CN111191543A

CN111191543A - Rape yield estimation method

Info

Publication number: CN111191543A
Application number: CN201911328528.0A
Authority: CN
Inventors: 周立波
Original assignee: Hongfujin Precision Industry Shenzhen Co Ltd
Current assignee: Hongfujin Precision Industry Shenzhen Co Ltd; Hunan City University
Priority date: 2019-12-20
Filing date: 2019-12-20
Publication date: 2020-05-22

Abstract

The invention discloses a rape yield estimation method, which comprises the following steps: 1) sampling: shooting rape canopy images in modes of an unmanned aerial vehicle and the like, and selecting sample data of a plurality of sample areas; 2) pretreatment: preprocessing the image; 3) setting key performance indexes: acquiring data according to the influence factors of the rape yield; 4) modeling: establishing a model according to the image, extracting features of the input image, and detecting key performance index information in each sample; the associated estimated yield is calculated. According to the rape yield estimation method, the characteristic of the existing image is found out by researching the plant type characteristic, the yield actual measurement model is established, and the construction and the training of the image recognition yield estimation model are guided.

Description

Rape yield estimation method

Technical Field

The invention belongs to the technical field of crop yield measurement, and particularly relates to a rape yield estimation method.

Background

In agricultural production, the estimation of the rape yield has strong practical significance. The rape yield is estimated, so that not only is the planting area of the rape adjusted by farmers at any time according to price change and the income of the farmers improved, but also the planting area of the rape can be enlarged or reduced in time according to the market condition of the rape.

The yield measurement of rape is a tedious work, the traditional method is manual estimation, the method is time-consuming and labor-consuming, the cost consumption is excessive, and various subjective errors can occur.

At present, the estimation of the yield of most rapes is mostly based on planting experience, and the deviation is very large, so that the problem that the estimation of the yield of the rapes by finding an effective method is urgently needed to be solved. For example, a large amount of peanuts are planted in northern China, the nutritional value of the peanuts is higher than that of grains, the peanuts contain a large amount of protein fat, and at present, no method for accurately estimating the peanut yield exists.

For some rape, yield estimation models have been established. For example, a number of wheat growth simulation systems are internationally available, such as the DSSAT system in the united states, the APSIM system in australia, the WheatGrow system in our country, and so on. The method quantitatively estimates the growth and development and yield formation process of wheat by analyzing the mechanism relationship between weather-soil-technical measures and the physiological and ecological process of wheat. However, the existing model for estimating the wheat yield has more used parameters, the data processing process is very complicated, and the estimation precision and speed are still to be improved.

In addition, when the rape yield is estimated, the time variable is in units of years or the whole growth period, and the difference of the rape caused by meteorological changes in different growth periods actually included in the growth period is ignored, so that the estimation accuracy of the rape yield is insufficient.

The rape yield estimation technology is divided into the following three major categories:

(one) classifying according to estimated method or model

The advanced measurement technology for the rape yield can be classified according to estimation modes, including statistical estimation, remote sensing estimation, dynamic growth model estimation and the like. When the connection between the rape growth and the external environment is not clear, a plurality of regression technologies are used, the connection between the rape growth and the external environment is clarified, a model is constructed after statistical results, and the model is subjected to subsequent analysis; the dynamic growth simulation estimation requires artificial simulation of the growth environment of the rape, the exchange of substances and energy is carried out all the time in the growth process of the rape, a relatively balanced state can be kept, the growth state of the rape is simulated in the whole process from the beginning to the end of the test according to the information of the substances and the energy in the actual growth condition of the rape, and the relation between the yield of the rape and the external environment is researched.

(II) classifying according to the estimated objects

The advanced measurement technology for rape yield can be classified according to research objects and can be divided into the following three types: grain crops, commercial crops and famous special-quality small crops. In China, the estimation of grain crops has a history of many years, so the application is very wide, the technology is relatively mature, the safety of the grain can be better guaranteed only by estimating the total production amount of the grain crops and the yield of each crop in a large range in China, the method has guiding significance in the process of preparing agricultural related policies, and the method has important value in maintaining the stability of the country. The economic crops mainly refer to cotton, rape, peanut and the like. The third famous superior small crops mainly refer to Chinese chestnut, mung bean, high-quality fruits and the like.

