CN113221446A

CN113221446A - Method and system for acquiring water and fertilizer strategies of saline soil, storage medium and electronic equipment

Info

Publication number: CN113221446A
Application number: CN202110431777.3A
Authority: CN
Inventors: 吴宗翰; 胡睿琦; 潘平波; 于叶露; 刘月
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: 2021-08-06
Anticipated expiration: 2041-04-21
Also published as: CN113221446B

Abstract

The invention relates to the technical field of agriculture, and provides a method and a system for acquiring a water and fertilizer strategy of saline soil, a storage medium and electronic equipment, wherein the method comprises the following steps: acquiring a preset value of soil conductivity at a preset time point for representing the soil salinization degree, and determining a target value of each parameter related to the soil salinization at the preset time point; and inputting the preset value and all the target values into the trained reverse residual error neural network to obtain a saline soil water-fertilizer strategy corresponding to the soil with the soil conductivity as the preset value. The constructed reverse residual error neural network can better fit the actual application scene of the farmland water conservancy, the reverse residual error neural network can directly simulate and optimize the saline soil water and fertilizer strategy according to the target soil salinity corresponding to the preset value of the soil conductivity, the model precision is high, the simulation requirement can be met, a powerful technical means is provided for optimizing the field irrigation and fertilization strategy, and the analysis efficiency is improved.

Description

Method and system for acquiring water and fertilizer strategies of saline soil, storage medium and electronic equipment

Technical Field

The invention relates to the technical field of agriculture, in particular to a method and a system for acquiring a water and fertilizer strategy of saline soil, a storage medium and electronic equipment.

Background

Saline soil is one of the most main low-yield and medium-yield soil types in China, the productivity level of the saline soil is closely related to the soil quality condition, and improper utilization of the saline soil can quickly cause soil degradation and productivity level reduction. In salinization irrigation areas with shortage of fresh water resources, brackish water utilization is inevitable, but excessive utilization can aggravate soil salinization, and an irrigation system for seeking safe utilization of brackish water is urgent in the field. Brackish water is used for drip irrigation, and the method is wide in application and high in water saving efficiency, and the drip irrigation mode can reduce the influence range of brackish water on soil salinity distribution. In addition, the reasonable irrigation quantity not only ensures the normal growth of crops, but also can reduce the negative effect of the brackish water on the soil. The mechanism of water-fertilizer coupling under salinity stress is more complex, the excessive application of nitrogen fertilizer can aggravate the secondary salinization of soil, further inhibit the growth of crops, and the reasonable fertilizing amount is drawn up to ensure the prevention of further deterioration of the salinized soil. Therefore, the soil salt environment is an important factor considered in the water and fertilizer optimization scheme of the salinization irrigation area, the accurate analysis of the complex relationship between the soil salt and the brackish water and the fertilizer becomes a scientific problem which needs to be solved urgently, and the method has important significance for improving the salinization soil, saving fresh water resources and improving the land utilization efficiency.

At present, many scholars obtain optimized water and fertilizer schemes by establishing linear methods such as multiple regression equations and the like. In the aspect of combining water and fertilizer with salinity, besides the method for establishing a linear equation to solve and optimize the water and fertilizer coupling mode, a HYDROUS-1D model is also used for simulating the water content and salinity of the salinized soil in different irrigation and fertilization modes, the optimized water and fertilizer is obtained by artificially analyzing the water and salt dynamics, and the equation solution in the model is a Calerkin linear finite element method. These conventional data analysis methods have been widely used for a long time, and unfortunately, these methods assume a linear relationship between indexes, which is difficult to be established in a real-world situation. The method cannot analyze the nonlinear relation among indexes, so that certain unknown characteristics are lost artificially, the comprehensiveness of analysis is limited to a certain extent, and particularly, the action mechanism of brackish water and fertilizer on saline soil is more complex, and the linear analysis method cannot obtain the complex characteristics of water, fertilizer and salt; in addition, the traditional model simulation method takes irrigation and fertilization as an input end, simulates and predicts the water and salt dynamics of the soil, further performs multiple screening analysis by combining the water and salt dynamics of the soil to obtain the optimized water and fertilizer, and has low analysis efficiency.

Disclosure of Invention

The invention aims to solve the technical problem of the prior art and provides a method and a system for acquiring a water and fertilizer strategy of saline soil, a storage medium and electronic equipment.

The technical scheme of the method for acquiring the water and fertilizer strategy of the saline soil is as follows:

acquiring a preset value of soil conductivity at a preset time point for representing the soil salinization degree, and determining a target value of each parameter related to the soil salinization at the preset time point;

and inputting the preset value and all target values into a trained reverse residual error neural network to obtain a saline soil water and fertilizer strategy which corresponds to the soil with the preset value of soil conductivity and comprises a target irrigation quantity and a target nitrogen application quantity, wherein the irrigation mainly aims at brackish water, and fresh water is also suitable.

The method for acquiring the water and fertilizer strategy of the saline soil has the following beneficial effects:

the constructed reverse residual error neural network can better fit the actual application scene of the farmland water conservancy, the reverse residual error neural network can directly simulate and optimize the saline soil water and fertilizer strategy according to the target soil salinity corresponding to the preset value of the soil conductivity, the model precision is high, the simulation requirement can be met, a powerful technical means is provided for optimizing the field irrigation and fertilization strategy, and the analysis efficiency is improved.

On the basis of the scheme, the method for acquiring the water and fertilizer strategy of the saline soil can be further improved as follows.

Further, still include:

building a residual error neural network;

training the residual error neural network by using a data set, and obtaining the reverse residual error neural network by combining with an Adam optimization algorithm, wherein the data set comprises: and at each historical preset time point, the historical target value of each parameter, the historical preset value of the soil conductivity, the historical target irrigation quantity and the historical target nitrogen application quantity.

Further, the building a residual error neural network further comprises:

and taking the ReLU function as an activation function of the built residual error neural network.

Further, the building of the residual error neural network further comprises:

introducing batch normalization into the built residual error neural network;

the Dropout algorithm is introduced into the constructed residual neural network.

The technical scheme of the acquisition system of the water and fertilizer strategy of the saline soil is as follows:

the device comprises a determining module and an obtaining module;

the determining module is used for acquiring a preset value of soil conductivity at a preset time point, and determining a target value of each parameter related to soil salinization at the preset time point;

and the acquisition module is used for inputting the preset value and all the target values into the trained reverse residual error neural network to obtain a saline soil water and fertilizer strategy which corresponds to the soil with the preset soil conductivity and comprises the target irrigation quantity and the target nitrogen application quantity.

The system for acquiring the water and fertilizer of the saline soil has the following beneficial effects:

On the basis of the scheme, the acquisition system of the water and fertilizer strategy of the saline soil can be further improved as follows.

The system further comprises a modeling module and a training module, wherein the modeling module is used for building a residual error neural network;

the training module is used for training the residual error neural network by using a data set and obtaining the reverse residual error neural network by combining an Adam optimization algorithm, wherein the data set comprises: and at each historical preset time point, the historical target value of each parameter, the historical preset value of the soil conductivity, the historical target irrigation quantity and the historical target nitrogen application quantity.

Further, the modeling module is further configured to:

introducing batch normalization into the built residual error neural network;

The storage medium stores instructions, and when a computer reads the instructions, the computer executes the method for acquiring the water and fertilizer strategy of the saline soil.

An electronic device of the present invention includes a processor and the storage medium, where the processor executes instructions in the storage medium.

