Hyperspectral Hybrid Maize Planting Area Extraction Method Based on the Fusion of Spectral and Spatial Features

Abstract

The present invention relates to a method for extracting high - spectral hybrid - seed maize planting areas based on the fusion of spectral and spatial features, including: obtaining data of the study area for pre - processing; performing masking to obtain the masked Sentinel - 2 data, and extracting maize planting distribution data to mask the pre - processed Zhuhai - 1 OHS hyperspectral data; constructing spectral features using the class - mean matrix clustering method; constructing spatial features based on the spectral features using the multi - scale morphology method; finding the optimal parameters of the classifier to obtain the hybrid - seed maize extraction model; respectively performing the extraction of hybrid - seed maize and evaluating it using the overall accuracy. The present invention uses the first - order statistic class - mean matrix and k - means clustering to extract spectral features, which has simple operations and increases the inter - class difference between hybrid - seed maize and field maize; the multi - scale morphology spatial features adopted by the present invention improve the "salt - and - pepper noise" phenomenon, retain a relatively complete plot structure, and improve the extraction accuracy of hybrid - seed maize.

Description

Technical Field

[0001] The present invention relates to the technical field of crop remote sensing identification, in particular to a method for extracting the planting area of high - spectral hybrid - seed maize based on the fusion of spectral and spatial features. Background Technique

[0002] At present, the maize planted in China is mainly hybrid varieties. The hybrid maize is called field maize, while the hybrid maize variety is obtained by crossing excellent inbred - line maize, and the inbred - line maize is called hybrid - seed maize. Currently, illegal seed production phenomena such as "private propagation and random production" are serious. To ensure the quality of agricultural seed supply, strengthen the supervision of seed production, and timely and accurately monitor the planting area and spatial distribution of hybrid - seed maize have become urgent problems to be solved. The traditional method for obtaining hybrid - seed information is through statistical surveys, which are usually limited by problems such as low efficiency, high cost, and large human influence, and it is difficult to meet the needs of modern seed industry development. Remote sensing technology has the characteristics of objective and accurate data, real - time acquisition, macroscopic dynamics, and low cost, and has great potential in monitoring the distribution of hybrid - seed maize.

[0003] The existing research on extracting the planting area of hybrid - seed maize based on remote sensing mainly focuses on the excavation of the texture features of high - resolution images. This is because the planting method and cultivation habit of hybrid - seed maize are different from those of field maize, forming unique texture structure features at the canopy scale. Although this method has achieved good results, it is difficult to synchronously cover a large monitoring area within a short critical discrimination window period, and the cost is relatively high. It is found that hybrid - seed maize is an inbred line, and its plants are shorter than those of field hybrid - line maize. This trait difference is correspondingly manifested as a spectral difference. Due to the similarity of species, the extraction of hybrid - seed maize planting is still affected by the low inter - class difference and high intra - class variance of field maize. Therefore, it is challenging to obtain the spectral feature differences of hybrid - seed maize.

[0004] The rich spectral information in hyperspectral remote sensing images provides a feasible method for the extraction of hybrid maize distribution. At the same time, hyperspectral image classification also faces the Hughes phenomenon and the curse of dimensionality. How to effectively extract classification features has become the key. Among feature extraction methods, supervised and unsupervised methods are more widely used. In unsupervised learning, methods represented by principal component analysis directly process hyperspectral data without considering the separability of actual categories. In supervised learning, popular methods such as linear discriminant analysis have problems relying on the prior knowledge provided by labeled samples. The application of a feature extraction method that combines the advantages of supervised and unsupervised methods in the field of remote sensing needs to be further explored. In addition, the spatial texture information contained in the three-dimensional data of hyperspectral images also has an auxiliary effect on solving the problems of "same object with different spectra" or "different objects with the same spectra" encountered in classification. For hyperspectral remote sensing images with medium and low spatial resolution, the expression of local texture features is not clear, while the expression of global structure information is more important. Among them, morphological methods are powerful tools for representing spectral spatial information and play an active role in the spatial classification of remote sensing images. Therefore, to solve the problem of difficult distinction between hybrid maize and field maize, constructing effective spectral and spatial features has become the key to the extraction of hybrid maize distribution. Summary of the Invention

[0005] To solve the problems of low efficiency, high cost and great influence of manual operation in traditional extraction methods, the purpose of the present invention is to provide a method for extracting the planting area of hybrid maize based on the fusion of spectral and spatial features with low cost and improved extraction accuracy of hybrid maize.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions: A method for extracting the planting area of hybrid maize based on the fusion of spectral and spatial features, the method includes the following steps in sequence:

[0007] (1) Obtain the Zhuhai-1 OHS hyperspectral data, Sentinel-2 multispectral data and the global land cover product of the European Space Agency in the study area, and preprocess the obtained data;

[0008] (2) Use the cultivated land distribution data in the global land cover product of the European Space Agency to mask the preprocessed Sentinel-2 data to obtain the masked Sentinel-2 data. Use the enhanced vegetation index threshold method to extract the maize planting distribution data from the masked Sentinel-2 data, and use the extracted maize planting distribution data to mask the preprocessed Zhuhai-1 OHS hyperspectral data;

[0009] (3) Construct spectral features using the class mean matrix clustering method;

[0010] (4) Construct spatial features based on spectral features using the multi-scale morphological method;

[0011] (5) Standardize the constructed spectral features and spatial features and input them into a support vector machine classifier to find the optimal parameters of the classifier, and obtain a seed production maize extraction model;

[0012] (6) Input the images of the seedling stage, jointing stage, and milk stage into the seed production maize extraction model, perform seed production maize extraction respectively, and use the overall accuracy for evaluation.

[0013] The specific steps of step (1) are as follows:

[0014] (1a) Obtain the Zhuhai-1 OHS hyperspectral data and perform preprocessing. The preprocessing includes radiometric calibration, atmospheric correction, and orthorectification processing. The radiometric calibration, atmospheric correction, and orthorectification are implemented based on ENVI software, and then crop it using the vector administrative boundary data to obtain the preprocessed Zhuhai-1 OHS hyperspectral data;

[0015] (1b) Obtain the Sentinel-2 multispectral data and perform preprocessing. Select Sentinel-2 images with a cloud cover of less than 5% on the engine platform, perform cloud removal, linear spatial interpolation, and calculate the EVI index, i.e., the enhanced vegetation index, to obtain the preprocessed Sentinel-2 data;

[0016] (1c) Obtain the global land cover product of the European Space Agency, extract the cultivated land distribution data, and the cultivated land distribution data type number is 40.

[0017] The specific steps of step (2) are as follows:

[0018] (2a) Use the cultivated land distribution data provided by the global land cover product of the European Space Agency to perform masking processing on the preprocessed Sentinel-2 data to obtain the masked Sentinel-2 data;

[0019] (2b) Screen the key time windows for the masked Sentinel-2 data, draw the sample mean curve of the enhanced vegetation index EVI, and screen the key time windows that can identify maize according to the place with the largest mean difference;

[0020] (2c) Draw a box plot of the sample enhanced vegetation index EVI according to the screened key time windows, determine the threshold conditions for maize extraction, and then perform threshold extraction on the masked Sentinel-2 data under the key time windows to obtain the maize planting distribution data;

[0021] (2d) Mask the preprocessed Zhuhai-1 OHS hyperspectral data with the maize planting distribution data.

[0022] The specific steps of step (3) are as follows:

[0023] (3a) Calculate the first-order statistics of the sample hyperspectral features from the Zhuhai-1 OHS hyperspectral data after masking to obtain the class mean matrix A:

[0024]

[0025] where, m ij is the average value of the training samples of the j-th class in the i-th feature, i = 1, 2, …, d, j = 1, 2, …, c, d represents the feature dimension, and c represents the number of classes; a i is the row vector of the i-th feature of the class mean matrix A;

[0026] (3b) Cluster the feature row vectors a i of the class mean matrix A using the k-means clustering algorithm, and cluster the approximate vectors into one cluster. The k-means clustering cluster G k is:

[0027] G k = {a1, …, a i} (2)

[0028] where, G1 represents the first clustering cluster, which is composed of the feature row vector a i ; G k represents the k-th clustering cluster, n represents the number of clustering clusters, and k ∈ 1, 2, …, n;

[0029] (3c) Finally, obtain the spectral features by calculating the mean value of each clustering cluster G k . The number of new spectral features is equal to the number of clustering clusters. The formula for extracting the new spectral features Z is as follows:

[0030]

[0031] where, z1 represents the first spectral feature, which is obtained by calculating the mean value of the clustering cluster G1; z k represents the k-th spectral feature, which is obtained by calculating the mean value of the clustering cluster G k ; k represents the number of clustering clusters.