(III) classifying according to forecast age

The advanced measurement technology for the rape yield can be classified according to time effect and can be divided into the following four types:

the first category is annual scene estimation. Different crops have different growth and development processes, the requirements on the external growth environment are different, and the influence of meteorological conditions on the yield of rape is judged by estimating the environmental condition in a period of time in the future, so that the estimation mode becomes annual scene estimation. Usually, the rape is released before or just after sowing, and the grower can adjust the rape planting plan in the year according to the released information.

The second category is trend estimation. In the period from before and after the rape is sowed to 60 days before the rape is matured, when the external environment changes, certain influence is caused on the growth and development of the rape, the climate change in a period of time in the future can be estimated, whether the rape can be harvested well or not can be judged according to the estimation result, and the like, and the estimation mode is called trend estimation. The rape seed oil is usually released 60 days before rape maturity, can guide the formulation of agricultural policies on the whole, and is also beneficial to the effective management of production.

The third category is quantitative estimation. After the rape is sowed in the field, the rape grows for a period of time till 30 days before maturity, the change of the external environment in the period of time can cause favorable influence or adverse influence on the growth of the rape, and the climate change of the next several days is estimated, so that the yield of the rape is judged, and the method is called quantitative estimation. Usually published one month before maturity, and has certain guiding significance in the aspects of agricultural overall policy establishment, handling, storage, transportation, flow, buying and selling and the like of the rapes.

The fourth category is dynamic estimation. The method is called dynamic estimation, and judges whether the influence on the normal growth and development of the rape can be generated or not by observing the change of the external environment condition at any time, thereby estimating the change of the yield. After the rape is sowed in the field, the change of climate can generate beneficial or harmful effect on the growth of the rape at each stage of the growth, and the crop yield is tracked and estimated in real time in one month, three months or one quarter by taking the weather condition of a future period as a basis for consideration. The method is used in combination with the change of the climatic environment, has strong timeliness, can cause the yield of the rape to fluctuate when the external environment is changed greatly, for example, under extreme severe conditions, and can be used for developing protective measures in time by a grower according to an estimation result, so that the adverse effect of disasters on the rape is avoided, the rape production is stable and efficient, and the method has guiding value on the grain safety production work of the whole country. Currently, this estimation has been used to estimate crop yield at the national level and to effectively serve production.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: aiming at the technical problems in the prior art, the invention provides a rape yield estimation method.

In order to solve the technical problems, the technical scheme provided by the invention is as follows:

a rape yield estimation method comprises the following steps:

1) sampling: shooting a rape panorama, and selecting sample data of a plurality of sample areas;

2) pretreatment: preprocessing the image;

3) setting key performance indexes: acquiring data according to the influence factors of the rape yield;

4) modeling: establishing a model according to the image, extracting features of the input image, and detecting key performance index information in each sample; the associated estimated yield is calculated.

As a further improvement of the technical proposal;

in the scheme, preferably, in the step 1), in the rape pod mature period, the aerial photography canopy of the unmanned aerial vehicle is adopted, and the high, medium and low altitude full coverage is performed, so that shooting is performed repeatedly.

In the foregoing solution, preferably, in the step 2), the image is subjected to one or more of cropping, brightness adjustment, contrast adjustment, color enhancement, rotation, and sharpening.

In the above scheme, preferably, the image in step 4) is modeled based on the YOLOv3 technology, the whole image is input into the image recognition model RYPY, and features are extracted from the input image through a feature extraction network.

In the above scheme, preferably, in the step 1), when sampling the rape sample, firstly, the canopy image is collected, and then the measurement information of the plant of the sample is taken.

In the foregoing scheme, preferably, in the step 3), the key performance indexes are the number of siliques, the number of grains per silique and the thousand-grain weight.

In the foregoing aspect, preferably, in the step 4), the features are extracted from the input image through a feature extraction network, and information of the number of pods, the number of branches, and the effective branch length in each rape sample is detected.

In the above scheme, preferably, in the step 4), according to the characteristics of the rape plants, taking the yield as a dependent variable Y and the branch length, the branch number and the effective branch length as independent variables X, performing multiple linear regression analysis on the measurement data of the rape plants by adopting a stepwise regression method, establishing a multiple linear regression equation, and calculating the estimated yield of the rape plants; the regression equation is:

Y＝0.811X₁+0.015X₃/X₂(R²＝0.820)。

wherein Y is the yield of the rape plants; x₁The effective branches are all long, and the applicable value range is 0<X₁<80；X₂Is the number of branches; x₃Total number of siliques, X₃/X₂The applicable value range is 0<X₃/X₂<90, respectively; coefficient of correlation R²The value was 0.820.