Drawings

FIG. 1 is a schematic flow chart of a method for acquiring a water and fertilizer strategy of saline soil according to an embodiment of the invention;

FIG. 2 is a graph comparing a neural network with a residual depth neural network;

FIG. 3 is a diagram of a Sigmiod function;

FIG. 4 is a schematic representation of the derivative of the Sigmiod function;

FIG. 5 is a diagram of a ReLU function;

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

FIG. 7 is one of the vertical distribution plots of the average conductivity of soil under different water and fertilizer strategies;

FIG. 8 is a second vertical distribution diagram of the average conductivity of the soil under different water and fertilizer strategies;

FIG. 9 is a third vertical distribution diagram of the average conductivity of soil under different water and fertilizer strategies;

FIG. 10 is a graph showing data for outputting a value of soil conductivity versus an actual value of soil conductivity;

FIG. 11 is a graph showing data of an output soil pH value versus an actual soil pH value;

FIG. 12 is a data diagram of an output value of soil moisture content versus an actual value of soil moisture content;

FIG. 13 is a data diagram of the target irrigation amount of the medlar in each growth period of the slightly saline soil, the moderately saline soil and the heavily saline soil;

FIG. 14 is a data diagram of the target nitrogen application amount for each growth period of Lycium barbarum in lightly saline soil, moderately saline soil, and heavily saline soil;

FIG. 15 is a graph of irrigation levels of lightly saline soils;

FIG. 16 is a graph of irrigation levels of moderately saline soils;

FIG. 17 is a diagram of irrigation level of heavily saline soil;

FIG. 18 is a schematic structural diagram of a system for acquiring water and fertilizer from saline soil according to an embodiment of the present invention;

Detailed Description

As shown in fig. 1, the method for acquiring a water and fertilizer strategy of saline soil according to the embodiment of the invention comprises the following steps:

s1, acquiring a preset value of soil conductivity at a preset time point for representing the soil salinization degree, and determining a target value of each parameter related to the soil salinization at the preset time point;

and S2, inputting the preset value and all target values into the trained reverse residual error neural network, and obtaining a saline soil water and fertilizer strategy which corresponds to the soil with the preset value of soil conductivity and comprises the target irrigation quantity and the target nitrogen application quantity.

Wherein, the parameters related to soil salinization comprise: depth of soil layer (depths), moisture content of soil, and pH value of soil.

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

s020, building a residual error neural network;

s021, training the residual error neural network by using a data set, and obtaining the reverse residual error neural network by combining an Adam optimization algorithm, wherein the data set comprises: and at each historical preset time point, the historical target value of each parameter, the historical preset value of the soil conductivity, the historical target irrigation quantity and the historical target nitrogen application quantity.

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

and S0200, taking the ReLU function as an activation function of the built residual error neural network.

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

s0201, introducing batch normalization into the built residual error neural network;

and S0202, introducing a Dropout algorithm into the built residual error neural network.

The method for acquiring the water and fertilizer strategy of the saline soil is explained in detail through the following embodiment:

s10, selecting a test area, specifically:

the test was carried out at the former flag red-wei test station (108 ° 45'-109 ° 36' E, 40 ° 30'-40 ° 40' N) of the inner mongolian Urait in 2016, 2017, which was located in the west of the three-lake river irrigation area at the most downstream of the river irrigation area. The research area belongs to moderate-temperature-zone continental multi-wind drought climate, the average precipitation for many years is 270mm, the average evaporation capacity for many years is 2383mm, the average temperature for many years is 7.9 ℃, the frost-free period is 146d, namely 146 days, and the accumulated temperature (more than 10 ℃) is 3200 h. The test area has a large amount of underground brackish water (the mineralization degree is about 3.84 g/L) resources, most of soil is filled soil and saline soil, the average water content of soil in a 0-100cm soil layer is 26.25%, the EC value of the soil is 1.58mS/cm, the pH value of the soil is 8.12, and the volume weight of the soil is 1.55g/cm³. The basic soil properties of the test area are shown in the following table 1:

TABLE 1

S11, experimental design, specifically:

the test plant is Ningqi No. 1 medlar, underground brackish water is adopted for carrying out drip irrigation, urea is adopted as a test fertilizer, diammonium phosphate and phosphorus nitrate are applied as base fertilizers before first water, urea is applied as a top dressing along with irrigation in a growth period, top dressing is carried out for 4 times, and the top dressing amount is 20%, 30%, 40% and 10% of the total top dressing amount in the whole year respectively. The brackish water drip irrigation system and the fertilization system are shown in table 2:

TABLE 2

The test factors are irrigation quantity and nitrogen application quantity, the irrigation quota is respectively set as 232.5, 285 and 375mm, and the nitrogen application quantity is respectively set as 210, 525, 750 and 975kg/hm²There were 12 treatments, each set of 3 replicates. The area of a single test cell is 39m²Each cell comprises 13 Chinese wolfberry plants, the plant spacing of the Chinese wolfberry plants is 1.0m, and the row spacing is 3.0 m. And a protective belt is arranged at the periphery of the test cell, and a 1m water-stop sheet is adopted for seepage-proofing isolation among the cells.

S12, defining input and output ends of the residual neural network, specifically:

1) the input of the residual neural network is defined as the matrix X ∈ R^n×dWhere n is the sample size, in this application n is 756; d is each input vector dimension, d being 3, { xi } in this application_{i＝1,2,3,4,5}E.X represents: the method comprises the steps of obtaining a vector set comprising a target value of soil water content, a preset value of soil EC (intrinsic safety) value, namely soil conductivity, a target value of soil pH value and target values of corresponding sampling time, namely a preset time point and soil depth, wherein the vector set comprises a historical target value of soil water content, a historical preset value of soil EC value, namely soil conductivity, a historical target value of soil pH value and a historical target value of corresponding sampling time, namely a historical preset time point and soil depth;

2) the output end of the residual error neural network is defined as a matrix Y epsilon R^n×2The matrix Y is formed by R^n×2Representing a series of fertigation levels with time and depth dimensions, the residual neural network aims to learn and train the mapping for a given input matrix

The parameter theta in the method enables Y' obtained by the trained model to be infinitely close to Y representing the real irrigation and fertilization level, so that a well-trained reverse residual error neural network is obtained; the matrix Y represents: the target amount of water to be poured and the target amount of nitrogen to be applied at the preset time point are understood to mean, at the time of training: historical target irrigation water amount and historical target nitrogen application amount at each historical preset time point;

s13, building a residual error neural network, specifically:

1) basic principle of residual neural network: the residual neural network is divided into an input layer, a hidden layer and an output layer. The method comprises the steps that soil data, namely soil sampling time (days), soil depth (depths), soil moisture content, soil EC (environmental protection) value and soil pH value, are input into an input layer, the size of a model is adjusted according to the size of the data through a pluggable Block module, namely a residual Block, and nonlinear characteristics of the residual Block are introduced into a neural network through a hidden layer activation function. The output end of the model is an optimized water and fertilizer coupling regulation and control mode, and the output function is f (x) -x to f (x), so that the optimization performance of the model is improved;

2) model mapping and residual Block, Block module, specifically:

an important objective of the rdr (reverse Deep resnet) model, namely the reverse residual neural network, constructed in the application is to efficiently analyze high-dimensional and large-scale irrigation and water conservancy data to obtain optimization capability, so that a structure with Deep layers and meeting efficient calculation requirements needs to be established. However, as the neural network becomes deeper, the problem of gradient vanishing/gradient explosion impeding model convergence occurs: i.e. as the model is deeper, the performance drops significantly after reaching saturation. The RDR proposed in this application solves the degradation problem by reconstructing the hidden layer as a learning residual function with reference to the identity map (layer input).

Suppose h₁(x) Is the output layer in the basic building block of the residual neural network, and proposes that the layer of the basic building block approximates a residual function: h (x) x. Thus, the original function can be converted into

Wherein

Is the output layer of the basic building Block (Block) of the residual neural network, h₁(x) Is the output layer of the block. By constructing f (x) -x operations as element-added spanning links, as shown in FIG. 2, with the structure of the neural network on the left and the structure of the residual neural network on the right in FIG. 2, it can be appreciated that the residual neural network is reversedThe differential neural network is obtained based on the residual neural network training, and the structure of the residual neural network is the structure of the reverse residual neural network;

3) hiding the layer:

because the residual neural network is a deep neural network structure with 6 hidden layers, it is critical to define affine transformations between layers of information flow. In general, the learning parameter θ of each layer is composed of a weight matrix w and a bias parameter b. Ith of residual neural network_thHidden layer h_lCan be defined as the following equation:

wherein W₁And b₁Is at l_thWeights and bias parameters that the layer can learn. h is_l-1Is a hidden layer representative of the previous layer. When l is 1, h₀Updating the rule layer by layer, wherein the given input data flow to the intermediate operation hidden layer and finally reach the output layer;

4) taking the ReLU function as an activation function of the built residual error neural network, specifically:

the activation function introduces nonlinear characteristics into the residual neural network, so that the residual neural network can arbitrarily approximate a complex natural function within the distribution range of the input. Without the activation function, the multilayer structure remains equivalent to a single layer structure due to its linear character. However, a great deal of current work does not accurately use the activation function, and they use the Sigmiod function in a shallow model, so that the simulation result is suboptimal. In this section, the present application will explain the reason why the sigmood function is not suitable for use in the field of agricultural land and water, and introduce a Rectified Linear Units (ReLU) function applied in RDR, so that:

first, a Sigmiod activation function, also called a logistic function, whose mathematical expression is:

FIGS. 3 and 4 reveal that the Sigmiod activation function compressively transforms the range of values to 0,1]In between, when a neuron in the network is saturated (the output implicit value in the neuron is close to 0 or 1,its gradient value is 0) makes the gradient more likely to approach 0. Obviously, a gradient of 0 results in that the weight parameters in the saturated neurons cannot be updated, and at the same time, the neurons connecting these saturated neurons will pass slowly, which is the phenomenon of gradient disappearance. Furthermore, the Sigmiod function is vulnerable to still another cause in current deep learning architectures: the e-x function in the Sigmiod function would make the overall computation of the model expensive if it were applied at each layer.

