CN110084145B - TensorFlow-based pest and disease damage time-frequency domain multi-scale identification system and operation method - Google Patents

Abstract

The invention discloses a pest and disease damage time-frequency domain multi-scale identification system and an operation method based on a TensorFlow, wherein the system comprises an OV7725 camera sub-node acquisition unit, a LoRa transmission sub-node transmission unit, a soil entropy sensor sub-node acquisition unit, a stm32 node processing unit and other hardware platforms based on the Internet of things technology, and the system transmits acquired images and farmland information to a monitoring center; and (3) carrying out a data statistical analysis method of the basic TensorFlow in the monitoring center, and carrying out analysis and judgment on the crop image and farmland information to obtain accurate identification and judgment of plant diseases and insect pests.

Description

A TensorFlow-based time-frequency domain multi-scale identification system and operation method for pests and diseases

technical field

The invention belongs to the field of identification of agricultural diseases and insect pests, and in particular relates to a time-frequency domain multi-scale identification system and operation method of diseases and insect pests based on TensorFlow (a symbolic mathematics system based on data flow programming).

Background technique

The monitoring of crop diseases and insect pests in farmland is an important step to realize agricultural informatization, and then realize agricultural intelligence. There are many steps and methods designed in the whole work process. In recent years, crop diseases and insect pests seem to be getting more and more serious, and various diseases continue to intersect. On the one hand, due to the indiscriminate use of pesticides in terms of types and dosages, crop pests gradually develop resistance; on the other hand, the price of agricultural products has been falling all the way, farmers are eager to increase their income, and the use of chemical fertilizers is increasing year by year, resulting in changes in soil composition. The growth of crop pests. Crop diseases and pests are becoming more and more difficult to control. It is inaccurate to simply think that it is caused by the amount of pesticides and fertilizers used. It is necessary to comprehensively analyze the leaf images of crop diseases and insect pests and various soil and air data. Among them, automatic identification of diseases and insect pests is one of the important issues. It is necessary to find a calculation method suitable for the identification of crop diseases and insect pests in the field.

The traditional methods of monitoring crop diseases and insect pests in farmland generally include field surveys by plant protection personnel, field sampling, etc., but these methods are destructive to the crops themselves, and are time-consuming and laborious. , it is difficult to meet the real-time monitoring requirements of large-scale pests and diseases, and it cannot be fully promoted and applied. In recent years, the identification technology of farmland crop diseases and insect pests has developed methods with modern technology, such as the application overview of Bayesian discriminant method, kernel K-means clustering method, and chromaticity moment method. Since various methods and technologies have their own application conditions and limitations, none of these methods can adaptively adjust the image size and features for judgment. Sometimes in practice, if only one parameter is used as the basis for interpretation, it may be difficult to obtain the expected results. Therefore, in practice, it is generally necessary to adopt a method of adaptive adjustment with different situations.

With the development of science and technology, the identification technology of crop pests and diseases in farmland has made great progress and development, and the accuracy and speed of pest identification have been continuously improved. However, various identification methods have their own application prerequisites and limitations. In practical applications, most of the current methods only idealize the environment, and there is no innovation in the method itself. Difficult to achieve better results. Therefore, in the application of farmland intelligent crop disease and pest monitoring, scientific and technical personnel should choose a reasonable data processing method according to the physical properties of the on-site medium and image characteristics.

Currently, there is no TensorFlow-based design scheme for integrated processing of image and data in field crop pest identification. In the construction of farmland monitoring projects, crop monitoring equipment based on cameras and data transmission modules is generally installed, which can analyze the collected farmland information, captured crop pictures, air temperature and humidity, crop pathogen spores, the number of pests and diseases Information shifts are monitored. However, after the existing images and data are acquired, most of them are only manually judged, which is greatly affected by subjective factors, and cannot flexibly realize the identification of diseases and insect pests in different farmlands, and is also not conducive to the popularization of agricultural intelligence.

