Unmanned aerial vehicle multi-source quantitative remote sensing rape growth state evaluation method

Abstract

According to the unmanned aerial vehicle multi-source quantitative remote sensing rape growth state evaluation method, rape is divided into a seedling stage, a flowering stage and a legume stage according to growth nodes, ground hyperspectral and multi-spectral data are synchronously obtained, resolution is increased and decreased, the data are converted into all-dimensional, multi-scale and multi-temporal data, a growth state factor inversion model of the rape is established, and the rape growth state is evaluated. Evaluating the contribution of different growth state factors to the growth vigor of the rape, and evaluating the growth state of the rape based on comprehensive evaluation indexes of leaf area index, overground biomass and chlorophyll content. The method comprises the following steps: establishing a model based on multi-layer linear regression and partial least square optimization models of three types of growth factors LAI, AGB and CC, multi-mode prediction based on entropy aggregation and hierarchical analysis of entropy loading expert knowledge in combination with expert knowledge and agronomic knowledge, and setting a model with an optimal verification index in different growth periods as a growth state evaluation model in the period. The rape growth state evaluation period is short, the pertinence is strong, and the accuracy is good.

Description

Technical Field

[0001] The present application relates to an unmanned aerial vehicle (UAV) remote sensing agricultural assessment method, and in particular to an unmanned aerial vehicle (UAV) multi-source quantitative remote sensing rapeseed growth status assessment method, belonging to the field of agricultural remote sensing technology. Background Art

[0002] Precision agriculture is the direction of agricultural modernization, and the use of remote sensing to digitally obtain crop phenotypic information provides new technical support for the production and management of precision agriculture. At present, the evaluation and monitoring of crop growth status has developed from a simple inversion of the NDVI index to a comprehensive model of satellite-aircraft-ground integrated collaborative observation. Different remote sensing platforms provide surface information of various scales, laying a solid foundation for large-scale quantitative evaluation of crop growth status. Rapeseed is an important oil crop, and its unique canopy structure during the flowering period is very characteristic. The long flowering period of rapeseed and the rapeseed flowers covering the entire canopy pose a great challenge to remote sensing observation.

[0003] With its excellent economic efficiency, drones have both high spatial resolution and temporal resolution, and can obtain high-resolution spectral images over a wide range. The convenient data acquisition method can obtain time series spectral images of vegetation in different growth cycles, which is of great significance for studying the growth status of vegetation in different growth periods.

[0004] Currently, remote sensing for crop growth status evaluation has gradually become a hot topic. For crop growth status evaluation, its importance lies in being able to describe the individual characteristics and group characteristics of crops at different times and the health status of crops. Crop growth status evaluation can also provide timely management information for large farms and large-scale cluster agriculture, and also provide a basis for agricultural producers and futures trading to estimate production.

[0005] The problems that need to be solved by the existing remote sensing rapeseed growth status assessment method and the key technical difficulties of this application include:

[0006] (1) The existing technology lacks a quantitative remote sensing method for evaluating the growth status of rapeseed. It is unable to meet the economical requirements while taking into account high spatial resolution and temporal resolution. The remote sensing images obtained are either too long or the resolution does not meet the requirements. In addition, the rapeseed flowering period is long and the rapeseed flowers cover the entire canopy, making remote sensing observation very difficult. The existing technology cannot accurately and quickly evaluate the growth status of rapeseed. Rapeseed is not divided into seedling stage, flowering stage and pod stage according to growth nodes based on quantitative remote sensing. It is impossible to simultaneously obtain ground hyperspectral and multispectral data, analyze rapeseed physiological and biochemical data and multispectral images of drones, and obtain all-round, multi-scale, and multi-temporal data. The inversion model of rapeseed growth status factors has not been established, and the contribution of different growth status factors to rapeseed growth cannot be evaluated, which cannot provide a quantitative decision-making basis for the management and production of precision agriculture.

[0007] (2) The prior art does not establish a rapeseed leaf area coefficient estimation model based on the different growth stages of rapeseed on the ground, and cannot scale up the empirical model established and verified based on the ground data to the drone, and the accuracy of the multispectral images obtained by the drone cannot meet the requirements. The prior art lacks the use of the measured aboveground dry matter mass and chlorophyll content in different growth stages and the hyperspectral data collected in the field to establish an estimation model for stem dry matter and leaf chlorophyll content. There is a lack of using the flower coverage rate during the flowering period to establish an estimation model for the fresh weight and dry weight of rapeseed fruit during the pod period. For the deviation between the model estimation results caused by drones of different scales, there is a lack of analysis of the causes of the errors, and it is impossible to analyze the relationship between local variance and error from the perspective of spatial heterogeneity and establish a regression model, and it is impossible to use the regression model to reduce the error caused by scale changes. There is a lack of a set of rapeseed growth status assessment models for different growth stages using the rapeseed growth status parameters at different growth stages as the driving factors of the model, and it is impossible to assess the growth status of rapeseed. The assessment of rapeseed growth status is poor in accuracy and has a long cycle.

[0008] (3) The existing technology does not use hyperspectral data as an intermediary to qualitatively and quantitatively analyze the applicability and inversion accuracy of the inversion model between remote sensing platforms of different scales. It has not established a regression model for the inversion error caused by the local variance of the image and the up-and-down scale, and cannot eliminate the inversion error caused by surface heterogeneity. The existing technology lacks the combination of the manually evaluated rapeseed growth status level with the leaf area index, aboveground biomass and chlorophyll content measured on the ground for elastic classification. It is unable to establish a model based on the multi-layer linear regression of three types of growth factors, LAI, AGB, and CC, partial least squares optimization model, and multi-mode prediction based on entropy aggregation and combined with expert knowledge and agronomic knowledge entropy value loading expert knowledge hierarchical analysis. There is a lack of models that set the best verification indicators in different growth periods as growth status evaluation models for that period. There is a lack of period LAI estimation models, poor generalization and applicability, weak model explanatory power, low confidence, and low evaluation accuracy. Summary of the invention

[0009] The application divides rapeseed into seedling stage, flowering stage and pod stage according to growth nodes, simultaneously acquires ground hyperspectral and multispectral data, analyzes rapeseed physiological and biochemical data and multispectral images of drones, performs resolution adjustment and conversion into all-round, multi-scale and multi-temporal data, establishes a growth state factor inversion model for rapeseed, evaluates the contribution of different growth state factors to the growth of rapeseed, evaluates the growth state of rapeseed based on comprehensive evaluation indicators of leaf area index, aboveground biomass and chlorophyll content, adopts elastic evaluation indicators with regional characteristics to conduct relative evaluation of rapeseed in the survey area, establishes a model based on multi-layer linear regression of three types of growth factors, LAI, AGB and CC, partial least squares optimization model, multi-mode prediction based on entropy aggregation and hierarchical analysis of expert knowledge combined with entropy value loading of expert knowledge and agronomic knowledge, sets the model with the best verification index in different growth periods as the growth state evaluation model for that period, and the rapeseed growth state evaluation cycle is short, highly targeted and accurate.

[0010] In order to achieve the above technical effects, the technical solutions adopted in this application are as follows:

[0011] The method of evaluating the growth status of rapeseed using multi-source quantitative remote sensing by unmanned aerial vehicles divides rapeseed into seedling stage, flowering stage and pod stage according to growth nodes, simultaneously acquires ground hyperspectral and multispectral data, analyzes rapeseed physiological and biochemical data and multispectral images of unmanned aerial vehicles, performs resolution adjustment and conversion, and converts them into all-round, multi-scale and multi-temporal data. The growth status factor inversion model of rapeseed is established to evaluate the contribution of different growth status factors to the growth of rapeseed. The model with the best verification index in different growth stages is the growth status evaluation model for that period.

[0012] 1) The SKR sensor was used to obtain long-term multispectral data of the entire growth cycle of rapeseed, and the changing trend of the rapeseed canopy spectrum during the growth period was analyzed. The LAI data of the LAI-2200C canopy analyzer was combined to establish a LAI estimation model for different periods. The high-spectral resolution spectral reflectance information of the rapeseed canopy measured by the FieldSpec 4 ground feature spectrometer was calculated according to the spectral response function of the SKR sensor and the MCA multispectral camera, and the common features of the two sensors for modeling were determined.

