CN112818831B - Hyperspectral image classification algorithm based on band clustering and improved domain transformation recursive filtering - Google Patents

Abstract

The invention aims to provide a hyperspectral image classification algorithm based on band clustering and improved recursive filtering, which is characterized in that a band clustering algorithm based on relative entropy is utilized to iteratively find out central bands in each band subset, the information of original spectral bands is reserved, compared with an original domain transformation recursive filtering algorithm, the hyperspectral image classification algorithm based on the band clustering and improved recursive filtering algorithm is used for performing Gaussian filtering on a central band set obtained by clustering as a guide image of domain transformation recursive filtering, the improved domain transformation recursive filtering algorithm is executed on the central bands of each set, and finally, the space-spectrum joint information of the hyperspectral image is fully acquired by obtaining a characteristic image set of all the central bands, so that the subsequent classification accuracy is improved.

Description

Hyperspectral image classification algorithm based on wave band clustering and improved domain transformation recursive filtering

Technical Field

The invention belongs to the technical field of image processing, and particularly relates to a hyperspectral image classification algorithm based on band clustering and improved domain transform recursive filtering.

Background

The hyperspectral remote sensing refers to a technology for acquiring relevant data of an object by utilizing a plurality of narrow electromagnetic wave bands, the technology detects and identifies a target through a plurality of spaces such as a spectrum space and a feature space, in recent years, a hyperspectral image is widely applied to the multidisciplinary crossing fields such as aerospace, agricultural science, geographic monitoring and environmental protection, and the like, wherein the improvement of the classification precision of the hyperspectral image is one of the current research hotspots. However, the existing hyperspectral image feature identification and classification algorithm can not meet the rapid development and demand of the hyperspectral remote sensing technology. First, the hyperspectral image is inevitably disturbed by noise. Secondly, the characteristics of multiple wave bands, over-high dimensionality and the like of the hyperspectral images bring great difficulty to the identification and classification of the hyperspectral images. Therefore, how to mine low-redundancy effective information from the high-dimensional hyperspectral image and improve the spatial quality and resolution of the image is a key for subsequent classification of hyperspectral data.

At present, aiming at the problem of overhigh dimension of a hyperspectral image, the existing dimension reduction technology is divided into two major categories: feature extraction and wave band selection. The related technologies of feature extraction include principal component analysis, local linear embedding, linear discrimination information and the like, and although the feature extraction technology can realize higher classification precision, the low-dimensional expression of feature extraction has no physical significance and is difficult to explain. The band selection technique generates a corresponding low-dimensional expression by selecting the most important spectral band, can retain the original information of the spectral band, but has limited improvement on classification accuracy. Therefore, how to improve the classification accuracy of the hyperspectral images while maintaining the original information of the hyperspectral images is a problem that needs to be solved urgently by technical personnel in the field.

Disclosure of Invention

The invention aims to provide a hyperspectral image classification algorithm based on wave band clustering and improved recursive filtering, which is used for improving the subsequent classification precision.

The technical scheme for solving the technical problems of the invention is as follows: a hyperspectral image classification algorithm based on band clustering and improved domain transform recursive filtering is characterized by comprising the following steps:

s1: inputting an original hyperspectral image X ^Q×D Q is the number of pixels on each wave band, D is the number of wave bands, and all D wave bands of the hyperspectral image are divided into K hyperspectral sub-wave band sets; computing a kth subband set P ^{k∈(1,...,K)} The calculation method comprises the following steps:

wherein X ═ X (X) ₁ ,...,X _D )∈R ^Q×D Representing the original hyperspectral image X with D bands in number and Q pixels per band,

represents the maximum integer step size not exceeding D/K;

s2: for each subset P in the band set obtained at S1 ^{k∈(1,...,K)} Determining clustering center band set M _i ，i∈(1,2,...,K)；

S3: for clustering center wave band set M _i I ∈ (1, 2.. K) is subjected to Gaussian filtering, and the specific formula is

In the formula, M _i For the input image, a and b are the index coordinates of the input image pixels,

is the regularization parameter, N (a) is the set of neighborhood pixels for a, σ _s Is the spatial standard deviation, G _i Is the result of gaussian filtering;

s4: initializing the guide image, i.e. setting the Gaussian filtered image as the initial guide image, F _i [0]＝G _i ；

S5: the central wave band set M obtained by clustering _i i ∈ (1, 2.. K.) as the input image set and the guide image F _i [0]Performing domain transform recursive filtering; the principle of recursive filtering of the domain transform is to apply a given transform domain D _t Will input the image M _i Conversion to D _t In (1), the formula of the recursive filtering is

