CN103823371B - Agriculture Tree Precise Fertilization system and fertilizing method based on neural network model - Google Patents

Abstract

The present invention relates to a kind of agriculture Tree Precise Fertilization system based on neural network model, including the sensor data acquisition module for gathering agricultural land soil parameter, its outfan pass sequentially through gateway, data transmission module with and the data base that is connected of central processing unit be connected, the input/output terminal of central processing unit respectively with RS remote sensing module, GPS locating module, GIS geography information module, DSS plant growth decision support module input/output terminal be connected.The invention also discloses the fertilizing method of a kind of agriculture Tree Precise Fertilization system based on neural network model.The present invention divides by functional module, System information, automatization level degree are high, merge the achievement of field of artificial intelligence neutral net research, utilize neutral net, can fully approach arbitrarily complicated non-linear relation, system uncertain with self adaptation can be learnt, set up neural network model, make it carry out Fertilization Decision according to ecological benefits, maximization of economic benefit, put into practice offer method foundation for variable fertilization.

Description

Agriculture Tree Precise Fertilization system and fertilizing method based on neural network model

Technical field

The present invention relates to precision agriculture system farmland fertilization technical field, the fertilizing method of especially a kind of agriculture Tree Precise Fertilization system based on neural network model.

Background technology

Traditional agriculture production technology carries out using big plot as administrative unit, plot environment that thousands of mus are big, edphic factor etc. are widely different, such production technology can cause grants the unbalanced of need, it is impossible to gives full play to plant growth potentiality, affects economic benefit and ecological benefits.Precision agriculture is based on global satellite positioning, present information management technique, crop aid decision supports the integrated Crop management technology assembled such as technology and agriculture project equipment technology, divide farmland into small cell, with the spatio-temporal difference of crop yield and growing environment condition for foundation, crop is carried out agricultural production cultivation management, it is precisely in that, utilize GPS, GIS, RS, DSS system detects crop immediately, farmland, soil parameters, according to administrative unit soil characteristic and crop growth needs, formulate and produce prescription map, sow with this, fertilising, spray medicine, management is thrown in irrigations etc..

Tree Precise Fertilization is the core content in precision agriculture technology; the enforcement of Tree Precise Fertilization can save fertilizer; increase grain yield; balanced soil nutrient, technology is based on soil nutrient status, crop regulation of fertilizer requirement and target output, regulated fertilization amount, nitrogen phosphoris and potassium fertilizer ratio and fertilizing time; to improve chemical fertilizer utilization ratio; maximally utilise land resource, obtain maximum output and maximum economic benefit, protecting agriculture ecological environment and natural resources with rational fertilizers input amount.The heavy difficult point of Tree Precise Fertilization is in that to formulate decision model, owing to target output is difficult to estimate accurately, in highly non-linear between yield and soil nutrient, dose and other influence factor, existing traditional Fertilization Model is many with current production practices incompatibility, and the mass data of accumulation is difficult to Instructing manufacture practice.

Summary of the invention

It is an object of the invention to provide the fertilizing method that a kind of intelligence degree is strong, have the agriculture Tree Precise Fertilization system based on neural network model of higher precision and robustness.

For achieving the above object, present invention employs techniques below scheme: the fertilizing method of a kind of agriculture Tree Precise Fertilization system based on neural network model, the method includes the step of following order:

(1) gathered the soil parameters of agricultural land soil to be detected by sensor data acquisition module, and sent to data base by gateway, data transmission module；

(2) central processing unit calls in data base the soil parameters of storage and carries out computing, first builds neuronal structure model, reselection activation primitive, finally carry out back-propagating study, obtain training objective, the sample data that training obtains, draw the neural network model based on back-propagating study；

(3) on the basis of neural network model, use GIS geography information module, and in conjunction with the Tree Precise Fertilization decision-making that DSS plant growth decision support module provides, provide precision agriculture fertilizing method；

The system implementing described fertilizing method includes the sensor data acquisition module for gathering agricultural land soil parameter, its outfan pass sequentially through gateway, data transmission module with and the data base that is connected of central processing unit be connected, the input/output terminal of central processing unit respectively with RS remote sensing module, GPS locating module, GIS geography information module, DSS plant growth decision support module input/output terminal be connected；Described sensor data acquisition module by soil temperature sensor, soil moisture measurement sensor and for measure Nitrogen In Soils, phosphorus, potassium content soil nutrient elements determination sensor form；Described sensor data acquisition module, RS remote sensing module, GPS locating module are bundled on mobile farm working machinery jointly；Described data transmission module adopts GPRS wireless network module；Described central processing unit is background computer.

