Insect biology parameter inversion method based on feature selection

Abstract

The application discloses an insect biological parameter inversion method based on feature selection; the application can be used for high-precision inversion of insect size parameters, and can help improve the insect species identification accuracy; the application firstly calculates multi-dimensional scattering features of insects, and based on a stepwise regression method, filters and respectively gives the best inversion feature combination for insect weight, body length and body width, and then based on the filtered feature combination, uses a random forest regression method to realize high-precision inversion of the insect size parameters; the effectiveness of the proposed inversion method is verified through measured data of a total of 76 species and 366 insects.

Description

Technical Field

[0001] This invention relates to the field of insect radar technology, and more specifically to a method for inverting insect biological parameters based on feature selection. Background Technology

[0002] Migration is an adaptive strategy evolved by insects and other animals to cope with seasonal or unpredictable changes in their habitats. Migratory insects facilitate the exchange of matter and energy between different regions. Furthermore, many migratory species are important vectors of economic pests or plant disease viruses, causing severe crop losses and profoundly impacting human life. To predict pests in a timely manner, formulate effective control strategies in advance, and prevent outbreaks, specialized pest monitoring is necessary. Radar, due to its ability to operate around the clock and in all weather conditions, and its wide detection range, has become a valuable tool for monitoring insect migration.

[0003] In radar observation, accurately estimating insect biological parameters such as weight, length, and width using radar echoes helps in insect species identification. Currently, a common method for estimating insect biological parameters in insect radar is to utilize the monotonic mapping relationship between the insect's RCS characteristics at a single frequency (mainly 9.4 GHz) and the biological parameters, obtaining empirical formulas for parameter estimation through polynomial fitting. However, traditional inversion methods use limited useful information (1-2 features), are simple, and have limited estimation accuracy. To further improve estimation accuracy, machine learning methods based on multidimensional scattering features have been studied for insect biological parameter estimation; however, these features are not screened, so the improvement is not significant.

[0004] Selecting features for a specific task is a crucial prerequisite for improving machine learning performance. By eliminating unimportant features, we can not only reduce computational load and the difficulty of the learning task, but also help guide the accurate estimation of insect biological parameters.

[0005] Therefore, combining feature selection and machine learning methods can fully utilize the multidimensional scattering information of insects, improve inversion accuracy, and enhance the radar's ability to identify insect species. Summary of the Invention

[0006] In view of this, the present invention provides a high-precision inversion algorithm for insect biological parameters based on feature selection. First, the optimal combination of feature parameters for body weight, body length, and body width is selected through stepwise regression. Then, random forest regression is used to learn and construct the optimal inversion models for these three types of biological parameters respectively, thereby achieving high-precision inversion of insect biological parameters.

[0007] To achieve the above objectives, the technical solution of the present invention includes the following steps:

[0008] Step 1: Extract the RCS scattering features and polarization invariant features of the insect target.

[0009] Step 2: Select the optimal scattering feature set using the forward stepwise regression algorithm; the specific process of the stepwise regression algorithm is as follows:

[0010] S201: Set the significance level as α E and α R The statistical significance level is represented by the p-value of the F-test; at each step, the p-value of the F-test is calculated to test whether there is a candidate variable.

[0011] S202: Fit each univariate model according to the set significance level, where the p-value is less than α. E Among the variables, the variable with the smallest p-value is selected as the optimal single variable. The optimal single variable is then incorporated into the stepwise model, thereby constructing the initial model.

[0012] S203: After constructing the initial model, continue to add new variables to the initial model, resulting in a p-value lower than α. E Each time a new variable is added, we take a step back to see if the new variable's entry into the stepwise model affects the significance of the old variables. That is, after adding a new variable, we check the p-value of the old variables. If the p-value is greater than α... R If so, it will be removed from the stepwise model.

[0013] S204: Repeat S203 above until adding a new variable cannot produce an output lower than α. E The p-value is calculated, and then the stepwise regression algorithm is terminated. Through stepwise regression, the optimal combination of variables for weight, length, and width is obtained respectively.

[0014] Step 3: After selecting the optimal combination of variables for weight, length, and width, these variables are combined to form a dataset, and training and test sets are constructed according to a set ratio. The training set is then subjected to regression using the random forest method to learn the best-fit model. Finally, the accuracy is tested on the test set, and the inversion accuracy of the insect biological parameters is statistically analyzed.

[0015] Furthermore, in step one, the RCS scattering features and polarization invariant features of the insect target are extracted, including 8 RCS scattering features and 3 polarization invariant features; specifically:

[0016] The eight RCS scattering features include three relevant parameters a0, a1, and a2 related to the insect's RCS, two parameters α2 and α4 characterizing the shape of the polarization pattern, and the maximum value σ of the insect's polarization pattern. xx and minimum value σ yy , σ xx and σ yy The ratio σ r .