[0013] 2) Calculate the local variance of MCA multispectral images with different resolutions to obtain the optimal resolution. Calculate and determine the factors that cause the scale effect as model nonlinearity, model driving factor nonlinearity and surface heterogeneity. Use Taylor series expansion to establish a rapeseed growth factor upscaling error model. Calculate the correlation between the inversion error caused by the scale effect and the local variance of the image. Establish a regression equation between the local variance and the error to correct the inversion results after upscaling and downscaling.

[0014] 3) The growth status of rapeseed was evaluated based on the comprehensive evaluation indicators of leaf area index, aboveground biomass and chlorophyll content. The elastic evaluation index with regional characteristics was used to make a relative evaluation of the rapeseed in the survey area. A model was established based on the multilayer linear regression of three types of growth factors, LAI, AGB and CC, the partial least squares optimization model, and the multimode prediction based on entropy aggregation, combined with the entropy value of expert knowledge and agronomic knowledge. The weights assigned to different growth status parameters represent the contribution of the growth status parameters to the evaluation of the growth status of rapeseed in this period. Multilayer linear regression of three types of growth factors was used for growth evaluation in the seedling stage, the partial least squares optimization model was used as the growth evaluation model in the flowering stage, and the multimode prediction based on entropy aggregation was set as the growth evaluation model in the pod stage.

[0015] Preferably, the best spatial resolution for rapeseed assessment: the spatiotemporal growth aggregation method is used to scale the leaf area coefficient images of the drone at different periods to obtain images with different resolutions of 1cm, 2cm, 4cm, 8cm, 16cm and 32cm, and the corresponding local variance images are obtained;

[0016] A 3×3 calculation template was used to calculate local variance images of different resolutions in three periods. When the spatial resolution was 4 cm in the seedling stage, the maximum local variance was 0.209, and the appropriate resolution for UAV observation in the seedling stage was 4 cm. During the flowering period, the flowers were distributed at the top of the rape canopy and were small in size. The maximum local mean square error was 0.339 when the resolution was 1 cm. In the pod stage, the pods were the observation objects and targets of the canopy, and the maximum local variance was 0.313 at a resolution of 2 cm.

[0017] Preferably, an upscaling error model of rapeseed growth factors is established based on Taylor series expansion: an upscaling model for a 50-meter UAV is established, and an upscaling error model of rapeseed growth factors is established based on Taylor series expansion;

[0018] Step 1: Spatiotemporal growth aggregation of rapeseed growth factors, the method is as follows:

[0019]

[0020] D 反射 =d 1反射 +d 2反射 +…+d n反射 Formula 2

[0021] D 入射 =d 1入射 +d 2入射 +…+d n入射 Formula 3

[0022]

[0023] Approximately considering the incident energy of each small pixel to be the same, we get:

[0024]

[0025] d i入射 is the incident energy received by the initial pixel, d i反射 is the energy reflected from the initial pixel to the sensor, G 升尺度 is the reflectivity of the pixel after spatiotemporal growth aggregation, and n is the total number of aggregated pixels;

[0026] Step 2: Based on the nearest canopy pixel clustering, assign the neighboring pixel value closest to the pixel to be found to the pixel to be found:

[0027] I(P)=I(N) Formula 7

[0028] x N =INT(x+0.5) Formula 8

[0029] The coordinates of the point P to be found are (x, y), INT is the rounding function, and N is the neighboring pixel;

[0030] Assume that the leaf area coefficients before and after scaling are LAI N and LAI L ,The LAI values ​​obtained by MCA after spatiotemporal growth aggregation combined with the ground estimation model inversion have a certain degree of underestimation and overestimation, and the spatial heterogeneity within the large-scale pixels causes upscaling errors;

[0031] Step 3: Definition Also define error = LAI N -LAI L :

[0032]

[0033] in, n is the number of pixels in the initial image corresponding to one pixel after upscaling, and Taylor's formula is used to calculate In CI green i =mCI green Expand at

[0034]

[0035] The ground LAI estimation model is a linear model with a second-order derivative of 0. is relatively The higher-order infinitesimal of is used as the error term;

[0036]

[0037] in It is the local variance of n pixels on the image. It can be deduced from the error term that the error is related to the local variance of the image. The second-order derivative of the ground estimation linear model is 0. A regression model is established by combining the absolute error with the local variance of the image. The errors caused by spatial heterogeneity of the images are eliminated after scaling the drone images of different scales.

[0038] Preferably, the scale transformation error is corrected based on the local variance: the local variance of the image reflects the spatial heterogeneity, and the EVI is determined based on the Taylor formula expansion 2 and CI green The correlation between the upscaling error and the local variance is calculated, and the correlation between the local variance and the error in the three periods is calculated;

[0039] About EVI 2 and CI green The local variance of the image is calculated, an odd number template is used, a 3×3 calculation template is set, random sampling points are set, the error value and the local variance of the sample area corresponding to the error value are calculated, and the error and local variance are used as variables for regression modeling. The empirical model of using local variance to estimate the error in three periods is obtained.

[0040] Preferably, the growth elasticity evaluation factor is constructed: all the leaf area index LAI, aboveground biomass AGB and chlorophyll content CC obtained by inversion are subjected to data standardization, the method is to use the maximum and minimum values ​​as the boundary, divide the difference by 10 as the level interval, and divide the inversion parameters into 1 to 10 levels:

[0041]

[0042] GS LAI is the LAI classification segment, INT is the rounding function, LAI estimate The leaf area index is obtained by UAV inversion;

[0043]

[0044] GS AGB is the AGB classification segment, INT is the rounding function, AGB estimate The aboveground biomass was obtained by UAV inversion;

[0045]

[0046] GS CC is the CC grading segment, INT is the rounding function, CC estimate The chlorophyll content is obtained by inversion using a drone.

[0047] Preferably, multi-layer linear regression of three types of growth factors: elastic evaluation of LAI, AGB, and CC measured is performed, and their levels are used as independent variables, and multi-layer linear regression is performed with the growth status level assessed by experts as the dependent variable;

[0048] Multilayer linear regression was used to obtain the coefficients of each parameter. The leaf area index level made an important contribution to the regression of the growth status level. Chlorophyll was used as a reference indicator when determining the regression. The contribution of aboveground biomass to the growth status rating during the pod stage was greater than that in the previous two periods. The model was verified for accuracy and T-tested. Neither AGB nor CC passed the significance test at the seedling stage, AGB at the flowering stage, and CC at the pod stage.

[0049] Preferably, partial least squares optimization model: the professional partial least squares analysis software SmartPLS 3.2 is used to build the growth status evaluation model. In three growth periods, 32 sample points are randomly set for modeling, and 16 sample points are used for model verification.

[0050] Preferably, multi-mode prediction based on entropy aggregation: Assume that for a certain indicator X of the same prediction object t (t=1,2,…,N), there are m single prediction methods to predict it, among which the prediction value of the i-th single prediction method at time t is X it (i=1,2,…,m;t=1,2,…,N), through K 1 ,K 2 ,…,K m The weight coefficients are used to weight the m individual prediction methods, where:

[0051] K 1 +K 2 +K 3 +…K m =1 Formula 20

[0052] The combined prediction X index is:

[0053] X t =K 1 ×X 1t +K 2 ×X 2t +K 3 ×X 3t +…+K m ×X mt Formula 21

[0054] The information entropy method is used to obtain the weight coefficient. Information entropy specifically reflects the variability of a certain indicator of the system and measures the amount of information. The greater the amount of information contained in a certain indicator of the system, the greater the impact of this indicator on the final decision. At this time, the greater the entropy value, the greater the weight should be given.

