Canopy chlorophyll content inversion method based on feature extraction and machine learning model

Abstract

The invention belongs to the technical field of chlorophyll content inversion, and discloses a canopy chlorophyll content inversion method based on feature extraction and a machine learning model, and the method comprises the following steps: employing a mode of combining an unmanned plane carrying a multispectral camera and ground actual measurement; collecting and processing the chlorophyll content data of the potato canopies, and planting a quilt index; carrying out feature selection by using a Pearson correlation analysis and competitive self-adaptive reweighted sampling algorithm; constructing a chlorophyll content inversion model based on machine learning; and carrying out chlorophyll content inversion model optimization based on a grey wolf optimization algorithm and a sparrow optimization algorithm, and carrying out precision evaluation. According to the canopy chlorophyll content inversion method based on feature extraction and the machine learning model, comparative study is carried out on the single growth period and the whole growth period of potatoes, the optimal inversion model suitable for the single growth period and the whole growth period is determined, and the accuracy and applicability of the study are guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of chlorophyll content inversion, and in particular to a canopy chlorophyll content inversion method based on feature extraction and machine learning models. Background Art

[0002] As the fourth largest food crop in the world, potatoes have significant economic and social value. my country's potato planting area is 4.218 million hectares, occupying an important position in global potato production. Chlorophyll plays a vital role in the growth of potatoes because it is an important component of photosynthesis. For crops such as potatoes, this is the basis for their growth and development. Therefore, it can be used as an important indicator for evaluating the growth status of potatoes.

[0003] Traditionally, chlorophyll content is generally measured by chemical instruments. Although this method is flexible to operate, it is costly and is also affected by factors such as light and temperature, which is not conducive to large-scale crop monitoring. Chlorophyll value measuring instruments are widely used in outdoor field environments to measure the relative chlorophyll content of leaf canopies. In recent years, they have received increasing attention in crop growth monitoring. Unmanned Aerial Vehicles (UAVs) equipped with multispectral cameras have red edge and near-infrared band detection functions. Compared with other general cameras, they have efficient and fast data processing and are relatively inexpensive, so they are widely used in crop yield prediction.

[0004] With the development of artificial intelligence, the application of machine learning (ML) algorithms has become more and more extensive. Machine learning models can provide support for the study of canopy chlorophyll content and improve the inversion accuracy. Although previous studies have shown that drone multispectral images combined with machine learning models significantly improve the accuracy of chlorophyll content inversion, and have been widely used in the study of various crops, most studies have only focused on a single growth period, lacking comparative studies for different growth periods, and different potato varieties have significant differences in their impact on agricultural research models. Therefore, the present invention collected drone and ground data of sixteen kinds of potatoes, and conducted comparative studies on the single growth period and the full growth period of potatoes. Summary of the invention

[0005] The purpose of the present invention is to provide a canopy chlorophyll content inversion method based on feature extraction and machine learning models, to construct an efficient chlorophyll content inversion model, to combine the characteristics of each development stage with measured data, to develop a model combination of different feature selection methods and parameter optimization, and to significantly improve the accuracy of model inversion; to conduct a comparative study on a single growth period and the entire growth period of potatoes, to determine the best inversion model suitable for a single growth period and the entire growth period, and to ensure the accuracy and applicability of the research.

[0006] To achieve the above object, the present invention provides a canopy chlorophyll content inversion method based on feature extraction and machine learning model, comprising the following steps:

[0007] Step S1, using a combination of a multispectral camera mounted on an unmanned aerial vehicle and ground measurement to collect and process the chlorophyll content data of potato canopies in the study area, and construct a vegetation index;

[0008] Step S2, using Pearson correlation analysis and competitive adaptive reweighted sampling CARS algorithm to perform feature selection on vegetation index;

[0009] Step S3: constructing a chlorophyll content inversion model based on machine learning;

[0010] Step S4, optimizing the chlorophyll content inversion model constructed in step S3 based on the gray wolf optimization algorithm and the sparrow optimization algorithm;

[0011] Step S5: perform accuracy evaluation on the chlorophyll content inversion model optimized in step S4.

[0012] Preferably, in step S1, a method combining a multispectral camera mounted on an unmanned aerial vehicle and ground measurement is used to collect and process the chlorophyll content data of potato canopies in the study area, and a vegetation index is constructed. The specific process is as follows:

[0013] Step S11: The UAV carries a multispectral camera to acquire and process image data;

[0014] Firstly, an unmanned aerial vehicle (UAV) was used to collect image data of the study area. The UAV was equipped with four multispectral sensor probes, corresponding to the spectral band ranges of green light GREEN, red light RED, red edge light REG, and near infrared light NIR.

[0015] Then, the collected images were input into Pix4Dmapper software for multispectral image synthesis and calibration, thereby obtaining orthorectified multispectral reflectance images;

[0016] Step S12: collecting and measuring ground data;

[0017] First, the study area was divided into 16 plots, each plot corresponding to one variety, and ground data of 16 potato varieties were collected. The area of ​​each plot was divided into 20m×100m, and data from 10 sample points were collected evenly in each plot. The leaf canopy chlorophyll content was collected from the potato planting base in the study area at the seedling stage and the mature stage, and 160 ground data were collected in each growth period.

[0018] Then, the 16 sample plots divided above were taken as objects, and 10 well-growing potato canopies were evenly selected as sample points in each sample plot. When measuring each leaf canopy, the veins were avoided, and 10 sample points were evenly collected at the mesophyll using a handheld chlorophyll meter. The average value was taken as the true value of a sample point. The readings obtained from the potato plants selected at the sample points of each sample plot represented the relative chlorophyll content SPAD of the sample plot.

[0019] Step S13: constructing a vegetation index based on the above acquired and processed data;

[0020] The orthophoto images obtained from the UAV after geometric correction were divided into 16 corresponding areas according to potato varieties. Ten sample points were taken in each area. Combined with the longitude and latitude coordinates of the ground measurement points, the reflectance of the four bands of GREEN, RED, REG and NIR corresponding to each sample point was extracted on ArcGIS. The vegetation index was constructed based on these reflectances to enhance the physiological characteristics of vegetation.

[0021] Preferably, in step S11, 20 reference points of a uniform white plate are selected to perform geometric correction on the image, and the reflectance of the reference points is converted into the reflectance of the multispectral image by using a pseudo-standard geometric correction method, so as to extract the reflectance of the accurate multispectral band. The correction formula is as follows:

[0022]

[0023] Among them, DN i is the reflectance value of the uniform white plate area; is the original mean before correction; is the average value of a uniform white board area; DN i,corrected is the corrected reflectance value.

[0024] Preferably, in step S2 vegetation index feature selection, the Pearson feature selection method is used to measure the linear relationship between vegetation index and potato canopy SPAD, and the Pearson correlation coefficient r X,Y The calculation formula is as follows:

[0025]

[0026] Where n is the number of samples; x i and i is the i-th data value of variables X and Y; and are the means of variables X and Y respectively.

