A Field Intelligent Prediction Method for Maize Kernel Moisture Content

Abstract

The present invention relates to the field of predicting the moisture content of corn, and particularly to an intelligent field prediction method for the moisture content of corn kernels. The method includes: collecting the spectral images of corn plants and the moisture content data of corn kernels, obtaining the preliminary spectral band screening results by performing stability screening on the reflectance data of spectral pixels; performing band weight analysis on the preliminary spectral band screening results to obtain the importance degree of each spectral band in the preliminary spectral band screening results; performing weighted optimization on the reflectance data of spectral pixels according to the importance degree of each spectral band, establishing a PLSR model through the optimized reflectance data of spectral pixels and the moisture content data of corn kernels, and predicting the moisture content of corn kernels by using the corn kernel moisture content prediction model for the spectral data in the corn field, so as to improve the prediction accuracy of the moisture content of corn kernels.

Description

Technical Field

[0001] The present invention relates to the technical field, and particularly relates to a method for intelligently predicting the water content of corn kernels in the field. Background Technique

[0002] Corn is an important food crop in China, and the water content of its kernels is a key factor affecting yield, quality, storage, and processing. Accurately and rapidly monitoring the water content of corn kernels is of great significance for guiding field management (such as irrigation, harvesting), optimizing storage conditions, and improving processing efficiency. Traditional methods for measuring the water content of corn kernels (such as the drying method) are time-consuming, damaging, and cannot achieve large-area, rapid, and non-destructive monitoring. In recent years, with the development of remote sensing technology, using unmanned aerial vehicles (UAVs) equipped with multi-spectral sensors to obtain the spectral information of the corn canopy and then inversely estimating crop physiological parameters has become an important precision agriculture technology means.

[0003] Multi-spectral remote sensing technology reflects the growth status and physiological parameters of crops by obtaining the reflectance information of crops in several specific bands (usually visible light, near-infrared, and short-wave infrared bands). Partial least squares regression (PLSR) is a commonly used multivariate statistical analysis method that can effectively handle the problem of multicollinearity and has been widely applied to spectral data analysis and modeling. Some studies have attempted to combine the PLSR method with multi-spectral remote sensing data to predict the water content or other physiological parameters of crops.

[0004] Although the partial least squares regression ( ) method has been widely applied to spectral data analysis, and multi-spectral remote sensing technology shows great potential in monitoring crop physiological parameters, in the specific application scenario of predicting the water content of corn kernels using UAV multi-spectral images, the existing algorithms and their commonly used variable selection methods still have limitations. As a result, the prediction accuracy, generalization ability, and interpretability of the model are limited. Specifically, although the number of bands in UAV multi-spectral images is relatively small, it still contains multiple bands. Some of these bands are closely related to the water content of corn kernels, while others contain noise, redundant information, or have a weak relationship with the water content. Traditional algorithms treat all bands equally and do not fully consider the contribution differences between bands, resulting in the model being vulnerable to the influence of noise and redundant information. More importantly, the existing variable selection methods (such as uninformative variable elimination , competitive adaptive reweighted sampling , variable importance in projection , etc.) often have difficulty taking into account multiple aspects such as the linear relationship, non-linear relationship between variables and the water content, and the correlation between variables when applied to multi-spectral data. For example, The method mainly selects based on the stability of variable regression coefficients, and it is easy to ignore variables that are non-linearly related to the target variable; Although the method simulates the principle of "survival of the fittest", the selection process has a certain degree of randomness, which easily causes instability of the results; The method can measure the contribution of variables to the overall model, but it does not consider the correlation between variables, resulting in redundancy in variable selection. These problems lead to existing and maize grain moisture content prediction models based on multispectral data having problems such as overfitting, insufficient generalization ability, and poor model interpretability. It is difficult to meet the requirements of precision agriculture for high-precision and high-reliability monitoring. Therefore, how to improve the variable selection method in the algorithm so that it can more effectively extract information related to maize grain moisture content from multispectral data, while reducing the influence of noise and redundant information and taking into account the linear and non-linear relationships between variables and moisture content is the key to improving the accuracy and interpretability of the prediction model. Summary of the Invention

[0005] In view of this, the present invention aims to propose a field intelligent prediction method for maize grain moisture content to optimize the variable screening of maize spectral data, thereby improving the accuracy of the prediction model.

[0006] To achieve the above object, the technical solution of the present invention is realized as follows:

[0007] A field intelligent prediction method for maize grain moisture content includes collecting maize plant spectral images and maize grain moisture content data to obtain spectral pixel reflectance data and maize grain moisture content data corresponding to the pixels, performing stability screening on the spectral pixel reflectance data to obtain a preliminary spectral band screening result; performing band weight analysis on the preliminary spectral band screening result to obtain the importance degree of each spectral band in the preliminary spectral band screening result; performing weighted optimization on the spectral pixel reflectance data through the importance degree of each spectral band to obtain optimized spectral pixel reflectance data, and establishing a PLSR model through the optimized spectral pixel reflectance data and maize grain moisture content data to obtain a maize grain moisture content prediction model; predicting the maize grain moisture content through the maize grain moisture content prediction model for maize field spectral data.

[0008] Further, according to performing stability screening on the spectral pixel reflectance data to obtain a preliminary spectral band screening result, the specific steps include:

[0009] Obtain maize plant spectral pixel reflectance data and maize grain moisture content data corresponding to each pixel, perform non-linear transformation on the maize plant spectral pixel reflectance data to obtain transformed maize plant spectral pixel reflectance data;

[0010] By separately evaluating the regression coefficients for each band in the spectral pixel reflectance data of the corn plants and the spectral pixel reflectance data of the transformed corn plants, the regression coefficients for each band in the spectral pixel reflectance data of the corn plants and the spectral pixel reflectance data of the transformed corn plants are obtained respectively;

[0011] By evaluating the stability of the regression coefficients for each band in the spectral pixel reflectance data of the corn plants and the spectral pixel reflectance data of the transformed corn plants, the stability levels corresponding to each band in the spectral pixel reflectance data of the corn plants and the spectral pixel reflectance data of the transformed corn plants are obtained respectively;

[0012] By conducting a comparative analysis of the stability levels corresponding to each band in the spectral pixel reflectance data of the corn plants and the spectral pixel reflectance data of the transformed corn plants, an evaluation of the stability differences for each band in the spectral pixel reflectance data of the corn plants is obtained;

[0013] By conducting a mapping analysis of the evaluation results of the stability differences, a band selection threshold is obtained, and the stability level of each band in the spectral pixel reflectance data of the corn plants is compared with the band selection threshold to obtain a preliminary spectral band selection result.