F _i [N]＝(1-d ^c )M _i [N]+d ^c F _i [N-1]

In the formula, F _i [N]For recursively filtered feature images, d ∈ [0,1 ]]Is a feedback coefficient, and

σ _s is the spatial standard deviation, c ═ ct (x) _n )-ct(x _n-1 ) To transform domain D _t Middle adjacent sample point x _n And x _n-1 The distance between the two, the formula of the transformation function ct (u) is

In the formula, σ _s Is the spatial standard deviation, σ _r Is the spectral standard deviation. M' _i (x) For an input image M _i First derivative function at pixel x, d as c increases ^c Approaches to 0, thereby playing the role of edge protection; s6: filtered image set F for the last time of recursive filtering of a domain transform _i [N]And i belongs to (1, 2.. multidot.K) and a support vector machine is used for classification, so that the classification precision is obtained.

The step S2 specifically includes:

s2.1: initializing cluster centers to

k is an integer and is the initial central band of each cluster group;

s2.2: setting any two wave bands X in high spectral data _p And X _q Where p ≠ q, then band X _p And X _q Is p, q ∈ (X) ₁ ,...,X _D )

Wherein m is _p (x)、m _q (x) Representing the probability corresponding to each pixel x in any two wave bands, wherein the larger the relative entropy between the two wave bands is, the larger the difference of the probability of random distribution corresponding to the pixels of the two wave bands is, and the lower the similarity between the wave bands is; otherwise, the similarity between the wave bands is higher;

s2.3: updating the cluster center based on the relative entropy, wherein the measurement formula for updating the cluster center is

Wherein, X _p ∈p ^k ，

Representing a subset of bands p ^k In (C) X _p Mean square of the sum of relative entropies between a band and other bands, X _q Representing a subset of bands p ^k In addition to X _p The other bands, B, represent the number of bands of the band subset. Taking the wave band with the minimum square mean value in the wave band subset as a clustering center wave band of the wave band subset;

s2.4: grouping arbitrary bands p ^k Non-clustering center band X _i ∈(X ₁ ,X ₂ ,..,X _i ,X _p-1 ,X _p+1 ,..,X _B ) With other groups p ⁱ Calculating the square mean value of the relative entropy sum according to S2.3 by using the clustering center wave bands of i belonging to (1,2, 1, i, K-1, K +1, K), and if the clustering center is not changed, taking the K clustering center wave bands as the feature image set M _i And i belongs to (1, 2.. K) otherwise, re-determining the clustering center. That is, each cluster center is equal to the average value of the corresponding cluster band, and the formula is

Wherein i ∈ (1, 2.. K), j ∈ (1, 2.. B).

The step 6 specifically includes:

s6.1: first, the filtered image set F obtained in S5 _i [N]i ∈ (1, 2.. K), and F ∈ (K) _i [N]Randomly dividing the training samples into training samples train _ samples and test _ samples, and performing cross validation on the training samples train _ samples, wherein the specific method of the cross validation comprises the steps of firstly randomly generating a penalty coefficient p' and two parameter sets, namely parameters, namely an rbf kernel function g, and then training a support vector machine by using the generated parameter sets and the training samples train _ samples;

s6.2: the function of the training support vector machine can be set as svmtrain (train _ samples, parameter), the output result is a structural body model containing parameters, the model and the test sample test _ samples are predicted by using the support vector machine, and the finally predicted ground object classification map is output;

s6.3: calculating a confusion matrix fusion _ matrix according to the predicted tag and the real tag, and then calculating the predicted tag according to the confusion matrixClassification precision N of ground object classification map _i ，i∈(1,2,...,K)；

S6.4: a decision fusion strategy based on majority voting is applied to K classification precisions to obtain a final classification result; the concrete formula is

Where l represents the label of one of the G possible surface feature classes of the test pixel, j is the classifier index, N (e) represents the number of times class e is predicted in the support vector machine, I represents the exponential function, j represents the number of times the class e is predicted in the support vector machine,

Indicating the strength of the vote, this step uses equal voting weights.

The invention has the beneficial effects that: the invention utilizes a band clustering algorithm based on relative entropy to iteratively find out the central band in each band subset, and retains the information of the original spectrum band.

Drawings

FIG. 1 is a flow chart of the present invention.

