Method for estimating cotton defoliation rate based on the change rate of UAV multispectral vegetation index

Abstract

The present invention discloses a method for estimating the defoliation rate of cotton based on the change rate of the multispectral vegetation index of an unmanned aerial vehicle, and belongs to the technical field of crop growth monitoring. The method for estimating the defoliation rate of cotton based on the change rate of the multispectral vegetation index of an unmanned aerial vehicle comprises: before and after the defoliation treatment, a drone is used to collect multispectral images of the cotton in the test area, the obtained multispectral data is pre-processed to obtain the vegetation reflectivity, and then the vegetation index before the defoliation treatment and the vegetation index after the defoliation treatment are calculated by the vegetation reflectivity before and after the defoliation treatment; the vegetation index change rate of the test area is calculated according to the vegetation index before the defoliation treatment and the vegetation index after the defoliation treatment; the vegetation index change rate of the test area is input into a trained random forest model, and the cotton defoliation rate of the test area is output. According to the method of the present invention, the defoliation rate is inverted by the vegetation index change rate, the change of the defoliation rate can be accurately predicted, and the defoliation process of cotton can be evaluated.

Description

Technical Field

[0001] The present invention relates to the technical field of crop growth monitoring, and more specifically to a method for estimating cotton defoliation rate based on the change rate of unmanned aerial vehicle multispectral vegetation index. Background Art

[0002] Cotton harvesting requires a lot of manual labor. In the context of the increasing shortage of agricultural labor, mechanized harvesting has become an inevitable trend. In actual agricultural production, defoliation and ripening agents are usually sprayed to cause the leaves of cotton plants to fall off as quickly as possible, so as to improve the efficiency of mechanical harvesting and reduce the impurity rate of seed cotton. The defoliation rate is used to evaluate the defoliation process of cotton.

[0003] Drones play an important role in crop monitoring. With the characteristics of high efficiency, high resolution, low cost, low risk and high flexibility, they will become an ideal tool for crop phenotyping. Drone images have recently been applied to crop growth monitoring, pest and disease monitoring, etc. In addition, higher resolution images can be obtained compared with satellite and airborne methods because they are taken at relatively low altitudes. Drone images are of great significance for the practical application of crop phenotyping research.

[0004] Currently, the vegetation index is used directly for monitoring when using drones. The vegetation index is a direct indicator that reflects the state of the canopy. However, the direct inversion using direct indicators lacks physical meaning and it is difficult to accurately obtain the results of defoliation rate. Summary of the invention

[0005] In response to the above problems, the present invention provides a method for estimating the cotton defoliation rate based on the change rate of the multi-spectral vegetation index of an unmanned aerial vehicle. According to the method of the present invention, the defoliation rate can be inverted by the change rate of the vegetation index, and the change of the defoliation rate can be accurately predicted, thereby evaluating the defoliation process of the cotton.

[0006] The object of the present invention is to provide a method for estimating cotton defoliation rate based on the change rate of multispectral vegetation index of unmanned aerial vehicle, comprising the following steps:

[0007] Before defoliation, a drone is used to collect multispectral images of cotton in a test area to obtain first multispectral data; the first multispectral data is preprocessed to obtain a first vegetation reflectance, and then the vegetation index of the test area before defoliation is calculated through the first vegetation reflectance.

[0008] After defoliation, a drone is used to collect multispectral images of the cotton in the test area to obtain second multispectral data; the second multispectral data is preprocessed to obtain a second vegetation reflectance, and then the second vegetation index after defoliation is calculated by the second vegetation reflectance; it should be noted that the defoliation treatment is: based on the planting conditions of the planting test area, the cotton in the test area is defoliated, and the vegetation index is collected 1 day before spraying the agent, and 5 days, 10 days, 15 days, and 20 days after spraying the agent.

[0009] The vegetation index change rate of the measured area is calculated based on the vegetation index before and after defoliation.

[0010] The vegetation index change rate of the tested area is input into the trained random forest model, and the cotton defoliation rate of the tested area is output.

[0011] In a preferred embodiment of the present invention, the training method of the random forest model is:

[0012] Before defoliation, a drone was used to collect multispectral images of cotton in the example area to obtain third multispectral data; the third multispectral data was preprocessed to obtain a third vegetation reflectance, and then the initial vegetation index of the example area before defoliation was calculated through the third vegetation reflectance.

[0013] After defoliation, a drone is used to collect multispectral images of cotton in the example area to obtain fourth multispectral data; the fourth multispectral data is preprocessed to obtain a fourth vegetation reflectance, and then the initial vegetation index of the example area after defoliation is calculated through the fourth vegetation reflectance.

[0014] The change rate of the initial vegetation index of the example area is calculated based on the initial vegetation index of the example area before the defoliation treatment and the initial vegetation index of the example area after the defoliation treatment.

[0015] The initial database was established by taking the initial vegetation index change rate of the sample area corresponding to the initial vegetation index type as the independent variable and the defoliation rate of the sample area as the dependent variable.

[0016] The minimum root mean square error is obtained by the change rate of the initial vegetation index of the initial database, the optimal number of decision trees in the random forest model and the optimal number of variables for the binary tree in the specified node are obtained, and then the trained random forest model is established.

[0017] In a preferred embodiment of the present invention, the method for obtaining the defoliation rate of the example area is: defoliating the cotton in the example area, and obtaining the defoliation rate of the example area by statistics.

