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

The omnidistance photosynthetic rate forecast model of cucumber based on neural network and method for building up

Technical field

The invention belongs to reading intelligent agriculture technical field, particularly the omnidistance photosynthetic rate forecast model of a kind of cucumber based on neural network and method for building up.

Background technology

Cucumber is one of vegetables of China's cultivation, and it is inseparable that quality and yield and its of cucumber carry out photosynthetic ability.Photosynthetic rate and chlorophyll content, temperature, CO ₂multiple factors such as concentration, intensity of illumination, relative humidity have remarkable relation.Wherein, chloroplast is that green plants carries out photosynthetic basal cell device, and chlorophyll is the basic composition material of chloroplast, most important in photosynthesis of plant, its content is the important indicator of photosynthesis of plant ability, nutrition condition and growth situation, activity, the stomatal conductance of Rubisco activating enzymes in object are made in temperature impact, CO ₂concentration directly affects the accumulation that crop carries out dark reaction speed and dry, and intensity of illumination is photosynthetic direct driving force and motive force, and relative humidity affects leaf stomatal conductance etc., and between each factor, existence influences each other.Therefore, consider multiple Effects of Factors, set up the omnidistance photosynthetic rate forecast model of multiple-factor coupling, to optimization cucumber luminous environment, there is vital role.

In recent years, numerous scholar has carried out correlative study setting up in photosynthetic rate model, and above-mentioned research all considers the association between the varying environment factor, but there is the deficiencies such as degree of fitting is lower, fitting formula is complicated, error is larger.And neural network has the advantage such as Nonlinear Mapping and adaptive learning ability, suitable matching and prediction complicated nonlinear system model, the photosynthetic rate modeling therefore based on neural network becomes study hotspot.Occur based on Hopfield network photosynthetic rate model, greenhouse tomato leaf stomatal conductance model based on BP neural network, tomato single leaf net photosynthesis in florescence rate prediction model based on WSN in the recent period, above-mentioned research from different perspectives by Application of Neural Network in photosynthetic rate modeling, but all do not consider the impact of different growing stages on crop, not yet set up omnidistance cucumber photosynthetic rate forecast model, and it is comparatively slow to there is training process, the deficiency that training error difference is larger.

Summary of the invention

In order to overcome the shortcoming of above-mentioned prior art, the object of the present invention is to provide the omnidistance photosynthetic rate forecast model of a kind of cucumber based on neural network and method for building up, design multiple-factor Nested simulation experiment, BP neural net model establishing is adopted by after data normalization process, under the prerequisite considering original environment and physiological parameter, innovatively using cucumber photosynthetic rate forecast model that growth phase information is distinguished as one dimension enter factor, omnidistance cucumber photosynthetic rate forecast model is set up, for basis is set up in the luminous environment regulation and control of industrialized agriculture by contrast verification.

To achieve these goals, the technical solution used in the present invention is:

Cucumber based on a neural network omnidistance photosynthetic rate forecast model, this Model Fusion growth period, model formation is wherein output signal T _orepresent the photosynthetic rate that neural computing obtains, input signal x'=(x ₁', x ₂' ..., x ₆') ^t; x ₁', x ₂', x ₃', x ₄', x ₅', x ₆' be respectively growth period, temperature, CO ₂concentration, intensity of illumination, relative humidity and chlorophyll content; M=8, n=6; v _ijthe weights of input layer to hidden layer, w _jbe hidden layer to output layer weight vector, what represent is that input layer is through the clean input quantity of weighed value adjusting to hidden layer; what represent is the input quantity of output layer.

The method for building up of the omnidistance photosynthetic rate forecast model of the described cucumber based on neural network, comprises the steps:

Step 1, obtain experimental data, process is as follows:

Adopt feeding block seedlings raising, treat that cucumber seedling grows up to two leaves wholeheartedly, selection growing way is even, the cucumber seedling of stem transverse diameter between 0.6 ~ 0.8cm, within plant height 10cm is tested, choose healthy and strong cucumber seedling 150 strain as test sample, treat that cucumber is in the phase of yielding positive results, choose the test sample of plant 150 strain as the phase of yielding positive results of about 50 centimetres, Flower node distance tap;