[0014] Furthermore, according to the separate evaluation of the regression coefficients for each band in the spectral pixel reflectance data of the corn plants and the spectral pixel reflectance data of the transformed corn plants, the regression coefficients for each band in the spectral pixel reflectance data of the corn plants and the spectral pixel reflectance data of the transformed corn plants are obtained respectively. The specific steps include:

[0015] Obtain the spectral pixel reflectance data of the corn plants. For the transformed spectral pixel reflectance data of the corn plants and the corn kernel moisture content data corresponding to each pixel, establish a first PLSR model based on the spectral pixel reflectance data of the corn plants and the corn kernel moisture content data corresponding to each pixel, and obtain the regression coefficient vector of the first PLSR model; establish a second PLSR model based on the transformed spectral pixel reflectance data of the corn plants and the corn kernel moisture content data corresponding to each pixel, and obtain the regression coefficient vector of the second PLSR model; by repeatedly establishing the first PLSR model and the second PLSR model, obtain multiple groups of regression coefficient vectors of the first PLSR model and multiple groups of regression coefficient vectors of the second PLSR model, calculate the mean values of the multiple groups of regression coefficient vectors of the first PLSR model and the multiple groups of regression coefficient vectors of the second PLSR model, and obtain the first regression coefficient vector of the first PLSR model and the first regression coefficient vector of the second PLSR model, so as to obtain the regression coefficients of each band in the spectral pixel reflectance data of the corn plants and the transformed spectral pixel reflectance data of the corn plants.

[0016] Further, based on the stability evaluation of the regression coefficients of each band in the spectral pixel reflectance data of the corn plants and the transformed spectral pixel reflectance data of the corn plants, respectively obtain the stability degrees corresponding to each band in the spectral pixel reflectance data of the corn plants and the transformed spectral pixel reflectance data of the corn plants. The specific steps are as follows:

[0017] Take the absolute value calculation result of the mean value of the regression coefficients of any band in the spectral pixel reflectance data of the corn plants as the first mean value of this band, and take the calculation result of dividing the standard deviation of the regression coefficients of any band in the spectral pixel reflectance data of the corn plants by the first mean value of this band as the stability degree of this band in the spectral pixel reflectance data of the corn plants; take the absolute value calculation result of the mean value of the regression coefficients of any band in the transformed spectral pixel reflectance data of the corn plants as the first mean value of this band, and take the calculation result of dividing the standard deviation of the regression coefficients of any band in the transformed spectral pixel reflectance data of the corn plants by the first mean value of this band as the stability degree of this band in the transformed spectral pixel reflectance data of the corn plants.

[0018] Further, based on the comparative analysis of the stability degrees corresponding to each band in the spectral pixel reflectance data of the corn plants and the transformed spectral pixel reflectance data of the corn plants, obtain the stability difference evaluation of each band in the spectral pixel reflectance data of the corn plants. The specific steps are as follows:

[0019] Obtain the stability degree of any wavelength band in the spectral pixel reflectance data of the corn plant and the transformed spectral pixel reflectance data of the corn plant, and use the calculation result of dividing the stability degree of any wavelength band in the transformed spectral pixel reflectance data of the corn plant by the stability degree of this wavelength band in the spectral pixel reflectance data of the corn plant as the stability difference evaluation of this wavelength band in the spectral pixel reflectance data of the corn plant.

[0020] Further, through mapping analysis of the stability difference evaluation results, obtain a wavelength band screening threshold, and compare the stability degree of each wavelength band in the spectral pixel reflectance data of the corn plant with the wavelength band screening threshold to obtain a preliminary spectral wavelength band screening result. The specific steps include:

[0021] Obtain the stability difference evaluation of each wavelength band in the spectral pixel reflectance data of the corn plant, the set first mapping parameter and the set adjustment coefficient. Use the mean value of the stability difference evaluations of all wavelength bands in the spectral pixel reflectance data of the corn plant as the second mean value. Use the calculation result of subtracting the second mean value from the stability difference evaluation of each wavelength band in the spectral pixel reflectance data of the corn plant as the first mean difference. Take the opposite number of the calculation result of multiplying the first mapping parameter by the first mean difference and perform an exponential mapping with the natural constant as the base to obtain the first exponential mapping; use the calculation result of adding the constant 1 to the first exponential mapping as the threshold optimization factor;

[0022] Obtain the stability degree of each wavelength band in the spectral pixel reflectance data of the corn plant. Use the mean value of the stability degrees of all wavelength bands in the spectral pixel reflectance data of the corn plant as the third mean value. Use the standard deviation of the stability degrees of all wavelength bands in the spectral pixel reflectance data of the corn plant as the first standard deviation. Use the calculation result of adding the adjustment coefficient multiplied by the first standard deviation to the third mean value as the basic screening threshold;

[0023] Use the calculation result of multiplying the threshold optimization factor by the basic screening threshold as the wavelength band screening threshold; for the stability degree of each wavelength band in the spectral pixel reflectance data of the corn plant, compare it with the wavelength band screening threshold. If the stability degree of the wavelength band in the spectral pixel reflectance data of the corn plant is less than or equal to the wavelength band screening threshold, retain the data of this wavelength band, thereby obtaining a preliminary spectral wavelength band screening result.