In step (2), described structure neuronal structure model refers to, the output y of neuron i_iFor:

$y_i=F\left(x\right)\·(\underset{j=i}{\overset{n}{\Σ}}{\mathrm{\ω}}_j{x}_j-{\mathrm{\θ}}_i),j\≠i$

Wherein, y_iBeing the output of neuron i, it is coupled together by power with other multiple neurons；X_jIt is network input, j=1,2 ..., n；W_jFor inputting weight, j=1,2 ... n；θ_iBeing the threshold values of neuron i, F is activation primitive.

In step (2), described selection activation primitive refers to, selects activation primitive Sigmoid function:

$F\left(x\right)=\frac{1}{1+e^{-ax}}$

α is function slope, and e is Euler's constant, and x is independent variable.

In step (2), described back-propagating study refers to, if Artificial Neural Network Structures input layer quantity is N, P, K, Yield tetra-, the i.e. nitrogen of soil nutrient, phosphorus, potassium content and target output, model output layer neuronal quantity is three, it is the nitrogen of fertilising, phosphorus, potassium application rate respectively, hidden layer neuron quantity t rule of thumb formula:

M is output neuron number, and n is input block number, and a selects 1 between 1 to 10, carries out following neutral net back-propagating study:

(1), data normalization processes: first the soil nutrient content gathered by sensor data acquisition module and target output data being normalized, the data collected be distributed between [-1,1], normalization formula is:

$P_n=2\×\frac{P-\min P}{\max P-\min P}-1$

Wherein, P is original input data, maximum in maxP and minP respectively P and minima, P_nFor the input data after normalization；

(2), random initializtion network weight and threshold value, set end condition as the output valve of neutral net and the mean square error of real output value less than a certain threshold value, when end condition is unsatisfactory for, be repeated below step:

Forward calculates input and the output of hidden layer or each unit j of output layer in each training sample, for

Jth unit inputsOutput

Wherein, e is natural logrithm, θ_jIt is each neuronic threshold value for changing neuronic activity, w_ijIndicate that the weights between preceding layer neuron and later layer neuron, be namely by the power of the unit i of last layer to the connection of unit j, O_iIt is the output of the unit i of last layer；

Then, the error of each unit j of output layer is calculated:

Err_j=O_j(1-O_j)(T_j-O_j)

Wherein, T_jIt is the j real output based on the known class label of given training sample；

By last to first hidden layer, for each unit j of hidden layer,

Err_j=O_j(1-O_j)∑_kErr_kw_kj

Wherein, w_kjIt is to the connection weight of unit j by unit k in next higher level, and Err_kIt it is the error of unit k；

Afterwards, w in network is updated_ijWeights, method is:

Δw_ij=lErr_jO_i

Wherein, l is learning rate, and after renewal, weights are

w_ij=w_ij+Δw_ij

For deviation θ each in network_j, rising in value is

Δθ_j=lErr_j

After renewal, deviation is θ_j=θ_j+Δθ_j；

Random initializtion network weight and neuronic threshold value, propagated forward calculating hidden layer neuron in layer and the neuronic input of output layer and output, back-propagating correction weights and threshold value, until end condition meets.

As shown from the above technical solution, the present invention divides by functional module, and System information, automatization level degree are high, the link such as including data acquisition, data transmission, data process, provides educated decisions support for agricultural Tree Precise Fertilization.In addition, the present invention has merged the achievement of field of artificial intelligence neutral net research, utilize neutral net, can fully approach arbitrarily complicated non-linear relation, system uncertain with self adaptation can be learnt, set up neural network model so that it is carry out Fertilization Decision according to ecological benefits, maximization of economic benefit, put into practice offer method foundation for variable fertilization.

Accompanying drawing explanation

Fig. 1 is the system structure schematic diagram of the present invention；

Fig. 2 is neuronal structure schematic diagram in the present invention；

Fig. 3,4 be neural network model of the present invention training process schematic.