[0027] Preferably, in step S2 vegetation index feature selection, a competitive adaptive reweighted sampling algorithm is used to evaluate the relevance of each vegetation index to the potato canopy SPAD prediction model through multiple sampling and analysis. The specific process is as follows:

[0028] Step S221, Monte Carlo sampling: Assume that the original vegetation index data set has n samples, the number of samples for each sampling is m (m≤n), and the number of sampling times is k; for the i-th sampling (i=1,2,…,k), the obtained sampling subset S i Contains m samples randomly drawn from the original data set;

[0029] Step S222: Establish a partial least squares regression PLS model: for the subset S obtained by the i-th sampling i , establish the PLS model; let X i is the vegetation index matrix m×p in the subset, where p is the number of vegetation indices; Y i is the corresponding target variable vector m×1; the PLS model is based on X i and Y i Decompose and regress to obtain the regression coefficient vector β i , as shown below:

[0030] X=TP T +E (3);

[0031] Y=UQ T +F (4);

[0032] Among them, T and U are the score matrices of X and Y respectively; P and Q are the loading matrices; E and F are the residual matrices;

[0033] Regression coefficient β i , as shown below:

[0034] β i =W(P T W) -1 (5);

[0035] Among them, W is the weight matrix, which is determined during the iteration process of the PLS algorithm;

[0036] Step S223, calculating the vegetation index importance index;

[0037] Assume ω i,j is the weight of the jth vegetation index in the PLS model of the i-th sampling, and the importance index I of the jth vegetation index is calculated. j , as shown below:

[0038]

[0039] Where n is the number of sampling times; a is the empirical index, where a=1 is set, and the vegetation index is selected according to the importance index.

[0040] Preferably, in step S3, a chlorophyll content inversion model is constructed based on machine learning, and the specific process is as follows:

[0041] Step S31, random forest RF: by establishing an RF model, an inversion experiment is conducted on the chlorophyll content of potato canopy, and a ten-fold cross-validation method is used to continuously change the selection of the test set so that each subset has a chance to become a test set, and the performance of the model on different data subsets is evaluated;

[0042] Different optimization algorithms were used to optimize the parameters of the RF model. In each round of cross-validation, different parameter combinations were used to train the model, and the model performance was evaluated on the test set to obtain an ideal chlorophyll content inversion model.

[0043] Step S32, support vector regression SVR: establish an SVR model to conduct an inversion study on the chlorophyll content of potato canopy, and use the RBF kernel function to fit the data. Its core expression is as follows:

[0044] K(x i ,x j ) = exp(-γ||x i -x j || 2 ) (7);

[0045] Among them, x i and x j is the input sample; γ is the parameter of the kernel function, which determines the distribution range of the sample in the feature space;

[0046] During the training process, the parameter optimization algorithm is used to continuously optimize the parameters according to the training data, and different hyperparameter combinations are evaluated to achieve chlorophyll content prediction and obtain the chlorophyll content inversion model;

[0047] Step S33, extreme gradient boosting XGBoost: the potato chlorophyll data set is divided into a training set and a test set in a ratio of 8:2. In the training phase, the XGBoost model uses the training set as the basis, the input features and the target variable, and uses the gradient boosting strategy to iteratively train the decision tree;

[0048] In each iteration, the XGBoost model first calculates the gradient of the loss function with respect to the predicted value, and then fits a new decision tree based on the obtained gradient information. The newly generated decision tree learns and corrects the errors made by the previous decision tree during the prediction process, gradually improving the prediction performance and accuracy of the entire model, and obtaining the optimal chlorophyll content inversion model.

[0049] Preferably, in step S4, the parameters of the RF, SVR and XGBoost chlorophyll content inversion models are optimized based on the Gray Wolf Optimization Algorithm to invert the chlorophyll content of the potato canopy, and the specific steps are as follows:

[0050] Step S411, inputting drone image data and measured potato canopy chlorophyll content;

[0051] Step S412: using Pearson correlation analysis method and CARS algorithm to perform feature selection on the constructed vegetation index;

[0052] Step S413, establishing RF, SVR and XGBoost chlorophyll inversion models;

[0053] Step S414, initializing the population and parameters of the gray wolf, and defining the fitness function of the gray wolf optimization algorithm to evaluate the pros and cons of each parameter combination;

[0054] Step S415, dividing the gray wolf individuals into social classes and calculating the distances between them to determine the update direction and position, and update the optimal position and fitness of the gray wolf;

[0055] Step S416, determine whether the stop condition is met, if the condition is met, execute step S417, otherwise execute step S415;

[0056] Step S417, outputting the optimal parameters of RF, SVR and XGBoost chlorophyll inversion models;

[0057] Step S418: training the chlorophyll inversion model, obtaining the result of the chlorophyll content inversion model, and evaluating the model through evaluation indicators.

[0058] Preferably, in step S4, the parameters of the RF, SVR and XGBoost chlorophyll inversion models are optimized based on the sparrow optimization algorithm, and the chlorophyll content is inverted. The specific steps are as follows:

[0059] Step S421, inputting drone image data and measured chlorophyll content data;

[0060] Step S422, using Pearson and CARS algorithms to select and construct vegetation indices;

[0061] Step S423, establishing RF, SVR and XGBoost chlorophyll content inversion models;

[0062] Step S424, initializing the position of the sparrow population and defining a fitness function;

[0063] Step S425, the fitness of each sparrow is calculated according to the fitness function, the sparrow with the best fitness is regarded as the discoverer, leading the search direction of the population, and the remaining sparrows are regarded as followers, and the followers update their positions according to the positions of the discoverer and themselves;

[0064] Step S426: randomly select a part of sparrows as scouts to explore the new search area and update their own positions, thereby updating the best position and fitness of the global sparrows;

[0065] Step S427, determine whether the termination condition is met, if the condition is met, execute step S428, otherwise execute step S425;

[0066] Step S428, output the best model parameters optimized by the sparrow algorithm;

[0067] Step S429: training the potato chlorophyll content inversion model, outputting the result of the chlorophyll content inversion model, and evaluating the accuracy of the model.

[0068] Preferably, in step S5, the chlorophyll content inversion model optimized in step S4 is evaluated for accuracy, and the specific process is as follows:

[0069] The data set was randomly divided into training set and test set with a ratio of 8:2. Ten-fold cross validation was used to evaluate the generalization ability of the model during training. The coefficient of determination R was used to 2 , root mean square error RMSE and mean absolute error MAE parameters are used to measure the accuracy of the model. The calculation formula is as follows:

[0070]

[0071]

[0072]

[0073] Where n is the number of samples; x i and i is the measured value; is the predicted value; and is the average of the measured values.

[0074] Therefore, the present invention adopts the above-mentioned canopy chlorophyll content inversion method based on feature extraction and machine learning model, uses data from different growth periods, multiple feature extraction methods and different optimization algorithms to optimize and improve the canopy SPAD inversion model, and finds that different feature extraction methods have significant differences in inversion accuracy in different growth periods. The CARS algorithm is better at extracting vegetation index in two single growth periods, and the Pearson correlation coefficient method in the entire growth period has higher accuracy. The accuracy of the optimization models using different optimization algorithms is significantly improved, and the SSA algorithm is better in each growth period. These application methods significantly improve the applicability and accuracy of canopy SPAD inversion research in different growth periods of potatoes.

