Hyperspectral image semi-supervised classification method and system based on multi-scale homogeneous region

Abstract

The application relates to a hyperspectral image semi-supervised classification method and system based on multi-scale homogeneous regions, wherein the method comprises the following steps: generating first homogeneous regions with a larger scale by using a first segmentation method and generating second homogeneous regions with a smaller scale by using a second segmentation method on a to-be-processed image, the first homogeneous regions comprise a plurality of first patches, and the second homogeneous regions comprise a plurality of second patches; the purity of the first homogeneous regions and the second homogeneous regions is evaluated to determine the feasibility of expanding training samples; the nearest point to an initial training sample in a first patch with a label in the first homogeneous regions is expanded as a training sample by using a K nearest neighbor algorithm; the label of a second patch with a label in the second patch is voted to expand training samples by using a majority voting strategy; and after classification is completed, the accuracy of a classification result is evaluated according to overall classification accuracy, average accuracy and a Kappa coefficient. The semi-supervised classification method can significantly improve classification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, in particular to a hyperspectral image semi-supervised classification method and system based on multi-scale homogeneous regions. BACKGROUND

[0002] With the progress of optical sensor technology, high spatial, spectral and temporal resolution can be achieved at present, especially the rich spectral information contained in hyperspectral images, which opens up a new application field. Hyperspectral images have very high spectral resolution and can obtain continuous spectral information of ground objects, providing more abundant ground information than traditional RGB images and multispectral images. Hyperspectral images can simultaneously obtain spectral information, ground object spatial information and radiation information, thus having great application value and broad development prospects, and playing an increasingly important role in agriculture, forestry, ocean, military and other fields.

[0003] In hyperspectral remote sensing, classification is an important issue worth attention. According to the use of training samples, the classification method of hyperspectral image is divided into supervised, unsupervised and semi-supervised classification. Unsupervised classification method does not need any prior knowledge training data, and automatically classifies pixels with similar spectral characteristics into a class. The cost of unsupervised clustering is the lowest, but the effect is often poor and the relationship between clusters and classes cannot be determined. Common unsupervised clustering includes K-means clustering, Gaussian mixture model, spectral clustering, etc.

[0004] Traditional supervised classification only uses labeled data to train classification, and uses a set of representative samples called training samples to classify the input data of each class. At present, the supervised classification methods mainly include maximum likelihood estimation method, nearest neighbor classifier, decision tree, random forest, support vector machine (SVM), etc. Supervised classification often needs a large number of training samples, and its classification ability increases with the increase of available training samples. However, obtaining reliable and accurate class labels for pixels is a costly and time-consuming task, which not only consumes a lot of manpower, but also consumes financial and material resources. Therefore, the supervised classification has the problem of limited labeled samples, and the semi-supervised classification method arises at the historic moment.

[0005] Semi-supervised classification of hyperspectral images is a combination of supervised classification and unsupervised classification. Semi-supervised classification method aims to learn the relationship between a large number of unlabeled samples and labeled samples to improve classification performance. Semi-supervised classification methods include co-training algorithm, generative model-based method, low-density partition method, and graph-based semi-supervised classification. Since a large amount of unlabeled sample information is used, semi-supervised classification method can obtain better accuracy than supervised classification with fewer labeled samples. Recent researches show that semi-supervised classification has important significance for small sample problem in hyperspectral classification.

[0006] At present, more and more researchers use semi-supervised learning method to realize the classification of hyperspectral images, and how to use a small amount of labeled samples to obtain high-precision classification results has important research significance. However, the existing semi-supervised classification method of homogeneous region expansion training sample only uses single-scale homogeneous region expansion sample, and does not consider the influence of noise points at the edge of large homogeneous region patches, which directly affects the number and accuracy of the expanded samples, and further directly affects the classification effect. SUMMARY

[0007] In view of the above-mentioned defects of the prior art, the purpose of the present application is to provide a hyperspectral image semi-supervised classification method and system based on multi-scale homogeneous region, which is used to solve the problem that it is difficult to obtain high-precision expanded samples in the prior art, thereby affecting the classification effect of hyperspectral images.

[0008] In order to achieve the above-mentioned purpose and other related purposes, the present application provides the following technical scheme:

[0009] A hyperspectral image semi-supervised classification method based on multi-scale homogeneous region, comprising:

[0010] Generating a first homogeneous region of a larger scale by using a first segmentation method and generating a second homogeneous region of a smaller scale by using a second segmentation method for the image to be processed, wherein the first homogeneous region comprises a plurality of first patches, and the second homogeneous region comprises a plurality of second patches;

[0011] Evaluating the purity of the first homogeneous region and the second homogeneous region to determine the feasibility of expanding training samples;

[0012] Expanding the nearest point to the initial training sample in the labeled first patch in the first homogeneous region as a training sample by using K nearest neighbor algorithm;

[0013] Expanding training samples by voting the labels of the labeled second patches in the second patches by using majority voting strategy;

[0014] After classification, the accuracy of the classification result is evaluated according to the overall classification accuracy, the average accuracy and the Kappa coefficient.

[0015] As a preferred scheme of the application, the first segmentation method uses JSEG segmentation method, and the second segmentation method uses SLIC segmentation method.

[0016] As a preferred scheme of the application, the first segmentation method uses JSEG segmentation method, and the second segmentation method uses SLIC segmentation method.

[0017] Extracting the first three principal components of the hyperspectral image as an RGB image and performing Kmeans clustering;

[0018] Selecting a suitable window size to calculate J value and generating a multi-scale J image;

[0019] The points with local J value less than a predetermined threshold are selected as seed points on the basis of the multi-scale J image, and the pixels similar to the seed points in the neighborhood of the seed points are merged together to finally obtain a JSEG segmentation image.

[0020] As a preferred scheme of the application, the generating of the second homogenous region of smaller scale by the second segmentation method specifically comprises:

[0021] K points are selected as clustering center points in the whole image, and the clustering center is moved to the position of the minimum gradient distance in the neighborhood;

[0022] All pixels in the image are subjected to clustering analysis, the distance between each pixel and all clustering centers is calculated, and the clustering center point with the minimum distance to each pixel is attributed to the clustering;

[0023] The average vector of all points in the clustering is recalculated as the center of the next clustering;

[0024] The above process is repeated until the super-pixel is output after convergence.

