized through PNN training, and the predicted value Y' approaches to the measured value Y infinitely.

S161, obtaining a structure of a neural network, specifically:

the first proposed Preferred Neural Network (PNN) of the present application is a typical Deep Learning (DL) model. PNN can be seen as an approximate natural function describing the complete dependency between soil organic matter content, soil total nitrogen content, soil salinity, pH and irrigation and nitrogen application, respectively. Specifically, the neural network optimizes the function by constructing a map y=f (x, θ) and learning a parameter θ to obtain a preferred neural network, the structure of the neural network is shown in fig. 2, and fig. 2 is also the structure of the preferred neural network; reLU, namely ReLU function, preference Matrix, namely preference matrix, and BatchNorm Layer, namely Layer for batch normalization in FIG. 2;

in the structure of the neural network, the well definition of affine transformation of the interlayer information flow is the key of the training of the neural network model. In general, the academic parameter θ for each layer contains a weight parameter w and a preference parameter b. Implicit representation h of layer l in neural network _l The definition is as follows:

wherein W is _l And b _l Respectively represent the learned l-layer weight and preference variable, h _l-1 Is an implicit representation of the upper layer, h when l=1 ₀ Using the hierarchical update rule, a given input data stream passes through each hidden layer with intermediate operations, finally reaching the output.

The structure of the obtained neural network comprises:

1) The ReLU function is used as an activation function of the neural network, specifically:

the activation functions introduce nonlinear mechanisms into the neural network, so that they can arbitrarily approximate natural functions with complex forms in the distribution range of the input, and the multi-layer structure without the activation functions is equivalent to a single-layer model due to the linear characteristics of the multi-layer structure. However, many current studies fail to employ the correct activation function, and most current studies introduce Sigmoid functions only in shallow models, resulting in suboptimal model performance (Cybenko, 1989). In this section, we will elucidate in detail the reasons why Sigmoid functions are unsuitable in formulating water and fertilizer optimization strategies, and introduce ReLU functions into the proposed neural network, in particular:

(1) the Sigmoid activation function is also called a logistic function, and the mathematical expression is as follows:

figures 3 and 4 show that the range of values for the compression transformation is between 0 and 1, which makes the model prone to a gradient approaching 0 at saturation (those neuron gradients with hidden values approaching 0 or 1 have a value of 0). Obviously, the weighting parameters of saturated neurons cannot be updated, and at the same time, the neurons connected with the saturated neurons propagate slowly Slow, a phenomenon known as gradient extinction. Furthermore, another reason that the Sigmoid function is not favored for application in a deep learning structure is that exp (·) operations in Sigmoid, as applied to each layer of the model, would make the Sigmoid function burdensome in computational complexity.

(2) The ReLU function can effectively cope with the above problem, and FIGS. 5 and 6 show the "half-tuning" feature of the ReLU function, whose mathematical expression is: reLU (x) =max (0, x), then: when x is<At 0, reLU activation causes the output to be 0, otherwise x remains the original value, so that a compact operation will make the propagation and convergence of the entire neural network more efficient. Since the ReLU liberates neurons from boundary constraints, at least a certain number of neurons back-propagate from the positive region, avoiding the problem of gradient extinction. Here we apply ReLU after each hidden layer to do the nonlinear transformation. This operation can be expressed as:

wherein (1)>

Is a hidden representation of the batch standardization, as will be described in more detail below. />

2) Batch normalization was introduced into neural networks, specifically:

while deep neural networks have become very effective due to the excellent generalization capability obtained in deep architecture, they have resulted in too complex training by the mechanism of converting input data through learnable parameters in all neural network layers. In this case, subtle changes in these parameters will be amplified indefinitely as the network goes deeper.

The input distribution of each layer is continuously changed, and each subsequent layer always obeys the updated distribution, so that the stability of the training process and the stability of the model are greatly weakened.

Considering that the preferred neural network PNN with a six-layer structure suffers from the same problems as described above, a batch normalization (Batch Normalization) is utilized to improve the iterative process of training and enhance model stability (Ioffe and Szegedy, 2015). Batch normalization (Batch Normalization) operations are divided into four steps:

(1) setting a mini batch B, and obtaining the average value of all elements in the mini batch B as follows:

where m is the number of samples in each mini-batch;

(2) obtaining a batch mean mu _B Corresponding variance

Specifically by->

Obtain the batch average value mu _B Corresponding variance->

(3) The normalization operation may be expressed as the mean and variance of the batch samples:

where l is a very small but non-0 number set to prevent σ from being equal to 0.

(4) Scaling and shifting the normalized hidden layer representation by two learnable parameters:

gamma represents a scale parameter for improving the expression capacity of the neural network, beta represents a displacement parameter for improving the expression capacity of the neural network, and both gamma and beta can be obtained through model training. The overall demonstration of the batch normalization transformation is as follows:

And (3) making:

minimum lot size l _th Hidden layer representation;

l: the number of layers;

m: the number of samples in each mini-batch;

gamma: scale parameters, obtainable by learning;

beta: the displacement parameters can be obtained through learning;

and (3) outputting:

1.for l＝1,2,3,······,L do；

2. updating the mini-batch mean:

3. updating the mini-batch variance:

4. implicit expression of standardized mini-batches:

5. panning and zooming the normalized implicit representation:

6.end for

3) The Dropout algorithm was introduced into neural networks, specifically:

the limited training data set typically results in complex depth structures having to be forced to adapt to the characteristics of the input data during training, a phenomenon commonly referred to as overfitting or co-adaptation. The Dropout algorithm is a solution to the above problem effectively at low computational cost (with a computational complexity of O (n)) by randomly suspending non-output neurons from a nonlinear model in each training iteration, and is specifically introduced in the form of a Dropout module.

Specifically, the Dropout algorithm multiplies the implicit representations generated by the input layer and hidden layer in the network by sampling a binary mask.Neurons whose output value is multiplied by 0 are temporarily suspended in the current training iteration. The samples for each layer of the mask are uncorrelated and the probability of a value of 1 in the mask is a predefined hyper-parameter of the whole network. The Dropout algorithm can be expressed mathematically in a hierarchical way:

Wherein (1)>

For element point multiplication operation, d _l For the Dropout mask of the first layer, p' is a pre-defined Dropout hyper-parameter before training.

4) A preference structure is introduced, in particular:

the existing model construction strategy based on BP neural network only directly obtains the mapping relation between the input end and the output end by virtue of a relatively simple structure, and essentially ignores the natural operation rule among related factors of agricultural water and soil engineering, so that the performance of the neural network is greatly reduced. In order to exert the priori knowledge advantage of the agricultural water and soil field and the excellent learning advantage of the artificial neural network to the greatest extent, the Preference Neural Network (PNN) provided herein makes use of the inherent correlation among the agricultural water and soil engineering factors to perform model performance optimization.

