CN113159446B

CN113159446B - Neural network-based soil nutrient and fruit quality relation prediction method

Info

Publication number: CN113159446B
Application number: CN202110509492.7A
Authority: CN
Inventors: 高志红; 黄霄; 陈涛; 倪照君; 侍婷
Original assignee: Nanjing Agricultural University
Current assignee: Nanjing Agricultural University
Priority date: 2021-05-11
Filing date: 2021-05-11
Publication date: 2024-02-02
Anticipated expiration: 2041-05-11
Also published as: CN113159446A

Abstract

The invention discloses a method for predicting the relation between soil nutrients and fruit quality based on a neural network, which comprises the steps of constructing a prediction model through a BP neural network, a Levenberg-Marquardt BP training function and a Log-sigmoid transfer function, and further predicting the relation between the content of soil mineral elements and fruit quality indexes; and detecting the sensitivity of the prediction model, screening the content of mineral elements in the soil which contributes to the prediction model to the greatest extent, exploring the proper range of the mineral elements with the greatest influence through 3D response surface analysis to obtain the optimal fruit quality, obtaining the law of absorbing the mineral elements by peach trees, and finally formulating a special fertilizer for peach formulas according to the law of absorbing the nutrients by the peach trees to balance the nutrient proportion of the mineral elements, thereby not only meeting the growth and development requirements of the peach trees, but also improving the fruit yield and quality. The invention can be widely applied to agriculture and forestry.

Description

Neural network-based soil nutrient and fruit quality relation prediction method

Technical Field

The invention belongs to the technical field of agricultural data analysis, and particularly relates to a neural network-based soil nutrient and fruit quality relation prediction method.

Background

The fertilizer brands and the formulas in the fertilizer market in China are numerous, farmers can only select the fertilizer to apply according to experience, so that the applied fertilizer can not meet the requirements of peach growth and development, and the development of a proper special fertilizer for peach formulas is an important measure for improving the quality of fruits. The existing special fertilizer for peach cannot meet the growth and development requirements of peach trees in specific areas, the mineral element nutrient proportions of soil in different areas are different, the nutrient requirements of special formula fertilizer are also different, and how to quickly select the special fertilizer suitable for local reality and capable of remarkably improving the quality of peach fruits is imperative.

At present, a plurality of researches are carried out on analyzing or evaluating soil by using a neural network, wherein an artificial neural network (figure 1) is a mathematical model for simulating distributed parallel information processing of human brain neural network behavior characteristics, and the structure is composed of a plurality of simple neurons which are widely connected with each other, and the aim of information processing is achieved by adjusting the interconnection relationship among a large number of nodes.

However, these prior art techniques for predicting soil properties through neural networks have the following problems:

1. the method does not need to destroy farmlands and crops, but only extracts partial characteristics of soil through the images, so that the final result is not accurate enough, and soil mineral elements cannot be really analyzed and predicted;

2. the existing prediction model can only simply give out soil fertility data, but cannot further determine the characteristics between soil fertility and fruit quality and the prediction result, and cannot further adjust a fertilization formula and the like according to actual conditions.

Disclosure of Invention

The invention aims to: the invention aims to solve the defects in the prior art and provides a neural network-based prediction method for the relation between soil nutrients and fruit quality, which is used for carrying out experimental investigation on the mineral nutrient content and fruit quality of a large number of orchards, and constructing a prediction model between the soil nutrients and the fruit quality through an artificial neural network.

