Plant Growth Parameter Inversion Method Based on Polarimetric SAR Images and Semi-Supervised Regression

Abstract

The present invention discloses a method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression, comprising: obtaining the original data of polarimetric SAR images containing at least one point to be measured; each point to be measured represents a target plant area; performing polarimetric target decomposition on the original data to obtain multi-dimensional features of each point to be measured; selecting the features of the target dimension from the multi-dimensional features of each point to be measured to obtain the key features of each point to be measured; the label of each sample point is the target growth parameter of the target plant in the target plant area represented by the sample point; using a trained random forest regressor to predict the target growth parameter according to the key features of each point to be measured, so as to obtain the target growth parameter of the target plant; the trained random forest regressor is trained by using training samples augmented by a semi-supervised learning algorithm. The present invention can effectively improve the prediction accuracy of plant growth parameters and has high robustness.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image ground object parameter inversion, and specifically relates to a plant growth parameter inversion method based on polarimetric SAR images and semi-supervised regression. Background Art

[0002] The growth parameters of crops (plant height, fresh weight, and water content) are important indicators for monitoring crop growth and estimating yield. Most traditional crop growth parameter measurement methods are based on point measurements. Since their measurement results do not have spatiotemporal continuity, they are not suitable for large-scale dynamic detection of crop growth parameters. At the same time, traditional manual measurement methods are too time-consuming and labor-intensive. With the development of remote sensing technology, remote sensing satellites can quickly obtain a wide range of ground information. Therefore, it is of great significance to use remote sensing to obtain crop growth parameters. Optical remote sensing satellites are restricted by weather conditions and cannot guarantee the timely acquisition of key remote sensing data, which hinders the monitoring of crop growth during the critical growth period. Synthetic aperture radar, as an active microwave remote sensing technology, not only has a high spatial resolution, but also can provide all-day, all-weather polarization information with richer information than traditional optical remote sensing and passive microwave remote sensing, which is more conducive to the study of crop growth parameter inversion.

[0003] In the study of crop growth parameter inversion, paddy field crops have been widely studied and the results achieved are relatively ideal, but dryland crops such as corn and wheat have been less studied and the monitoring effect is not ideal. Since the acquisition of wheat growth parameters is time-consuming and labor-intensive, current research focuses on small sample regression tasks, and deep learning algorithms driven by large amounts of data are difficult to play a role. When the sample size is extremely small and the growth parameters and input features are low in correlation, many machine learning algorithms such as support vector machine regression and ridge regression are also difficult to achieve good results.

[0004] As an ensemble learning method with strong generalization performance, random forest regression has been proven to fit nonlinear data well, and can still play a good role in the case of few samples and low correlation between features and parameters. In some studies combining optical image and SAR image data, the random forest regression method has achieved relatively satisfactory results. However, when the optical image cannot be obtained smoothly due to weather constraints, when it is difficult to obtain rich multi-source data and the sample size is very small, the effect of using the random forest regression method is still not good. Therefore, the wheat growth parameter inversion method that is less dependent on data multi-source and more robust still needs to be studied. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a plant growth parameter inversion method based on polarimetric SAR images and semi-supervised regression.

[0006] The technical problem to be solved by the present invention is realized through the following technical solutions:

[0007] The present invention provides a method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression, including:

[0008] Obtain the original data of polarimetric SAR images containing at least one measurement point; each measurement point represents a target plant area;

[0009] Perform polarimetric target decomposition on the original data to obtain multi-dimensional features of each measurement point;

[0010] Select features of the target dimension from the multi-dimensional features of each measurement point to obtain the key features of each measurement point; wherein, the target dimension is obtained by performing feature selection processing on the multi-dimensional features of labeled sample points in polarimetric SAR sample images using the recursive feature elimination method with cross-validation; the label of each sample point is the target growth parameter of the target plant in the target plant area represented by the sample point;

[0011] Use a trained random forest regressor to predict the target growth parameter according to the key features of each measurement point to obtain the target growth parameter of the target plant; the trained random forest regressor is trained using training samples augmented by a semi-supervised learning algorithm.

[0012] Compared with the prior art, the beneficial effects of the present invention are:

[0013] In the case where the polarimetric SAR image is used as a single data source, the present invention obtains multi-dimensional features through polarimetric target decomposition, fully excavates the information contained in the polarimetric SAR image data, alleviates the difficulties caused by the single data source, and furthermore, has less dependence on data multi-source. When the correlation between the features obtained by polarimetric target decomposition and the growth parameters is low, and the multi-dimensional features obtained by polarimetric target decomposition are highly redundant, the present invention selects key features through the recursive feature elimination method with cross-validation, weakens the influence of mutual interference between multi-dimensional features, and effectively improves the prediction accuracy of plant growth parameters. When the number of labeled growth parameter samples is small, the present invention uses training samples augmented by a semi-supervised learning algorithm to train the prediction model, so that the training effect of the prediction model can be better, and the predicted plant growth parameters are more accurate and have higher robustness.

[0014] The following will further describe the present invention in detail with reference to the accompanying drawings and specific embodiments. Description of the Drawings

[0015] Figure 1It is a schematic flowchart of a method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression provided by an embodiment of the present invention;

[0016] Figure 2 It is a schematic diagram of a polarimetric SAR image collected in a certain area in the simulation experiment provided by an embodiment of the present invention;

[0017] Figure 3 It is a schematic diagram showing the distribution of 35 small winter wheat plots used in the simulation experiment provided by an embodiment of the present invention;

[0018] Figure 4 It is a result graph of 10-fold cross-validation of the fresh weight parameter of winter wheat in the simulation experiment provided by an embodiment of the present invention;

[0019] Figure 5 It is a result graph of 10-fold cross-validation of the plant height parameter of winter wheat in the simulation experiment provided by an embodiment of the present invention;

[0020] Figure 6 It is a result graph of 10-fold cross-validation of the water content parameter of winter wheat in the simulation experiment provided by an embodiment of the present invention. Detailed implementation manners

[0021] The present invention will be further described in detail below in conjunction with specific embodiments, but the implementation manners of the present invention are not limited thereto.