Specifically, the application firstly proposes the concept of preference values, and the preference matrix calculated based on the concept clearly expresses the inherent dependence relationship between each index of the input end and the water nitrogen allocation strategy. PNN fuses the preference matrix into the neural network structure by Hadamard product (Hadamard product) of the preference matrix with the learnable weights between input and output. In this case, the neural network learns the map based on the a priori knowledge of the agricultural water and soil engineering as guided by the strong constraints of the preference matrix. Referring to the formula in hierarchical affine transformation:

The mathematical expression of the preference constraint of PNN is: h is a _l (h _l-1 ；W _l ,b ₁ )＝h ^T _l-1 W _l ⊙P+b _l Wherein P is ∈R by the formula->

And the preference matrix calculated in table 8. The symbol ". Alt is the Hadamard product between the corresponding elements of the matrix.

For the nodes in the traditional fully-connected neural network, all nodes in the next layer are connected, and further, the iterative updating of the weight value and the bias value of each layer is carried out in the reverse propagation stage of the error. Although fully-connected neural networks have proven to well establish unknown relationships between variables, their overly complex structure tends to slow the model's speed of operation. In addition, the variable with extremely low correlation and the variable with very close relation are interfered by the model because the nodes corresponding to the variable and the variable have equal connection numbers in the fully connected network, and convergence is finally affected.

The preference structure, namely the preference mechanism of the preference neural network PNN can well solve the problems, and the preference structure builds a concept which is derived from a big degree, namely the learning process of the neural network can be manually interfered and guided by changing the connection weight matrix of the corresponding node of the variable at the input end of the model, so that the relation among certain variables is emphasized and the other relations are weakened. Specifically, through correlation between independent variables and dependent variables reflected by measured data, the influence of each independent variable on the dependent variables is primarily evaluated, the relationship between the influence and the influence is embedded into a neural network in a preference matrix form, and finally the convergence speed of the neural network is greatly optimized, as shown in fig. 7;

In fig. 7, the number of connecting lines between different input nodes and hidden layers represents the connection strength obtained by the corresponding variable of the point according to the self preference, and the hadamard product is performed by using the weight matrix and the preference value matrix, so that the connection relationship between each input end node and the hidden layer has the preference according to the natural law, the input end variable (such as soil salt content) with a stronger relationship with the output end obtains the larger connection strength, and the depth of the network in the figure is only for illustrating the preference connection structure, and is not the depth of the PNN used in the experiment.

S17, training, specifically:

(1) reverse training scheme (Reverse Training Scheme), specifically:

in the past, the BP neural network (BP neural network) is used for solving the problem of water nitrogen optimization of farmland water conservancy, and a forward training (Forward Training Scheme) parameter model from treatment to index is established. The neural network model obtained based on the forward training scheme (Forward Training Scheme) is limited by that the input and output ends are respectively an implementation strategy and target parameters, and better treatment is required to be put forward through a relatively large amount of water and fertilizer strategies after the model is established, so that the optimal strategy cannot be directly formulated according to the target value of the required parameters, the end-to-end efficient decision output is realized, and the advantages of the deep learning framework in the artificial intelligence cannot be fully utilized.

Objective causal relationships between farmland water and fertilizer strategies and target parameters are verified by a plurality of field test researches, and the mapping corresponding to the relationships is not changed due to the change of fruit sequence. The essence of the model research is to optimize the actual production strategy through revealing and learning the causal relationship objectively existing in the nature. The neural network trained under the traditional mapping relation cannot directly provide the optimal strategy, and only the target parameter corresponding to a certain strategy can be provided.

The first proposed reverse training scheme (Reverse Training Scheme) is a neural network training scheme which applies a causal relationship reverse mapping relationship to a model training process so as to directly formulate an optimal strategy through target parameters, and solves the problems of low structural efficiency and insufficient representativeness of the proposed strategy caused by a forward training process in decision research, thereby obtaining more efficient decision performance.

(2) -an objective function (Objective Function), in particular:

root Mean Square Error (RMSE) and standard deviation (MSE) which are widely used in the field of irrigation and water conservancy tend to have a negative effect on model performance due to excessively sensitive detection of data outliers. To increase the robustness of the model, we use the Huber loss (also called the smoothed L1 loss) as an objective function for computing the gradient and updating the model herein. Huber losses essentially balance the sensitivity of RMSE with the weak sensitivity of Mean Absolute Error (MAE), which is defined in terms of the segmentation used:

Where k represents a balancer performing decision operations, there is a scaling down penalty value that may scale up the penalty value when the penalty value is greater than k, but less than k.

(3) Optimization, specifically:

unlike the previous research that the repeated iteration with large slope caused by adopting a random gradient descent method (SGD) is not converged and that all neural network parameters can only adopt the problem of uniform learning rate, the Preferential Neural Network (PNN) can stably and smoothly reach a convergence state in a training stage by adopting Adam operation, namely an Adam algorithm. Motion term v _t And the exponentially weighted moving average term s is initialized with 0 as the two parts of Adam's foremost kernel (also called "drain-average term").

Wherein, adam's algorithm is as follows:

and (3) making:

B＝{x ¹ ,…,x ^m y respectively corresponding to ⁽ⁱ⁾ ；

m: the number of samples in a mini-batch;

t: iteration times;

v: a momentum term;

s: average missing items;

β _v : v non-negative super parameters;

β _s : s non-negative super parameters;

epsilon: a stability constant;

initializing:

v＝0，s＝0

1.for t＝0,1,2,......,T do

2. gradient generation:

3.t＝t+1；

4. calculating a motion term: v _t ＝β _v v _t-1 +(1-β _v )g _t ；

5. Calculating a leakage average term:

6. deviation correction of the motion term:

7. deviation correction of the leakage average term:

8. rescaling the fade:

9. updating parameters:

10. ending;

the process of training the resulting preferred neural network is shown in table 5 below:

And (3) making:

B _x ＝{x ¹ ,...,x ^m }: mini-batch (mini-batch) in training dataset;

l is the number of layers;

m: number of samples in mini-batch;

1.for iterator＝1,2,3,······,T do：

2. by corresponding y ⁱ For a micro-batch (B) containing m examples in the training dataset _x ) The sampling is carried out and the sample is taken,

3.for l＝1,2,3,......,L do：

4. using layer-by-layer affine transformation:

5.

6. application of nonlinear activation transformation

7. Application of signal discarding module (Dropout module)

8. Batch standardized application

9.end for

10. Derived loss function gradient update favoring neural networks (PNNs)

11.end for

12. And returning the y' value according to the expected water and fertilizer filling level.

S18, model verification, specifically: model performance is typically assessed using a standard of Root Mean Squared Error Root Mean Square Error (RMSE), mean Squared Error Mean Square Error (MSE) and mean absolute error Mean Absolute Error (MAE). Wherein,,

wherein n represents the number of data pairs, y _i 、/>

Respectively representing a predicted value and an actual value, +.>

Representing the average of the measured values, the range of values for MSE is 0, ++ infinity), the closer the MSE is to 0, the higher the accuracy of the model. RMSE is used as the square of the MSE, and the magnitude of the model precision can be compared more intuitively, and the value range and the precision evaluation standard are consistent with the MSE. MAE is in the range of [0, + ], mae=0 indicates that the analog value completely matches the measured value. For a prediction model perfectly adapted to all samples, it should appear rmse=0, mse=0, mae=0.

