Posterior probability hyperspectral image classification method based on multiscale entropy rate superpixels

Abstract

The present application relates to hyperspectral remote sensing image processing technical field, specifically to posterior probability hyperspectral image classification method based on multiscale entropy rate superpixel, including preprocessing hyperspectral image data, obtaining processing image, initial classification of hyperspectral image, obtaining multiple initial classification label spectral information, principal component analysis on preprocessing image, obtaining first three main components, image processing on three main components, obtaining superpixel image under different scales, fusion processing on superpixel image, obtaining dimensionality reduction hyperspectral image, domain transformation interpolation convolution filtering processing on multiple initial classification labels and dimensionality reduction hyperspectral image, obtaining rearrangement classification label, classification on hyperspectral image, then superpixel segmentation operation, feature extraction and correction, obtaining final classification label, thereby improving the problem of insufficient spatial information extraction and neglecting superpixel edge information of existing spectral image classification technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hyperspectral remote sensing image processing, and particularly relates to a posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels. BACKGROUND

[0002] With the continuous development of remote sensing technology, hyperspectral sensing technology, as an organic integration of target spectrum acquisition and target space imaging, describes the spectral characteristics of the spatial characteristics and ground object information of the earth's surface. The imaging high resolution characteristics and multi-dimensional imaging characteristics of hyperspectral images (HSI) can express the target in the form of an image, and can also perform spectral detection with the aid of fine electromagnetic spectrum, thereby providing technical support for the automatic identification and fine classification of ground objects.

[0003] However, while the hyperspectral image is widely applied, the processing work thereof still faces great challenges. In particular, the imaging characteristics of remote sensing still make this type of data have specific defects, which makes it difficult to extract features and classify technologies. The difficulties are as follows: 1. The continuous spectral band imaging, while improving the spectral resolution, also increases the data volume, causing information redundancy and increasing the difficulty of data processing; 2. The high dimensionality of data makes the classification effect poor for small sample data. Traditional machine learning methods (support vector machine, random forest, K nearest neighbor) only use spectral information for direct classification, which cannot meet the technical requirements of fine classification. At present, the research on the classification processing technology of hyperspectral images still needs to be further improved.

[0004] Most of the initial methods for classifying hyperspectral images only use single spectral information or spatial information. For example, the pixel-level classifier-support vector machine (SVM) directly processes the hyperspectral image, and the effect can only reach an accuracy of 80%-86%. Moreover, the classification result is accompanied by a large amount of noise and misclassification, misclassification, etc., which cannot achieve satisfactory results. The extended multi-attribute morphological profile (EMAPs) performs morphological operation on the hyperspectral data, extracts the spatial information of HSI under different attributes, and combines the two to achieve better classification results than the classification methods that only consider spectral information or spatial information.

[0005] Most of the literature and research results show that the method based on superpixel segmentation is better in image space information extraction, the superpixel segmentation algorithm is based on the assumption that adjacent pixels have similar structures, can divide the image into a single homogeneous region according to the similarity of similar texture, color, brightness and other information of the image, and the pixels are grouped according to the similar features between the pixels, the whole image is finely divided into a plurality of non-overlapping homogeneous sub-regions, the structure properties of the object in the same superpixel structure have high similarity, which can reduce the error of judging the surrounding pixels, the method based on superpixel better utilizes the structure space information between adjacent pixels, provides stronger and more identifiable features, and has higher classification accuracy, however, the method based on superpixel also has certain defects, because the spatial information in each superpixel is represented by the mean value of the pixels, the spatial information in such superpixel can be uniformly represented, but the information of the edge of the superpixel is lost, and under a single segmentation scale, the spatial information of the image cannot be fully extracted, thereby reducing the classification accuracy. SUMMARY

[0006] The present application aims to provide a posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels, aiming at improving the problem that the existing spectral image classification technology ignores the edge information of superpixels.

[0007] To achieve the above object, the present application provides a posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels, comprising the following steps:

[0008] Pretreating the hyperspectral image data to obtain a processed image;

[0009] Initially classifying the hyperspectral image to obtain an initial classification label containing spectral information;

[0010] Performing principal component analysis on the pretreated image to obtain the first three principal components;

[0011] Performing image processing on the first three principal components to obtain superpixel images under different scales;

[0012] Fusing the superpixel images to obtain a reduced dimension hyperspectral image;

[0013] Performing domain transformation interpolation convolution filtering processing on the plurality of initial classification labels and the reduced dimension hyperspectral image to obtain a rearranged classification label.

[0014] The specific way of initially classifying the hyperspectral image to obtain a plurality of initial classification labels containing spectral information is:

[0015] Dividing the hyperspectral image into a training set and a test set to obtain divided data;

[0016] Classify the divided data by using a support vector machine to obtain a plurality of initial classification labels.

