Hyperspectral Image Classification Method Integrating Superpixel and Extended Multi-Attribute Contour Features

Abstract

The present invention relates to the technical field of hyperspectral remote sensing image, and specifically relates to a hyperspectral image classification method that fuses superpixel and extended multi-attribute profile features, obtaining features from the superpixel level and the pixel level respectively. The method for obtaining superpixel-level features is to segment the hyperspectral image into individual homogeneous regions, then perform principal component analysis on each homogeneous region, and then reconstruct the image. Pixel-level feature extraction uses the extended multi-attribute profile method (EMAP) to extract texture features from the first three principal components of the HSI, and connects the two features along the spectral dimension to achieve information complementarity. Finally, it is input into a support vector machine for classification. The present invention extracts the low-dimensional inherent features of the HSI while reducing information redundancy, and at the same time extracts the texture information of the HSI using the EMAP method, solving the problem of insufficient information extraction of superpixel features at a single scale.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral remote sensing images, and particularly to a hyperspectral image classification method that fuses superpixel and extended multi-attribute contour features. Background Art

[0002] As a type of remote sensing data, hyperspectral images are in the form of three-dimensional cubes, consisting of dozens to hundreds of continuous narrow-band frequency bands, covering the visible and infrared light ranges, and having the characteristic of combining spectrum and image. Therefore, they have numerous applications in fields such as mineral exploration, hydrological monitoring, and environmental protection. Hyperspectral images have high resolution and rich information, so the processing of hyperspectral data has always been a hot topic studied by many scholars.

[0003] Due to the particularity of its data form, it contains rich spatial and spectral information. At the same time, the high correlation between the information of its bands also brings image data redundancy, resulting in the frequent occurrence of the Hughes phenomenon in the process of hyperspectral data processing, which is particularly obvious in the case of high-dimensional data and insufficient labeled samples. At the same time, due to the high-dimensionality of its data and large data volume, the complexity of data calculation is very high. In addition, the noise in hyperspectral images will also reduce the accuracy of the classification model. Therefore, in the process of hyperspectral data processing, it is necessary to extract its main information. In current hyperspectral classification technologies, the classification results of simply extracting spectral information for processing are not ideal.

[0004] Recently, due to its high efficiency, the method based on superpixel segmentation has been widely applied to the field of hyperspectral image processing. Based on the assumption that adjacent pixels have similar spatial structures, it uses superpixels to divide the image into multiple non-overlapping homogeneous regions, and each homogeneous region is regarded as similar-structured ground objects of the same type. In most literatures, many scholars regard it as a preprocessing process. Therefore, the superpixel segmentation method can be regarded as characterizing the HSI at the superpixel level, improving the recognition of the HSI.

[0005] Although the superpixel method has been successfully applied to the feature extraction of hyperspectral images, due to noise and lack of spatial information, the accuracy of the model will still be affected. Therefore, in order to remove the noise in the image, an image filtering model is usually introduced into the superpixel method, which can effectively reduce the noise and smooth it globally. However, the definitions of filters are all defined globally. They usually use various windows to filter the image and may fail when dealing with complex boundary regions. The processing of ground object edge information overly relies on the parameters of superpixels. Too low parameters will lead to incomplete decomposition, and too high parameters will cause over-decomposition of the image. At the same time, the classification effect is also affected by the different spatial sizes and distributions of ground objects. Summary of the Invention

[0006] The object of the present invention is to provide a hyperspectral image classification method integrating superpixel and extended multi-attribute profile features, aiming to extract two different features in two ways to achieve the purpose of complementary information, fully extract the spatial features of HSI, and solve the problem of over-smoothing of complex boundary regions in the hyperspectral image classification process.

[0007] To achieve the above object, the present invention provides a hyperspectral image classification method integrating superpixel and extended multi-attribute profile features, including the following steps:

[0008] Step 1: Perform principal component analysis on the hyperspectral image to obtain the first principal component image and the first three principal components respectively.

[0009] Step 2: Set the number of superpixels for the first principal component image, perform entropy rate superpixel segmentation, and split each homogeneous region after segmentation into separate structures.

[0010] Step 3: Perform principal component analysis on each split structure, and then recombine and restore these structures to form a new dimension-reduced HSI.

[0011] Step 4: Construct EMAP features based on the first three principal components.

[0012] Step 5: Perform principal component analysis on the EMAP features and fuse the dimension-reduced HSI obtained in Step 3 to obtain fused features.

[0013] Step 6: Perform recursive filtering on the fused features.

[0014] Step 7: Divide the training set and the test set, input them into a support vector machine, and obtain classification labels.

[0015] Among them, in the process of performing principal component analysis on the hyperspectral image, it is first necessary to reduce the number of spectral dimension bands of the image.