Detailed description of the invention

A kind of agriculture Tree Precise Fertilization system based on neural network model, including the sensor data acquisition module for gathering agricultural land soil parameter, its outfan pass sequentially through gateway, data transmission module with and the data base that is connected of central processing unit be connected, the input/output terminal of central processing unit respectively with RS remote sensing module, GPS locating module, GIS geography information module, DSS plant growth decision support module input/output terminal be connected.Described sensor data acquisition module by soil temperature sensor, soil moisture measurement sensor and for measure Nitrogen In Soils, phosphorus, potassium content soil nutrient elements determination sensor form.Described data transmission module adopts GPRS wireless network module, and described central processing unit is background computer, as shown in Figure 1.

As shown in Figure 1, described sensor data acquisition module, RS remote sensing module, GPS locating module is bundled on mobile farm working machinery jointly, when Work machine is when Tanaka's operation, community, sensor data acquisition module Real-time Collection each farmland moisture, the data such as fertility, it is sent in data base by data transmission module, simultaneously, the GPS of community, farmland positions information, RS remote sensing image hierarchical information is also correspondingly saved in data base, by information integration and process, unification is called through central processing unit, it is sent to GIS geography information module, via GIS geography information module analysis, process, form farmland spatial geographic information figure, yield spatial distribution map, and carry out the system decision-making in conjunction with DSS plant growth decision support module.

As it is shown in figure 1, described sensor data acquisition module is responsible for gathering in agricultural land soil fertile soil parameters such as nitrogen, phosphorus, potassium, temperature, moisture；Have the soil nutrient elements determination sensor based on the exploitation of soil liquid photoelectric colorimetry, based on the soil moisture measurement sensor of near-infrared (NIR) spectral technique and transfer impedance transformation theory, these sensors can be detected simultaneously by agricultural land soil the instant data of the soil of main nutrient elements.Described data transmission module is that the data coming from sensor data acquisition module are passed through the integrated gateway that is sent to, and will be transferred to the Internet by gateway unification by GPRS wireless network module or 3G network, to reach remotely to receive data function.Described GPS locating module is responsible for obtaining farmland latitude and longitude information data, in order to the later stage carries out agricultural land soil spatial data data management in conjunction with GIS GIS-Geographic Information System module, implements accurate positioning fertilising.Described RS remote sensing module is used for obtaining field numerical map, generates field-crop Yield distribution in space figure by Production rate.

As shown in Figure 1, agricultural land soil is being carried out intelligence sample by described GIS geography information module, after the data acquisitions such as soil nutrient, water content and Soil Nitrogen, phosphorus, potassium are got off, data transmission module sends these data to central processing unit with corresponding GPS location data information, and in conjunction with information such as spatial geographical locations figure, soil nutrient, output distribotion, fertilizer application modes, it is updated to the GIS information geography system module that can be used for native system, instructs follow-up Tree Precise Fertilization to produce.Described central processing unit is by setting up neutral net and by training the data that transmission obtains to draw effective neural network model, just Fertilization Model is formed on this neural network model basis, excavate information potential, useful in data, and generate accurate space Fertilization prescription chart in conjunction with field-crop yield and geographical location information, instruct and carry out accurate farmland operation.Described data base is responsible for storage system and is correlated with various data, positions information, agricultural land soil parameter, output distribotion information including GPS, is connected with central processing unit, it is provided that data are transmitted and Informational support.Described DSS plant growth decision support module is developed based on crop expert knowledge library, including growth course and the nutrient demand for describing crop such as crop growth model, crop alimentary knowledge model.

This method comprises the following steps: the first step, sensor data acquisition module gather the soil parameters of agricultural land soil to be detected, and sent to data base by gateway, data transmission module；Second step, central processing unit calls the soil parameters of storage in data base and carries out computing, first builds neuronal structure model, reselection activation primitive, finally carries out back-propagating study, obtains training objective, the sample data that training obtains, draws the neural network model based on back-propagating study；3rd step, on the basis of neural network model, uses GIS geography information module, and in conjunction with the Tree Precise Fertilization decision-making that DSS plant growth decision support module provides, provides precision agriculture fertilizing method.

In second step, in step (2), described structure neuronal structure model refers to, the output y of neuron i_iFor:

$y_i=F\left(x\right)\·(\underset{j=i}{\overset{n}{\Σ}}{\mathrm{\ω}}_j{x}_j-{\mathrm{\θ}}_i),j\≠i$

Wherein, y_iBeing the output of neuron i, it is coupled together by power with other multiple neurons；X_jIt is network input, j=1,2 ..., n；W_jFor inputting weight, j=1,2 ... n；θ_iBeing the threshold values of neuron i, F is activation primitive.