[0017] The specific manner of image processing on the three main components to obtain the superpixel image is:

[0018] The specific manner of image processing on the three main components to obtain the superpixel image is:

[0019] Different numbers of superpixels are set for the three main component images to obtain setting data.

[0020] Based on the setting data, entropy rate superpixel segmentation is performed to obtain superpixel images at different scales.

[0021] The specific manner of fusion processing on the superpixel image to obtain the reduced hyperspectral image is:

[0022] Decision fusion is performed on the superpixel segmentation image to obtain spatial information.

[0023] Based on the spatial information, principal component analysis is performed to form a reduced hyperspectral image.

[0024] The specific manner of domain transformation interpolation convolution filtering processing on the plurality of initial classification labels and the reduced hyperspectral image to obtain the rearranged classification label is:

[0025] The plurality of initial classification labels are taken as the correction target, and the reduced hyperspectral image is taken as the guide image, and domain transformation interpolation convolution filtering is performed to obtain the rearranged classification label.

[0026] After the plurality of initial classification labels and the reduced hyperspectral image are processed in step to obtain the rearranged classification label, the method further includes:

[0027] The accuracy of the rearranged classification label is evaluated by using average accuracy, overall accuracy and Kappa coefficient to obtain an evaluation result.

[0028] The posterior probability hyperspectral image classification method based on the multiscale entropy rate superpixel of the application carries out preprocessing on the hyperspectral image data, obtains a pretreated image, carries out initial classification on the hyperspectral image, obtains a plurality of initial classification labels, carries out principal component analysis on the pretreated image, obtains the first three main components, carries out image processing on the first three main components, obtains a superpixel image, carries out principal component analysis processing on the superpixel image, obtains a reduced dimension hyperspectral image, firstly carries out initial probability classification on the hyperspectral image through a multi-classification support vector machine, then carries out superpixel segmentation operation of different scales on the first three main components of the image, and carries out fusion to obtain a final reduced dimension image, inputs the reduced dimension image and the initial probability distribution map into a domain transformation interpolation convolution filter to carry out filtering to modify the initial probability, and obtains the final classification label after modification, carries out experiments on the IndianPines data set respectively through the proposed method and the traditional RBF-SVM method, the superBF method and the RF method, compared with the traditional other methods, the overall precision of the proposed method is increased by 2.46%, and the Kappa coefficient is increased by 2.59%, which shows that the proposed method can obtain good effect in actual test, has obvious advantages compared with other methods, and thus solves the problem that the existing spectral image classification technology ignores the superpixel edge information. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, hereinafter, the drawings needed to be used in the embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0030] Figure 1 is a flowchart of the posterior probability hyperspectral image classification method based on the multiscale entropy rate superpixel provided by the application.

[0031] Figure 2 is a classification framework schematic diagram.

[0032] Figure 3 is a first principal component image obtained by principal component analysis on an original image.

[0033] Figure 4 is a real ground object classification map of an experimental image.

[0034] Figure 5 is a superpixel segmentation map.

[0035] Figure 6 is a superpixel multiscale segmentation schematic diagram.

[0036] Figure 7 Comparison diagram before filtering.

[0037] Figure 8 Image without superpixel segmentation.

[0038] Figure 9 It is a superpixel segmentation effect image.

[0039] Figure 10 It is a comparison of the precision of each type of experiment. Detailed Implementation

[0040] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein 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 intended to explain the present invention, and should not be construed as limiting the present invention.

[0041] Please see Figures 1 to 10 This invention provides a posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels, comprising the following steps:

[0042] S1 preprocesses the hyperspectral image data to obtain the processed image;

[0043] Specifically, select a hyperspectral image and its accompanying ground truth map like Figure 2 The hyperspectral image data is normalized and preprocessed to obtain the preprocessed image, where M and N are the rows and columns of the image, and L represents the number of bands in the spectral dimension of the image.

[0044] Within the dataset, x i,j For any point, x max x min The maximum and minimum values ​​are respectively the values ​​of x in the dataset. i,j The normalization method is as follows:

[0045]

[0046] S2 performs initial classification on the hyperspectral image to obtain multiple initial classification labels (SVMs) containing spectral information. map 。;

[0047] S21 divides the hyperspectral image into a training set and a test set to obtain partitioned data;

[0048] Specifically, a test set and a training set are divided for the hyperspectral image, and a same proportion of test samples is obtained from each type of surface feature, and the proportion can be selected as 1%, 3%, 5%, 10%, and 30%. Half of the total number of samples with a smaller number is taken to obtain the divided data.

[0049] S22 classifies the divided data by using a support vector machine to obtain a plurality of initial classification labels.

[0050] Specifically, the hyperspectral image is initially classified by using a multi-classification support vector machine to obtain a plurality of initial classification labels SVM map .