[0055] Preferably, the overall multi-mode prediction algorithm based on entropy aggregation is as follows:

[0056] Step 1: Normalize the prediction error:

[0057] The proportion of the relative error of the forecast is calculated by calculating the relative error between the forecast result of the i-th single forecast at time t and the actual value:

[0058] e it =|(X t -X it ) / X t Formula 22

[0059] where e it is the prediction relative error of the i-th prediction method at the t-th time, and the proportion of the prediction relative error of the i-th single prediction method at the t-th time is described as:

[0060]

[0061] Step 2, calculate the entropy value:

[0062] The entropy value of the relative error of the prediction of the i-th single prediction method is:

[0063]

[0064] Where k is a constant greater than zero, k = 1 / InN, h i The value range of is 0≤h≤1;

[0065] Step 3, calculate the coefficient of variation:

[0066] The coefficient of variation of the relative error of the prediction of the i-th single prediction method is:

[0067] d i =1-h i Formula 26

[0068] Step 4, calculate the entropy weight:

[0069] The entropy weight of the i-th single prediction method is:

[0070]

[0071] Step 5, calculate the weight coefficient:

[0072]

[0073] For the three growth periods, linear regression models were established with leaf area index level, aboveground biomass level and chlorophyll content level as independent variables and growth level as dependent variable. The linear model was established to prevent the scale effect caused by the nonlinearity of the model when the scale changed, which would cause the model's inadaptability and uncontrollable errors.

[0074] Preferably, entropy values ​​load the hierarchical analysis of expert knowledge:

[0075] Step 1: Establish a rapeseed evaluation structural model: First, after analyzing the rapeseed evaluation, the relevant factors are divided into several levels from top to bottom according to different attributes. Among the many factors in the same layer, the factors in the upper layer or have an influence on the factors in the upper layer and at the same time dominate the factors in the lower layer or are affected by the factors in the lower layer. The factors in the same layer are independent of each other. The top is the target layer with only one element, the bottom is the object layer, and there are multiple levels in between.

[0076] Step 2: Construct the growth judgment matrix of each level: Starting from the second level of the hierarchical structure model, use the paired comparison method for the factors of the same level that belong to each factor of the previous level, with a comparison scale from 1 to 9, to construct a comparison matrix until the bottom level;

[0077] Step 3: Evaluate the consistency calculation test:

[0078] Calculate the evaluation weight vector and perform consistency test. Compare the matrix to calculate the maximum eigenvalue root and the maximum eigenvector. Perform consistency test based on the consistency index CI and the random consistency index RI and the ratio of the two, that is, the consistency ratio. If it is less than 0.1, the consistency test passes. Then, normalize the eigenvalues ​​to get the weight vector. If the test fails, reconstruct the comparison matrix.

[0079] According to the scores given by invited experts and combined with agricultural knowledge, the obvious errors in the scores were eliminated and the weight of the matrix was judged by the geometric mean method. Three judgment matrices were established for the different growth periods of rapeseed;

[0080] Construct the judgment matrix A of the seedling stage m :

[0081]

[0082] Construct the judgment matrix A of flowering period h :

[0083]

[0084] Construct the judgment matrix A of flowering period j :

[0085]

[0086] Then calculate the eigenvector W of each judgment matrix and perform normalization to obtain the relative importance;

[0087]

[0088] Verify the consistency of the judgment matrix and use λ max The difference with n is used to test consistency, where λ max To determine the maximum eigenvalue of matrix A, the calculation deviation consistency index CI is defined as:

[0089] CI=(λ max -n) / (n-1) Formula 35

[0090] The consistency index is defined as:

[0091] CR=CI / RI Formula 36

[0092] In the formula, RI is the randomness index. When n=3, RI=0.58. When CR<0.1, it meets the consistency requirement. Otherwise, the judgment matrix vector needs to be readjusted. The weights of the hierarchical analysis method are calculated according to the judgment matrix of each period.

[0093] Combined with the established single parameter growth assessment model, and using the weights of each parameter given by hierarchical analysis, a combined evaluation model was established and verified.

[0094] Compared with the prior art, the innovations and advantages of this application are:

[0095] (1) This application divides rapeseed into seedling stage, flowering stage and pod stage according to growth nodes, simultaneously acquires ground hyperspectral and multispectral data, analyzes rapeseed physiological and biochemical data and multispectral images of drones, performs resolution adjustment and conversion, and converts them into all-round, multi-scale, and multi-phase data. The growth state factor inversion model of rapeseed is established to evaluate the contribution of different growth state factors to the growth of rapeseed. The model with the best verification index in different growth periods is the growth state evaluation model of that period. The innovation points include: First, using hyperspectral data as an intermediary, qualitatively and quantitatively analyzing the applicability and inversion accuracy of the inversion model between remote sensing platforms of different scales. The verification results show that the relative error of reflectance of MCA and SKR sensors in each observation band in different growth periods does not exceed 15%. Second, a regression model of the inversion error caused by local variance of the image and the scale adjustment is established to eliminate the inversion error caused by surface heterogeneity. Third, the manually evaluated rapeseed growth status level is combined with the leaf area index, aboveground biomass and chlorophyll content measured on the ground and elastically graded. The model is established based on multilayer linear regression of three types of growth factors, LAI, AGB and CC, partial least squares optimization model, and multi-mode prediction based on entropy aggregation, combined with expert knowledge and agronomic knowledge entropy loading expert knowledge hierarchical analysis. The weights assigned to different growth status parameters represent the contribution of the growth status parameters to the evaluation of rapeseed growth status in this period. The model with the best verification indicators set in different growth periods is the growth status evaluation model for this period. The rapeseed growth status evaluation cycle is short, highly targeted and accurate.

[0096] (2) This application establishes a rapeseed leaf area coefficient estimation model based on different growth stages of rapeseed on the ground, and scales the empirical model established and verified based on ground data to drones, and verifies that the accuracy of the multispectral images obtained by drones can meet the requirements after applying the ground empirical model. The estimation model of stem dry matter and leaf chlorophyll content was established using the measured aboveground dry matter mass and chlorophyll content at different growth stages and the hyperspectral data collected on the spot, and the estimation model of fresh weight and dry weight of rapeseed fruit at the pod stage was established using the flower coverage rate at the flowering stage, and the accuracy was verified. For the deviations between the model estimation results caused by drones of different scales, the causes of the errors were analyzed, and from the perspective of spatial heterogeneity, the relationship between local variance and error was analyzed and a regression model was established. The regression model was used to reduce the errors caused by scale changes, and the model was verified. Using the rapeseed growth status parameters at different growth stages as the driving factors of the model, a set of rapeseed growth status evaluation models for different growth stages was established to evaluate the growth status of rapeseed. The UAV of this application takes into account high spatial resolution and time resolution with its excellent economy, and can obtain high-resolution spectral images over a wide range. The convenient data acquisition method can obtain time series spectral images of vegetation in different growth cycles, and accurately evaluate the growth status of rapeseed in different growth stages. It has good promotion and applicability, strong model explanatory power, high confidence, and high evaluation efficiency.

[0097] (3) The present application collects long-term ground spectral information according to different growth stages of rapeseed, and the period-based LAI estimation model established by this has high generalizability and applicability. The LAI estimation model for rapeseed seedling stage is Y LAI =1.7242X EVI2 +1.8042, R of the model 2=0.7693, RMSE =0.061, upscaled to UAV data, the RMSE of the inversion result is 0.651, and the accuracy is 74.9%. The second is to calculate the extreme value of the local variance of the UAV image and obtain the optimal resolution, which is 4cm in the seedling stage, 1cm in the flowering stage, and 2cm in the pod stage. The Taylor series expansion is used to analyze the cause of the error in the inversion result after upscaling and downscaling, and the correlation between the local variance and the inversion error is analyzed, and a model is established to eliminate the error. Third, the three growth status parameters are elastically rated. The model is established based on multilayer linear regression of three types of growth factors, LAI, AGB, and CC, partial least squares optimization model, and multi-mode prediction based on entropy aggregation, combined with expert knowledge and agronomic knowledge entropy loading expert knowledge hierarchical analysis. According to the agronomic significance and model verification indicators of the model in different periods, a multivariate regression model is set up on rapeseed seedlings to evaluate their growth. The growth status evaluation models for the three periods have a good mathematical and agronomic foundation, strong model explanatory power, and high confidence. After regional extrapolation verification, it has good evaluation accuracy, can describe the individual characteristics and group characteristics of crops at different periods as well as the health status of crops, which is conducive to providing timely management information for large farms and large-scale cluster agriculture, and also provides a basis for agricultural producers and futures trading to estimate production. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 It is the local variance map of LAI images with different resolutions at the rapeseed seedling stage.