[0032] The specific content of step (4) is: Perform opening and closing operations with structural element scales of 2, 4, and 6 on the constructed spectral features to obtain multi-scale morphological space features. The calculation formulas for morphological opening and closing operations are shown in formulas (4) and (5):

[0033]

[0034]

[0035] where, MP γ represents the image after the opening operation; MPΦ Denotes the image after closing operation; γ s (I and Φ s (I) respectively denote the opening operation and the closing operation performed on the image I; s is the size of the structuring element. When s = 0, it means that no operation is performed on the original image I.

[0036] The said step (5) specifically includes the following steps:

[0037] (5) Normalize the constructed spectral features and spatial features and input them into the support vector machine classifier to find the optimal parameters of the classifier, and obtain the extraction model of hybrid maize seeds.

[0038] (5a) Perform Z-Score normalization on the spectral features and spatial features, and calculate the Fisher discriminant ratio to measure the separability of the features. The Fisher discriminant ratio formula:

[0039]

[0040]

[0041] where, FR i represents the Fisher discriminant ratio of class i under the given features, μ i is the mean of the features of class i, is the variance of the features of class i, i * is the class with the smallest distance from the mean of class i to the means of all other classes;

[0042] (5b) Input the normalized features into the support vector machine classifier, and use the grid search method combined with the 5-fold cross-validation method to find the optimal parameters of the classifier, and obtain the extraction model of hybrid maize seeds.

[0043] As can be seen from the above technical solutions, the beneficial effects of the present invention are as follows: First, the present invention adopts a hierarchical and step-by-step extraction strategy to realize the extraction of the hybrid maize seed planting area, which can significantly reduce the interference of non-maize ground object types on the extraction result of hybrid maize seeds; Second, the present invention uses the first-order statistic class mean matrix and k-means clustering to extract spectral features, which is simple in operation and increases the inter-class difference between hybrid maize seeds and field maize; Third, the multi-scale morphological spatial features adopted by the present invention improve the "salt and pepper noise" phenomenon, retain a relatively complete plot structure, and improve the extraction accuracy of hybrid maize seeds; Fourth, the scheme of the present invention for extracting features based on class mean matrix clustering and multi-scale morphology can more accurately realize the extraction of the distribution of hybrid maize seeds compared with the three schemes of only using PCA-LDA, only using class mean matrix clustering, and based on PCA-LDA and multi-scale morphology. At the same time, the application of the "Zhuhai-1" OHS hyperspectral data in the classification and extraction of hybrid maize seeds is also explored in the present invention. Description of the Drawings

[0044] Figure 1 is the method flowchart of the present invention;

[0045] Figure 2 is the EVI box plot under the key time window;

[0046] Figure 3 is the schematic diagram of multi-scale morphological opening and closing operations;

[0047] Figure 4 is the Fisher feature discrimination ratio diagram. Specific implementation manners

[0048] As Figure 1 shown, a method for extracting the planting area of hybrid seed production maize based on the fusion of spectral and spatial features, the method includes the following steps in sequence:

[0049] (1) Obtain the Zhuhai-1 OHS hyperspectral data, Sentinel-2 multispectral data of the study area, and the global land cover product of the European Space Agency, and preprocess the obtained data;

[0050] (2) Mask the preprocessed Sentinel-2 data with the cultivated land distribution data in the global land cover product of the European Space Agency to obtain the masked Sentinel-2 data, use the enhanced vegetation index threshold method to extract the maize planting distribution data from the masked Sentinel-2 data, and mask the preprocessed Zhuhai-1 OHS hyperspectral data with the extracted maize planting distribution data;

[0051] (3) Construct spectral features using the class mean matrix clustering method;

[0052] (4) Construct spatial features based on the spectral features using the multi-scale morphological method;

[0053] (5) Normalize the constructed spectral features and spatial features and input them into the support vector machine classifier to find the optimal parameters of the classifier, and obtain the hybrid seed production maize extraction model;

[0054] (6) Input the images of the seedling stage, jointing stage, and milk ripening stage into the hybrid seed production maize extraction model, perform hybrid seed production maize extraction respectively, and use the overall accuracy for evaluation.