Contents of the invention

The identification methods of crop diseases and insect pests in farmland include manual identification, MATLAB identification, machine vision identification and other transmission methods. The application scenarios and effective distances of different types of pest identification methods are often different, and the perception technology of their external performance is also different. Therefore, it is necessary to first study the image information and farmland information of crop pests and diseases. In order to more accurately and timely identify diseases and insect pests of farmland crops in complex environments, it is necessary to cooperatively perceive the surrounding environment or its own state, and perform adaptive, time-frequency domain, and multi-scale processing methods to obtain effective information from farmland soil and crop images. Information, and actively respond according to the corresponding rules, carry out preliminary treatment and identification of crop diseases and insect pests. Therefore, the research on the identification technology of crop diseases and insect pests in complex scenarios is one of the important links to ensure the intelligentization of agriculture.

In summary, the present invention combines the actual situation of agricultural pest identification, analyzes the farmland information and crop information that lead to agricultural pest identification, and proposes a multi-scale identification system and method based on TensorFlow in the time-frequency domain for agricultural pest identification. : 1. The occurrence of soil diseases and insect pests on crops will lead to obvious temperature and humidity differences between the place where the diseases and insect pests occur and the surrounding medium, so it is necessary to collect soil entropy values for comparative analysis; 2. The characteristics of leaves when crops are affected by diseases and insect pests The characteristics of the leaves are different, so it is necessary to collect images of plant leaves in the farmland through the camera for comparative analysis; 3. The captured images of crops cannot be guaranteed to be very clear and accurate, so multi-scale transformation is required to lock the occurrence of diseases and insect pests in the images 4. There will be subtle differences between crops with pests and diseases and crops without pests and diseases, which are generally difficult for human eyes to detect, so it is necessary to perform time-frequency domain analysis on crop images and farmland information. 5. Comprehensively process the processed farmland information and image information, adopt the method based on TensorFlow to automatically change the judgment threshold, adjust the parameters adaptively, and obtain accurate pest identification and judgment.

A TensorFlow-based time-frequency domain multi-scale identification system for pests and diseases, including OV7725 camera sub-node acquisition unit, LoRa transmission sub-node transmission unit, soil entropy sensor sub-node acquisition unit, stm32 node processing unit and other hardware platforms based on Internet of Things technology , and transmit the collected images and farmland information to the monitoring center; the TensorFlow-based data statistical analysis method is carried out in the monitoring center, and the crop images and farmland information are analyzed and judged to obtain accurate pest identification and judgment.

An operation method of a TensorFlow-based time-frequency domain multi-scale identification system for pests and diseases, comprising the following steps:

The first step is to obtain the gradient distribution of the temperature field and humidity field of the crop field;

The second step is multi-scale transformation of farmland crop images;

The third step is time-domain transformation of farmland crop images;

The fourth step is the frequency domain transformation of farmland crop images;

The fifth step is data fusion. The temperature and humidity of farmland soil obtained in the first, third and fourth steps, and the time-frequency domain information of crop images are statistically analyzed using the method of double-layer propagation neural network to obtain accurate identification and judgment of pests and diseases. .

Further, the TensorFlow method is a mathematical statistical analysis and processing method based on adaptive changes.

Further, the specific operation of calculating the gradient distribution of the temperature field and the humidity field of the crop field is as follows:

The process of collecting and processing the soil temperature of crops for identification of crop diseases and insect pests is as follows:

Obtain the temperature value W(t) _n of the current measuring point n at the current moment t and the temperature value W(t) n-1 of the previous measuring point n _-1 for gradient calculation, the calculation formula is:

T _t =W(t) _n -W(t) _n-1

W(t) _n is the temperature value of the current measuring point n at the current time t; W(t) _n-1 is the temperature value of the previous measuring point n-1 at the current time t; T _t is the temperature gradient value. After the T _t temperature gradient value is calculated, the gradient value of each measuring point is described, and the temperature gradient field distribution of the corresponding point is made;

The process of collecting and processing soil moisture in crops for identification of crop diseases and insect pests is as follows:

Obtain the humidity value H(t) _n of the current measuring point n at the current time t and the humidity value H(t) _{n-1 of the previous measuring point n-1} for gradient calculation, and the calculation formula is:

D _t ＝H(t) _n -H(t) _n-1

H(t) _n is the humidity value of the current measuring point n at the current time t; H(t) _n-1 is the humidity value of the previous measuring point n-1 at the current time t; D _t is the humidity gradient value. After the D _t humidity gradient value is calculated, the gradient value of each measuring point is depicted, and the humidity gradient field distribution of the corresponding point is made.

