CN105389452A - Cucumber whole-course photosynthetic rate prediction model based on neural network, and establishment method - Google Patents

Abstract

The invention discloses a cucumber whole-course photosynthetic rate prediction model based on a neural network. A multifactor nesting experiment is utilized for obtaining cucumber seedling photosynthetic rate test data, an LM (Levenberg-Marquardt) training method is adopted to carry out model training, and a cucumber whole-course photosynthetic rate model which combines a growing stage is established and is subjected to model performance parameter comparison and accuracy verification with a photosynthetic rate model of a single growing period and a whole-course photosynthetic rate model which does not combines a growing period stage parameter. A training result indicates that the whole-course photosynthetic rate model established in a way that the growing period is added to serve as a one-dimensional input quantity can effectively pass over local flat areas, and the whole-course photosynthetic rate model has an obvious superiority, meets a training requirement that errors are smaller than 0.0001, and is verified in an xor checkout way, so that a determination coefficient of a model prediction value and an actual measurement value is 0.9897, an error is smaller than 6.559%, and a theoretical basis and technical support can be provided for facility and crop luminous environment regulation and control.

Description

Cucumber whole-course photosynthetic rate prediction model based on neural network and establishment method

Technical Field

The invention belongs to the technical field of intelligent agriculture, and particularly relates to a cucumber whole-course photosynthetic rate prediction model based on a neural network and an establishment method.

Background

The cucumber is one of the main vegetables cultivated in China, and the quality and the yield of the cucumber are indistinguishable from the photosynthesis capacity thereof. Photosynthetic Rate and chlorophyll content, temperature, CO₂A plurality of factors such as concentration, illumination intensity, relative humidity and the like have significant relations. Wherein, the chloroplast is the basic organelle of green plants for photosynthesis, the chlorophyll is the basic component of the chloroplast and is important in the photosynthesis of the plants, the content of the chlorophyll is an important indicator factor of the photosynthesis capacity, the nutritional status and the growth situation of the plants, and the temperature influences the activity of Rubisco activating enzyme in crops, the stomatal conductance and the CO conductivity₂The concentration directly influences the dark reaction rate of crops and the accumulation of dry matters, the illumination intensity is the direct power and the pushing force of photosynthesis, the relative humidity influences the porosity conductivity of leaves and the like, and all factors have mutual influence. Therefore, the influence of multiple factors is comprehensively considered, and a multi-factor coupling whole-course photosynthetic rate prediction model is established, so that the method has an important role in optimizing the cucumber luminous environment.

In recent years, many scholars have conducted relevant research in establishing photosynthetic rate models, and the research considers the correlation between different environmental factors, but the research has the disadvantages of low fitting degree, complex fitting formula, large error and the like. The neural network has the advantages of nonlinear mapping, adaptive learning capability and the like, and is suitable for fitting and predicting a nonlinear complex system model, so that the photosynthetic rate modeling based on the neural network becomes a research hotspot. Recently, a Hopfield network photosynthetic rate model, a BP neural network-based greenhouse tomato leaf stomata conductance model and a WSN-based tomato flowering phase single leaf net photosynthesis rate prediction model have appeared, and the researches apply the neural network to photosynthetic rate modeling from different angles, but do not consider the influence of different growth periods on crops, and do not establish a whole-course cucumber photosynthetic rate prediction model, and the defects of slow training process and large training error difference exist.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention aims to provide a cucumber whole-course photosynthetic rate prediction model based on a neural network and an establishment method, a multi-factor nesting test is designed, data is normalized and then modeled by adopting a BP neural network, a cucumber photosynthetic rate prediction model which is obtained by innovatively distinguishing growth stage information as one-dimensional input factors is established on the premise of considering the original environment and physiological parameters, the cucumber photosynthetic rate prediction model in the whole course is established through comparison and verification, and a foundation is established for light environment regulation and control of facility agriculture.