[0022] Figure 1 It is a schematic flowchart of a method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression provided by an embodiment of the present invention, as Figure 1 shown, the method includes:

[0023] S101. Obtain the original data of the polarimetric SAR image containing at least one measurement point; each measurement point represents a target plant area.

[0024] Here, the target plant can be any plant, for example, it can be dryland crops, such as wheat, corn, etc. The polarimetric SAR image is a polarimetric SAR image of a certain area, and a target plant area in this area is a point in the polarimetric SAR image. The target plant area is a small plot in the target plant planting area, for example, a small plot in a farmland or a test field.

[0025] S102. Perform polarimetric target decomposition on the original data to obtain multi-dimensional features of each measurement point.

[0026] Here, by performing polarimetric target decomposition on the original data of the polarimetric SAR image, multi-dimensional features of each point in the polarimetric SAR image can be obtained. Therefore, when there are points to be measured in the polarimetric SAR image, multi-dimensional features of each point to be measured can be obtained.

[0027] S103. Select the features of the target dimension from the multi-dimensional features of each point to be measured to obtain the key features of each point to be measured; among them, the target dimension is obtained by performing feature selection processing on the multi-dimensional features of the labeled sample points in the polarimetric SAR sample image using the recursive feature elimination method of cross-validation; the label of each sample point is the target growth parameter of the target plant in the target plant area represented by the sample point.

[0028] Here, the target growth parameter can be any plant growth parameter. For example, it can be plant height, fresh weight, water content, etc.

[0029] S104. Use the trained random forest regressor to predict the target growth parameter according to the key features of each point to be measured, and obtain the target growth parameter of the target plant; the trained random forest regressor is trained using the training samples augmented by the semi-supervised learning algorithm.

[0030] Here, the training samples used to train the random forest regressor can be the training samples augmented by the improved COREG semi-supervised learning algorithm.

[0031] In the present invention, S102 is implemented through the following steps:

[0032] S1021. Perform filtering processing on the original data to obtain the polarimetric covariance matrix C.

[0033] For example, the original data can be filtered using a refined lee 3*3 filtering window to obtain the polarimetric covariance matrix C. Here, since the original data is a polarimetric covariance matrix, the data obtained after filtering is still a polarimetric covariance matrix.

[0034] S1022. Extract the diagonal elements of the polarimetric covariance matrix C to obtain the backscattering coefficients of each point in the polarimetric SAR image.

[0035] Here, the backscattering coefficient of each point is a 3D feature.

[0036] S1023. Perform Freeman-Durden decomposition, Cloude decomposition, non-negative eigenvalue decomposition, TSVM decomposition, and Yamaguchi decomposition on the polarimetric covariance matrix C respectively to obtain different features of different dimensions of each point.

[0037] Specifically, performing Freeman-Durden decomposition on the polarization covariance matrix C can obtain three-dimensional features, namely surface scattering feature, double-bounce scattering feature, and volume scattering feature; performing Cloude decomposition on the polarization covariance matrix C can obtain three-dimensional features, namely scattering entropy, average scattering angle, and anisotropy; performing non-negative eigenvalue decomposition on the polarization covariance matrix C can obtain three-dimensional features, namely odd-bounce scattering feature, even-bounce scattering feature, and volume scattering feature; performing TSVM decomposition on the polarization covariance matrix C can obtain four-dimensional features, namely helicity, maximum amplitude, and two polar coordinates describing the symmetric scattering type; performing Yamaguchi decomposition on the polarization covariance matrix C can obtain four-dimensional features, namely surface scattering feature, double-bounce scattering feature, volume scattering feature, and helix scattering feature. It should be noted that when using different decomposition methods to perform eigen-decomposition on the polarization covariance matrix C, even if the same-named features are obtained through decomposition, the specific values are different. For example, although the surface scattering feature can be obtained through both Freeman-Durden decomposition and Yamaguchi decomposition of the polarization covariance matrix C, the values of the surface scattering feature obtained are different; the same principle applies to the double-bounce scattering feature and the volume scattering feature.

[0038] S1024. Use the backscattering coefficient of each point to be measured and the feature set composed of different features in different dimensions as the multi-dimensional features of each point to be measured.

[0039] For each point to be measured, use the backscattering coefficient (three-dimensional feature) of this point to be measured and the feature set composed of the 17-dimensional features obtained through Freeman-Durden decomposition, Cloude decomposition, non-negative eigenvalue decomposition, TSVM decomposition, and Yamaguchi decomposition as the multi-dimensional features of this point to be measured, that is, the 20-dimensional features of this point to be measured are obtained.

[0040] In the present invention, the target dimension represents multiple dimension numbers in the dimension numbers of the multi-dimensional features. For example, when the multi-dimensional feature is a 20-dimensional feature, the dimension numbers of the multi-dimensional feature are from 1 to 20, and the target dimension is multiple different integers from 1 to 20. Therefore, when screening the important features of each point, the features corresponding to the target dimension can be used as the key features of this point.

[0041] In the present invention, before performing the above S103, the following steps are further included:

[0042] S01. Obtain the original sample data of the polarimetric SAR sample image, and perform polarimetric target decomposition on the original sample data to obtain the multi-dimensional features of each labeled sample point.

[0043] Here, the polarimetric SAR sample image is a polarimetric SAR image serving as the sample image. The area represented by this polarimetric SAR sample image may be the same as or different from the area represented by the above-mentioned polarimetric SAR image containing the point to be measured. The present invention does not limit this.

[0044] Here, the specific principle of performing polarimetric target decomposition on the original sample data is the same as that of S102 above, and will not be elaborated here.

[0045] S02. Using the 10-fold cross-validation method, divide the set F composed of the multi-dimensional features of multiple labeled sample points and the labels of these multiple labeled sample points into 10 initial training sets and 10 initial test sets corresponding one by one to these 10 initial training sets, and create a dimension set Dim that records the dimension numbers of all features in the multi-dimensional features.

[0046] In the present invention, both the initial training set and the initial test set are composed of multiple samples. Each sample is composed of the multi-dimensional features of a sample point and the label corresponding to this sample point.

[0047] For example, when the multi-dimensional feature is a 20-dimensional feature, the elements in the created dimension set Dim are positive integers from 1 to 20.