[0018] In a preferred embodiment of the present invention, the initial vegetation index change rate of the example area is screened for importance in the random forest model to obtain the optimal vegetation index change rate, and the vegetation index type corresponding to the optimal vegetation index change rate is determined.

[0019] The vegetation index change rate of the test area is calculated by the vegetation index type corresponding to the optimal vegetation index change rate, and the vegetation index change rate of the test area is input into the trained random forest model to output the cotton defoliation rate of the test area.

[0020] In a preferred embodiment of the present invention, the initial vegetation index type is a normalized difference vegetation index, a normalized difference red edge vegetation index, a terrestrial chlorophyll index, a leaf chlorophyll index, a normalized difference vegetation index between green and red bands, a normalized difference red edge index, a structure-insensitive pigment vegetation index, a normalized near-infrared and blue band difference vegetation index, a normalized near-infrared and red band difference vegetation index, a green normalized difference vegetation index, a normalized difference yellowness index, a visible light difference vegetation index, a normalized difference chlorophyll index, Modified chlorophyll absorption reflectance index, triangular vegetation index, converted chlorophyll absorption reflectance index, simple ratio vegetation index, green ratio vegetation index, near-infrared red edge ratio vegetation index, greenness index, red edge green band vegetation index, crop vegetation index, difference vegetation index, near-infrared and green band difference vegetation index, near-infrared and red edge band difference vegetation index, red edge soil adjusted vegetation index, optimized soil adjusted vegetation index, enhanced vegetation index, modified chlorophyll content and crop condition index and normalized square root vegetation index.

[0021] In a preferred embodiment of the present invention, a drone is used to acquire a planting test area for multispectral image acquisition to obtain first multispectral data. The specific operation method is: determine the spatial range of the test area, and the DJI GS Pro software will automatically generate a flight path. The DJI P4M is equipped with a multispectral camera for multispectral image acquisition. The multispectral camera includes 1 color sensor for visible light imaging and 5 monochrome sensors for multispectral imaging. The single sensor has 2.08 million effective pixels and 2.12 million total pixels. The 5 multispectral bands are blue band, green band, red band, red edge band, and near infrared band; the central wavelength of the blue band is 450nm±16nm; the central wavelength of the green band is 560nm±16nm; the central wavelength of the red band is 650nm±16nm; the central wavelength of the red edge band is 730nm±16nm; the central wavelength of the near infrared band is 840nm±26nm.

[0022] The multi-spectral field of view of the drone is 62.7°, the flight altitude is set to 100m, the flight speed is designed to be 5m / s, the heading overlap rate is designed to be 80%, the side overlap rate is designed to be 70%, the drone photography mode is designed to hover at the waypoint, and other flight parameters are default. It should be noted that the blue band is recorded as B, the green band is recorded as G, the red band is recorded as R, the red edge band is recorded as RE, and the near infrared band is recorded as NIR.

[0023] The first multispectral data preprocessing method is:

[0024] The first multispectral data was radiometrically corrected using a UAV radiation whiteboard, and PIX4DMapper was used to puzzle agricultural multispectral images. After puzzlement, band synthesis was performed in a complete remote sensing image processing platform, and then normalization was performed to obtain the first vegetation reflectance.

[0025] The method for preprocessing the first multispectral data, the second multispectral data, the third multispectral data and the fourth multispectral data is the same.

[0026] In a preferred embodiment of the present invention, the method for obtaining the optimal number of decision trees in the random forest model and the optimal number of variables for the binary tree in the specified node is: screening the number of variables for the binary tree in the specified node or the number of decision trees contained in the specified random forest to obtain the optimal number of decision trees contained in the specified random forest and the optimal number of variables for the binary tree in the specified node.

[0027] It should be noted that the optimal number of variables used for binary trees in the specified node is denoted as mtry, and the optimal number of decision trees contained in the specified random forest is denoted as ntree.

[0028] In a preferred embodiment of the present invention, the optimal number of decision trees contained in the random forest is specified to be 1000 trees.

[0029] In a preferred embodiment of the present invention, the optimal number of variables for the binary tree in the specified node is selected as 6 according to the minimum root mean square error.

[0030] In a preferred embodiment of the present invention, the agent used for defoliation treatment is Hinseli.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The vegetation index itself can characterize the physiological and biochemical state of the plant canopy, while the defoliation rate is a relative quantity. Even if a drone is used for direct shooting, leaf shedding data can only be obtained at an extremely low altitude, and only upper leaf shedding data can be obtained, and the obtained leaf shedding data is incomplete. Therefore, the present invention uses drone remote sensing monitoring technology to obtain the vegetation index before and after drug administration, calculates the vegetation index change rate, and uses the vegetation index change rate instead of the vegetation index as the independent variable for inverting the defoliation rate, which is more reasonable. The random forest regression modeling method is used to construct a model, and the defoliation rate is inverted by the vegetation index change rate, which is more practical. The prediction accuracy of cotton defoliation rate is high, because both the vegetation index and the plant physiological and biochemical indicators are direct indicators, and the inversion accuracy is not high.

[0033] UAV remote sensing monitoring technology can replace manual work to quickly monitor the cotton defoliation rate over a large area, providing a reference for dynamic monitoring of cotton defoliation and timely harvesting. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is the random forest importance screening result of the vegetation index change rate and defoliation rate selected when the density is 3000 plants / mu at 25m in the study area.