The method utilizes 2017 field actual measurement data to train and verify the preference neural network PNN, takes the soil organic matter content, the soil salt content, the pH value and the total nitrogen content of soil which are measured in 0-100cm of soil after three times of watering and fertilizing in the growth period as the input end of a model, predicts the watering amount and the nitrogen application amount, and finally verifies the performance of the model by comparing the obtained simulation value with the actual watering and fertilizing level, namely the actual watering amount and the actual nitrogen application amount. The model training set and the test set are established according to a random sampling method and are established according to a ratio of 4:1, and the model is enabled to be stable and smooth to tend to converge after multiple iterations through Adam operation.

The present application adopts the same data set as the preferred neural network PNN to train the linear SVR, poly SVR, rbf SVR, LR, LOR and traditional BP neural network, and verifies the performance of each model by comparing RMSE, MSE, MAE between models, and the results are shown in the following table 3:

TABLE 3 Table 3

As can be seen from table 6:

1) RMSE, MSE, MAE when the PNN of the preference neural network predicts the irrigation quantity is 0.012, 0.009 and 0.011 respectively, and is 88.78% -99.18%, 89.57% -99.28% and 88.81% -99.25% higher than linear SVR, poly SVR, rbf SVR, LR, LOR and traditional BP neural networks respectively. Therefore, the prediction accuracy of the preference neural network PNN of the application on the irrigation quantity is obviously superior to that of other models.

2) When the preferred neural network PNN predicts the predicted nitrogen application amount, RMSE, MSE, MAE of the PNN is 0.008, 0.007 and 0.007 respectively, compared with linear SVR, poly SVR, rbf SVR, LR, LOR and the traditional BP neural network, 87.47% -98.76%, 92.82% -99.11% and 86.99% -96.78% are respectively higher, and therefore, the prediction accuracy of the preferred neural network PNN to the nitrogen application amount is obviously superior to other models.

In summary, the preferred neural network PNN disclosed by the application has remarkable superiority in the aspect of multidimensional fertility target prediction water nitrogen strategies based on soil organic matter content, total soil nitrogen content, soil salt content, pH value and the like of specific soil. PNN is able to derive a water nitrogen strategy directly by entering specific fertility targets, which benefits from the inverse structure that the preferred neural network PNN of the present application has.

More advantageously, the reverse and forward structures of the preference neural network PNN can be flexibly converted according to actual demands, and the application further shows the simulation condition of the forward PNN model on the soil average value of soil organic matter content, soil total nitrogen content, soil salt content and pH value between 0cm and 40cm under different water nitrogen conditions. The forward PNN model respectively simulates the change condition of each index of soil under different irrigation amounts and nitrogen application amounts of 17 days, 59 days and 87 days after corn sowing, and the irrigation amount and the nitrogen application amount are consistent with the field test setting. The fitting relation between the predicted value and the measured value of each index is shown in fig. 8-11;

The results shown in FIGS. 8-11 demonstrate that: the simulation values and the actual measurement values of the soil organic matter content, the soil total nitrogen content, the soil salt content and the pH value are well matched, the slope of the fitted straight line is approximately 1, and the determination coefficients R2 of the four-index regression analysis are 0.8162, 0.6401, 0.5916 and 0.4763 (P < 0.05) respectively. Among them, forward PNN predicts the organic matter most accurately, but the accuracy of the prediction for pH is slightly poorer than other indexes, and the difference may be related to the stability of four indexes in soil. In conclusion, the dual verification of the forward PNN and the reverse PNN proves the superiority and advancement of the preference neural network PNN in the multi-objective water and fertilizer strategy optimization aspect.

S19, optimizing water nitrogen strategy formulation based on multidimensional fertility targets, and specifically:

1) Soil fertility standard: soil fertility refers to all soil factors that determine crop productivity (Havlin et al 2005). Crop growth is severely affected by soil salinization (Chaves M et al, 2009), soil pH (Rengasamy and Pichu, 2016), soil organic content (Jiang H et al, 2018), and total nitrogen (Harper L A, 1987) content. Specifically:

(1) nitrogen is the most demanding nutrient for plants except carbon and plays a central role in plant metabolism as a component of proteins, nucleic acids, chlorophyll, coenzymes, plant hormones and secondary metabolites (Hawkesford M et al 2012). Its lack will result in slow plant growth, thin plants, slender stalks, small leaves and premature abscission of old leaves. The nitrogen deficiency can decompose chloroplast and inhibit formation of chloroplast, and the green deficiency caused by the nitrogen deficiency is uniformly distributed on the whole leaf blade, and the serious nitrogen deficiency can cause necrosis of the leaf blade. Total nitrogen in soil refers to the sum of various forms of nitrogen content in soil, including organic nitrogen and inorganic nitrogen. As an index capable of intuitively reflecting the nitrogen deficiency in the soil, the total nitrogen in the soil can be used for effectively evaluating the soil fertility.

(2) Soil organic matters can release and fix soil nutrients, increase farmland production potential, soil water holding capacity and effectiveness of soil nitrogen on plant generation, and improve soil structure (Johnston a E, 2007). Therefore, soil organic matter is widely used in evaluation of soil fertility in the field.

(3) The pH is a measure of the concentration of hydrogen ions, more precisely of the activity of hydrogen ions. In order to maintain normal metabolic function, the pH of the plant root dermis cytoplasm must be maintained between 7.0 and 7.5. Plant growth is inhibited by H+ concentrations in the cytoplasm being higher or lower than those of the cell wall and the outside (Marschner H, 1991). The obvious difference between plant populations on alkaline soil and acidic soil is due to H ⁺ And OH (OH) ^- The ion concentration interferes with plant metabolism. Therefore, the pH value is one of the most valuable indexes for evaluating soil fertility.

(4) Soil salinization is an important factor determining crop productivity, and the salt content is widely applied to a soil fertility evaluation system as an important index for representing the soil salinization degree. Na+ and Cl-ions in soil salts can cause osmotic stress on plants and are prone to nutrient imbalance ultimately affecting plant growth (Sairam R K and Tyagi A et al., 2004).

Therefore, the research comprehensively characterizes the influence rules of different water-nitrogen strategies on the soil fertility by four indexes of soil organic matter content, soil total nitrogen content, soil salt content and pH value, namely, the influence rules of different water-nitrogen strategies on the soil fertility are characterized by four parameters related to the soil fertility, namely, the soil organic matter content, the soil total nitrogen content, the soil salt content and the pH value.

The method is characterized in that four indexes of soil organic matter content, total nitrogen content, salt content and pH value are classified and graded according to a soil formula fertilization test result (see Table 7) of an inner Mongolia river cover irrigation area 3414 and an inner Mongolia autonomous region local standard DB 15/T1086-2016 farmland soil fertility grading technical specification.