[0048] S03. Before performing the p-th round of deletion processing on the initial training set, the initial test set, and Dim, obtain the training set of the (p - 1)-th round, the test set of the (p - 1)-th round, and the dimension set Dim of the (p - 1)-th round p-1 ; each sample point in the training set of the (p - 1)-th round and the test set of the (p - 1)-th round has M-dimensional features; where p is an integer greater than or equal to 1. When p is 1, the training set of the (p - 1)-th round is the 10 initial training sets, the test set of the (p - 1)-th round is the 10 initial test sets, the dimension set of the (p - 1)-th round is Dim, and M = N, where N represents the total dimension of each multi-dimensional feature.

[0049] S04. When M is equal to 1, use the dimension number in Dim p-1 as the target dimension; when M is greater than 1, perform the p-th round of deletion processing, and in the x-th loop of the p-th round of deletion processing, use the x-th dimensional feature in each M-dimensional feature as the feature to be deleted in the x-th time. Based on the initial random forest regressor, the training set of the (p - 1)-th round, and the test set of the (p - 1)-th round, determine the x-th coefficient change amount to obtain M coefficient change amounts in the p-th round of deletion processing; x is a positive integer greater than or equal to 1, and the value of x ranges from 1 to M.

[0050] Specifically, the "determine the x-th coefficient change amount based on the initial random forest regressor, the training set of the (p - 1)-th round, and the test set of the (p - 1)-th round" in S04 above is implemented through the following steps:

[0051] Step a: Delete the feature to be deleted at the x-th time in the training set of the (p - 1)-th round and the test set of the (p - 1)-th round respectively, to obtain the training set of the x-th time in the (p - 1)-th round and the test set of the x-th time in the (p - 1)-th round.

[0052] Step b: Use the training set of the (p - 1)-th round and the training set of the x-th time in the (p - 1)-th round respectively to train the initial random forest regressor, to obtain the first random forest regressor of the x-th time and the second random forest regressor of the x-th time.

[0053] Step c: Use the first random forest regressor of the x-th time to predict the test set of the (p - 1)-th round, and calculate R 2 Coefficient of determination, to obtain the first R of the x-th time 2 Coefficient.

[0054] Step d: Use the second random forest regressor of the x-th time to predict the test set of the x-th time in the (p - 1)-th round, and calculate R 2 Coefficient of determination, to obtain the second R of the x-th time 2 Coefficient.

[0055] Step e: Take the difference between the second R of the x-th time 2 Coefficient and the first R of the x-th time 2 Coefficient as the change amount of the x-th coefficient.

[0056] S05: When the M change amounts of coefficients meet the preset coefficient change condition, take the dimension serial number in Dim p-1 as the target dimension; when the M change amounts of coefficients do not meet the preset coefficient change condition, perform feature deletion processing on the training set of the (p - 1)-th round, the test set of the (p - 1)-th round and Dim p-1 to obtain the training set of the p-th round, the test set of the p-th round and the dimension set Dim of the p-th round p-1 , and let p = p + 1 and then perform the deletion processing of the (p + 1)-th round until the target dimension is obtained.

[0057] Specifically, when there is a coefficient change amount greater than or equal to 0 among the M change amounts of coefficients, it indicates that the M change amounts of coefficients do not meet the preset coefficient change condition; while when there is no coefficient change amount greater than or equal to 0 among the M change amounts of coefficients, it indicates that the M change amounts of coefficients meet the preset coefficient change condition.

[0058] Specifically, the above S05 includes the following steps:

[0059] Step A: When there is a coefficient change amount greater than or equal to 0 among the M change amounts of coefficients, it indicates that the M

[0060] Step B: Among the M coefficient change amounts, delete the x-th feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0 from both the training set in the (p - 1)-th round and the test set in the (p - 1)-th round, respectively obtain the training set in the p-th round and the test set in the p-th round, and delete x from Dim, obtain the dimension set Dim in the p-th round. After that, let p = p + 1, and perform the deletion process in the (p + 1)-th round based on the training set in the p-th round, the test set in the p-th round, and Dim until the target dimension is obtained. Here, the value of x is a positive integer in the range of 1 to N, and the x-th feature to be deleted is the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0. max Step B: Among the M coefficient change amounts, delete the x-th feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0 from both the training set in the (p - 1)-th round and the test set in the (p - 1)-th round, respectively obtain the training set in the p-th round and the test set in the p-th round, and delete x from Dim, obtain the dimension set Dim in the p-th round. After that, let p = p + 1, and perform the deletion process in the (p + 1)-th round based on the training set in the p-th round, the test set in the p-th round, and Dim until the target dimension is obtained. Here, the value of x is a positive integer in the range of 1 to N, and the x-th feature to be deleted is the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0. max from Dim p-1 to obtain the dimension set Dim in the p-th round. After that, let p = p + 1, and perform the deletion process in the (p + 1)-th round based on the training set in the p-th round, the test set in the p-th round, and Dim until the target dimension is obtained. Here, the value of x is a positive integer in the range of 1 to N, and the x-th feature to be deleted is the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0. x in the p-th round, and then let p = p + 1, perform the deletion process in the (p + 1)-th round based on the training set in the p-th round, the test set in the p-th round, and Dim until the target dimension is obtained. Here, the value of x is a positive integer in the range of 1 to N, and the x-th feature to be deleted is the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0. p in the p-th round until the target dimension is obtained. Here, the value of x is a positive integer in the range of 1 to N, and the x-th feature to be deleted is the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0. max is a positive integer in the range of 1 to N, and the x-th feature to be deleted is the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0. max The x-th feature to be deleted is the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0.

[0061] Here, the x-th coefficient change amount corresponds to the x-th feature to be deleted.

[0062] In the present invention, the above S104 is implemented through the following steps:

[0063] S1041: Input the key features of each point to be measured into the trained first random forest regressor to obtain the first target growth parameter prediction value.

[0064] S1042: Input the key features of each point to be measured into the trained second random forest regressor to obtain the second target growth parameter prediction value.

[0065] S1043: Take the average of the first target growth parameter prediction value and the second target growth parameter prediction value as the target growth parameter of the target plant in the target plant area corresponding to each point to be measured.