Further, the multi-scale transformation of the crop image in the farmland, the specific operation is as follows:

The process of crop disease and insect pest recognition image segmentation is as follows:

Obtain the image R occupying the entire spatial area, and divide the image into four regions R ₁ , R ₂ , R ₃ , and R ₄ , the process can be expressed as:

R ₁ +R ₂ +R ₃ +R ₄ =R

R is the entire spatial area occupied by the image; R ₁ is the upper left image block of image segmentation, R ₂ is the upper right image block of image segmentation, R ₃ is the lower left image block of image segmentation, and R ₄ is the lower right image block of image segmentation. This time, the image is divided twice, and the divided layer is enlarged to achieve the purpose of multi-scale transformation;

The process of zooming in and transforming the recognition image of crop diseases and insect pests is as follows:

Obtain the one-dimensional abscissa function B(x) of the original image, and obtain the function B(x) obtained by shortening the one-dimensional abscissa function of the original image by one time after processing, and shift B(2x) to the right by one unit to obtain the function B( 2x-1), the calculation formula is:

A(x)=2[B(2x)+B(2x-1)]

A(x) is the double magnification function on the one-dimensional abscissa of the image, B(x) is the one-dimensional abscissa function of the original image, and B(2x) is obtained by shortening the one-dimensional abscissa function of the original image by one time function, B(2x-1) is a function obtained by shifting B(2x) to the right by one unit. The one-dimensional ordinate function is also processed according to this method, and finally the image can be enlarged to four times its original size.

The multi-scale transformation of the present invention has two image segmentation and image magnification processes. As an optimization, the process of image segmentation first, image magnification, image segmentation again, and image magnification again is adopted. The final image obtained is the original image enlarged in length and width. Four times, the area is magnified sixteen times. Since then, the image multi-scale transformation of the present invention has been completed.

Further, the specific operation process of the time-domain transformation of the crop image in the farmland is as follows:

Obtain the coordinates (r, g, b) representing the red, green, and blue images of crop diseases and insect pests, the largest max among r, g, and b and the smallest min among r, g, b, and the calculation formula is as follows:

S=((max-min)/max)*100/255

V=max*100/255

H is the hue on the red, green and blue coordinates (r, g, b) of the image, S is the saturation on the red, green and blue coordinates (r, g, b) of the image, V is the red of the image , green, and blue brightness on the coordinates (r, g, b). After obtaining H, S, and V, calculate the remainder of the row and column pixels of the quantized image, and match them with 72 color regions. Finally, extract the color features of the image and match them with the color library to find out whether there are pests or diseases.

Further, the specific operation process of the frequency domain transformation of the crop image in the farmland is as follows:

Obtain the direction and scale μ and ν of the Gabor filter, the image pixel point coordinates (x, y), represented by z, the Gaussian envelope σ in the frequency domain, the function k _{μ, ν} that controls the width of the Gaussian window, the wavelength of the oscillation part, and the direction . Calculated as follows:

Ψ _μ,v (z) is the Gabor kernel amplitude characteristic; μ and ν represent the direction and scale of the Gabor filter respectively; z represents the image pixel coordinates (x, y); || · || represents the norm; σ is Gaussian envelope; k _{μ, ν} controls the width of the Gaussian window, the wavelength and direction of the oscillating part, defined as

k _ν = k _max /f _υ is the sampling frequency of the filter, k _max is the maximum frequency, and f _υ is the interval factor limiting the distance of the kernel function in the frequency domain;

After the Gabor kernel amplitude characteristic Ψ _μ,v (z) is obtained, it is convolved with the original image to obtain obvious image features, so as to judge diseases and insect pests.

Further, the fifth step is to obtain the farmland temperature field, humidity field, time-domain processing value, and frequency-domain processing value respectively through the above steps, and use a double-layer propagation neural network to perform data statistical analysis, and return the input vector X _i to The calculation formula is as follows:

_Xi is the input vector of each neuron, that is, the soil moisture field, temperature field, time domain calculation value, and frequency domain calculation value of this invention. X _i ^* is the normalization processing done on the input vector X _i .

Not all input neuron nodes can transmit data to the first layer of neural network, only the optimal value can, the calculation formula is as follows:

为当前测点n从输入神经元到第一层神经网输入向量X_i做的归一化连接加权值，归一化方法与X_i归一化方法相同。S_j(n)为X_i ^*与

乘积和的最大值，也即第一层最优值。X _i ^* is the normalization process for the current measurement point n from the input neuron to the input vector X _i of the first layer of neural network.