Measure Net Photosynthetic Rate, in process, utilize temperature control module to set 16,20,24,28,32 DEG C of totally 5 thermogrades; Utilize CO ₂injection module setting carbon dioxide volume ratio is 300,600,900,1200,1500 μ L/L totally 5 gradients; LED light source module is utilized to obtain 0,20,50,100,200,300,500,700,1000,1200,1500 μm of ol/ (m ²s) totally 11 photon flux density gradients, carry out 275 groups of tests altogether in a nesting relation, often group test does repeated test on 3 strain plant of random selecting, records leaf room relative humidity in test, and record tested chlorophyll content in leaf blades, thus formed with chlorophyll content, temperature, CO ₂concentration, intensity of illumination, relative humidity are input, and Net Photosynthetic Rate is the 1650 groups of test figures exported, i.e. Seedling Stage 825 groups, phase 825 groups of yielding positive results;

Step 2, Modling model

Step 2.1 training method

Input signal is x'=(x ₁', x ₂' ..., x ₆') ^t; x ₁', x ₂', x ₃', x ₄', x ₅', x ₆' be respectively growth period, temperature, CO ₂concentration, intensity of illumination, relative humidity and chlorophyll content, output signal T _orepresent the photosynthetic rate obtained by neural computing, corresponding actual measurement photosynthetic rate is teacher signal T _d, set up by BP gradient coaching method and add the omnidistance cucumber seedling photosynthetic rate model of growth period as the one dimension factor;

Step 2.2 training process

Random assignment input layer is to hidden layer weight vector initial value V and hidden layer to output layer weight vector initial value W; Run BP neural network procedure, input training set sample and according to $T_o=f\[\underset{j=0}{\overset{m}{\Σ}}{w}_{j}f\left(\underset{i=0}{\overset{n}{\Σ}}{v}_{ij}{x}_i\right)\]$ The output T of computational grid _o;

Based on teacher signal T _dwith network output signal T _o, system total error in formula, for training sample actual value, for training sample network output valve, P is training sample number, and l is output layer number;

Based on teacher signal T _d, network output signal T _o, hidden layer to the output component of output layer weight vector and hidden layer, output layer error signal δ ^o=(T _d-T _o) T _o(1-T _o), neuron error signal in formula, ω _jfor hidden layer is to the weight vector of output layer, y _jfor the output of each layer;

Adopt LM coaching method to carry out network training, input layer is to the weight vector of hidden layer hidden layer is to output layer weights ω _j+=Δ ω, in formula, η is learning rate, and Δ ω is weighed value adjusting vector, Δ w=-(J _n ^tj _n+ η _ni) ^-1j _n ^tr _n, be used to the Hessian matrix of approximate objective function, I is unit matrix, η _nfor the parameter that LM coaching method inside is greater than 0, for accelerating the training speed of network, work as η _nclose to zero time, LM algorithm is close to Gauss-Newton method; Along with η _ncontinuous increase, LM algorithm is similar to method of steepest descent;

Step 2.3 model is set up

Work as E _rESbe less than setting error amount or study number of times reach setting step number time, training stopping obtain final forecast model.

In described step 2.2,

To training set sample carry out the normalized in 0.2 ~ 0.9 interval, arranging neural network the number of hidden nodes is 10, and Random assignment input layer is to hidden layer weight vector initial value V and hidden layer to output layer weight vector initial value W; Then BP neural network procedure is run, input training set sample and according to calculate the output y of hidden node _j; According to calculate the output of output layer; Pass through judge whether neural network reaches training precision, if do not reached, reselect sample and start training, otherwise training stops, model has been set up.

The present invention sets up the mapping of envirment factor and physiological factor and photosynthetic rate, thus effectively can carry out the regulation and control of luminous environment, significant to the volume increase of crop.

Compared with prior art, the invention has the beneficial effects as follows:

1) the omnidistance photosynthetic rate forecast model of cucumber based on neural network is proposed, by adding one-dimensional growth phase input quantity, effectively distinguish cucumber and yielded positive results the photosynthetic rate value difference at different conditions of phase, local flat district can be effectively crossed in training process, do not occur repeatedly shaking, convergence rapidly, precision is higher than the forecast model of mixed growth phase.

2) with the neural network prediction model that LM coaching method is set up, its coefficient of determination is 0.9897, has good fit effect, can realize the prediction for plant different growing stages photosynthetic rate value.The omnidistance cucumber seedling photosynthetic rate forecast model built can be the regulation and control of gold-tinted Seedling Stage luminous environment and provides theoretical.