[0017] The three polarization invariant features include: the determinant d, the trace Trg, and the Frobenius norm of the insect Graves power scattering matrix GPSM.

[0018] Further, in step one, the RCS scattering features and polarization invariant features of the insect target are extracted, including 8 RCS scattering features and 3 polarization invariant features. The above 11 features are squared, expanding the number of features to 22. Among them, the 8 RCS scattering features include 3 related parameters a0, a1 and a2 related to the insect's RCS, 2 parameters α2 and α4 characterizing the shape of the polarization pattern, and the maximum value σ of the insect's polarization pattern. xx and minimum value σ yy , σ xx and σ yy The ratio σ r ;

[0019] The three polarization invariant features include: the determinant d, the trace Trg, and the Frobenius norm of the insect Graves power scattering matrix GPSM.

[0020] Furthermore, the eight RCS scattering features and three polarization invariant features were extracted using the following method:

[0021] For a single-station fully polarimetric radar, the scattering matrix SM of the insect target is obtained directly. The scattering matrix SM is represented by S in formula (1):

[0022]

[0023] In the formula, s 11 s 12 s 21 and s 22 Let HH, HV, VH, and VV represent the square roots of the polarization RCS of the scattering matrix of the insect target, respectively, in meters. 2 ;β is s 12 With s 11 The relative phase; β' is s 21 With s 11 The relative phase; γ is s 22 With s 11 The relative phase;

[0024] Since insect radar is single-station, s 12 =s 21 , β=β'; Next, the insect pattern σ(α) measured by the fully polarimetric radar is calculated using SM:

[0025]

[0026] Where h(α) is the normalized effective length of the radar antenna; α is the polarization direction of the antenna; θ1 is the insect's orientation; and θ2 is the harmonic angle of θ1.

[0027] Based on the polarization pattern, three relevant parameters a0, a1, and a2 related to the insect's RCS are obtained:

[0028]

[0029]

[0030]

[0031] Based on a0, a1, and a2, the two parameters α2 and α4 that characterize the shape of the polarization pattern can be calculated:

[0032]

[0033]

[0034] Where Δφ is the relative phase, defined as the phase difference between two eigenvalues ​​of SM; σ xx σ represents the maximum value of the insect polarization pattern. yy To obtain the minimum value, both are calculated from a0, a1, and a2:

[0035] σ xx =a0+a1+a2 (8)

[0036] σ yy =a0-a1+a2 (9)

[0037] σ xx and σ yy The ratio is defined as σ r :

[0038]

[0039] Furthermore, the determinant d of the insect Graves power scattering matrix GPSM is highly correlated with the insect's body size; the trace and Frobenius norm of GPSM are also invariants, just like the determinant d; the determinant d, trace Trg, and Frobenius norm are the three polarization invariants, expressed as:

[0040]

[0041] Trg=σ xx +σ yy (12)

[0042]

[0043] This yielded 11 scattering characteristics, including α0~α4, α xx α yy σ r , d, Trg and Frobenius norms.

[0044] To enrich the information dimensions, it is required to square the above 11 scattering features, expanding the number of features to 22.

[0045] Furthermore, in step two, the significance level is set to α. E =0.05, α R =0.10.

[0046] Furthermore, in step three, the training set and the test set are constructed according to the set ratio, specifically: the training set and the test set are constructed according to the ratio of 75% and 25%, respectively.

[0047] Beneficial effects:

[0048] This invention proposes a high-precision inversion method for insect biological parameters based on feature selection. This method can screen out combinations of independent variables (RCS scattering features) that have explanatory significance for the dependent variable (biological parameters). While increasing the utilization rate of insect scattering information, it reduces redundant feature information, and based on the random forest regression method, it can achieve high-precision estimation of insect biological parameters, laying a solid foundation for radar-based insect species identification. Attached Figure Description

[0049] Figure 1 This is an iterative flowchart for feature selection;

[0050] Figure 2 This is a flowchart of the random forest regression inversion process. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0052] This invention provides

[0053] Step 1: Feature Extraction

[0054] First, in order to select features, it is necessary to calculate the eight RCS scattering features and three polarization invariant features commonly used in insect radar science, as shown in Table 1.

[0055] Table 1 Extracted Feature Parameters

[0056]

[0057]

[0058] For a single-station fully polarimetric radar, the scattering matrix SM of insect targets can be obtained directly, usually expressed by formula (1):

[0059]

[0060] In the formula, s 11 s 12 s 21 and s 22 Represent the RCS of HH, HV, VH and VV polarizations respectively (unit: m) 2 The square root of ). β, β', and γ are the relative phases with reference to the HH phase, where β is the square root of s. 12 With s 11 The relative phase; β' is s 21 With s 11 The relative phase; γ is s 22 With s 11 The relative phase; Since insect radar is monostationary, s 12 =s 21 , β=β'. Next, the insect pattern measured by the fully polarimetric radar is calculated using SM:

[0061]

[0062] Where h(α) is the normalized effective length of the radar antenna; α is the polarization direction of the antenna; θ1 is the insect's orientation; and θ2 is the harmonic angle of θ1.