[0024] Further, perform wavelength band weight analysis on the preliminary spectral wavelength band screening result to obtain the importance degree of each spectral wavelength band in the preliminary spectral wavelength band screening result. The specific steps include:

[0025] Obtain the spectral band screening results of the spectral pixel reflectance data of corn plants, integrate the spectral band screening results into a spectral matrix, perform correlation analysis on the spectral matrix to obtain the correlation coefficient matrix of the spectral matrix; perform eigenvalue analysis through the correlation coefficient matrix of the spectral matrix to obtain the eigenvalues of each spectral band in the spectral matrix; perform feature difference evaluation on the features of each spectral band in the spectral matrix to obtain the importance degree of each spectral band in the preliminary spectral band screening results.

[0026] Further, according to the above-mentioned performing feature difference evaluation on the features of each spectral band in the spectral matrix to obtain the importance degree of each spectral band in the preliminary spectral band screening results, the specific steps include:

[0027] Obtain the correlation coefficient matrix and the importance degree adjustment parameter; extract the matrix eigenvector of the correlation coefficient matrix to obtain the eigenvalue vector of the correlation coefficient matrix; take the difference between the eigenvalue of any spectral band in the eigenvalue vector and the mean value of all eigenvalues in the eigenvalue vector of the correlation coefficient matrix as the second mean difference; perform exponential mapping with the natural constant as the base on the calculation result of multiplying the importance degree adjustment parameter by the second mean difference to obtain the second exponential mapping; take the reciprocal of the calculation result of adding the constant 1 to the second exponential mapping as the importance degree of the spectral band.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] A field intelligent prediction method for the water content of corn kernels according to the present invention collects the spectral images of corn plants and the data of the water content of corn kernels to obtain the spectral pixel reflectance data and the data of the water content of corn kernels corresponding to the pixels. Through stability screening of the spectral pixel reflectance data, preliminary spectral band screening results are obtained; band weight analysis is performed on the preliminary spectral band screening results to obtain the importance degree of each spectral band in the preliminary spectral band screening results; the spectral pixel reflectance data is weighted and optimized through the importance degree of each spectral band to obtain the optimized spectral pixel reflectance data, and a PLSR model is established through the optimized spectral pixel reflectance data and the data of the water content of corn kernels to obtain a prediction model for the water content of corn kernels; the water content of corn kernels is predicted through the prediction model for the water content of corn kernels for the spectral data in the corn field. Among them, stability screening is performed on the spectral pixel reflectance data to eliminate the bands greatly affected by the environment, improve the reliability of the input data, and then further quantify the contribution of each spectral band to the prediction of the water content of corn kernels through weight analysis of each band and perform weighted optimization accordingly, enhancing the influence of key spectral features on the prediction model, thereby improving the prediction accuracy of the model. Description of the Drawings

[0030] The accompanying drawings that form a part of the present invention are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0031] Figure 1 It is a flowchart of a method for intelligent prediction of corn kernel moisture content in the field according to an embodiment of the present invention; Detailed implementation manners

[0032] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0033] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "upper", "lower", "inner", "back", etc. is based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0034] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0035] See Figure 1 , which is a flowchart of a method for intelligent prediction of corn kernel moisture content in the field provided by Embodiment 1 of the present invention. As Figure 1 described, a method for intelligent prediction of corn kernel moisture content in the field may include:

[0036] Step S101, collect the spectral image of the corn plant and the corn kernel moisture content data to obtain the spectral pixel reflectance data and the corn kernel moisture content data corresponding to the pixels.

[0037] Obtain aerial photographs of the test field by using a drone equipped with a multispectral sensor during the critical growth period of corn. After preprocessing such as radiometric calibration, atmospheric correction, and geometric correction, the reflectance values of each sampling point in each band are obtained to form a multispectral data matrix.

[0038] Synchronously perform ground sampling during the flight of the drone. Collect corn kernel samples and determine their moisture content by the drying method to obtain moisture content data, and determine the corresponding relationship between the moisture content data and the multispectral data matrix.

[0039] Specifically, for the multi-spectral data matrix, each data in the matrix corresponds to the spectral data of a pixel (voxel) in the multi-spectral image. For the moisture content data of the corn kernel samples, the average moisture content data of the corn kernel samples at the corresponding position of each pixel in each spectral image is obtained as the actual measured value of the moisture content of the corn kernels corresponding to the pixel.

[0040] Step S102: Through stability screening of the spectral voxel reflectance data, a preliminary spectral band screening result is obtained.

[0041] When using unmanned aerial vehicle multi-spectral images for predicting the moisture content of corn kernels, the existing algorithms face challenges in variable selection. The traditional PLSR algorithm treats all spectral bands (variables) equally and does not consider the contribution differences of different bands to moisture content prediction, which may lead to the model containing a large amount of noise and redundant information, affecting the prediction accuracy, generalization ability, and interpretability. To solve this problem, an effective variable selection method is needed. The uninformative variable elimination method ( ) is a commonly used variable selection method based on the stability of the model regression coefficients, but it mainly focuses on the linear relationship between variables and the target variable. However, there are often complex non-linear relationships between spectral reflectance and crop physiological parameters, and the standard UVE method may wrongly eliminate variables that are non-linearly related to the moisture content.

[0042] To overcome the limitations of the standard method, it is necessary to identify variables that are non-linearly related to the moisture content by comparing the stability of variables on the original spectral data and the spectral data after non-linear transformation. Specifically, during implementation, first perform appropriate non-linear transformation on the original spectral data (such as logarithmic transformation, square root transformation, reciprocal transformation, etc., and the specific selection depends on the characteristics of the data and prior knowledge) to convert potential non-linear relationships into linear relationships or enhance existing non-linear relationships. Then, establish models based on the original spectral data and the transformed spectral data respectively, and calculate the stability of each variable in both cases (by multiple modelings, calculate the ratio of the standard deviation to the mean of the regression coefficients). If a variable shows higher stability on the transformed data than on the original data, it indicates that there is a non-linear relationship between this variable and the moisture content and should be retained. Through this improved method, the importance of variables can be evaluated more comprehensively and accurately, laying a foundation for constructing a better PLSR prediction model in the follow-up.

