A pesticide residue content inversion method and system based on multi-dimensional optical data

Abstract

A pesticide residue content inversion method and system based on multi-dimensional optical data relate to the field of hyperspectral remote sensing, and solve the problems that existing pesticide residue inversion technology based on hyperspectral is easily disturbed by soil state, and has low inversion efficiency and precision.A pesticide residue content inversion method obtains an absorption peak range and a spectral sample set based on a spectral image of soil to be inverted; obtains a Pearson correlation coefficient of each wavelength based on spectral reflectivity and polarization parameters of each wavelength; obtains a characteristic wavelength set according to the Pearson correlation coefficient r and the absorption peak range; fuses spectral information and polarization information of the characteristic wavelength set to obtain a fused characteristic vector set; trains a supervised learning model through the characteristic vector set to obtain a pesticide residue content inversion model; and obtains pesticide residue content through the model.The method is suitable for the field of pesticide residue content inversion of farmland soil.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of hyperspectral remote sensing, in particular to the field of image processing. BACKGROUND

[0002] In recent years, spectral technology plays an important role in agricultural and forestry remote sensing. Hyperspectral data can be used to retrieve the characteristics of soil organic matter, nutrients, water content, and salt content, providing soil management information for agricultural production. Pesticide residue retrieval is an important part of soil management. Applying spectral technology to pesticide residue retrieval is beneficial to variable fertilization, targeted pesticide application, and other aspects, improving agricultural production efficiency and crop quality, and thus ensuring food safety and achieving precision agricultural management.

[0003] However, single hyperspectral technology cannot distinguish between targets with similar compositions but different surface states. Pesticides have very low content in soil, and their spectral signals are highly overlapped with the spectral characteristics of soil organic matter, clay minerals, and water, so they are easily disturbed by soil state, resulting in low retrieval efficiency and accuracy.

[0004] In summary, the existing hyperspectral-based pesticide residue retrieval technology has the problems of being easily disturbed by soil state, low retrieval efficiency, and low accuracy. SUMMARY

[0005] The present application provides the following solutions to the problems of existing hyperspectral-based pesticide residue retrieval technology being easily disturbed by soil state, low retrieval efficiency, and low accuracy.

[0006] Solution one: a pesticide residue content retrieval method based on multi-dimensional optical data, comprising the following steps:

[0007] Step S01: based on the spectral image of the soil to be retrieved, obtaining the absorption peak range and the spectral sample set; based on the spectral reflectance and polarization parameters of each wavelength in the spectral sample set, obtaining the Pearson correlation coefficient of each wavelength;

[0008] The polarization parameters include the degree of polarization and the polarization angle;

[0009] Step S02: selecting a number of candidate characteristic wavelengths according to the Pearson correlation coefficient r of each wavelength and the absorption peak range, and constructing a candidate set of characteristic wavelengths;

[0010] Step S03: screening the candidate set of characteristic wavelengths to obtain a set of characteristic wavelengths;

[0011] Step S04: performing spectral information and polarization information fusion on each characteristic wavelength in the set of characteristic wavelengths to obtain a set of fused feature vectors;

[0012] Step S05, selecting a type of supervised learning model based on the candidate set of characteristic wavelengths;

[0013] Step S06, training the supervised learning model obtained in step S05 by the set of feature vectors to obtain a pesticide residue content inversion model;

[0014] Step S07, fusing the spectral reflectance and polarization parameters of the soil to be inverted, and obtaining the pesticide residue content through the pesticide residue content inversion model.

[0015] Further, in an embodiment of the present application, in step S05

[0016] If the wavelengths of the candidate set of characteristic wavelengths are all wavelengths with Pearson correlation coefficient |r|>correlation threshold, a PLSR model is selected;

[0017] Otherwise, an SVR model is selected.

[0018] Further, in an embodiment of the present application, the candidate set of characteristic wavelengths in step S02 is constructed by the following steps:

[0019] Step S021, inputting wavelengths with Pearson correlation coefficient |r|≧high correlation threshold to the candidate set of characteristic wavelengths;

[0020] Step S022, screening wavelengths with Pearson correlation coefficient |r|<low correlation threshold by a polarization-spectrum difference method, inputting the screened wavelengths to the candidate set of characteristic wavelengths, and completing the construction of the candidate set of characteristic wavelengths.