The ReLU function effectively solves the above problems. Fig. 5 and 6 illustrate the semi-correcting property of the ReLU function, whose mathematical expression is as follows: re L U (x) ═ max (0, x), and if x < 0, the output of the ReLU activation function is 0, otherwise the output is still its original value x. This precise operation makes the propagation and transmission of the entire neural network more efficient. Because ReLU liberates neurons from a limited range, a large number of neurons are back-propagated in the positive region, avoiding the gradient vanishing problem. According to the application, a ReLU function is applied behind each hidden layer to complete nonlinear transformation. This operation can be formulated as follows:

wherein the content of the first and second substances,

is a hidden layer representation of batch normalization, as will be described in detail below.

5) Introducing batch normalization into the constructed residual neural network, specifically:

although deep neural networks achieve good generalization capability from deep architectures, training is achieved by a mechanism in which input data is transformed from learnable parameters of all layers in the neural network, complicating the training process. In this case, as the network becomes deeper, the slight variations in these parameters in the structure will be amplified. The change of the input distribution of each layer requires that the subsequent layer continuously obeys the updated distribution, thereby reducing the stability of the training process and the stability of the model.

Considering that a residual neural network having a six-layer structure may suffer from the same problems as described above, the stability of the training iteration is enhanced and the model stability is improved using batch normalization. The batch normalization operation has 4 layers:

given a small batch B, obtaining the average value of all elements in the small batch B by the following formula:

where m represents the number of samples in each small lot B.

② calculating mu_BCorresponding variance

Specifically, the method comprises the following steps:

calculating to obtain;

usable mu for batch normalization operation_BAnd

expressed as:

where ε is a small value to prevent

The problem of unstable parameters caused by time;

the normalized hidden layer representation can be scaled and shifted by two learnable parameters γ and β:

wherein gamma represents a scaling parameter, beta represents a displacement parameter for enhancing the expression capability of a hidden layer in the residual neural network, and both gamma and beta can be obtained by training and learning, wherein the specific calculation flow of batch normalization conversion is as follows:

order:

minimum batch size_thHidden layer representation；

L: the number of layers;

m: the number of samples in each small batch;

γ: scaling parameters, which can be learned;

beta: displacement parameters which can be obtained by learning;

and (3) outputting:

1: for L ═ 1,2,3, ·, L;

2: updating mu_B，

3: updating

4: small batch normalization of hidden layer representations:

5: scaling and translating the normalized hidden layer representation

6: finishing;

6) introducing a Dropout algorithm into the constructed residual neural network, specifically:

deep complex architectures with limited training data sets typically force the network to adapt to input characteristics during training, which is referred to as overfitting or co-adaptation. Dropout is an algorithm with low computational cost (computational complexity o (n)), and in each training iteration, non-output neurons are randomly suspended from a nonlinear model, so that effective solution is realized. In particular, the dropout mechanism extracts the binary mask multiplied by the output from the neural networkThe hidden representation generated in the hidden layer and the hidden layer, and the neuron with the output value multiplied by 0 is temporarily deleted in the current iteration process. The mask samples of each layer are independent, and the probability that the sample value in the mask is 1 is a predefined hyper-parameter of the whole network. With reference to the formula:

the mathematical expression of the dropout algorithm can be represented in a hierarchical manner:

d～Bernoulli(p′)；

wherein

Is a dot product operation by elements, d₁Is a_thThe dropout mask of a layer, p' is a pre-defined dropout hyper-parameter before training.

S14, training, specifically:

training a strategy reversely, specifically:

the problem of optimizing the irrigation and fertilization system by using the traditional water conservancy method needs to be solved through a complex process of inputting field measured data (including indexes such as irrigation and fertilization systems, soil physicochemical properties and the like), simulating soil index rules under different irrigation systems and further calculating and optimizing the irrigation system, and the defects of zigzag solving process, low calculation efficiency, artificial subjective analysis errors and the like exist. The causal relationship between irrigation and fertilization and soil salinity objectively exists, namely the soil salinity is changed due to irrigation of brackish water and application of nitrogen fertilizer, the objective causal relationship is not changed due to reverse solution, and therefore the irrigation and fertilization system can be efficiently and accurately predicted and optimized by utilizing the reverse solution of the neural network model.

Based on the method, a reverse training strategy is firstly proposed and applied to a reverse residual error neural network (RDR) model, and after field data are input, an irrigation and fertilization system can be efficiently solved and optimized directly according to target soil indexes. Specifically, the inverse training strategy changes the output end of the traditional model into the input end, reversely learns the mapping between the two, and the input end of the traditional model into the output end.

It is worth mentioning that the reverse training strategy provided by the application is not only applied to the RDR model, but also can be widely adapted to other water conservancy models based on a deep learning framework so as to achieve the purpose of efficiently and accurately optimizing the irrigation scheme.

Secondly, setting an objective function, specifically: the square error MSE, mean absolute error MAE and root mean square error RMSE are widely used in farmland water and soil, but they tend to sensitively fit outliers in the data, which may negatively impact model performance. To enhance the robustness of the RDR proposed by the present application, the present application takes the Huber loss function (also called smooth L1 loss function) as an objective function to calculate the gradient and eventually update the model. The Huber loss essentially balances the contour sensitivity of RMSE and the contour kurtosis of the mean absolute error MAE by employing the following piecewise definitions:

wherein k is a balancer, and if the value of y-f (x) is greater than k, the loss is amplified using linear operations, and if the value is less than or equal to k, the loss is reduced using quadratic equations.

(iii) model optimization, specifically:

different from the problems that the repeated iteration of a large slope is not converged and all neural network parameters can only adopt a uniform learning rate due to the adoption of a random gradient descent method (SGD) in the previous research, the reverse residual error neural network disclosed by the application realizes stable convergence in the training process by utilizing Adma algorithm optimization. The two basic components of the Adma optimizer, the Adma algorithm, are a momentum term v and an exponentially weighted moving average term s (also called leakage average), both of which are typically initialized to zero. Momentum v accumulates the previous gradient elements:

to reduce the gradient variance, the formula is as follows:

v_t＝β_vv_t-1+(1-β_v)g_t；

to obtain

Where t is the iteration step size, and β v ∈ [0,1) controls the number of previous gradient gradients that would affect the direction of descent of the current step size. The exponentially weighted moving average term s accumulates the previous gradient variance element by element to give each parameter an individual learning rate:

notably, since v₀And s₀Are initialized to zero and thus a large number of deviations will initially tend primarily to small values. To correct this problem, v is_tAnd s_tIs re-normalized so that the sum of the term values for all steps is 1, called bias correction, specifically:

v is to be_tThe deviation is corrected as follows:

similarly, s is_tThe deviation is corrected as follows:

using corrected deviations

And

the gradient of each element in the network parameters may be rescaled as follows:

where δ is a predefined learning rate and ε is a very small constant for numerical stability. The point-by-point element update can be written as follows:

the algorithm flow of the Adma algorithm optimization is as follows:

order: b ═ x¹,…,x^mAre respectively corresponding to y⁽ⁱ⁾；

m: the number of samples in each small batch;

t: the number of iteration steps;

v: a momentum term;

s: the leakage average term is an exponential weighted moving average term;