Detailed Description

The technical solution in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

As shown in fig. 1, the present invention comprises the steps of:

s1: inputting an original hyperspectral image X ^Q×D Q is the number of pixels on each wave band, D is the number of wave bands, and D are all the hyperspectral imagesDividing the wave band into K hyperspectral sub-wave band sets; computing a kth subband set P ^{k∈(1,...,K)} The calculation method comprises the following steps:

wherein X ═ X (X) ₁ ,...,X _D )∈R ^Q×D Representing an original hyperspectral image X with D number of bands and Q pixels per band,

represents the maximum integer step size not exceeding D/K;

s2: for each subset P in the band set obtained at S1 ^{k∈(1,...,K)} Determining clustering center band set M _i ，i∈(1,2,...,K)；

The step S2 specifically includes:

s2.1: initializing cluster centers to

k is an integer and is the initial central band of each cluster group;

s2.2: setting any two wave bands X in high spectral data _p And X _q Where p ≠ q, then band X _p And X _q Is p, q ∈ (X) ₁ ,...,X _D )

Wherein m is _p (x)、m _q (x) Representing the probability corresponding to each pixel x in any two wave bands, wherein the larger the relative entropy between the two wave bands is, the larger the difference of the random distribution probability corresponding to the pixels of the two wave bands is, and the lower the similarity between the wave bands is; otherwise, the similarity between the wave bands is higher;

s2.3: updating the cluster center based on the relative entropy, wherein the measurement formula for updating the cluster center is

Wherein, X _p ∈p ^k ，

Representing a subset p of bands ^k In (C) X _p Mean square of the sum of relative entropies between a band and other bands, X _q Representing a subset of bands p ^k In addition to X _p The other bands, B, represent the number of bands of the band subset. Taking the wave band with the minimum square mean value in the wave band subset as a clustering center wave band of the wave band subset;

s2.4: grouping arbitrary bands p ^k Non-clustering center band X _i ∈(X ₁ ,X ₂ ,..,X _i ,X _p-1 ,X _p+1 ,..,X _B ) With other groups p ⁱ Calculating the square mean value of the relative entropy sum according to S2.3 by using the clustering center wave bands of i belonging to (1,2, 1, i, K-1, K +1, K), and if the clustering center is not changed, taking the K clustering center wave bands as the feature image set M _i And i belongs to (1, 2.. K) otherwise, re-determining the clustering center. That is, each cluster center is equal to the average value of the corresponding cluster band, and the formula is

Wherein i ∈ (1, 2.. K), j ∈ (1, 2.. B). S3: for clustering center wave band set M _i I ∈ (1, 2.. K) is subjected to Gaussian filtering, and the specific formula is

In the formula, M _i For the input image, a and b are the index coordinates of the input image pixels,

is the regularization parameter, N (a) is the set of neighborhood pixels for a, σ _s Is the spatial standard deviation, G _i Is the result of gaussian filtering;

s4: initializing the guide image, i.e. setting the Gaussian filtered image as the initial guide image, F _i [0]＝G _i ；

S5: the central wave band set M obtained by clustering _i i ∈ (1, 2.. K.) as the input image set and the guide image F _i [0]Performing a domain transformationRecursive filtering; the principle of recursive filtering of the domain transform is to apply a given transform domain D _t Will input the image M _i Conversion to D _t In (1), the formula of the recursive filtering is

F _i [N]＝(1-d ^c )M _i [N]+d ^c F _i [N-1]

In the formula, F _i [N]For recursively filtered feature images, d ∈ [0,1 ]]Is a feedback coefficient, and

σ _s is the spatial standard deviation, c ═ ct (x) _n )-ct(x _n-1 ) To transform domain D _t Middle adjacent sample point x _n And x _n-1 The distance between the two, the formula of the transformation function ct (u) is

In the formula, σ _s Is the spatial standard deviation, σ _r Is the spectral standard deviation. M _i ' (x) is an input image M _i First derivative function at pixel x, d as c increases ^c Approaching to 0, thereby playing the role of edge protection; s6: filtered image set F for the last time of recursive filtering of a domain transform _i [N]And i belongs to (1, 2., K) and a support vector machine is used for classification to obtain classification precision.

The step 6 specifically comprises:

s6.1: first, the filtered image set F obtained in S5 _i [N]i ∈ (1, 2.. K), and F ∈ (K) _i [N]Randomly dividing the training samples into training samples train _ samples and test _ samples, and performing cross validation on the training samples train _ samples, wherein the specific method of the cross validation comprises the steps of firstly randomly generating a penalty coefficient p' and two parameter sets, namely parameters, namely an rbf kernel function g, and then training a support vector machine by using the generated parameter sets and the training samples train _ samples;

s6.2: the function of the training support vector machine can be set as svmtrain (train _ samples, parameter), the output result is a structural body model containing parameters, the model and the test sample test _ samples are predicted by using the support vector machine, and the finally predicted ground object classification map is output;

s6.3: calculating a confusion matrix fusion _ matrix according to the predicted prediction label and the real label, and then calculating the classification precision N of the predicted ground object classification map according to the confusion matrix _i ，i∈(1,2,...,K)；

S6.4: a decision fusion strategy based on majority voting is applied to K classification precisions to obtain a final classification result; the concrete formula is

Where l represents the label of one of the G possible feature classes for the test pixel, j is the classifier index, N (e) represents the number of times class e is predicted in the support vector machine, I represents the exponential function, b,

Indicating the strength of the vote, this step uses equal voting weights.