In order to achieve the purpose, the invention adopts the technical scheme that:

a cucumber whole-course photosynthetic rate prediction model based on a neural network fuses a growing period and has a model formula ofWherein the output signal T_oRepresenting the photosynthetic rate calculated by the neural network, the input signal x ═ x₁',x₂',......,x₆')^T；x₁′、x₂′、x₃′、x₄′、x₅′、x₆Respectively growth period, temperature, CO₂Concentration, light intensity, relative humidity and chlorophyll content; m is 8, n is 6; v. of_ijIs the weight, w, of the input layer to the hidden layer_jIs the hidden layer to output layer weight vector,the net input quantity of the input layer adjusted to the hidden layer by the weight value is shown;the input quantities of the output layers are indicated.

The method for establishing the cucumber whole-course photosynthetic rate prediction model based on the neural network comprises the following steps of:

step 1, acquiring experimental data, wherein the process is as follows:

adopting a nutrition pot for seedling culture, selecting cucumber seedlings with uniform growth vigor, stem cross diameter of 0.6-0.8 cm and plant height within 10cm for test when the cucumber seedlings grow into two leaves and one heart, selecting 150 robust cucumber seedlings as test samples, and selecting 150 plants with flowering nodes 50 cm away from a tap as the test samples in the flowering fruiting period when the cucumbers are in the flowering fruiting period;

measuring net photosynthetic rate, and setting 5 temperature gradients of 16, 20, 24, 28 and 32 ℃ by using a temperature control module in the process; by using CO₂The injection module sets the volume ratio of carbon dioxide to be 300, 600, 900, 1200 and 1500 mu L/L for 5 gradients; obtaining 0, 20, 50, 100, 200, 300, 500, 700, 1000, 1200, 1500 [ mu ] mol/(m) by using the LED light source module²S) 11 photon flux density gradients, a total of 275 trials were performed in a nested fashion, each trial was repeated on 3 plants selected at random, and the chamber relative humidity and chlorophyll content of the leaves tested were recorded in the trials to yield a plot of chlorophyll content, temperature, CO₂1650 groups of test data with input of concentration, illumination intensity and relative humidity and output of net photosynthetic rate, namely 825 groups in seedling stage and 825 groups in flowering and fruiting stage;

step 2, establishing a model

Step 2.1 training method

The input signal is x ═ x₁',x₂',......,x₆')^T；x₁'、x₂'、x₃'、x₄'、x₅'、x₆Respectively growth period, temperature, CO₂Concentration, light intensity, relative humidity and leavesChlorophyll content, output signal T_oRepresenting the photosynthetic rate calculated by the neural network, and the corresponding actually measured photosynthetic rate being a teacher signal T_dEstablishing a whole-course cucumber seedling photosynthetic rate model with a growth period as a one-dimensional factor by a BP gradient training method;

step 2.2 training procedure

Randomly distributing an initial value V of a weight vector from an input layer to a hidden layer and an initial value W of the weight vector from the hidden layer to an output layer; operating BP neural network program, inputting training set sampleAnd according to $T_o=f\[\underset{j=0}{\overset{m}{\Σ}}{w}_{j}f\left(\underset{i=0}{\overset{n}{\Σ}}{v}_{ij}{x}_i\right)\]$ Output of a computing network T_o；

Based on teacher signal T_dAnd a network output signal T_oTotal error of systemIn the formula,in order to train the true value of the sample,the output value of the training sample network is obtained, P is the number of training samples, and l is the number of output layers;

based on teacher signal T_dNetwork output signal T_oWeight vector from hidden layer to output layer, output component of hidden layer, error signal of output layer^o＝(T_d-T_o)T_o(1-T_o) Neuron error signalIn the formula, ω_jAs weight vector from hidden layer to output layer, y_jIs the output of each layer;

adopting LM training method to train network, inputting weight vector from layer to hidden layerWeight ω from hidden layer to output layer_jWhere η is the learning rate, Δ ω is the weight adjustment vector, and Δ w is- (J)_n ^TJ_n+η_nI)^-1J_n ^Tr_n，Is a Hessian matrix used to approximate the objective function, I is an identity matrix, η_nThe parameter greater than 0 in LM training method is used to speed up the training of network, when η_nWhen approaching zero, LM algorithm approaches Gauss-Newton method, with η_nThe LM algorithm is similar to the steepest descent method when the LM is continuously increased;

step 2.3 model building

When E is_RESWhen the error value is smaller than the set error value or the learning frequency reaches the set step number, the training is stopped to obtain the final prediction model.