Soil organic matter content, total nitrogen content and salt content of soil corresponding to soil fertility grading in soil formulated fertilization test result are shown in tables 4 and 5:

grade	Acute deficiency	Lack of supply	In (a)	High height
					Organic matter (g.kg) ^-1 )	Below 10	10-20	20 or more
Total nitrogen (g.kg) ^-1 )	<0.41	0.41-0.87	0.87-1.60	>1.60

TABLE 4 Table 4

Grade	Non-salinized soil	Light salinized soil	Middle salinized soil	Heavy salinized soil	Saline soil
						Concentration (g.kg) ^-1 )	<2	2-4	4-6	6-10	>10

TABLE 5

The pH value is classified into 4 grades according to the membership degree of the pH value to the soil productivity evaluation in an evaluation system of soil productivity in a classification technical Specification of DB 15/T1086-2016 cultivated land force and the like, and the membership degree is specifically shown in Table 6:

Grade	1	2	3	4
					pH value of	>8.5	＜6.5	7.5-8.5	6.5-7.5

TABLE 6

According to the method, a mathematical statistical relationship between the measured salt contents of 105 soil samples in a mathematical statistical analysis test area and the electrical conductivity Ec1:5 of the soil leaching liquor is adopted, SPSS software is used for analysis, a linear relationship between the Electrical Conductivity (EC) of the soil and the salt content of the soil is established, and a statistical formula (R2=0.988) is obtained as shown in FIG. 12: s is S _i ＝2.591E _C1:5 +0.4682, where S _i Is the salt content (g.kg) of the i-th layer soil ^-1 ) I.e. the mass fraction of salt; ec (Ec) _1∶5 The conductivity (ms/cm) of the soil leaching solution is 1:5 of the soil water ratio. On the basis, the soil salt content of each soil layer is obtained through the soil conductivity values (EC) of different water nitrogen modes monitored by field tests.

Before the field test starts, the soil fertility evaluation index related to the study is subjected to initial sample measurement, and the nutrient content of each soil layer is shown in table 7:

depth of	Organic matter (g.kg) ^-1 )	Total nitrogen (g.kg) ^-1 )	Full salt content (g.kg) ¹ )	pH
					0-20	19.52	1.04	1.08	8.20
20-40	13.95	0.86	0.90	8.15
					40-60	13.17	0.68	1.09	8.10
60-80	9.72	0.59	1.25	8.15
					80-100	10.57	0.55	1.37	8.04

TABLE 7

It can be seen that the initial organic matter content of each soil layer in the test area is shown to be absent except 60cm-80cm, the rest soil layers are at a medium level, and all soil layers of the total nitrogen of the initial soil are at a medium level. The initial salinization level of the test area is lower, the test area belongs to non-salinized soil, and the pH value grade is grade 3. The water filling and fertilization has obvious influence on the four fertility indexes, and the aim of adopting the optimized water nitrogen mode is to: the negative effects of fertilization and irrigation on the soil are maintained at a low level while the nutrient content of the soil is increased. Namely: the total nitrogen content of the soil organic matters and the soil reaches 20 g.kg ^-1 And 1.6 g.kg ^-1 The salt content of the soil is reduced to 2 g.kg ^-1 Hereinafter, the pH is maintained between 6.5 and 7.5.

2) And (3) simulating a water and fertilizer strategy under the change of fertility indexes:

PNN is used as a deep neural network with six-dimensional input vectors, and the final output water-nitrogen strategy comprehensively considers the factors of irrigation time and fertilization time, namely preset time point, soil organic matter content, soil total nitrogen content, soil salt content, pH value and the like, so that the multidimensional coupling change mechanism exceeds the capability range which can be displayed by the image. In order to embody scientific rules, five dimensions of 6-dimensional variables in an input model are set as a set value in the model simulation training process, and then the change condition of a corresponding water-nitrogen strategy caused by single fertility index change, namely parameter change, is revealed. Specifically:

the fertigation time was set to 17, 59 and 87 days after sowing, consistent with the field test, that is, the preset time point was set to 17, 59 and 87 days after sowing; taking the average value of 0-40cm of the corn root system layer by the soil depth; and defining three of the soil organic matter content, the soil total nitrogen content, the soil salt content and the pH value as fixed values, namely determining a target value, and researching the corresponding water and fertilizer strategy obtained by the model when the rest parameter is at different levels. Wherein the target value of the organic matter content of the soil is 20 g.kg ^-1 The target value of the total nitrogen content of the soil is 1.6g.kg ^-1 The target value of the salt content in the soil is 2 g.kg ^-1 The target value of the pH was 7.5. Specifically:

(1) the target value of the total nitrogen content of the soil is set to be 1.6g.kg ^-1 Setting the target value of the salt content of the soil to 2 g.kg ^-1 Setting the target pH value at 7.5 and the target organic matter content in the simulated soil at 5-30 g.kg ^-1 As shown in figures 13-15, the target water filling amount and the target nitrogen application amount of the obtained water-nitrogen strategy generally show an increasing trend along with the increase of the target value of the organic matter content of the soil, but the trend reaches 28 g.kg in total nitrogen ^-1 And substantially stop. This is because irrigation helps to accumulate organic carbon, reduce the decomposition rate, and nitrogen application can reduce the loss of organic matters. The organic matter content of the soil is 20 g.kg ^-1 In the above process, three target water filling amounts of corn in seedling stage, jointing stage and grouting stage are 84.36-120.03mm, 87.94-110.94mm and 90.16-114.67mm, respectively, and target nitrogen applying amounts are 95.11-131.57kg·hm, respectively ^-2 、86.70-99.27kg·hm ^-2 And 71.73-110.13 kg-hm ^-2 . The nitrogen application amount tends to decrease along with the advancing of the growing period, which is probably due to the fact that more water and nitrogen are needed for straw decomposition in the soil in the early growing period, and the soil organic matters are relatively high in the late growing period due to the fact that the straw is mostly decomposed, and the influence of fertilization on the index is gradually reduced.

(2) Setting a target value of the organic matter content of the soil to 20 g.kg ^-1 The target value of the salt content of the soil is set to be 2 g.kg ^-1 Setting the target value of pH value to 7.5, and setting the target value of the total nitrogen content of the simulated soil to 0.5-2 g.kg ^-1 The corresponding water nitrogen strategy changes in the range, as shown in fig. 16-18, in the range of the nitrogen application amount and the irrigation amount given by the PNN, the total nitrogen content of the soil in the growing period generally tends to increase along with the increase of the nitrogen application amount and the irrigation amount. The difference is that the upper limit value of the total nitrogen content of the soil at 17d is 1.74 g.kg ^-1 About, the nitrogen application amount and the water irrigation amount are continuously increased, the lifting effect is not obvious, and the total nitrogen content of the soil can be increased to 2.00 g.kg by increasing the nitrogen application amount and the water irrigation amount at 59d and 87d ^-1 Left and right. This is probably due to the fact that at 17d the crop is in the seedling stage, the crop root layer in the 0-40cm soil is not fully developed, the soil porosity is lower, and the nitrogen holding capacity is poorer than that of the corn in the jointing stage and the milk ripening stage. The total nitrogen in the soil is 1.6g.kg ^-1 In the above cases, the three times of target irrigation amounts are 89.30-97.30mm, 86.31-90.76mm, 90.78-93.24mm, and the target nitrogen application amounts are 93.65-98.89mm, 81.38-93.74mm, 76.13-79.12kg·hm ^-2 . The upper and lower limit values of the irrigation quantity are firstly reduced and then increased along with the promotion of the growth period, and the upper and lower limit values of the nitrogen application quantity are gradually reduced.