[0035] Figure 2 This is the random forest prediction result of the vegetation index change rate and defoliation rate selected when the density is 3000 plants / mu at 25m in the study area, where (a) is the training set and (b) is the validation set.

[0036] Figure 3 This is the random forest importance screening result of the vegetation index change rate and defoliation rate selected when the density is 6000 plants / mu at 25m in the study area.

[0037] Figure 4 This is the random forest prediction result of the vegetation index change rate and defoliation rate selected when the density is 6000 plants / mu at 25m in the study area, where (a) is the training set and (b) is the validation set.

[0038] Figure 5 This is the random forest importance screening result of the vegetation index change rate and defoliation rate selected when the comprehensive density is 3000 plants / mu and 6000 plants / mu at 25m in the study area.

[0039] Figure 6 This is the random forest prediction result diagram of the vegetation index change rate and defoliation rate selected when the comprehensive density is 3000 plants / mu and 6000 plants / mu at 25m in the study area, where (a) is the training set and (b) is the validation set. DETAILED DESCRIPTION

[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] Table 1 Calculation formulas for different vegetation indices

[0042]

[0043] The present invention obtains the vegetation index at different days before and after medication, calculates the vegetation index change rate, and calculates the defoliation rate by investigating the number of leaves at different days before and after medication, and then performs random forest variable importance screening, and models are built through the vegetation index change rate and defoliation rate.

[0044] Defoliation rate = (number of leaves before medication - number of leaves on a certain day after medication) / number of leaves before medication.

[0045] Vegetation index change rate = (initial vegetation index of the sample area before defoliation treatment - initial vegetation index of the sample area after defoliation treatment) / initial vegetation index of the sample area before defoliation treatment.

[0046] The vegetation reflectance is calculated according to the formula in Table 1 to obtain the vegetation index.

[0047] The following is an explanation through specific embodiments.

[0048] Example 1

[0049] Step 1: Establishment of trained random forest model

[0050] Step 1.1, defoliate the cotton in the sample area, and obtain the defoliation rate of the sample area by statistics; the details are as follows.

[0051] Based on the planting conditions in the experimental planting area, the cotton was defoliated and the defoliation rate was calculated statistically.

[0052] Each plot in the sample area is 9 meters long, 8 meters wide, and has a row spacing of 1 meter. In order to increase the amount of data and consider the consistency of the cotton plant population status, each plot is divided into two parts.

[0053] The experiment was of split-plot design, with density as the main factor and hinseli concentration as the secondary factor.

[0054] The density is designed to be 3,000 plants per mu.

[0055] When spraying Hinseli, a control group was designed, recorded as ck, i.e. 0g / mu, and an equal amount of clean water was sprayed. Then different concentrations of Hinseli were sprayed in 4 areas, namely: the concentration was 50ml / mu, i.e. 71.5g / mu, recorded as T1; the concentration was 100ml / mu, i.e. 143g / mu, recorded as T2; the concentration was 200ml / mu, i.e. 286g / mu, recorded as T3; the concentration was 300ml / mu, i.e. 429g / mu, recorded as T4. It should be noted that the data obtained in this embodiment was obtained by testing four different concentrations under the condition of a density of 3000 plants / mu. On the one hand, it is to meet the needs of different dosages in actual production, and on the other hand, it is to make the data more reasonable.

[0056] The leaf data of cotton in the sample areas at different periods were collected one day before spraying the pesticide, and 5, 10, 15 and 20 days after spraying the pesticide. The leaf defoliation rate of the sample areas at different days after spraying the pesticide was calculated according to the formula.

[0057] Step 1.2: The method for obtaining the initial vegetation index change rate of the sample area is specifically as follows.

[0058] Before and after defoliation, drones were used to obtain multispectral images of the planting test area to obtain multispectral data; the obtained multispectral data were preprocessed and calculated to obtain the initial vegetation index of the sample area before defoliation and the initial vegetation index of the sample area after defoliation. Specifically, the spatial range of the test area was determined first, and the DJI GS Pro software automatically generated the flight path. The DJI P4M is equipped with a multispectral camera for multispectral image acquisition. The multispectral camera includes 1 color sensor for visible light imaging and 5 monochrome sensors for multispectral imaging. Single sensor: 2.08 million effective pixels, 2.12 million total pixels. The 5 multispectral bands are blue band, green band, red band, red edge band, and near infrared band; the central wavelength of the blue band is 450nm±16nm; the central wavelength of the green band is 560nm±16nm; the central wavelength of the red band is 650nm±16nm; the central wavelength of the red edge band is 730nm±16nm; the central wavelength of the near infrared band is 840nm±26nm.

[0059] The multi-spectral field of view of the drone is 62.7°, the flight altitude is set to 100m, the flight speed is designed to be 5m / s, the heading overlap rate is designed to be 80%, the side overlap rate is designed to be 70%, the drone photography mode is designed to hover at the waypoint, and other flight parameters are default. It should be noted that the blue band is recorded as B, the green band is recorded as G, the red band is recorded as R, the red edge band is recorded as RE, and the near infrared band is recorded as NIR.