[0021] Further, in an embodiment of the present application, in step S022, the method of screening wavelengths with Pearson correlation coefficient |r|<low correlation threshold by a polarization-spectrum difference method refers to sequentially performing first-level screening and second-level screening on the wavelengths to be screened:

[0022] The first-level screening is that if the spectral reflectance difference of the wavelength is significant and at least one polarization difference is significant, the wavelength is input to the candidate set of characteristic wavelengths;

[0023] The second-level screening is that for wavelengths that do not meet the above conditions, further judgment is made that if the spectral reflectance difference or any one polarization difference of the wavelength is significant and other differences are close to the threshold, the wavelength is input to the candidate set of characteristic wavelengths;

[0024] Otherwise, further judgment is made that if the wavelength belongs to the absorption peak range, the wavelength is input to the candidate set of characteristic wavelengths;

[0025] The polarization difference refers to the difference in degree of polarization or the difference in polarization angle of the wavelength.

[0026] Further, in an embodiment of the present application, the spectral reflectance difference is obtained by

[0027]

[0028] is obtained, wherein, is the reflectance mean value at wavelength is the reflectance mean value of the blank sample at wavelength

[0029] The polarization difference is obtained by

[0030]

[0031] is obtained, wherein, is the polarization mean value at wavelength is the polarization mean value of the blank sample at wavelength

[0032] The polarization angle difference is obtained by

[0033]

[0034] is obtained, wherein, is the polarization angle mean value at wavelength is the polarization angle mean value of the blank sample at wavelength

[0035] Further, in an embodiment of the present application, the feature wavelength set in step S03 is constructed by the following steps:

[0036] Step S031, for all wavelengths in the feature wavelength candidate set with Pearson correlation coefficient |r|> high correlation threshold, the projection variance of each wavelength is obtained using successive projections algorithm SPA, if the number of wavelengths is greater than or equal to a preset value, the preset value of the wavelength with the largest projection variance is input into the feature wavelength set, and the construction of the feature wavelength set is completed; otherwise, all wavelengths are input into the feature wavelength set, and then step S032 is executed;

[0037] Step S032, for all wavelengths in the feature wavelength candidate set with Pearson correlation coefficient |r|≤ high correlation threshold, the wavelengths are screened by competitive adaptive reweighted sampling CARS method, and the screened wavelengths are input into the feature wavelength set.

[0038] ​​​​​​Scheme two, a pesticide residue content inversion system based on multi-dimensional optical data, comprising the following modules:

[0039] Module one, for obtaining an absorption peak range and a spectral sample set based on a spectral image of a soil to be inverted; and obtaining a Pearson correlation coefficient of each wavelength based on a spectral reflectance and a polarization parameter of each wavelength in the spectral sample set;

[0040] The polarization parameter comprises a degree of polarization and a polarization angle;

[0041] Module two, for selecting a plurality of candidate characteristic wavelengths according to the Pearson correlation coefficient r of each wavelength and the absorption peak range, and constructing a characteristic wavelength candidate set;

[0042] Module three, for screening the characteristic wavelength candidate set to obtain a characteristic wavelength set;

[0043] Module four, for fusing spectral information and polarization information of each characteristic wavelength in the characteristic wavelength set to obtain a fused characteristic vector set;

[0044] Module five, for selecting a type of supervised learning model based on the characteristic wavelength candidate set;

[0045] Module six, for training the supervised learning model obtained by module five through the characteristic vector set to obtain a pesticide residue content inversion model;

[0046] Module seven, for fusing the spectral reflectance and the polarization parameter of the soil to be inverted, and obtaining the pesticide residue content through the pesticide residue content inversion model.

[0047] Scheme three, an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete mutual communication through the communication bus;

[0048] The memory is used for storing a computer program;

[0049] The processor is used for executing the program stored on the memory to realize the steps of any one of the above methods.

[0050] Scheme four, a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the steps of any one of the above methods.

[0051] The pesticide residue content inversion method and system based on multi-dimensional optical data effectively solve the problems of existing pesticide residue inversion techniques based on hyperspectrum, such as being easily disturbed by soil state, low inversion efficiency and low precision. Specific beneficial effects include:

[0052] The pesticide residue content inversion method provided by the application fuses hyperspectral technology and polarization technology, breaks through the information limitation of traditional imaging technology, and realizes more detailed and in-depth feature extraction and analysis of a target. Through double verification of "spectrum fingerprint locking component" and "polarization feature analyzing surface", the anti-interference capability, inversion efficiency and accuracy of pesticide residue content inversion are greatly improved.