[0055] The step (1) specifically includes the following steps:

[0056] (1a) Obtain the Zhuhai-1 OHS hyperspectral data and perform preprocessing, where the preprocessing includes radiometric calibration, atmospheric correction, and orthorectification processing. The radiometric calibration, atmospheric correction, and orthorectification are implemented based on ENVI software, and then clip it using vector administrative boundary data to obtain the preprocessed Zhuhai-1 OHS hyperspectral data;

[0057] (1b) Obtain the Sentinel-2 multispectral data and perform preprocessing. Select Sentinel-2 images with cloud cover less than 5% on the engine platform. After cloud removal and linear spatial interpolation, calculate the EVI index, i.e., the enhanced vegetation index, to obtain the preprocessed Sentinel-2 data;

[0058] (1c) Obtain the global land cover product of the European Space Agency and extract the cultivated land distribution data, where the cultivated land distribution data type number is 40.

[0059] The specific steps of step (2) include the following steps:

[0060] (2a) Use the cultivated land distribution data provided by the global land cover product of the European Space Agency to perform masking on the preprocessed Sentinel-2 data to obtain the masked Sentinel-2 data;

[0061] (2b) Screen the key time windows for the masked Sentinel-2 data, draw the curve of the sample mean of the enhanced vegetation index EVI, and screen the key time windows that can identify corn according to the place with the largest mean difference;

[0062] (2c) Draw the box plot of the sample enhanced vegetation index EVI according to the screened key time windows, determine the threshold conditions for extracting corn, and then perform threshold extraction on the masked Sentinel-2 data under the key time windows to obtain the corn planting distribution data; As Figure 2 shown, determine the threshold conditions for extracting corn EVI 5.26 <0.3, EVI 7.30 >0.5, EVI 9.20 <0.4 to obtain the corn planting distribution data;

[0063] (2d) Mask the preprocessed Zhuhai-1 OHS hyperspectral data with the corn planting distribution data.

[0064] The specific steps of step (3) include the following steps:

[0065] (3a) Calculate the first-order statistics of the sample hyperspectral features in the masked Zhuhai-1 OHS hyperspectral data to obtain the class mean matrix A:

[0066]

[0067] Among them, m ij is the average value of the training samples of the j-th class in the i-th feature, where i = 1, 2, …, d, j = 1, 2, …, c, d represents the feature dimension, which is 32 in the present invention; c represents the number of classes, which is 3 in the present invention; a i is the row vector of the i-th feature of the class mean matrix A;

[0068] (3b) The class mean matrix A performs clustering on the feature row vector a i using the k-means clustering algorithm, clustering the approximate vectors into one cluster, and the k-means clustering cluster G k is:

[0069] G ｋ = {a1, …, a ｉ} (2)

[0070] Among them, G1 represents the first clustering cluster, which is composed of the feature row vector a ｉ ; G ｋ represents the k-th clustering cluster, n represents the number of clustering clusters, and k ∈ 1, 2, …, n;

[0071] (3c) Finally, the spectral features are obtained by calculating the mean value of each clustering cluster G k . The number of the new spectral features is equal to the number of clustering clusters. The formula for extracting the new spectral feature Z is as follows:

[0072]

[0073] Among them, z1 represents the first spectral feature, which is obtained by calculating the mean value of the clustering cluster G1; z k represents the k-th spectral feature, which is obtained by calculating the mean value of the clustering cluster G k ; k represents the number of clustering clusters, and k is 5 in the present invention.

[0074] The specific step (4) refers to: performing opening and closing operations with structural element scales of 2, 4, and 6 on the constructed spectral features. The results are shown as Figure 3 shown, and the multi-scale morphological space features are obtained. The calculation formulas for the morphological opening and closing operations are shown in Formulas (4) and (5):

[0075]

[0076]

[0077] Among them, MP γ represents the image after the opening operation; MP Φ represents the image after the closing operation; γ s (I and Φ s(I) represents the opening operation and the closing operation performed on the image I respectively; s is the size of the structural element. When s = 0, it means that no operation is performed on the original image I.

[0078] The morphological closing operation image gradually fills the small dark elements inside the object, while the image after the opening operation suppresses the adjacent small bright elements. Figure 3 It shows that the object elements in the image after the morphological operation are more homogeneous, and the object edge features are clearer. As the SE gets larger and larger, the spatial details of the small objects gradually disappear.

[0079] Step (5) specifically includes the following steps:

[0080] (5) Normalize the constructed spectral features and spatial features and input them into the support vector machine classifier to find the optimal parameters of the classifier, and obtain the extraction model of hybrid maize seeds;

[0081] (5a) Perform Z-Score normalization on the spectral features and spatial features, and calculate the Fisher discriminant ratio to measure the separability of the features. As Figure 4 shown, the Fisher discriminant ratio formula:

[0082]

[0083]

[0084] Among them, FR i represents the Fisher discriminant ratio of class i under the given feature, μ i is the mean value of the features of class i, is the variance of the features of class i, and i * is the class with the smallest distance from the mean value of class i to the mean values of all other classes; the maximum value of FR for each broken line belongs to the morphological features, indicating that the morphological features have better separability in the classification of hybrid maize seeds.