[0060] First, the multispectral data is calibrated using a UAV radiation whiteboard, and the agricultural multispectral image is mosaicked using PIX4DMapper software. After mosaicking, band synthesis is performed in a complete remote sensing image processing platform, and then normalization is performed to obtain vegetation reflectance information.

[0061] The vegetation index was calculated by the vegetation reflectivity according to the calculation formula in Table 1 to obtain the vegetation index of cotton canopy in different plots.

[0062] One day before spraying the pesticide, the initial vegetation index of the sample area before defoliation treatment was obtained by collecting, preprocessing and calculating using the above method using an unmanned aerial vehicle.

[0063] 5, 10, 15 and 20 days after spraying the pesticide, the initial vegetation index of the sample area after defoliation treatment was obtained by collecting, preprocessing and calculating using the above method using a drone.

[0064] Step 1.3: Use random forest regression modeling to build a model. The specific steps are as follows:

[0065] Step 1.3.1, using the vegetation index change rate of the sample area as the independent variable and the defoliation rate of the sample area as the dependent variable, establish an initial database, as shown below.

[0066] The relevant data of vegetation index change rate and defoliation rate in the experiment were collected, and the initial database was established with the vegetation index change rate as the independent variable and the defoliation rate as the dependent variable.

[0067] Step 1.3.2: In the random forest model, the vegetation index change rate of the initial database is screened by importance to obtain the final vegetation index change rate, and the vegetation index type corresponding to the optimal vegetation index change rate is determined; the details are as follows.

[0068] In the random forest, the vegetation index change rate of the initial database is ranked by importance, and the number of variables is determined according to the degree of importance attenuation, that is, the importance of variables is ranked according to their importance. The variable importance ratio has a turning point, and the relative size of the variable importance no longer changes sharply. Below the turning point, the importance of the variables is not much different.

[0069] like Figure 1 As shown, the vegetation index change rates finally screened out are NDRE change rate, GI change rate, SIPI change rate, MTCI change rate, CIREG change rate, GRVI1 change rate, NPCI change rate, NDVI change rate, EVI change rate and NLI change rate.

[0070] Step 1.3, obtain the minimum root mean square error through the vegetation index change rate of the initial database, and obtain the optimal number of decision trees in the random forest model and the optimal number of variables for the binary tree in the specified node, as shown below.

[0071] Select 80% of the data in the initial database as the training set, and the remaining 20% ​​of the data as the test set. Input the data of the training set, set mtry or ntree to unknown, obtain the minimum root mean square error, and obtain the corresponding optimal number of decision trees contained in the specified random forest and the optimal number of variables used for binary trees in the specified node. The optimal number of decision trees contained in the specified random forest is 1000 trees, and the optimal number of variables used for binary trees in the specified node is selected as 6 according to the minimum root mean square error. The random forest model is constructed with the determined optimal parameters, which is recorded as the optimal random forest model, that is, the trained model is obtained.

[0072] Step 2: Prediction of cotton defoliation rate in the tested area

[0073] Before defoliation, a drone is used to collect multispectral images of cotton in a test area to obtain first multispectral data; the obtained first multispectral data is preprocessed to obtain a first vegetation reflectance, and then the vegetation index of the test area before defoliation is calculated through the first vegetation reflectance.

[0074] After defoliation, a drone is used to collect multispectral images of cotton in the test area to obtain second multispectral data; the obtained second multispectral data is preprocessed to obtain a second vegetation reflectance, and then the vegetation index of the test area after defoliation is calculated through the second vegetation reflectance.

[0075] The vegetation index change rate of the test area is calculated by the vegetation index type corresponding to the optimal vegetation index change rate. Specifically, the vegetation index change rate of the test area is calculated according to the vegetation index of the test area before defoliation treatment and the vegetation index of the test area after defoliation treatment.

[0076] The vegetation index change rate of the tested area is input into the trained random forest model, and the cotton defoliation rate of the tested area is output.

[0077] Step 3: The accuracy is evaluated by the coefficient of determination, root mean square error, and relative root mean square error, as shown below.

[0078] The coefficient of determination is denoted as R 2 , the root mean square error is recorded as RMSE, and the relative root mean square error is recorded as RRMSE. 2 The calculation formulas for , RMSE and RRMSE are shown below.

[0079] Formula 1;

[0080] Formula 2;

[0081] Formula 3;

[0082] Among them, i represents the sample of actual value, m represents the sample of predicted value, P i represents the model prediction value of the i-th sample vegetation index, O i represents the actual value of the i-th sample vegetation index, O m represents the average value of the actual value of the vegetation index of the mth sample, and N represents the number of samples.

[0083] It should be noted that the training set is used to predict the test area according to the trained random forest model, and the validation set is used to perform statistics on the test area, and the defoliation rate of the validation set is compared with the defoliation rate of the validation set. The accuracy of this method is evaluated by the determination coefficient, root mean square error, and relative root mean square error. The determination coefficient is denoted as R 2 The root mean square error is recorded as RMSE, and the relative root mean square error is recorded as RRMSE.

[0084] from Figure 2 As can be seen from (a) and (b) in the figure, the accuracy of both the training set and the validation set has achieved good results. 2 Reached 0.978, RMSE: 3.575%, RRMSE reached 9.016%. Validation set R 2 Reached 0.910, RMSE: 10.446%, and RRMSE reached 21.047%.