[0099] Figure 2 It's LAI N and LAI L Plots of the two upscaling results.

[0100] Figure 3 This is a graph of the error and local variance regression model results for three periods.

[0101] Figure 4 This is a schematic diagram of the AGB classification intervals of aboveground biomass in three periods.

[0102] Figure 5 This is a schematic diagram of the CC classification intervals of chlorophyll content in three periods.

[0103] Figure 6 It is a schematic diagram of the partial least squares growth status evaluation model and model verification.

[0104] Figure 7 This is a schematic diagram of modeling single growth factors in each period using entropy aggregation and multi-mode prediction.

[0105] Figure 8 This is a schematic diagram of multi-mode prediction and verification based on entropy aggregation.

[0106] Fig. 9It is a pairwise comparison scaling diagram of entropy value loading expert knowledge in the hierarchical analysis seedling stage.

[0107] Fig.10 It is a schematic diagram of density segmentation and false color display of three growth parameters.

[0108] Fig.11 This is a schematic diagram of the evaluation results of rapeseed growth during the flowering period in the experimental area. DETAILED DESCRIPTION

[0109] The following, in conjunction with the accompanying drawings, further describes the technical solution of the method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles provided in the present application, so that those skilled in the art can better understand the present application and implement it.

[0110] This application divides rapeseed into seedling stage, flowering stage and pod stage according to growth nodes, simultaneously acquires ground hyperspectral and multispectral data, analyzes rapeseed physiological and biochemical data and multispectral images of drones, and obtains all-round, multi-scale and multi-temporal data. Through the analysis of remote sensing data, an inversion model of rapeseed's growth state factors is established, and the contribution of different growth state factors to the growth of rapeseed is evaluated, providing a quantitative decision-making basis for the management and production of precision agriculture.

[0111] 1) The SKR sensor was used to obtain long-term multispectral data of the entire growth cycle of rapeseed, and the changing trend of the rapeseed canopy spectrum during the growth period was analyzed. The LAI data of the LAI-2200C canopy analyzer was combined to establish a LAI estimation model for different periods. The high-spectral resolution spectral reflectance information of the rapeseed canopy measured by the FieldSpec 4 ground feature spectrometer was calculated according to the spectral response function of the SKR sensor and the MCA multispectral camera, and the common features of the two sensors for modeling were determined.

[0112] 2) The local variance of MCA multispectral images with different resolutions was calculated to obtain the optimal resolution. The factors that caused the scale effect were determined to be model nonlinearity, model driving factor nonlinearity and surface heterogeneity. The upscaling error model of rapeseed growth factor was established based on Taylor series expansion. The correlation between the inversion error caused by the scale effect and the local variance of the image was calculated, and a regression equation of the local variance and error was established to correct the inversion results after upscaling and downscaling.

[0113] 3) The growth status of rapeseed was evaluated based on comprehensive evaluation indicators of leaf area index, aboveground biomass and chlorophyll content. Due to different climates, planting methods and rapeseed varieties in different regions, plastic grading indicators were not used, but elastic evaluation indicators with regional characteristics were used to relatively evaluate the rapeseed in the survey area. A model was established based on multilayer linear regression of three types of growth factors, LAI, AGB and CC, partial least squares optimization model, multi-mode prediction based on entropy aggregation, and hierarchical analysis of expert knowledge loaded with entropy of expert knowledge and agronomic knowledge. The weights assigned to different growth status parameters represent the contribution of the growth status parameters to the evaluation of rapeseed growth status in this period. The model with the best verification index in different growth periods was set as the growth status evaluation model for this period. Multilayer linear regression of three types of growth factors was used for growth evaluation in the seedling stage, partial least squares optimization model was used as the growth evaluation model in the flowering stage, and multi-mode prediction based on entropy aggregation was set as the growth evaluation model in the pod stage.

[0114] 1. Analysis of scale effect of rapeseed growth factors and scale conversion method

[0115] As rapeseed undergoes large phenotypic variations at different growth stages, and the main parts of the canopy are leaves, flowers, and pods, it is necessary to set the optimal resolution according to the growth stage, adjust the resolution up or down, and reduce the error caused by the scale effect, so as to adjust the resolution of the drone image to the best.

[0116] Optimal spatial resolution for rapeseed assessment

[0117] The spatiotemporal growth aggregation method is used to scale the leaf area coefficient images of the drone at different times to obtain images with different resolutions of 1cm, 2cm, 4cm, 8cm, 16cm and 32cm, and the corresponding local variance images are obtained, such as Figure 1 .

[0118] A 3×3 calculation template was used to calculate local variance images of different resolutions in three periods. When the spatial resolution was 4 cm in the seedling stage, the maximum local variance was 0.209, and the appropriate resolution for UAV observation in the seedling stage was 4 cm. During the flowering period, the flowers were distributed at the top of the rape canopy and were small in size. The maximum local mean square error was 0.339 when the resolution was 1 cm. In the pod stage, the pods were the observation objects and targets of the canopy, and the maximum local variance was 0.313 at a resolution of 2 cm.

[0119] (II) Establishing a scaling error model for rapeseed growth factors based on Taylor series expansion

[0120] The rapeseed production area is huge, and the drone's endurance is limited. The drone's aerial photography of the rapeseed production area is divided into different areas. Due to the interference of complex factors such as different operators and weather conditions, it is impossible to ensure that the flight altitude of each flight zone is consistent. In actual work, in order to ensure large-scale monitoring, the drone's flight altitude is 50 meters to 200 meters. The drone rapeseed observation flight altitude is 50 meters. An upscaling model for the 50-meter drone is established, and a rapeseed growth factor upscaling error model is established based on Taylor series expansion.

[0121] The upscaling of the MCA sensor is based on the nearest canopy pixel clustering method and the spatiotemporal growth aggregation method. The nearest canopy pixel clustering method retains the original scale image information to the maximum extent, and the spatiotemporal growth aggregation method simulates the process of mixing small pixels into large pixels after upscaling. The nearest canopy pixel clustering method downsamples the 1.3cm resolution image, and the spatiotemporal growth aggregation method uses a template to traverse the entire image to generate a new mean image as the result of upscaling.

[0122] Step 1: Spatiotemporal growth aggregation of rapeseed growth factors, the method is as follows:

[0123]

[0124] D 反射 =d 1反射 +d 2反射 +…+d n反射 Formula 2

[0125] D 入射 =d 1入射 +d 2入射 +…+d n入射 Formula 3

[0126]

[0127] Approximately considering the incident energy of each small pixel to be the same, we get:

[0128]

[0129] d i反射 is the incident energy received by the initial pixel, d i反射 is the energy reflected from the initial pixel to the sensor, G 升尺度 is the reflectivity of the pixel after spatiotemporal growth aggregation, and n is the total number of aggregated pixels;

[0130] Step 2: Based on the nearest canopy pixel clustering, assign the neighboring pixel value closest to the pixel to be found to the pixel to be found:

[0131] I(P)=I(N) Formula 7

[0132] x N=INT(x+0.5) Formula 8

[0133] The coordinates of the point P to be found are (x, y), INT is the rounding function, and N is the neighboring pixel;

[0134] Depend on Figure 2 It can be concluded that there is an error in the leaf area coefficient before and after the scaling. Let the leaf area coefficients before and after the scaling be LAI N and LAI L ,The LAI values ​​obtained by MCA after spatiotemporal growth aggregation combined with the ground estimation model inversion have a certain degree of underestimation and overestimation, and the spatial heterogeneity within the large-scale pixels causes upscaling errors;

[0135] Step 3: Definition Also define error = LAI N -LAI L :

[0136]

[0137]

[0138] in, n is the number of pixels in the initial image corresponding to one pixel after upscaling, and Taylor's formula is used to calculate In CI green i =mCI green Expand at

[0139]

[0140] The ground LAI estimation model is a linear model with a second-order derivative of 0. is relatively The higher-order infinitesimal of is used as the error term;

[0141]

[0142] in It is the local variance of n pixels on the image. It can be deduced from the error term that the error is related to the local variance of the image. The second-order derivative of the ground estimation linear model is 0. A regression model is established by combining the absolute error with the local variance of the image. The errors caused by spatial heterogeneity of the images are eliminated after scaling the drone images of different scales.