[0066] Exemplarily, the target growth parameter of the target plant in the target plant area corresponding to a point to be measured is expressed as: Wherein, represents the trained first random forest regressor, represents the trained second random forest regressor, and x d represents the key features of this point to be measured.

[0067] In the present invention, the method for augmenting training samples based on the semi-supervised learning algorithm includes the following steps:

[0068] S1: Screen out the unlabeled sample points from the polarimetric SAR sample image.

[0069] Here, for each labeled sample point in the polarimetric SAR sample image, in the polarimetric SAR sample image, a preset number of unlabeled sample points are selected from a preset range centered on the labeled sample point. The preset number can be set according to actual needs. For example, it can be 20. Thus, when there are 10 labeled sample points in the polarimetric SAR sample image, 10 * 20 unlabeled sample points can be screened out from the polarimetric SAR sample image.

[0070] S2. Construct a first initial kNN regressor and a second initial kNN regressor with different hyperparameters.

[0071] For example, the first initial kNN regressor is denoted as kNN1(k1, D1, L1), the second initial kNN regressor is denoted as kNN2(k2, D2, L2), and these two kNN regressors perform predictions in the way of nearest neighbor average prediction. k1 and k2 represent different hyperparameters of the number of k nearest neighbors. For example, k1 = 3 and k2 = 5; D1 and D2 represent two different distance metrics. For example, D1 uses the Manhattan distance and D2 uses the Euclidean distance. L1 and L2 represent two sample sets that are initially the same but may be different after expansion. Initially, L1 = L2 = L, where the labeled sample points in the polarimetric SAR sample image form the set L.

[0072] S3. At the e-th iteration, if e is greater than the preset number of iteration rounds E, stop the iteration, and obtain the expanded training samples based on the sample set obtained in the (e - 1)-th round. If e is less than or equal to the preset number of iteration rounds E, randomly sample a subset u from the set U composed of the screened unlabeled sample points to obtain the subset u in the e-th round. e ; e is a positive integer, and the value of e ranges from 1 to E; when e is 1, the sample set obtained in the (e - 1)-th round is the set L composed of the labeled sample points in the polarimetric SAR sample image.

[0073] For example, E = 50. Here, u can contain 200 sample points.

[0074] S4. At the iter-th iteration in the e-th round, if iter is less than the preset number of iterations iters, let the variable j = 1.

[0075] For example, iters = 10.

[0076] S5. Determine whether j is less than or equal to 2.

[0077] S6. If j is greater than 2, let iter = iter + 1, and return to the above S4.

[0078] S7. If j is less than or equal to 2, then based on u e for each labeled sample point x inu the j-th initial kNN regressor updated in the (e - 1)-th round For the i-th initial kNN regressor updated in the (e - 1)-th round the pseudo-label sample set Temp i Update it, let j = j + 1, and then return to S5 above. After executing the iters-th iteration, the pseudo-label sample set and Based on the pseudo-label sample set respectively determine the sample sets of the first and second initial kNN regressors in the e-th round, and use the sample sets in the e-th round to and Update to obtain the first initial kNN regressor updated in the e-th round and the second initial kNN regressor iter is a positive integer, and the value of iter ranges from 1 to iters; i ∈ {1, 2}, i ≠ j.

[0079] Specifically, the sample sets obtained in the (e - 1)-th round include: the sample set of the first initial kNN regressor in the (e - 1)-th round and the sample set of the second initial kNN regressor in the (e - 1)-th round The above S7 is specifically implemented through the following steps:

[0080] S71: Use to determine u from the set L e pool(m) e For each labeled sample point x u in, obtain the set Ω of the k nearest neighbors u , and use and the key features of x u to obtain the pseudo-label of x u k is a positive integer greater than 0. Specifically,

[0081] in this formula, x represents the key features of x u u .

[0082] S72: Use x u and to update to obtain

[0083] Specifically,

[0084] S73: Use and to respectively process the set Ω​u Predict each element in it, and calculate x based on the prediction results u The corresponding predicted error change

[0085] Specifically, y i is the label of x i In this formula, x i represents the key features of x i The key features

[0086] S74. When there is a value greater than 0 in u e The corresponding If there is a value greater than 0 in When there is a value greater than 0, take the largest value greater than 0 The corresponding x u And the largest value greater than 0 The corresponding x u The pseudo-label As a pseudo-label sample, and add the pseudo-label sample to The pseudo-label sample set Temp i ; When there is no value greater than 0 in u e The corresponding If there is no value greater than 0 in When there is no value greater than 0, do not update Temp i .

[0087] S75. Let j = j + 1 and then return to the above S5. After executing the iters-th iteration, obtain the pseudo-label sample set of the first initial kNN regressor And the pseudo-label sample set of the second initial kNN regressor

[0088] S76. Add To To obtain the sample set of the e-th round of the first initial kNN regressor Add To To obtain the sample set of the e-th round of the second initial kNN regressor

[0089] S77. Use To update To obtain And use To update To obtain

[0090] Specifically, use The key features, and the labels corresponding to the key features, to To update, and obtain Similarly, using the key features, and the tags corresponding to the key features, for perform an update to obtain

[0091] S8. Train and validate a random forest regressor based on the sample set in the e-th round, and when the validation result meets the preset conditions, obtain an augmented training sample based on the sample set in the e-th round; otherwise, set e = e + 1 and then return to the above S3.

[0092] Specifically, the sample set in the e-th round includes: the sample set of the first initial kNN regressor in the e-th round and the sample set of the second initial kNN regressor in the e-th round The above S8 is specifically implemented through the following steps:

[0093] S81. Respectively use and to train the initial random forest regressor respectively, and obtain the first trained random forest regressor in the e-th round and the second trained random forest regressor in the e-th round

[0094] Specifically, use the key features and the tags corresponding to the key features to train the initial random forest regressor, and obtain the first trained random forest regressor in the e-th round and, use the key features and the tags corresponding to the key features to train the initial random forest regressor, and obtain the second trained random forest regressor in the e-th round

[0095] S82. Respectively use and to jointly predict the validation set, and calculate and the common R 2 coefficient of determination R 2 val,e ; the validation set is the set composed of the key features and tags of some labeled sample points in set L.