[0085] Example 2

[0086] Step 1: Establishment of trained random forest model

[0087] Step 1.1, defoliate the cotton in the sample area, and obtain the defoliation rate of the sample area by statistics; the details are as follows.

[0088] Based on the planting conditions in the experimental planting area, the cotton was defoliated and the defoliation rate was calculated statistically.

[0089] Each plot in the sample area is 9 meters long, 8 meters wide, and has a row spacing of 1 meter. In order to increase the amount of data and consider the consistency of the cotton plant population status, each plot is divided into two parts.

[0090] The experiment was of split-plot design, with density as the main factor and hinseli concentration as the secondary factor.

[0091] The density is designed to be 6,000 plants per mu.

[0092] When spraying Hinseli, the concentration was designed as a control group, recorded as ck, i.e. 0g / mu, and an equal amount of clean water was sprayed. Then different concentrations of Hinseli were sprayed in 4 areas, namely: the concentration was 50ml / mu, i.e. 71.5g / mu, recorded as T1; the concentration was 100ml / mu, i.e. 143g / mu, recorded as T2; the concentration was 200ml / mu, i.e. 286g / mu, recorded as T3; the concentration was 300ml / mu, i.e. 429g / mu, recorded as T4. It should be noted that the data obtained in this embodiment was obtained by testing four different concentrations under the condition of a density of 6000 plants / mu. On the one hand, it is to meet the needs of different dosages in actual production, and on the other hand, it is to make the data more reasonable.

[0093] The leaf data of cotton in the sample areas at different periods were collected one day before spraying the pesticide, and 5, 10, 15 and 20 days after spraying the pesticide. The leaf defoliation rate of the sample areas at different days after spraying the pesticide was calculated according to the formula.

[0094] Step 1.2: The method for obtaining the initial vegetation index change rate of the sample area is specifically as follows.

[0095] Before and after defoliation, drones were used to obtain multispectral images of the planting test area to obtain multispectral data; the obtained multispectral data were preprocessed and calculated to obtain the initial vegetation index of the sample area before defoliation and the initial vegetation index of the sample area after defoliation. Specifically, the spatial range of the test area was determined first, and the DJI GS Pro software automatically generated the flight path. The DJI P4M is equipped with a multispectral camera for multispectral image acquisition. The multispectral camera includes 1 color sensor for visible light imaging and 5 monochrome sensors for multispectral imaging. Single sensor: 2.08 million effective pixels, 2.12 million total pixels. The 5 multispectral bands are blue band, green band, red band, red edge band, and near infrared band; the central wavelength of the blue band is 450nm±16nm; the central wavelength of the green band is 560nm±16nm; the central wavelength of the red band is 650nm±16nm; the central wavelength of the red edge band is 730nm±16nm; the central wavelength of the near infrared band is 840nm±26nm.

[0096] The multi-spectral field of view of the drone is 62.7°, the flight altitude is set to 100m, the flight speed is designed to be 5m / s, the heading overlap rate is designed to be 80%, the side overlap rate is designed to be 70%, the drone photography mode is designed to hover at the waypoint, and other flight parameters are default. It should be noted that the blue band is recorded as B, the green band is recorded as G, the red band is recorded as R, the red edge band is recorded as RE, and the near infrared band is recorded as NIR.

[0097] First, the multispectral data is calibrated using a UAV radiation whiteboard, and the agricultural multispectral image is mosaicked using PIX4DMapper software. After mosaicking, band synthesis is performed in a complete remote sensing image processing platform, and then normalization is performed to obtain vegetation reflectance information.

[0098] The vegetation index was calculated by the vegetation reflectivity according to the calculation formula in Table 1 to obtain the vegetation index of cotton canopy in different plots.

[0099] One day before spraying the pesticide, the initial vegetation index of the sample area before defoliation treatment was obtained by collecting, preprocessing and calculating using the above method using an unmanned aerial vehicle.

[0100] 5, 10, 15 and 20 days after spraying the pesticide, the initial vegetation index of the sample area after defoliation treatment was obtained by collecting, preprocessing and calculating using the above method using a drone.

[0101] Step 1.3: Use random forest regression modeling to build a model. The specific steps are as follows:

[0102] Step 1.3.1, using the vegetation index change rate of the sample area as the independent variable and the defoliation rate of the sample area as the dependent variable, establish an initial database, as shown below.

[0103] The relevant data of vegetation index change rate and defoliation rate in the experiment were collected, and the initial database was established with the vegetation index change rate as the independent variable and the defoliation rate as the dependent variable.

[0104] Step 1.3.2: In the random forest model, the vegetation index change rate of the initial database is screened by importance to obtain the final vegetation index change rate, and the vegetation index type corresponding to the optimal vegetation index change rate is determined; the details are as follows.

[0105] In the random forest, the vegetation index change rate of the initial database is ranked by importance, and the number of variables is determined according to the degree of importance attenuation, that is, the importance of variables is ranked according to their importance. The importance ratio of variables has a turning point, and the importance of variables below the turning point is not much different.

[0106] like Figure 3 As shown, the vegetation index change rates finally screened out are MTCI change rate, CIREG change rate, NDRE change rate, SIPI change rate, NLI change rate, RESAVI change rate, OSAVI change rate, NDYI change rate, DVIREG change rate and RDVI change rate.

[0107] Step 1.3, obtain the minimum root mean square error through the vegetation index change rate of the initial database, and obtain the optimal number of decision trees in the random forest model and the optimal number of variables for the binary tree in the specified node, as shown below.