[0143] (III) Correcting scale transformation errors based on local variance

[0144] The canopy morphology and spectrum of rapeseed change significantly in different growth stages. Different leaf area coefficient estimation models are established in different growth stages. The multispectral imager carried by the UAV is also affected by the spectral changes of the canopy morphology. An error correction model is established based on the growth stage to limit the impact of different growth stages to the minimum range. Each pixel of the small-scale UAV multispectral imager belongs to the same ground object, and its spectrum reflects the characteristics of the ground object. However, after the scale is increased, each pixel no longer contains a single ground object, but becomes a mixed pixel composed of multiple end members. To unmix the mixed pixels, it is necessary to predetermine the type of each end member and the ground spectrum for linear or nonlinear unmixing. However, the resolution of the UAV image is high, the area is large, and the texture characteristics are single. It is determined that the impact of the mixed pixels on the UAV spectral image is less than the impact of the spatial heterogeneity of the spectral image.

[0145] Based on the local variance of the image to reflect the spatial heterogeneity, the EVI is determined based on the Taylor formula expansion 2 and CI green The correlation between the upscaling error and the local variance is calculated, and the correlation between the local variance and the error in the three periods is calculated;

[0146] About EVI 2 and CI green The local variance of the image is calculated, an odd number template is used, a 3×3 calculation template is set, random sampling points are set, the error value and the local variance of the sample area corresponding to the error value are calculated, and the error and local variance are used as variables for regression modeling. The empirical model of using local variance to estimate the error in three periods is obtained, and the scale effect error obtained in the three growth periods of rapeseed is regressed with the local variance model, as shown in the figure. Figure 3 .

[0147] 2. Rapeseed Growth Status Assessment Model

[0148] It is divided into evaluation model and diagnosis model. The diagnosis model detects whether the vegetation is threatened by pests and diseases, nutrients and climate environment. The evaluation model is divided into year-on-year comparison model and grade model. The year-on-year comparison model compares and analyzes the measured values ​​or vegetation index obtained this year with the measured values ​​and vegetation index under good growth in previous years.

[0149] 1. Constructing growth elasticity evaluation factors

[0150] Based on the uncertainty of annual environmental changes, the upgrading and improvement of rapeseed varieties in the planting areas, and the large differences in climate in different planting areas, a more adaptable elastic evaluation factor is established to adapt to different operating sample areas and enhance the regional extrapolation capability and robustness of the evaluation model.

[0151] Modeling requires data standardization of all inverted leaf area index LAI, aboveground biomass AGB and chlorophyll content CC. The method is to use the maximum and minimum values ​​as boundaries, divide their difference by 10, and use it as the level interval to divide the inversion parameters into 1 to 10 levels:

[0152]

[0153] GS LAI is the LAI classification segment, INT is the rounding function, LAI estimate The leaf area index is obtained by UAV inversion;

[0154]

[0155] GS AGB is the AGB classification segment, INT is the rounding function, AGB estimate The aboveground biomass is obtained by UAV inversion; Figure 4 shown.

[0156]

[0157] GS CC is the CC grading segment, INT is the rounding function, CC estimate The chlorophyll content is obtained by UAV inversion. Figure 5 shown.

[0158] 2. Establishing a rapeseed evaluation model

[0159] Rapeseed growth is affected and restricted by various factors. These factors play different roles in different growth stages of rapeseed. For example, the growth characteristics of rapeseed in different growth stages provide different information for yield estimation. If a single growth state factor is used, some useful information may be lost, which will affect the precision and accuracy of rapeseed growth state assessment. It is more scientific and reasonable to comprehensively consider the growth state parameters of each period to conduct rapeseed growth state assessment.

[0160] The contribution of the information carried by the growth status factors of rapeseed in different growth stages to its growth potential is converted into a weight determination problem. When the growth status factors are inverted using remote sensing images of rapeseed in multiple periods, the size of the weight reflects the amount of growth information carried by the growth status in each period, and the most important growth status factors for growth evaluation are found.

[0161] This application uses three types of plant growth factors to evaluate the growth of rapeseed, comprehensively considers the contribution of leaf area index LAI, aboveground biomass AGB and chlorophyll content CC to the growth status, and establishes a model based on LAI, AGB, CC multi-layer linear regression, partial least squares optimization and mathematical analysis entropy method combined with hierarchical analysis of expert knowledge and agronomic knowledge.

[0162] 1. Multi-layer linear regression of three types of growth factors

[0163] The measured LAI, AGB, and CC were elastically evaluated, with their levels as independent variables and the growth status level assessed by experts as the dependent variable for multi-layer linear regression.

[0164] Multilayer linear regression was used to obtain the coefficients of each parameter. The leaf area index grade made an important contribution to the regression of the growth status grade. The chlorophyll content coefficient was high during the flowering period, and chlorophyll was used as a reference indicator when determining the regression. The contribution of aboveground biomass to the growth status rating during the pod stage was greater than that in the previous two periods. The model was verified for accuracy and T-tested. Neither AGB nor CC passed the significance test during the seedling stage, AGB during the flowering period, and CC during the pod stage.

[0165] 2. Partial Least Squares Optimization Model

[0166] The professional partial least squares analysis software SmartPLS 3.2 was used to build the growth status evaluation model. In the three growth periods, 32 sample points were randomly set for modeling, and 16 sample points were used for model verification. Figure 6 It is a partial least squares growth state evaluation model and model verification.

[0167] The partial least squares optimization evaluation model has a good determination coefficient, but its model verification index root mean square error is much larger than that of other modeling methods.

[0168] 3. Multi-mode prediction based on entropy aggregation

[0169] Assume that for a certain indicator X of the same prediction object t (t=1,2,…,N), there are m single prediction methods to predict it, among which the prediction value of the i-th single prediction method at time t is X it (i=1,2,…,m;t=1,2,…,N), through K 1 ,K 2 ,…,K m The weight coefficients are used to weight the m individual prediction methods, where:

[0170] K 1 +K 2 +K 3 +…K m =1 Formula 20

[0171] The combined prediction X index is:

[0172] X t =K 1 ×X 1t +K2 ×X 2t +K 3 ×X 3t +…+K m ×X mt Formula 21

[0173] The information entropy method is used to obtain the weight coefficient. Information entropy specifically reflects the variability of a certain indicator of the system and measures the amount of information. The greater the amount of information contained in a certain indicator of the system, the greater the role of this indicator in the final decision. At this time, the greater the entropy value, the greater the weight should be given. The overall algorithm of multi-mode prediction based on entropy aggregation is as follows:

[0174] Step 1: Normalize the prediction error:

[0175] The proportion of the relative error of the forecast is calculated by calculating the relative error between the forecast result of the i-th single forecast at time t and the actual value:

[0176] e it =|(X t -X it ) / X t Formula 22

[0177] where e it is the prediction relative error of the i-th prediction method at the t-th time, and the proportion of the prediction relative error of the i-th single prediction method at the t-th time is described as:

[0178]

[0179] Step 2, calculate the entropy value:

[0180] The entropy value of the relative error of the prediction of the i-th single prediction method is:

[0181]

[0182] Where k is a constant greater than zero, k = 1 / InN, h i The value range of is 0≤h≤1;

[0183] Step 3, calculate the coefficient of variation:

[0184] The coefficient of variation of the relative error of the prediction of the i-th single prediction method is:

[0185] d i =1-h i Formula 26

[0186] Step 4, calculate the entropy weight:

[0187] The entropy weight of the i-th single prediction method is:

[0188]

[0189] Step 5, calculate the weight coefficient:

[0190]

[0191] For the three growth periods, linear regression models were established with leaf area index level, aboveground biomass level and chlorophyll content level as independent variables and growth level as dependent variable. The linear model was established to prevent the scale effect caused by the nonlinearity of the model when the scale changed, which would cause the model's inadaptability and uncontrollable errors.