[0025] As a preferred scheme of the application, after the generating of the first homogenous region of larger scale by the first segmentation method, the method further comprises purifying the impure first patch by using the mean shift algorithm;

[0026] During the purifying, the average variance is used as an index for measuring the purity of the first patch, and the calculation formula of the average variance is specifically as follows:

[0027]

[0028] wherein c represents the total number of bands of the hyperspectral image, N represents the number of pixels in the first patch, xi represents the feature vector of the i-th element in the first patch, i xi represents the feature vector of the i-th element in the first patch, xi represents the feature vector of the i-th element in the first patch,

[0029] As a preferred scheme of the application, the expanding of the nearest point to the initial training sample in the first homogenous region as a training sample by using the K nearest neighbor algorithm specifically comprises:

[0030] The distance between the feature vector of the k-th unlabeled point and the feature vector of the i-th labeled point is calculated according to the formula wherein n represents the number of elements in the second patch, j represents the number of labeled points, m represents the number of feature vectors of each element, the feature vector of the k-th unlabeled point is U k = u1, u2,..., um, m the feature vector of the i-th labeled point is L i = l1, l2,..., lm.m ;

[0031] ranking the distances dist between the n-k unlabeled points and the i-th labeled point in the second patch in descending order, wherein dist = sort (dist 1,i ,dist 2,i ......dist n-k,i );

[0032] selecting K unlabeled points closest to the labeled point and assigning the K unlabeled points with the same label as the labeled point;

[0033] until the j labeled points are traversed, counting the label frequency of the unlabeled points, and removing the unlabeled points with different labels and the repeated unlabeled points with the same label if the unlabeled points appear multiple times.

[0034] As a preferred scheme of the application, the label of the second patch in the second patch is voted and augmented by using the majority voting strategy, and the training sample specifically includes:

[0035] When there is only one labeled point in the second patch, the label mode of the elements in the second patch is taken as the label of the labeled point, and all the points in the second patch are assigned with the same label as the label mode;

[0036] When there are multiple labeled points of the same category in the second patch, the label of the category is taken as the label of the second patch, and all the points in the second patch are assigned with the same label as the labeled point;

[0037] When there are labeled points of different categories in the second patch, the label of all the labeled points is voted, the label mode of the labeled points is taken as the label of the second patch, and the label of the labeled points is kept unchanged, and the other elements in the second patch are assigned with the same label as the label of the second patch.

[0038] As a preferred scheme of the application, the purity of the first homogeneous region and the second homogeneous region is evaluated to determine the feasibility of the augmented training sample, and specifically includes:

[0039] The purity is calculated by using the formula , wherein n represents the number of patches segmented from the hyperspectral image, m i represents the label mode of the labeled samples in the i-th patch, S i represents the total number of labeled samples in the i-th patch.

[0040] As a preferred scheme of the application, the accuracy of the classification result is evaluated according to the overall classification accuracy, the average accuracy and the Kappa coefficient, and specifically includes:

[0041] The overall classification accuracy is calculated by formula The overall classification accuracy is calculated by formula The overall classification accuracy is calculated by formula

[0042] The average accuracy is calculated by formula The average accuracy is calculated by formula The average accuracy is calculated by formula

[0043] The Kappa coefficient is calculated by formula The Kappa coefficient is calculated by formula The Kappa coefficient is calculated by formula

[0044] Wherein, c is the number of categories of hyperspectral remote sensing image, the size of the confusion matrix is c rows and c columns, the total number of test samples of remote sensing image is n, the confusion matrix is M, M ij Indicates the number of samples of the jth misclassification into the ith category.

[0045] To solve the above technical problems, the embodiment of the present application also provides a hyperspectral image semi-supervised classification system based on multi-scale homogeneous regions, comprising:

[0046] An image segmentation unit is configured to generate larger-scale first homogeneous regions by using a first segmentation method and smaller-scale second homogeneous regions by using a second segmentation method for the image to be processed, wherein the first homogeneous regions comprise a plurality of first patches, and the second homogeneous regions comprise a plurality of second patches.

[0047] A first expansion unit is configured to expand the nearest point to the initial training sample in the labeled first patch in the first homogeneous region as a training sample by using a K-nearest neighbor algorithm.

[0048] A second expansion unit is configured to vote for the label of the labeled second patch in the second patch by using a majority voting strategy to expand the training sample.

[0049] A first evaluation unit is configured to evaluate the purity of the first homogeneous regions and the second homogeneous regions to determine the feasibility of expanding the training sample.

[0050] A second evaluation unit is configured to evaluate the accuracy of the classification result according to the overall classification accuracy, the average accuracy and the Kappa coefficient after the classification is completed.

[0051] As described above, the present application has the following beneficial effects:

[0052] The embodiment of the present application provides a hyperspectral image semi-supervised classification method and system based on multi-scale homogeneous regions, wherein the method comprises the following steps: generating a first homogeneous region with a larger scale and a second homogeneous region with a smaller scale by using a first segmentation method and a second segmentation method on a to-be-processed image, the first homogeneous region comprises a plurality of first patches, and the second homogeneous region comprises a plurality of second patches; using a K nearest neighbor algorithm to expand a point closest to an initial training sample in a first patch with a label in the first homogeneous region as a training sample; using a majority voting strategy to vote for the label of a second patch with a label in the second patch to expand the training sample; evaluating the purity of the first homogeneous region and the second homogeneous region to determine the feasibility of expanding the training sample; and evaluating the accuracy of the classification result according to the overall classification accuracy, the average accuracy and the Kappa coefficient after the classification is completed. By using different segmentation methods to obtain the first homogeneous region with a larger scale and the second homogeneous region with a smaller scale, more training samples are combined and expanded, the K nearest neighbor algorithm is used to expand the training sample in the first homogeneous region with a larger scale, only K points closest to the labeled sample in the first patch are expanded, the influence of edge noise points far away from the labeled sample is reduced, the expansion of the error sample is reduced, the sample accuracy is improved, the increase of the number of training samples with high accuracy can improve the model generalization ability, and therefore the classification accuracy is improved, which has important significance for land use change monitoring and target identification. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A method flowchart of the embodiment one of the present application is shown.

[0054] Figure 2 A method flowchart for generating homogeneous regions with different scales in the embodiment one of the present application is shown.

[0055] Figure 3 A method flowchart for generating the first homogeneous region in the embodiment one of the present application is shown.

[0056] Figure 4 A method flowchart for generating the second homogeneous region in the embodiment one of the present application is shown.