The technical scheme is as follows: the invention discloses a neural network-based soil nutrient and fruit quality relation prediction method, which comprises the following steps:

s1, data processing

Respectively preprocessing the collected soil mineral element values and the corresponding fruit quality indexes, and then respectively grouping the two groups of data into a training sample, a test sample and a test sample; obtaining a corresponding characteristic value;

wherein the soil mineral elements include N, P, K, ca, mg, fe, mn, cu, zn and B; the fruit quality parameters comprise single fruit weight, soluble solids content, titratable acid content, solid-acid ratio and edible rate;

s2, constructing a prediction model based on the BP neural network, taking a training sample as input of the BP neural network, and training network parameters;

the BP neural network comprises an input layer, a hidden layer and an output layer; each mineral element value in the soil is an input layer independent variable, namely the number of input layer nodes is 10, and each fruit quality index value is an output layer dependent variable, namely the number of output layer nodes is 5; the hidden layer uses a Levenberg-Marquardt training function and a Log-sigmoid transfer function; in the construction process, a test sample is used for testing a prediction model;

the Log-sigmoid transfer function has the expression:

s3, carrying out classified prediction on the test sample by using the constructed prediction model;

s4, eliminating independent variables one by one based on the obtained prediction model to analyze and evaluate the stability of the prediction model; the independent variable refers to the content of each mineral element in the soil;

s5, obtaining the soil mineral element value range with the largest influence on different quality indexes of the fruits through sensitivity analysis and response surface analysis.

The sensitivity analysis refers to the calculation by eliminating input variables one by one based on the obtained prediction model, namely, N, P, K, ca, mg, fe, mn, cu, zn and B are removed one by one to respectively calculate corresponding indexes (single fruit weight, soluble solid content, titratable acid content, solid-acid ratio and edible rate) of different fruit qualities. And the response surface analysis is performed by constructing a 3D surface graph based on several elements with the largest contribution degree obtained by the sensitivity analysis and the fruit quality index.

In the above step S2, when the network parameters are set, firstly, the training samples, the test samples and the test samples have data amounts of 70%,15% and 15% of soil mineral element content data samples and fruit quality index value samples, respectively, are put together for simulation training, and the mineral element content is used as input data, and the fruit quality data is used as target data.

In addition, matlab software is used for continuously adjusting the number of hidden layers and selecting different training functions and transfer functions to simulate the model so as to obtain the optimal model. In addition, the number of hidden layers can be continuously adjusted when a prediction model is constructed using Matlab.

Further, the data of the input layer and the output layer in the prediction model are normalized as follows:

in the above formula, T is the original data, tn is the normalized output value or input value, T _min And T _max Is the minimum and maximum of the relevant variables.

Further: and (3) randomly selecting 70% of the mineral elements in each soil in the step (S1) as training samples, randomly selecting 15% of the mineral elements as test samples, and randomly selecting 15% of the mineral elements as test samples.

Further, in the step S4, the stability of the prediction model is analyzed and evaluated by the following indexes: mean absolute error MAE, root mean square error RMSE, relative standard error RSE, fitting coefficient R ² And a mean square error MSE;

n is the number of data, -is the average of the variables, mi and Pi represent each fruit quality index measurement and corresponding fruit quality index prediction, respectively.

The beneficial effects are that: according to the method, a prediction model is built through a BP neural network, a Levenberg-Marquardt BP training function and a Log-sigmoid transfer function, so that the relation between the mineral element content of soil and the fruit quality index is predicted; and detecting the sensitivity of the prediction model, screening the content of mineral elements in the soil which contributes to the prediction model to the greatest extent, exploring the proper range of the mineral elements with the greatest influence through 3D response surface analysis to obtain the optimal fruit quality, obtaining the law of absorbing the mineral elements by peach trees, and finally formulating a special fertilizer for peach formulas according to the law of absorbing the nutrients by the peach trees to balance the nutrient proportion of the mineral elements, thereby not only meeting the growth and development requirements of the peach trees, but also improving the fruit yield and quality.

Drawings

FIG. 1 is a diagram of the overall network architecture of the present invention;

FIG. 2 is a schematic diagram of predicting single fruit weight in an embodiment;

FIG. 3 is a schematic diagram of predicted soluble solids content in an example;

FIG. 4 is a schematic representation of predicted titratable acid content in an example;

FIG. 5 is a schematic diagram showing the predicted solid acid ratio content in the examples;

FIG. 6 is a schematic diagram of predicted fruit feeding rates in the examples;

FIG. 7 is a schematic diagram of sensitivity analysis in the examples;

FIG. 8 is a graph of response curve analysis between single fruit weight and soil mineral element content in the examples;

FIG. 9 is a graph showing a response curve analysis between the soluble solids content and the soil mineral element content in the examples;

FIG. 10 is a graph of response curves for titratable acid content and soil mineral element content for the example;

FIG. 11 is a graph showing a response curve analysis between the solid-acid ratio and the mineral element content of the soil in the examples;

FIG. 12 is a graph of response curve analysis between the edible rate and the mineral element content of the soil in the example.