[0053] The method provided by the application is suitable for the field of pesticide residue content inversion of farmland soil. BRIEF DESCRIPTION OF DRAWINGS

[0054] The above and / or additional aspects and advantages of the application will become apparent and be readily understood from the following description, taken in connection with the accompanying drawings, in which:

[0055] Figure 1 is a flowchart of the pesticide residue content inversion method according to the first embodiment. DETAILED DESCRIPTION

[0056] The various embodiments of the application will be described in detail below with reference to the accompanying drawings. The embodiments described by reference to the accompanying drawings are exemplary and are intended to explain the application, and cannot be understood as limiting the application.

[0057] In the first embodiment, the pesticide residue content inversion method based on multi-dimensional optical data includes the following steps as shown in the figure: Figure 1

[0058] Step S01: Based on the spectral image of the soil to be inverted, the absorption peak range and the spectral sample set are obtained; based on the spectral reflectance and the polarization parameter of each wavelength in the spectral sample set, the Pearson correlation coefficient of each wavelength is obtained.

[0059] The polarization parameter includes the degree of polarization and the polarization angle.

[0060] Step S02: According to the Pearson correlation coefficient r of each wavelength and the absorption peak range, a plurality of candidate characteristic wavelengths are selected to construct a characteristic wavelength candidate set;

[0061] Step S03: The characteristic wavelength candidate set is screened to obtain a characteristic wavelength set.

[0062] Step S04: The spectral information and polarization information of each characteristic wavelength in the characteristic wavelength set are fused to obtain a fused characteristic vector set.

[0063] Step S05: Based on the characteristic wavelength candidate set, the type of the supervised learning model is selected.

[0064] ​Step S06, training the supervised learning model obtained in step S05 by the feature vector set, to obtain a pesticide residue content inversion model;

[0065] Step S07, fusing the spectral reflectance and polarization parameters of the soil to be inverted, and obtaining the pesticide residue content through the pesticide residue content inversion model.

[0066] In this embodiment, the absorption peak range in step S01 is obtained by superimposing the spectrum of the soil to be inverted and the spectrum of the soil background without the target pesticide, marking the pesticide spectrum absorption peak outside the overlapping range.

[0067] In this embodiment, the spectral reflectance is preferably pre-processed, and the polarization parameter is preferably calibrated and denoised.

[0068] The pre-processing includes SG filtering, baseline correction and normalization.

[0069] In this embodiment, the instrument system error is corrected by a standard polarizer (such as a linear polarizer) to ensure the accuracy of the degree of polarization DoP and the angle of polarization AoP.

[0070] The degree of polarization DoP is obtained by

[0071]

[0072] obtained, wherein, and are the maximum and minimum polarized light intensities at the same wavelength, respectively.

[0073] The angle of polarization AoP is obtained by

[0074]

[0075] obtained, wherein, is the measured value of the angle of polarization, is the instrument angle of polarization offset (calibrated by a standard sample).

[0076] In this embodiment, the fusion of spectral information and polarization information of the characteristic wavelength in step S04 is preferably basic fusion or enhanced fusion;

[0077] The basic fusion is performed by

[0078]

[0079] fusion, wherein, is the spectral reflectance of the nth characteristic wavelength, is the degree of polarization of the nth characteristic wavelength, is the angle of polarization of the nth characteristic wavelength.

[0080] The enhanced fusion is a derivative transformation on the spectral reflectance and a gradient transformation on the polarization parameter based on the basic fusion, further strengthening the feature sensitivity, and forming an enhanced feature vector.

[0081] In the embodiment, the spectral sample set is preferably a sample set with different pesticide contents, and the number of samples is greater than 30 to ensure statistical reliability.

[0082] In the embodiment, the Pearson correlation coefficient (PCC) is a statistical quantity for measuring the degree of linear correlation between two continuous variables, and the value range is [-1, 1], wherein,

[0083] r=1: complete positive linear correlation, that is, one variable increases synchronously with the increase of another variable;

[0084] r=-1: complete negative linear correlation, that is, one variable decreases synchronously with the increase of another variable;

[0085] r=0: no linear correlation, and the linear relationship between the two is not significant.