[0057] Figure 5a A JSEG segmentation result diagram of PaviaU data in the embodiment two of the present application is shown.

[0058] Figure 5b A JSEG segmentation result diagram of Salinas data in the embodiment two of the present application is shown.

[0059] Figure 5c A JSEG segmentation result diagram of IndianP data in the embodiment two of the present application is shown.

[0060] Figure 6aSLIC segmentation result map of PaviaU data in Embodiment Two of the present application is shown.

[0061] Figure 6b SLIC segmentation result map of Salinas data in Embodiment Two of the present application is shown.

[0062] Figure 6c SLIC segmentation result map of IndianP data in Embodiment Two of the present application is shown.

[0063] Figure 7a Classification effect map of IndianP data in Embodiment Five of the present application obtained by using SVM method is shown.

[0064] Figure 7b Classification effect map of IndianP data in Embodiment Five of the present application obtained by using CNN-PF method is shown.

[0065] Figure 7c Classification effect map of IndianP data in Embodiment Five of the present application obtained by using AGR method is shown.

[0066] Figure 7d Classification effect map of IndianP data in Embodiment Five of the present application obtained by using AS-SVM method is shown.

[0067] Figure 8a Classification effect map of Salinas data in Embodiment Five of the present application obtained by using SVM method is shown.

[0068] Figure 8b Classification effect map of Salinas data in Embodiment Five of the present application obtained by using CNN-PF method is shown.

[0069] Figure 8c Classification effect map of Salinas data in Embodiment Five of the present application obtained by using AGR method is shown.

[0070] Figure 8d Classification effect map of Salinas data in Embodiment Five of the present application obtained by using AS-SVM method is shown.

[0071] Figure 9a Classification effect map of PaviaU data in Embodiment Five of the present application obtained by using SVM method is shown.

[0072] Figure 9b Classification effect map of PaviaU data in Embodiment Five of the present application obtained by using CNN-PF method is shown.

[0073] Figure 9cThe classification effect diagram obtained by using the AGR method for the PaviaU data in Embodiment Five of the present application is shown.

[0074] Figure 9d The classification effect diagram obtained by using the AS-SVM method for the PaviaU data in Embodiment Five of the present application is shown.

[0075] Figure 10 The basic structure schematic diagram of Embodiment Six of the present application is shown.

[0076] Figure 11 The basic structure block diagram of the computer equipment of Embodiment Seven of the present application is shown. DETAILED DESCRIPTION

[0077] The embodiments of the present application are described below by specific embodiments, and other advantages and effects of the present application can be easily understood by those skilled in the art from the content disclosed in the specification.

[0078] Please refer to Figures 1 to 11 It should be understood that the structures, proportions, sizes, etc. shown in the drawings attached to the specification are only used to understand and read the content disclosed in the specification by those skilled in the art, and are not used to limit the conditions that the present application can be implemented, and therefore do not have technical significance. Any modification of the structure, change of the proportional relationship or adjustment of the size, without affecting the effects that the present application can produce and the purposes that the present application can achieve, should still fall within the scope of the technical content disclosed in the present application. At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" in the specification are only for the convenience of clear description, and are not used to limit the scope of the present application, and the change or adjustment of the relative relationship, without substantially changing the technical content, is also considered as the scope of the present application.

[0079] Embodiment One

[0080] Please refer to Figure 1 The present application provides a hyperspectral image semi-supervised classification method based on multi-scale homogeneous regions, comprising:

[0081] S100, a first segmentation method is used to generate a first homogeneous region of a larger scale, and a second segmentation method is used to generate a second homogeneous region of a smaller scale, the first homogeneous region comprises a plurality of first patches, and the second homogeneous region comprises a plurality of second patches, wherein the first segmentation method adopts a JSEG segmentation method, and the second segmentation method adopts a SLIC segmentation method.

[0082] The JSEG segmentation method is a JSEG image segmentation algorithm, mainly including color quantization and spatial segmentation. The color quantization is to combine the similar colors which are not important in the original image into one color, and the purpose is to extract the representative colors in the image to distinguish adjacent regions. In the color quantization graph, the pixel value is the label of the quantized color class, and different pixel values represent different quantized colors. The SLIC segmentation method is a simple linear iterative clustering (SLIC) algorithm proposed by Achanta et al. in 2010. The SLIC algorithm can better express the texture features of ground object information and improve the adhesion of ground object boundaries. The SLIC superpixel can obtain a very high-purity homogeneous region in the case of a large number of segmented patches. Therefore, the SLIC algorithm is used to generate a small-size homogeneous region in this paper.

[0083] As shown in Figure 2 , Figure 2 A method flowchart for generating homogeneous regions of different scales is shown. The segmentation process mainly includes: extracting the first three principal components of the pre-hyperspectral image, using the JSEG segmentation method for segmentation processing, finding impure patches on the basis of JSEG segmentation, purifying the impure patches by using the mean shift algorithm, and finally merging the isolated points with the patches of the surrounding first homogeneous region to generate the JSEG homogeneous region, i.e. the first homogeneous region. At the same time, after extracting the first three principal components of the pre-hyperspectral image, a suitable number of segmented patches is set, and the SLIC segmentation method is used for simple superpixel segmentation to obtain the SLIC homogeneous region, i.e. the second homogeneous region.

[0084] Specifically, as shown in Figure 3 , Figure 3 A method flowchart for generating the first homogeneous region is shown. The first segmentation method for generating a larger-scale first homogeneous region specifically includes:

[0085] S101, extract the first three principal components of the hyperspectral image as an RGB image and perform Kmeans clustering. Kmeans clustering specifically refers to a k-means clustering algorithm, which is an iterative clustering analysis algorithm. The steps are as follows: first, divide the data into K groups, then randomly select K objects as the initial cluster centers, then calculate the distance between each object and each seed cluster center, and assign each object to the cluster center closest to it. The cluster center and the objects assigned to them represent a cluster. Each sample is assigned, and the cluster center of the cluster is recalculated according to the existing objects in the cluster. This process will be repeated until a certain termination condition is met. The termination condition can be that no (or a minimum number of) objects are reassigned to different clusters, no (or a minimum number of) cluster centers change, or the error sum of squares is locally minimized.

[0086] S102, select a suitable window size to calculate J value, generate multi-scale J image, J image is also called J value image, which is composed of J value, in J image, small J value represents the center of the region, and large J value represents the boundary of the region, the calculation method of specific J value pair is prior art, which is not described here;

[0087] S103, select the point with local J value less than the predetermined threshold value as the seed point on the basis of the multi-scale J image, and merge the pixels similar to the seed point in the neighborhood to obtain the JSEG segmentation map finally.