[0043] The specific preliminary spectral band screening process includes:

[0044] Obtain the spectral pixel reflectance data of corn plants and the corn kernel moisture content data corresponding to each pixel. Perform a non-linear transformation on the spectral pixel reflectance data of corn plants to obtain the transformed spectral pixel reflectance data of corn plants. By evaluating the regression coefficients for each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants respectively, obtain the regression coefficients for each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants. By evaluating the stability of the regression coefficients for each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants respectively, obtain the stability levels corresponding to each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants. By conducting a comparative analysis on the stability levels corresponding to each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants, obtain the stability difference evaluation for each band in the spectral pixel reflectance data of corn plants. By performing a mapping analysis on the stability difference evaluation results, obtain the band screening threshold. Compare the stability level of each band in the spectral pixel reflectance data of corn plants with the band screening threshold to obtain the preliminary spectral band screening result.

[0045] Among them, by evaluating the regression coefficients for each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants respectively, obtaining the regression coefficients for each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants respectively includes:

[0046] Obtain the spectral pixel reflectance data of corn plants, the transformed spectral pixel reflectance data of corn plants, and the corn kernel moisture content data corresponding to each pixel. Establish a first PLSR model using the spectral pixel reflectance data of corn plants and the corn kernel moisture content data corresponding to each pixel, and obtain the regression coefficient vector of the first PLSR model. Establish a second PLSR model using the transformed spectral pixel reflectance data of corn plants and the corn kernel moisture content data corresponding to each pixel, and obtain the regression coefficient vector of the second PLSR model. By establishing the first PLSR model and the second PLSR model multiple times, obtain multiple groups of regression coefficient vectors of the first PLSR model and multiple groups of regression coefficient vectors of the second PLSR model. Calculate the mean of the multiple groups of regression coefficient vectors of the first PLSR model and the multiple groups of regression coefficient vectors of the second PLSR model to obtain the first regression coefficient vector of the first PLSR model and the first regression coefficient vector of the second PLSR model, thereby obtaining the regression coefficients for each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants.

[0047] It should be noted that these two The model is not used for final moisture content prediction, but serves as an intermediate step in variable selection. Its role is to provide a preliminary assessment to help determine which variables have strong linear or non-linear relationships with the moisture content, and to provide relatively reliable information about variable stability through these two models to support preliminary variable screening. Using the original spectral data matrix and the corresponding measured moisture content values, by calling the existing algorithm library, the first model is established. Then, non-linear transformation (such as logarithmic transformation, square root transformation, etc.) is performed on the original spectral data . In this embodiment, the non-linear transformation process is carried out through logarithmic transformation to obtain the transformed spectral data matrix; the second model is established using the transformed spectral data matrix and the measured moisture content values, and the regression coefficient vectors are extracted from these two models respectively. In order to ensure the stability of the regression coefficients, the method of averaging multiple modelings is adopted, that is, the above modeling process is repeated multiple times and the regression coefficient vectors obtained each time are averaged. Finally, the regression coefficients of each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants are obtained.

[0048] After obtaining the regression coefficients of each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants, it is necessary to evaluate the stability of the regression coefficients of each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants, and the corresponding stability degrees of each band in the spectral pixel reflectance data of corn plants and the transformed spectral pixel reflectance data of corn plants are obtained respectively. The specific steps include:

[0049] The absolute value of the mean of the regression coefficients of any band in the spectral pixel reflectance data of corn plants is calculated as the first mean of this band, and the result of dividing the standard deviation of the regression coefficients of any band in the spectral pixel reflectance data of corn plants by the first mean of this band is used as the stability degree of this band in the spectral pixel reflectance data of corn plants; the absolute value of the mean of the regression coefficients of any band in the transformed spectral pixel reflectance data of corn plants is calculated as the first mean of this band, and the result of dividing the standard deviation of the regression coefficients of any band in the transformed spectral pixel reflectance data of corn plants by the first mean of this band is used as the stability degree of this band in the transformed spectral pixel reflectance data of corn plants.

[0050] In one embodiment, assume that the regression coefficient of the th band in the spectral pixel reflectance data of corn plants is , and the regression coefficient of the th band in the transformed spectral pixel reflectance data of corn plants is , the calculation expression for the stability degree of the -th band in the spectral pixel reflectance data of corn plants and the stability degree of the -th band in the spectral pixel reflectance data of corn plants after transformation is:

[0051]

[0052]

[0053] Wherein, represents the stability degree of the -th band in the spectral pixel reflectance data of corn plants; represents the stability degree of the -th band in the spectral pixel reflectance data of corn plants after transformation; represents the standard deviation of the regression coefficient of the -th band in the spectral pixel reflectance data of corn plants; represents the standard deviation of the regression coefficient of the -th band in the spectral pixel reflectance data of corn plants after transformation; represents the mean value of the regression coefficient of the -th band in the spectral pixel reflectance data of corn plants; represents the mean value of the regression coefficient of the -th band in the spectral pixel reflectance data of corn plants after transformation.

[0054] It should be noted that and respectively calculate the stability of each variable on the original spectral data and the spectral data after non - linear transformation. Here, the idea of the traditional method is borrowed, and the ratio of the standard deviation to the mean value of the regression coefficient (i.e., the coefficient of variation) is used to measure the stability of the variable. The coefficient of variation can eliminate the influence of the dimension between different variables, making the stability of different variables comparable. The reason for calculating the stability of the variable on the original data and the transformed data is that there is often a non - linear relationship between the spectral reflectance and the water content. By introducing non - linear transformation, the potential non - linear relationship can be transformed into a linear relationship or the strength of the non - linear relationship can be enhanced so that the model can better capture these relationships.