In the step 2.2, the step of the method,

for training setBook (I)Carrying out normalization processing in an interval of 0.2-0.9, setting the number of hidden nodes of the neural network to be 10, and randomly distributing an initial value V of a weight vector from an input layer to a hidden layer and an initial value W of the weight vector from the hidden layer to an output layer; then operating BP neural network program and inputting training set sampleAnd according toComputing the output y of hidden nodes_j(ii) a According toCalculating the output of the output layer; by passingAnd judging whether the neural network achieves the training precision, if not, reselecting the sample to start training, otherwise, stopping training, and finishing the model establishment.

The invention establishes the mapping of the environmental factors and the plant physiological factors with the photosynthetic rate, thereby effectively regulating and controlling the luminous environment and having important significance for increasing the yield of crops.

Compared with the prior art, the invention has the beneficial effects that:

1) the method provides a cucumber whole-course photosynthetic rate prediction model based on a neural network, effectively distinguishes the difference of photosynthetic rate values of a cucumber seedling stage and a flowering and fruiting stage under different conditions by adding a one-dimensional growth period input quantity, can effectively cross a local flat area in the training process, has no repeated oscillation, and is rapid in convergence and higher in precision than the prediction model in a mixed growth period.

2) The neural network prediction model established by the LM training method has the determination coefficient of 0.9897, has good fitting effect, and can realize the prediction of the photosynthetic rate values of plants in different growth periods. The constructed whole-course cucumber seedling photosynthetic rate prediction model can provide a theory for yellow light seedling stage luminous environment regulation and control.

The whole-course photosynthetic rate prediction model provided by the invention can provide a theoretical basis for cucumber luminous environment regulation and control, and can be expanded to be applied to the establishment of photosynthetic optimization regulation and control models of different crops so as to improve the photosynthetic capacity of greenhouse crops.

Drawings

FIG. 1 is a flow chart of the neural network based algorithm of the present invention.

FIG. 2 is a graph showing the variation of the error in different growth phases of the present invention.

FIG. 3 is a diagram illustrating the correlation between the measured and simulated values of the light-combination rate in the model verification according to the present invention.

Detailed Description

The embodiments of the present invention will be described in detail below with reference to the drawings and examples.

The invention relates to a cucumber whole-course photosynthetic rate prediction model based on a neural network, which is established in the following process:

1. materials and methods

The test is carried out in scientific research greenhouse of northwest agriculture and forestry science and technology university from 4 months to 7 months in 2014. The cucumber variety to be tested is vinc-stichopus japonicus, the soaked seeds are subjected to germination acceleration in a culture dish, low-temperature treatment is carried out when germination is needed, and seedlings are grown in a hole tray with 50 holes (540mm, 280mm and 50mm) by using a nutrition pot. The seedling substrate is a special substrate for agricultural seedling. During the seedling cultivation period, the water and fertilizer are kept sufficient, when the cucumber seedlings grow into two leaves and one heart, the cucumber seedlings with uniform growth vigor, stem cross diameter of 0.6-0.8 cm and plant height of less than 10cm are selected for testing. 150 healthy and strong cucumber seedlings were selected as test samples. During the test period, normal field cultivation management is carried out, no pesticide and hormone are sprayed, and 150 plants with the flowering pitch being about 50 centimeters from the tap are selected as test samples during the flowering fruiting period when the cucumbers are in the flowering fruiting period.