[0088] Meanwhile, after JSEG segmentation, it also includes:

[0089] S104, purify the first patch by using the mean shift algorithm, when purifying, the average variance is used as the index to measure the purity of the first patch, and the calculation formula of the average variance is specifically:

[0090]

[0091] Wherein, c represents the total band number of hyperspectral image, N is the number of pixels in the first patch, x i represents the feature vector of the i-th element in the first patch, represents the feature center of the first patch.

[0092] A large number of experimental results show that the proportion of pure patches is much higher than 50%, and all impure patches can be purified by taking a value less than 50% as a threshold value, and in order to avoid the trouble caused by adjusting parameters, 50% is always taken as the threshold value in this embodiment.

[0093] As Figure 4 shown, Figure 4 the method flow chart for generating the second homogeneous region is shown, and the second homogeneous region with smaller scale is generated by using the second segmentation method, which specifically includes:

[0094] S105, select K points as clustering center points in the whole image, and move the clustering center to the position of the minimum gradient distance in the neighborhood, so as to prevent the clustering center from being located at the boundary;

[0095] S106, perform clustering analysis on all pixels in the image, calculate the distance between each pixel and all clustering centers, and attribute the clustering center point with the minimum distance to each pixel to the clustering;

[0096] S107, recalculate the average vector of all points in the clustering as the center of the next clustering;

[0097] S108, repeat the above process until the superpixel is output after convergence.

[0098] S200, evaluate the purity of the first homogeneity region and the second homogeneity region to determine the feasibility of expanding the training sample.

[0099] Specifically, since the purity of the homogeneity region of the hyperspectral remote sensing image directly determines the accuracy of the subsequent image information extraction, especially in the classification based on image segmentation, high-precision homogeneity region is required, and the quality of the homogeneity region needs certain evaluation index, and the good or bad of the evaluation index directly affects the feasibility of the subsequent homogeneity region for expanding the training sample, in the embodiment, the formula is used to calculate the purity, wherein the number of patches of the hyperspectral image segmentation is n, m i represents the class mode of the labeled sample in the i-th patch, S i represents the total number of labeled samples in the i-th patch.

[0100] S300, using K-neighbor algorithm to expand the nearest point to the initial training sample in the labeled first patch in the first homogeneity region as a training sample.

[0101] For the large-scale first homogeneity region generated by the JSEG segmentation method, there are more elements in some first patches and under-segmentation phenomenon may occur, at this time, if all elements in the labeled first patch are directly assigned the same label as the labeled point, the number of error points will be increased. The labeled first patch refers to the first patch containing the labeled sample. In order to reduce the error rate of the homogeneity region expansion point caused by the impurity of the patch and ensure the quality of the expanded sample, the K-neighbor algorithm is used to expand the training sample in the first homogeneity region, only the nearest point to the initial training sample in the labeled first patch is expanded, and K has no fixed value, which is determined by the number of elements in the first patch. In the embodiment, the value of K of all labeled first patches is half of the number of elements in the first patch, and a large number of experiments prove that the effect is best at this time.

[0102] S300 specifically includes:

[0103] S301, according to the formula , the distance between the feature vector of the k-th unlabeled point and the feature vector of the i-th labeled point is calculated, wherein the number of elements in the second patch is n, the number of labeled points is j, the number of feature vectors of each element is m, the feature vector of the k-th unlabeled point is U k = u1, u2... u m , and the feature vector of the i-th labeled point is L i = l1, l2... l m ;

[0104] S302, the distances dist of the n-k unlabeled points in the second patch to the i-th labeled point are arranged in descending order, wherein dist = sort (dist1,i dist 2,i dist n-k,i );

[0105] S303, selecting K unmarked points closest to the marked point, and assigning the K unmarked points with the same label as the marked point;

[0106] S304, repeating steps S301-S303 until all j marked points are traversed, counting the label frequency of the unmarked points, and removing the unmarked points with different labels and the repeated unmarked points with the same label if the unmarked points appear multiple times.

[0107] S400, using a majority voting strategy to vote for the label of the marked second patch in the second patch.

[0108] The SLIC segmentation method produces a large number of second patches and the number of elements in the second patch is relatively uniform. This segmentation method produces a large number of over-segmented patches, making the purity of the second patch extremely high. Therefore, a simple majority voting strategy is used to vote for the label of the marked second patch, and the mode of the label of the marked sample in the second patch is found as the label of the second patch, which specifically includes:

[0109] S401, when there is only one marked point in the second patch, the mode of the label of the elements in the second patch is taken as the label of the marked point, and all points in the second patch are assigned the same label as the mode;

[0110] S402, when there are multiple marked points of the same category in the second patch, the label of the category is taken as the label of the second patch, and all points in the second patch are assigned the same label as the marked point;

[0111] S403, when there are marked points of different categories in the second patch, the labels of all marked points are voted, the mode of the label of the marked point is selected as the label of the second patch, and the label of the marked point is kept unchanged, and the other elements in the second patch are assigned the same label as the label of the second patch.

[0112] S500, after classification, the accuracy of the classification result is evaluated according to the overall classification accuracy, average accuracy and Kappa coefficient.

[0113] Specifically, the classification result needs to be measured by certain standards, and the classification result evaluation index in the embodiment is three commonly used indexes: overall classification accuracy (OA), average accuracy (AA) and Kappa coefficient; the basis for calculating OA, AA and Kappa coefficient is a confusion matrix, which is a matrix for measuring the accuracy of the classifier classification, comparing the number of predicted categories and true categories, the sum of a certain row of the matrix represents the number of ground true samples of the category, the sum of a certain column represents the number of samples predicted as the category, and the data on the main diagonal line represents the number of samples of each category correctly classified.

[0114] Suppose the number of categories of the hyperspectral remote sensing image is c, then the size of the confusion matrix is c rows and c columns, the total number of test samples of the remote sensing image is n, the confusion matrix is M, the number of samples of the jth category misclassified as the ith category, and the specific calculation process of the three classification evaluation indexes is as follows:

[0115] The overall classification accuracy is calculated by the formula The overall classification accuracy is the percentage of the total number of correctly classified test samples of the hyperspectral image to the total number of test samples, and the larger the value, the higher the classification accuracy of the hyperspectral image;

[0116] The average accuracy is calculated by the formula The average accuracy is the average value of the accuracy of each category, and the larger the value, the higher the accuracy;

[0117] The Kappa coefficient is calculated by the formula The Kappa coefficient can quantitatively evaluate the consistency of the hyperspectral classification map and the reference data, and the larger the value, the better the classification result.