β_v: v is a non-negative hyperparameter;

β_s: a non-negative hyperparameter of s;

epsilon: small constants for numerical stability;

initialization: v is 0, s is 0;

t0, 1,2, T performing a for loop;

2. producing a gradual change

3.t＝t+1；

4. Calculating momentum items: v. of_t＝β_vv_t-1+(1-β_v)g_t；

5. Calculating the leakage average term:

6. deviation correction of momentum term:

7. bias correction of the leakage average term:

8. rescaling the fade:

9. updating parameters:

10. finishing;

the process of applying the inverse residual neural network is as follows:

order:

B_x＝{x¹,,,,,x^m}: a small batch B of training data sets;

l: number of layers

m: the number of samples in each small batch;

t0, 1,2,3, T performing a for loop;

2. from band with corresponding yⁱFor mini-batch B containing m data in the training dataset of (1)_xSampling is carried out;

l1, 2,3, L performs a for loop;

4. using affine transformations, reference, layer by layer

Obtaining: h is₁＝x_iW₁+b₁，

5. Reference formula

Applying nonlinear activation function conversion;

6. reference to

d-Bernoulli (p') applies dropout mechanism;

7. reference formula

Applying batch normalization;

8. end up

9. According to the formula:

updating the RDR by using gradient calculation loss;

10. finishing;

11. returning to obtain y' according to the expected value;

s15, model test:

1) the model test contains 540(108 × 5) data sets, representing a total of 108 groups, each group containing 5 data, respectively: and at any historical preset time, randomly extracting 80% of an original data set as a training set of RDR (resource data record) and 20% of the original data set as a verification set, wherein the historical target value of the water content of the soil, the historical preset value of the EC (environmental impact) value of the soil, namely the historical preset value of the conductivity of the soil, the historical target value of the pH value of the soil, the historical target value of the depth of the soil layer, and the corresponding target irrigation quantity and target nitrogen application quantity. Representing a set of included vectors;

2) the model performance is evaluated by Mean Squared Error (MSE), Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The mean square error can reflect the degree of difference between the analog value and the measured value; the average absolute error can avoid the mutual offset of errors, so that the actual prediction error can be accurately reflected; the root mean square error can be used to estimate the variance of the error, independent of sample size, reflecting the degree of dispersion of the model, the smaller the value of the mean square error, the more stable the model, wherein,

wherein S is_iIs an analog value; m_iIs an actual measurement value; i is an observation point; n 'is the total number of observation points, n' is540 (a).

3) Based on two groups of variables of irrigation quantity and nitrogen application quantity, table 3 contrasts and analyzes MSE, MAE and RMSE values of a linear support vector machine, a poly support vector machine, an rbf support vector machine, LOR, LR, a traditional BP neural network and a reverse residual deep neural network, namely an RDR model, provided by the application.

TABLE 3

As can be seen from table 3, the inverse residual deep neural network proposed in the present application performs better than the conventional model. Specifically, the method comprises the following steps: in the aspect of water filling, the MSE of the RDR model is averagely improved by 60.72 percent compared with the traditional model, the MAE of the RDR model is improved by 48.94 percent compared with the traditional model, and the RMSE of the RDR model is improved by 68.65 percent compared with the traditional model; in the aspect of fertilization, the MSE of the RDR model is improved by 75.82% compared with that of the traditional model, the MAE of the RDR model is improved by 62.91% compared with that of the traditional model, and the RMSE of the RDR model is improved by 79.90% compared with that of the traditional model.

The reverse residual deep neural network (RDR) model constructed by the method has the following advantages:

1) creatively adopting pluggable Block design, constructing a hidden layer into a residual function to achieve high-dimensional mapping of a model, and highly fitting an actual application scene of irrigation and water conservancy; the model adopts a reverse training strategy, can directly simulate and optimize a water and fertilizer strategy according to the salt content of target soil, and is an innovation of farmland water conservancy.

2) The RDR model after training and adjustment can well predict the irrigation quantity and the nitrogen application quantity, the mean square error, the mean absolute error and the root mean square error of the RDR model are verified, the reverse RMSE value range of the model is 0.0024-0.0049, the model precision is high, the simulation requirement is met, and a powerful technical means is provided for optimizing the field irrigation and fertilization strategy.

3) The method is used for designing and constructing a brand-new deep neural network model and a brand-new training strategy aiming at the specific problems of the irrigation and water conservancy, primarily exploring artificial intelligence to solve the specific problems of the irrigation and water conservancy field based on a machine learning theory and field actual measurement data, and has important significance for improving salinized soil and saving fresh water resources.

S16, applying the trained inverse residual deep neural network, specifically:

1) source of model data, specifically:

irrigation and fertilization data: the tests were conducted in the former flag red-wei test station (108 ° 45'-109 ° 36' E, 40 ° 30'-40 ° 40' N) of inner mongolian Ura in 2016, 2017. In the test, brackish water is adopted for carrying out drip irrigation in different growth periods of the medlar for 8 times all the year, the irrigation quota, namely the irrigation quantity is set to be 232.5mm, 285mm and 375mm, the total three levels are three, the nitrogen fertilizer applied in the test is urea, diammonium phosphate and phosphorus nitrate, and the nitrogen application quantity is set to be 210kg/hm²、525kg/hm²、750kg/hm²、975kg/hm²Total 4 levels. And 3 times of repeated tests are set for each treatment, and a 1m water-stop sheet is adopted for seepage-proofing isolation among cells.

Wherein, different growth periods include: the method comprises the following steps of (1) a germination period, a spring shoot growing period, a flowering period, a fruit expansion initial period, a fruit expansion period, a summer full fruit initial period, a summer full fruit period and a summer full fruit later period, wherein the germination period is carried out from the 1 st day, the spring shoot growing period is started from the 4 th day, the flowering period is started from the 28 th day, the fruit expansion initial period is started from the 42 th day, the fruit expansion period is started from the 60 th day, the summer full fruit initial period is started from the 79 th day, the summer full fruit period is started from the 94 th day, the summer full fruit later period is started from the 125 th day, and the 1 st day, the 4 th day, the 28 th day, the 42 th day, the 60 th day, the 79 th day, the 94 th day and the 125 th day are preset time points;

soil is taken by a soil drill before and after each irrigation, the depth of the soil taking layer is 100cm, wherein 0-40 cm per 10cm is a soil taking layer, 40-100 cm per 20cm is a soil taking layer, and the total number of the soil taking layers is 7. The observation items include soil moisture content, soil conductivity, namely soil EC value and soil pH value.

Selecting target soil salinity parameters:

the soil conductivity can represent the soil salinization degree, the soil pH value can represent the soil pH value, and no unified division standard is available for the soil salinization degree and the soil pH value. The standard of the salinization degree of the soil in the inner Mongolia river-sleeve irrigation area is divided according to the conductivity and the total salt content of the soil, and the standard of the EC value of the soil of 0-100cm is shown in Table 4. The alkaline earth classification standard which is characterized by pH in certain area of inner Mongolia is summarized by referring to the division of alkaline earth at home and abroad, and as shown in table 5, the reference standard is adopted for soil division in the test area.

TABLE 4

Type of soil	Mild alkaline earth	Moderate alkaline soil	Heavy alkaline earth	Alkaline earth
					pH value	7.5～8.5	8.5～9.5	9.5～10.0	＞10.0

TABLE 5

The average EC value of 0-100cm of the current soil is 1.58mS/cm, and the soil belongs to heavily salinized soil; the pH value is 8.12, the soil belongs to mild alkaline soil, and the average water content of each soil layer is 26.25%. Therefore, the target EC values of the heavily saline soil, the moderately saline soil and the lightly saline soil are respectively set to be 1.28mS/cm, 0.85mS/cm and 0.61mS/cm, and the target pH average value of the three types of saline soil is 8. And on the basis, an optimized water-fertilizer scheme, namely a water-fertilizer strategy of the saline soil, aiming at the severe saline soil, the moderate saline soil and the mild saline soil, is constructed by utilizing RDR.