The omnidistance photosynthetic rate forecast model that the present invention proposes can be the regulation and control of cucumber luminous environment and provides theoretical foundation, and the photosynthetic Optimum Regulation model that easily extensible is applied to Different Crop is set up, to improve the photosynthetic capacity of chamber crop.

Accompanying drawing explanation

Fig. 1 the present invention is based on neural network algorithm process flow diagram.

Fig. 2 is the error change curve of different growing stages of the present invention.

Fig. 3 is the dependency diagram in modelling verification of the present invention between photosynthetic rate measured value and the analogue value.

Embodiment

Embodiments of the present invention are described in detail below in conjunction with drawings and Examples.

The process of establishing of the omnidistance photosynthetic rate forecast model of a kind of cucumber based on neural network of the present invention is as follows:

1, materials and methods

This is tested and carries out in Xibei Univ. of Agricultural & Forest Science & Technology's scientific research greenhouse in April, 2014 to July.Supply examination cucumber variety to be " Chang Chun Mi Ci ", in double dish, carrying out vernalization by soaking swollen seed, in time will sprouting, carrying out sub zero treatment, in the dish of 50 holes (540mm280mm50mm) cave, adopt feeding block seedlings raising.Seedling medium is agricultural cultivation dedicated substrate.During seedling culture, maintenance liquid manure is sufficient, treats that cucumber seedling grows up to two leaves wholeheartedly, and selection growing way is even, the cucumber seedling of stem transverse diameter between 0.6 ~ 0.8cm, within plant height 10cm is tested.Choose healthy and strong cucumber seedling 150 strain as test sample.Duration of test, carries out normal field planting management, does not spray any agricultural chemicals and hormone, treats that cucumber is in the phase of yielding positive results, and chooses the test sample of plant 150 strain as the phase of yielding positive results of about 50 centimetres, Flower node distance tap.

The portable photosynthetic instrument of the Li-6400XT type adopting LI-COR company of the U.S. to produce measures Net Photosynthetic Rate, the parameter such as temperature, CO2 concentration, intensity of illumination around the multiple submodule control on demand blades adopting photosynthetic instrument to match in process of the test.Wherein, temperature control module is utilized to set 16,20,24,28,32 DEG C of totally 5 thermogrades; Utilize CO ₂injection module setting carbon dioxide volume ratio is 300,600,900,1200,1500 μ L/L totally 5 gradients; LED light source module is utilized to obtain 0,20,50,100,200,300,500,700,1000,1200,1500 μm of ol/ (m ²s) totally 11 photon flux density (Photofluxdensity, PFD) gradient, carry out 275 groups of tests altogether in a nesting relation, often group test does repeated test on 3 strain plant of random selecting, leaf room relative humidity is recorded in test, and adopt the tested chlorophyll content in leaf blades of SPAD-502Plus type chlorophyll meter record of Japanese Konica company, thus formed with chlorophyll content, temperature, CO2 concentration, intensity of illumination, relative humidity as input, Net Photosynthetic Rate is the 1650 groups of test figures exported, i.e. Seedling Stage 825 groups, yields positive results the phase 825 groups.

2, model is set up

2.1 training method

In order to set up optimum photosynthetic rate forecast model, difference in growth period for cucumber adopts same modeling method to build together vertical four kinds of models, be respectively only for cucumber forecast model, only to yield positive results the forecast model of phase, the photosynthetic rate forecast model of cucumber whole process and the difference in growth period is set up the forecast model of cucumber whole process as one dimension input for cucumber.Input signal is x'=(x ₁' x ₂' ... x ₅') ^t; x ₁', x ₂', x ₃', x ₄', x ₅' be respectively temperature, CO ₂concentration, intensity of illumination, relative humidity and chlorophyll content, the 4th kind of model adds and inputs as one dimension growth period, and output signal all uses T _orepresent the photosynthetic rate that network calculations obtains, often group corresponding actual measurement photosynthetic rate is teacher signal T _d.Omnidistance cucumber seedling photosynthetic rate model T is set up by BP gradient coaching method _d' (X').

As shown in Figure 1, when BP neural network procedure is run, according to network connection value and threshold value, Random assignment input layer is to hidden layer weight vector initial value V and hidden layer to output layer weight vector initial value W; Run BP neural network procedure, input training set sample and according to the output T of computational grid _o, and trigger following process:

Based on teacher signal and network output signal, system total error computing formula is

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

In formula, for training sample actual value, for training sample network output valve, P is training sample number, and l is output layer number; Based on teacher signal T _d, network output signal T _o, hidden layer to the output component of output layer weight vector and hidden layer,

Output layer error signal:

δ ^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 formula, ω _jfor hidden layer is to the weight vector of output layer, y _jfor the output of each layer.