[0118] The embodiment of the application obtains larger-scale first homogeneous regions and smaller-scale second homogeneous regions by using different segmentation methods, thereby combining and expanding more training samples, and further purifying impure homogeneous region patches by using the mean shift algorithm, thereby improving the purity of large-scale homogeneous regions. At the same time, the K nearest neighbor algorithm is used to expand training samples in the first homogeneous region of larger scale, and only the K points closest to the marker sample in the first patch are expanded, which can reduce the influence of edge noise points far from the marker sample, thereby reducing the expansion of false samples and improving the sample accuracy. For the same marker sample, different segmentation processes of different methods can obtain different homogeneous regions containing the marker sample, and combining these homogeneous regions can obtain more expanded samples. The small-scale homogeneous region can expand more training samples on the basis of the large-scale homogeneous region expansion samples, further increasing the number of training samples. The increase of the number of high-precision expanded training samples can improve the model generalization ability, thereby improving the classification accuracy, which has important significance for land use change monitoring and target recognition.

[0119] The research shows that the increase of the number of training samples can avoid the overfitting of the classifier and improve the generalization ability of the model, the method for rapidly expanding a large number of high-precision training samples of homogeneous regions of different scales is studied, the SVM classifier is successfully trained under the condition of a small number of actual labeled samples, and the hyperspectral data is effectively classified, experiments are conducted on three common hyperspectral data, and the results show that the method can obtain a large number of high-precision training samples on the Pavia University, Salinas and Indian Pines three kinds of hyperspectral data, and the expansion sample precision reaches more than 99%, 99% and 97% respectively. The SVM classification precision based on the expansion sample of the embodiment is much higher than that of the SVM classification precision using only the original labeled point, compared with the classical SVM, CNN-PF and AGR methods, the overall classification precision, average classification precision and Kappa coefficient of the embodiment are improved to different degrees.

[0120] Embodiment two

[0121] In order to more clearly illustrate the present application, a specific example is described below, in the embodiment, three commonly used hyperspectral image data sets are selected as test data, which are Pavia University data, Salinas data and Indian Pines data. The PaviaU data image size is 610x340, a total of 42776 labeled samples containing 9 categories, the data contains 103 spectral bands, the spatial resolution is 1.3m, and the spectral resolution is 4nm to 12nm; the Salinas data image size is 512x217, the spatial resolution is 3.7m, a total of 204 bands are used in the experiment, the data contains 54129 labeled samples, and contains 16 categories; the Indian Pines data image size is 145x145, the spatial resolution is 20m, a total of 224 bands are contained, 200 bands are used in the experiment after removing 20 bands that cannot be reflected by water, the data contains 10249 labeled samples of 16 categories, the embodiment generates homogeneous regions of different scales on the three data and expands the training samples, and finally performs semi-supervised classification of SVM and compares and analyzes with other three semi-supervised classification methods.

[0122] In order to obtain homogeneous regions of different scales, the JSEG segmentation method is used to generate large-scale JSEG homogeneous regions, and the SLIC segmentation method is used to generate small-scale SLIC homogeneous regions.

[0123] (1) JSEG homogeneous region generation result

[0124] The embodiment carries out traditional JSEG segmentation on three common hyperspectral images, purifies the impure homogeneous region patches in the segmentation results by using the mean shift method, and carries out region merging on the over-segmented patches, so as to finally obtain the large-scale homogeneous region results based on JSEG. The specific results can be referred to the following table, which shows the purity and patch quantity of JSEG segmentation of the three hyperspectral data.

[0125]

[0126] As can be seen from the table, the large-scale generation method based on JSEG segmentation has achieved good results on the three hyperspectral data, in which the segmentation purity of PaviaU and Salinas data reaches more than 99%, and the accuracy of IndianP data with lower resolution also reaches more than 98%, which shows that this method has good applicability and robustness on hyperspectral data.

[0127] Figures 5a-5c The figure is the JSEG segmentation result of PaviaU, Salinas and IndianP data. As can be seen from the figure, although the number of elements in the patch obtained by JSEG segmentation is uneven, good segmentation effect is obtained.

[0128] (2) SLIC homogeneous region generation result

[0129] In order to obtain better segmentation results, a suitable superpixel segmentation number K needs to be selected, and the patch quantity and patch purity are comprehensively considered. Finally, the segmentation number of PaviaU data, Salinas data and IndianP data in this embodiment is selected as 5000, 3000 and 1400 respectively. The specific results are shown in the following table, and the following table is the purity and actual patch number of the homogeneous region obtained by SLIC segmentation of the three hyperspectral data.

[0130]

[0131] As can be seen from the table, the purity of the homogeneous region obtained by SLIC segmentation of the three data is more than 99%, which shows that the purity of the homogeneous region patch obtained by SLIC segmentation is very high.

[0132] Figures 6a-6c The figure is the SLIC homogeneous region result of the three hyperspectral data. As can be seen from the figure, the number of patches obtained by SLIC segmentation is large and the area is small, the size of the homogeneous region patch is uniform and has high homogeneity, and the SLIC segmentation method obtains good homogeneous region segmentation results on the three hyperspectral data.

[0133] Embodiment three

[0134] The embodiment discloses examples of different scale homogeneity region expansion samples, and the two different scale homogeneity region expansion training samples generated in the embodiment are used for sample expansion.

[0135] (1) JSEG homogeneity region expansion training sample

[0136] The JSEG homogeneity region has a large patch area and many elements in the patch. In order to obtain better expansion samples, the K nearest neighbor algorithm is used for JSEG homogeneity region expansion training sample. In order to reduce the influence of accidental errors, all experiments in the embodiment are randomly performed 10 times, and the average value is taken. The specific results are shown in the following table. The following table is the number and accuracy of the training samples expanded by the K nearest neighbor algorithm for the JSEG homogeneity region when the number of initial training samples of each class is 5, 10, 15 and 20 for PaviaU, Salinas and IndianP data. The accuracy of the expanded training samples is obtained by calculating the proportion of the same prediction class and the real class of the expanded training samples.