Compared with the prior art, the rape yield estimation method provided by the invention has the following advantages:

according to the rape yield estimation method, the characteristic of the existing image is found out by researching the plant type characteristic, the yield actual measurement model is established, and the construction and the training of the image recognition yield estimation model are guided.

According to the rape yield estimation method, the final stage of the rape growth and development process is the mature period of the pod growth, the time is about 30 days, the leaves completely fall off in the period, and the growth conditions of the pod and branches have high visibility in the aerial photography overlooking angle. By aerial photographing the canopy image of the rape sample by the unmanned aerial vehicle, counting and analyzing the growth conditions of the siliques and the branches by using an image recognition technology, measuring the actual yield of the sample, and establishing a correlation model of the canopy image of the rape and the actual yield, a rapid yield estimation method is provided for the field rape.

Drawings

FIG. 1 is a technical scheme of the present invention.

FIG. 2 is a schematic diagram of rape plant type parameters.

Detailed Description

The following describes in detail specific embodiments of the present invention. It should be understood that the detailed description and specific examples, while indicating the present invention, are given by way of illustration and explanation only, not limitation.

FIGS. 1 and 2 show the method for estimating the yield of rape of the present invention, and illustrate the application of the method for estimating the yield of rape of the present invention.

The yield estimation method of the embodiment adopts the unmanned aerial vehicle to shoot the images, and comprises the following steps:

1) selecting a sample

A field block which is uniform in planting and good in growth vigor is selected as a reference field block, 20 samples with the square meters of 0.25 are designed in the south-north direction and 10 samples with the square meters of 0.25 are designed in the east-west direction. The number of the sample is 300, wherein the sample numbers are 01-10. The south and north sample prescription is B01-B20, the east and west sample prescription is Y01-Y10, and the longitude and latitude of the center position of the sample prescription are 112.68415 and 28.91786. The samples were spread over an area 1000 meters long and 500 meters wide.

2) Sampling

When sampling a rape sample, firstly, acquiring a canopy image, and specifically comprising the following steps:

(1) the method comprises the steps of using an unmanned aerial vehicle to take an aerial panorama, taking 3 times back and forth according to the full coverage of a space with a height of 10m, a middle height of 20m and a low height of 5m, selecting five sampling points at four corners and the center, taking 1 picture from the height of 1m above a canopy to 20m, and taking a video at the height of 1m per liter.

(2) And selecting a sample, starting to select the sample at a distance of more than 2m from the sample boundary, and numbering the samples.

(3) Firstly, sticking the sample plot number and the sample plot number sticker on a sample square frame, and then placing the sample square frame right above the sample square.

(4) A herringbone ladder is used for taking more than 5 photos from the position 1m higher than the canopy layer by using a camera from different angles.

(5) And measuring the planting density.

(6) The plant of the same side is taken and marked with the sample plot number and the sample side number.

(7) The plant height, the number of branches, and the height of the branch position of each branch on the main diameter were measured.

(8) Cut each branch from the main inflorescence down and number each branch from top to bottom, for example with the number: 00. 01, 02 ….

(9) The total length of the effective branches (the branches referred to herein are those with healthy silique production) were measured and the branch number and branches were photographed for 5 or more from each angle.

(10) The effective branch length (branch length of fruit part) was measured.

(11) The quantity of each branch was counted and the siliques were removed and weighed.

(12) And (4) respectively bagging the hornfruits of each sample square and numbering.

(13) The siliques are dried in the sun according to the prescription, and the shell is removed, the weight is weighed, the thousand seed weight is measured, and the yield of the rapeseeds of each prescription is recorded.

(14) And (6) counting the results.

All parties require more than 2m from the same-land boundary to ensure stability and representativeness of data. Data were collected for a total of 10 plots and 300 plots. Through preliminary screening, 3720 aerial images of unmanned aerial vehicles at different heights, 2170 images of the sample squares and 19740 images of 3366 branches in 300 sample squares are obtained, and 25630 effective images are obtained in total. 300 varieties including rape plant type parameters, planting density, variety yield and the like are collected. Data were collected for yield, silique number, branch length, silique weight, etc. of 3366 branches.