[0086] The Pearson correlation coefficient is obtained by

[0087]

[0088] wherein, is the standard deviation of the variable is the standard deviation of the variable is the variable and the variable is the covariance of the variables and the variable .

[0089] The covariance is used to measure the consistency of the overall change trend of the two, and the standard deviation is used to measure the dispersion degree of the variable.

[0090] In the embodiment, the Pearson correlation coefficient of each wavelength in step S01 is obtained by

[0091]

[0092] wherein, is the number of samples of the spectral sample set, is the the spectral reflectance, is the corresponding mean value, is the the polarization parameter, is the corresponding mean value.

[0093] In the embodiment, the training set and the validation set are preferably divided in a ratio of 7:3 or 8:2, 5-fold or 10-fold cross-validation is used to evaluate the generalization ability of the model and avoid overfitting.

[0094] In the embodiment, the accuracy of the single spectral feature model and the spectral-polarization fusion feature model is preferably compared to verify the gain effect of the polarization information.

[0095] In the embodiment, the obtained pesticide residue content inversion model is preferably evaluated, and the determination coefficient R2, the root mean square error RMSE and the mean absolute error MAE are selected for evaluation.

[0096] The determination coefficient R2 is

[0097] ,

[0098] wherein, y is the predicted value, y is the result value, and n is the sample number. The closer to 1 indicates that the inversion effect is better. The root mean square error RMSE is

[0099]

[0100] The smaller the value, the smaller the error. The mean absolute error MAE is

[0101]

[0102]

[0103]

[0104] Reflects the average deviation of the predicted value and the true value.

[0105] The pesticide residue content inversion method based on multi-dimensional optical data in the embodiment uses the Pearson correlation coefficient as a statistical quantity for measuring the linear correlation degree of spectral reflectance and polarization parameters, and provides high-specificity fusion features for subsequent content inversion. After selecting the characteristic wavelength of the pesticide, the core logic of combining polarization information to invert the pesticide content in the soil is to construct a fusion input vector of "spectral features + polarization features", and to establish a mapping relationship between the features and the pesticide content through a quantitative regression model.

[0106] Embodiment two, the pesticide residue content inversion method of the embodiment one is further limited, in the embodiment, in step S05

[0107] ​​​​​If the wavelengths of the characteristic wavelength candidate set are all wavelengths with Pearson correlation coefficient |r|>correlation threshold, a PLSR model is selected;

[0108] Otherwise, an SVR model is selected.

[0109] In this embodiment, the correlation threshold is preferably 0.3.

[0110] In this embodiment, the PLSR model comprises the following steps:

[0111] Step S0511, by

[0112] standardizing and , wherein , are the mean and standard deviation of X, respectively, , and the same applies.

[0113] Step S0512, the polarization parameters (DoP, AoP) are taken as column vectors of the X matrix, and the spectral parameters (such as reflectivity, derivative spectrum) of the characteristic wavelength are jointly used for latent variable extraction, and the contribution of the polarization information is reflected by the weight vector w (the polarization parameter with greater contribution to the inversion corresponds to greater weight).

[0114] Step S0513, extract the linear combination of the first pair of latent variables ( ) and ( linear combination), with the constraint condition being to satisfy , wherein is the weight vector of , is the weight vector of , and has the maximum correlation with .

[0115] Step S0514, residual error calculation is performed on and : , wherein is the loading vector of , is the loading vector of .

[0116] Step S0515, repeat steps S053 to S054 to extract latent variables until the residual error satisfies the convergence condition (such as cumulative variance contribution rate≥95%).

[0117] Step S0516, assuming that k latent variables are extracted in total, the final regression model is: where , , is the error term.

[0118] De-normalization gives the final content .

[0119] In this embodiment, the SVR model is

[0120] Let the training sample be ( the fusion feature vector, is the corresponding pesticide content.

[0121] The goal of SVR is to find a regression function ( is the kernel function mapping, is the weight vector, is the bias).

[0122] The constraint condition is to satisfy where , is the slack variable.

[0123] The optimization goal is: where C is the penalty coefficient (balance model complexity and error).

[0124] By introducing the Lagrange multiplier , the optimization problem is converted into a dual problem:

[0125] where is the kernel function.

[0126] The spectrum-polarization fusion feature is directly used as the input of the SVR, and the kernel function automatically mines the synergistic effect of the polarization information and the spectral information (such as the correction of the DoP difference on the nonlinear relationship) by calculating the similarity of the fusion features.