[0096] Here, the validation set can be the set composed of the key features and tags of 10% of the sample points in set L. Specifically, x val represents the key features of the labeled sample points in the validation set, and y val represents the tag corresponding to x val (i.e., the true tag), represents xval The corresponding predicted label.

[0097] S83. When R 2 val,e is less than or equal to the R obtained in the previous e rounds 2 the maximum R in the coefficient of determination 2 the coefficient of determination R 2 best , let count = count + 1; when R 2 val,e is greater than R 2 best , update R 2 best to R 2 val,e , and let count = 0; where the initial value of count is 0.

[0098] Here, the initial value of R 2 best is -1, and when e is 1, the corresponding R 2 best is -1. count represents the cumulative number of rounds in which R val on the validation set x 2 has not improved (the initial setting of count is 0).

[0099] S84. Judge whether count is greater than or equal to H; H is a preset positive integer. When count is greater than or equal to H, use the key features and labels of and

[0100] the key features and labels of

[0101] as the augmented training samples; when count is less than H, let e = e + 1 and then return to the above S3.

[0102] (5) Set hyperparameters and create two kNN (k-nearest neighbor) regressors: kNN1(k1, D1, L1) and kNN2(k2, D2, L2) represent two kNN regressors with different hyperparameters; and set the following parameters:

[0103] epochs represents the total number of rounds of semi-supervised learning, epoch represents the current training round. In the present invention, the initial setting is epoch = 0, epochs = 50;

[0104] iters represents the number of iterations required for each update of kNN within an epoch, iter represents the current iteration number. In the present invention, iter is initially set to 0 and iters is set to 10;

[0105] m represents the number of sample points randomly sampled from U by the random sampling pool pool each time. In the present invention, m = 200;

[0106] u represents the set composed of m samples randomly sampled from the set U by the random sampling pool pool each time;

[0107] Temp i (i ∈ {1, 2}) is used to temporarily store the pseudo-label samples predicted by another kNN);

[0108] k1 and k2 represent different hyperparameters of the number of k-nearest neighbors, k1 = 3 and k2 = 5;

[0109] D1 and D2 represent two different distance metric methods. D1 uses the Manhattan distance and D2 uses the Euclidean distance;

[0110] L1 and L2 represent two labeled sample sets that are initially the same but may be different after expansion. Initially, L1 = L2 = L;

[0111] x val represents the validation set that accounts for 10% of the training set size divided from the training set, and y val is the label of the validation set;

[0112] represents the highest R val calculated on x 2 value ( initially set to -1);

[0113] count represents the cumulative number of rounds in which R val on the validation set x 2 does not improve (count is initially set to 0);

[0114] represents the labeled sample set finally expanded by semi-supervised learning. Initially, let represents the labeled sample set expanded in the current round of semi-supervised learning;

[0115] (6a) Determine whether the current epoch is less than epochs: If so, sequentially execute step (6b); Otherwise, jump to step (6j);

[0116] (6b) Sample a random subset u from the set U and set the variable iter = 0;

[0117] (6c) Determine whether iter is less than iters. If so, set the variable j = 1 and sequentially execute step (6d); if not, jump to step (6h);

[0118] (6d) Determine whether j ≤ 2: If so, let h j represent the j-th kNN regressor: h j ← kNN j (k j , D j , L j )(j ∈ {1, 2}); if not, set iter = iter + 1 and jump to step (6c);

[0119] (6e) Loop through each element x in u u , and for each element x u sequentially execute the following steps:

[0120] 1) Calculate the set Ω of the k nearest neighbors of x u in the set L: Ω u ← Neighbors(x u , L u , k j , D j , D j );

[0121] 2) Use h j to predict the pseudo-label of x u ;

[0122] 3) Create

[0123] 4) Respectively use h j and h j ' to predict each element x u in Ω i and calculate

[0124] (6f) Determine whether there exists an element x in u such that : u :

[0125] 1) If there is no element such that ;

[0126] 2) If there exists 's element x u : where

[0127] (6g) Let j = j + 1 and jump to step 6(d).

[0128] (6h) Update the labeled samples {L1, L2} and the kNN regressors {h1, h2}:

[0129] L1 ← L1 ∪ Temp1; L2 ← L2 ∪ Temp2;

[0130] h1 ← kNN1(k1, D1, L1); h2 ← kNN2(k2, D2, L2);

[0131] (6i) Limit the number of rounds of semi - supervised learning using early stopping and save the best result:

[0132] Specifically, the following steps are sequentially executed after step (6h):

[0133] 1) Respectively use (i.e., L1 obtained in step (6h)) and (i.e., L2 obtained in step (6h)) to train the random forest regressors RFR1 and RFR2;

[0134] 2) Use RFR1 and RFR2 to predict on the validation set x val :

[0135] 3) Calculate R 2 on the validation set:

[0136] 4) Judge whether R 2 val ≤ R 2 best :

[0137] If R 2 val ≤ R 2 best : Let count = count + 1;

[0138] If R 2 val > R 2 best : Let R 2 best = R 2 val ,

[0139] 5) Judge whether count >= 10: If so, jump to step (6j); if not, let epoch = epoch + 1 and jump to step (6a);

[0140] (6j) Take As the finally expanded set of labeled training samples, the execution of the improved COREG semi-supervised learning algorithm is terminated.

[0141] In the case where the number of labeled samples involved in the present invention is small, the correlation between features and growth parameters is low, and the data source is single, directly using the COREG semi-supervised learning algorithm will not only not improve the inversion result, but will instead make the accuracy of the inversion result lower. This is because the original COREG semi-supervised learning algorithm is prone to introducing pseudo-samples full of noise in the case of few samples and high noise, causing the distribution of the expanded labeled samples to deviate from the original distribution. Based on the above problems, the present invention has made certain improvements to the strategy of the COREG semi-supervised learning algorithm, making its noise tolerance significantly improved in this case. Compared with the original COREG semi-supervised learning algorithm, the improvement of the COREG semi-supervised learning algorithm in the present invention lies in: in the original COREG semi-supervised learning algorithm, after kNN1(k1,D1,L1) and kNN2(k2,D2,L2) predict credible pseudo-labels for each other in each round of iteration, the two kNNs are immediately updated. However, in the case of high noise and low correlation between features and growth parameters, it is very easy for the original data distribution in the kNN to deviate. Therefore, in the present invention, the pseudo-labels predicted by kNN1(k1,d1,L1) and kNN2(k2,d2,L2) for each other are first temporarily stored in Temp i≠j (i ∈ {1, 2}), and the kNN is updated when the appropriate number (for example, 10) is temporarily stored. The reason for doing this is that after predicting multiple pseudo-labels, the prediction offsets in different directions in the pseudo-labels can cancel each other out to a certain extent, so that the data distribution in the updated kNN still roughly conforms to the original distribution.