The net photosynthetic rate was measured using a Li-6400XT model portable photosynthetic apparatus manufactured by LI-COR, USA, and during the test, parameters such as temperature around the leaves, CO2 concentration, and illumination intensity were controlled as required using a plurality of submodules selected by the photosynthetic apparatus. Wherein, a temperature control module is used for setting 5 temperature gradients of 16, 20, 24, 28 and 32 ℃; by using CO₂The injection module sets the volume ratio of carbon dioxide to be 300, 600, 900, 1200 and 1500 mu L/L for 5 gradients; obtaining 0, 20, 50, 100, 200, 300, 500, 700, 1000, 1200, 1500 [ mu ] mol/(m) by using the LED light source module²S) 11 Photon Flux Density (PFD) gradients, a total of 275 sets of tests were performed in a nested fashion, each set of tests was tested repeatedly on 3 randomly selected plants, the relative humidity of the leaf chamber was recorded in the tests, and the chlorophyll content of the leaf blade to be tested was recorded using SPAD-502Plus type chlorophyll meter from konica, japan, to form 1650 sets of test data with chlorophyll content, temperature, CO2 concentration, light intensity, and relative humidity as inputs and net photosynthetic rate as output, i.e., 825 sets at the seedling stage and 825 at the flowering result stage.

2. Model building

2.1 training method

In order to establish an optimal photosynthetic rate prediction model, the same modeling method is adopted to establish four models aiming at different growth periods of the cucumber, namely a prediction model only aiming at the seedling stage of the cucumber, a prediction model only aiming at the flowering and fruiting stage of the cucumber, a photosynthetic rate prediction model in the whole course of the cucumber and a prediction model establishing the whole course of the cucumber by taking the different growth periods as one-dimensional input. The input signal is x ═ x₁'x₂'...x₅')^T；x₁'、x₂'、x₃'、x₄'、x₅' temperature, CO respectively₂Concentration, light intensity, relative humidity and chlorophyll content, fourthThe seed model is added into the growing period as one-dimensional input, and output signals all use T_oThe photosynthetic rate obtained by network calculation is represented, and the measured photosynthetic rate corresponding to each group is teacher signal T_d. Establishing a whole-course cucumber seedling photosynthetic rate model T by a BP gradient training method_d'(X')。

As shown in fig. 1, when the BP neural network program runs, randomly allocating an initial value V of a weight vector from an input layer to a hidden layer and an initial value W of a weight vector from the hidden layer to an output layer according to a network connection value and a threshold; operating BP neural network program, inputting training set sampleAnd according toOutput of a computing network T_oAnd triggers the following process:

based on teacher signal and network output signal, the total error of system is calculated by the formula

$E_{RES}=\sqrt{\frac{1}{P}\underset{P=1}{\overset{P}{\Σ}}\underset{k=1}{\overset{l}{\Σ}}{(T_d^p-T_o^P)}^2}---\left(1\right)$

In the formula,in order to train the true value of the sample,the output value of the training sample network is obtained, P is the number of training samples, and l is the number of output layers; based on teacher signal T_dNetwork output signal T_oThe hidden layer to output layer weight vectors and the output component of the hidden layer,

output layer error signal:

^o＝(T_d-T_o)T_o(1-T_o)(2)

neuron error signal:

${\mathrm{\δ}}_j^y=\left({\mathrm{\δ}}^0{\mathrm{\ω}}_j\right)y_j(1-y_j)---\left(3\right)$

in the formula, ω_jAs weights from hidden layer to output layerVector, y_jIs the output of each layer.

The LM training method is adopted for network training, and the calculation formulas of the weights from the input layer to the hidden layer and the weights from the hidden layer to the output layer are

$v_{ij}=v_{ij}+{\mathrm{\η\δ}}_j^y{x}_i---\left(4\right)$

ω_j＝ω_j+Δω(5)

In the formula v_ijFor the weight vector from the input layer to the hidden layer, η is the learning rate, Δ ω is the weight adjustment vector, Δ w is calculated as:

Δw＝-(J_n ^TJ_n+η_nI)^-1J_n ^Tr_n(6)

wherein,is a Hessian matrix used to approximate the objective function, I is an identity matrix η_nParameters greater than 0 in LM training method for speeding up training of network when η_nWhen approaching zero, LM algorithm approaches Gauss-Newton method, with η_nIncreasing, the LM algorithm approximates the steepest descent method.