(3) Setting a target value of the organic matter content of the soil to 20 g.kg ^-1 The target value of the total nitrogen content of the soil was set to 1.6g.kg ^-1 Setting the target value of pH value to 7.5 and the target value of simulated soil salt content to 1.0-3.0 g.kg ^-1 The corresponding water-nitrogen strategy changes in the range, as shown in fig. 19-21, show that the soil salinity changes represent obvious water-nitrogen interaction effects. When the nitrogen application amount is at a low level, the soil salt content does not significantly change or shows a decreasing trend with the increase of the irrigation amount, as shown in fig. 20, whereas when the nitrogen application amount is high, the soil salt content slightly increases with the increase of the irrigation amount. This may be due to the low nitrogen application rate of irrigation which tends to leach the nitrogen fertilizer below 0-40cm, and the crop root layer has a smaller salt content due to the excessive fertilization. More fertilizer is easy to accumulate on the surface layer under high fertilization level, so that salt accumulation is caused. The salt content of the soil is 2 g.kg ^-1 In the following, the three target irrigation amounts should be 85.18mm-92.27mm, 87.06mm-94.93mm and 78.40mm-94.61mm, respectively, and the three target nitrogen application amounts should be 86.42-103.68, 80.78-86.96 and 69.71-83.10kg·hm, respectively ^-2 . As can be seen, the upper limit value of the irrigation quantity is increased and then decreased along with the advancement of the growth period, the lower limit value is increased and then slightly decreased, and the upper limit value and the lower limit value of the nitrogen application quantity are gradually decreased.

(4) Setting a target value of the organic matter content of the soil to 20 g.kg ^-1 The target value of the total nitrogen content of the soil was set to 1.6g.kg ^-1 The target value of the salt content of the soil is set to be 2 g.kg ^-1 The target value of the pH value of the simulated soil is 6.5-8.5g.kg ^-1 The corresponding water nitrogen strategy varies within the range. As shown in fig. 22-24, the water nitrogen strategy corresponding to the target value of the pH value of the soil with the whole growth period of 0-40 cm. The pH value of the soil is obviously influenced by the interaction of water and nitrogen, when the nitrogen application amount of the soil is smaller, the pH value of the soil is less influenced by the irrigation level, and when the nitrogen application amount is about 17dThe nitrogen content is larger and the trend of firstly decreasing and then increasing is presented, so that the pH value of the soil in the early stage of fertility is obviously increased by the high-nitrogen low-water. The soil pH of 59d and 87d generally increased with the amount of irrigation and nitrogen application. The pH value increases slowly and then quickly at 59d, and finally tends to stabilize, and increases rapidly and then tends to stabilize at 87d, and decreases slightly after reaching the peak value. When the pH value of the soil is maintained at 6.5-7.5, the three times of target irrigation amounts are 77.94-90.68, 76.41-88.61 and 81.65-91.71mm respectively, and the target nitrogen application amounts are 89.88-99.32, 73.37-86.70 and 70.50-78.38 kg-hm-2 respectively. The upper and lower limit values of the irrigation quantity are firstly reduced and then increased along with the promotion of the growth period. The upper and lower limits of the nitrogen application amount gradually decrease along with the promotion of the growth period.

3) Optimizing water nitrogen strategy formulation, specifically:

according to the method, the water and nitrogen optimization ranges when the soil organic matter content, the soil total nitrogen content, the soil salt content and the pH value reach the high fertility standard are respectively obtained through the water and nitrogen interaction curved surfaces, and the water and nitrogen optimization ranges of four indexes are intersected to obtain the optimal water and nitrogen strategy. Specifically:

(1) as shown in fig. 25, the range of the target irrigation amount on the 17 th day after sowing is: 89.30-90.68mm; the range of target irrigation at day 59 after sowing was: 87.91mm-88.61mm; the range of target irrigation at 87 days after sowing is: 90.78mm-91.70mm;

(2) as shown in fig. 26, the range of target nitrogen application amounts on the 17 th day after sowing is: 95.11-98.89kg hm ^-2 The method comprises the steps of carrying out a first treatment on the surface of the The range of target nitrogen application at day 59 post-sowing was: 86.69-86.70 kg/hm ^-2 The method comprises the steps of carrying out a first treatment on the surface of the The range of target nitrogen application at day 87 after sowing was: 76.12-78.38kg hm ^-2 ；

Then, the organic matter content of the soil can be kept at 20 g.kg ^-1 The total nitrogen content of the soil is kept to be 1.2 g.kg ^-1 The salt content of the soil is 2 g.kg ^-1 The pH value is between 6.5 and 7.5, and the soil pH value belongs to the soil pH range under high soil fertility. As can be seen, the optimal target filling quantity fluctuation is less than 2mm, and the optimal target nitrogen application quantity fluctuation is less than 3kg & hm ^-2 The formulated optimal water nitrogen strategy is accurate. Optimizing irrigation quantity along with growth periodAnd further, the nitrogen application amount is continuously reduced, which is consistent with the practical situation that the base fertilizer is applied more and the topdressing is less in production.

Straw deep-buried returning is a key means for controlling salinization and improving soil fertility, and is attracting attention in recent years. On the one hand, researchers have conducted intensive studies on the characteristics of the straw layer itself: zhao Y (2016) et al found that 40cm was the optimum depth for straw burial in the river set irrigation zone, as a result of studies on the relationship between straw burial depth and soil water salt distribution. Zhang H (2020) et al have studied the change of the saline soil water-salt exchange flux under different straw layer thicknesses, and have shown that when the straw burying depth is 5cm, the soil can obtain better permeability and inhibit salt backflow. As can be seen, most of researches on deep burying of the straw are mainly aimed at spreading of soil water salt, but the influence of the deep burying of the straw on the soil environment is multidimensional in reality. On the other hand, the interaction of nitrogen in the deep buried sewage of the straw is also becoming more important. Rasol G (2020) reveals the influence trend of the coupling of the nitrogen in the deep buried water of the straw on the quality of tomatoes and the salinity of soil by setting a straw-water filling quantity-nitrogen application quantity cross test and combining with TOPSIS sequencing analysis. The research reveals the influence trend of straw deep burial and water nitrogen irrigation on a plurality of indexes of soil, but because of the limitation of field workload, the research irrigation and nitrogen application are only set at two levels, and finally a relatively accurate water nitrogen optimization strategy cannot be obtained. The research absorbs the optimization results of the burying depth and thickness of the straws by the predecessor, buries the straws with the thickness of 5cm at the soil layer of 40cm, develops the cross test of the field with three irrigation levels and four fertilization levels, constructs the complex mapping relation between the multi-dimensional soil fertility index and the irrigation and nitrogen application strategy through the PNN model, finally obtains the accurate range of the water and nitrogen irrigation level, and makes up the research blank of accurately formulating the water and nitrogen strategy under the multi-dimensional target at present.