[0075] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a flow chart of the canopy chlorophyll content inversion method based on feature extraction and machine learning model of the present invention;

[0077] Figure 2 It is the drone image preprocessing process of the present invention;

[0078] Figure 3 It is the flow chart of the gray wolf optimization algorithm of the present invention;

[0079] Figure 4 It is the flow chart of the sparrow optimization algorithm of the present invention;

[0080] Figure 5 is a descriptive statistical graph of SPAD of potato canopies at different growth stages in an embodiment of the present invention;

[0081] Figure 6 is the CARS selection result of the RF model of the original image in the mature stage in the embodiment of the present invention;

[0082] Figure 7 : It is a scatter plot of the SPAD actually measured at the seedling stage in the embodiment of the present invention and the SPAD predicted by the random forest (RF) model; wherein, (a)-(f) are scatter plots corresponding to the random forest (RF) model; (g)-(l) are scatter plots corresponding to the support vector regression (SVR) model; (m)-(r) are scatter plots corresponding to the extreme gradient boosting (XGBoost) model;

[0083] Figure 8 : are scatter plots of the SPAD actually measured in the mature stage and the SPAD predicted by each model in the embodiment of the present invention; wherein, (a)-(f) are scatter plots corresponding to the random forest (RF) model; (g)-(l) are scatter plots corresponding to the support vector regression (SVR) model; (m)-(r) are scatter plots corresponding to the extreme gradient boosting (XGBoost) model;

[0084] Fig. 9 It is a scatter plot of the SPAD actually measured during the whole growth period in the embodiment of the present invention and the SPAD predicted by each model; among them, (a)-(f) are scatter plots corresponding to the random forest (RF) model; (g)-(l) are scatter plots corresponding to the support vector regression (SVR) model; (m)-(r) are scatter plots corresponding to the extreme gradient boosting (XGBoost) model. DETAILED DESCRIPTION

[0085] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0086] like Figure 1 As shown in the figure, the canopy chlorophyll content inversion method based on feature extraction and machine learning model includes the following steps:

[0087] Step S1: using an unmanned aerial vehicle (UAV) equipped with a multispectral camera and ground measurement to collect and process the potato chlorophyll content data in the study area, and construct a vegetation index.

[0088] Step S11: The unmanned aerial vehicle (UAV) carries a multispectral camera to acquire and process image data.

[0089] First, a drone was used to collect image data of the study area, with a take-off weight of 1.356 kg, a maximum horizontal flight speed of 20 m / s, and a maximum take-off altitude of 6000 m. The UAV was equipped with four multispectral sensor probes, corresponding to the spectral band ranges of green light (GREEN), red light (RED), red edge light (REG), and near infrared light (NIR), and the image sensor had an effective pixel of 12 million.

[0090] Data collection was performed in clear and windless weather to prevent the image from being affected by strong light and temperature. The drone took off from the same preset location each time. A calibration standard plate was placed before each takeoff to collect accurate image data. The height was set to 30m, the moving speed was set to 5m / s, the lateral angle was 70, and the heading was 80.

[0091] Then, the collected images are input into Pix4Dmapper software for multispectral image synthesis and calibration, so as to obtain orthorectified multispectral reflectance images. The UAV image preprocessing process is as follows: Figure 2 shown.

[0092] 20 reference points of a uniform white plate are selected to perform geometric correction on the image. The reflectance of the reference points is converted into the reflectance of the multispectral image using the pseudo-standard geometric correction method, which is convenient for extracting the accurate reflectance of the multispectral band. The correction formula is as follows:

[0093]

[0094] Among them, DN i is the reflectance value of the uniform white plate area; is the original mean before correction; is the average value of a uniform white board area; DN i,corrected is the corrected reflectance value.

[0095] Step S12: collecting and measuring ground data.

[0096] Firstly, the study area was divided into 16 plots, each plot corresponding to one potato variety. A total of 16 ground data of potato varieties were collected. The area of ​​each plot was 20m×100m, and data from 10 sample points were evenly collected in each plot. The chlorophyll content of leaf canopy was collected from the potato planting base in the study area at the seedling stage and mature stage respectively. A total of 160 ground data were collected in each growth period.

[0097] Then, the correlation between the image data obtained by the UAV and the measured data on the ground was studied. After the UAV image collection, the 16 sample plots divided above were taken as the objects, and 10 well-growing potato canopies were evenly selected as sample points in each sample plot. The veins were avoided when measuring each leaf canopy. A handheld chlorophyll meter was used to evenly collect 10 sample points in the mesophyll, and the average value was taken as the true value of a sample point. The readings obtained from the potato plants selected at the sample points of each sample plot represented the relative chlorophyll content of the sample plot (Soil-Plant Analysis Development, SPAD).

[0098] Step S13: constructing a vegetation index based on the above acquired and processed data.

[0099] Vegetation index can significantly reflect the changes in the real physiological characteristics of vegetation. Generally, the potato canopy image data collected by drones will be affected by factors such as atmospheric interference, light intensity, and internal physiological characteristics. The vegetation index can significantly reduce these influences. Therefore, 18 vegetation indices were initially selected, as shown in Table 1.

[0100] Table 1 18 vegetation indices used for SPAD estimation in the present invention

[0101] Vegetation Index name formula GRVI Green-Red Vegetation Index GRVI=(GR) / (G+R) MCARI Modified Chlorophyll Absorption Ratio Index MCARI=(REG-R)-(0.2*(REG-G))*(REG / R) DVI Difference Vegetation Index DVI=NIR-R MTVI Modified Triangular Vegetation Index 1.5*(1.2*(REG-G)-2.1*(RG)) WDV Wide Dynamic Range Vegetation Index WDRVI=(0.12*NIR-R) / (0.12*NIR+R) EVI2 Two-band enhanced vegetation index EVI2=2.5*(NIR-R) / (NIR+2.4*R+1) RECI Red Edge Chlorophyll Index RECI=(NIR / REG)-1 GCI Green Chlorophyll Index GCI=(NIR / G)-1 NDVI Normalized Difference Vegetation Index NDVI = (NIR-R) / (NIR+R) GNDVI Green Normalized Difference Vegetation Index GNDVI=(NIR-G) / (NIR+G) RVI Ratio Vegetation Index RVI=NIR / R NDGI Normalized Difference Greenness Index NDGI=(REG-G) / (REG+G) MSRI Modified Simple Ratio Index MSR = (NIR / R-1) / (NIR / R+1) OSAVI Optimizing Soil Adjustment Vegetation Index OSAVI=(NIR-R) / (NIR+R+0.16) SRI Simple ratio index SR=NIR / REG NDRE Normalized difference red edge index NDRE=(NIR-REG) / (NIR+REG) NLI Nonlinear vegetation index NLI = (NIR*NIR-R) / (NIR*NIR+R) TVI Triangular Vegetation Index TVI=0.5*(120*(NIR-REG)-200*(R-REG))

[0102] Among them, G represents the green band value, R represents the red band value, REG represents the red edge band value, and NIR represents the near infrared band value.

[0103] The orthophoto images obtained from the UAV after geometric correction were divided into 16 corresponding areas according to potato varieties. Ten sample points were taken in each area. Combined with the longitude and latitude coordinates of the ground measurement points, the reflectance of the four bands of GREEN, RED, REG and NIR corresponding to each sample point was extracted on ArcGIS. The vegetation index was constructed based on these reflectances to enhance the physiological characteristics of vegetation.

[0104] Step S2: Use Pearson correlation analysis and competitive adaptive reweighted sampling (CARS) algorithm to perform feature selection on the index.

[0105] Since characteristic variables such as vegetation indices usually show complexity and diversity, which leads to information redundancy and increases research errors, it is necessary to select the initially selected vegetation characteristics to determine the optimal variable characteristics for inverting potato chlorophyll content.