Neuron is the basic computational ele-ment of neutral net, is the non-linear unit of multi input, single output, and neuron has excited and suppresses two kinds of duties.One neuron can regard an information process unit as, is made up of three fundamentals: connecting line, adder, activation primitive.Neuronal structure is as shown in Figure 2.

In second step, described selection activation primitive refers to, selects activation primitive Sigmoid function:

$F\left(x\right)=\frac{1}{1+e^{-ax}}$

α is function slope, and e is Euler's constant, and x is independent variable.Activation primitive, for the result of calculation of sum unit is carried out functional operation, obtains neuron output, and advantage is as x=-∞, and F (x) extreme value is 0, and as x=∞, F (x) extreme value is 1, and can directly suppress or activate weight factor.

In second step, described back-propagating study refers to, if Artificial Neural Network Structures input layer quantity is N, P, K, Yield tetra-, the i.e. nitrogen of soil nutrient, phosphorus, potassium content and target output, model output layer neuronal quantity is three, it is the nitrogen of fertilising, phosphorus, potassium application rate respectively, hidden layer neuron quantity t rule of thumb formula:

M is output neuron number, and n is input block number, and a selects 1 between 1 to 10, carries out following neutral net back-propagating study, as shown in Figure 3,4:

(1), data normalization processes: first the soil nutrient content gathered by sensor data acquisition module and target output data being normalized, the data collected be distributed between [-1,1], normalization formula is:

$P_n=2\×\frac{P-\min P}{\max P-\min P}-1$

Wherein, P is original input data, maximum in maxP and minP respectively P and minima, P_nFor the input data after normalization；

(2), random initializtion network weight and threshold value, set end condition as the output valve of neutral net and the mean square error of real output value less than a certain threshold value, when end condition is unsatisfactory for, be repeated below step:

Forward calculates input and the output of hidden layer or each unit j of output layer in each training sample, for

Jth unit inputsOutput

Wherein, e is natural logrithm, θ_jIt is each neuronic threshold value for changing neuronic activity, w_ijIndicate that the weights between preceding layer neuron and later layer neuron, be namely by the power of the unit i of last layer to the connection of unit j, O_iIt is the output of the unit i of last layer；

Then, the error of each unit j of output layer is calculated:

Err_j=O_j(1-O_j)(T_j-O_j)

Wherein, T_jIt is the j real output based on the known class label of given training sample；

By last to first hidden layer, for each unit j of hidden layer,

Err_j=O_j(1-O_j)∑_kErr_kw_kj

Wherein, w_kjIt is to the connection weight of unit j by unit k in next higher level, and Err_kIt it is the error of unit k；

Afterwards, w in network is updated_ijWeights, method is:

Δw_ij=lErr_jO_i

Wherein, l is learning rate, and after renewal, weights are

w_ij=w_ij+Δw_ij

For deviation θ each in network_j, rising in value is

Δθ_j=lErr_j

After renewal, deviation is θ_j=θ_j+Δθ_j；

Random initializtion network weight and neuronic threshold value, propagated forward calculating hidden layer neuron in layer and the neuronic input of output layer and output, back-propagating correction weights and threshold value, until end condition meets.

In sum, the core of the present invention is in that first to be gathered the soil parameters of agricultural land soil to be detected by sensor data acquisition module, and is sent to data by gateway, data transmission module；Central processing unit calls the soil parameters of storage in data base and carries out computing, first builds neuronal structure model, reselection activation primitive, finally carry out back-propagating study, obtain training objective, the sample data that training obtains, draw the neural network model based on back-propagating study；On the basis of neural network model, use GIS geography information module, and in conjunction with the Tree Precise Fertilization decision-making that DSS plant growth decision support module provides, provide precision agriculture fertilizing method.