Detailed Description

The technical scheme of the present invention is described in detail below, but the scope of the present invention is not limited to the embodiments.

The invention discloses a neural network-based soil nutrient and fruit quality relation prediction method, which comprises the following steps:

s1, data processing

Respectively preprocessing the collected soil mineral element values and the corresponding fruit quality indexes, and then respectively grouping the two groups of data into a training sample, a test sample and a test sample;

s2, as shown in FIG. 1, constructing a prediction model based on the BP neural network, taking a training sample as input of the BP neural network, and training network parameters;

the BP neural network comprises an input layer, a hidden layer and an output layer; each mineral element value in the soil is an input layer independent variable, namely the number of input layer nodes is 10, and each fruit quality index value is an output layer dependent variable, namely the number of output layer nodes is 5; the hidden layer uses a Levenberg-Marquardt training function and a Log-sigmoid transfer function; and using the test sample to test the model;

s4, based on the obtained prediction model, eliminating independent variables one by one to analyze and evaluate the stability of the prediction model; the independent variable refers to the content of each mineral element in the soil;

The Levenberg-Marquardt training function and the Log-sigmoid transfer function in the embodiment can enable the prediction model to obtain the best performance. The Levenberg-Marquardt BP training function has second-order convergence rate and fewer iteration times, and the convergence rate and stability of the algorithm are greatly improved. The Log-sigmoid transfer function has no limitation on the input variable, which can be any value between plus and minus infinity, and the output is normalized to the (0-1) range. It is commonly used in multi-layer networks trained with back propagation algorithms, mainly because the function is easily distinguishable and is slightly scalable. It has excellent mathematical properties, and many numerical optimization algorithms can be directly used to solve the optimal solution.

In this embodiment, other functions, such as a BFG training function, a CGB training function, a CGP training function, an SCG training function, a linear transfer function r, a Tangent-sigmoid transfer function, etc., may be used as the training function and the transfer function in the prediction model, but the prediction accuracy and the final detection evaluation value of the prediction model constructed by using these training functions and transfer functions are not the same as those of the prediction model of the present invention. As shown in table 1, taking predicted fruit weight per fruit as an example, the comparison results are as follows:

TABLE 1 comparison of the effects of the model of the invention on predicting single fruit weight index of fruits with other models

When predicting other quality indexes of fruits, other transfer functions and training functions can be adopted in the prediction model, but the final prediction result is the same as the embodiment, and the prediction precision and the evaluation test precision are not the same as those of the prediction model of the invention.

In this embodiment, the data of the input layer and the output layer of the prediction model are normalized as follows:

in the above formula, T is the original data, tn is the normalized output value or input value, T _min And T _max Is the minimum and maximum of the relevant variables. In each soil mineral element in step S1, 70% is randomly selected as a training sample, 15% is randomly selected as a test sample, and 15% is randomly selected as a test sample.

In step S4, the stability of the prediction model is analyzed and evaluated by the following indexes: mean absolute error MAE, root mean square error RMSE, relative standard error RSE, fitting coefficient R ² And a mean square error MSE;

example 1

The first stage:

soil and corresponding peach fruits were collected and analyzed in 75 peach orchards in Jiangsu 75 different areas as shown in tables 2 and 3.