In the mature period of rape pod development, the pod is not only the place for rape seed growth and development, but also responsible for the main photosynthesis in the period. The horns are very thick and opaque, so that the distribution of the horns can naturally avoid mutual shielding and covering, each horn can be ensured to receive light as much as possible, and the visibility of the horns is very high from the aerial angle. These typical canopy image features are the basis for the image processing technique to estimate rape yield.

3) Image pre-processing

In the embodiment, enough rape sample images are obtained by adopting multi-angle, multi-time-period and multi-azimuth shooting.

Considering that the problems of uneven illumination, inconsistent illumination intensity, low definition and the like can bring adverse effects to subsequent rape feature recognition, a series of preprocessing such as cutting, brightness adjustment, contrast adjustment, color enhancement, rotation, sharpening and the like needs to be carried out on the rape sample side image, and the calculated amount during image feature extraction is reduced, so that the overall detection speed is improved, the training set can be increased, and the detection accuracy is improved.

4) Key performance index setting

(1) Rape plant type characteristic analysis

The main research on rape plant characteristics includes plant height, branch number, branch length, effective branch length (with healthy silique branch length), silique number, branch height, area, etc. Correlation analysis was performed on each parameter using Pearson correlation coefficient method, and the correlation data processing was performed in Excel 2016 and SPSS 23.0.

a. Rape pod number correlation analysis

The institute measured 300 samples of rape plants with 20% lower branches, 30% upper branches and 50% average branches. Typical silique distribution of homogeous oilseed rape plants is shown in table 1.

TABLE 1 typical plant silique distribution table for homoeogenic branches

The main inflorescence branches have the best illumination conditions, so that the silique yield is high, the siliques are heavy, and the density of the branched siliques is high. As the branch height decreases, the silique yield appears to be normally distributed from low to high and then low. The weight of the horns is relatively stable in average, and the weight of the horns with the lowest branches is relatively light.

The secondary branching does not have independent record processing, so that the density of the branched siliques is increased abnormally, and the average weight of the siliques is reduced abnormally due to the low quality of the secondary branching siliques. From the perspective of canopy image identification, there is no material impact on yield estimation. The basic information for the estimation of silique yield is derived from branches, taking the typical silique and branch parameters of homoeogenic branches as an example, the corresponding relationship is shown in Table 2.

TABLE 2 relation table of typical plants of homoeogenic branches for silique and branch parameters

The collected sample data is sorted, and statistics of the related information such as the number of the branch fruits is shown in table 3. The average yield of the No. 00 main inflorescence is the highest, accounting for 11.8 percent of the total yield, and the density of the branched pod and the effective branched pod density are also the highest. The average yield of No. 11 branches is reduced because the average branch number of the plant types of the homogenic branches and the lower branches is about 11, and the lowest branch yield is low, so that the statistical data of No. 11 branches are abnormally reduced, and the surface of 12 branches which relatively rise is unreasonable. The statistical data trend of the density of the branched siliques and the density of the effective branched siliques and the yield of the siliques synchronously fluctuate, and influence caused by plant type difference is mutually verified.

TABLE 3 statistical table of the distribution of the silique branches of rape

Research shows that the quantity of the siliques and the branching positions have no obvious correlation, and the quantity of the siliques tends to decrease as the branching positions decrease on the whole; the main inflorescence has high yield and is stable. There was no clear correlation between the distribution of the weight of each silique in the number of different branches.

Statistics of the correlation between the amount of siliques from rape and the branching parameters are shown in Table 4. And counting the average data of the branch length and the effective branch length. And calculating the crown layer difference and the branch growth amount, wherein the crown layer difference is the difference between the highest point of the branch and the highest point of the main inflorescence, and the specific calculation mode is as follows: the difference in the canopy height-branch length-branch height, and the illumination of the branches can be analyzed according to the difference in the canopy height. The branch growth amount is the increase value of branch biomass represented by subtracting the previous branch length from the present branch length.

TABLE 4 relationship table of rape pod and branch parameters

The study showed that the number of branch angles was not significantly correlated with branch length, but the overall trend was positively correlated. The branch angle number and the canopy difference accord with the illumination distribution rule, but the correlation is not high. The correlation between the number of branch siliques and the amount of branch growth is not strong. The correlation between the quantity of the siliques and the height of the siliques is extremely small, and the yield distribution of the siliques of the rape branches is relatively balanced on the whole.

b. Correlation analysis of branch parameters

The number of branches, the length of the branches and the effective length of the branches determine the yield of the siliques of the rape. The study first measured the parameters associated with 3 different plant types of canola and then statistically analyzed the relationship between the parameters and the branches.