[0127] The commonly used radial basis kernel function (RBF) is: where γ is the kernel function parameter. By solving the dual problem, we get , and then get the regression function

[0128] .

[0129] The embodiment is a further limitation of step S05, which exemplifies the solution of step S05, which is based on the selection of a supervised model by Pearson correlation coefficient, wherein the Pearson correlation coefficient |r|>0.7 represents a linear relationship dominant scene, the wavelength and the pesticide content are approximately linearly related, and there is a multiple collinearity (such as a high correlation between the spectrum and the polarization feature), which is suitable for preliminary inversion and benchmark model construction, and therefore the PLSR model is selected.

[0130] The Pearson correlation coefficient |r|<0.3 represents a nonlinear relationship dominant scene, and when the soil matrix interference is complex (such as organic matter and water content fluctuation), the spectrum-polarization feature and the pesticide content are nonlinearly related, and the SVR is mapped to a high-dimensional space through a kernel function to establish a nonlinear regression model, and the inversion accuracy is higher.

[0131] Embodiment three, the pesticide residue content inversion method of the embodiment one is further limited, in the embodiment, the feature wavelength candidate set in step S02 is constructed by the following steps;

[0132] Step S021, the wavelengths with the Pearson correlation coefficient |r|≧high correlation threshold are input into the feature wavelength candidate set;

[0133] Step S022, the wavelengths with the Pearson correlation coefficient |r|<low correlation threshold are screened by the polarization-spectrum difference method, and the screened wavelengths are input into the feature wavelength candidate set, and the construction of the feature wavelength candidate set is completed.

[0134] In the embodiment, if the number of wavelengths is too large, a preset value can be set, and when the number of wavelengths in the feature wavelength candidate set is greater than the preset value, the input of the wavelengths into the feature wavelength candidate set is stopped.

[0135] In the embodiment, the high correlation threshold is greater than the low correlation threshold, and the high correlation threshold is preferably 0.7, and the low correlation threshold is preferably 0.3,

[0136] |r|>0.7 represents that the spectral reflectance and polarization parameters of the pesticide content in the soil are strongly linearly related at this wavelength, the spectrum and polarization signal of such wavelength have strong cooperativity, can reflect the chemical and physical properties of the pesticide, and have strong anti-interference ability;

[0137] 0.3<|r|≤0.7 represents a moderate linear correlation of the pesticide content in the soil at this wavelength;

[0138] |r|≤0.3 represents a weak linear correlation of the pesticide content in the soil at this wavelength.

[0139] In the embodiment, if the Pearson correlation coefficient r satisfies the condition: the high correlation threshold > |r| >= the low correlation threshold, the wavelength is input into the characteristic wavelength candidate set according to experience, or the wavelength is screened through the polarization-spectrum difference analysis method and input into the characteristic wavelength candidate set.

[0140] The embodiment is a further limitation of step S02, and the characteristic wavelength candidate set is exemplified. In the soil pesticide detection scene, the correlation of the spectral reflectance and the polarization parameter at the same wavelength reflects the consistency of the “chemical-physical” dual response of the wavelength to the pesticide: if |r| is large (such as |r| > 0.7), it indicates that the spectral signal and the polarization signal of the wavelength can cooperatively reflect the presence or content change of the pesticide (the pesticide changes the spectral absorption and the polarization state), and the wavelength is suitable for being used as a characteristic wavelength; if |r| is small (such as |r| < 0.3), it indicates that the spectral signal and the polarization signal of the wavelength are irrelevant, and the wavelength may be disturbed by the background (such as the soil organic matter only affects the spectrum but does not affect the polarization), and the wavelength needs to be removed or used with caution.

[0141] Therefore, the method is beneficial to better integrating the polarization parameter into the overall inversion process by constructing the characteristic wavelength candidate set based on the Pearson correlation coefficient. In the selection of the characteristic wavelength, different screening methods are applied to the wavelengths under different linear relationships by calculating the linear relationship between the polarization parameter and the spectral reflectance, so that the screened wavelengths perform more excellently in the final fusion characteristic inversion process. To prevent the problem that the effect of screening the characteristic wavelength by using the method of calculating the Pearson correlation coefficient is not good in the case of low linear correlation

[0142] Embodiment four, the pesticide residue content inversion method in the embodiment is a further limitation of the pesticide residue content inversion method in the embodiment three. In the embodiment, the method for screening the wavelength with the Pearson correlation coefficient |r| < the low correlation threshold through the polarization-spectrum difference method in step S022 is that the wavelength to be screened is sequentially subjected to primary screening and secondary screening.