TABLE 2 analysis of fruit quality index in peach orchard

	Weight per gram of single fruit	Soluble solids/%	Titratable acid/%	Ratio of solid to acid	Edible rate/%
						Average value of	281.988	13.271	0.322	44.416	94.526
Maximum value	393.300	17.960	0.536	95.280	96.253
						Minimum value	162.200	9.600	0.140	19.730	92.696
Standard deviation of	46.001	1.496	0.084	14.284	0.905
						Coefficient of variation (%)	16.313	11.277	0.261	32.159	0.957

TABLE 3 analysis of mineral element content in peach orchard soil

And a second stage:

and constructing a prediction model based on the BP neural network, wherein the prediction model comprises a Log-sigmoid transfer function and a Levenberg-Marquardt training function.

And then predicting the single fruit weight, the soluble solid content, the titratable acid content, the solid-acid ratio and the edible rate of the peach fruits by utilizing the prediction model through the mineral element content in the soil.

The analysis was then compared with the actual acquisition analysis data of the first stage (tables 2 and 3).

First, predicting single fruit weight of fruit

Prediction accuracy R of the present embodiment ² 0.9735, RMSE 0.0482, MSE is 0.0023, MAE is 0.0308 and RSE is 0.1155. Further comparisons of predicted and measured values of single fruit weights (fig. 2) showed that during the training phase, their distribution patterns were very close (fig. 2 a) while they had nearly identical box-plot structure (fig. 2 b). At the same time, all predicted and measured values of single fruit weights present similar line graph results (fig. 2 c), indicating that the predictive model of the present invention is capable of accurately predicting single fruit weights.

(II) predicting soluble solids content of fruit

Prediction accuracy R of the present embodiment ² For 0.9607, its prediction model evaluation coefficient RMSE is 0.0598, mse is 0.0036, mae is 0.0401 and RSE is 0.1432.

Further comparisons of predicted and measured values of soluble solids content (fig. 3) showed that during the training phase they were distributed in very close proximity (fig. 3 a) while they had a similar box-plot structure (fig. 3 b). At the same time, all predicted and measured values of the soluble solids content present similar line graph results (fig. 3 c), indicating that the predictive model of the present invention can accurately predict the soluble solids content of the fruit.

(III) predicting the titratable acid content of fruits

Prediction accuracy R of the present embodiment ² For 0.9036, the prediction model evaluation coefficient RMSE is 0.1045, mse is 0.0109, mae is 0.0765 and RSE is 0.2502.

Here, the predicted and measured values of titratable acid content are further compared (fig. 4), and the result shows that during the training phase, their distribution patterns are very close (fig. 4 a) while they have a similar box-pattern structure (fig. 4 b). At the same time, all predicted and measured values of titratable acid content present similar line graph results (fig. 4 c), and the model-based method of the invention can accurately predict the titratable acid content of fruits.

(IV) predicting solid-acid ratio of fruit

Prediction accuracy R of the present embodiment ² For 0.9660, the prediction model evaluation coefficient RMSE was 0.0594, mse was 0.0035, mae was 0.0428 and RSE was 0.1422.

Further comparisons of predicted and measured values of the solid acid ratio index (fig. 5) show that during the training phase, their distribution patterns are very close (fig. 5 a) while they have almost the same box pattern structure (fig. 5 b). Meanwhile, the predicted values and measured values of all the solid acid ratio indexes show similar line graph results (figure 5 c), which shows that the prediction model of the invention can accurately predict the solid acid ratio of fruits.

Fifth, fruit edibility is predicted

Prediction accuracy R of the present embodiment ² For 0.9735, the model evaluation coefficient RMSE was 0.0917, mse was 0.0084, mae was 0.0688 and RSE was 0.2195.

Here, the predicted and measured values of fruit feeding rate are then compared (fig. 6), which shows that during the training phase, their distribution patterns are very close (fig. 6 a) while they have almost the same box-pattern structure (fig. 6 b). At the same time, all predicted and measured values of fruit eating rate exhibited similar line graph results (fig. 6 c), indicating that the predictive model of the present invention can accurately predict fruit eating rate.

The result shows that the soil nutrient and fruit quality relation prediction model can accurately predict the fruit quality index.

Example 2

This example further verifies which mineral element content in the peach garden soil has the greatest effect on fruit quality index.