[0192] Figure 7 Modeling a single growth factor for each period, where y 1 ,y 2 ,y 3 is the estimation model of the growth parameters at the seedling stage for the growth grade, y 4 ,y 5 ,y 6 is the estimation model of the growth parameters of flowering period for the growth grade, y 7 ,y 8 ,y 9 is the estimation model of the growth parameters for the growth grade during the pod stage, x lai , x agb , x cc It is leaf area index level, aboveground biomass level, and chlorophyll content level.

[0193] The information entropy method was used to establish the growth status evaluation model of rapeseed at different stages, and the model evaluation and verification were carried out. Figure 8 In the three growth periods, 32 sample points were randomly set for modeling, and 16 sample points were used for model verification.

[0194] Figure 8 The entropy weighted growth model and its verification are presented. The entropy method weights different growth factors according to the information they carry. The model has a good mathematical foundation and strong interpretability. The determination coefficients of the three-period model are all greater than 0.5, and the model is available. The model can better reflect the importance of different growth factors in judging the growth of rapeseed, among which the leaf area index is more important in these three periods.

[0195] 4. Hierarchical analysis of expert knowledge loaded with entropy

[0196] Step 1: Establish a rapeseed evaluation structural model: First, after analyzing the rapeseed evaluation, the relevant factors are divided into several levels from top to bottom according to different attributes. Among the many factors in the same layer, the factors in the upper layer or have an influence on the factors in the upper layer and at the same time dominate the factors in the lower layer or are affected by the factors in the lower layer. The factors in the same layer are independent of each other. The top is the target layer with only one element, the bottom is the object layer, and there are multiple levels in between.

[0197] Step 2: Construct the growth judgment matrix of each level: Starting from the second level of the hierarchical structure model, use the paired comparison method for the factors of the same level that belong to each factor of the previous level, with a comparison scale from 1 to 9, to construct a comparison matrix until the bottom level;

[0198] Step 3: Evaluate the consistency calculation test:

[0199] The evaluation weight vector is calculated and a consistency test is performed. The maximum eigenvalue root and the maximum eigenvector are calculated by comparing the matrix. A consistency test is performed based on the consistency index CI and the random consistency index RI and the ratio of the two, that is, the consistency ratio. If it is less than 0.1, the consistency test passes. Then, the characteristic items are normalized to obtain the weight vector. If the test fails, the comparison matrix is ​​reconstructed.

[0200] According to the scores given by invited experts and combined with agricultural knowledge, the obvious errors in the scores were eliminated and the weight of the matrix was judged by the geometric mean method. Three judgment matrices were established for the different growth periods of rapeseed;

[0201] During the seedling stage, rapeseed is between the four-leaf stage and the eight-leaf stage, and the leaf area index changes greatly. The change of leaf area index can better reflect the level of rapeseed growth change. At the same time, rapeseed undergoes wintering, and the chlorophyll content reflects the level of frost damage suffered by rapeseed. However, the difference between each plant in aboveground biomass is not obvious, and its ability to characterize growth is weak. Fig. 9 .

[0202] Construct the judgment matrix A of the seedling stage m :

[0203]

[0204] During the flowering period, rapeseed goes through the budding and bolting period, the plant grows densely, the stem growth is completed, and rapeseed flowers bloom on the top, and the leaf area index reaches the maximum. The leaves are the site of photosynthesis, and the leaf area is positively correlated with the intercepted light energy. Leaf area plays an important role in evaluating the growth status of rapeseed during the flowering period. At this time, the aboveground biomass is directly reflected in the stems and leaves of the rapeseed, and also reflects the accumulation of organic matter, which is indicative of the growth. Affected by the canopy flowers, the visible light part of the canopy spectrum is raised as a whole, and chlorophyll is not easy to detect and is greatly affected, so it has limited effect on evaluating the growth.

[0205] Construct the judgment matrix A of flowering periodh :

[0206]

[0207] During the pod stage, the rapeseed inflorescence ends, the leaves gradually fall off, and the leaf area index weakens its role in indicating growth. At the same time, the pods begin to grow, and the accumulation of biomass enters a peak period. The pod skin becomes the core site of photosynthesis, and the aboveground biomass has a significant role in indicating the growth of rapeseed. As the pods mature, the yellow chlorophyll of rapeseed gradually fades, and its effect on growth is weak.

[0208] Construct the judgment matrix A of flowering period j :

[0209]

[0210] Then calculate the eigenvector W of each judgment matrix and perform normalization to obtain the relative importance;

[0211]

[0212] Verify the consistency of the judgment matrix and use λ max The difference with n is used to test consistency, where λ max To determine the maximum eigenvalue of matrix A, the calculation deviation consistency index CI is defined as:

[0213] CI=(λ max -n) / (n-1) Formula 35

[0214] The consistency index is defined as:

[0215] CR=CI / RI Formula 36

[0216] In the formula, RI is the randomness index. When n=3, RI=0.58. When CR<0.1, it meets the consistency requirement. Otherwise, the judgment matrix vector needs to be readjusted. The weights of the hierarchical analysis method are calculated according to the judgment matrix of each period.

[0217] Combined with the established single parameter growth assessment model, and using the weights of each parameter given by the hierarchical analysis, a combined evaluation model was established and verified;

[0218] In the three growth periods, 32 sample points were randomly set for modeling, and 16 sample points were used for model verification. Fig.10 ,The hierarchical analysis involves professionals and the empowerment process adds agronomy background, so the method has better theoretical ,relevance and is easy to explain.

[0219] 2. Assessing the accuracy and adaptability of the growth status evaluation model

[0220] Using remote sensing to evaluate rapeseed growth status over a large area has become a hot topic in the current agricultural remote sensing field. UAV remote sensing images have become the first choice in actual production due to their excellent performance. By using the established evaluation model and combining the two based on the inversion of growth status factors, a new technical means is provided for agricultural remote sensing evaluation of crop growth.

[0221] 1. Combination of UAV inversion parameters and growth evaluation model

[0222] According to the growth parameter rating method, the remote sensing image inversion results of leaf area index, aboveground biomass and chlorophyll content at the seedling stage, flowering stage and pod stage were used to construct elastic evaluation indicators.

[0223] The inverted growth parameters of each period were used as driving variables of the model to input the estimated growth results in the growth evaluation model established above. All growth result graphs were enhanced and displayed using MATLAB's pseudo-color graph.

[0224] 2. UAV growth evaluation product accuracy assessment

[0225] The average value of the absolute relative error (ARE), the root mean square error (RMSE), and the estimation accuracy are used as verification indicators to evaluate the accuracy of the evaluation products obtained by the four models in different growth periods.

[0226] (1) Seedling stage

[0227] The relative error distributions of the two linear regression models are discrete. The multi-layer linear regression growth evaluation model underestimates the growth, and the relative error of the growth evaluation of rapeseed plots with poor or above grades is negative; the partial least squares regression optimization overestimates the growth at the seedling stage. Both of them are more biased towards leaf area index in terms of weighting. The partial least squares regression gives a negative weight to the aboveground biomass, and the partial least squares ignores the nutritional stress of rapeseed seedlings due to different soil nutrients at the seedling stage; the multi-mode prediction based on entropy aggregation and the hierarchical analysis of expert knowledge with entropy loading show a high degree of similarity. Both overestimate the growth at the poor and medium levels, and underestimate the growth at the better and good levels.

[0228] Comprehensive evaluation at the seedling stage, multivariate regression growth evaluation was higher than the other three models in the three verification indicators. Therefore, multilayer linear regression of three types of growth factors was used for growth evaluation at the seedling stage.

[0229] (2) Flowering period

[0230] The multi-layer linear regression and partial least squares optimization models of three types of growth factors overestimated the growth of medium and lower grades. Since both models have a large weight on chlorophyll, and the flowering period is affected by the spectrum of the top rapeseed flowers, the red and green bands, which are critical for chlorophyll inversion, will be disturbed, so the two linear models of flowering period are more seriously affected. Although the relative errors of the entropy combination method and the hierarchical analysis method are quite similar, the former gives almost the same weight to the three parameters, while the latter pays more attention to leaf area index and aboveground biomass.