2) Response of soil conductivity to different water and fertilizer strategies and distribution rules thereof (non-RDR model analysis): as can be seen from FIGS. 7 to 9, the maximum values of the conductivity of the treatments occur in the soil layers of 0 to 10cm and 80 to 100cm, and most of the salt is accumulated in the surface layer or deep layer of the soil. The average conductivity value of the W3N0 treatment in each treatment was the smallest and 1.13mS/cm, and the average conductivity of the W2N0 treatment was the largest and 2.27 mS/cm. In 8 treatments of W1 and W2, the EC values show the trend that the EC values are reduced firstly and then increased and then slowly reduced along with the increase of the soil layer depth; and the EC value under the treatment of W3 is reduced in a 0-30cm soil layer and is basically kept unchanged in a 30-100cm soil layer. In the W1 treatment, the soil salt is divided into N3 > N1 > N0 > N2, in the W2 treatment, the soil salt is divided into N0 > N1 > N3 > N2, in the W3 treatment, the soil salt is divided into N3 > N1 > N2 > N0, and the nitrogen application level under different irrigation levels has different influences on the EC value of the soil.

Irrigation and fertilization of brackish water in salinized irrigation areas have important influence on soil salinity. The balance relation between the salt content of the salinized soil and the salt content brought into/leached by the brackish water is complex, and the mechanism that the saline content of the soil is influenced by changing the soil structure by applying fertilizer under the condition of brackish water irrigation needs to be further disclosed. The average conductivity value of the soil under different water and fertilizer treatments is analyzed by variance, as shown in fig. 6, the average conductivity of the soil under the treatment of W1N3 is the largest and is 1.8643mS/cm, and the soil under the treatment of W1 has no significant difference in different fertilization levels; and the average conductivity of the soil treated by W2N2, W3N2 and W2N3 is respectively and obviously reduced by 38.01 percent, 33.41 percent and 32.18 percent compared with that of the soil treated by W1N 3. The two-factor variance significance test is carried out on the soil conductivity of different irrigation and fertilization treatments, the irrigation level difference is significant (P is less than 0.05), and the influence of fertilization and water and fertilizer interaction on the soil conductivity is not significant. Further, based on multiple comparisons of soil EC values with water irrigation by Bonferroni method, as shown in table 7, significant differences between W1 and W2, and between W1 and W3, and no significant difference between W2 and W3 were found. Judging W2N2 to be better treatment based on the analysis of variance result, and W3N2 times; it can be seen that the traditional analysis method can only select the optimal treatment from the existing water irrigation and fertilization levels, the limited field test limits the number of test treatments, and the simulation of the water irrigation and fertilization amount in each treatment room (such as within the W1 and W2 levels) cannot be carried out.

TABLE 6

(I) Irrigation water	(J) Irrigation water	Average difference (I-J)	Significance of
				W1	W2	0.6029*	0.000
	W3	0.3733*	0.016
				W2	W1	-0.6029*	0.000
	W3	-0.2295	0.237
				W3	W1	-0.3733*	0.016
	W2	0.2295	0.237

TABLE 7

3) Field data verification based on the RDR model:

when the RDR model is used for solving the field practical problem, the validity of the model needs to be verified in advance. Considering that the output end of the reverse model RDR is different water and fertilizer strategies, namely different water and fertilizer strategies of the saline soil, including different target irrigation amounts and different target nitrogen application amounts; in the actual field test, 3-4 variables are set in advance by using irrigation and fertilization processing, and the number of the variables is not consistent with the number of output ends of a reverse training model; the input end of the reverse model is the data of light, medium and heavy salinized soil, and the test field is the heavy salinized soil at the moment and cannot provide the data of the salt environment of the light, medium and heavy salinized soil. Therefore, the field data verification part adjusts the RDR structure to be in the forward direction, the input end of the model is the actual field irrigation quantity and nitrogen application quantity, the value of the soil conductivity, the value of the soil pH value and the value of the soil moisture content are output in a simulation mode, and the actual soil data, namely the actual value of the soil conductivity, the actual value of the soil pH value and the actual value of the soil moisture content, are compared and analyzed, and the results are shown in the figures 10 to 12. The result shows that the RMSE value range is between 0.0131 and 0.1814, and the precision is high.

4) Optimizing water and fertilizer coupling strategy simulation of saline soil of different degrees based on RDR, specifically:

the RDR model can simulate the total irrigation amount and the total nitrogen application amount under different salinized soil targets, can also simulate the irrigation amount and the fertilization amount in different growth periods, and controls the irrigation time and the irrigation frequency.

1) The slightly saline soil is taken as a target for explanation, and specifically: the preset value of the soil conductivity of the slightly saline soil is 0.61mS/cm, the target value of the soil pH value is 8.0, the target value of the soil water content is 26.25%, the preset time point is set as the time point when the growth period is finished, and by combining with an RDR model, the obtained target irrigation quantity, namely the optimized total irrigation quantity is 284.17mm, the target nitrogen application quantity, namely the optimized total fertilizer application quantity is 603.01kg/hm²That is, the total water irrigation during the whole growth period is 284.17mm, and the total water irrigation during the whole growth period is 603.01kg/hm²；

2) The medium saline soil is taken as a target for explanation, and specifically: the preset value of the soil conductivity of the moderately saline soil is 0.85mS/cm, the target value of the soil pH value is 8.0, the target value of the soil water content is 26.25%, the preset time point is set as the time point when the growth period is finished, the RDR model is combined, the obtained target irrigation quantity, namely the optimized total irrigation quantity is 280.06mm, the target nitrogen application quantity, namely the optimized total fertilizer application quantity is 588.95kg/hm²That is, the total water irrigation during the whole growth period is 280.06mm, and the total water irrigation during the whole growth period is 588.95kg/hm²；

3) The description is given by taking heavily saline soil as an object, and specifically comprises the following steps: the preset value of the soil conductivity of the moderately saline soil is 1.28mS/cm, the target value of the soil pH value is 8.0, the target value of the soil water content is 26.25%, the preset time point is set as the time point when the growth period is ended, the RDR model is combined, the obtained target irrigation quantity, namely the optimized total irrigation quantity is 267.61mm, and the target nitrogen application quantity is namely the target nitrogen application quantityThe optimized total fertilizing amount is 537.80kg/hm²That is, the total water irrigation during the whole growth period is 267.61mm, and the total water irrigation during the whole growth period is 537.80kg/hm²；

4) Taking 8 times of drip irrigation as an example, the preset time points of each time are set on the 1 st, 10 th, 28 th, 42 th, 60 th, 79 th, 94 th and 125 th days to correspond to the time points of entering different growth periods, then:

the slightly saline soil is used as an objective for explanation, and specifically: the preset value of the soil conductivity of the slightly saline soil is 0.61mS/cm, the target value of the soil pH value is 8.0, and the target value of the soil water content is 26.25%, and by combining an RDR model, the target irrigation amounts on the 1 st day, the 10 th day, the 28 th day, the 42 th day, the 60 th day, the 79 th day, the 94 th day and the 125 th day are obtained, wherein the preset value of the soil conductivity of the slightly saline soil is the average value of the soil conductivity of 0-100cm soil layers, the soil pH value of the slightly saline soil is the average value of the soil pH values of 0-100cm soil layers, and the soil water content of the slightly saline soil is the average value of the soil water content of 0-100cm soil layers;

secondly, medium saline soil is taken as a target for explanation, and specifically: the preset value of the soil conductivity of the moderate saline soil is 0.85mS/cm, the target value of the soil pH value is 8.0, and the target value of the soil water content is 26.25%, and by combining an RDR model, the target irrigation quantity on the 1 st day, the 10 th day, the 28 th day, the 42 th day, the 60 th day, the 79 th day, the 94 th day and the 125 th day is obtained, wherein the preset value of the soil conductivity of the moderate saline soil is the average value of the soil conductivity of 0-100cm soil layers, the soil pH value of the moderate saline soil is the average value of the soil pH value of 0-100cm soil layers, and the soil water content of the moderate saline soil is the average value of the soil water content of 0-100cm soil layers;

thirdly, explaining by taking the heavily saline soil as a target, specifically: the preset value of the soil conductivity of the moderately saline soil is 1.28mS/cm, the target value of the soil pH value is 8.0, and the target value of the soil water content is 26.25%, and by combining an RDR model, the target irrigation quantity on the 1 st day, the 10 th day, the 28 th day, the 42 th day, the 60 th day, the 79 th day, the 94 th day and the 125 th day is obtained, wherein the preset value of the soil conductivity of the heavily saline soil is the average value of the soil conductivity of 0-100cm soil layers, the soil pH value of the heavily saline soil is the average value of the soil pH values of 0-100cm soil layers, and the soil water content of the heavily saline soil is the average value of the soil water content of 0-100cm soil layers;

when the soil is slightly saline soil, moderate saline soil and severe saline soil, the target irrigation quantity and the target nitrogen application quantity of the medlar in each growth period are shown in figures 13 and 14, and the irrigation difference under different degrees of saline soil is mainly reflected in the initial fruit expansion stage and the early fruit bearing stage in summer; the fertilization difference is also at the early fruit enlargement stage and at the early fruit summer stage. Therefore, the irrigation and fertilization control in the two growth periods has important significance for controlling soil salinization.