[0055] After obtaining the stability degree of each band in the spectral pixel reflectance data of corn plants before and after transformation, the evaluation of the stability difference of each band in the spectral pixel reflectance data of corn plants can be obtained by comparing and analyzing the corresponding stability degrees of each band in the spectral pixel reflectance data of corn plants and the spectral pixel reflectance data of corn plants after transformation, including:

[0056] Obtain the stability degree of any wavelength band in the spectral pixel reflectance data of the corn plant and the spectral pixel reflectance data of the transformed corn plant. Take the calculation result of dividing the stability degree of any wavelength band in the spectral pixel reflectance data of the transformed corn plant by the stability degree of this wavelength band in the spectral pixel reflectance data of the corn plant as the evaluation of the stability difference of this wavelength band in the spectral pixel reflectance data of the corn plant.

[0057] Finally, through mapping analysis of the stability difference evaluation results, obtain the band screening threshold. Compare the stability degree of each wavelength band in the spectral pixel reflectance data of the corn plant with the band screening threshold to obtain the preliminary spectral band screening results, including:

[0058] Obtain the stability difference evaluation of each wavelength band in the spectral pixel reflectance data of the corn plant, the set first mapping parameter and the set adjustment coefficient. Take the mean value of the stability difference evaluations of all wavelength bands in the spectral pixel reflectance data of the corn plant as the second mean value. Take the calculation result of subtracting the stability difference evaluation of each wavelength band in the spectral pixel reflectance data of the corn plant from the second mean value as the first mean difference. Take the opposite number of the calculation result of multiplying the first mapping parameter by the first mean difference and perform exponential mapping with the natural constant as the base to obtain the first exponential mapping. Take the calculation result of adding the constant 1 to the first exponential mapping as the threshold optimization factor;

[0059] Obtain the stability degree of each wavelength band in the spectral pixel reflectance data of the corn plant. Take the mean value of the stability degrees of all wavelength bands in the spectral pixel reflectance data of the corn plant as the third mean value. Take the standard deviation of the stability degrees of all wavelength bands in the spectral pixel reflectance data of the corn plant as the first standard deviation. Take the calculation result of adding the third mean value to the product of the adjustment coefficient and the first standard deviation as the basic screening threshold;

[0060] Take the calculation result of multiplying the threshold optimization factor by the basic screening threshold as the band screening threshold; for the stability degree of each wavelength band in the spectral pixel reflectance data of the corn plant, compare it with the band screening threshold. If the stability degree of the wavelength band in the spectral pixel reflectance data of the corn plant is less than or equal to the band screening threshold, retain the data of this wavelength band, thereby obtaining the preliminary spectral band screening results.

[0061] In one embodiment, assume that the stability difference evaluation of the th wavelength band in the spectral pixel reflectance data of the corn plant is , then the calculation expression of the band screening threshold of the th wavelength band in the spectral pixel reflectance data of the corn plant is:

[0062]

[0063] Among them, represents the band screening threshold of the th band in the spectral pixel reflectance data of corn plants; represents the first mapping parameter; represents the stability difference evaluation of the th band in the spectral pixel reflectance data of corn plants; represents the mean value of the stability difference evaluations of all bands in the spectral pixel reflectance data of corn plants; represents the mean value of the stability degrees of all variables on the original data; represents the adjustment coefficient; represents the standard deviation of the stability degrees of all variables on the original data.

[0064] It should be noted that, based on and , the stability ratio of the variable is further calculated. represents the degree of change in the stability of the variable on the transformed data relative to its stability on the original data. If the value of a variable is large, it indicates that the stability of the variable on the transformed data has been significantly improved, which means that there is likely a non - linear relationship between the variable and the water content. The traditional method only judges the importance of a variable based on its stability on the original data and ignores the non - linear relationship. The present invention incorporates the non - linear relationship into the consideration scope of variable selection by introducing . In order to convert into a variable selection threshold, an exponential mapping function ( ) is introduced. It can map the value to a range between . When is much larger than the offset parameter , the output of the exponential mapping function is close to ; when is close to , the output is close to ; when is much smaller than , the output is close to . By introducing the exponential mapping function, the non - linear transformation of is realized, making the threshold smoother with respect to the change of . In addition, the steepness parameter of the exponential mapping function controls the sensitivity of the threshold to the change of , and the larger the value, the more sensitive the threshold is to the change of . The offset parameter Fine-tuning can be performed according to the actual situation. In this embodiment, it is set that , , and the parameter setting can be adjusted according to the actual scenario without any requirements.

[0065] Finally, multiply the output value of the function by a reference threshold to obtain the final variable selection threshold . This reference threshold is jointly determined by the average value of the stability of all variables on the original data and the standard deviation , that is . Among them, the average value reflects the overall level, and the standard deviation reflects the degree of dispersion. The adjustment coefficient is used to control the overall level of the threshold. The larger it is, the higher the threshold and the fewer variables are selected. For variables linearly related to the water content, is close to , is close to the reference threshold; for variables non-linearly related to the water content, is greater than , will be higher than the reference threshold, making these variables more likely to be selected.

[0066] In this step, through non-linear transformation, the non-linear features in the spectral data are enhanced, and the response ability of the spectral information to the change of water content is improved; subsequently, by evaluating the regression coefficients of the original spectral data and the transformed spectral data, the contribution degree of different bands to the prediction of grain water content is identified, and further through stability analysis, the spectral bands with robust contributions to the prediction are screened out to reduce noise interference. In addition, through stability difference evaluation and mapping analysis, the band screening threshold is accurately set to ensure that only the spectral bands that contribute stably and importantly to the water content prediction are retained, avoiding errors introduced by redundant bands or low-contribution bands, thereby improving the generalization ability of the model. Finally, based on the optimized and screened spectral data, a high-precision PLSR water content prediction model is established, which can effectively improve the prediction accuracy of the spatial distribution of maize grain water content in the field, provide a scientific basis for precision harvesting and water management, reduce grain losses caused by water differences, and improve the intelligent level of agricultural production.

[0067] Step S103: Perform band weight analysis on the preliminary spectral band screening results to obtain the importance degree of each spectral band in the preliminary spectral band screening results.