2) The application develops a reverse neural network with a preference mechanism, namely a preference neural network, and establishes a complex mapping relation of four indexes of soil organic matter content, soil total nitrogen content, soil salt content and pH value, namely four parameters, three-time irrigation amount and nitrogen application amount on the basis of the reverse neural network. Fitting results show that the prediction performance of the preference neural network PNN is obviously superior to that of linear SVR, poly SVR, rbf SVR, LR, LOR and the traditional BP neural network.

The excellent performance of the preferred neural network PNN of the present application depends on its particular network architecture. First, the Dropout module can significantly reduce PNN overfitting risk and computational cost by randomly suspending non-output neurons in a nonlinear model. While Batch Normalization keeps PNN from being overly sensitive to small changes in the network's learned parameters, enhancing the stability of the model during training. In addition, the activation function ReLU adopted by the PNN model has the characteristic of semi-adjustment, and the gradient vanishing risk can be effectively reduced. Most importantly, the PNN has a unique preference mechanism, can actively learn in the training process and connect the prior rules in the persistent data through the preference among neurons, greatly improves the model convergence, and remarkably reduces the error rate and the calculation cost in the model training. Compared with other artificial intelligent models, the PNN model has better prediction precision, and the reverse training structure can also obtain the optimal water-nitrogen strategy through an end-to-end mode directly based on the multidimensional fertility target value, so that the applicability and popularization of the model are greatly enhanced. More importantly, the reverse and forward structures of the PNN can be flexibly converted according to actual requirements, and the simulation process of the forward PNN model on soil organic matters, total nitrogen, salt content and pH value under different water and nitrogen conditions is further shown, and good fitting performance is obtained. The double verification of forward and reverse PNN proves the superiority and advancement of PNN model in constructing complex mapping relation between water-nitrogen strategy and soil fertility.

The PNN has the advantage of forward and reverse flexible conversion, and is endowed with multiple practical functions. Specifically:

(1) if the forward PNN can be used for accurately quantifying the change conditions of soil salinity, organic matters and other environmental factors under the influence of human factors such as irrigation, fertilization and the like so as to accurately evaluate the feasibility of a certain production strategy, the reverse PNN can be used for formulating an accurate human operation scheme based on the environmental targets such as the soil salinity, the organic matters and the like so as to guide actual production.

(2) The PNN model structure is formed, and can be trained by different index data sets to prepare optimization schemes in different fields. Therefore, PNN is not limited in function to the formulation of water-nitrogen optimization strategies, but other quantifiable embodiments can be formulated, and the goal is not limited to maintaining higher soil fertility, but rather improving crop quality, yield, etc.

(3) It is worth mentioning that besides the organic matter content of the soil, the total nitrogen content of the soil, the salt content of the soil and the pH value, the input vector of the preference neural network PNN of the application also contains consideration of irrigation time, fertilization time and soil layer depth, so that the PNN can not only formulate a water nitrogen optimization strategy based on a soil fertility target of 0-40cm, but also can formulate more flexible water nitrogen strategy aiming at the salt content, the pH value, the organic matter content and the total nitrogen content of different spaces under different irrigation time. Accurate control of PNN over time and space factors makes it possible to apply it as a core decision tool in precision agriculture.

According to field actual measurement data of organic matter content, total nitrogen content, salt content and pH value of soil in each growth period of corn, which are obtained through 2017 straw deep buried sewer nitrogen cross test, the simulation performance of the constructed reverse preference neural network PNN on the water nitrogen irrigation strategy is obviously superior to that of linear SVR, poly SVR, rbf SVR, LOR, LR and a traditional BP neural network, and the simulation of forward PNN on the change of soil organic matter, total nitrogen, salt content and pH value under different water nitrogen strategies also has higher precision.

The optimal water-nitrogen strategy obtained by the preference neural network PNN based on the multidimensional fertility target is as follows: the range of target nitrogen application amount on the 17 th day after sowing is as follows: 95.11-98.89kg hm ^-2 The method comprises the steps of carrying out a first treatment on the surface of the The range of target nitrogen application at day 59 post-sowing was: 86.69-86.70 kg/hm ^-2 The method comprises the steps of carrying out a first treatment on the surface of the The range of target nitrogen application at day 87 after sowing was: 76.12-78.38kg hm ^-2 The method comprises the steps of carrying out a first treatment on the surface of the Then, the organic matter content of the soil can be kept at 20 g.kg ^-1 The total nitrogen content of the soil is kept to be 1.2 g.kg ^-1 The salt content of the soil is 2 g.kg ^-1 In the following, the pH value is between 6.5 and 7.5, and in summary, the preferred neural network PNN of the present application performs well in formulating an optimized water nitrogen strategy based on a specific fertility objective.

The development of Chinese agriculture is very dependent on fertilizer input (Wang QB et al, 1996), and excessive irrigation and nitrogen application not only lead to low utilization efficiency of crop fertilizer, but also lead to soil acidification (Miao YX et al, 2011), soil nitrogen leaching (Lu J et al, 2019), organic matter content reduction (Sun H et al, 2018) and secondary salinization (Machado RMA and Serralheiro RP, 2017), and finally lead to low soil fertility. Studies have shown that the nitrogen loss increases by 1% for every 10% increase in nitrogen application (Zhang SH et al 2020). The deep burying of the straw can relieve nitrogen loss (Yang H et al, 2016) and salinization (Yonggan Zhao et al, 2016) of the soil, and simultaneously increase carbon reserves (Lu Fei, 2015) of the soil. Therefore, the optimization of the water-nitrogen strategy under the condition of deeply burying the straw is significant for improving the soil fertility, reducing the non-point source pollution and preventing and controlling the salinization, but the research of comprehensively formulating the water-nitrogen strategy by taking the soil fertility as a target under the condition of deeply burying the straw is less at present.

The optimization of water and nitrogen is always a research hotspot in the field of farmland water conservancy. Over the last two decades, water nitrogen application strategies have been largely available from limited field treatments (Yin F,2007;Wang Y et al,2019;Ning D et al,2019;Zhang MM et al,2018). To reduce field trial costs and improve accuracy of water nitrogen strategies, physical models have been developed such as HYDRUS (Karandish F et al,2018; hu A, 2019) and SWAP (Wu Y et al,2017;Sonneveld M P W et al,2003). The model can simulate the influence of different water nitrogen allocation embodiments on soil solutes and plant growth under specific boundary conditions, and the accuracy of the water nitrogen strategy is improved by increasing the encryption water nitrogen setting level of the simulation times. However, whether through field experiments or physical models, the proposed water-nitrogen optimization strategy is mostly based on the forward research concept of treatment-index-treatment, namely, the influence trend of water-nitrogen measures on soil indexes is determined first, and then the trend is utilized to reversely infer better water-nitrogen measures. The accuracy of the water and nitrogen strategy formulated under the concept is very dependent on the range and density of water and nitrogen variables of a field arrangement or input model, and many researches only carry out preferential screening in limited water and nitrogen treatment, but the effectiveness of water and nitrogen combinations which are not covered in the experimental design or the input range of the model cannot be fully verified. Thus, the process-index-process concept is no longer suitable for the formulation of water-nitrogen strategies due to its high test cost, low accuracy and representativeness. An innovative research idea capable of directly providing a water nitrogen optimization strategy according to a determined target is needed to be provided.