5) The method comprises the following steps of simulating water, fertilizer and salt under saline soil of different degrees based on an RDR model, specifically:

the soil conductivity and different water and fertilizer coupling strategies, namely the saline soil water and fertilizer strategy, present a complex nonlinear relation, and the RDR model can better fit the complex relation among salinity, irrigation quantity and fertilization. As can be seen from fig. 15 to 17:

if the degree of salinization of soil is reduced, the irrigation amount and the fertilization amount need to be increased to a certain degree. Although brackish water contains some salt as a source of irrigation water, the salt flushed out by irrigation is still more than the salt carried into the soil by the brackish water itself, for example:

if the original heavy saline soil is improved into light saline soil, the proper brackish water irrigation level is 283-284 mm, and the fertilization level is 601-605 kg/hm². As can be seen from fig. 15, the soil EC value and the irrigation level map indicating the irrigation amount and nitrogen application amount are flat, and the minimum value appears at the edge.

If the original heavy saline soil is improved to medium saline soil, the irrigation amount is preferably 277-281 mm, and the nitrogen application amount is 581-595 kg/hm²It is more suitable. As can be seen from FIG. 16, the soil EC values and the irrigation level map indicating the irrigation amount and nitrogen application amount are convex, and the minimum value appears at the edge. The proper irrigation level of the heavily saline soil is 266-268 mm, and the proper fertilization level is 531-543 kg/hm². As can be seen from FIG. 17, the soil EC values and irrigation levels are concave, with the minimum occurring in the irrigation rangeAnd (4) the following steps.

In salinized soil, when the irrigation quantity is constant, the EC value of the soil shows three rules: firstly, the EC value of the soil is increased along with the increase of the fertilizing amount (slightly saline soil); secondly, the EC value of the soil is decreased and then increased along with the increase of the fertilizing amount (moderate and severe saline soil); and thirdly, the EC value of the soil is reduced along with the increase of the fertilizing amount (the soil is heavily saline). When the fertilizing amount is fixed, the EC value of the soil shows three different rules: firstly, the EC value of the soil is reduced along with the increase of the irrigation quantity (slightly saline soil); the EC value of the soil rises firstly and then falls (moderately saline soil) along with the increase of the irrigation quantity; and thirdly, the EC value of the soil is firstly reduced and then increased along with the increase of the irrigation quantity (the soil is heavily saline).

As shown in FIG. 15, the irrigation amount of the slightly saline soil is 283.6-284.8 mm, and the soil salinity is increased to increase the nitrogen application amount when the irrigation amount is constant, so that the whole soil is in an upward trend; when the fertilizing amount is constant, the EC value of the soil in the slightly saline soil is in a whole descending trend along with the increase of the irrigation amount. The soil salinity in the slightly saline soil is less, the soil is greatly influenced by two factors of water filling and fertilizer application, on one hand, the EC value of the soil is increased by applying the fertilizer at a certain time according to the water filling amount, on the other hand, the water filling amount of the slightly saline water is less, the salt brought into the soil by the slightly saline water is limited when the fertilizer is applied at a certain time, and the effect of washing the soil salinity by the slightly saline water is exerted to a greater extent.

In moderately saline soil, the soil salinization rule of a part of soil is the same as that of slightly saline soil when the irrigation quantity is constant, and the soil EC value change rule of moderately and severely saline soil is complex. When the water irrigation amount is constant, the EC value of the soil under the partial water irrigation amount (279.5-281 mm) after the nitrogen fertilizer application is increased rises at different rising rates, and the EC value of the soil under the other water irrigation amounts (278-279.5 mm) firstly falls and then rises along with the increase of the fertilizer application amount, as shown in figure 16. The part of the irrigation quantity which is increased by a certain fertilizing amount and the salt content which is reduced shows that the soil leached by the brackish water has more salt content, when the fertilizing amount is increased to a certain degree, the salt content leached by the brackish water is less than the salt content which is brought into the soil by the nitrogen fertilizer, and the salt content is increased along with the increase of the fertilizing amount. When the fertilizing amount is constant, the EC value of the soil in the moderately saline soil integrally rises and then falls along with the increase of the watering amount, and the salt rises and falls at different rates under different fertilizing amounts. The soil salt leaching capability of the brackish water is gradually increased along with the irrigation quantity. When the irrigation quantity is small, the salt leaching capacity is small, the EC value of the soil is in an ascending trend, and at the moment, the soil salt comes from two aspects of brackish water and nitrogen fertilizer. Along with the increase of the irrigation quantity, when the leaching capacity of brackish water is increased, the EC value of the soil is in a descending trend.

In the heavily saline soil, the salinization rule of partial soil is the same as that of moderately saline soil when the irrigation quantity is constant. In heavily saline soils, a "diagonal line" appears, as shown in FIG. 17. And when the irrigation amount is in the range of 267-268.5 mm and the irrigation amount is unchanged, the EC value of the soil is reduced when the fertilizing amount is increased, and at the moment, the slightly saline water leaches more soil salt. In addition, the content of salt in soil has great influence on the quantity of soil aggregates, the increase of salinization degree can cause the quantity of soil plough layer aggregates to be reduced [19], and the increase of fertilizing amount in heavily salinized soil can cause the soil structure to be looser and easy to wash salt. When the irrigation quantity is within the range of 266-267 mm, the soil EC value firstly decreases along with the increase of the fertilization quantity, and after the soil EC value decreases over the broken line, the soil EC value increases. The EC value of the soil is reduced to show that the soil structure is loosened by fertilization at the moment, and more salt is leached by water; and reach the function department that "broken line" represents, the ability that represents fertilizer soil loosening structure reaches the threshold value, and soil is too tight at this moment, and difficult salinity is difficult to be rinsed, and soil salinity increases. The loosening capacity threshold of the fertilizer to soil is different under different irrigation quantities, so that a 'broken line' function is formed. When the fertilizing amount is fixed, the EC value of the soil is firstly reduced and then increased along with the increase of the irrigation amount. With the increase of irrigation quantity, the salt content carried into the soil by the brackish water is gradually more than that leached by the brackish water, and the EC value of the soil is increased.

Compared with traditional fresh water irrigation, brackish water irrigation provides water needed by crop growth on the one hand, and increases salinity in soil on the other hand, and influences soil quality and crop growth. Researches show that after the growth period of crops is over, the soil in a soil layer of 0-20 cm after brackish water treatment has no obvious phenomena of desalination and salt accumulation; soil layers of 20-40 cm and 40-120 cm are in salt accumulation states of different degrees. Research results show that 3g/L of brackish water can be used as irrigation water for winter wheat, but continuous use can cause salt accumulation in soil. The field test result shows that the saline water irrigation is carried out on the saline soil, the soil in the growth period is in a salt accumulation state, but the saline water irrigation does not have great threat to the growth and the yield of crops. Therefore, the long-term utilization of brackish water for irrigation can generate adverse effects on soil, and the scientific, reasonable, efficient and safe utilization of brackish water is always the core problem concerned by farmland irrigation.