[0068] In the preliminary spectral band screening, some noise and variables (spectral bands) with weak correlation with the water content have been removed. However, there may still be high correlations and redundancies among the remaining variables, which are mainly reflected in the following two aspects:

[0069] 1. Information redundancy problem caused by spectral continuity: Although the number of spectral bands in UAV multispectral images is not as large as that of hyperspectral images, they still have a certain degree of continuity. The physical characteristics (such as light absorption and reflection) between adjacent bands are often very similar, resulting in highly correlated spectral values. This correlation means that there is a large amount of redundancy in the information carried by adjacent bands. If these highly correlated bands are all included in the model, it will lead to an overly complex model and be prone to overfitting.

[0070] 2. Information overlap problem caused by potential factors: Even non-adjacent bands may show a certain degree of correlation due to being affected by some common factors (such as light, atmosphere, etc.) or being related to some potential variables (such as chlorophyll content, cell structure, etc.). This means that there is an information overlap problem among the spectral information of different bands. That is, multiple bands reflect the changes of the same potential factor. This information overlap will also lead to overfitting of the model.

[0071] The above two problems (information redundancy caused by spectral continuity and information overlap caused by potential factors) will both lead to overfitting of the PLSR model. Therefore, to solve the above problems, it is necessary to further refine the variables after preliminary screening to reduce the correlation and redundancy between variables.

[0072] The method used in the present invention is to analyze the correlation coefficient matrix between variables. The correlation coefficient matrix describes the degree of linear correlation between pairs of variables. If the correlation coefficient between two variables is close to or , it indicates that there is a strong positive or negative correlation between them. If a variable has a strong correlation with multiple other variables, then this variable is redundant. However, variables cannot be simply eliminated based on the magnitude of the correlation coefficient. Because the correlation coefficient can only reflect the relationship between two variables and cannot reflect the complex relationship among multiple variables. Therefore, to more accurately evaluate the correlation and redundancy between variables, it is necessary to analyze the overall structure of the correlation coefficient matrix. Eigenvalue decomposition can decompose the correlation coefficient matrix into a set of eigenvectors and corresponding eigenvalues. Each eigenvector represents a variable combination, and each eigenvalue represents the variance of the variable combination. The larger the eigenvalue, the greater the variance of the variable combination, indicating that the information contained is more; the smaller the eigenvalue, the smaller the variance of the variable combination, indicating that the information contained is less, pointing to noise or redundant information.

[0073] Finally, by analyzing the eigenvalues and eigenvectors of the correlation coefficient matrix, independent variable combinations are identified. For those variable combinations with smaller eigenvalues, the corresponding original variables are very likely to be highly correlated with other variables, indicating that they are redundant. According to this criterion, the weights of these variables are reduced.

[0074] Specifically, perform band weight analysis on the preliminary spectral band screening results to obtain the importance of each spectral band in the preliminary spectral band screening results, including:

[0075] Obtain the spectral band screening results of the spectral pixel reflectance data of corn plants, integrate the spectral band screening results into a spectral matrix, perform correlation analysis on the spectral matrix to obtain the correlation coefficient matrix of the spectral matrix; perform eigenvalue analysis through the correlation coefficient matrix of the spectral matrix to obtain the eigenvalues of each spectral band in the spectral matrix; evaluate the feature differences of each spectral band in the spectral matrix to obtain the importance of each spectral band in the preliminary spectral band screening results.

[0076] Among them, evaluating the feature differences of each spectral band in the spectral matrix to obtain the importance of each spectral band in the preliminary spectral band screening results specifically includes:

[0077] Obtain the correlation coefficient matrix and the importance adjustment parameter; extract the matrix eigenvectors of the correlation coefficient matrix to obtain the eigenvalue vector of the correlation coefficient matrix; take the difference between the eigenvalue of any spectral band in the eigenvalue vector and the mean of all eigenvalues in the eigenvalue vector of the correlation coefficient matrix as the second mean difference; perform exponential mapping with the natural constant as the base on the calculation result of multiplying the importance adjustment parameter by the second mean difference to obtain the second exponential mapping; take the reciprocal of the calculation result of adding the constant 1 to the second exponential mapping as the importance of the spectral band.

[0078] In an embodiment, assume the eigenvalue of the th band is , then the calculation expression for the importance of the

[0079]

[0080] th band in the spectral matrix is: represents the importance of the th band in the spectral matrix; represents the importance adjustment parameter; represents the eigenvalue of the th band in the spectral matrix; represents the mean of all eigenvalues in the eigenvalue vector of the correlation coefficient matrix; represents the constant ; represents the natural constant.

[0081] It should be noted that the eigenvector represents a linear combination of variables, and the eigenvalue represents the variance of these linear combinations. The larger the eigenvalue, the greater the variance of the corresponding variable combination and the more information it contains; the smaller the eigenvalue, the smaller the variance of the corresponding variable combination and the less information it contains. At the same time, the weights are not directly determined according to the size of the eigenvalues, but the average value of the eigenvalues is calculated. , and then calculate each eigenvalue and The difference, this difference reflects the degree of deviation of the eigenvalue from the average level. If an eigenvalue is less than , it means that the corresponding variable combination contains less information, and the weights of these variables need to be reduced. Finally, the weight of each variable is calculated through the formula so that is sensitive to the difference between the eigenvalue and the mean, the smaller, the closer to ; the larger, the closer to . In this way, variables with smaller eigenvalues (usually corresponding to redundant information) will be given lower weights, while variables with larger eigenvalues (usually containing more valid information) will be given higher weights.

[0082] Step S104, weight and optimize the spectral pixel reflectance data according to the importance of each spectral band to obtain the optimized spectral pixel reflectance data, and establish a PLSR model with the optimized spectral pixel reflectance data and the corn kernel moisture content data.

[0083] First, weight the spectral reflectance corresponding to each band in the spectral matrix according to the obtained band importance to obtain a weighted spectral data matrix. Then, use the weighted spectral data matrix and the corresponding measured moisture content value vector , and construct model. The construction process of the model belongs to the prior art, so it will not be elaborated here. Decompose the spectral data matrix and the moisture content vector into linear combinations of several latent variables, and these latent variables can explain the variation of the spectral data to the greatest extent and be related to the moisture content to the greatest extent. The model determines the number of latent variables and the load vector corresponding to each latent variable through iterative calculation to establish the regression relationship between the spectral data and the moisture content. To avoid overfitting of the model, cross-validation needs to be used to determine parameters such as the optimal number of latent variables of the model.