Claims (4)

1., based on a fertilizing method for the agriculture Tree Precise Fertilization system of neural network model, the method includes the step of following order:

(1) gathered the soil parameters of agricultural land soil to be detected by sensor data acquisition module, and sent to data base by gateway, data transmission module；

(2) central processing unit calls in data base the soil parameters of storage and carries out computing, first builds neuronal structure model, reselection activation primitive, finally carry out back-propagating study, obtain training objective, the sample data that training obtains, draw the neural network model based on back-propagating study；

(3) on the basis of neural network model, use GIS geography information module, and in conjunction with the Tree Precise Fertilization decision-making that DSS plant growth decision support module provides, provide precision agriculture fertilizing method；

The system implementing described fertilizing method includes the sensor data acquisition module for gathering agricultural land soil parameter, its outfan pass sequentially through gateway, data transmission module with and the data base that is connected of central processing unit be connected, the input/output terminal of central processing unit respectively with RS remote sensing module, GPS locating module, GIS geography information module, DSS plant growth decision support module input/output terminal be connected；Described sensor data acquisition module by soil temperature sensor, soil moisture measurement sensor and for measure Nitrogen In Soils, phosphorus, potassium content soil nutrient elements determination sensor form；Described sensor data acquisition module, RS remote sensing module, GPS locating module are bundled on mobile farm working machinery jointly；Described data transmission module adopts GPRS wireless network module；Described central processing unit is background computer.

2. fertilizing method according to claim 1, it is characterised in that: in step (2), described structure neuronal structure model refers to, the output y of neuron i_iFor:

$y_i=F\left(x\right)\·(\underset{j=i}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ω}}_j{x}_j-{\mathrm{\θ}}_i),j\≠i$

Wherein, y_iBeing the output of neuron i, it is coupled together by power with other multiple neurons；X_jIt is network input, j=1,2 ..., n；W_jFor inputting weight, j=1,2 ... n；θ_iBeing the threshold values of neuron i, F is activation primitive.

3. fertilizing method according to claim 1, it is characterised in that: in step (2), described selection activation primitive refers to, selects activation primitive Sigmoid function:

$F\left(x\right)=\frac{1}{1+e^{-ax}}$

α is function slope, and e is Euler's constant, and x is independent variable.

4. fertilizing method according to claim 1, it is characterized in that: in step (2), described back-propagating study refers to, if Artificial Neural Network Structures input layer quantity is N, P, K, Yield tetra-, the i.e. nitrogen of soil nutrient, phosphorus, potassium content and target output, model output layer neuronal quantity is three, is the nitrogen of fertilising, phosphorus, potassium application rate respectively, hidden layer neuron quantity t rule of thumb formula:M is output neuron number, and n is input block number, and a selects 1 between 1 to 10, carries out following neutral net back-propagating study:

(1), data normalization processes: first the soil nutrient content gathered by sensor data acquisition module and target output data being normalized, the data collected be distributed between [-1,1], normalization formula is:

$P_n=2\×\frac{P-\min P}{\max P-\min P}-1$

Wherein, P is original input data, maximum in maxP and minP respectively P and minima, P_nFor the input data after normalization；

(2), random initializtion network weight and threshold value, set end condition as the output valve of neutral net and the mean square error of real output value less than a certain threshold value, when end condition is unsatisfactory for, be repeated below step:

Forward calculates input and the output of hidden layer or each unit j of output layer in each training sample, inputs for jth unitOutput

Wherein, e is natural logrithm, θ_jIt is each neuronic threshold value for changing neuronic activity, w_ijIndicate that the weights between preceding layer neuron and later layer neuron, be namely by the power of the unit i of last layer to the connection of unit j, O_iIt is the output of the unit i of last layer；

Then, the error of each unit j of output layer is calculated:

Err_j=O_j(1-O_j)(T_j-O_j)

Wherein, T_jIt is the j real output based on the known class label of given training sample；

By last to first hidden layer, for each unit j of hidden layer,

Err_j=O_j(1-O_j)∑_kErr_kw_kj

Wherein, w_kjIt is to the connection weight of unit j by unit k in next higher level, and Err_kIt it is the error of unit k；

Afterwards, w in network is updated_ijWeights, method is:

Δw_ij=lErr_jO_i

Wherein, l is learning rate, and after renewal, weights are

w_ij=w_ij+Δw_ij

For deviation θ each in network_j, rising in value is

Δθ_j=lErr_j

After renewal, deviation is θ_j=θ_j+Δθ_j；

Random initializtion network weight and neuronic threshold value, propagated forward calculating hidden layer neuron in layer and the neuronic input of output layer and output, back-propagating correction weights and threshold value, until end condition meets.