[0085] (5b) Input the normalized features into the support vector machine classifier, and use the grid search method combined with the 5-fold cross-validation method to find the optimal parameters of the classifier, and obtain the extraction model of hybrid maize seeds.

[0086] Taking the field-collected samples as the validation set, calculate the confusion matrix for the extraction results of the hybrid maize seed planting areas generated by four different feature combination schemes (see Table 1), conduct accuracy verification, compare the models based on the overall accuracy, and examine the mapping effect of the optimal extraction model on hybrid maize seeds. The overall spatial planting distribution laws of the four groups of models for hybrid maize seeds are consistent. It can be seen that the classification results of Model 4 and Model 2 are more homogeneous inside the plots than those of Model 1 and Model 3.

[0087] Table 1 Four extraction schemes for hybrid maize seeds

[0088]

[0089] Table 2 shows the extraction accuracies of four models. After comparison, Model 4, which is the solution proposed by the present invention, is the model with the best performance in the extraction of hybrid maize seeds. Its overall accuracy is as high as 94.10%. The mapping accuracy (producer accuracy, PA) and user accuracy (user accuracy, UA) of hybrid maize seeds are 91.28% and 92.43% respectively. The accuracy of the model using only spectral features has been improved, which indicates that the fusion of spectral and spatial features has greatly improved the classification accuracy.

[0090] Comparison of the extraction accuracies of four models in Table 2

[0091]

[0092] For the optimal extraction model, that is, the present invention, the images of the seedling stage, jointing stage and milk-ripe stage are input respectively for the extraction of hybrid maize seeds, and the overall accuracy is used for evaluation.

[0093] Comparison of the data accuracy of the optimal extraction model input with data of different dates in Table 3

[0094]

[0095] Table 3 gives the accuracy results obtained by inputting the data of three periods under the optimal extraction model. The results show that the classification accuracy of hybrid maize seeds is the highest at the milk-ripe stage, followed by the jointing stage, and the worst at the seedling stage. This shows that there are significant differences between hybrid maize seeds and field maize at the milk-ripe stage, while the recognition effect of hybrid maize seeds at the seedling stage is not good. Therefore, the recognition time of hybrid maize seeds in the present invention is at the milk-ripe stage.

[0096] In summary, the present invention uses the first-order statistical class mean matrix and k-means clustering to extract spectral features, which is simple in operation and increases the inter-class difference between hybrid maize seeds and field maize; the multi-scale morphological spatial features adopted by the present invention improve the "salt and pepper noise" phenomenon, retain a relatively complete plot structure, and improve the extraction accuracy of hybrid maize seeds.

Claims

1. A method for extracting the planting area of hybrid maize seeds based on the fusion of spectral and spatial features, characterized in that: The method includes the following steps in sequence: (1) Obtain the Zhuhai-1 OHS hyperspectral data, Sentinel-2 multispectral data of the study area, and the global land cover product of the European Space Agency, and preprocess the obtained data; (2) Use the cultivated land distribution data in the global land cover product of the European Space Agency to mask the preprocessed Sentinel-2 data to obtain the masked Sentinel-2 data. Use the enhanced vegetation index threshold method to extract the corn planting distribution data from the masked Sentinel-2 data, and use the extracted corn planting distribution data to mask the preprocessed Zhuhai-1 OHS hyperspectral data; (3) Construct spectral features using the class mean matrix clustering method; (4) Construct spatial features based on the spectral features using the multi-scale morphology method; (5) Normalize the constructed spectral features and spatial features and input them into the support vector machine classifier to find the optimal parameters of the classifier, and obtain the hybrid corn extraction model; (6) Input the images of the seedling stage, jointing stage, and milk ripening stage into the hybrid corn extraction model, perform hybrid corn extraction respectively, and use the overall accuracy for evaluation.