[0084] This step evaluates the importance of each band in the spectral matrix, so that in subsequent steps, the reflectance data of each band in the spectral matrix is weighted according to the importance, thereby eliminating information redundancy caused by spectral continuity in the spectral matrix and information overlap caused by potential factors, improving the prediction accuracy of the PLSR model, and more accurately predicting the moisture content of corn kernels.

[0085] Step S105: Predict the moisture content of corn kernels using the corn kernel moisture content prediction model for the corn field spectral data.

[0086] For a new sample, first obtain its multispectral data and perform the same preliminary spectral band screening, band importance evaluation, and weighting as in the above steps to obtain the weighted spectral data. Then input it into the constructed model, and the predicted value of the moisture content of the corn kernels of this sample can be obtained. Specifically, for each pixel in the image, extract its multispectral reflectance value and perform variable selection and weighting according to the method in the present invention to obtain the weighted spectral data. Input the weighted spectral data of each pixel into the model to obtain the corresponding predicted value of the moisture content of the kernels. Then, arrange the predicted values of the moisture content of each pixel according to their spatial positions in the image to generate a spatial distribution map of the moisture content of corn kernels, and use different colors or gray levels to represent different moisture content ranges. This moisture content distribution map can provide an important reference for precision agricultural production. Identify the moisture content differences in different regions within the field according to the moisture content distribution map to guide precision irrigation. It is also possible to determine the maturity differences of corn kernels in different regions within the field according to the moisture content distribution to guide timely harvesting. Prioritize harvesting for areas where the moisture content has reached the harvesting standard, and appropriately postpone the harvesting time for areas with still relatively high moisture content, thereby improving the harvesting efficiency, reducing kernel losses, and ensuring the quality of corn.

[0087] In summary, for the field intelligent prediction method of corn kernel moisture content described in the present invention, spectral images of corn plants and corn kernel moisture content data are collected to obtain spectral pixel reflectance data and corn kernel moisture content data corresponding to the pixels. Through stability screening of the spectral pixel reflectance data, a preliminary spectral band screening result is obtained; band weight analysis is performed on the preliminary spectral band screening result to obtain the importance of each spectral band in the preliminary spectral band screening result; the spectral pixel reflectance data is weighted and optimized according to the importance of each spectral band to obtain the optimized spectral pixel reflectance data, and a PLSR model is established by using the optimized spectral pixel reflectance data and the corn kernel moisture content data to obtain a corn kernel moisture content prediction model; the corn kernel moisture content is predicted by using the corn kernel moisture content prediction model for the spectral data in the corn field. Among them, stability screening is performed on the spectral pixel reflectance data to eliminate the bands greatly affected by the environment, improve the reliability of the input data, and then further quantify the contribution of each spectral band to the prediction of kernel moisture content through weight analysis of each band and perform weighted optimization accordingly, enhance the influence of key spectral features on the prediction model, and thus improve the prediction accuracy of the model.

[0088] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A field intelligent prediction method for corn kernel moisture content, comprising: The spectral image of corn plants and the water content data of corn kernels are collected to obtain the spectral pixel reflectance data and the water content data of corn kernels corresponding to the pixels, wherein the method is characterized in that the spectral pixel reflectance data is subjected to stability screening to obtain a preliminary spectral band screening result; the preliminary spectral band screening result is subjected to band weight analysis to obtain the importance of each spectral band in the preliminary spectral band screening result; the spectral pixel reflectance data is weightedly optimized according to the importance of each spectral band to obtain the optimized spectral pixel reflectance data; a PLSR model is established by using the optimized spectral pixel reflectance data and the water content data of corn kernels to obtain a water content prediction model of corn kernels; the water content of corn kernels is predicted by using the water content prediction model of corn kernels for predicting the water content of corn kernels on the spectral data of corn fields; The stability screening of the spectral pixel reflectance data to obtain preliminary spectral band screening results specifically includes: Acquire spectral pixel reflectance data of corn plants and corn kernel moisture content data corresponding to each pixel, and perform nonlinear transformation on the spectral pixel reflectance data of corn plants to obtain transformed spectral pixel reflectance data of corn plants; The regression coefficients of each band in the spectral pixel reflectance data of corn plants and the spectral pixel reflectance data of corn plants after transformation are evaluated respectively, and the regression coefficients of each band in the spectral pixel reflectance data of corn plants and the spectral pixel reflectance data of corn plants after transformation are obtained respectively; By evaluating the stability of the regression coefficient of each band in the spectral pixel reflectance data of corn plants and the spectral pixel reflectance data of corn plants after transformation, the stability of each band in the spectral pixel reflectance data of corn plants and the spectral pixel reflectance data of corn plants after transformation is obtained respectively; By comparing and analyzing the stability of each band in the spectral pixel reflectance data of the corn plant and the transformed spectral pixel reflectance data of the corn plant, a stability difference evaluation of each band in the spectral pixel reflectance data of the corn plant is obtained; By mapping and analyzing the stability difference evaluation results, a band screening threshold is obtained, and the stability of each band in the corn plant spectral pixel reflectance data is compared with the band screening threshold to obtain a preliminary spectral band screening result.