In this embodiment, the stability of the model is analyzed by removing the input variables (i.e., the content of the corresponding mineral elements in the soil) one by one and performing detection.

(1) Among the mineral element content and single fruit weight prediction models of this example, the prediction model that does not contain Cu content has the lowest RMSE value and the prediction model that does not contain B content has the highest RMSE value. In sensitivity analysis, RMSE value magnitude represents the relative contribution of an input variable to an output variable. The higher the RMSE value, the higher the importance of the input variable that is rejected. Therefore, the influence of mineral element content in peach garden soil on single fruit weight is B > Ca > N > K > P > Fe > Mn > Mg > Zn > Cu sequentially from large to small.

(2) Among the predictive models of mineral element content and soluble solids content of peach fruits in this example, the predictive model containing no Zn content had the lowest RMSE value and the predictive model containing no Fe content had the highest RMSE value. The influence of the mineral element content in the peach garden soil on the content of the soluble solids is that Fe > K > B > Ca > Mn > P > N > Mg > Cu > Zn in sequence from large to small.

(3) Among the mineral element content and peach fruit titratable acid content prediction models of this example, the prediction model containing no Cu content had the lowest RMSE value and the prediction model containing no Ca content had the highest RMSE value. The influence of the mineral element content in the peach garden soil on the content of titratable acid is Ca > N > B > K > P > Fe > Mg > Zn > Mn > Cu in sequence from large to small.

(4) Among the predictive models of mineral element content and peach fruit solid acid ratio in this example, the predictive model not containing K content has the lowest RMSE value and the predictive model not containing B content has the highest RMSE value. The influence of mineral element content in peach garden soil on solid-acid ratio is B > N > Fe > Ca > P > Mg > Mn > Zn > Cu > K sequentially from big to small.

(5) Among the mineral element content and peach fruit eating rate prediction models of this example, the prediction model not containing Zn content has the lowest RMSE value and the prediction model not containing Ca content has the highest RMSE value. The influence of the mineral element content in the peach garden soil on the edible rate is Ca > Fe > N > Mn > P > K > Mg > B > Cu > Zn in sequence from large to small. In summary, N, P, K, ca, fe and B content in peach garden soil has the greatest effect on fruit quality.

As shown in FIG. 7, the sensitivity analysis results of the above examples revealed that the B, ca, N, K, P, fe content in the soil had the greatest effect on the fruit weight.

To explore the appropriate range of these elements further, we performed a response surface analysis (fig. 8).

As shown in FIG. 8a, the response curve analysis of the B, ca content and the single fruit weight in the soil shows that the higher single fruit weight can be obtained when the B content in the soil is 0.2-0.96mg/kg and the Ca content is 204.0-296.0mg/kg, and the higher single fruit weight can be obtained when the B content in the soil is 0.8-1.2mg/kg and the Ca content is 130.0-156.0 mg/kg.

As shown in FIG. 8b, the response curve analysis of the N, K content and the single fruit weight in the soil shows that the higher single fruit weight can be obtained when the N content in the soil is 116.0-182.0mg/kg and the K content is 450.0-600.0mg/kg, and the higher single fruit weight can be obtained when the N content in the soil is 204.0-276.0mg/kg and the K content is 490.0-585.0mg/kg, but the single fruit weight index of the fruit is obviously reduced when the K content in the soil is more than 910.0 mg/kg.

As shown in FIG. 8c, the response curve analysis of the P, fe content and the single fruit weight in the soil shows that the higher single fruit weight can be obtained when the P content in the soil is 28.0-98.0mg/kg and the Fe content is 16.0-44.0mg/kg, and the higher single fruit weight can be obtained when the P content in the soil is 40.0-137.0mg/kg and the Fe content is 112.0-140.0 mg/kg.

As is clear from the sensitivity analysis results of the above model, the Fe, K, B, ca content in the soil has the greatest effect on the soluble solid content of the fruits.

Example 3:

to further obtain the content range of mineral elements in the soil which have the greatest influence on other different quality indexes of fruits, the response surface method analysis is adopted in the embodiment (fig. 9 to 12).