2. The method for extracting the planting area of hybrid seed production maize based on the fusion of spectral and spatial features according to claim 1, wherein: The specific steps of step (1) include the following steps: (1a) Obtain the Zhuhai-1 OHS hyperspectral data and perform preprocessing. The preprocessing includes radiometric calibration, atmospheric correction, and orthorectification processing. The radiometric calibration, atmospheric correction, and orthorectification are implemented based on the ENVI software, and then crop using the vector administrative boundary data to obtain the preprocessed Zhuhai-1 OHS hyperspectral data; (1b) Obtain the Sentinel-2 multispectral data and perform preprocessing. Select Sentinel-2 images with a cloud cover of less than 5% on the engine platform, perform cloud removal, linear spatial interpolation, and calculate the EVI index, i.e., the enhanced vegetation index, to obtain the preprocessed Sentinel-2 data; (1c) Obtain the global land cover product of the European Space Agency and extract the cultivated land distribution data. The cultivated land distribution data type number is 40.

3. The method for extracting the planting area of hybrid seed production maize based on the fusion of spectral and spatial features according to claim 1, wherein: The specific steps of step (2) include the following steps: (2a) Use the cultivated land distribution data provided by the global land cover product of the European Space Agency to mask the preprocessed Sentinel-2 data to obtain the masked Sentinel-2 data; (2b) Screen the key time windows for the masked Sentinel-2 data, draw the sample mean curve of the enhanced vegetation index EVI, and select the key time windows that can identify corn according to the place with the largest mean difference; (2c) Draw the box plot of the sample enhanced vegetation index EVI according to the selected key time windows, determine the threshold conditions for extracting corn, and then perform threshold extraction of the corn planting distribution data on the masked Sentinel-2 data under the key time windows; (2d) Mask the preprocessed Zhuhai-1 OHS hyperspectral data with the corn planting distribution data.

4. The method for extracting the planting area of hybrid maize seeds based on the fusion of spectral and spatial features according to claim 1, wherein: The specific steps of step (3) include the following steps: (3a) Calculate the first-order statistics of the sample hyperspectral features in the masked Zhuhai-1 OHS hyperspectral data to obtain the class mean matrix A: where m ij is the average value of the training samples of the j-th class in the i-th feature, i = 1, 2, …, d, j = 1, 2, …, c, d represents the feature dimension, and x represents the number of classes; a i is the row vector of the i-th feature of the class mean matrix A; (3b) The class mean matrix A performs clustering on the feature row vector a through the k-means clustering algorithm i to cluster the approximate vectors into one cluster. The k-means clustering cluster G k is as follows: G k = {a1,…,a i} (2) Among them, G1 represents the first clustering cluster, which is composed of the feature row vector a i ; G k represents the k-th clustering cluster, n represents the number of clustering clusters, and k ∈ 1, 2, …, n; (3c)Finally, the spectral features are obtained by calculating the mean of each clustering cluster G k . The number of new spectral features is equal to the number of clustering clusters. The formula for extracting the new spectral feature Z is as follows: Among them, z1 represents the first spectral feature, which is obtained by taking the mean of the clustering cluster G1; z k represents the k-th spectral feature, which is obtained by taking the mean of the clustering cluster G k ; k represents the number of clustering clusters.

5. The method for extracting the planting area of hybrid maize seeds based on the fusion of spectral and spatial features according to claim 1, characterized in that: The specific content of step (4) is as follows: perform opening and closing operations with structural element scales of 2, 4, and 6 on the constructed spectral features to obtain multi-scale morphological space features. The calculation formulas for morphological opening and closing operations are shown in Formulas (4) and (5): Among them, MP Υ represents the image after the opening operation; MP Φ represents the image after the closing operation; Υ s (I) and Φ s (I) respectively represent the opening operation and the closing operation performed on the image I; s is the size of the structuring element. When s = 0, it means that no operation is performed on the original image I.

6. The method for extracting the planting area of hybrid maize seeds based on the fusion of spectral and spatial features according to claim 1, wherein: The specific content of step (5) includes the following steps: (5) Normalize the constructed spectral features and spatial features and input them into a support vector machine classifier to find the optimal parameters of the classifier, and obtain a seed production maize extraction model; (5a) Perform Z-Score normalization on the spectral features and spatial features, and calculate the Fisher discriminant ratio to measure the separability of the features. The Fisher discriminant ratio formula: Among them, FR i represents the Fisher discriminant ratio of class i under a given feature, μ i is the mean of the features of class i, is the variance of the features of class i, and i * is the class with the smallest distance from the mean of class i to the means of all other classes; (5b) Input the normalized features into a support vector machine classifier, and use the grid search method combined with the 5-fold cross-validation method to find the optimal parameters of the classifier, and obtain a seed production maize extraction model.

Images