2. The method for intelligent field prediction of corn grain moisture content according to claim 1, characterized in that: According to the method, by evaluating the regression coefficients of each band in the spectral pixel reflectance data of the corn plant and the spectral pixel reflectance data of the corn plant after transformation, the regression coefficients of each band in the spectral pixel reflectance data of the corn plant and the spectral pixel reflectance data of the corn plant after transformation are obtained respectively, and the specific steps include: Acquire the spectral pixel reflectance data of the corn plant, the spectral pixel reflectance data of the corn plant after transformation, and the corn kernel moisture content data corresponding to each pixel, establish a first PLSR model through the spectral pixel reflectance data of the corn plant and the corn kernel moisture content data corresponding to each pixel, and obtain the regression coefficient vector of the first PLSR model; establish a second PLSR model through the spectral pixel reflectance data of the corn plant after transformation and the corn kernel moisture content data corresponding to each pixel, and obtain the regression coefficient vector of the second PLSR model; by repeatedly establishing the first PLSR model and establishing the second PLSR model, obtain multiple groups of regression coefficient vectors of the first PLSR model and multiple groups of regression coefficient vectors of the second PLSR model, calculate the mean of the multiple groups of regression coefficient vectors of the first PLSR model and the multiple groups of regression coefficient vectors of the second PLSR model, obtain the first regression coefficient vector of the first PLSR model and the first regression coefficient vector of the second PLSR model, thereby obtaining the regression coefficient of each band in the spectral pixel reflectance data of the corn plant and the spectral pixel reflectance data of the corn plant after transformation.

3. The method for intelligent field prediction of corn grain moisture content according to claim 1, characterized in that: According to the method of evaluating the stability of the regression coefficient of each band in the spectral pixel reflectance data of the corn plant and the spectral pixel reflectance data of the corn plant after transformation, the stability corresponding to each band in the spectral pixel reflectance data of the corn plant and the spectral pixel reflectance data of the corn plant after transformation is obtained, and the specific steps include: The absolute value calculation result of the mean of the regression coefficient of any band in the spectral pixel reflectance data of the corn plant is used as the first mean of the band, and the standard deviation of the regression coefficient of any band in the spectral pixel reflectance data of the corn plant is divided by the first mean of the band as the stability of the band in the spectral pixel reflectance data of the corn plant; the absolute value calculation result of the mean of the regression coefficient of any band in the spectral pixel reflectance data of the corn plant after transformation is used as the first mean of the band, and the standard deviation of the regression coefficient of any band in the spectral pixel reflectance data of the corn plant after transformation is divided by the first mean of the band as the stability of the band in the spectral pixel reflectance data of the corn plant after transformation.

4. The method for intelligent field prediction of corn grain moisture content according to claim 1, characterized in that: According to the comparative analysis of the stability corresponding to each band in the spectral pixel reflectance data of the corn plant and the transformed spectral pixel reflectance data of the corn plant, the stability difference evaluation of each band in the spectral pixel reflectance data of the corn plant is obtained, and the specific steps include: The stability of any band in the spectral pixel reflectance data of the corn plant and the spectral pixel reflectance data of the corn plant after transformation is obtained, and the stability of any band in the spectral pixel reflectance data of the corn plant after transformation is divided by the stability of the band in the spectral pixel reflectance data of the corn plant as a calculation result of the stability difference evaluation of the band in the spectral pixel reflectance data of the corn plant.

5. The method for intelligent field prediction of corn grain moisture content according to claim 1, characterized in that: According to the mapping analysis of the stability difference evaluation results, a band screening threshold is obtained, and the stability of each band in the corn plant spectral pixel reflectance data is compared with the band screening threshold to obtain a preliminary spectral band screening result, the specific steps of which include: Obtaining a stability difference assessment of each band in the spectral pixel reflectance data of the corn plant, setting a first mapping parameter and a set adjustment coefficient, taking the mean of the stability difference assessments of all bands in the spectral pixel reflectance data of the corn plant as a second mean, taking a calculation result of subtracting the stability difference assessment of each band in the spectral pixel reflectance data of the corn plant from the second mean as a first mean difference, performing an exponential mapping with a natural constant as the base on the inverse of a calculation result of multiplying the first mapping parameter by the first mean difference to obtain a first exponential mapping; taking a calculation result of adding a constant 1 to the first exponential mapping as a threshold optimization factor; Obtaining the stability of each band in the spectral pixel reflectance data of the corn plant, taking the mean of the stability of all bands in the spectral pixel reflectance data of the corn plant as the third mean, taking the standard deviation of the stability of all bands in the spectral pixel reflectance data of the corn plant as the first standard deviation, and taking the third mean plus the adjustment coefficient multiplied by the first standard deviation as the basic screening threshold; The calculation result of multiplying the threshold optimization factor by the basic screening threshold is used as the band screening threshold; the stability of each band in the spectral pixel reflectance data of the corn plant is compared with the band screening threshold; if the stability of the band in the spectral pixel reflectance data of the corn plant is less than or equal to the band screening threshold, the band data is retained to obtain a preliminary spectral band screening result.

6. The method for intelligent field prediction of corn grain moisture content according to claim 1, characterized in that: According to the band weight analysis of the preliminary spectral band screening result, the importance of each spectral band in the preliminary spectral band screening result is obtained, and the specific steps include: The spectral band screening results of the spectral pixel reflectance data of the corn plant are obtained, and the spectral band screening results are integrated into a spectral matrix. By performing correlation analysis on the spectral matrix, a correlation coefficient matrix of the spectral matrix is ​​obtained; by performing eigenvalue analysis on the correlation coefficient matrix of the spectral matrix, the eigenvalue of each spectral band in the spectral matrix is ​​obtained; by performing feature difference evaluation on the characteristics of each spectral band in the spectral matrix, the importance of each spectral band in the preliminary spectral band screening results is obtained.

7. The method for intelligent field prediction of corn grain moisture content according to claim 6, characterized in that: According to the feature difference evaluation of the features of each spectral band in the spectral matrix, the importance of each spectral band in the preliminary spectral band screening result is obtained, and the specific steps include: Acquire the correlation coefficient matrix and the importance adjustment parameter; perform matrix eigenvector extraction on the correlation coefficient matrix to obtain the eigenvalue vector of the correlation coefficient matrix; use the difference between the eigenvalue of any spectral band in the eigenvalue vector and the mean of all eigenvalues ​​in the eigenvalue vector of the correlation coefficient matrix as the second mean difference; perform exponential mapping with a natural constant as the base on the result of multiplying the importance adjustment parameter by the second mean difference to obtain a second exponential mapping; use the inverse of the result of adding the constant 1 to the second exponential mapping as the importance of the spectral band.

Images