Under the condition of a certain underlying surface and drainage conditions, the soil salinity increment depends on the mineralization degree of irrigation water, the irrigation times and the irrigation quantity. Some researchers adopt a 'brackish water' irrigation mode to carry out long-term brackish water irrigation, and the result shows that the desalination amount of the whole research area in the non-growth period is larger than the salt deposition amount in the growth period, and the salinity is mainly discharged outside the area through a drainage ditch system, so that the soil is in a desalination trend overall. From the aspect of irrigation quantity, researches show that brackish water infiltrates to wash the alkaline soil, the salt content of the same soil layer is gradually reduced along with the increase of infiltration water quantity, and the salt washing depth is increased. The application explores the suitable brackish water irrigation quantity based on the soil salinity, not only considers the factor that the brackish water brings the salinity into the soil, but also analyzes the leaching effect of the brackish water on the soil salinity, and has certain scientific basis.

The fertilization also has certain influence on the soil salinity, and when the irrigation amount is constant, the soil salinity index can be improved by increasing the fertilization fertilizer. There is a paper that "whether light or moderately saline soil, the depth to which salt is washed increases with increasing irrigation. When the irrigation amount is the same, the soil is more saline when nitrogen is applied. "substantially in agreement with the results presented in fig. 15 and 16 of the present application. There are other documents that also conclude that "increasing nitrogen fertilizer at the same irrigation level increases the relative salinity of the soil". Studies have shown that over-fertilization during early plant development may lead to salinization and reduced pod yield. In the application, the fertilizer used for topdressing in the test area is urea, and researches show that the hydrolyzed urea is mainly NH4⁺、NO ^3-The (ionic) form exists in soil for crops to absorb and utilize, so that the excessive application of urea in saline soil increases the soil salinity except the absorption and utilization of crops. The present application is based on the above research resultsThe complex nonlinear relation between brackish water irrigation-fertilization and soil salinity is considered, the influence of the irrigation and fertilization factors on the soil salinity is comprehensively analyzed by adopting an efficient means, and the soil salinity dynamic rule is more comprehensively and accurately explained.

Most of the traditional water and fertilizer optimizing methods are based on the change of the soil conductivity under different irrigation strategies in the growth period so as to select the optimized water and fertilizer amount, and the optimal treatment is W2N2 based on traditional analysis means such as variance analysis. Although such methods are feasible, certain disadvantages remain to be improved. The optimized value under the method can only be screened in the existing water and fertilizer treatment, and whether the optimal irrigation quantity or fertilization quantity appears between the two treatments cannot be ensured.

Further, many scholars seek a suitable application range by establishing a water-fertilizer interaction curved surface through a simulation means. At present, the curved surface of the water and fertilizer interaction effect is researched aiming at crop yield, quality and the like, and the soil salinity is analyzed innovatively.

In the application, the drawing method of the water-nitrogen interaction effect curved surface is different from that of the predecessor. The two-factor interactive effect curved surface which is currently researched is mostly obtained by two-dimensional relationship calculation, and a two-dimensional regression model is established by using Matlab so as to draw a three-dimensional water-nitrogen interactive effect curved surface; the method includes the steps that an artificial intelligence model RDR learning data rule is utilized to directly obtain irrigation and fertilization values under target saline soil, namely a saline soil water and fertilizer strategy corresponding to soil with soil conductivity as a preset value, namely a preset value of the soil conductivity at a preset time point for representing the soil salinization degree is determined, and a target value of each parameter related to the soil salinization at the preset time point is determined; inputting the preset values and all target values into a trained reverse residual error neural network, obtaining a saline soil water and fertilizer strategy corresponding to the soil with the soil conductivity as the preset values and comprising target irrigation quantity and target nitrogen application quantity, further drawing a water and nitrogen interaction effect three-dimensional graph, and obtaining a RDR (resource data record) optimized water and fertilizer strategy based on the light saline soil, wherein the RDR optimized water and fertilizer strategy comprises the following steps: the total irrigation amount is 284.17mm and the total fertilizer application amount is 603.01kg/hm2 all the year round, namely the irrigation is basically maintained at the original water level, and the fertilizer application amount needs to be reduced by 19.85 percent on the basis of the original medium fertilizer application. It can be seen that the model analysis results are significantly different from those of the conventional means at the fertilization level. In the traditional analysis means, the difference between brackish water irrigation in a salinized irrigation area and fertilization treatment under a nitrogen fertilizer application condition is not obvious, so that the mechanism of influencing soil salinity by different fertilization levels cannot be further analyzed, but the RDR can accurately learn any characteristics of the influence of irrigation and fertilization on soil salinity, and further more strictly optimize the nitrogen application amount between N1 and N2.

The continuity of soil layers of different depths is mostly considered in the fitting relation between the current related water fertilizer and salinity, and the corresponding water fertilizer strategy is analyzed by taking the soil salinized soil of different degrees as continuous variables. The method is primarily explored aiming at water and fertilizer strategies under different salinized soils, and complex fitting curve relations are presented between irrigation, fertilization and soil EC values in the process from severe salinized soil to mild salinized soil.

Traditional artificial intelligence models such as a linear support vector machine, a BP neural network and the like are forward training models, and have the defects that the number of layers is too shallow, and an activation function is not used or cannot be used correctly. The RDR model breaks through a forward training mode of a traditional model, and a reverse training strategy is used for directly solving and optimizing a water and fertilizer strategy according to target soil salinity. The RDR has a complex network structure (the model has 6 layers and 768 nodes), and the nonlinear transformation of the model is completed by using a ReLU activation function with less computation loss and no gradient disappearance. The model uses batch normalization to strengthen the stability of training iteration and improve the stability of the model, and a dropout module is used to prevent overfitting of the model. In addition, the RDR innovatively constructs a pluggable Block module, and the recombination mapping relation is x to f (x) -x, so that the optimization and training performance of the model is improved. Therefore, the machine learning model data of two variable groups of irrigation and fertilization prove that the RDR model is more authentic and effective. The water irrigation fertilization-salinity three-dimensional graph further drawn by using RDR model simulation data can not only fully analyze the two-dimensional relationship among indexes, but also explore the complex rule which is not reflected by the two-dimensional relationship to a certain extent.

The salinized soil can be divided into different types such as soda type, halogenated type and the like, and the non-salinized soil is also distinguished due to different soil fertility, soil matrix and the like, so the farmland soil has stronger regional characteristics. The traditional model needs to manually input basic data required by the model before simulating a non-test area irrigation system, and the acquisition of the basic data consumes manpower and material resources in the field sampling process and the like. On the premise that enough irrigation and construction systems and soil data exist, the RDR obtained by big data training has stronger generalization capability, and the optimal irrigation and construction system of a certain area can be simulated and predicted by a target value under the condition that no basic data exists in the new area.

Under the premise that the factors meet the condition that objective causal relationship exists, target data indexes of the reverse training plan are not limited to soil salinity, can also be set as soil fertility indexes (such as nitrogen, phosphorus and potassium), crop indexes (including crop yield and quality) and the like, and a reverse training system can be established based on sufficient multi-dimensional data support. The reverse training system can learn the mapping relation between the multidimensional input end and the output end, the response degree (expressed by the weight w) of soil salinity, fertility, crop quality and the like to the irrigation system can be clarified, the multidimensional indexes occupy different weights, and the optimized irrigation system and the realization probability thereof can be obtained by inputting expected target data.

The optimized water and fertilizer treatment obtained based on the traditional analysis means is W2N2, the water and fertilizer value which is not set for treatment can be considered in the RDR model simulation result, the result is more rigorous compared with the traditional analysis means, and the learning capability of RDR is not restricted by the significance of factors.

RDR simulation results show that complex fitting curve relations are presented among irrigation, fertilization and soil EC values. If the soil in the test area is improved to be slightly saline soil (0.65mS/cm), the fertilizing amount needs to be reduced by 19.85 percent on the basis of the original medium fertilizing amount, and the total brackish water irrigation amount is basically maintained at the original medium irrigation level.

On the premise of large enough data volume and enough data types, the target data of the RDR model is not limited to soil salinity any more, and the mapping relations (such as soil fertility, crop quality and irrigation level) of other multi-dimensional input ends and output ends can be further learned, so that other requirements in the field are met, and the RDR model has a good development prospect.