[0108] Select 80% of the data in the initial database as the training set, and the remaining 20% ​​of the data as the test set. Input the data of the training set, set mtry or ntree to unknown, obtain the minimum root mean square error, and obtain the corresponding optimal number of decision trees contained in the specified random forest and the optimal number of variables used for binary trees in the specified node. The optimal number of decision trees contained in the specified random forest is 1000 trees, and the optimal number of variables used for binary trees in the specified node is selected as 6 according to the minimum root mean square error. The random forest model is constructed with the determined optimal parameters, which is recorded as the optimal random forest model, that is, the trained model is obtained.

[0109] Step 2: Prediction of cotton defoliation rate in the tested area

[0110] Before defoliation, a drone is used to collect multispectral images of cotton in a test area to obtain first multispectral data; the obtained first multispectral data is preprocessed to obtain a first vegetation reflectance, and then the vegetation index of the test area before defoliation is calculated through the first vegetation reflectance.

[0111] After defoliation, a drone is used to collect multispectral images of cotton in the test area to obtain second multispectral data; the obtained second multispectral data is preprocessed to obtain a second vegetation reflectance, and then the vegetation index of the test area after defoliation is calculated through the second vegetation reflectance.

[0112] The vegetation index change rate of the test area is calculated by the vegetation index type corresponding to the optimal vegetation index change rate. Specifically, the vegetation index change rate of the test area is calculated according to the vegetation index of the test area before defoliation treatment and the vegetation index of the test area after defoliation treatment.

[0113] The vegetation index change rate of the tested area is input into the trained random forest model, and the cotton defoliation rate of the tested area is output.

[0114] Step 3: The accuracy is evaluated by the coefficient of determination, root mean square error, and relative root mean square error, as shown below.

[0115] It should be noted that the training set is used to predict the test area according to the trained random forest model, and the validation set is used to perform statistics on the test area, and the defoliation rate of the validation set is compared with the defoliation rate of the validation set. The accuracy of this method is evaluated by the determination coefficient, root mean square error, and relative root mean square error. The determination coefficient is denoted as R 2 The root mean square error is recorded as RMSE, and the relative root mean square error is recorded as RRMSE.

[0116] from Figure 4As can be seen from (a) and (b) in the figure, the accuracy of both the training set and the validation set has achieved good results. 2 Reached 0.980, RMSE: 3.454%, RRMSE reached 8.837%. Validation set R 2 Reached 0.899, RMSE: 7.530%, and RRMSE reached 20.192%.

[0117] Example 3

[0118] Step 1: Establishment of trained random forest model

[0119] Step 1.1, defoliate the cotton in the sample area, and obtain the defoliation rate of the sample area by statistics; the details are as follows.

[0120] Based on the planting conditions in the planting test area, the cotton was defoliated and the defoliation rate was calculated statistically, as described below.

[0121] Each plot in the sample area is 9 meters long, 8 meters wide, and has a row spacing of 1 meter. In order to increase the amount of data and consider the consistency of the cotton plant population status, each plot is divided into two parts.

[0122] The experiment was of split-plot design, with density as the main factor and hinseli concentration as the secondary factor.

[0123] The density is designed to be 3,000 plants / mu and 6,000 plants / mu.

[0124] When spraying Hinseli, a control group was designed, recorded as ck, i.e. 0g / mu, and an equal amount of clean water was sprayed. Then different concentrations of Hinseli were sprayed in 4 areas, namely: the concentration was 50ml / mu, i.e. 71.5g / mu, recorded as T1; the concentration was 100ml / mu, i.e. 143g / mu, recorded as T2; the concentration was 200ml / mu, i.e. 286g / mu, recorded as T3; the concentration was 300ml / mu, i.e. 429g / mu, recorded as T4. It should be noted that the data obtained in this embodiment was obtained by testing four different concentrations under the conditions of density of 3000 plants / mu and 6000 plants / mu. On the one hand, it is to meet the needs of different dosages in actual production, and on the other hand, it is to make the data more reasonable.

[0125] The leaf data of cotton in the sample areas at different periods were collected one day before spraying the pesticide, and 5, 10, 15 and 20 days after spraying the pesticide. The leaf defoliation rate of the sample areas at different days after spraying the pesticide was calculated according to the formula.

[0126] Step 1.2: The method for obtaining the initial vegetation index change rate of the sample area is specifically as follows.

[0127] Before and after defoliation, drones were used to obtain multispectral images of the planting test area to obtain multispectral data; the obtained multispectral data were preprocessed and calculated to obtain the initial vegetation index of the sample area before defoliation and the initial vegetation index of the sample area after defoliation. Specifically, the spatial range of the test area was determined first, and the DJI GS Pro software automatically generated the flight path. DJI P4M is equipped with a multispectral camera for multispectral image acquisition. The multispectral camera includes 1 color sensor for visible light imaging and 5 monochrome sensors for multispectral imaging. Single sensor: 2.08 million effective pixels and 2.12 million total pixels. The 5 multispectral bands are blue band, green band, red band, red edge band, and near infrared band; the central wavelength of the blue band is 450nm±16nm; the central wavelength of the green band is 560nm±16nm; the central wavelength of the red band is 650nm±16nm; the central wavelength of the red edge band is 730nm±16nm; the central wavelength of the near infrared band is 840nm±26nm.