The measurement data related to the branch parameters of the 8 th sample side of the lower branch plant type and the 10 th sample side of the longitudinal direction are shown in Table 5. Including sample size, number of branches, effective stem length, etc., wherein the effective stem length is the difference between the highest branch height and the lowest branch height of the stem and is the effective distribution interval of the branches on the stem. The lower branch type has more branches and shorter plant height, part of the lowest branches grow from the base of the plant, and the effective length of the main stem is longer. In the three samples of Y03, Y06 and Y10 in the table, the lowest branch is relatively high, the corresponding branch building is also obviously reduced, the effective length of the main stem is also short, and the three are related to each other.

TABLE 5 TABLE for the related measurement of branch number of branch plant type sample plot

The measurement data related to the branch parameters of No. 10 samples longitudinally ahead of No. 7 samples of the homoeogenic branch plants are shown in Table 6. The measurement parameters are as in Table 5. The number of branches of the homogenetic branch plant is moderate, the plant height is slightly high, the lowest branch is 144-410mm, the effective length of the main stem is longer, and the number of branches and the area of the check square have certain positive correlation.

The measurement data related to the branch parameters of the No. 4 sample longitudinal direction and the No. 10 sample longitudinal direction of the superior branch plant type are shown in Table 7. The measurement parameters are as in Table 5. The upper branch type has less branches, higher plant height, the lowest branch is 478-857mm, and the effective length of the main stem is shorter.

TABLE 6 correlation measuring table for branch number of homogemc branch plant type sample plot

TABLE 7 correlation measurement table for branch number of branch plant type sample plot

Statistics of data related to various plant type branching parameters are shown in Table 8, and include the lowest branch height, the highest branch height, the effective length of the main stem, the length of the branch node, the plant height, and the plant area per plant, which may be correlated with the number of branches. Branch node length refers to the length of the main stem occupied by each branch. Table 8 all data are statistical averages. Research shows that the number of branches is in linear positive correlation with the effective length of the main stem. The branch number and the lowest branch are linearly and negatively correlated in trend and is consistent with the phenomenon that the branch number of the lower branch type rape is large. The number of branches is minimally correlated with the highest branch height. The correlation of the number of branches with plant height, school square area and branch node length is small. The number of branches has no direct relation with the characteristics of the plant types and the like.

TABLE 8 statistical table of the relationship between the number of branches and the plant type

c. Modeling of rape yield and plant characteristics

The yield of rape is composed of three factors of the quantity of siliques, the quantity of kernels per silique and the weight of thousand kernels. The variation of the three factors is limited by the characteristics of varieties on one hand and is influenced by cultivation management measures and climatic factors on the other hand. The factor with the largest variation range and the largest influence on yield is the quantity of siliques, and the quantity of siliques and the grain weight are mainly related to the characteristics of varieties and have smaller variation under different cultivation conditions. The diameter analysis of the three elements shows that the direct effect on the yield is maximum in the number of horns and fruits, multiple in the number of horns and grains and minimum in the weight of thousand grains. It can therefore be concluded that the pod growth has a decisive effect on rape yield. The measured data show that the quantity of the siliques is in a positive linear correlation with the yield.

Rape pod grows on the main stem and branches, so that the length and yield of the pod branches are physiologically connected. According to the statistics of the measured data, the cumulative length of the silique branches is in extremely obvious linear positive correlation with the yield. The cumulative length of the silique branches has a very good linear relation with the yield, and the length of the silique branches can be used as an important parameter of a yield estimation model after being extracted in the image processing process. In the canopy image, the identification degree is better, and a reliable identification model can be trained through a large number of samples.

The sample used in the study was 0.25m in area²The average number of effective branches (with healthy siliques) of a sample plant is 11.2, the effective branches are distributed in the sample square in a more balanced way and are easy to identify in a canopy image. The measured data show that the effective branch number is in positive exponential correlation with the yield, R²Comprises the following steps: 0.5247, the number of effective branches correlated very significantly with yield (P < 0.001).