[0137]

[0138]

[0139] As can be seen from the table, a small amount of initial training samples can be expanded into a large number of high-precision labeled sample points through the JSEG homogeneity region. When 5, 10, 15 and 20 samples are taken for each class, the expansion sample accuracy of PaviaU data and Salinas data is higher than 99%, and the expansion sample accuracy of IndianP data with lower resolution is also higher than 97%. Although the samples expanded by the JSEG homogeneity region contain a small amount of error samples, the number of samples can be greatly increased.

[0140] (2) SLIC homogeneity region expansion training sample

[0141] Compared with the JSEG homogeneity region, the SLIC homogeneity region has more patches and smaller patch area. The majority voting strategy is used for sample expansion. The average number and accuracy of the small scale homogeneity region expansion samples based on SLIC segmentation are shown in the following table:

[0142]

[0143] As can be seen from the table, when the SLIC homogeneity region is expanded, the expansion sample accuracy of PaviaU data and Salinas data is more than 99%, and the expansion sample accuracy of IndianP data with lower resolution is more than 98%, which indicates that the sample accuracy expanded by the SLIC homogeneity region is very high.

[0144] Further, the following table is a comparison of JSEG homogeneity region and SLIC homogeneity region expansion samples:

[0145]

[0146]

[0147] As can be seen from the table, although the SLIC homogeneity region and the JSEG homogeneity region expansion samples have repeated parts, the probability of the same class of the repeated part of the PaviaU and Salinas data is more than 99%, and the consistency of the class of the repeated part of the IndianP data is more than 98%, which shows that the expansion samples of the two methods have high accuracy. In the experiment, the repeated samples and the samples with different labels in the expansion samples of different scale homogeneity regions are excluded, which reduces the expansion of error points, and the small scale homogeneity region adds a large number of samples to the expansion samples of the large scale homogeneity region, which further increases the number of training samples.

[0148] Example Four

[0149] This embodiment discloses an example of support vector machine using expansion training samples, and the SVM classification without parameter selection has the characteristics of fast classification speed and high efficiency. In order to prove the effectiveness of the JSEG homogeneity region expansion training sample on classification, linear SVM without parameter selection is used for classification, and compared with other three classical semi-supervised classification methods.

[0150] (1) Influence of JSEG homogeneity region expansion sample on SVM classification

[0151] Under the condition of determining the initial sample, the linear SVM classification is carried out by using the JSEG homogeneity region expansion training sample, the influence of the JSEG homogeneity region expansion sample on the SVM classification is studied, 10 experiments are carried out for each class of three commonly used hyperspectral data under the condition of selecting 5, 10, 15 and 20 initial samples respectively, and the average value is taken, and the specific results are shown in the following table, and the following table shows the comparison of SVM classification accuracy before and after the expansion of large scale homogeneity region samples.

[0152]

[0153] As can be seen from the table, the classification accuracy of SVM of the three hyperspectral data is greatly improved after the training samples are expanded by JSEG homogeneous regions, and the classification accuracy of IndianP data is improved most obviously. When 5 initial samples are selected for each class, the classification accuracy of IndianP data is improved by 22.24%, the classification accuracy of Salinas data is improved by 9.33%, and the classification accuracy of PaviaU data is improved by 18.17%, which indicates that the training samples expanded by JSEG homogeneous regions have good effect on the improvement of SVM classification. When 20 initial sample points are selected for each class, the SVM classification accuracy of IndianP, Salinas and PaviaU data after the samples are expanded by JSEG is 83.67%, 91.86% and 89.08% respectively.

[0154] (2) Effect of SLIC homogeneous region on SVM classification

[0155] In order to prove the effectiveness of the training samples expanded by SLIC homogeneous regions, this embodiment compares the classification effect of SVM trained by the samples expanded by JSEG homogeneous regions only and the samples expanded by JSEG and SLIC homogeneous regions together, 5, 10, 15 and 20 samples are selected for each class of the three hyperspectral data for 10 experiments and the average value is calculated, and the specific results are shown in the following table. The table shows the comparison of SVM classification accuracy before and after the addition of samples in small-scale homogeneous regions based on the expansion of samples in large-scale homogeneous regions.

[0156]

[0157] As can be seen from the table, the classification accuracy of linear SVM is further improved after the training samples are expanded by small-scale homogeneous regions, and the effect of IndianP data is most obvious among the three data. The classification accuracy of Salinas data and PaviaU data is improved to different degrees. When 5 initial samples are selected for each class, the classification accuracy of IndianP, Salinas and PaviaU data is improved by 1.79%, 1.42% and 0.46% respectively.

[0158] Example Five

[0159] The embodiment discloses examples of semi-supervised comparison experiments, in order to prove the effectiveness of the method of the application, three other classical classification methods are used to carry out comparison experiments on three common hyperspectral data. The SVM is a classification method based on original training samples; the AS-SVM method refers to the SVM classification method of the application based on different scale homogenous region amplification samples; the CNN-PF method is a semi-supervised deep network classification method proposed by Wei Li, which uses a new pixel pair method to expand the training set and classifies the image; and the AGR method is a semi-supervised classification method based on anchor point mapping proposed by Wei Liu in 2010. The specific comparison results are shown in the following table, and the table shows the classification accuracy comparison of different methods on IndianP data.

[0160]

[0161] From the table, it can be seen that when the number of initial samples of each class is 5, 10, 15 and 20, the accuracy of the AS-SVM method of the present application is the highest, followed by the CNN-PF method, which shows the effectiveness of the SVM classification method based on training sample amplification of the present application.

[0162] Figures 7a-7d The figure shows the comparison of classification effect diagrams obtained by using different semi-supervised classification methods on IndianP data. From the figure, it can be seen that there are obvious and large amounts of noise in the classification effect diagrams obtained by using the SVM, CNN-PF and AGR methods, and the classification effect diagram of the AS-SVM method has the least noise and maintains good edge properties. Among the above four classification methods, the AS-SVM method has the best classification effect on the IndianP data.

[0163] The following table shows the classification accuracy comparison of different methods on Salinas data:

[0164]

[0165] From the table, it can be seen that the method with the highest classification accuracy is the AS-SVM method of the present application, followed by the CNN-PF method, which shows the effectiveness of the AS-SVM method of the present application for classification of Salinas data.

[0166] Figures 8a-8d The figure shows the comparison of classification effect diagrams obtained by using different semi-supervised classification methods on Salinas data. From the figure, it can be seen that the AS-SVM method has the best edge retention effect and the least noise in the classification diagram, and has the best classification effect.