As shown in FIG. 9a, the response curve analysis of the Fe content and the K content in the soil shows that the soil can obtain higher soluble solid content when the Fe content is 196.0-272.0mg/kg and the K content is 218.0-391.0mg/kg, and the soil can obtain higher soluble solid content when the Fe content is 60.0-140.0mg/kg and the K content is 400.0-836.0 mg/kg. However, when the Fe content in the soil is less than 50.0mg/kg, the soluble solids content is significantly reduced.

Response curve analysis of B, ca content and soluble solids content in soil as shown in fig. 9B, higher soluble solids content can be obtained when the B content in soil is 0.58-0.90mg/kg and the Ca content is 140.0-174.0mg/kg, but the soluble solids content is significantly reduced when the Ca content in soil is greater than 266.0 mg/kg.

Response surface analysis of Ca and N contents and titratable acid contents in soil is shown in FIG. 10a, lower titratable acid contents can be obtained when Ca contents in soil are 168.0-306.0mg/kg and N contents are 62.0-108.0mg/kg, and lower titratable acid contents can be obtained when Ca contents in soil are 168.0-198.0mg/kg and N contents are 71.0-176.0 mg/kg.

Response surface analysis of B, K content and titratable acid content in soil is shown in fig. 10B, wherein lower titratable acid content can be obtained when B content in soil is 0.16-0.62mg/kg and K content is 80.0-160.0mg/kg, lower titratable acid content can be obtained when B content in soil is 0.78-1.06mg/kg and K content is 240.0-360.0mg/kg, and lower titratable acid content can be obtained when B content in soil is 0.80-1.10mg/kg and K content is 720.0-1000.0 mg/kg.

Response surface analysis of P, fe content and titratable acid content in soil As shown in FIG. 10c, lower titratable acid content can be obtained when P content in soil is 11.0-32.0mg/kg and Fe content is 48.0-252.0mg/kg, lower titratable acid content can be obtained when P content in soil is 92.0-167.0mg/kg and Fe content is 200.0-258.0mg/kg, but when Fe content in soil is less than 48mg/kg, the titratable acid content of peach fruits is significantly increased.

As shown in FIG. 11a, the response curve analysis of the B, N content and the solid-acid ratio in the soil shows that the higher solid-acid ratio can be obtained when the B content in the soil is 0.145-0.40mg/kg and the N content is 17.0-115.0mg/kg, and the higher solid-acid ratio can be obtained when the B content in the soil is 0.63-1.03mg/kg and the N content is 162.0-194.0 mg/kg.

As shown in FIG. 11b, the response curve analysis of the Fe content and Ca content in the soil with the solid-acid ratio shows that when the Fe content in the soil is 100.0-155.0mg/kg and the Ca content is 180.0-235.0mg/kg, the higher solid-acid ratio can be obtained.

As shown in FIG. 12a, the response curve analysis of Ca and Fe contents and the edible rate in the soil shows that the higher edible rate can be obtained when the Ca content in the soil is 174.0-234.0mg/kg and the Fe content is 220.0-270.0mg/kg, and the higher edible rate can be obtained when the Ca content in the soil is 170.0-240.0mg/kg and the Fe content is 130.0-170.0 mg/kg.

As shown in FIG. 12b, the response curve analysis of the N, mn content and the single fruit weight in the soil shows that the higher edible rate can be obtained when the N content in the soil is 140.0-210.0mg/kg and the Mn content is 105.0-152.0mg/kg, and the higher edible rate can be obtained when the N content in the soil is 220.0-250.0mg/kg and the Mn content is 37.0-72.0 mg/kg.

As shown in FIG. 12c, the response curve analysis of the P, K content and the single fruit weight in the soil shows that the higher edible rate can be obtained when the P content in the soil is 44.0-140.0mg/kg and the K content is 440.0-650.0 mg/kg. In summary, when the N in the soil is 71-108mg/kg, the P is 92.0-137.0mg/kg, the K is 490.0-585.0mg/kg, the Ca is 170.0-198.0mg/kg, the Fe content is 125-140mg/kg, the Mn content is 37-72mg/kg, the B content is 0.80-1.02mg/kg, and the peach fruit quality index can be obviously improved.