[0106] Step S21: Pearson feature selection method.

[0107] The Pearson feature selection method is a statistical indicator method based on the Pearson correlation coefficient to measure the degree of linear correlation between two variables. In vegetation index selection, it can be used to measure the linear relationship between vegetation index and potato canopy SPAD. If the Pearson correlation coefficient between the vegetation index and the SPAD measured in the potato canopy is close to +1 or -1, it means that there is a strong linear relationship between them. This vegetation index is of great value in estimating SPAD. The Pearson correlation coefficient r X,Y The calculation formula is as follows:

[0108]

[0109] Where n is the number of samples; x i and i is the i-th data value of variables X and Y; and are the means of variables X and Y respectively.

[0110] The vegetation index is used as the free variable X and the potato canopy SPAD is used as the target variable Y to calculate, and then the r X,Y The vegetation index is selected by the value of X,Y When the absolute value of is greater than a certain threshold, the variable is considered to be significantly correlated and the corresponding vegetation index is selected.

[0111] In the present invention, the threshold value selected in the seedling stage and the mature stage is 0.3, and the threshold value selected in the whole growth period is 0.7, and the selected vegetation index is used for research.

[0112] Step S22: Competitive Adaptive Reweighted Sampling (CARS) algorithm.

[0113] CARS is a feature selection method that combines Monte Carlo sampling and partial least squares regression (PLS). When selecting vegetation indices, the core idea is to evaluate the relevance of each vegetation index to the potato canopy SPAD prediction model through multiple sampling and analysis. It simulates a "competition" process, where each vegetation index fights for the opportunity to be retained in the model, selects vegetation indices based on correlation indicators, eliminates feature variables with weak relationships, and thus selects vegetation indices with strong correlations. The specific process is as follows:

[0114] Step S221: Monte Carlo sampling.

[0115] Assume that the original vegetation index data set has n samples, the number of samples for each sampling is m (m≤n), and the number of sampling times is k. For the i-th sampling (i=1,2,…,k), the obtained sampling subset S i It contains m samples randomly selected from the original data set. This process is mainly achieved through a random sampling algorithm.

[0116] Step S222: Establish a partial least squares regression (PLS) model.

[0117] For the subset S obtained by the i-th sampling i , establish the PLS model. Let X i is the vegetation index matrix in the subset (m×p, p is the number of vegetation indices), Y i is the corresponding target variable vector (m×1). The PLS model is constructed by i and Y i Decompose and regress to obtain the regression coefficient vector β i This process involves an iterative decomposition using a complex PLS algorithm, as shown below:

[0118] X=TP T +E (3);

[0119] Y=UQ T +F(4);

[0120] Among them, T and U are the score matrices of X and Y respectively; P and Q are the loading matrices; E and F are the residual matrices.

[0121] Regression coefficient β i , as shown below:

[0122] β i =W(P T W) -1 (5);

[0123] Among them, W is the weight matrix, which is determined during the iteration process of the PLS algorithm.

[0124] Step S223: Calculate the vegetation index importance index.

[0125] Assume ω i,j is the weight (or regression coefficient or other related indicators) of the jth vegetation index in the PLS model of the i-th sampling, and the importance index I of the jth vegetation index is calculated. j , as shown below:

[0126]

[0127] Where n is the number of sampling times; a is the empirical index, where a=1 is set, and the vegetation index is selected according to the importance index.

[0128] Step S3: construct a chlorophyll content inversion model based on machine learning.

[0129] Step S31: Random Forest (RF).

[0130] Random Forest (RF) is an integrated learning algorithm that constructs multiple independent decision trees. When all decision trees are constructed, for a new sample, it is input into each decision tree for prediction, and then the predicted values ​​are averaged to obtain the final chlorophyll content prediction result.

[0131] In the present invention, an inversion experiment on the chlorophyll content of potato canopy is carried out by establishing an RF model, and a ten-fold cross-validation method is used to continuously change the selection of the test set so that each subset has a chance to become a test set, thereby comprehensively evaluating the performance of the model on different data subsets.

[0132] The parameters of the RF model were optimized using different optimization algorithms. In each round of cross-validation, different parameter combinations were used to train the model, and the model performance was evaluated on the test set to obtain an ideal chlorophyll content inversion model. The tuning parameters of the RF model in the present invention are shown in Table 2.

[0133] Step S32: Support Vector Regression (SVR).

[0134] Support Vector Regression (SVR) is a regression algorithm based on the principle of Support Vector Machine (SVM). The key to this model lies in the selection of kernel function. Commonly used kernel functions include linear kernel function, polynomial kernel function, radial basis function (RBF), etc.

[0135] In the present invention, an SVR model is established to invert the chlorophyll content of potato canopy, and the RBF kernel function is used to fit the data. The core expression is as follows:

[0136] K(x i ,x j ) = exp(-γ||x i -x j || 2 ) (7);

[0137] Among them, x i and x j is the input sample; γ is the parameter of the kernel function, which determines the distribution range of the sample in the feature space.

[0138] During the training process, a parameter optimization algorithm is used to continuously optimize parameters according to the training data, and different hyperparameter combinations are evaluated so that the model can more accurately predict the chlorophyll content. The parameters tuned by the SVR model in the present invention are shown in Table 2.

[0139] Step S33: Extreme Gradient Boosting (XGBoost).

[0140] Extreme Gradient Boosting (XGBoost) is a machine learning algorithm based on gradient-boosted decision trees (GBDT). The XGBoost model has a significant improvement in learning efficiency. Its core idea is to gradually add new learners (usually decision trees) on the basis of the previous round of models by minimizing the loss function to improve the predictive ability of the overall model.

[0141] In the present invention, the potato chlorophyll data set is divided into a training set and a test set in a ratio of 8:2. In the training phase, the XGBoost model uses the gradient boosting strategy to iteratively train the decision tree based on the training set, input features and target variables. In each iteration, the XGBoost model first calculates the gradient of the loss function with respect to the predicted value, and then fits a new decision tree based on the obtained gradient information, with the aim of effectively reducing the loss value. The newly generated decision tree can learn and correct the errors generated by the previous decision tree in the prediction process, thereby gradually improving the prediction performance and accuracy of the entire model.

[0142] Since the XGBoost model has many parameters, the present invention uses two optimization algorithms to optimize its parameters to obtain the optimal chlorophyll content inversion model. The parameters of the XGBoost model improved in the present invention are shown in Table 2.

[0143] Table 2 Parameters and meanings of the RF, SVR and XGBoost models

[0144]

[0145]

[0146] Step S4: Optimize the chlorophyll content inversion machine learning model constructed in step S3 based on the gray wolf optimization algorithm and the sparrow optimization algorithm.

[0147] Step S41: Grey wolf optimization algorithm performs parameter optimization.

[0148] GreyWolfOptimizer (GWO) is a heuristic optimization algorithm based on the predation behavior of grey wolf groups. Grey wolf groups have a strict hierarchy, and in the algorithm, these hierarchies are used to guide the search direction of the wolf group.

[0149] In the present invention, the position of the gray wolf population is randomly initialized. The position of each gray wolf represents a set of parameters of the machine learning model. According to the selected evaluation index and machine learning model, the fitness function is defined, and the wolf group is graded and the position is continuously updated. After the position is updated, the fitness and relative optimal solution of each gray wolf are repeatedly calculated. Each solution can be regarded as the position of a wolf. The wolf group continuously updates its position to approach the optimal parameters. The flowchart of the gray wolf optimization algorithm to improve the chlorophyll content inversion model is as follows: Figure 3 As shown in the figure, by improving the model, the accuracy of RF, SVR and XGBoost models is improved, and their complex parameter problems are also significantly improved.