The method for acquiring the saline soil water and fertilizer strategy aims to overcome the defects of the traditional water and fertilizer strategy optimization solving and construct a reverse residual error deep neural network (RDR). The RMSE value of the soil salinity index of the RDR model is positively verified to be between 0.0131 and 0.1814, and the simulation precision is high. 79.17% of soil EC value in the test area is higher than the standard of heavily salinized soil (1.24-1.32 mS/cm), and the water and fertilizer optimization strategy is reversely solved according to ideal soil salinity indexes in the salinization irrigation area based on RDR, and the result shows that: if the soil in the test area is improved to be slightly saline soil (0.65mS/cm), the fertilizing amount needs to be reduced by 19.85 percent on the basis of the original medium fertilizing amount so as to reduce soil salinity brought by the fertilizing amount, and the total irrigation amount of brackish water is basically maintained at the original medium irrigation level. The RDR fits a complex nonlinear relation presented by the soil EC value and different water and fertilizer coupling strategies, and reveals a dynamic law between soil salt content and brackish water-fertilizer under saline soil of different degrees.

In recent decades, Artificial Intelligence (AI) technology has advanced greatly in the fields of computer vision, natural language processing, machine translation, medical imaging, medical information processing, robotics, biological information control, and the like, and has higher learning accuracy in many aspects such as data prediction in various scenes. The farmland water and soil environment is a complex soil-plant system, the artificial intelligence is applied to the farmland water and soil field, the trend is not blocked, especially the deep learning with the learning capability can learn the complex nonlinear relation more efficiently and accurately, and the defects which cannot be solved by the traditional analysis method are overcome.

Most of artificial intelligence applied to farmland water and soil at present is a BP neural network, and particularly in the aspect of water, fertilizer and salt coupling relation, a BP neural network model is established to predict multiple progresses of crop yield, but the following defects exist:

firstly, the existing BP neural network is only provided with one or two hidden layers, the number of the layers is too shallow, so that the model has no good learning ability in the training process, the generalization ability of the model is basically and completely lost, and the significance of using the neural network is lost;

most of the existing related works do not use the activation function correctly, and the activation function sigmiod used by the BP neural network which is widely applied at present is not suitable for data analysis and prediction in the field of farmland water and soil due to the design defect of neuron saturation and high calculation cost;

the existing BP neural network solves the problem by a local classification method, and the advantage of solving the end-to-end problem by artificial intelligence cannot be fully exerted;

and fourthly, the traditional data analysis method usually needs to input different water and fertilizer combinations for multiple times to predict the crop yield, and the calculation efficiency is low based on a better yield reverse-deducing optimization water and fertilizer mode. In addition, most of artificial intelligence models established in the field range of the current salinization irrigation area are limited to crop yield, and the research based on soil salinity is rare. The research about soil salinity is mostly different machine learning model performance of regional scale comparison, or the artificial intelligence model about soil salinity is established and regional scale prediction salinity is carried out, the artificial intelligence model that farmland scale and soil salinity are relevant is not established yet.

Based on the method, the nonlinear inverse residual error Deep neural network (RDR) with the multilayer Deep network is established for the first time. The method is different from the water and fertilizer optimization mode of the traditional model, the reverse neural network training strategy used in RDR can directly obtain a water irrigation and fertilizer application system by inputting the soil salinity environment, the information acquisition efficiency is improved, and the advantages of artificial intelligence from end to end are fully exerted; a dropout module is added into the RDR to prevent overfitting, and meanwhile, a large number of parameters learned in batch standardized models of batch normalization layers (batch normalization layers) are utilized to guarantee stability of model output. By using the ReLU activation function, the RDR realizes nonlinear analysis, and can solve the difficult problem of complex farmland environment. Meanwhile, the model has strong learning ability and generalization, and the model output function f (x) is skillfully changed into f (x) -x, so that the model precision is improved; the newly constructed pluggable Block (Block) in the model can be widely applied to field data analysis in other areas. The application learns the actually measured field data through a water-fertilizer coupling field test of planting the Chinese wolfberry by drip irrigation with brackish water for two years, and an RDR model is constructed. An optimized water-fertilizer coupling mode beneficial to soil salinity is obtained based on RDR, an important reference basis can be provided for optimizing a water irrigation and fertilization system and soil salt control in a salinized irrigation area, and a powerful and effective scheme is provided for research on comprehensive effects of water and fertilizer in salinized soil.

The main contributions of the present application are summarized below:

1) providing a brand new artificial intelligence model RDR which can directly obtain an optimized water and fertilizer scheme according to ideal soil salinity indexes;

2) and a brand-new artificial intelligence model training strategy beneficial to sustainable development of the saline soil of the farmland is provided.

In the above embodiments, although the steps are numbered as S1, S2, etc., but only the specific embodiments are given in this application, 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 protection scope of the present invention, it is understood that some embodiments may include some or all of the above embodiments.

As shown in fig. 18, an acquiring system 200 for a water and fertilizer strategy of saline soil according to an embodiment of the present invention includes a determining module 210 and an acquiring module 220;

the determining module 210 is configured to obtain a preset value of soil conductivity at a preset time point, where the preset value is used to indicate a soil salination degree, and determine a target value of each parameter associated with soil salination at the preset time point;

the obtaining module 220 is configured to input the preset value and all the target values into the trained inverse residual neural network, so as to obtain a saline soil water and fertilizer strategy including a target irrigation amount and a target nitrogen application amount corresponding to the soil with the preset soil conductivity.

Preferably, in the above technical solution, the method further comprises a modeling module and a training module, wherein the modeling module is used for building a residual neural network;

Preferably, in the above technical solution, the modeling module is further configured to:

introducing batch normalization into the built residual error neural network;

The above steps for realizing the corresponding functions of the parameters and the unit modules in the system 200 for acquiring water and fertilizer from saline soil according to the present invention may refer to the parameters and the steps in the above embodiment of the method for acquiring water and fertilizer from saline soil, which are not described herein again.

The storage medium stores instructions, and when the instructions are read by a computer, the computer executes the method for acquiring the water and fertilizer strategy of the saline soil.

An electronic device according to an embodiment of the present invention includes a processor and the storage medium, where the processor executes instructions in the storage medium. Wherein, the electronic equipment can be selected from computer, mobile phone, etc

As will be appreciated by one skilled in the art, the present invention may be embodied as a system, method or computer program product.

Accordingly, the present disclosure may be embodied in the form of: the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.) or an embodiment combining software and hardware aspects that may all generally be referred to herein 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 having computer-readable program code embodied in the medium.

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. A computer readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the foregoing. 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 of the foregoing. In the context of this application, 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.

Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made to the above embodiments by those of ordinary skill in the art within the scope of the present invention.

Claims

1. A method for acquiring a water and fertilizer strategy of saline soil is characterized by comprising the following steps:

determining a preset value of soil conductivity at a preset time point for representing the soil salinization degree, and determining a target value of each parameter related to the soil salinization at the preset time point;

and inputting the preset value and all the target values into the trained reverse residual error neural network to obtain a saline soil water and fertilizer strategy which corresponds to the soil with the preset value of soil conductivity and comprises the target irrigation quantity and the target nitrogen application quantity.

2. The method for acquiring the water and fertilizer strategy of the saline soil as claimed in claim 1, further comprising:

building a residual error neural network;

3. The method for acquiring the saline soil water and fertilizer strategy according to claim 2, wherein the constructing of the residual neural network further comprises:

4. The method for acquiring the saline soil water and fertilizer strategy according to claim 2 or 3, wherein the building of the residual neural network further comprises the following steps:

introducing batch normalization into the built residual error neural network;

5. A system for acquiring a water and fertilizer strategy of saline soil is characterized by comprising a determining module and an acquiring module;

the determining module is used for determining a preset value of the soil conductivity at a preset time point, which is used for representing the soil salinization degree, and determining a target value of each parameter related to the soil salinization at the preset time point;

6. The system for acquiring the saline soil water and fertilizer strategy is characterized by further comprising a modeling module and a training module, wherein the modeling module is used for building a residual error neural network;

7. The system for acquiring saline soil water and fertilizer strategies according to claim 6, wherein the modeling module is further configured to:

8. The system for acquiring saline soil water and fertilizer strategies according to claim 6 or 7, wherein the modeling module is further configured to:

introducing batch normalization into the built residual error neural network;

9. A storage medium, wherein instructions are stored in the storage medium, and when the instructions are read by a computer, the instructions cause the computer to execute a method for acquiring the water and fertilizer strategy of saline soil according to any one of claims 1 to 4.

10. An electronic device comprising a processor and the storage medium of claim 9, the processor executing instructions in the storage medium.