In recent years, artificial Intelligence (AI) models have been successfully used in the prediction of farm indicators such as yield (Su YX et al 2017,Senthilnath J et al,2016), soil water nitrogen (Shekofteh H et al, 2013), and the like. Successful predictions of citrus and coffee fruit numbers by SVM models trained on crop images have been studied (Ramos PJ et al,2017;Sengupta S et al,2014); high resolution data collected by satellite sensors predicts wheat yield (Pantazi XE et al 2016); the SVM model is trained with near infrared spectral data to predict soil total nitrogen, organic carbon and moisture content (Morellos a ET al, 2016), or more accurately predict crop ET0 (Sanikhani H ET al, 2019) and canola yield (Niedbala G, 2019) using a multi-layer perceptron (MLP). In addition to the prediction on the time scale, yuhong Dong (2019) predicts the corn yield under each fertilization treatment by using a wavelet-BP neural network, and establishes a mapping relationship between fertilization and yield. Gu J (2017) and the like predict crop yield under different irrigation volumes by using BP neural network. The combination of artificial intelligence and field tests provides a new thought for the formulation of the water-nitrogen strategy, but the model is mostly used for the prediction of indexes such as yield and the like, and the accurate water-nitrogen strategy can be rarely formulated. In addition, the BP neural network is simple in structure, the problem of overfitting caused by limited test data volume cannot be effectively solved, and the defect of insufficient generalization capability of a model is also a great disadvantage of the neural network with fewer layers.

Therefore, a Preferential Neural Network (PNN) is innovatively developed, and a reverse training mechanism of the PNN enables a model to directly determine an optimal water nitrogen irrigation strategy each time according to soil organic matter content, soil total nitrogen content, soil salt content and pH value, so that an end-to-end prediction mechanism is truly realized; the Dropout module of the PNN enables the model to perform random and repeated training by using limited field data, breaks through the limitation of data quantity, and greatly reduces the risk of over-fitting; the network structure of the 2 hidden layers can greatly improve the compliance of PNN to the preference structure.

The traditional BP neural network is used as a linear full-connection model, and all neurons have the same connection number, so that the change of each dimension of an input vector is treated equally when signals are propagated. It is needless to say that there is a large gap between the interaction mechanism between such signal propagation process and the actual variables in the field. For example, in the present study, four indexes of total soil nitrogen, organic matter (SOM), total salt content and pH input by the model have significant differences in sensitivity to changes of irrigation and nitrogen application, if the indexes with strong sensitivity to water nitrogen application are fully connected with the indexes with weak sensitivity in the same way, the model cannot accurately identify the strong and weak relation between the two indexes and the water nitrogen strategy in signal propagation, and finally, the model operation cost and the error risk may be increased.

Therefore, the study provides a preference value concept for better representing the dependency degree among different indexes in farmland water conservancy for the first time, and the preference value is the centralized representation of the relativity and the significance among variables. Based on the preference among variables in the actual measurement data, a neuron connection preference module is created for PNN, and the preference matrix and the weight matrix can be subjected to Hadamard product according to the correlation strength between the processing level of the actual measurement data set and the index so as to control the connection strength among neurons in advance, so that the model convergence is greatly optimized, and meanwhile, the prediction precision is effectively improved. The PNN model trained by four soil fertility data, namely parameters, namely soil organic matter content, soil total nitrogen content, soil salt content and pH value, collected in 2017 under different water and nitrogen measures is used for constructing complex mapping relations between each index and a water and nitrogen application strategy, and the reliability of the water and fertilizer application strategy based on the PNN is finally proved through verification of forward and reverse PNN. And further provides an optimized water-nitrogen strategy capable of maintaining the high fertility of the crop root soil. The main contributions of the present application are as follows:

1. establishing a water-nitrogen optimization strategy based on a multidimensional target to construct a Preferential Neural Network (PNN);

2. Creatively provides an artificial neural network reverse training scheme suitable for a water-nitrogen optimization scheme;

3. providing an optimized water nitrogen distribution strategy based on farmland soil fertility under deep straw burial;

4. the effectiveness of the models presented herein and their training strategies has been validated by field trials.

The application develops a 2 hidden layer deep learning model with a reverse training mechanism and a preference structure for the first time based on 2017 field actual measurement data, namely a preference neural network (Preference neural network, PNN). The unique reverse training mechanism of PNN can directly output a water irrigation and nitrogen application strategy in the growth period based on a multi-dimensional target (organic matters, total nitrogen, salt content and pH value) of 0-40cm soil fertility in the corn root zone, and the 2 hidden layer structure can well fit with the preference mechanism of PNN. The preference value and the calculation mode thereof proposed for the first time show the correlation and significance among variables in a concentrated way, and the preference mechanism established based on the correlation and significance can naturally fuse the neural network structure with the priori rule of the data in the nature, so that the model convergence is greatly improved, and the prediction precision is effectively improved. In addition, the novel adoption of Dropout and Batch Normalization modules of the PNN effectively reduces the error risk and the operation cost, and the gradient vanishing problem of the deep neural network is solved by adopting a ReLU function. The mean square error of the PNN prediction water nitrogen strategy is 0.007-0.009, the average absolute error is 0.007-0.011, the root mean square error is 0.008-0.012, and the method is remarkably superior to linear SVR, poly SVR, rbf SVR, LOR, LR and a traditional BP neural network trained under the same condition. Through PNN simulation, 17d, 59d and 87d after sowing are respectively irrigated with 89.30-90.68mm, 87.91-88.61mm and 90.78-91.70mm, nitrogen is applied with 95.11-98.89kg hm-2, 86.69-86.70kg hm-2 and 76.12-78.38kg hm-2, organic matters and total nitrogen content of soil with the growth period of 0-40cm can be respectively maintained to be more than 20 g.kg-1 and 1.6 g.kg-1, the salt content is maintained to be less than 2 g.kg-1, the pH value is between 6.5 and 7.5, the requirement of high soil fertility index is met, and PNN can be practically applied to the establishment of an optimal water nitrogen strategy.

In the above embodiments, although steps S1, S2, etc. are numbered, only specific embodiments are given herein, and those skilled in the art may adjust the execution sequence of S1, S2, etc. according to the actual situation, which is also within the scope of the present invention, and it is understood that some embodiments may include some or all of the above embodiments.