[0142] In addition, some studies use a distance-weighted kNN regressor in the implementation of the COREG semi-supervised learning algorithm. However, in the usage scenario of the present invention, due to the low correlation between features and prediction parameters and the high noise of the original polarimetric SAR data, if the distance-weighted kNN algorithm is used, it is easier to introduce noise, resulting in the expanded samples quickly deviating from the original distribution in the early stage of semi-supervised learning and reducing the prediction performance. Based on the above considerations, the present invention adopts an average k-nearest neighbor kNN prediction strategy, averaging the values of the k nearest neighbor parameters as the pseudo-label. It has been found through experiments that the improved method can significantly improve the noise tolerance in the case of low correlation.

[0143] In the case where the polarimetric SAR image is used as a single data source, the present invention obtains multi-dimensional features through polarimetric target decomposition, fully excavates the information contained in the polarimetric SAR image data, alleviates the difficulties caused by the single data source, and moreover, has less dependence on data multi-source. When the features obtained by polarimetric target decomposition have a low correlation with the growth parameters and the multi-dimensional features obtained by polarimetric target decomposition are highly redundant, the present invention selects key features through a recursive feature elimination method based on cross-validation, weakens the influence of mutual interference between multi-dimensional features, and effectively improves the prediction accuracy of plant growth parameters. When the number of labeled growth parameter samples is small, the present invention uses training samples augmented by a semi-supervised learning algorithm to train the prediction model, so that the training effect of the prediction model can be better, and the predicted plant growth parameters are more accurate and have higher robustness.

[0144] The following further illustrates the effect of the present invention in combination with simulation experiments.

[0145] 1. Simulation experiment conditions:

[0146] The simulation of the present invention is carried out in a hardware environment with an Intel(R) Core(TM) i5-13500H CPU with a main frequency of 2.60 GHZ and 16 GB of memory, and a software environment of python (3.10).

[0147] 2. Simulation content:

[0148] The simulation experiment for retrieving winter wheat growth parameters from polarimetric SAR images based on random forest regression and semi-supervised learning of the present invention is carried out on polarimetric SAR images of the same area in different periods in a certain area by using the method proposed by the present invention. The three periods are April 16th, May 10th, and June 30th respectively. The size of the image is 3644×6213, and the resolution is 8 m. Exemplarily, Figure 2 is a schematic diagram of the polarimetric SAR image collected in this certain area in the simulation experiment of the present invention.

[0149] The labeled samples used in the present invention are 7 wheat quadrats extracted from the research area for measuring the growth parameters of crops. The specific area of the quadrat depends on the natural boundary of the field plot. The above quadrats are large quadrats, and 5 small quadrats are set inside the large quadrat for measuring growth parameters. Each small quadrat is measured three times, and the average is taken as the growth parameter value of the small quadrat. The distribution map of all 35 winter wheat small quadrats is as shown in Figure 3 shown, where the white small squares represent the small quadrats used for measuring growth parameters. A total of 105 labeled sample data are obtained for all 35 winter wheat small quadrats in three periods, and the labeled growth parameters include plant height, fresh weight, and plant water content.

[0150] For the polarization target decomposition features and measurement results of growth parameters of the 105 acquired sample points, the method proposed in the present invention is used, and the experimental results are evaluated through 10-fold cross-validation 10 times.

[0151] 3. Analysis of simulation results:

[0152] Figure 4 It is the result diagram of 10-fold cross-validation of the fresh weight parameter of winter wheat in the simulation experiment; Figure 5 It is the result diagram of 10-fold cross-validation of the plant height parameter of winter wheat in the simulation experiment; Figure 6 It is the result diagram of 10-fold cross-validation of the water content parameter of winter wheat in the simulation experiment.

[0153] As shown in Table 1 is the result comparison of using RFR and the method proposed in the present invention for the fresh weight parameter; as shown in Table 2 is the result comparison of using RFR and the method proposed in the present invention for the plant height parameter; as shown in Table 3 is the result comparison of using RFR and the method proposed in the present invention for the water content parameter.

[0154] Table 1 Comparison table of the results of using RFR regression and the method proposed in the present invention for the fresh weight parameter

[0155]

[0156]

[0157] It can be seen from Table 1 that the RMSE of the inversion result of the fresh weight parameter of winter wheat by the Cross-Verified Recursive Feature Elimination (CVRFE) method used in the present invention is reduced by 0.023 g, and at the same time its R 2 coefficient is increased by 0.051; the improved COREG semi-supervised learning algorithm further reduces the RMSE of the inversion result of the fresh weight parameter of winter wheat by 0.03 g on the basis of CVRFE, and at the same time its R 2 coefficient is increased by 0.02.

[0158] Table 2 Comparison table of the results of using RFR regression and the method proposed in this paper for the plant height parameter

[0159]

[0160] It can be seen from Table 2 that the RMSE of the inversion result of the plant height parameter of winter wheat by the CVRFE method used in the present invention is reduced by 0.574 cm, and at the same time its R 2 coefficient is increased by 0.058; the improved COREG semi-supervised learning algorithm further reduces the RMSE of the inversion result of the plant height parameter of winter wheat by 0.038 cm on the basis of CVRFE, and at the same time its R2 The coefficient has increased by 0.05.

[0161] Table 3 Comparison table of the results of using RFR regression and the method proposed in this paper for the water content parameter

[0162]

[0163] As can be seen from Table 3, the RMSE of the inversion result of the winter wheat water content parameter using the CVRFE method in the present invention has decreased by 0.28%, and at the same time its R 2 coefficient has increased by 0.089; the improved COREG semi-supervised learning algorithm further reduces the RMSE of the inversion result of the winter wheat water content parameter by 0.02% on the basis of CVRFE, and at the same time its R 2 coefficient has increased by 0.03.