[0231] Although the multi-mode prediction based on entropy aggregation and the hierarchical analysis of entropy-loaded expert knowledge are based on mathematical calculations and agronomy, they have a high relative error in growth assessment. Therefore, the partial least squares optimization model is used as the growth evaluation model during the flowering period.

[0232] (3) Pod stage

[0233] The multi-layer linear regression of three types of growth factors underestimated the overall growth. Although the partial least squares optimization model was relatively uniform, the leaves of rapeseed gradually fell during the silique period. The partial least squares optimization model did not reduce the weight, but gave a higher weight to the leaf area index. The multi-mode prediction based on entropy aggregation and the hierarchical analysis of entropy-loaded expert knowledge showed a certain degree of similarity, which was reflected in the fact that both gave a higher weight to the aboveground biomass, and believed that the accumulation of rapeseed pods and stem dry matter became a more important evaluation indicator during this period.

[0234] Comprehensive analysis shows that although the multi-mode prediction based on entropy aggregation and the hierarchical analysis of entropy-loaded expert knowledge both have a solid mathematical and agronomic foundation and good model explanatory power, considering the verification indicators, the multi-mode prediction based on entropy aggregation was set as the growth evaluation model in the pod stage.

[0235] (III) Regional extrapolation of rapeseed growth evaluation model

[0236] The optimal resolution of drone images during the flowering period is 1 cm. The multispectral data of aerial drones were downscaled to 1 cm resolution, and geometric correction and radiation calibration preprocessing were completed. The three types of factors calculated were classified according to the elastic index classification method, and enhanced display was performed using the density segmentation function of ENVI.

[0237] In order to verify the applicability of the model extrapolated to other production areas, experts combined their own agricultural knowledge with rapeseed phenotypic factors in the flowering period and soil environmental factors such as fertilizer application amount to obtain the rapeseed growth grade as the verification truth value.

[0238] The analysis determined that the partial least squares optimization model was used to evaluate the growth status of rapeseed during its flowering period, and the growth status parameter level was used as the input of the model to obtain the result image after evaluation, and the density segmentation was used for enhanced display. Fig.11 .

[0239] In order to compare the model evaluation results with the expert evaluation results, the estimated results of 36 fields were averaged and then divided into 5 levels using elastic classification. The classified data were calculated and the confusion matrix was compiled to reflect the error. The confusion matrix was proposed to compare the number of pixels in a specific category with the number of pixels classified into that category in the image.

[0240] The total number of samples is set as n, the number of evaluation levels is set as m, and n ij The value of is used as the number of samples (i is the number of rows and j is the number of columns) that are divided into level i (i=1,2,…,m) during the model evaluation process and rated as level j (j=1,2,…,m) by experts.

[0241] The number of samples in the sample that are rated as level i by the model is:

[0242]

[0243] The number of samples rated as level j by experts is:

[0244]

[0245] The overall accuracy OA is:

[0246]

[0247] The producer accuracy PA is:

[0248]

[0249] The user accuracy UA is:

[0250]

[0251] According to the above accuracy calculation formula and the confusion matrix of growth status evaluation level, the evaluation accuracy of rapeseed growth during flowering period after regional extrapolation was calculated. The overall accuracy is 0.625, which proves that the growth status of rapeseed can be evaluated relatively accurately by extrapolating the experimental area where the model is established to other experimental sample areas, and the model has strong portability. The model only overestimates different growth states. From the user's perspective, the model has a good evaluation ability for the three levels of good, medium and poor, but the evaluation effect is not good for the two more subjective and vague levels of good and poor. From the producer's perspective, the model's evaluation accuracy for rapeseed with poor growth is not high. In summary, regardless of whether the growth of rapeseed in the operation area is uniform or uneven, the rapeseed growth status evaluation model can accurately evaluate and grade its growth.

Claims

1. A method for evaluating rapeseed growth status using multi-source quantitative remote sensing by unmanned aerial vehicles, characterized in that: Rapeseed is divided into seedling stage, flowering stage and pod stage according to growth nodes. Ground hyperspectral and multispectral data are acquired simultaneously, and rapeseed physiological and biochemical data and multispectral images of drones are analyzed. The resolution is raised and lowered, and converted into all-round, multi-scale, and multi-temporal data. An inversion model of rapeseed growth status factors is established to evaluate the contribution of different growth status factors to the growth of rapeseed. The model with the best verification index in different growth stages is the growth status evaluation model for that period. 1) The SKR sensor was used to obtain long-term multispectral data of the entire growth cycle of rapeseed, and the changing trend of the rapeseed canopy spectrum during the growth period was analyzed. The LAI data of the LAI-2200C canopy analyzer was combined to establish a LAI estimation model for different periods. The high-spectral resolution spectral reflectance information of the rapeseed canopy measured by the FieldSpec 4 ground feature spectrometer was calculated according to the spectral response function of the SKR sensor and the MCA multispectral camera, and the common features of the two sensors for modeling were determined. 2) Calculate the local variance of MCA multispectral images with different resolutions to obtain the optimal resolution. Calculate and determine the factors that cause the scale effect as model nonlinearity, model driving factor nonlinearity and surface heterogeneity. Use Taylor series expansion to establish a rapeseed growth factor upscaling error model. Calculate the correlation between the inversion error caused by the scale effect and the local variance of the image. Establish a regression equation between the local variance and the error to correct the inversion results after upscaling and downscaling. 3) The growth status of rapeseed was evaluated based on the comprehensive evaluation indicators of leaf area index, aboveground biomass and chlorophyll content. The elastic evaluation index with regional characteristics was used to make a relative evaluation of the rapeseed in the survey area. A model was established based on the multilayer linear regression of three types of growth factors, LAI, AGB and CC, the partial least squares optimization model, and the multimode prediction based on entropy aggregation, combined with the entropy value of expert knowledge and agronomic knowledge. The weights assigned to different growth status parameters represent the contribution of the growth status parameters to the evaluation of the growth status of rapeseed in this period. Multilayer linear regression of three types of growth factors was used for growth evaluation in the seedling stage, the partial least squares optimization model was used as the growth evaluation model in the flowering stage, and the multimode prediction based on entropy aggregation was set as the growth evaluation model in the pod stage.

2. According to claim 1, the method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles is characterized in that: Optimal spatial resolution for rapeseed assessment: The spatiotemporal growth aggregation method was used to scale the leaf area coefficient images of drones at different times to obtain images with different resolutions of 1 cm, 2 cm, 4 cm, 8 cm, 16 cm and 32 cm, and the corresponding local variance images were obtained; A 3×3 calculation template was used to calculate local variance images of different resolutions in three periods. When the spatial resolution was 4 cm in the seedling stage, the maximum local variance was 0.209, and the appropriate resolution for UAV observation in the seedling stage was 4 cm. During the flowering period, the flowers were distributed at the top of the rape canopy and were small in size. The maximum local mean square error was 0.339 when the resolution was 1 cm. In the pod stage, the pods were the observation objects and targets of the canopy, and the maximum local variance was 0.313 at a resolution of 2 cm.