[0150] The Gray Wolf Optimization Algorithm optimizes the parameters of the RF, SVR, and XGBoost models. The specific steps for SPAD inversion of potato canopy are as follows:

[0151] Step S411, inputting drone image data and measured potato canopy chlorophyll content.

[0152] Step S412: Use the Pearson correlation analysis method and the CARS algorithm to perform feature selection on the constructed vegetation index.

[0153] Step S413: Establish RF, SVR and XGBoost chlorophyll inversion models.

[0154] Step S414: Initialize the population and parameters of the gray wolf, and define the fitness function of the gray wolf optimization algorithm to evaluate the pros and cons of each parameter combination.

[0155] Step S415: divide the gray wolf individuals into social classes and calculate the distances between them to determine the update direction and position, and update the optimal position and fitness of the gray wolf.

[0156] Step S416: determine whether the stop condition is met, if the condition is met, execute step S417, otherwise execute step S415.

[0157] Step S417: output the optimal parameters of the RF, SVR and XGBoost models.

[0158] Step S418: train the inversion model to obtain the result of the chlorophyll content inversion model, and evaluate the model through evaluation indicators.

[0159] Step S42: Optimize parameters using the sparrow optimization algorithm.

[0160] Sparrow Search Algorithm (SSA) is an optimization algorithm proposed to simulate the foraging and anti-predation behavior of sparrow groups.

[0161] In the chlorophyll inversion model of the present invention, a group of "sparrows" is randomly initialized, that is, corresponding to a group of model parameter combinations, and sparrows with different roles such as discoverers, followers and scouts are set. The discoverer is responsible for exploring the search space, and the follower conducts a local search around the discoverer. At the same time, a warning mechanism is introduced. When encountering danger, some sparrows will randomly move to other locations to avoid falling into the local optimum. The scout is responsible for exploring the information of the new area and providing updated information on the population position to improve the search efficiency of the population, thereby optimizing and improving the parameter accuracy of the entire chlorophyll content inversion model. The flow chart of the SSA algorithm to improve the chlorophyll content inversion model is as follows: Figure 4 shown.

[0162] The specific steps of using the Sparrow Optimization Algorithm to optimize the parameters of the RF, SVR, and XGBoost models and invert the chlorophyll content are as follows:

[0163] Step S421, inputting drone image data and measured chlorophyll content data.

[0164] Step S422: Use the Pearson and CARS algorithms to select and construct vegetation indices.

[0165] Step S423: Establish RF, SVR and XGBoost chlorophyll content inversion models.

[0166] Step S424: Initialize the position of the sparrow population and define a fitness function.

[0167] Step S425, the fitness of each sparrow is calculated according to the fitness function, the sparrow with the best fitness is regarded as the discoverer, which leads the search direction of the population, and the remaining sparrows are regarded as followers, which update their positions according to the discoverer and their own positions.

[0168] Step S426: Randomly select a part of sparrows as scouts to explore the new search area and update their own positions, thereby updating the best position and fitness of the global sparrows.

[0169] Step S427, determine whether the termination condition is met, if the condition is met, execute step S428, otherwise execute step S425.

[0170] Step S428: output the best model parameters optimized by the sparrow algorithm.

[0171] Step S429: training the potato chlorophyll content inversion model, outputting the result of the chlorophyll content inversion model, and evaluating the accuracy of the model.

[0172] Step S5: perform accuracy evaluation on the chlorophyll content inversion model optimized in step S4.

[0173] The data set is randomly divided, and the ratio of training set to test set is 8:2. In order to enhance the performance of the model and prevent overfitting, the ten-fold cross-validation method is used during training to evaluate the generalization ability of the model, which can more comprehensively utilize the data to evaluate the performance of the model.

[0174] In each test, the accuracy and other evaluation indicators are calculated, and the average value of the evaluation indicators is calculated after multiple tests to evaluate the generalization ability of the model. 2 , root mean square error RMSE and mean absolute error MAE parameters to measure the accuracy of the model. 2 The closer it is to 1, the better the stability and accuracy of the model, and the higher the degree of fit; the smaller the RMSE value, the higher the prediction accuracy of the model; MAE is used to test the prediction ability of the model, and a lower MAE value indicates that the model has better prediction accuracy; its calculation formula is as follows:

[0175]

[0176]

[0177]

[0178] Where n is the number of samples; x i and i is the measured value; is the predicted value; and is the average of the measured values.

[0179] Example

[0180] The potato planting base research area of ​​this embodiment is located at 41°9'36"N, 111°36'10"E. Figure 1 As shown, it belongs to the temperate continental monsoon climate zone, and is also a combination of plains, hills, mountains, hills and plains, providing a variety of soil types and climate environments. Influenced by the southeast monsoon in summer and the Siberian high pressure in winter, the annual average temperature is around 6°C, the annual sunshine time exceeds 3,000 hours, and the annual precipitation can reach 360 mm. These specific geographical and climatic characteristics provide good light conditions for the growth of potatoes in the region.

[0181] The study area was divided into 16 plots, each plot corresponding to one variety, and ground data of 16 potato varieties were collected. The area of ​​each plot was divided into 20m×100m, and data from 10 sample points were collected evenly in each plot, and 160 ground data were collected in each growth period. The drone collected data between 8:30 and 11:30 a.m. on July 10 and between 9:00 and 11:00 a.m. on August 12 in clear and windless weather conditions to prevent the image from being affected by strong light and temperature; during the actual ground collection and measurement, the canopy chlorophyll content was collected at the potato planting base in the study area at the seedling stage (July 10, 2024) and the maturity stage (August 12, 2024).

[0182] The research image data was obtained by using a multispectral camera mounted on a drone, and the multispectral images were spliced ​​using Pix4Dmapper software. The orthophoto images were then corrected and imported into the ENVI5.3 software. The collected data points were annotated and used as the region of interest (RIO), which was the sampling point of the chlorophyll content closest to the measured potato canopy, so as to extract the longitude and latitude information of the points of interest. Then, the longitude and latitude were entered into the ArcGIS software and the reflectivity of the four bands of GREEN, RED, REG and NIR was extracted. Based on this, the vegetation characteristic index was constructed to reflect the structural changes and serve as the characteristic variable of the model input.

[0183] 1. SPAD descriptive statistics of potato canopy.

[0184] The values ​​of potato canopy SPAD in different growth stages are slightly different, as shown in Table 3. The canopy SPAD in the seedling stage ranged from 40.30 to 56.00, with an average value of 48.47, a standard deviation of 3.70, and a coefficient of variation of 7.63%; the canopy SPAD in the mature stage ranged from 28.20 to 54.20, with an average value of 43.20, a standard deviation of 5.40, and a coefficient of variation of 12.50%; the canopy SPAD in the full growth period ranged from 28.20 to 56.00, with an average value of 45.84, a standard deviation of 5.33, and a coefficient of variation of 11.63%.