[0143] The primary screening is that if the spectral reflectance difference of the wavelength is significant and at least one polarization difference is significant, the wavelength is input into the characteristic wavelength candidate set.

[0144] The secondary screening is that for the wavelength that does not satisfy the above condition, further judgment is made. If the spectral reflectance difference or any one polarization difference of the wavelength is significant, and other differences are close to the threshold, the wavelength is input into the characteristic wavelength candidate set.

[0145] Otherwise, further judgment is made. If the wavelength belongs to the absorption peak range, the wavelength is input into the characteristic wavelength candidate set.

[0146] The polarization difference value refers to a difference in degree of polarization or a difference in polarization angle.

[0147] In this embodiment, the difference in the first screening is significantly greater than three times the standard deviation, for example, the spectral reflectance difference is significant, which can be represented as At least one polarization difference is significant, which can be represented as Or ;

[0148] In the second screening, the other difference is close to the threshold value, which can be represented as a difference of about three times the standard deviation, for example, the degree of polarization difference is close to the threshold value, which can be represented as .

[0149] Wherein, is the spectral reflectance difference, is the degree of polarization difference, is the polarization angle difference, is the spectral reflectance standard deviation, is the degree of polarization standard deviation, is the polarization angle standard deviation.

[0150] In this embodiment, the blank sample does not contain pesticides; has the same soil matrix, moisture content, particle size and other conditions as the soil to be inverted; and the number of samples is at least 30 to ensure statistical significance.

[0151] The spectral data of the blank sample and the soil to be inverted should be subjected to SG filtering, baseline correction, SNV normalization; and the polarization data should be subjected to instrument calibration, noise suppression, and polarization angle correction. Ensure that the spectral data and polarization data correspond to each other in wavelength, and avoid wavelength shift leading to error in difference calculation.

[0152] This embodiment is a further limitation of step S022, and the polarization-spectral difference method is exemplified. When |r|≤low correlation threshold, the polarization-spectral difference method is used to screen characteristic wavelengths. The core logic of this method is to screen characteristic wavelengths that are sensitive to pesticides and insensitive to background interference by quantifying the spectral and polarization parameter differences between "soil samples containing pesticides" and "pure soil blank samples". This method uses the characteristics that "pesticides will change the spectral absorption and polarization state of the sample, while the parameters of the pure soil background are relatively stable", and accurately locks the wavelength band that can reflect the pesticide signal.

[0153] This method performs two screenings. The first screening (core wavelength band) must meet the "spectral difference is significant + at least one polarization difference is significant" wavelength, which takes into account the chemical absorption (spectrum) and physical structure change (polarization) of pesticides, and has the strongest specificity.

[0154] Secondary screening requires the wavelength to meet the condition of single significant difference but other differences close to the threshold. If the wavelength belongs to the characteristic absorption peak of the pesticide (such as near 1700nm), it can be included in the candidate set.

[0155] Pruning rule: the difference does not reach the threshold, or belongs to the strong absorption band of background interference (such as soil moisture 1450nm, organic matter 1680nm), and is directly pruned.

[0156] Embodiment five, this embodiment is a further limitation of the pesticide residue content inversion method of embodiment four, in this embodiment, the spectral reflectance difference is obtained by

[0157]

[0158] wherein, is the reflectance mean at wavelength , and is the reflectance mean of the blank sample at wavelength .

[0159] The polarization difference is obtained by

[0160]

[0161] wherein, is the polarization mean at wavelength , and is the polarization mean of the blank sample at wavelength .

[0162] The polarization angle difference is obtained by

[0163]

[0164] wherein, is the polarization angle mean at wavelength , and is the polarization angle mean of the blank sample at wavelength .

[0165] This embodiment is a further limitation of embodiment four, and the spectral reflectance difference, the polarization difference and the polarization angle difference are exemplified, the spectral reflectance difference is a positive value indicating that the pesticide increases the reflectivity at this wavelength (such as the specular reflection of liquid pesticide film), and a negative value indicating that the reflectivity decreases (such as the characteristic absorption of the pesticide), and the greater the absolute value, the more significant the influence of the pesticide on the spectrum.