As shown in fig. 27, an artificial intelligence system 200 for acquiring a water nitrogen policy according to an embodiment of the present invention includes a first acquisition module 210 and a second acquisition module 220;

the first obtaining module 210 is configured to obtain a target value of each parameter associated with soil fertility at a preset time point;

the second obtaining module 220 is configured to input a preset time point and all target values into a preference neural network with a preference mechanism, so as to obtain a water-nitrogen strategy including a target water filling amount and a target nitrogen application amount, where the preference mechanism is used to characterize a relationship between each parameter and the water filling amount and the nitrogen application amount respectively.

Preferably, in the above technical solution, the device further comprises a building module and a training module;

the building module is used for building a neural network;

the training module is used for: and obtaining and normalizing the preference value between each parameter and the water filling amount and the nitrogen application amount respectively to obtain a preference matrix, introducing the preference matrix into the neural network, and combining with an Adam algorithm to train to obtain the preference neural network with a preference mechanism.

Preferably, in the above technical solution, the building module is further configured to use the ReLU function as an activation function of the neural network, and use the ReLU function as an activation function of the neural network, so as to solve the gradient vanishing problem of the neural network.

Preferably, in the above technical solution, the building module is further configured to: the Dropout algorithm is introduced into the neural network, the batch standardization is introduced into the neural network, and the Dropout algorithm and the batch standardization are introduced into the neural network, so that the error risk and the operation cost are effectively reduced, and the operation efficiency is improved.

The steps for implementing the corresponding functions of the parameters and the unit modules in the artificial intelligence system 200 for obtaining the water-nitrogen policy according to the present invention may refer to the parameters and the steps in the embodiment of the artificial intelligence method for obtaining the water-nitrogen policy according to the present invention, which are not described herein.

The electronic equipment comprises a memory, a processor and a program stored in the memory and running on the processor, wherein the processor realizes the steps of the artificial intelligence method for acquiring the water-nitrogen strategy implemented by any one of the above when executing the program.

The electronic device may be a computer, a mobile phone, or the like, and the program is computer software or mobile phone APP, and the parameters and steps in the above electronic device according to the present invention may refer to the parameters and steps in the above embodiment of an artificial intelligence method for obtaining a water-nitrogen policy, which are not described herein.

Those skilled in the art will appreciate that the present invention may be implemented as a system, method, or computer program product.

Accordingly, the present disclosure may be embodied in the following forms, namely: either entirely hardware, entirely software (including firmware, resident software, micro-code, etc.), or entirely software, or a combination of hardware and software, referred to herein generally as a "circuit," module "or" system. Furthermore, in some embodiments, the invention may also be embodied in the form of a computer program product in one or more computer-readable media, which contain computer-readable program code.

Any combination of one or more computer readable media may be employed. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium can be, for example, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of the computer-readable storage medium include an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

While embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and not to be construed as limiting the invention, and that variations, modifications, alternatives and variations may be made to the above embodiments by one of ordinary skill in the art within the scope of the invention.

Claims

1. An artificial intelligence method for obtaining a water nitrogen strategy, comprising:

inputting a preset time point and all target values into a preference neural network with a preference mechanism to obtain a water-nitrogen strategy comprising target water filling quantity and target nitrogen application quantity, wherein the preference mechanism is used for representing the relation between each parameter and the water filling quantity and the nitrogen application quantity respectively;

further comprises:

building a neural network;

obtaining and normalizing preference values between each parameter and the water filling amount and the nitrogen application amount respectively to obtain a preference matrix, introducing the preference matrix into the neural network, and training by combining with an Adam algorithm to obtain a preference neural network with a preference mechanism;

the process for building the neural network comprises the following steps:

defining the input end of the neural network as a matrix

Where n is the sample size and d is the dimension of each input vector, where { xi }, i=1, …, n ε X generationTable the target values of organic matter content, total nitrogen content, salt content, pH and output of the neural network are defined as matrix representing irrigation and nitrogen application >

Obtaining the structure of a neural network;

wherein, the preference value calculation formula is:

wherein PV _ij As variable X _i And Y _j Preference value, X _i Represents the ith variable related to the experimental treatment, Y _j Representing the j-th index variable obtained by test monitoring under different treatments, specifically: soil organic matter content, total nitrogen content of soil, soil salt content, pH value, and soil sampling time point and depth, e is constant and e is 0.001, and r is pearson correlation coefficient.

2. An artificial intelligence method for obtaining a water nitrogen strategy according to claim 1, wherein said building a neural network further comprises:

the ReLU function is used as an activation function of the neural network.

3. An artificial intelligence method for obtaining a water nitrogen strategy according to claim 1 or 2, characterized in that said building a neural network further comprises:

introducing a Dropout algorithm into the neural network;

batch normalization was introduced into the neural network.

4. An artificial intelligence system for acquiring a water nitrogen strategy is characterized by comprising a first acquisition module and a second acquisition module;

The second acquisition module is used for inputting a preset time point and all target values into a preference neural network with a preference mechanism to obtain a water-nitrogen strategy comprising target water filling quantity and target nitrogen application quantity, wherein the preference mechanism is used for carrying out focused learning on the neural network parameters according to objective relations between each parameter and the water filling quantity and the nitrogen application quantity respectively;

the device also comprises a building module and a training module;

the building module is used for building a neural network;

the training module is used for: obtaining and normalizing preference values between each parameter and the water filling amount and the nitrogen application amount respectively to obtain a preference matrix, introducing the preference matrix into the neural network, and training by combining with an Adam algorithm to obtain a preference neural network with a preference mechanism;

the process of building the neural network by the building module comprises the following steps:

defining the input end of the neural network as a matrix

Wherein n is the sample capacity, d is the dimension of each input vector, wherein { xi }, i=1, …, n ε X represent the target value of soil organic matter content, the target value of total nitrogen content, the target value of soil salt content, the target value of pH value for measuring soil fertility, and the output end of the neural network is defined as the matrix representing irrigation and nitrogen application amount ∈ >

Obtaining the structure of a neural network;

wherein, the preference value calculation formula is:

wherein PV _ij As variable X _i And Y _j Preference value, X _i Represents the ith variable related to the experimental treatment, Y _j Representing the j-th index variable obtained by test monitoring under different treatments, specifically: soil organic matter content, soilTotal nitrogen content, soil salinity, pH, and soil sampling time point, depth, e is constant and e takes 0.001, r is pearson correlation coefficient.

5. The artificial intelligence system for obtaining a water nitrogen strategy according to claim 4, wherein the building module is further configured to use a ReLU function as an activation function of the neural network.

6. An artificial intelligence system for obtaining a water nitrogen strategy according to claim 4 or 5, characterized in that the building module is further adapted to: the Dropout algorithm is introduced into the neural network and batch normalization is introduced into the neural network.

7. An electronic device comprising a memory, a processor and a program stored on the memory and running on the processor, wherein the processor, when executing the program, implements the steps of an artificial intelligence method for obtaining a water nitrogen policy as claimed in any one of claims 1 to 3.