[0128] The multi-spectral field of view of the drone is 62.7°, the flight altitude is set to 100m, the flight speed is designed to be 5m / s, the heading overlap rate is designed to be 80%, the side overlap rate is designed to be 70%, the drone photography mode is designed to hover at the waypoint, and other flight parameters are default. It should be noted that the blue band is recorded as B, the green band is recorded as G, the red band is recorded as R, the red edge band is recorded as RE, and the near infrared band is recorded as NIR.

[0129] First, the multispectral data is calibrated using a UAV radiation whiteboard, and the agricultural multispectral image is mosaicked using PIX4DMapper software. After mosaicking, band synthesis is performed in a complete remote sensing image processing platform, and then normalization is performed to obtain vegetation reflectance information.

[0130] The vegetation index was calculated by the vegetation reflectivity according to the calculation formula in Table 1 to obtain the vegetation index of cotton canopy in different plots.

[0131] One day before spraying the pesticide, the initial vegetation index of the sample area before defoliation treatment was obtained by collecting, preprocessing and calculating using the above method using an unmanned aerial vehicle.

[0132] 5, 10, 15 and 20 days after spraying the pesticide, the initial vegetation index of the sample area after defoliation treatment was obtained by collecting, preprocessing and calculating using the above method using a drone.

[0133] Step 1.3: Use random forest regression modeling to build a model. The specific steps are as follows:

[0134] Step 1.3.1, using the vegetation index change rate of the sample area as the independent variable and the defoliation rate of the sample area as the dependent variable, establish an initial database, as shown below.

[0135] The relevant data of vegetation index change rate and defoliation rate in the experiment were collected, and the initial database was established with the vegetation index change rate as the independent variable and the defoliation rate as the dependent variable.

[0136] Step 1.3.2: In the random forest model, the vegetation index change rate of the initial database is screened by importance to obtain the final vegetation index change rate, and the vegetation index type corresponding to the optimal vegetation index change rate is determined; the details are as follows.

[0137] In the random forest, the vegetation index change rate of the initial database is ranked by importance, and the number of variables is determined according to the degree of importance attenuation, that is, the importance of variables is ranked according to their importance. The importance of variables has a turning point, and the importance of variables below the turning point is not much different.

[0138] like Figure 5 As shown, the vegetation index change rates finally screened out are SIPI change rate, RDVI change rate, EVI change rate, NLI change rate, OSAVI change rate, MCARI1 change rate, DVIREG change rate, RESAVI change rate, NPCI change rate and NDVI change rate.

[0139] Step 1.3, obtain the minimum root mean square error through the vegetation index change rate of the initial database, and obtain the optimal number of decision trees in the random forest model and the optimal number of variables for the binary tree in the specified node, as shown below.

[0140] Select 80% of the data in the initial database as the training set, and the remaining 20% ​​of the data as the test set. Input the data of the training set, set mtry or ntree to unknown, obtain the minimum root mean square error, and obtain the corresponding optimal number of decision trees contained in the specified random forest and the optimal number of variables used for binary trees in the specified node. The optimal number of decision trees contained in the specified random forest is 1000 trees, and the optimal number of variables used for binary trees in the specified node is selected as 6 according to the minimum root mean square error. The random forest model is constructed with the determined optimal parameters, which is recorded as the optimal random forest model, that is, the trained model is obtained.

[0141] Step 2: Prediction of cotton defoliation rate in the tested area

[0142] Before defoliation, a drone is used to collect multispectral images of cotton in a test area to obtain first multispectral data; the obtained first multispectral data is preprocessed to obtain a first vegetation reflectance, and then the vegetation index of the test area before defoliation is calculated through the first vegetation reflectance.

[0143] After defoliation, a drone is used to collect multispectral images of cotton in the test area to obtain second multispectral data; the obtained second multispectral data is preprocessed to obtain a second vegetation reflectance, and then the vegetation index of the test area after defoliation is calculated through the second vegetation reflectance.

[0144] The vegetation index change rate of the test area is calculated by the vegetation index type corresponding to the optimal vegetation index change rate. Specifically, the vegetation index change rate of the test area is calculated according to the vegetation index of the test area before defoliation treatment and the vegetation index of the test area after defoliation treatment.

[0145] The vegetation index change rate of the tested area is input into the trained random forest model, and the cotton defoliation rate of the tested area is output.

[0146] Step 3: The accuracy is evaluated by the coefficient of determination, root mean square error and relative root mean square error, as shown below.

[0147] It should be noted that the training set is used to predict the test area according to the trained random forest model, and the validation set is used to perform statistics on the test area, and the predicted defoliation rate of the validation set is compared with the actual defoliation rate of the validation set. The accuracy of this method is evaluated by the determination coefficient, root mean square error, and relative root mean square error. The determination coefficient is denoted as R 2 , the root mean square error is recorded as RMSE, and the relative root mean square error is recorded as RRMSE.

[0148] from Figure 6 As can be seen from (a) and (b) in the figure, the accuracy of both the training set and the validation set has achieved good results. Considering the density of 3000 plants / mu and the density of 6000 plants / mu, the training set R 2 The validation set R 2 It reaches 0.902, the RMSE is 7.466%, and the RRMSE reaches 18.085%.