[0166] The polarization difference , the absolute value is taken because the effect of pesticides on DoP can be an increase or a decrease, and the focus is on the size of the difference rather than the direction, the greater the absolute value, the more significant the change in the polarization state of the soil surface by the pesticide.

[0167] Polarization angle difference , the refractive index of the pesticide is different from that of the soil, which will cause the polarization angle to shift, the greater the absolute value, the more significant the shift, and the more sensitive the polarization response at this wavelength.

[0168] Embodiment six, this embodiment is a further limitation of the pesticide residue content inversion method described in embodiment one, in this embodiment, the characteristic wavelength set in step S03 is constructed by the following steps:

[0169] Step S031, for all wavelengths in the characteristic wavelength candidate set with Pearson correlation coefficient |r|> high correlation threshold, use the successive projection method SPA to obtain the projection variance of each wavelength, if the number of wavelengths is greater than or equal to the preset value, input the preset value of the wavelength with the largest projection variance into the characteristic wavelength set, complete the construction of the characteristic wavelength set; Otherwise, input all wavelengths into the characteristic wavelength set, and then execute step S032;

[0170] Step S032, for all wavelengths in the characteristic wavelength candidate set with Pearson correlation coefficient |r|≤ high correlation threshold, select them by the competitive adaptive reweighted sampling CARS method, and input the selected wavelengths into the characteristic wavelength set.

[0171] In this embodiment, the high correlation threshold is preferably 0.7.

[0172] In this embodiment, the successive projection method SPA includes:

[0173] Set the maximum number of projection steps (usually 50-100 steps), the target number of selected wavelengths (10-30);

[0174] Start from any wavelength, calculate the projection vector of each unselected wavelength and the selected wavelength, select the wavelength with the largest projection variance to join the subset, and repeat until the target number of wavelengths is reached;

[0175] Output the optimal wavelength subset (such as 1695nm, 1700nm, 1705nm, etc.) and the corresponding wavelength index.

[0176] In this embodiment, the competitive adaptive reweighted sampling CARS is iterative sampling, including

[0177] Round 1: Resample the wavelengths based on the weight coefficient (calculated from the PLSR model regression coefficient), the larger the weight, the higher the probability of being selected;

[0178] Round 2: Construct PLSR model for the sampled waveband, calculate the root mean square error of cross-validation (RMSECV);

[0179] Round 3: Eliminate the 5%~10% wavebands with the lowest weight, and repeat the iteration until the number of remaining target wavebands;

[0180] Finally, the waveband subset corresponding to the iteration round with the smallest RMSECV is selected as the final feature wavelength.

[0181] This embodiment is a further limitation of step S03, and the scheme of the feature wavelength set is exemplified. For a linear scene with Pearson correlation coefficient |r|>high correlation threshold, the wavelengths that can “maximize the explanation of the target variable (pesticide content) and minimize the redundancy between wavebands” can be screened out by vector projection operation using the successive projections algorithm SPA. Essentially, a “projection path” is found in a high-dimensional space, so that the spectral reflectance and polarization parameter variance of the projected wavelength are maximized and the correlation is minimized.

[0182] For a linear scene with Pearson correlation coefficient |r|

Claims

1. A method for inverting pesticide residue content based on multidimensional optical data, characterized in that, Includes the following steps: Step S01: Based on the spectral image of the soil to be inverted, obtain the absorption peak range and spectral sample set; Based on the spectral reflectance and polarization parameters of each wavelength in the spectral sample set, the Pearson correlation coefficient for each wavelength is obtained; The polarization parameters include the degree of polarization and the polarization angle; Step S02: Based on the Pearson correlation coefficient r and absorption peak range of each wavelength, select several candidate characteristic wavelengths to construct a candidate set of characteristic wavelengths; Step S03: Filter the candidate set of characteristic wavelengths to obtain the set of characteristic wavelengths; Step S04: For each feature wavelength in the feature wavelength set, spectral information and polarization information are fused to obtain a fused feature vector set; Step S05: Based on the candidate set of feature wavelengths, select the type of supervised learning model; Step S06: Train the supervised learning model obtained in step S05 using the feature vector set to obtain the pesticide residue content inversion model; Step S07: The spectral reflectance and polarization parameters of the soil to be inverted are fused, and the pesticide residue content is obtained through the pesticide residue content inversion model. In step S05 If all wavelengths in the candidate set of feature wavelengths are wavelengths with a Pearson correlation coefficient |r| > correlation threshold, then the PLSR model is selected; Otherwise, select the SVR model.