[0185] Table 3 Descriptive statistics of SPAD of potato canopy at different growth stages

[0186] Growth period Number of samples Minimum Maximum average value Standard Deviation Coefficient of variation Seedling stage 162 40.30 56.00 48.47 3.70 7.63％ Maturity 162 28.20 54.20 43.20 5.40 12.50％ Whole reproductive period 324 28.20 56.00 45.84 5.33 11.63％

[0187] The difference in coefficient of variation reflects the changes in SPAD of potato canopy at different developmental stages. It is precisely because of these differences that the feasibility and necessity of studying the inversion of SPAD of potato canopy at different developmental stages are established. The descriptive statistics of SPAD of potato canopy at different growth stages are as follows: Figure 5 shown.

[0188] 2. Selection of vegetation index.

[0189] Orthophoto fusion and correction were performed on UAV images of the seedling stage, maturity stage and full growth period to remove noise and redundant information, and the orthophoto images for research were obtained. The characteristic vegetation indices constructed by corresponding to the measured ground SPAD in the three growth periods and each band were selected using the Pearson correlation analysis method and the CARS algorithm.

[0190] (1) Pearson correlation analysis method.

[0191] For Pearson correlation analysis, different development stages have different correlation characteristics. From the seedling stage to the mature stage, the correlation gradually decreases. This is related to the characteristics of SPAD and vegetation index. The correlation coefficients between each vegetation index and SPAD at each development stage are shown in Table 4.

[0192] Table 4 Pearson correlation coefficient between chlorophyll content and vegetation index

[0193]

[0194] During the seedling stage, the vegetation indices with the strongest correlation with SPAD were EVI2, NDVI, and OSAVI, with correlation coefficients of 0.40. There were 10 indices with correlations greater than the threshold of 0.3. The vegetation features selected in the dataset were: WDRVI, EVI2, RECI, NDVI, GNDVI, MSRI, OSAVI, SRI, NDRE, and NLI.

[0195] In the mature stage, the vegetation index with the strongest correlation with SPAD is TVI, with a correlation coefficient of 0.35. There are 7 vegetation features with correlation greater than the threshold of 0.3. The vegetation features selected in the dataset are: DVI, EVI2, NDVI, GNDVI, MSRI, OSAVI and TVI.

[0196] During the entire growth period, the vegetation index with the strongest correlation with SPAD is WDRVI, with a correlation coefficient of 0.80. There are 10 indexes with correlations greater than the threshold of 0.7. The vegetation features selected in the dataset are: GRVI, DVI, MTVI, WDRVI, EVI2, NDVI, NDGI, MSRI, OSAVI and TVI.

[0197] The above coefficients are the correlation coefficients between vegetation index and SPAD at the significance level of 0.01, which are of significant reliability. In summary, the correlation of the whole growth period has been significantly improved compared with that of the single growth period. This is because the characteristics of the whole growth period can capture the seasonal laws and periodic changes in time series data, thus improving the correlation between the two.

[0198] (2)CARS algorithm.

[0199] For the CARS algorithm, an exponential decay function is used to select variables, and the performance of the variable subset obtained by each sampling is evaluated through the cross-validation method. It can quickly select variables that have important contributions to the model from a large number of feature variables to improve the predictive ability and efficiency of the model.

[0200] In this embodiment, the number of Monte Carlo (MC) samples is set to 50, and the number of extracted feature variables is determined based on the minimum value of the root mean square error (RMSE). Taking the CARS-RF model of the original image of the mature stage as an example, Figure 6 As shown in the figure, the number of selected feature variables gradually decreases, and the change speed is faster in the early stage and slower in the later stage. There are two stages: "primary selection" and "fine selection". This is mainly due to the characteristics of the algorithm. Figure 6It can be seen that as the number of Monte Carlo samples gradually increases, the value of the root mean square error also changes continuously. When the number of samples is 35, the RMSE reaches the minimum value of 3.12. Four characteristic variables are selected, corresponding to the vegetation index MCARI, DVI, GCI and NLI. The vegetation index of the corresponding model of each growth period obtained after selection by the CARS algorithm is shown in Tables 5, 6 and 7.

[0201] Table 5 CARS algorithm selection results at seedling stage

[0202]

[0203] Table 6 Results of CARS algorithm selection in mature stage

[0204]

[0205]

[0206] Table 7 CARS algorithm selection results for the entire growth period

[0207]

[0208]

[0209] In summary, the best vegetation index for model research was selected by two feature selection methods, Pearson correlation analysis and CARS algorithm. According to Table 5, most of the vegetation indices selected by different optimization models using the CARS feature selection method in each growth period contain the near infrared band (NIR), because in this band, the reflection, absorption and scattering characteristics of objects to light will produce specific spectral characteristics, which is conducive to the inversion study of potato canopy SPAD.

[0210] For the two single growth periods, the seedling stage and the mature stage, the vegetation features selected by the CARS algorithm are better than those by the Pearson correlation analysis method. This is because the CARS algorithm can better explore the relationship between variables in a single growth period and has a higher sensitivity to features that are important indicators of vegetation growth status. The Pearson algorithm usually performs linear correlation analysis on the overall data. Because the amount of data in a single growth period is small, the Pearson method is not as sensitive to such local, smaller changes as the CARS algorithm.

[0211] In the long time series data of the whole growth period, the results of the Pearson algorithm have high stability and interpretability. The linear relationship between the vegetation index and the canopy SPAD can intuitively reflect the growth trend of vegetation throughout the growth period. The CARS algorithm will be affected by factors such as overfitting during the whole growth period. It will pay too much attention to the variable selection of a certain stage and ignore the important variable information of other growth stages, resulting in its vegetation index extraction effect during the whole growth period being inferior to that of the Pearson algorithm.

[0212] 3. Select the best optimization model for inverting potato canopy SPAD.

[0213] (1) The best model for inverting canopy SPAD during the seedling stage.

[0214] This embodiment is based on the multispectral image bands of potato seedlings, maturity and full growth period (seedlings and maturity), constructs vegetation index to invert its canopy SPAD, uses Pearson correlation analysis and CARS algorithm to select vegetation index to remove redundant and interfering feature information, and uses gray wolf algorithm and sparrow algorithm to optimize and improve three machine learning models respectively, combines the two feature selection methods with the model improved by the optimization algorithm, and compares the accuracy with the original model. The results are shown in Table 8.

[0215] Table 8 SPAD inversion results of potato canopy during seedling stage

[0216]

[0217]

[0218] For the Pearson correlation coefficient method, the GWO-RF model has the highest inversion accuracy, and its test set determination coefficient R 2 The value of root mean square error (RMSE) is 2.75, and the value of mean absolute error (MAE) is 2.05. For the CARS algorithm, the SSA-RF model has the highest inversion accuracy, and its test set determination coefficient R 2 The SPAD values ​​of the actual measured values ​​in the seedling stage and the SPAD predicted by each model are shown in the following figure. Figure 7 shown.

[0219] Therefore, in the seedling stage, the effect of the CARS feature extraction method is slightly better than that of the Pearson correlation analysis method, and the model with the highest inversion accuracy is the CARS-SSA-RF model.

[0220] (2) The best model for inverting canopy SPAD at maturity.

[0221] During the potato maturity period, the previously selected vegetation index was input into different machine learning models as feature variables, and the models were optimized and improved using the gray wolf optimization algorithm and the sparrow optimization algorithm. The results are shown in Table 9.