In summary, the fertilizer for peach formula suitable for peach orchard in example 1 comprises the following nutrients: comprises 70-100 parts of nitrogen, 90-140 parts of phosphorus, 500-600 parts of potassium, 170-200 parts of calcium, 10-15 parts of iron, 4-7 parts of manganese and 1-1.5 parts of boron. Wherein, the nitrogen is derived from urea, the phosphorus is derived from dipotassium hydrogen phosphate, the potassium is derived from potassium sulfate, the calcium is derived from calcium sulfate, the iron is derived from ferrous sulfate, the manganese is derived from manganese sulfate, and the boron is derived from borax. Raw materials are taken according to the proportion of nutrient elements, crushed by a crusher, sieved by a 60-mesh sieve, put into a disc granulator, mixed and granulated, dried by a dryer after granulation, cooled by a rotary sieve, and finally the special fertilizer (particle 1-2 mm) for the peach formula is obtained.

Example 4:

according to the above formula, the special fertilizer for peach formula is prepared according to the embodiment, and a certain peach garden in Changzhou city of Jiangsu province is tested, the test variety is lake Jing Milou, 100 kg of the special fertilizer for peach formula is applied to each mu of the test tree, 100 kg of the compound fertilizer (the three-element compound fertilizer with fifteen contents of nitrogen, phosphorus and potassium) is applied to each mu of the test tree by using a common fertilization method, and the test result shows that the land mass, the peach quality index and the like of the special fertilizer for peach formula are obviously improved compared with the common fertilization mode. As shown in Table 4, the weight of the single fruit is improved by more than 7.96%, the content of soluble solids is improved by 3.99%, the solid acid ratio is improved by 33.70%, the edible rate is improved by 0.56%, the titratable acid content affecting the quality of the fruit is obviously reduced by 8%, the comprehensive quality of the fruit is obviously improved, and the special fertilizer for the peach formula has a relatively obvious effect of improving the quality.

Table 4 comparative test of fertilizer effect for peach

Claims

1. A neural network-based soil nutrient and fruit quality relation prediction method is characterized by comprising the following steps of: the method comprises the following steps:

s1, data processing

s3, carrying out classification prediction on the test sample by utilizing the constructed prediction model, wherein the classification prediction comprises predicting the single fruit weight of the fruit, predicting the soluble solid content of the fruit, predicting the titratable acid content of the fruit, predicting the solid-acid ratio of the fruit and predicting the edible rate of the fruit;

s5, obtaining a soil mineral element value range with the greatest influence on different quality indexes of fruits through sensitivity analysis and response surface method analysis, wherein the sensitivity analysis refers to respectively calculating corresponding indexes of different fruit qualities based on the obtained prediction model by removing N, P, K, ca, mg, fe, mn, cu, zn and B one by one, and the corresponding indexes comprise single fruit weight, soluble solid content, titratable acid content, solid-acid ratio and edible rate; the response surface method analysis is performed by constructing a 3D curved surface graph based on several elements with the largest contribution degree obtained by sensitivity analysis and fruit quality indexes.

2. The neural network-based soil nutrient and fruit quality relationship prediction method according to claim 1, wherein: the data of the input layer and the output layer in the prediction model are subjected to standardized processing as follows:

3. The neural network-based soil nutrient and fruit quality relationship prediction method according to claim 1, wherein: and (3) randomly selecting 70% of the mineral elements in each soil in the step (S1) as training samples, randomly selecting 15% of the mineral elements as test samples, and randomly selecting 15% of the mineral elements as test samples.

4. The neural network-based soil nutrient and fruit quality relationship prediction method according to claim 1, wherein: in the step S4, the stability of the prediction model is analyzed and evaluated by the following indexes: mean absolute error MAE, root mean square error RMSE, relative standard error RSE, fitting coefficient R ² And a mean square error MSE;