[0063] Based on the polarization pattern, three parameters a0, a1, and a2 related to the insect's RCS can be obtained:

[0064]

[0065]

[0066]

[0067] Based on a0, a1, and a2, the two parameters α2 and α4 that characterize the shape of the polarization pattern can be calculated:

[0068]

[0069]

[0070] Δφ represents the relative phase, defined as the phase difference between two eigenvalues ​​of SM. σ xx σ represents the maximum value of the insect polarization pattern. yy To obtain the minimum value, both can be calculated from a0, a1, and a2:

[0071] σxx =a0+a1+a2 (8)

[0072] σ yy =a0-a1+a2 (9)

[0073] Their ratio is defined as σ r :

[0074]

[0075] Furthermore, the determinant *d* of the insect Graves power scattering matrix (GPSM) is highly correlated with the insect's body size. The trace and Frobenius norm of the GPSM are also invariants, just like the determinant *d*. The determinant *d*, trace *Trg*, and Frobenius norm can be expressed as:

[0076]

[0077] Trg=σ xx +σ yy (12)

[0078]

[0079] To date, all 11 scattering features have been calculated. To enrich the information dimensions, the squares of these 11 scattering features are taken, expanding the number of features to 22.

[0080] Step 2: Feature Filtering

[0081] This invention employs a forward stepwise regression algorithm to select the optimal scattering feature set. The general idea of ​​forward stepwise regression is to build a regression model from a set of candidate predictor variables by progressively adding variables to the model until there is no reasonable reason to add new variables. In each step, a variable is considered for addition to the candidate variable set based on certain pre-specified conditions. This invention starts with a model containing no variables, then progressively adds variables that are most significant according to the F-test, whose addition will statistically bring the most significant improvement in fit, and repeats this process until all important variables are included in the model. The iterative flowchart of the entire feature selection process is as follows: Figure 1 As shown, the detailed process of stepwise regression can be summarized as follows.

[0082] (1) Set the significance level. To determine when to add or remove a variable in the stepwise regression model, the significance level is set using α. E and α R The statistical significance level of this invention is represented by the p-value of the F-test. At each step, the p-value of the F-test is calculated to test whether a candidate variable is present or absent. This invention sets α... E =0.05, αR =0.10.

[0083] (2) Fit the initial model. Once the significance level is specified, then fit each univariate model, i.e., regress y in x1, y in x2, ... and other candidate variables. When the p-value is less than α... E Among the variables, the first variable to enter the stepwise model is the one with the smallest p-value, which is considered the "best" single variable.

[0084] (3) Iterative testing and regression. After building the initial model, new variables are added to the model until the p-value is lower than α. E =0.05. For each new variable added, step back one step to see if the new variable's inclusion in the stepwise model affects the significance of the old variables. That is, after adding a new variable, check the p-value of the old variables. If the p-value is greater than α... R If the value is 0.10, it is removed from the stepwise model.

[0085] (4) End the regression. Continue with step 3 above until adding a new variable does not produce an outcome lower than α. E =0.05 p-value.

[0086] Through stepwise regression, the optimal combination of variables for weight, length, and width can be obtained respectively.

[0087] Step 3: Random Forest Regression

[0088] After selecting the optimal combination of variables for weight, body length, and body width, these variables are combined to form a dataset, and training and test sets are constructed in a 75% and 25% ratio, respectively. Random forest regression is applied to the training set to learn the best-fit model. Finally, accuracy testing is performed on the test set, and the inversion accuracy of insect biological parameters is statistically analyzed. Specific implementation steps are as follows: Figure 2 As shown.

[0089] Example:

[0090] To verify the effectiveness of the insect biological parameter inversion method based on feature selection in this invention, inversion accuracy was tested using single-frequency (GHz) data of 366 insects from 76 species. First, 22-dimensional scattering features of all insects were extracted based on step one. Second, a stepwise regression method was used to extract the optimal combination of feature variables for body weight, body length, and body width from the insect dataset. Finally, based on the selected variable combinations, a random forest regression method was used for model fitting and inversion accuracy testing. The iterative results of the optimal inversion feature combinations for biological parameters obtained after screening are shown in Table 2. The first step of the regression step is the best single feature obtained after verification, and the last step of the regression step is the optimal feature combination selected through feature screening.