[0149] In summary, it can be seen that whether it is a single analysis of 3,000 plants / mu or 6,000 plants / mu, or a comprehensive analysis of 3,000 plants / mu and 6,000 plants / mu, the results of modeling based on the random forest model using the vegetation index change rate and defoliation rate are very stable and relatively accurate.

[0150] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0151] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index, characterized in that: include: Before defoliation, a drone is used to collect multispectral images of cotton in the test area to obtain first multispectral data; Preprocessing the first multispectral data to obtain a first vegetation reflectance, and then calculating a vegetation index of the test area before defoliation treatment through the first vegetation reflectance; After defoliation, a drone is used to collect multispectral images of cotton in the test area to obtain second multispectral data; the second multispectral data is preprocessed to obtain a second vegetation reflectance, and then the vegetation index of the test area after defoliation is calculated by the second vegetation reflectance; The vegetation index change rate of the tested area is calculated according to the vegetation index of the tested area before defoliation treatment and the vegetation index of the tested area after defoliation treatment; The vegetation index change rate of the tested area is input into the trained random forest model, and the cotton defoliation rate of the tested area is output; The training method of the random forest model is: Before defoliation, a drone is used to collect multispectral images of cotton in the sample area to obtain third multispectral data; the third multispectral data is preprocessed to obtain a third vegetation reflectance, and then the initial vegetation index of the sample area before defoliation is calculated by the third vegetation reflectance; After defoliation, a drone is used to collect multispectral images of cotton in the sample area to obtain fourth multispectral data; the fourth multispectral data is preprocessed to obtain a fourth vegetation reflectance, and then the initial vegetation index of the sample area after defoliation is calculated by the fourth vegetation reflectance; The change rate of the initial vegetation index of the sample area is calculated according to the initial vegetation index of the sample area before defoliation treatment and the initial vegetation index of the sample area after defoliation treatment; The initial database is established by taking the initial vegetation index change rate of the sample area corresponding to the initial vegetation index type as the independent variable and the defoliation rate of the sample area as the dependent variable; The minimum root mean square error is obtained by the change rate of the initial vegetation index of the initial database, the optimal number of decision trees in the random forest model and the optimal number of variables for the binary tree in the specified node are obtained, and then the trained random forest model is established.

2. The method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index according to claim 1 is characterized in that: The method for obtaining the defoliation rate of the example area is: defoliating the cotton in the example area, and obtaining the defoliation rate of the example area by statistics.

3. The method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index according to claim 1, characterized in that: In the random forest model, the initial vegetation index change rate of the sample area is screened by importance to obtain the best vegetation index change rate, and the vegetation index type corresponding to the best vegetation index change rate is determined; The vegetation index change rate of the test area is calculated by the vegetation index type corresponding to the optimal vegetation index change rate, and the vegetation index change rate of the test area is input into the trained random forest model to output the cotton defoliation rate of the test area.

4. The method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index according to claim 1, characterized in that: The initial vegetation index types are normalized difference vegetation index, normalized difference red edge vegetation index, terrestrial chlorophyll index, leaf chlorophyll index, normalized difference vegetation index between green band and red band, normalized difference red edge index, structure-insensitive pigment vegetation index, normalized near-infrared and blue band difference vegetation index, normalized near-infrared and red band difference vegetation index, green normalized difference vegetation index, normalized difference yellowness index, visible light difference vegetation index, normalized difference chlorophyll index, modified chlorophyll absorption index, and normalized difference vegetation index. Reflectance index, triangular vegetation index, conversion chlorophyll absorption reflectance index, simple ratio vegetation index, green ratio vegetation index, near-infrared red edge ratio vegetation index, greenness index, red edge green band vegetation index, crop vegetation index, difference vegetation index, near-infrared and green band difference vegetation index, near-infrared and red edge band difference vegetation index, red edge soil adjusted vegetation index, optimized soil adjusted vegetation index, enhanced vegetation index, corrected chlorophyll content and crop condition index and normalized square root vegetation index.

5. The method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index according to claim 1, characterized in that: The first multispectral data preprocessing method is: The first multispectral data was radiometrically corrected using a UAV radiation whiteboard, and agricultural multispectral images were mosaicked using PIX4DMapper. After mosaicking, band synthesis was performed in a complete remote sensing image processing platform. Then, normalization was performed after mosaicking to obtain the first vegetation reflectance. The method for preprocessing the first multispectral data, the second multispectral data, the third multispectral data and the fourth multispectral data is the same.

6. The method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index according to claim 1, characterized in that: The optimal number of decision trees in the random forest model and the optimal number of variables for the binary tree in a given node are obtained as follows: 80% of the data in the initial database are selected as the training set, and the remaining 20% ​​of the data are selected as the test set. The number of variables used for the binary tree in the specified node or the number of decision trees contained in the specified random forest is screened to obtain the optimal number of decision trees contained in the specified random forest and the optimal number of variables used for the binary tree in the specified node.

7. The method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index according to claim 6, characterized in that: The optimal number of decision trees in the random forest is specified to be 1000 trees.

8. The method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index according to claim 6, characterized in that: The optimal number of variables used in the binary tree in the specified node is selected as 6 according to the minimum root mean square error.

9. The method for estimating cotton defoliation rate based on the change rate of drone multispectral vegetation index according to claim 1, characterized in that: The agent used for defoliation treatment is Hinseli.

Images