The significant characteristics of the effective branches in the canopy image can effectively count the number of the effective branches through an image recognition technology, and the significant correlation between the effective branches and the yield is an important parameter for applying the image processing technology to rape yield estimation.

In the present example, the thousand-grain weight of rape seeds in all the sample plots is measured by randomly selecting 3 sample plots from each sample plot, taking 1000 rapeseeds from each sample plot, weighing the rapeseeds by an electronic scale with the precision of 0.001g, and averaging the values. The measurement results are shown in table 9. The maximum thousand-grain weight is 4.82g, the minimum is 3.69g, and the average is 4.00 g. The measured data shows that the material weight and the yield are in a linear positive correlation, R²Comprises the following steps: 0.1126, but the correlation was not significant.

Table 9 sample data measuring meter

The above studies show that 4 parameters of the rape pod number, the effective branch number, the branch length and the effective branch length are in extremely obvious linear correlation with yield, and in a canopy image, the rape pod number, the effective branch number, the branch length and the effective branch length have obvious image characteristics. Therefore, the rape yield estimation method based on the image recognition technology has a very good theoretical basis. And the three parts are supported logically, so that the rape yield estimation process can be tested mutually, and the reliability of the estimation result is improved.

5) Training test set for branch artificial labeling image

The rape yield estimation model training test database consists of an artificially labeled image set and an unlabeled image set. The artificial annotation images include branch images and silique images.

(1) Artificial labeling of branch images

The manual labeling of the sample branch image is shown in table 10 as the first 20 branch labeling information, and the pixel scale of the sample image is adjusted to 1151 × 648 after the sample image is processed. The image labeling information mainly comprises a central coordinate X, a central coordinate Y, a frame width and a frame height. The frame length of the image branch label is between 133 pixels and 278 pixels, the average is 209 pixels, the frame height is 113 pixels and 336 pixels, the average height is 241 pixels, the maximum is 86772 pixels, and the average is 51406 pixels.

TABLE 10 Manual annotation of branched images

(2) Artificial labeling of silique images

The manual labeling of the silique image is shown in table 11. Is the first 20 corner labeling information, and the pixel size of the sample image after processing is 1151 × 648. The image labeling information mainly comprises a central coordinate X, a central coordinate Y, a frame width and a frame height. The image corner marks have a border length of 29-120 pixels, an average of 95 pixels, a border height of 53-223 pixels, an average height of 131 pixels, and a total of 3004 pixels, a maximum of 14961 pixels, and an average of 7499 pixels.

TABLE 11 artificial annotation of silique images

6) Feature extraction

A CNN (Convolutional Neural Network) algorithm is adopted, the essential characteristics of a CNN abstract data set are utilized, the characterization information is mined, a regressor is used for replacing a CNN classifier, a quantitative regression estimation model is constructed, and the specified data estimation can be input.

Convolutional layer convolution operations are shown below:

in the above formula, the first and second carbon atoms are,

represents the ith characteristic diagram of the ith layer,

the jth convolution kernel, which represents the l-1 th and l-th layers in the network, the symbol "x" represents the convolution operation,

representing the bias vector for layer i. And finally, completing nonlinear activation by the linear combination, thereby forming a final characteristic diagram of the l layers.

7) Building models

In the embodiment, a model is established based on the YOLOv3 technology to complete the image recognition process in the sample, and the YOLO algorithm directly treats the monitoring problem of the target object as a regression problem of position coordinates and confidence scores, so that the YOLO algorithm can estimate the types and positions of a plurality of targets in real time at one time.

For an input image, the model of the embodiment maps the input image to output tensors of 3 scales, which respectively represent the probability that various objects exist at each position of the image. For example, for an input image of 416 × 416, 3 prior frames are set for each grid of the feature map at each scale, for a total of 13 × 3+26 × 3+52 × 3 — 10647 estimates. Each estimate is an 85-dimensional vector consisting of (4+1+80) frame coordinates (4 values), frame confidence (1 value), and object class probability (for the COCO dataset, there are 80 objects). The model constructed in this example was also named RYPY.

(1) Inputting the whole image into an image recognition model RYPY, and extracting features of the input image through a feature extraction network to obtain a feature map with a certain size;

(2) and detecting rape information such as pod, branch length, effective branch length and the like in each sample prescription, and calculating by the following formula to obtain the estimated rape yield value.