It is the normalized connection weight value for the current measurement point n from the input neuron to the input vector X _i of the first layer neural network. The normalization method is the same as that of _Xi normalization. S _j (n) is _Xi ^* and

The maximum value of the sum of products, that is, the optimal value of the first layer.

In order to obtain the optimal value of the first layer, it is necessary to adjust the weighted value multiple times. The weighted value correction calculation formula of the connection weight value of the current measurement point n from the input neuron to the first layer of neural network is as follows:

为当前测点n从输入神经元到第一层神经网4个输入值的归一化连接加权值，初始赋值分别为0.1，0.2，0.3，0.4。α为学习率，随着训练不断降低，是一个0—1之间的值，初始赋值0.3。X_i ^*为对当前测点n从输入神经元到第一层神经网输入向量X_i做的归一化处理。

为修正后的当前测点n从输入神经元到第一层神经网的连接加权值。反复修正权值，直至权值不变，即

等于

此时

等于零。

It is the normalized connection weight value of the current measurement point n from the input neuron to the 4 input values of the first layer neural network, and the initial assignments are 0.1, 0.2, 0.3, 0.4 respectively. α is the learning rate, which decreases as the training continues, and is a value between 0 and 1, with an initial value of 0.3. X _i ^* is the normalization process for the current measurement point n from the input neuron to the input vector X _i of the first layer of neural network.

It is the weighted value of the connection from the input neuron to the first layer of the neural network for the current measurement point n after correction. Correct the weight repeatedly until the weight remains unchanged, that is,

equal

at this time

is equal to zero.

Similarly, not all neuron nodes of the first layer of neural network can transmit data to the second layer of neural network, only the optimal value can, the calculation formula is as follows:

乘积和的最大值，也即第一层神经网最优值。L_j(n)为V_jk与S_j(n)乘积和的最大值，也即第二层神经网最优值；V _jk is the connection weight value of the current measurement point n from the neuron node of the first layer of neural network to the second layer of neural network. S _j (n) is _Xi ^* and

The maximum value of the sum of products, which is the optimal value of the first layer of neural network. L _j (n) is the maximum value of the product sum of V _jk and S _j (n), which is the optimal value of the second layer neural network;

In order to obtain the optimal value of the second layer, it is necessary to adjust the weighted value multiple times. The calculation formula of the weighted value correction between the first layer neural network and the second layer neural network is as follows:

V _jk (n+1)＝V _jk (n)+βb _j [(y _k -D _k ]

V _jk (n) is the connection weight value of the current measuring point n from the first layer neural network to the second layer neural network, and the initial assignment values are both 0.5 and 0.5. β is the learning rate, which decreases as the training continues, and is a value between 0 and 1, with an initial value of 0.4. b _j is the binary output vector of the competition layer, when L _j (n) is obtained, b _j =1; otherwise, b _j =0. D _k is the expected output value of the second layer neural network. V _jk (n+1) is the weighted connection weight value of the current measuring point n from the first layer of neural network to the second layer of neural network after correction. y _k is the actual output value of the second layer neural network, and the calculation formula is as follows:

V _kj is the connection weight value from the first layer neural network to the second layer neural network, V _kj ^* is the product of the connection weight value V _kj and S _j (n) from the first layer neural network to the second layer neural network, when When L _j (n) is obtained, that is, when the product sum of the optimal value of the first layer neural network and the connection weight value of the second layer neural network is the largest, b _j = 1 at this time, and V _kj ^* is L _j (n); no or b _j =0, y _k =0.

Beneficial effect:

The invention overcomes the problem of difficult identification of farmland crop diseases and insect pests in existing complex scenarios, and can scientifically identify farmland crop disease and insect pests.

Description of drawings

Fig. 1 is the time-frequency domain multi-scale identification system and method flow chart of pests and diseases based on TensorFlow of the present invention;

Fig. 2 is the enlarged schematic diagram of multi-scale segmentation of farmland diseases and insect pest images of the present invention;

Fig. 3 is a schematic diagram of a time-domain HSV model of farmland diseases and insect pest images of the present invention;

Figure 4 is a schematic diagram of a two-layer propagation neural network.