[0222] Table 9 SPAD inversion results of potato canopy at maturity

[0223]

[0224]

[0225] For the Pearson correlation coefficient method, the SSA-RF model has the highest inversion accuracy, and its test set determination coefficient R 2 The value of root mean square error (RMSE) is 3.59, and the value of mean absolute error (MAE) is 2.97. For the CARS algorithm, the SSA-XGBoost model has the highest inversion accuracy, and its test set determination coefficient R 2 It reached 0.66, the root mean square error RMSE value was 3.23, and the mean absolute error MAE value was 2.75. The scatter plot of the actual SPAD measured in the mature stage and the SPAD predicted by each model is as follows: Figure 8 shown.

[0226] Therefore, in the mature stage, the effect of the CARS feature extraction method is also slightly higher than that of the Pearson correlation analysis method, and the model with the highest inversion accuracy is the CARS-SSA-XGBoost model.

[0227] (3) The optimal model for inverting canopy SPAD during the entire growth period.

[0228] In order to study the growth process of potato throughout its entire growth period and thus evaluate the situation of potato throughout its entire growth cycle, it is necessary to fuse the data of the two growth periods and combine the characteristics of the vegetation index with the optimization algorithm to invert the potato SPAD. The results are shown in Table 10.

[0229] Table 10 SPAD inversion results of potato canopy during the whole growth period

[0230]

[0231]

[0232] For the Pearson correlation coefficient method, the SSA-XGBoost model has the highest inversion accuracy, and its determination coefficient R 2The value of root mean square error (RMSE) is 2.39, and the value of mean absolute error (MAE) is 1.99. For the CARS algorithm, the SSA-XGBoost model has the highest inversion accuracy, and its coefficient of determination R 2 The SPAD values ​​of the whole growth period are actually measured and the SPAD values ​​predicted by each model are shown in the figure below. Fig. 9 shown.

[0233] Therefore, during the entire growth period, the features selected by the Pearson correlation analysis method were better than those selected by the CARS algorithm, and the model with the highest inversion accuracy was the Pearson-SSA-XGBoost model.

[0234] In summary, in the present invention, two optimization algorithms combined with two feature selection methods were used to study three machine learning models, and multiple groups of comparative studies were set up, which is more reliable and convincing. The results show that in the seedling stage, the best results were obtained using the sparrow optimization algorithm (SSA), and the model with the highest research accuracy was the CARS-SSA-RF model, with a determination coefficient R 2 The value of the root mean square error (RMSE) was 2.63, and the value of the mean absolute error (MAE) was 2.00. In the mature stage, the best effect was achieved by using the sparrow optimization algorithm. The model with the highest accuracy was the CARS-SSA-XGBoost model, and the determination coefficient R 2 The value of the root mean square error (RMSE) was 3.23, and the value of the mean absolute error (MAE) was 2.75. In the whole growth period, the best effect was achieved by using the sparrow optimization algorithm. The Pearson-SSA-XGBoost model with the highest accuracy was studied, and the determination coefficient R 2 It reached 0.87, the root mean square error RMSE was 2.39, and the mean absolute error MAE was 1.99.

[0235] In each growth period, the sparrow optimization algorithm (SSA) has a more flexible search strategy, the behavior of the discoverer and follower, and the vigilance mechanism of the scout, which makes the information exchange between individuals more frequent and efficient, and can converge to the optimal solution faster. The gray wolf optimization algorithm is prone to fall into the local optimum during the search process, especially when dealing with complex functions, and may not be able to find the global optimal solution. According to the inversion results of the potato canopy SPAD mentioned above, if Figure 7 , Figure 8 and Fig. 9As shown in the figure, the accuracy of the model improved by the optimization algorithm has been significantly improved compared with the unimproved model. Different growth periods and research models must be combined with specific feature extraction methods and optimization algorithms to obtain the best prediction effect, which is of great significance for the inversion of potato canopy SPAD by UAV multispectral technology.

[0236] Therefore, the present invention adopts the above-mentioned canopy chlorophyll content inversion method based on feature extraction and machine learning model, uses data from different growth periods, multiple feature extraction methods and different optimization algorithms to optimize and improve the canopy SPAD inversion model, and finds that different feature extraction methods have significant differences in inversion accuracy in different growth periods. The CARS algorithm is better at extracting vegetation index in two single growth periods, and the Pearson correlation coefficient method in the entire growth period has higher accuracy. The accuracy of the optimization models using different optimization algorithms is significantly improved, and the SSA algorithm is better in each growth period. These application methods significantly improve the applicability and accuracy of canopy SPAD inversion research in different growth periods of potatoes.

[0237] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A canopy chlorophyll content inversion method based on feature extraction and machine learning model, characterized in that: The following steps are involved: Step S1, using a combination of a multispectral camera mounted on an unmanned aerial vehicle and ground measurement to collect and process the chlorophyll content data of potato canopies in the study area, and construct a vegetation index; Step S2, using Pearson correlation analysis and competitive adaptive reweighted sampling CARS algorithm to perform feature selection on vegetation index; Step S3: constructing a chlorophyll content inversion model based on machine learning; Step S4, optimizing the chlorophyll content inversion model constructed in step S3 based on the gray wolf optimization algorithm and the sparrow optimization algorithm; Step S5: perform accuracy evaluation on the chlorophyll content inversion model optimized in step S4.

2. The canopy chlorophyll content inversion method based on feature extraction and machine learning model according to claim 1 is characterized in that: In step S1, a combination of a multispectral camera mounted on an unmanned aerial vehicle and ground measurement was used to collect and process the chlorophyll content data of potato canopies in the study area, and a vegetation index was constructed. The specific process is as follows: Step S11: The UAV carries a multispectral camera to acquire and process image data; Firstly, an unmanned aerial vehicle (UAV) was used to collect image data of the study area. The UAV was equipped with four multispectral sensor probes, corresponding to the spectral band ranges of green light GREEN, red light RED, red edge light REG, and near infrared light NIR. Then, the collected images were input into Pix4Dmapper software for multispectral image synthesis and calibration, thereby obtaining orthorectified multispectral reflectance images; Step S12: collecting and measuring ground data; First, the study area was divided into 16 plots, each plot corresponding to one variety, and ground data of 16 potato varieties were collected. The area of ​​each plot was divided into 20m×100m, and data from 10 sample points were collected evenly in each plot. The leaf canopy chlorophyll content was collected from the potato planting base in the study area at the seedling stage and the mature stage, and 160 ground data were collected in each growth period. Then, the 16 sample plots divided above were taken as objects, and 10 well-growing potato canopies were evenly selected as sample points in each sample plot. When measuring each leaf canopy, the veins were avoided, and 10 sample points were evenly collected at the mesophyll using a handheld chlorophyll meter. The average value was taken as the true value of a sample point. The readings obtained from the potato plants selected at the sample points of each sample plot represented the relative chlorophyll content SPAD of the sample plot. Step S13: constructing a vegetation index based on the above acquired and processed data; The orthophoto images obtained from the UAV after geometric correction were divided into 16 corresponding areas according to potato varieties. Ten sample points were taken in each area. Combined with the longitude and latitude coordinates of the ground measurement points, the reflectance of the four bands of GREEN, RED, REG and NIR corresponding to each sample point was extracted on ArcGIS. The vegetation index was constructed based on these reflectances to enhance the physiological characteristics of vegetation.