[0167] The following table shows the classification accuracy comparison of different methods on PaviaU data:

[0168]

[0169] As shown in the table, the AS-SVM method has the highest accuracy when the initial samples are 5, 10, 15 and 20. When 20 samples are selected for each class, the accuracy of the AS-SVM method is 89.41%, which is 14.25% higher than the accuracy of the SVM method and 32.54% higher than the classification accuracy of the AGR method.

[0170] Figures 9a-9d This is a comparison of classification results obtained using different semi-supervised classification methods on the PaviaU dataset. As can be seen from the figures, the AS-SVM method proposed in this invention achieves the best classification results, with the fewest noise points and the best edge preservation in the classification image.

[0171] Example 6

[0172] This embodiment provides a semi-supervised classification system for hyperspectral images based on multi-scale homogeneous regions, such as... Figure 10 As shown, it includes: an image segmentation unit, used to generate a larger-scale first homogeneous region from the image to be processed using a first segmentation method, and to generate a smaller-scale second homogeneous region using a second segmentation method, wherein the first homogeneous region includes a plurality of first patches, and the second homogeneous region includes a plurality of second patches; a first amplification unit, used to amplify the points closest to the initial training samples in the labeled first patches in the first homogeneous region using the K-nearest neighbor algorithm as training samples; a second amplification unit, used to vote on the labels of the labeled second patches in the second patches to amplify the training samples using a majority voting strategy; a first evaluation unit, used to evaluate the purity of the first homogeneous region and the second homogeneous region to determine the feasibility of amplifying the training samples; and a second evaluation unit, used to evaluate the accuracy of the classification results based on the overall classification accuracy, average accuracy, and Kappa coefficient after classification is completed.

[0173] The embodiment of the present application obtains a first homogenous region of a larger scale and a second homogenous region of a smaller scale by using different segmentation methods by the image segmentation unit, thereby combining and expanding more training samples, and further purifying the impure homogenous region patches by using the mean shift algorithm, thereby improving the purity of the large-scale homogenous region, and the first expansion unit uses the K nearest neighbor algorithm to expand training samples of the first homogenous region of a larger scale, and only expands the K points closest to the labeled sample in the first patch, which can reduce the influence of the edge noise points far away from the labeled sample, thereby reducing the expansion of false samples and improving the sample accuracy; for the same labeled sample, different segmentation processes of different methods can obtain different homogenous regions containing the labeled sample, and combining these homogenous regions can obtain more expanded samples, and the small-scale homogenous region can expand more training samples on the basis of the large-scale homogenous region expansion samples, thereby further increasing the number of training samples, and the increase of the number of high-precision expanded training samples can improve the model generalization ability, thereby improving the classification accuracy, which has important significance for land use change monitoring and target recognition.

[0174] Embodiment seven

[0175] The present embodiment provides a computer device, please refer to Figure 11 , Figure 11 The basic structure block diagram of the computer device of the present embodiment is shown in the figure.

[0176] As shown in Figure 11 , the internal structure diagram of the computer device. The computer device includes a processor, a non-volatile storage medium, a memory and a network interface connected by a system bus. Among them, the non-volatile storage medium of the computer device stores an operating system, a database and computer readable instructions, the database can store control information sequence, the computer readable instructions are executed by the processor, which can make the processor realize the above-mentioned semi-supervised classification method. The processor of the computer device is used to provide computing and control ability to support the operation of the whole computer device. The memory of the computer device can store computer readable instructions, which can make the processor execute a semi-supervised classification method when the computer readable instructions are executed by the processor. The network interface of the computer device is used to connect and communicate with the terminal. Those skilled in the art can understand Figure 11 that the structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or less components than those shown in the figure, or combine certain components, or have different component arrangement.

[0177] The processor in the present embodiment is used to execute Figure 10The specific functions of the image segmentation unit, the first expansion unit, the second expansion unit, the first evaluation unit and the second evaluation unit are stored in the memory with program codes and various data required for executing the above modules. The network interface is used for data transmission between the user terminal or the server. The server can call the program codes and data of the server to execute the functions of all sub-modules.

[0178] The computer device obtains the first homogenous region of a larger scale and the second homogenous region of a smaller scale by using different segmentation methods through the image segmentation unit, so as to combine and expand more training samples, and the mean shift algorithm is used to further purify the impure homogenous region patches, thereby improving the purity of the large-scale homogenous region. Meanwhile, the first expansion unit uses the K nearest neighbor algorithm to expand training samples of the first homogenous region of a larger scale, and only the K points closest to the labeled sample in the first patch are expanded, so as to reduce the influence of the edge noise points far away from the labeled sample, thereby reducing the expansion of the error samples and improving the sample precision. For the same labeled sample, different segmentation processes of different methods can obtain different homogenous regions containing the labeled sample, and combining these homogenous regions can obtain more expanded samples. The small-scale homogenous region can expand more training samples on the basis of the large-scale homogenous region expansion samples, and further increase the number of training samples. The increase of the number of high-precision expansion training samples can improve the model generalization ability, thereby improving the classification precision, which has important significance for land use change monitoring and target recognition.

[0179] Correspondingly, the application also provides a storage medium storing computer readable instructions, which are executed by one or more processors to make the one or more processors execute the steps of the information encryption method of any one of the above embodiments.

[0180] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a computer readable storage medium, and the program can include the processes of the above-mentioned embodiments when executed. The storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0181] The above embodiments are only illustrative of the principles and effects of the application, and are not used to limit the application. Any person skilled in the art can modify or change the above embodiments without departing from the spirit and scope of the application. Therefore, all equivalent modifications or changes made by those skilled in the art without departing from the spirit and technical thought of the application should be covered by the claims of the application.