[0091] Table 2 Iterative results of optimal feature combinations for insect biological parameters

[0092]

[0093] The above results indicate that the optimal feature combination for weight inversion in this insect dataset is σ. yy , α2, a0, t, a1, The optimal feature combination for body length inversion is σ yy , α2, f,a1; The optimal feature combination for volume width inversion is (σ yy ) 2 The dataset, a2, α2, is constructed based on the optimal feature combination. It is then divided into training and test sets at a ratio of 75% and 25%, respectively. The inversion accuracy is verified using the random forest regression method.

[0094] The final verification result showed that the average relative error of weight inversion was 18.83% (R0). 2 =0.9837), the average error in body length is 11.37% (R 2 =0.9623), and the average relative error of body width is 16.87% (R). 2 =0.7157), this result is better than traditional univariate / binary fitting methods and machine learning methods without feature selection.

[0095] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for inverting insect biological parameters based on feature selection, characterized in that, Includes the following steps Step 1: Extracting the target insect. RCS Scattering characteristics and polarization invariant characteristics; in step one, the insect target is extracted. RCS Scattering characteristics and polarization invariant characteristics, including 8 RCS Scattering characteristics and three polarization invariant characteristics; Specifically: 8 RCS Scattering characteristics include insects RCS Three relevant parameters , and Two parameters characterizing the shape of the polarization pattern and The maximum value of the insect polarization pattern and minimum value , and ratio ; The three polarization invariant features include: insects Graves Power scattering matrix GPSM determinant d ,trace Trg and Frobenius Norm, and square the above 11 features, expanding the number of features to 22; Step 2: Select the optimal scattering feature set using the forward stepwise regression algorithm; the specific process of the stepwise regression algorithm is as follows: S 201: Set the significance levels as follows: and The statistical significance level was changed from F Inspection Value representation; at each step, F Inspection The value is calculated to test whether or not a candidate variable exists; S 202: Fit each univariate model according to the set significance level. Value less than Among the variables, with The variable with the smallest value is the best single variable, which is then incorporated into the stepwise model to construct the initial model. S 203: After constructing the initial model, continue to add new variables to the initial model to obtain... Value lower than Each time a new variable is added, we take a step back to see if the new variable's inclusion in the stepwise model affects the significance of the old variables. In other words, after adding a new variable, we check the significance of the old variables. Value, if Value greater than If so, remove it from the stepwise model; S 204: Repeat the above. S 203, until adding a new variable cannot produce a value lower than 103. of The value is then determined, and the stepwise regression algorithm ends. Through stepwise regression, the optimal combination of variables for weight, length, and width is obtained respectively. Step 3: After selecting the optimal combination of variables for weight, length, and width, these variables are combined to form a dataset, and training and test sets are constructed according to a set ratio. The training set is then subjected to regression using the random forest method to learn the best-fit model. Finally, the accuracy is tested on the test set, and the inversion accuracy of the insect biological parameters is statistically analyzed.

2. The method for inverting insect biological parameters based on feature selection as described in claim 1, characterized in that, The eight RCS Scattering features and three polarization invariant features were extracted using the following method: For a single-station fully polarimetric radar, the scattering matrix of insect targets can be obtained directly. SM scattering matrix SM Using formula (1) S express: （1） In the formula, , , and These represent the scattering matrices of the insect targets, respectively. HH , HV , VH and VV polarization RCS The square root, unit: ; for and The relative phase; for and The relative phase; for and The relative phase; Because insect radar is single-station, , Next, using SM Calculate the insect pattern obtained by fully polarimetric radar : （2） in, The normalized effective length of the radar antenna; This represents the polarization direction of the antenna. Orientation of the insect; for The harmonic angle; Based on the polarization pattern, we obtain the relationship with insects. RCS Three relevant parameters , and : （3） （4） （5） according to , and Two parameters characterizing the shape of the polarization pattern can be calculated. and : （6） （7） in For relative phase, it is defined as SM Phase difference between two eigenvalues; This represents the maximum value of the insect polarization pattern. For the minimum value, both are derived from , and The calculation yielded: （8） （9） and The ratio is defined as : （10） In addition, insects Graves Power scattering matrix GPSM determinant Highly correlated with insect body size; GPSM traces and Frobenius Norm and determinant Both are invariants; determinant d ,trace Trg and Frobenius Norms are the three polarization invariants, expressed as: （11） （12） （13） This yields 11 scattering characteristics, including α 0 ~α 4 、α xx , α yy , , d , Trg and Frobenius Norm; To enrich the information dimensions, it is required to square the above 11 scattering features, expanding the number of features to 22.

3. The insect biological parameter inversion method based on feature selection as described in claim 1 or 2, characterized in that, In step two, the significance level is set to =0.05, =0.

10.

4. The insect biological parameter inversion method based on feature selection as described in claim 1, characterized in that, In step three, the training set and the test set are constructed according to the set ratio, specifically: the training set and the test set are constructed according to the ratio of 75% and 25%, respectively.

Images