2. The pesticide residue content inversion method according to claim 1, characterized in that, The candidate set of characteristic wavelengths mentioned in step S02 is constructed through the following steps; Step S021: Input the wavelengths with Pearson correlation coefficients |r| ≥ high correlation threshold into the feature wavelength candidate set; Step S022: Wavelengths with Pearson correlation coefficients |r| < low correlation threshold are screened using the polarization-spectral difference method, and the screened wavelengths are input into the characteristic wavelength candidate set to complete the construction of the characteristic wavelength candidate set.

3. The pesticide residue content inversion method according to claim 2, characterized in that, The method described in step S022, which involves screening wavelengths with a Pearson correlation coefficient |r| < low correlation threshold using the polarization-spectral difference method, refers to performing first-level screening and second-level screening on the wavelengths to be screened sequentially. The first-level screening is as follows: if the spectral reflectance difference of the wavelength is significant and at least one polarization difference is significant, then the wavelength is input into the feature wavelength candidate set. The secondary screening involves further determining whether a wavelength does not meet the above conditions. If the difference in spectral reflectance or any polarization difference of the wavelength is significant, and other differences are close to the threshold, then the wavelength is input into the candidate set of feature wavelengths. Otherwise, further determination is made, and if the wavelength belongs to the absorption peak range, then the wavelength is input into the characteristic wavelength candidate set; The polarization difference refers to the difference in the degree of polarization or the difference in the polarization angle of the wavelength.

4. The pesticide residue content inversion method according to claim 3, characterized in that, The difference in spectral reflectance ,pass Obtain, among which, wavelength The average reflectance under the following conditions wavelength Mean reflectance of the blank sample; The polarization degree difference pass Obtain, among which, wavelength The mean degree of polarization under the following conditions wavelength Mean polarization degree of the blank sample below; The polarization angle difference pass Obtain, among which, wavelength The mean polarization angle under the following conditions wavelength The mean polarization angle of the blank sample.

5. The pesticide residue content inversion method according to claim 1, characterized in that, The feature wavelength set mentioned in step S03 is constructed through the following steps; Step S031: For all wavelengths in the candidate set of feature wavelengths where the Pearson correlation coefficient |r| > the high correlation threshold, use the continuous projection method SPA to obtain the projection variance of each wavelength. If the number of wavelengths is greater than or equal to a preset value, then input the preset number of wavelengths with the largest projection variance into the set of feature wavelengths to complete the construction of the set of feature wavelengths. Otherwise, input all wavelengths into the set of characteristic wavelengths, and then execute step S032; Step S032: For all wavelengths in the candidate set of feature wavelengths with a Pearson correlation coefficient |r| ≤ a high correlation threshold, a competitive adaptive reweighted sampling CARS method is used for screening, and the screened wavelengths are input into the feature wavelength set.

6. A pesticide residue content inversion system based on multidimensional optical data, characterized in that, Includes the following modules: Module 1 is used to obtain the absorption peak range and spectral sample set based on the spectral image of the soil to be inverted; and to obtain the Pearson correlation coefficient for each wavelength based on the spectral reflectance and polarization parameters of each wavelength in the spectral sample set. The polarization parameters include the degree of polarization and the polarization angle; Module 2 is used to select several candidate characteristic wavelengths based on the Pearson correlation coefficient r and absorption peak range of each wavelength, and construct a candidate set of characteristic wavelengths. Module 3 is used to filter the candidate set of characteristic wavelengths to obtain a set of characteristic wavelengths; Module 4 is used to fuse spectral and polarization information for each feature wavelength in the feature wavelength set to obtain a fused feature vector set; Module 5 is used to select the type of supervised learning model based on the set of candidate feature wavelengths; Module 6 is used to train the supervised learning model obtained in Module 5 using the feature vector set to obtain a pesticide residue content inversion model; Module 7 is used to fuse the spectral reflectance and polarization parameters of the soil to be inverted, and obtain the pesticide residue content through the pesticide residue content inversion model; Module 5 further includes: The unit is used to select the PLSR model when all wavelengths in the candidate set of characteristic wavelengths are wavelengths with a Pearson correlation coefficient |r| > the correlation threshold; otherwise, the SVR model is selected.

7. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-5.

Images