3. The canopy chlorophyll content inversion method based on feature extraction and machine learning model according to claim 2 is characterized in that: In step S11, 20 reference points of a uniform white plate are selected to perform geometric correction on the image. The reflectance of the reference points is converted into the reflectance of the multispectral image using a pseudo-standard geometric correction method, so as to extract the accurate reflectance of the multispectral band. The correction formula is as follows: Among them, DN i is the reflectance value of the uniform white plate area; is the original mean before correction; is the average value of a uniform white board area; DN i,corrected is the corrected reflectance value.

4. The canopy chlorophyll content inversion method based on feature extraction and machine learning model according to claim 1 is characterized in that: In step S2 vegetation index feature selection, the Pearson feature selection method was used to measure the linear relationship between vegetation index and potato canopy SPAD. The Pearson correlation coefficient r X,Y The calculation formula is as follows: Where n is the number of samples; x i and i is the i-th data value of variables X and Y; and are the means of variables X and Y respectively.

5. The canopy chlorophyll content inversion method based on feature extraction and machine learning model according to claim 1, characterized in that: In step S2 vegetation index feature selection, a competitive adaptive reweighted sampling algorithm was used to evaluate the relevance of each vegetation index to the SPAD prediction model of potato canopy through multiple sampling and analysis. The specific process is as follows: Step S221, Monte Carlo sampling: Assume that the original vegetation index data set has n samples, the number of samples for each sampling is m (m≤n), and the number of samplings is k; for the i-th sampling (i=1,2,…,k), the obtained sampling subset S i Contains m samples randomly drawn from the original data set; Step S222: Establish a partial least squares regression PLS model: for the subset S obtained by the i-th sampling i , establish the PLS model; let X i is the vegetation index matrix m×p in the subset, where p is the number of vegetation indices; Y i is the corresponding target variable vector m×1; the PLS model is based on X i and Y i Decompose and regress to obtain the regression coefficient vector β i , as shown below: X=TP T +E (3); Y=UQ T +F (4); Among them, T and U are the score matrices of X and Y respectively; P and Q are the loading matrices; E and F are the residual matrices; Regression coefficient β i , as shown below: β i =W(P T W) -1 (5); Among them, W is the weight matrix, which is determined during the iteration process of the PLS algorithm; Step S223, calculating the vegetation index importance index; Assume ω i,j is the weight of the jth vegetation index in the PLS model of the i-th sampling, and the importance index I of the jth vegetation index is calculated. j , as shown below: Where n is the number of sampling times; a is the empirical index, where a=1 is set, and the vegetation index is selected according to the importance index.

6. The canopy chlorophyll content inversion method based on feature extraction and machine learning model according to claim 1, characterized in that: In step S3, a chlorophyll content inversion model is constructed based on machine learning. The specific process is as follows: Step S31, random forest RF: by establishing an RF model, an inversion experiment is conducted on the chlorophyll content of the potato canopy, and a ten-fold cross-validation method is used to continuously change the selection of the test set so that each subset has a chance to become a test set, and the performance of the model on different data subsets is evaluated; Different optimization algorithms were used to optimize the parameters of the RF model. In each round of cross-validation, different parameter combinations were used to train the model, and the model performance was evaluated on the test set to obtain an ideal chlorophyll content inversion model. Step S32, support vector regression SVR: establish an SVR model to invert the chlorophyll content of potato canopy, use RBF kernel function to fit the data, and its core expression is as follows: K(x i ,x j )=exp(-γ||x i -x j || 2 ) (7); Among them, x i and x j is the input sample; γ is the parameter of the kernel function, which determines the distribution range of the sample in the feature space; During the training process, the parameter optimization algorithm is used to continuously optimize the parameters according to the training data, and different hyperparameter combinations are evaluated to achieve chlorophyll content prediction and obtain the chlorophyll content inversion model; Step S33, extreme gradient boosting XGBoost: the potato chlorophyll data set is divided into a training set and a test set in a ratio of 8:

2. In the training phase, the XGBoost model uses the training set as the basis, the input features and the target variable, and uses the gradient boosting strategy to iteratively train the decision tree; In each iteration, the XGBoost model first calculates the gradient of the loss function with respect to the predicted value, and then fits a new decision tree based on the obtained gradient information. The newly generated decision tree learns and corrects the errors made by the previous decision tree during the prediction process, gradually improving the prediction performance and accuracy of the entire model, and obtaining the optimal chlorophyll content inversion model.

7. The canopy chlorophyll content inversion method based on feature extraction and machine learning model according to claim 1, characterized in that: In step S4, the parameters of the RF, SVR and XGBoost chlorophyll content inversion models are optimized based on the Gray Wolf Optimization Algorithm to invert the chlorophyll content of potato canopy. The specific steps are as follows: Step S411, inputting drone image data and measured potato canopy chlorophyll content; Step S412: using Pearson correlation analysis method and CARS algorithm to perform feature selection on the constructed vegetation index; Step S413, establishing RF, SVR and XGBoost chlorophyll inversion models; Step S414, initializing the population and parameters of the gray wolf, and defining the fitness function of the gray wolf optimization algorithm to evaluate the pros and cons of each parameter combination; Step S415, dividing the gray wolf individuals into social classes and calculating the distances between them to determine the update direction and position, and update the optimal position and fitness of the gray wolf; Step S416, determine whether the stop condition is met, if the condition is met, execute step S417, otherwise execute step S415; Step S417, outputting the optimal parameters of RF, SVR and XGBoost chlorophyll inversion models; Step S418: training the chlorophyll inversion model, obtaining the result of the chlorophyll content inversion model, and evaluating the model through evaluation indicators.

8. The canopy chlorophyll content inversion method based on feature extraction and machine learning model according to claim 1, characterized in that: In step S4, the parameters of the RF, SVR and XGBoost chlorophyll inversion models are optimized based on the sparrow optimization algorithm, and the chlorophyll content is inverted. The specific steps are as follows: Step S421, inputting drone image data and measured chlorophyll content data; Step S422, using Pearson and CARS algorithms to select and construct vegetation indices; Step S423, establishing RF, SVR and XGBoost chlorophyll content inversion models; Step S424, initializing the position of the sparrow population and defining a fitness function; Step S425, the fitness of each sparrow is calculated according to the fitness function, the sparrow with the best fitness is regarded as the discoverer, leading the search direction of the population, and the remaining sparrows are regarded as followers, and the followers update their positions according to the positions of the discoverer and themselves; Step S426: randomly select a part of sparrows as scouts to explore the new search area and update their own positions, thereby updating the best position and fitness of the global sparrows; Step S427, determine whether the termination condition is met, if the condition is met, execute step S428, otherwise execute step S425; Step S428, output the best model parameters optimized by the sparrow algorithm; Step S429: training the potato chlorophyll content inversion model, outputting the result of the chlorophyll content inversion model, and evaluating the accuracy of the model.

9. The canopy chlorophyll content inversion method based on feature extraction and machine learning model according to claim 1, characterized in that: In step S5, the chlorophyll content inversion model optimized in step S4 is evaluated for accuracy. The specific process is as follows: The data set was randomly divided into training set and test set with a ratio of 8:

2. Ten-fold cross validation was used to evaluate the generalization ability of the model during training. The coefficient of determination R was used to 2 , root mean square error RMSE and mean absolute error MAE parameters are used to measure the accuracy of the model. The calculation formula is as follows: Where n is the number of samples; x i and i is the measured value; is the predicted value; and is the average of the measured values.

Images