[0016] Among them, the number of superpixels in the entropy rate superpixel segmentation is set manually, and the number of homogeneous regions is the same as the number of superpixels.

[0017] Among them, the new dimension-reduced HSI is an image I with superpixel features. sp In the process of extracting superpixel features, first, for the image I sp perform principal component analysis on the homogeneous regions in it to make the dimension of each homogeneous region reach the preset dimension, and then recombine it into the original size.

[0018] Among them, in the process of constructing EMAP features based on the first three principal components, an extended morphological profile method is used to extract pixel features, which are constructed according to different attribute rules for the first three principal components.

[0019] Among them, the fused feature is generated by connecting the EMAP feature and the new dimension-reduced HSI along the spectral dimension to form a new feature set, and recursive filtering is used to remove redundant information.

[0020] Among them, the process of dividing the training set and the test set, inputting them into the support vector machine, and obtaining the classification labels is specifically as follows: First, divide the original image into the training set and the test set, then input them into the support vector machine for training to obtain a structure, and finally use the structure and the test samples as inputs for prediction to obtain the predicted classification labels.

[0021] The present invention provides a hyperspectral image classification method that fuses superpixel and extended multi-attribute profile features, obtaining features from two levels: the superpixel level and the pixel level. The method for obtaining the superpixel-level features is to segment the hyperspectral image into individual homogeneous regions, then perform principal component analysis on each homogeneous region, and then recombine the image. At the pixel level, the extended multi-attribute profile method (EMAP) is used to extract texture features from the first three principal components of the HSI. Connecting the two types of features along the spectral dimension achieves information complementarity, improving the current situation of insufficient extraction of hyperspectral images using a single HSI representation. Finally, it is input into the support vector machine for classification. This method applies the principal component analysis method to each homogeneous region, reducing information redundancy while extracting the low-dimensional intrinsic features of the HSI. At the same time, the EMAP method is used to extract the texture information of the HSI, supplementing the problem of insufficient information extraction of superpixel features at a single scale. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for description in the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0023] Figure 1 It is a schematic flowchart of a hyperspectral image classification method that fuses superpixel and extended multi-attribute profile features of the present invention.

[0024] Figure 2 It is a schematic diagram of the method principle of a specific embodiment of the present invention.

[0025] Figure 3 It is a schematic diagram of the ground objects of the experimental image in a specific embodiment of the present invention.

[0026] Figure 4 It is a true ground object classification map of the experimental image in a specific embodiment of the present invention.

[0027] Figure 5 It is a schematic diagram of superpixel-level feature extraction in a specific embodiment of the present invention.

[0028] Figure 6 It is a schematic diagram of pixel-level feature extraction in a specific embodiment of the present invention.

[0029] Figure 7 It is an example diagram of the classification effect of the control experiment in a specific embodiment of the present invention.

[0030] Figure 8 It is an example diagram of the classification effect of the present invention in a specific embodiment of the present invention. Detailed implementation manners

[0031] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to explain the present invention, but should not be construed as limiting the present invention.

[0032] Please refer to Figure 1 , the present invention provides a hyperspectral image classification method based on extracting complementary information of superpixel-level features and pixel-level features, including the following steps:

[0033] S1: Perform principal component analysis on the hyperspectral image to respectively obtain the first principal component image and the first three principal components;

[0034] S2: Set the number of superpixels for the first principal component image, perform entropy rate superpixel segmentation, and split each homogeneous region after segmentation into separate structures;

[0035] S3: Perform principal component analysis on each split structure, and then recombine and restore these structures to form a new dimensionality-reduced HSI;

[0036] S4: Construct EMAP features based on the first three principal components;

[0037] S5: Perform principal component analysis on the EMAP features, and fuse the dimensionality-reduced HSI obtained in step S3 to obtain fused features;

[0038] S6: Perform recursive filtering on the fused features;

[0039] S7: Divide the training set and the test set, input into the support vector machine, and obtain classification labels.

[0040] Furthermore, the present invention also proposes specific embodiments and further explains in combination with the implementation steps:

[0041] AsFigure 2 and Figure 3 As shown, a hyperspectral image H is selected M×N×B and its accompanying corresponding ground truth G M×N , and in this example, the publicly available University of Pavia dataset is selected, with a size of 610×340×103, that is, M = 610, N = 340, B = 103; M and N are the two-dimensional rows and columns, and B represents the number of spectral dimension bands of the image. The ground truth classification map is as Figure 4 shown.

[0042] First, perform principal component analysis on the original hyperspectral image H M×N×B to reduce the dimension to where P < B.

[0043] I = {P c1 , P c2 ,..., P cn} = PCA{H}

[0044] Take the first three principal components of the image, which are and set the number of superpixels to S. Perform entropy rate superpixel segmentation (ERS) on I3 to obtain the segmented image I sp

[0045] I SP = [I1, I2, I3,..., I S

[0046] At this time, homogeneous regions have been divided within the hyperspectral image, and there are similar structures within the homogeneous regions.