2.2 Performance analysis

Based on the test sample set, performing network training by adopting an LM training method to obtain four models, wherein fig. 2a is a cucumber prediction model only established for a seedling stage, fig. 2b is a model only established for a flowering growth stage, fig. 2c is a full-course model, and fig. 2d is a model added into the growth stage as a one-dimensional factor. The comparison and analysis of the training results show that, by 57 steps in fig. 2a, the network reaches the expected error level, no oscillation and local flat region occurs in the training process, the error function is 0.0000658, by 38 steps in fig. 2b, the network reaches the expected error level, no oscillation and local flat region occurs in the training process, the error function is 0.0000993, by 1000 steps, the network does not reach the expected error level, the error function is 0.00030153, by 13 steps in fig. 2d, the network reaches the expected error level, no oscillation and local flat region occurs in the training process, and the error function is 0.000028408.

Based on the results, the model established by adding the growth period as a one-dimensional factor has obvious effect, can provide theoretical basis and technical support for regulation and control of the light environment, and simplifies the operation of the light environment equipment.

3 analysis of model validation results

The test sample set obtained by adopting the multi-factor nested test is 1650 two groups, the samples are divided into a training set and a test set, wherein 660 groups are used for establishing the model, the rest 165 groups are used for forming the test set and account for about 20 percent of the total samples, the model verification is carried out by adopting an abnormal checking method, and the correlation analysis of the measured photosynthetic rate value and the predicted value is obtained as shown in the figure. As can be seen from fig. 3, the determination coefficient of correlation analysis between the model measured value and the predicted value based on the LM training method in fig. 3a is 0.987, the slope of the straight line is 1.031, and the intercept is 0.343, the determination coefficient of correlation analysis between the model measured value and the predicted value based on the LM training method in fig. 3b is 0.9922, the slope of the straight line is 1.0211, and the intercept is 1.4331, the determination coefficient of correlation analysis between the model measured value and the predicted value based on the LM training method in fig. 3c is 0.8796, the slope of the straight line is 0.9424, and the intercept is 0.04474, the determination coefficient of correlation analysis between the model measured value and the predicted value based on the LM training method in fig. 3d is 0.9897, the slope of the straight line is 0.9982, and the intercept is 0.002729. The linearity of the model established in the growing period is considered to be obviously higher, and the fitting degree is better.

The error analysis of the test result shows that the maximum relative error between the measured value and the simulated value of the whole-growth-period photosynthetic rate prediction model established in the growing period is less than +/-6.559%, which indicates that the model established in the text can carry out the photosynthetic rate model prediction in the whole growing period and has good precision.

Claims (3)

1. A cucumber whole-course photosynthetic rate prediction model based on a neural network is characterized in that the model fuses a growing period and has a model formula ofWherein the output signal T_oRepresenting the photosynthetic rate calculated by the neural network, the input signal x ═ x₁′,x₂′,......,x₆′)^T；x₁′、x₂′、x₃′、x₄′、x₅′、x₆Respectively growth period, temperature, CO₂Concentration, light intensity, relative humidity and chlorophyll content; m is 8, n is 6; v. of_ijIs the weight, w, of the input layer to the hidden layer_jIs the hidden layer to output layer weight vector,the net input quantity of the input layer adjusted to the hidden layer by the weight value is shown;the input quantities of the output layers are indicated.