Adopt LM coaching method to carry out network training, input layer to the weights of hidden layer and hidden layer to output layer weight computing formula is

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

ω _j＝ω _j+Δω(5)

V in formula _ijfor input layer is to the weight vector of hidden layer, η is learning rate, and Δ ω is weighed value adjusting vector, and Δ w computing formula is:

Δw＝-(J _n ^TJ _n+η _nI) ^-1J _n ^Tr _n(6)

Wherein, be used to the Hessian matrix of approximate objective function, I is unit matrix.η _nfor the parameter that LM coaching method inside is greater than 0, for accelerating the training speed of network.Work as η _nclose to zero time, LM algorithm is close to Gauss-Newton method; Along with η _ncontinuous increase, LM algorithm is similar to method of steepest descent.

2.2 performance evaluation

Based on above-mentioned test sample collection, adopt LM coaching method to carry out network training, obtain four kinds of models, Fig. 2 a is namely only for the cucumber forecast model that Seedling Stage is set up, Fig. 2 b is namely only for the model that growth period of blooming sets up, and Fig. 2 c is omnidistance model, and Fig. 2 d adds the model of growth period as the one dimension factor.Comparative analysis training result can find, by the end of 57 steps in Fig. 2 a, network reaches the error level of expectation, there is not concussion and flat region, local in training process, error function is 0.0000658, by the end of 38 steps in Fig. 2 b, network reaches the error level of expectation, there is not concussion and flat region, local in training process, error function is 0.0000993, in Fig. 2 c there is local flat district in training process, by the end of 1000 steps, network does not reach the error level of expectation, error function is 0.00030153, by the end of 13 steps in Fig. 2 d, network reaches the error level of expectation, there is not concussion and flat region, local in training process, error function is 0.000028408.

Based on the above results, it is remarkable to add the modelling effect set up as the one dimension factor growth period, can regulate and control provide fundamental basis and technical support for luminous environment, simplify the operation of luminous environment equipment.

3 modelling verification interpretations of result

Adopt the test sample collection totally 1650 Ge Liang groups that multiple-factor Nested simulation experiment obtains, sample is divided into training set and test set, wherein 660 groups are used for the foundation of model, remain 165 groups for forming test set, account for 20% of total sample, adopt different method of calibration to carry out modelling verification, obtain photosynthetic rate measured value and predicted value correlation analysis as shown in the figure.Can find from Fig. 3, 0.987 based on the model actual measurement value of LM coaching method and the coefficient of determination of predicted value correlation analysis in Fig. 3 a, straight slope is 1.031, intercept is 0.343, 0.9922 based on the model actual measurement value of LM coaching method and the coefficient of determination of predicted value correlation analysis in Fig. 3 b, straight slope is 1.0211, intercept is 1.4331, 0.8796 based on the model actual measurement value of LM coaching method and the coefficient of determination of predicted value correlation analysis in Fig. 3 c, straight slope is 0.9424, intercept is 0.04474, 0.9897 based on the model actual measurement value of LM coaching method and the coefficient of determination of predicted value correlation analysis in Fig. 3 d, straight slope is 0.9982, intercept is 0.002729.Consider that the linearity of Modling model in growth period is obviously higher, fitting degree is better.

Error analysis is carried out to test findings known, consider the measured value of omnidistance photosynthetic rate forecast model that growth period sets up and analogue value maximum relative error be less than ± 6.559%, show that the model set up can carry out the photosynthetic rate model prediction in full growth period herein, have good precision.

Claims (3)

1. based on the omnidistance photosynthetic rate forecast model of cucumber of neural network, to it is characterized in that, this Model Fusion growth period, model formation is wherein output signal T _orepresent the photosynthetic rate that neural computing obtains, input signal x '=(x ₁', x ₂' ..., x ₆') ^t; x ₁', x ₂', x ₃', x ₄', x ₅', x ₆' be respectively growth period, temperature, CO ₂concentration, intensity of illumination, relative humidity and chlorophyll content; M=8, n=6; v _ijthe weights of input layer to hidden layer, w _jbe hidden layer to output layer weight vector, what represent is that input layer is through the clean input quantity of weighed value adjusting to hidden layer; what represent is the input quantity of output layer.