[0051] S3 performs principal component analysis on the preprocessed image to obtain the first three principal components

[0052] Specifically, principal component analysis is performed on the normalized hyperspectral image to obtain three principal components that are not correlated and have the largest contribution rate.

[0053] S4 performs image processing on the first three principal components to obtain superpixel images at different scales.

[0054] S41 sets different numbers of superpixels for the first three principal component images to obtain setting data.

[0055] Specifically, for the three principal components, and the number of superpixels is set, and the optional value is S=[50, 100, 150, 200, 250, 300), and entropy rate superpixel segmentation (ERS) is performed on Y1 to obtain

[0056] Y=[Y1,Y2,Y3,...,Y S ]

[0057] S42 performs entropy rate superpixel segmentation based on the setting data to obtain superpixel images at different scales.

[0058] Specifically, Figure 6 As shown in the figure, S images are obtained, each image contains different superpixels, and a similar structure of homogeneous regions is constructed

[0059]

[0060] S5 performs fusion processing on the superpixel images to obtain a reduced hyperspectral image.

[0061] S51 performs decision fusion on the superpixel segmentation images to obtain spatial information.

[0062] Specifically, decision fusion is performed on the obtained multi-scale superpixel segmentation images to obtain the spatial information of the images.

[0063] S52 performs principal component analysis based on the spatial information to generate a dimension-reduced hyperspectral image.

[0064] Specifically, principal component analysis was performed, and the fusion strategy using the averaging fusion method was as follows:

[0065]

[0066] Ultimately, a dimensionality-reduced HSI is formed.

[0067] S6 applies the initial classification label SVM map The reduced-dimensional hyperspectral image is then subjected to domain transformation, interpolation, convolution, and filtering to obtain rearranged classification labels.

[0068] Specifically, domain transform convolution filtering is performed on the multiple initial classification labels and the reduced-dimensional hyperspectral image.

[0069] Specifically, regarding the original image Divide the training set and the test set into X train ,Y train ,X test ,Y test The input is fed into a support vector machine for training, resulting in a parameterized struct type structure model. The model and test samples are then used as input for prediction, yielding the predicted land cover classification map label SVM. map Finally, SVM map As input to the domain transform interpolation convolution filter, the dimension-reduced HSI is used as a reference image to determine the edges of ground features, correct the initial labels, and finally obtain the final classification labels.

[0070]

[0071] S7 uses average accuracy, overall accuracy, and Kappa coefficient to evaluate the accuracy of the rearranged classification labels and obtain the evaluation results.

[0072] Specifically, the accuracy is evaluated using average accuracy, overall accuracy, and the Kappa coefficient to obtain the evaluation results.

[0073] The above-disclosed embodiments are merely preferred embodiments of the posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels of the present invention. Of course, they should not be construed as limiting the scope of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention still fall within the scope of the invention.

Claims

1. A posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels, characterized in that, Includes the following steps; The hyperspectral image data is preprocessed to obtain the processed image; The hyperspectral image is initially classified to obtain multiple initial classification labels containing spectral information; Principal component analysis was performed on the preprocessed image to obtain the first three principal components; Image processing is performed on the first three main components to obtain superpixel images at different scales, specifically as follows: Different numbers of superpixels were set for the first three principal component images to obtain the setting data; Based on the set data, entropy rate superpixel segmentation is performed to obtain superpixel images at different scales; The superpixel image is fused to obtain a dimension-reduced hyperspectral image, specifically as follows: Decision fusion is performed on the superpixel segmented image to obtain spatial information; Principal component analysis is performed based on the spatial information to generate a dimension-reduced hyperspectral image; The multiple initial classification labels and the reduced-dimensional specular image are subjected to domain transformation, interpolation, convolution, and filtering to obtain the corrected rearranged classification labels.

2. The posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels as described in claim 1, characterized in that, The specific method for performing initial classification on the hyperspectral image to obtain multiple initial classification labels containing spectral information is as follows: The hyperspectral image is divided into a training set and a test set to obtain partitioned data; The partitioned data is classified using a support vector machine to obtain multiple initial classification labels.

3. The posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels as described in claim 1, characterized in that, The specific method for performing domain transform interpolation convolution filtering on the multiple initial classification labels and the reduced-dimensional specular image to obtain rearranged classification labels is as follows: Using the multiple initial classification labels as correction targets and the reduced-dimensional hyperspectral image as a guide image, domain transformation, interpolation, convolution, and filtering are performed to obtain rearranged classification labels.

4. The posterior probability hyperspectral image classification method based on multi-scale entropy rate superpixels as described in claim 1, characterized in that, After processing the multiple initial classification labels and the reduced-dimensionality specular image to obtain rearranged classification labels, the method further includes: The accuracy of the rearranged classification labels was evaluated using average accuracy, overall accuracy, and Kappa coefficient, and the evaluation results were obtained.

Images