Claims

1. A semi-supervised classification method for hyperspectral images based on multi-scale homogeneous regions, characterized in that, include: The image to be processed is used to generate a first homogeneous region of a larger scale using a first segmentation method and a second homogeneous region of a smaller scale using a second segmentation method. The first homogeneous region includes a plurality of first patches and the second homogeneous region includes a plurality of second patches. The first segmentation method uses the JSEG segmentation method, and the second segmentation method uses the SLIC segmentation method. The process of generating a smaller-scale second homogeneous region using the second segmentation method specifically includes: selecting K points as cluster centers in the entire image and moving the cluster centers to the location of the minimum gradient distance within their neighborhood; performing cluster analysis on all pixels in the image, calculating the distance between each pixel and all cluster centers, and assigning the cluster center with the smallest distance to each pixel to that cluster; recalculating the average vector of all points in that cluster as the center for the next cluster; repeating the above process until convergence, and then outputting a superpixel. The purity of the first homogeneous region and the second homogeneous region is evaluated to determine the feasibility of amplifying the training samples; The K-nearest neighbor algorithm is used to amplify the points in the first patch of the first homogeneous region that are closest to the initial training samples as training samples. A majority voting strategy is employed to amplify the training samples by voting on the labels of the marked second patches. Specifically, this includes: when there is only one marked point in the second patch, the mode of the labels of the elements in the second patch is used as the label of that marked point, and all points in the second patch are assigned the same label as the mode of the labels; when there are multiple marked points of the same category in the second patch, the label of that category is used as the label of the second patch, and all points in the second patch are assigned the same label as the marked point; when there are marked points of different categories in the second patch, the labels of all marked points are voted on, and the mode of the labels of the marked points is selected as the label of the second patch, while keeping the labels of the marked points unchanged, and other elements in the second patch are assigned the same label as the label of the second patch. After classification is completed, based on the overall classification accuracy, average accuracy, and... The coefficients evaluate the accuracy of the classification results.

2. The semi-supervised classification method for hyperspectral images based on multi-scale homogeneous regions according to claim 1, characterized in that, The specific steps of generating a larger-scale first homogeneous region using the first segmentation method include: The top three principal components of the hyperspectral image are extracted as RGB images and then K-means clustering is performed. Calculate the J value by selecting an appropriate window size, and generate a multi-scale J image; Based on the multi-scale J image, points with local J values ​​less than a predetermined threshold are selected as seed points, and pixels in their neighborhoods that are similar to the seed points are merged together to finally obtain the JSEG segmentation map.

3. The semi-supervised classification method for hyperspectral images based on multi-scale homogeneous regions according to claim 1, characterized in that, After generating a large-scale first homogeneous region using the first segmentation method, the method further includes: purifying the impure first patch using a mean shift algorithm; During purification, the average variance is used as an indicator to measure the purity of the first patch. The specific formula for calculating the average variance is as follows: ， in, Represents the total number of bands in the hyperspectral image. This represents the number of pixels in the first patch. Representing the first plaque The feature vector of each element The characteristic center of the first patch.

4. The semi-supervised classification method for hyperspectral images based on multi-scale homogeneous regions according to claim 1, characterized in that, The step of using the K-nearest neighbor algorithm to amplify the points in the first patch of the first homogeneous region that are closest to the initial training samples as training samples specifically includes: According to the formula Calculate the first The eigenvector of the unmarked point and the... The distance between the feature vectors of the marked points, where the number of elements in the second patch is . The number of marked points is The number of feature vectors for each element is , No. The feature vector of the unlabeled points is , No. The feature vector of each labeled point is ; For the second patch From the unmarked point to the... Distance between marked points Arranged in descending order, among which... ; Select the K nearest unmarked points to the marked point, and assign the K unmarked points the same label as the marked point; Until the traversal is complete After identifying all labeled points, the frequency of labels appearing in unlabeled points is counted. If an unlabeled point appears multiple times, unlabeled points with different labels and duplicate unlabeled points with the same label are removed.

5. The semi-supervised classification method for hyperspectral images based on multi-scale homogeneous regions according to claim 1, characterized in that, The specific steps for evaluating the purity of the first homogeneous region and the second homogeneous region to determine the feasibility of amplifying the training samples are as follows: Using formula Calculate purity, where the number of patches segmented from the hyperspectral image is . , Indicates the first The mode of the class of labeled samples within each patch, Indicates the first The total number of labeled samples in each patch.

6. The semi-supervised classification method for hyperspectral images based on multi-scale homogeneous regions according to claim 1, characterized in that, The basis is the overall classification accuracy, average accuracy, and The coefficients used to evaluate the accuracy of classification results specifically include: With formula Calculate the overall classification accuracy; With formula Calculate the average accuracy; With formula calculate coefficient; Where c is the number of categories in the hyperspectral remote sensing image, the confusion matrix is ​​c rows and c columns, the total number of test samples in the remote sensing image is n, and the confusion matrix is ​​M. This represents the number of samples that are misclassified from class j to class i.

7. A semi-supervised classification system for hyperspectral images based on multi-scale homogeneous regions, characterized in that, include: An image segmentation unit is used to generate a larger-scale first homogeneous region from the image to be processed using a first segmentation method, and to generate a smaller-scale second homogeneous region using a second segmentation method. The first homogeneous region includes several first patches, and the second homogeneous region includes several second patches. The first segmentation method uses the JSEG segmentation method, and the second segmentation method uses the SLIC segmentation method. Specifically, generating the smaller-scale second homogeneous region using the second segmentation method includes: selecting K points as cluster centers in the entire image and moving the cluster centers to the location of the minimum gradient distance in their neighborhood; performing cluster analysis on all pixels in the image, calculating the distance between each pixel and all cluster centers, and assigning the cluster center with the smallest distance to each pixel to that cluster; recalculating the average vector of all points in the cluster as the center for the next cluster; repeating the above process until convergence, and then outputting superpixels. The first amplification unit is used to amplify the points in the first patch of the first homogeneous region that are closest to the initial training samples using the K-nearest neighbor algorithm as training samples. The second amplification unit is used to amplify the training samples by voting on the labels of the marked second patches using a majority voting strategy. Specifically, it includes: when there is only one marked point in the second patch, the mode of the labels of the elements in the second patch is used as the label of the marked point, and all points in the second patch are assigned the same label as the mode of the labels; when there are multiple marked points of the same category in the second patch, the label of that category is used as the label of the second patch, and all points in the second patch are assigned the same label as the marked point; when there are marked points of different categories in the second patch, the labels of all marked points are voted on, the mode of the labels of the marked points is selected as the label of the second patch, while keeping the labels of the marked points unchanged, and the other elements in the second patch are assigned the same label as the label of the second patch. The first evaluation unit is used to evaluate the purity of the first homogeneous region and the second homogeneous region to determine the feasibility of amplifying the training samples. The second evaluation unit is used to assess the overall classification accuracy, average accuracy, and... after classification is completed. The coefficients evaluate the accuracy of the classification results.

Images