[0047]

[0048] Perform principal component analysis on each homogeneous region separately, and finally recombine these individual structures to form an image I with superpixel features sp . Please refer to Figure 5 .

[0049] I sp contains S homogeneous regions. Further split the homogeneous regions. At this time, each homogeneous region still contains d dimensions. Perform principal component analysis on each homogeneous region to make the dimension of each homogeneous region reach the preset dimension, and then recombine it to the original size, which ensures the separability of the object edges.

[0050] Please refer to Figure 6 , and pixel-level features are extracted using the extended morphological profile method (EMAP). The steps are to take the first three principal components of the original image, which are ​Different attribute criteria are adopted, including the area criterion "a" and the standard deviation "s".

[0051] I EMAP = EMAP(I3)

[0052] Connect I EMAP and I SP along the spectral dimension and perform recursive filtering.

[0053]

[0054] Divide the original image into training set and test set as X train , Y train , X test . Y test , input them into the support vector machine for training to obtain a struct-type structure model with parameters, and use the model and test samples as inputs for prediction to obtain the predicted classification labels.

[0055] label = SVM(O) RBF

[0056] The present invention is experimentally compared with traditional methods such as SVM, PCA, and LFDA on the University of Pavia dataset. Compared with other traditional methods, the overall accuracy of this method is improved by 2.46% and the kappa coefficient is improved by 2.59%. The specific data are shown in Table 1. The effects of other control classification methods are as Figure 7 shown. It can be seen that the misclassification phenomenon is relatively serious in other classification methods, and there are noise points at the edges of ground objects that are not cleared. In Figure 8 , the misclassification and wrong classification phenomena are greatly reduced, which comprehensively shows that this method has achieved good results in actual tests and has obvious advantages over other methods.

[0057] Table 1 Comparison of various accuracies of comparison methods

[0058]

[0059]

[0060] The above-disclosed is only a preferred embodiment of the present invention. Of course, the scope of the rights of the present invention cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of implementing the above embodiments, and the equivalent changes made according to the claims of the present invention still fall within the scope covered by the invention.

Claims

1. A hyperspectral image classification method integrating superpixel and extended multi-attribute contour features, characterized in that, Including the following steps: Step 1: Perform principal component analysis on the hyperspectral image to obtain the first principal component image and the first three principal components respectively; Step 2: Set the number of superpixels for the first principal component image, perform entropy rate superpixel segmentation, and split each homogeneous region after segmentation into separate structures; Step 3: Perform principal component analysis on each split structure, and then recombine and restore these structures to form a new dimension-reduced HSI; Step 4: Construct EMAP features based on the first three principal components; Step 5: Perform principal component analysis on the EMAP features and fuse the dimension-reduced HSI obtained in Step 3 to obtain fused features; Step 6: Perform recursive filtering on the fused features; Step 7: Divide the training set and the test set, input them into the support vector machine, and obtain classification labels.

2. The hyperspectral image classification method integrating superpixels and extended multi-attribute profile features according to claim 1, characterized in that In the process of performing principal component analysis on the hyperspectral image, it is first necessary to reduce the number of spectral dimension bands of the image.

3. The hyperspectral image classification method integrating superpixels and extended multi-attribute profile features according to claim 1, characterized in that The number of superpixels in the entropy rate superpixel segmentation is set manually, and the number of homogeneous regions is the same as the number of superpixels.

4. The hyperspectral image classification method integrating superpixels and extended multi-attribute profile features according to claim 1, characterized in that The new dimensionality-reduced HSI is the image I with superpixel features sp , in the process of extracting superpixel features, first perform principal component analysis on the homogeneous regions in the image I sp , so that the dimension of each homogeneous region reaches the preset dimension, and then reorganize it into the original size.

5. The hyperspectral image classification method integrating superpixels and extended multi-attribute profile features according to claim 1, characterized in that In the process of constructing EMAP features based on the first three principal components, an extended morphological profile method is used to extract pixel features, which are constructed according to different attribute rules for the first three principal components.

6. The hyperspectral image classification method integrating superpixels and extended multi-attribute profile features according to claim 1, characterized in that The fused features are generated by connecting the EMAP features and the new dimension-reduced HSI along the spectral dimension to form a new feature set, and recursive filtering is used to remove redundant information.

7. The hyperspectral image classification method integrating superpixels and extended multi-attribute profile features according to claim 1, characterized in that The process of dividing the training set and the test set, inputting them into the support vector machine, and obtaining classification labels is specifically as follows: first divide the original image into the training set and the test set, then input them into the support vector machine for training to obtain a structure, and finally use the structure and the test samples as inputs for prediction to obtain the predicted classification labels.

Images