3. The method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles according to claim 1 is characterized in that: Establishing the upscaling error model of rapeseed growth factors based on Taylor series expansion: Establishing the upscaling model for 50-meter UAVs, and establishing the upscaling error model of rapeseed growth factors based on Taylor series expansion; Step 1: Spatiotemporal growth aggregation of rapeseed growth factors, the method is as follows: D 反射 =d 1反射 +d 2反射 +…+d n反射 Formula 2 D 入射 =d 1入射 +d 2入射 +…+d n入射 Formula 3 Approximately considering the incident energy of each small pixel to be the same, we get: d i入射 is the incident energy received by the initial pixel, d i反射 is the energy reflected from the initial pixel to the sensor, G 升尺度 is the reflectivity of the pixel after spatiotemporal growth aggregation, and n is the total number of aggregated pixels; Step 2: Based on the nearest canopy pixel clustering, assign the neighboring pixel value closest to the pixel to be found to the pixel to be found: I(P)=I(N) Formula 7 x N =INT(x+0.5) Formula 8 The coordinates of the point P to be found are (x, y), INT is the rounding function, and N is the neighboring pixel; Assume that the leaf area coefficients before and after scaling are LAI N and LAI L ,The LAI values ​​obtained by MCA after spatiotemporal growth aggregation combined with the ground estimation model inversion have a certain degree of underestimation and overestimation, and the spatial heterogeneity within the large-scale pixels causes upscaling errors; Step 3: Definition Also define error = LAI N -LAI L : in, n is the number of pixels in the initial image corresponding to one pixel after upscaling, and Taylor's formula is used to calculate In CI green i =mCI green Expand at The ground LAI estimation model is a linear model with a second-order derivative of 0. is relatively The higher-order infinitesimal of is used as the error term; in It is the local variance of n pixels on the image. It can be deduced from the error term that the error is related to the local variance of the image. The second-order derivative of the ground estimation linear model is 0. A regression model is established between the absolute error and the local variance of the image. The errors caused by the spatial heterogeneity of the images are eliminated after scaling the drone images of different scales.

4. The method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles according to claim 1 is characterized in that: Correcting the scale transformation error based on local variance: The local variance of the image reflects the spatial heterogeneity, and the EVI2 and CI are determined based on the Taylor formula expansion. green The correlation between the upscaling error and the local variance is calculated, and the correlation between the local variance and the error in the three periods is calculated; EVI2 and CI green The local variance of the image is calculated, an odd number template is used, a 3×3 calculation template is set, random sampling points are set, the error value and the local variance of the sample area corresponding to the error value are calculated, the error and local variance are used as variables for regression modeling, and an empirical model for estimating the error using local variance in three periods is obtained.

5. The method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles according to claim 1 is characterized in that: Construct growth elasticity evaluation factors: All the inverted leaf area index LAI, aboveground biomass AGB and chlorophyll content CC are subjected to data standardization. The method is to use the maximum and minimum values ​​as boundaries, divide their difference by 10, and use it as the level interval to divide the inverted parameters into 1 to 10 levels: GS LAI is the LAI classification segment, INT is the rounding function, LAI estimate The leaf area index is obtained by UAV inversion; GS AGB is the AGB classification segment, INT is the rounding function, AGB estimate The aboveground biomass was obtained by UAV inversion; GS CC is the CC grading segment, INT is the rounding function, CC estimate The chlorophyll content is obtained by inversion using a drone.

6. The method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles according to claim 1, characterized in that: Multilayer linear regression of three types of growth factors: elastic evaluation of LAI, AGB, and CC measured was performed, with their levels as independent variables and the level of growth status assessed by experts as the dependent variable for multilayer linear regression; Multilayer linear regression was used to obtain the coefficients of each parameter. The leaf area index level made an important contribution to the regression of the growth status level. Chlorophyll was used as a reference indicator when determining the regression. The contribution of aboveground biomass to the growth status rating during the pod stage was greater than that in the previous two periods. The model was verified for accuracy and T-tested. Neither AGB nor CC passed the significance test at the seedling stage, AGB at the flowering stage, and CC at the pod stage.

7. The method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles according to claim 1 is characterized in that: Partial least squares optimization model: Professional partial least squares analysis software SmartPLS 3.2 was used to build the growth status evaluation model. In the three growth periods, 32 sample points were randomly set for modeling, and 16 sample points were used for model verification.

8. The method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles according to claim 1, characterized in that: Multi-mode prediction based on entropy aggregation: Assuming that for a certain indicator X of the same prediction object t (t=1, 2, ..., N), there are m single prediction methods to predict it, among which the prediction value of the i-th single prediction method at time t is X it (i=1,2,…,m;t=1,2,…,N), through K1,K2,…,K m The weight coefficients are used to weight the m individual prediction methods, where: K1+K2+K3+…K m =1 Formula 20 The combined prediction X index is: X t =K1×X 1t +K2×X 2t +K3×X 3t +…+K m ×X mt Formula 21 The information entropy method is used to obtain the weight coefficient. Information entropy specifically reflects the variability of a certain indicator of the system and measures the amount of information. The greater the amount of information contained in a certain indicator of the system, the greater the impact of this indicator on the final decision. At this time, the greater the entropy value, the greater the weight should be given.

9. The method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles according to claim 8, characterized in that: The overall algorithm of multi-mode prediction based on entropy aggregation is as follows: Step 1: Normalize the prediction error: The proportion of the relative error of the forecast is calculated by calculating the relative error between the forecast result of the i-th single forecast at time t and the actual value: e it =|(X t =X it ) / X t | Formula 22 where e it is the prediction relative error of the i-th prediction method at the t-th time, and the proportion of the prediction relative error of the i-th single prediction method at the t-th time is described as: Step 2, calculate the entropy value: The entropy value of the relative error of the prediction of the i-th single prediction method is: Where k is a constant greater than zero, k = 1 / InN, h i The value range of is 0≤h≤1; Step 3, calculate the coefficient of variation: The coefficient of variation of the relative error of the prediction of the i-th single prediction method is: d i =1-h i Formula 26 Step 4, calculate the entropy weight: The entropy weight of the i-th single prediction method is: Step 5, calculate the weight coefficient: For the three growth periods, linear regression models were established with leaf area index level, aboveground biomass level and chlorophyll content level as independent variables and growth level as dependent variable. The linear model was established to prevent the scale effect caused by the nonlinearity of the model when the scale changed, which would cause the model's inadaptability and uncontrollable errors.

10. The method for evaluating rapeseed growth status by multi-source quantitative remote sensing using unmanned aerial vehicles according to claim 1, characterized in that: Hierarchical analysis of expert knowledge loaded with entropy value: Step 1: Establish a rapeseed evaluation structural model: First, after analyzing the rapeseed evaluation, the relevant factors are divided into several levels from top to bottom according to different attributes. Among the many factors in the same layer, the factors in the upper layer or have an influence on the factors in the upper layer and at the same time dominate the factors in the lower layer or are affected by the factors in the lower layer. The factors in the same layer are independent of each other. The top is the target layer with only one element, the bottom is the object layer, and there are multiple levels in between. Step 2: Construct the growth judgment matrix of each level: Starting from the second level of the hierarchical structure model, use the paired comparison method for the factors of the same level that belong to each factor of the previous level, with a comparison scale from 1 to 9, to construct a comparison matrix until the bottom level; Step 3: Evaluate the consistency calculation test: Calculate the evaluation weight vector and perform consistency test. Compare the matrix to calculate the maximum eigenvalue root and the maximum eigenvector. Perform consistency test based on the consistency index CI and the random consistency index RI and the ratio of the two, that is, the consistency ratio. If it is less than 0.1, the consistency test passes. Then, normalize the eigenvalues ​​to get the weight vector. If the test fails, reconstruct the comparison matrix. According to the scores given by invited experts and combined with agricultural knowledge, the obvious errors in the scores were eliminated and the weight of the matrix was judged by the geometric mean method. Three judgment matrices were established for the different growth periods of rapeseed; Construct the judgment matrix A of the seedling stage m : Construct the judgment matrix A of flowering period h : Construct the judgment matrix A of flowering period j : Then calculate the eigenvector W of each judgment matrix and perform normalization to obtain the relative importance; Verify the consistency of the judgment matrix and use λ max The difference with n is used to test consistency, where λ max To determine the maximum eigenvalue of matrix A, the calculation deviation consistency index CI is defined as: CI = (λ max - n) / (n - 1) Equation 35 The consistency index is defined as: CR=CI / RI Formula 36 In the formula, RI is the randomness index. When n=3, RI=0.

58. When CR<0.1, it meets the consistency requirement. Otherwise, the judgment matrix vector needs to be readjusted. The weights of the hierarchical analysis method are calculated according to the judgment matrix of each period. Combined with the established single parameter growth assessment model, and using the weights of each parameter given by hierarchical analysis, a combined evaluation model was established and verified.

Images