2. The method for establishing the cucumber whole-course photosynthetic rate prediction model based on the neural network as claimed in claim 1, characterized by comprising the following steps:

step 1, acquiring experimental data, wherein the process is as follows:

adopting a nutrition pot for seedling culture, selecting cucumber seedlings with uniform growth vigor, stem cross diameter of 0.6-0.8 cm and plant height within 10cm for test when the cucumber seedlings grow into two leaves and one heart, selecting 150 robust cucumber seedlings as test samples, and selecting 150 plants with flowering nodes 50 cm away from a tap as the test samples in the flowering fruiting period when the cucumbers are in the flowering fruiting period;

measuring net photosynthetic rate, and setting 5 temperature gradients of 16, 20, 24, 28 and 32 ℃ by using a temperature control module in the process; by using CO₂The injection module sets the volume ratio of carbon dioxide to be 300, 600, 900, 1200 and 1500 mu L/L for 5 gradients; obtaining 0, 20, 50, 100, 200, 300, 500, 700, 1000, 1200, 1500 [ mu ] mol/(m) by using the LED light source module²S) 11 photon flux density gradients, a total of 275 trials were performed in a nested fashion, each trial was repeated on 3 plants selected at random, and the chamber relative humidity and chlorophyll content of the leaves tested were recorded in the trials to yield a plot of chlorophyll content, temperature, CO₂1650 group test with input of concentration, illumination intensity and relative humidity and output of net photosynthetic rateData, namely 825 groups at seedling stage and 825 groups at flowering and fruiting stage;

step 2, establishing a model

Step 2.1 training method

The input signal is x ═ x₁′,x₂′,......,x₆′)^T；x₁′、x₂′、x₃′、x₄′、x₅′、x₆Respectively growth period, temperature, CO₂Concentration, illumination intensity, relative humidity and chlorophyll content, and output signal T_oRepresenting the photosynthetic rate calculated by the neural network, and the corresponding actually measured photosynthetic rate being a teacher signal T_dEstablishing a whole-course cucumber seedling photosynthetic rate model with a growth period as a one-dimensional factor by a BP gradient training method;

step 2.2 training procedure

Randomly distributing an initial value V of a weight vector from an input layer to a hidden layer and an initial value W of the weight vector from the hidden layer to an output layer; operating BP neural network program, inputting training set sampleAnd according to $T_o=f\[\underset{j=0}{\overset{m}{\Σ}}{w}_{j}f\left(\underset{i=0}{\overset{n}{\Σ}}{v}_{ij}{x}_i\right)\]$ Output of a computing network T_o；

Based on teacher signal T_dAnd a network output signal T_oTotal error of systemIn the formula,in order to train the true value of the sample,the output value of the training sample network is obtained, P is the number of training samples, and l is the number of output layers;

based on teacher signal T_dNetwork output signal T_oWeight vector from hidden layer to output layer, output component of hidden layer, error signal of output layer^o＝(T_d-T_o)T_o(1-T_o) Neuron error signalIn the formula, ω_jAs weight vector from hidden layer to output layer, y_jIs the output of each layer;

adopting LM training method to train network, inputting weight vector from layer to hidden layerWeight ω from hidden layer to output layer_jWhere η is the learning rate, Δ ω is the weight adjustment vector, and Δ w is- (J)_n ^TJ_n+η_nI)^-1J_n ^Tr_n，Is a Hessian matrix used to approximate the objective function, I is an identity matrix, η_nThe parameter greater than 0 in LM training method is used to speed up the training of network, when η_nWhen approaching zero, LM algorithm approaches Gauss-Newton method, with η_nThe LM algorithm is similar to the steepest descent method when the LM is continuously increased;

step 2.3 model building

When E is_RESWhen the error value is smaller than the set error value or the learning frequency reaches the set step number, the training is stopped to obtain the final prediction model.

3. The method for establishing the cucumber global photosynthetic rate prediction model based on the neural network as claimed in claim 2, wherein in the step 2.2,

for training set sampleCarrying out normalization processing in an interval of 0.2-0.9, setting the number of hidden nodes of the neural network to be 10, and randomly distributing an initial value V of a weight vector from an input layer to a hidden layer and an initial value W of the weight vector from the hidden layer to an output layer; then operating BP neural network program and inputting training set sampleAnd according toComputing the output y of hidden nodes_j(ii) a According toCalculating the output of the output layer; by passingAnd judging whether the neural network achieves the training precision, if not, reselecting the sample to start training, otherwise, stopping training, and finishing the model establishment.