[0164] It should be noted that the terms "first" and "second" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined.

[0165] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0166] In the specification, the word "including" does not exclude other components or steps, and "a" or "one" does not exclude the case of a plurality. Certain measures are described in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0167] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression, characterized in that, Including: Obtaining the original data of a polarimetric SAR image including at least one point to be measured; Each point to be measured represents a target plant area; Performing polarimetric target decomposition on the original data to obtain multi-dimensional features of each point to be measured; Selecting features of a target dimension from the multi-dimensional features of each point to be measured to obtain key features of each point to be measured; wherein, the target dimension is obtained by performing feature selection processing on the multi-dimensional features of labeled sample points in a polarimetric SAR sample image using a recursive feature elimination method with cross-validation; the label of each sample point is the target growth parameter of the target plant in the target plant area represented by the sample point; Using a trained random forest regressor to predict the target growth parameter according to the key features of each point to be measured to obtain the target growth parameter of the target plant; the trained random forest regressor is trained using training samples augmented by a semi-supervised learning algorithm; Among them, the method for augmenting training samples based on a semi-supervised learning algorithm includes: S1. Screening unlabeled sample points from the polarimetric SAR sample image; S2. Constructing a first initial kNN regressor and a second initial kNN regressor with different hyperparameters; S3. At the e-th iteration, if e is greater than the preset number of iterations E, stop the iteration and obtain the augmented training samples based on the sample set obtained in the (e - 1)-th iteration. If e is less than or equal to the preset number of iterations E, randomly sample a subset u from the set U composed of the filtered unlabeled sample points to obtain the subset u in the e-th iteration e ; e is a positive integer, and the value range of e is from 1 to E; when e is 1, the sample set obtained in the (e - 1)-th iteration is the set L composed of the labeled sample points in the polarimetric SAR sample image S4. In the iter-th iteration of the e-th round, if iter is less than the preset number of iterations iters, then set the variable j = 1; S5. Judging whether j is less than or equal to 2; S6. If j is greater than 2, then set iter = iter + 1 and return to the above S4; S7. If j is less than or equal to 2, then based on the u e each labeled sample point x u in, and the j-th initial kNN regressor updated in the (e - 1)-th round for the i-th initial kNN regressor updated in the (e - 1)-th round the pseudo-labeled sample set Temp i is updated, and after setting j = j + 1, return to the above S5. After the iters-th iteration is completed, the pseudo-labeled sample set and Based on the pseudo-labeled sample set respectively determine the sample sets of the e-th round of the first initial kNN regressor and the second initial kNN regressor, and use the sample sets of the e-th round to update the and the to obtain the first initial kNN regressor updated in the e-th round and the second initial kNN regressor iter is a positive integer, and the value of iter ranges from 1 to iters; i ∈ {1, 2}, i ≠ j; S8. Training and validating a random forest regressor according to the sample set of the e-th round, and when the validation result reaches a preset condition, obtaining augmented training samples based on the sample set of the e-th round, otherwise, set e = e + 1 and return to the above S3.

2. The plant growth parameter inversion method based on polarimetric SAR images and semi-supervised regression according to claim 1, wherein The performing polarimetric target decomposition on the original data to obtain multi-dimensional features of each point to be measured includes: Performing filtering processing on the original data to obtain a polarimetric covariance matrix C; Extracting diagonal elements of the polarimetric covariance matrix C to obtain the backscattering coefficients of each point in the polarimetric SAR image; Performing Freeman-Durden decomposition, Cloude decomposition, non-negative eigenvalue decomposition, TSVM decomposition, and Yamaguchi decomposition on the polarimetric covariance matrix C respectively to obtain different features of different dimensions of each point; Taking the feature set composed of the backscattering coefficients and different features of different dimensions of each point to be measured as the multi-dimensional features of each point to be measured.

3. The method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression according to claim 1, wherein Before selecting features of a target dimension from the multi-dimensional features of each point to be measured to obtain key features of each point to be measured, the method further includes: Obtaining the original sample data of the polarimetric SAR sample image and performing polarimetric target decomposition on the original sample data to obtain multi-dimensional features of each labeled sample point; Using the 10-fold cross-validation method, a set F composed of the multi-dimensional features of multiple labeled sample points and the labels of the multiple labeled sample points is divided into 10 initial training sets and 10 initial test sets corresponding one by one to the 10 initial training sets, and a dimension set Dim recording the dimension numbers of all features in the multi-dimensional features is created; Before the p-th round of deletion processing on the initial training set, the initial test set, and Dim, obtain the training set of the (p - 1)-th round, the test set of the (p - 1)-th round, and the dimension set Dim of the (p - 1)-th round p-1 ; each sample point in the training set of the (p - 1)-th round and the test set of the (p - 1)-th round has M-dimensional features; where p is an integer greater than or equal to 1, when p is 1, the training set of the (p - 1)-th round is the 10 initial training sets, the test set of the (p - 1)-th round is the 10 initial test sets, the dimension set of the (p - 1)-th round is Dim, and M = N, where N represents the total dimension of each multi-dimensional feature; When M is equal to 1, use the dimension serial number in the said Dim p-1 as the target dimension; when M is greater than 1, perform the p-th round of deletion processing, and in the x-th loop of the p-th round of deletion processing, use the x-th dimension feature in each of the M-dimensional features as the feature to be deleted for the x-th time. Based on the initial random forest regressor, the training set of the (p - 1)-th round, and the test set of the (p - 1)-th round, determine the x-th coefficient change amount, and obtain M such coefficient change amounts in the p-th round of deletion processing; x is a positive integer greater than or equal to 1, and the value of x ranges from 1 to M; When the M coefficient change amounts satisfy a preset coefficient change condition, use the dimension serial number in the Dim p-1 as the target dimension; when the M coefficient change amounts do not satisfy the preset coefficient change condition, perform feature deletion processing on the training set of the (p - 1)-th round, the test set of the (p - 1)-th round, and the Dim p-1 to obtain the training set of the p-th round, the test set of the p-th round, and the dimension set Dim p-1 of the p-th round, and after setting p = p + 1, perform the deletion processing of the (p + 1)-th round until the target dimension is obtained.