Detailed ways

Below in conjunction with specific embodiment the present invention is described in further detail:

Fig. 1 is the time-frequency domain multi-scale identification system and method flow chart of pests and diseases based on TensorFlow (a symbolic mathematics system based on data flow programming) of the present invention. The system first initializes and measures the soil temperature and humidity of the current measuring point n and crop image information; then it performs the gradient solution process of the temperature field of the farmland soil, the gradient solution process of the humidity field of the farmland soil, the first segmentation process of the crop image, and the crop image information. The first enlargement process of the image, the second segmentation process of the crop image, the second enlargement process of the crop image, the time domain transformation process of the crop image, and the frequency domain transformation process of the crop image; finally, the temperature and humidity gradient field and the crop image are respectively obtained through the above steps For time-frequency domain information, the data is statistically analyzed and processed using the double-layer propagation neural network method to determine whether pests and diseases occur at the current measuring point n.

Fig. 2 is a schematic diagram of multi-scale segmentation and enlargement of farmland diseases and insect pest images according to the present invention. 1 is the first layer, which is the original image captured or the image that has undergone simple physical processing. The first layer is divided into four parts, and the upper right layer is taken and enlarged to 2 to form the second layer. The upper right part of the first layer is magnified four times in area. Continue to divide the second layer into four parts, take the upper right layer and enlarge it to 3 to form the third layer. At this time, the area of the upper right part of the second layer is enlarged by four times, and the area of the upper right part of the first layer is magnified sixteen times. According to this method, each part of the first layer can be enlarged, that is, multi-scale transformation is realized.

Fig. 3 is a schematic diagram of a time-domain HSV model of an image of farmland diseases and insect pests according to the present invention. 1 is the angle of the red hue; 2 is the angle of the yellow hue; 3 is the angle of the green hue; 4 is the angle of the cyan hue; 5 is the angle of the blue hue; 6 is the angle of the magenta hue. 7 is the hue (H) rotation direction. 8 is the extension direction of saturation (S). 9 is the extending direction of brightness (V). According to the combination of 7 (hue), 8 (saturation), and 9 (brightness), the crop images can be classified and processed.

Fig. 4 is a schematic diagram of a two-layer propagation neural network of the present invention. 1 is the temperature gradient value of the current measuring point n; 2 is the humidity gradient value of the current measuring point n; 3 is the time domain processing value of the crop image of the current measuring point n; 4 is the frequency domain processing value of the crop image of the current measuring point n; 5 and 6 are two kinds of judgment results (yes or not); 7 is an input neuron node, 8 is a first-layer neural network node, and 9 is a second-layer neural network node. The measurement point data is optimized from nodes 1, 2, 3, and 4, and the weights are adjusted. The optimal value enters the first layer of neural network; after optimization, the weight is adjusted, and the optimal value enters the second layer of neural network. Find out whether pests and diseases have occurred at the measuring point on TensorFlows.

Although the embodiments of the present invention have been shown and described, those skilled in the art can understand that various changes, modifications and substitutions can be made to these embodiments without departing from the principle and spirit of the present invention. and modifications, the scope of the invention is defined by the appended claims and their equivalents.

Claims (1)

1. The operation method of the pest and disease damage time-frequency domain multi-scale identification system based on the TensorFlow is characterized in that the system comprises an OV7725 camera sub-node acquisition unit, a LoRa transmission sub-node transmission unit, a soil entropy sensor sub-node acquisition unit, a stm32 node processing unit and other hardware platforms based on the Internet of things technology, and acquired images and farmland information are transmitted to a monitoring center; a data statistical analysis method based on TensorFlow is carried out in a monitoring center, crop images and farmland information are analyzed and judged, and accurate pest and disease identification judgment is obtained;

the operation method comprises the following steps:

step one, obtaining gradient distribution of a temperature field and a humidity field of a crop farmland;

secondly, carrying out multi-scale transformation on farmland crop images;

thirdly, performing time domain transformation on farmland crop images;

fourthly, transforming the farmland crop image frequency domain;

fifthly, data fusion, namely carrying out data statistics analysis on farmland soil temperature and humidity and crop image time-frequency domain information obtained in the first step, the third step and the fourth step by using a double-layer transmission neural network method to obtain accurate plant disease and insect pest identification judgment;

the TensorFlow method is a mathematical statistical analysis processing method based on self-adaptive change;

the gradient distribution of the temperature field and the humidity field of the crop farmland is obtained, and the specific operation is as follows:

the temperature acquisition and treatment process of the crop pest and disease damage identification farmland soil is as follows:

acquiring a temperature value W (t) of a current measuring point n at a current moment t _n And the temperature value W (t) of the last measuring point n-1 _n-1 Gradient operation is carried out, and the calculation formula is as follows:

T _t ＝W(t) _n -W(t) _n-1

W(t) _n the temperature value of the current measuring point n at the current moment t; w (t) _n-1 The temperature value of the last measuring point n-1 at the current time t; t (T) _t For the temperature gradient value, T is _t After the temperature gradient value is calculated, the gradient value of each measuring point is drawn, and the temperature gradient field distribution of the corresponding point is made;

the crop pest and disease damage identification farmland soil humidity acquisition and treatment process comprises the following steps:

acquiring a humidity value H (t) of a current measuring point n at a current moment t _n And the humidity value H (t) of the last measuring point n-1 _n-1 Gradient operation is carried out, and the calculation formula is as follows:

D _t ＝H(t) _n -H(t) _n-1

H(t) _n the humidity value of the current measuring point n at the current moment t; h (t) _n-1 The humidity value of the last measuring point n-1 at the current moment t; d (D) _t For the humidity gradient value, D _t After the humidity gradient value is calculated, the gradient value of each measuring point is drawn, and the humidity gradient field distribution of the corresponding point is made;

the farmland crop image multi-scale transformation is operated as follows:

the crop pest identification image segmentation processing process comprises the following steps:

acquiring an image R occupying the whole spatial region, dividing the image into R ₁ 、R ₂ 、R ₃ 、R ₄ Four regions, the process of which can be expressed as:

R ₁ +R ₂ +R ₃ +R ₄ ＝R

r is the whole space area occupied by the image; r is R ₁ Upper left image block for image segmentation, R ₂ Upper right image block for image segmentation, R ₃ Lower left image block for image segmentation，R ₄ For the lower right image block of the image segmentation, the image is segmented twice, and the segmented image layer is amplified;

the crop pest identification image amplification and transformation processing process comprises the following steps:

obtaining a one-dimensional abscissa function B (x) of an original image, processing to obtain a function B (x) obtained by shortening the one-dimensional abscissa function of the original image by one time, and translating the function B (2 x-1) obtained by shifting the function B (2 x) to the right by one unit, wherein the calculation formula is as follows:

A(x)＝2[B(2x)+B(2x-1)]

a (x) is a double amplification function on a one-dimensional abscissa of an image, B (x) is a one-dimensional abscissa function of an original image, B (2 x) is a function obtained by shortening the one-dimensional abscissa function of the original image by one time, B (2 x-1) is a function obtained by translating B (2 x) to the right by one unit, and a one-dimensional ordinate function is treated according to the method, so that the image can be amplified to be four times of the original image finally;

the multi-scale transformation comprises two image segmentation and image amplification processes, namely image segmentation, image amplification, image segmentation, image amplification again and image amplification again, wherein the finally obtained image is four times longer and wider than the original image, and the area is sixteen times larger, so that the multi-scale transformation of the image is completed;

the farmland crop image time domain transformation specifically comprises the following steps:

coordinates (r, g, b) representing red, green and blue of the crop pest images are obtained, and the maximum one of r, g and b is max and the minimum one of r, g and b is min, and the calculation formula is as follows:

S＝((max-min)/max)*100/255

V＝max*100/255

h is the hue size on the red, green and blue coordinates (r, g, b) of the image, S is the saturation size on the red, green and blue coordinates (r, g, b) of the image, V is the brightness size on the red, green and blue coordinates (r, g, b) of the image, after H, S and V are obtained, the quantized row pixels and column pixels of the image are subjected to redundancy and are matched with 72 color areas, and finally the color characteristics of the image are extracted and matched with a color library to obtain whether diseases and insect pests exist or not;

the farmland crop image frequency domain transformation comprises the following specific operation processes:

the directions and the scales mu and v of the Gabor filter are obtained, the coordinates (x, y) of the pixel points of the image are expressed by z, the frequency domain Gaussian envelope sigma is expressed by z, and the width of the Gaussian window, the wavelength of the oscillation part and the function k of the directions are controlled _μ,ν The calculation formula is as follows:

Ψ _μ,v (z) is Gabor kernel amplitude characteristic; μ and ν represent the direction and scale, respectively, of the Gabor filter; z represents the image pixel coordinates (x, y); I.I representing norms; sigma is a gaussian envelope; k (k) _μ,ν The width, the wavelength of the oscillation part and the direction of the Gaussian window are controlled and are defined as

k _ν ＝k _max /f _υ For the filter sampling frequency, k _max At maximum frequency f _υ A spacing factor that limits the kernel distance in the frequency domain,

obtaining Gabor kernel amplitude characteristic ψ _μ,v After (z), convolving with the original image to obtain obvious image characteristics so as to judge the plant diseases and insect pests,

further, the fifth step is to obtain farmland temperature field, humidity field, time domain processing value and frequency domain processing value respectively, and then use double-layer propagation neural network to perform data statistical analysis to input vector X _i The normalization processing is carried out, and the calculation formula is as follows:

X _i for the input vector of each neuron, namely the soil humidity field, the temperature field, the time domain calculated value, the frequency domain calculated value and X of the invention _i ^* For input vector X _i Performing normalization treatment;

not all input neuron nodes can transmit data to the first layer neural network, only the optimal value can be calculated as follows:

X _i ^* to input vector X from the input neuron to the first layer neural network for the current measuring point n _i Normalization processing of W _j ^* (n) input vector X from input neuron to first layer neural network for current measurement point n _i Normalized connection weighted value, normalization method and X _i The normalization method is the same, S _j (n) is X _i ^* And (3) with

The maximum value of the product sum, i.e. the first layer optimum;

in order to obtain the first-layer optimal value, the weighting value needs to be adjusted for a plurality of times, and the connection weighting value correction calculation formula of the current measuring point n from the input neuron to the first-layer neural network is as follows:

the normalized connection weight values of the current measuring point n from the input neuron to the 4 input values of the first layer neural network are respectively 0.1,0.2,0.3,0.4,alpha is learning rate, is a value between 0 and 1 along with continuous decrease of training, and is initially assigned with 0.3, X _i ^* To input vector X from the input neuron to the first layer neural network for the current measuring point n _i The normalization processing is carried out, so that the data of the data are processed,

for the connection weighted value of the corrected current measuring point n from the input neuron to the first layer neural network, repeatedly correcting the weighted value until the weighted value is unchanged, namely +.>

Equal to->

At this time->

Equal to zero;

also, not all first layer neural network neuron nodes can transmit data to the second layer neural network, only the optimal value can be calculated as follows:

V _jk for the connection weight value of the current measuring point n from the first layer neural network neuron node to the second layer neural network, S _j (n) is X _i ^* And (3) with

The maximum value of the product sum, i.e. the first layer neural net optimum, L _j (n) is V _jk And S is equal to _j (n) the maximum value of the product of the sums, i.e., the second layer neural net optimum;

in order to obtain the second-layer optimal value, the weighted value needs to be adjusted for a plurality of times, and the weighted value correction calculation formula between the first-layer neural network and the second-layer neural network is as follows:

V _jk (n+1)＝V _jk (n)+βb _j [(y _k -D _k ]

V _jk (n) the connection weight value of the current measuring point n from the first layer neural network to the second layer neural network is 0.5, the initial assignment is 0.5, the beta is the learning rate, the value between 0 and 1 is continuously reduced along with training, the initial assignment is 0.4, b _j As the binary output vector of the competitive layer, when L _j (n) when taken, b _j =1; whether or not b _j ＝0，D _k V is the expected output value of the second layer neural network _jk (n+1) is the connection weighted value of the corrected current measuring point n from the first layer neural network to the second layer neural network, y _k For the actual output value of the second-layer neural network, the calculation formula is as follows:

V _kj weighting the connection of the first layer of neural network to the second layer of neural network, V _kj ^* Weighting value V for connection of a first layer of neural network to a second layer of neural network _kj And S is equal to _j (n) when L _j (n) when the sum of the weighted value multiplication of the first layer neural network optimal value and the second layer neural network connection weighted value is maximum, b _j ＝1，V _kj ^* Is L _j (n); whether or not b _j ＝0，y _k ＝0。

Images