According to the rape plant characteristics, taking yield as a dependent variable (Y) and branch length, branch number and effective branch length as independent variables (X), performing multiple linear regression analysis on the measurement data of the rape plant by adopting a stepwise regression method, and establishing an optimal multiple linear regression equation as follows:

Y＝0.811X₁+0.015X₃/X₂(R²0.820) in the above formula: y is the yield (g) of the rape plants; x₁The effective branch is equal length (mm), and the applicable value range is 0<X₁<80；X₂Number of branches (mm); x₃Total number of siliques (mm), X₃/X₂The applicable value range is 0<X₃/X₂<90；R²To measure the correlation coefficient of production.

8) Model validation

Comparing the yield estimation result of the step 7) with the actual manually measured yield of the training test set in the step 5), and verifying the accuracy of the model.

Table 12 shows the comparison between the estimated and the measured values of rape yield in 10 test fields calculated by the model RYPY.

Comparison of the estimated and measured values of rape yield in the same plot in Table 12

Sample plot number	Estimate value (Kg/mu)	Measured value (Kg/mu)
			01	156.54	135.01
02	85.26	112.58
			03	157.73	128.43
04	81.92	104.53
			05	146.63	160.39
06	98.53	65.90
			07	169.63	148.88
08	91.48	117.03
			09	159.93	132.13
10	117.35	87.80

In addition, 0.25m of random sampling will be performed²The yield estimation value of the corresponding image in the area and the yield actual value in the area are counted and subjected to error analysis, as shown in table 13.

TABLE 13 sample yield estimation error analysis

The error calculation formula is as follows:

relative error (%):

mean absolute error:

average relative error (%):

wherein x is_iModel estimates for the ith image;

the artificial measured value of the ith image; m is the number of images.

The calculation shows that the AE and RE of the model are 23.35 Kg/mu and 22.81% respectively in 10 test fields. Therefore, 0.25m in the estimated field²The accuracy of the yield estimation model is 77.19%, which is already high, when the yield of the rape in the region is in the range of (1), and the value shows that the model has higher estimation accuracy.

The above embodiments are merely preferred embodiments of the present invention, which is not intended to limit the present invention in any way. Although the present invention has been described with reference to the preferred embodiments, it is not intended to be limited thereto. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical spirit of the present invention should fall within the protection scope of the technical scheme of the present invention, unless the technical spirit of the present invention departs from the content of the technical scheme of the present invention.

Claims

1. A rape yield estimation method is characterized by comprising the following steps:

1) sampling: shooting a rape canopy image, and selecting sample data of a plurality of sample areas;

2) pretreatment: preprocessing the image;

2. The method for estimating rape yield according to claim 1, wherein the step 1) is to shoot the rape with high, middle and low altitude full coverage repeatedly by using the unmanned aerial vehicle aerial canopy in the rape pod maturation stage.

3. The method of claim 2, wherein in the step 2), the image is processed by one or more of cropping, adjusting brightness, adjusting contrast, enhancing color, rotating and sharpening.

4. The method of claim 2, wherein the image in step 4) is modeled based on a YOLOv3 technique, the whole image is input into an image recognition model RYPY, and features of the input image are extracted through a feature extraction network.

5. The method of claim 2, wherein in step 1), when sampling the rape seed, the canopy image is collected first, and then the measurement information of the rape seed is taken.

6. The method of claim 5, wherein the key performance indicators in step 3) are silique number, kernel number per silique and thousand kernel weight.

7. The method of claim 6, wherein in the step 4), the features of the input image are extracted through a feature extraction network, and the information of the number of pods, the number of branches and the effective branch length of each rape sample is detected.

8. The method of claim 7, wherein in the step 4), according to the characteristics of the rape plants, the yield is taken as a dependent variable Y, the branch length, the branch number and the effective branch length are taken as independent variables X, and the multiple linear regression analysis is performed on the measurement data of the rape plants by adopting a stepwise regression method to establish a multiple linear regression equation to calculate the estimated yield of the rape plants; the regression equation is:

Y＝0.811X₁+0.015X₃/X₂(R²＝0.820)。

wherein Y is the yield of the rape plants; x₁For effective branches to be equally long, the applicable value range is 0<X₁<80；X₂Is the number of branches; x₃Total number of siliques, X₃/X₂The applicable value range is 0<X₃/X₂<90, respectively; coefficient of correlation R²The value was 0.820.