4. The method for inverting plant growth parameters based on polarimetric SAR images and semi-supervised regression according to claim 3, wherein Determining the x-th coefficient change amount based on the initial random forest regressor, the training set of the (p - 1)-th round, and the test set of the (p - 1)-th round includes: Respectively deleting the x-th feature to be deleted in the training set of the (p - 1)-th round and the test set of the (p - 1)-th round, to obtain the training set of the x-th time in the (p - 1)-th round and the test set of the x-th time in the (p - 1)-th round; Respectively using the training set of the (p - 1)-th round and the training set of the x-th time in the (p - 1)-th round to train the initial random forest regressor, to obtain the first random forest regressor of the x-th time and the second random forest regressor of the x-th time; Use the first random forest regressor of the x-th time to predict the test set of the (p - 1)-th round, and calculate R according to the prediction result 2 Coefficient of determination, and obtain the first R of the x-th time 2 Coefficient; Use the second random forest regressor of the x-th time to predict the test set of the x-th time in the (p - 1)-th round, and calculate R according to the prediction result 2 The coefficient of determination to obtain the second R of the x-th time 2 Coefficient; Take the difference between the second R coefficient of the x-th time and the first R coefficient of the x-th time as the coefficient change amount of the x-th one. 2 2 ​​ 5. The method for inverting plant growth parameters based on polarimetric SAR images and semi-supervised regression according to claim 3, characterized in that When the M coefficient change amounts do not meet the preset coefficient change condition, for the training set of the (p - 1)-th round, the test set of the (p - 1)-th round, and Dim p-1 perform feature deletion processing to obtain the training set of the p-th round, the test set of the p-th round, and the dimension set Dim p-1 of the p-th round, and after setting p = p + 1, perform the deletion processing of the (p + 1)-th round until the target dimension is obtained, including: When there is a coefficient change amount greater than or equal to 0 among the M coefficient change amounts, it indicates that the M coefficient change amounts do not meet the preset coefficient change condition; Among the M coefficient change amounts, delete the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0 in the x max -th time from both the training set of the (p - 1)-th round and the test set of the (p - 1)-th round, respectively obtain the training set of the p-th round and the test set of the p-th round, and delete x max from Dim p-1 to obtain the dimension set Dim p of the p-th round. After that, let p = p + 1, and perform the deletion process of the (p + 1)-th round based on the training set of the p-th round, the test set of the p-th round, and the Dim p until the target dimension is obtained; the feature to be deleted corresponding to the x max -th time is the feature to be deleted corresponding to the largest coefficient change amount that is greater than or equal to 0.

6. The method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression according to claim 1, wherein The trained random forest regressor includes: the trained first random forest regressor and the trained second random forest regressor; Using the trained random forest regressor to predict the target growth parameter according to the key features of each point to be measured, to obtain the target growth parameter of the target plant, includes: Inputting the key features of each point to be measured into the trained first random forest regressor to obtain the first target growth parameter prediction value; Inputting the key features of each point to be measured into the trained second random forest regressor to obtain the second target growth parameter prediction value; Taking the average value of the first target growth parameter prediction value and the second target growth parameter prediction value as the target growth parameter of the target plant in the target plant area corresponding to each point to be measured.

7. The method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression according to claim 1, characterized in that, The sample set obtained in the (e - 1)-th round includes: the sample set of the (e - 1)-th round of the first initial kNN regressor and the sample set of the (e - 1)-th round of the second initial kNN regressor The above S7 includes: S71. Determine the set Ω of the k nearest neighbors of each labeled sample point x in the set L using the following steps: Determine the u from the set L e among each labeled sample point x u and use the key features of the x u to obtain the pseudo-label of the x and the x u where k is a positive integer greater than 0; u pseudo-label k is a positive integer greater than 0; S72. Adopt the said x u and update the said to obtain S73. Using the said and the said to predict each element in the set Ω u respectively, and calculating the change amount of the prediction error corresponding to the said x u based on the prediction results S74. When the e corresponding has a value greater than 0 in it, take the largest value greater than 0 u and the corresponding x and the corresponding x u pseudo-label as a pseudo-label sample, and add the pseudo-label sample to the pseudo-label sample set Temp i ; When the u e corresponding has no value greater than 0 in it, do not update the Temp i ; S75. After setting \(j = j + 1\), return to the above S5 until after the iters-th iteration is completed, obtaining the pseudo-label sample set of the first initial kNN regressor and the pseudo-label sample set of the second initial kNN regressor S76. Add the to the to obtain the sample set of the e-th round of the first initial kNN regressor Add the to the to obtain the sample set of the e-th round of the second initial kNN regressor S77. Adopt the Update the Obtain the And adopt the Update the Obtain the 8. The method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression according to claim 1, characterized in that The sample set of the e-th round includes: the sample set of the e-th round of the first initial kNN regressor and the sample set of the e-th round of the second initial kNN regressor The above S8 includes: S81. Respectively use the and the to perform respective training on the initial random forest regressor, and obtain the first random forest regressor trained in the e-th round and the second random forest regressor S82. Use the and the to jointly predict the validation set, and calculate the and the common R 2 coefficient of determination The validation set is a set composed of the key features and labels of some labeled sample points in the set L; S83. When the R 2 val,e is less than or equal to the maximum R in the coefficient of determination obtained in the previous e rounds 2 in the coefficient of determination 2 of the coefficient of determination R 2 best , let count = count + 1; when the R 2 val,e is greater than the R 2 best , update the R 2 best to the R 2 val,e , and let count = 0; where the initial value of count is 0; S84. Determine whether count is greater than or equal to H; H is a preset positive integer. When count is greater than or equal to H, use the key features and labels of the and the key features and labels of the as the augmented training samples; when count is less than H, increment e by 1 and then return to the above S3.

9. The method for retrieving plant growth parameters based on polarimetric SAR images and semi-supervised regression according to claim 1, wherein The target plant includes wheat, and the target growth parameter includes any one of wheat fresh weight, water content, and plant height.

Images