2., according to claim 1 based on the method for building up of the omnidistance photosynthetic rate forecast model of cucumber of neural network, it is characterized in that, comprise the steps:

Step 1, obtain experimental data, process is as follows:

Adopt feeding block seedlings raising, treat that cucumber seedling grows up to two leaves wholeheartedly, selection growing way is even, the cucumber seedling of stem transverse diameter between 0.6 ~ 0.8cm, within plant height 10cm is tested, choose healthy and strong cucumber seedling 150 strain as test sample, treat that cucumber is in the phase of yielding positive results, choose the test sample of plant 150 strain as the phase of yielding positive results of about 50 centimetres, Flower node distance tap;

Measure Net Photosynthetic Rate, in process, utilize temperature control module to set 16,20,24,28,32 DEG C of totally 5 thermogrades; Utilize CO ₂injection module setting carbon dioxide volume ratio is 300,600,900,1200,1500 μ L/L totally 5 gradients; LED light source module is utilized to obtain 0,20,50,100,200,300,500,700,1000,1200,1500 μm of ol/ (m ²s) totally 11 photon flux density gradients, carry out 275 groups of tests altogether in a nesting relation, often group test does repeated test on 3 strain plant of random selecting, records leaf room relative humidity in test, and record tested chlorophyll content in leaf blades, thus formed with chlorophyll content, temperature, CO ₂concentration, intensity of illumination, relative humidity are input, and Net Photosynthetic Rate is the 1650 groups of test figures exported, i.e. Seedling Stage 825 groups, phase 825 groups of yielding positive results;

Step 2, Modling model

Step 2.1 training method

Input signal is x '=(x ₁', x ₂' ..., x ₆') ^t; x ₁', x ₂', x ₃', x ₄', x ₅', x ₆' be respectively growth period, temperature, CO ₂concentration, intensity of illumination, relative humidity and chlorophyll content, output signal T _orepresent the photosynthetic rate obtained by neural computing, corresponding actual measurement photosynthetic rate is teacher signal T _d, set up by BP gradient coaching method and add the omnidistance cucumber seedling photosynthetic rate model of growth period as the one dimension factor;

Step 2.2 training process

Random assignment input layer is to hidden layer weight vector initial value V and hidden layer to output layer weight vector initial value W; Run BP neural network procedure, input training set sample and according to $T_o=f\[\underset{j=0}{\overset{m}{\Σ}}{w}_{j}f\left(\underset{i=0}{\overset{n}{\Σ}}{v}_{ij}{x}_i\right)\]$ The output T of computational grid _o;

Based on teacher signal T _dwith network output signal T _o, system total error in formula, for training sample actual value, for training sample network output valve, P is training sample number, and l is output layer number;

Based on teacher signal T _d, network output signal T _o, hidden layer to the output component of output layer weight vector and hidden layer, output layer error signal δ ^o=(T _d-T _o) T _o(1-T _o), neuron error signal in formula, ω _jfor hidden layer is to the weight vector of output layer, y _jfor the output of each layer;

Adopt LM coaching method to carry out network training, input layer is to the weight vector of hidden layer hidden layer is to output layer weights ω _j+=Δ ω, in formula, η is learning rate, and Δ ω is weighed value adjusting vector, Δ w=-(J _n ^tj _n+ η _ni) ^-1j _n ^tr _n, be used to the Hessian matrix of approximate objective function, I is unit matrix, η _nfor the parameter that LM coaching method inside is greater than 0, for accelerating the training speed of network, work as η _nclose to zero time, LM algorithm is close to Gauss-Newton method; Along with η _ncontinuous increase, LM algorithm is similar to method of steepest descent;

Step 2.3 model is set up

Work as E _rESbe less than setting error amount or study number of times reach setting step number time, training stopping obtain final forecast model.

3., according to claim 2 based on the method for building up of the omnidistance photosynthetic rate forecast model of cucumber of neural network, it is characterized in that, in described step 2.2,

To training set sample carry out the normalized in 0.2 ~ 0.9 interval, arranging neural network the number of hidden nodes is 10, and Random assignment input layer is to hidden layer weight vector initial value V and hidden layer to output layer weight vector initial value W; Then BP neural network procedure is run, input training set sample and according to calculate the output y of hidden node _j; According to calculate the output of output layer; Pass through judge whether neural network reaches training precision, if do not reached, reselect sample and start training, otherwise training stops, model has been set up.