In the present invention, as shown in table 1, the proposed method and the conventional SVM, PCA method and EMP-PLS-SVM, BS _ EPF _ SDAE method are performed on Indian Pines data set, and the overall accuracy of the method is improved by 16.93% and 18.2% compared to the conventional SVM, PCA method. Compared with the EMP-PLS-SVM and the BS _ EPF _ SDAE algorithm, the overall precision is improved by 6.07 percent and 2.4 percent. The method has a bad expression on classification precision.

The invention utilizes a band clustering algorithm based on relative entropy to iteratively find out the central band in each band subset, and retains the information of the original spectrum band.

Claims (3)

1. A hyperspectral image classification algorithm based on band clustering and improved domain transform recursive filtering is characterized by comprising the following steps:

s1: inputting an original hyperspectral image X ^Q×D Q is the number of pixels on each wave band, D is the number of wave bands, and all D wave bands of the hyperspectral image are divided into K hyperspectral sub-wave band sets; computing a kth subband set P ^{k∈(1,...,K)} The calculation method comprises the following steps:

wherein X ═ X (X) ₁ ,...,X _D )∈R ^Q×D Representing an original hyperspectral image X with D number of bands and Q pixels per band,

represents the maximum integer step size not exceeding D/K;

s2: for each subset P in the band set obtained at S1 ^{k∈(1,...,K)} Determining clustering center band set M _i ，i∈(1,2,...,K)；

S3: for clustering center wave band set M _i I ∈ (1, 2.. K) is subjected to Gaussian filtering, and the specific formula is

In the formula, M _i For the input image, a and b are the index coordinates of the pixels of the input image,

is the regularization parameter, N (a) is the set of neighborhood pixels for a, σ _s Is the spatial standard deviation, G _i Is the result of gaussian filtering;

s4: initializing the guide image, i.e. setting the Gaussian filtered image as the initial guide image, F _i [0]＝G _i ；

S5: the central wave band set M obtained by clustering _i i ∈ (1, 2.. K.) as the input image set and the guide image F _i [0]Performing domain transform recursive filtering; the principle of recursive filtering of the domain transform is to apply a given transform domain D _t Will input the image M _i Conversion to D _t In (1), the formula of the recursive filtering is

F _i [N]＝(1-d ^c )M _i [N]+d ^c F _i [N-1]

In the formula, F _i [N]For recursively filtered feature images, d ∈ [0,1 ]]Is a feedback coefficient, and

σ _s is the spatial standard deviation, c ═ ct (x) _n )-ct(x _n-1 ) To transform domain D _t Middle adjacent sample point x _n And x _n-1 The distance between the two, the formula of the transformation function ct (u) is

In the formula, σ _s Is the spatial standard deviation, σ _r Is the spectral standard deviation, M _i ' (x) is an input image M _i First derivative function at pixel x, d as c increases ^c Approaching to 0, thereby playing the role of edge protection;

s6: filtered image set F for the last time of recursive filtering of a domain transform _i [N]And i belongs to (1, 2.. multidot.K) and a support vector machine is used for classification, so that the classification precision is obtained.

2. The hyperspectral image classification algorithm based on band clustering and improved domain transform recursive filtering according to claim 1, wherein the step S2 specifically comprises:

s2.1: initializing cluster centers to

k is an integer and is the initial central band of each cluster group;

s2.2: setting any two wave bands X in high spectral data _p And X _q Where p ≠ q, then band X _p And X _q Is p, q ∈ (X) ₁ ,...,X _D )

Wherein m is _p (x)、m _q (x) Representing the probability corresponding to each pixel x in any two wave bands, wherein the larger the relative entropy between the two wave bands is, the larger the difference of the random distribution probability corresponding to the pixels of the two wave bands is, and the lower the similarity between the wave bands is; otherwise, the similarity between the wave bands is higher;

s2.3: updating the cluster center based on the relative entropy, wherein the measurement formula for updating the cluster center is

Wherein X _p ∈p ^k ，AV _Xp Representing a subset of bands p ^k In (C) X _p Mean square of the sum of relative entropies between a band and other bands, X _q Representing a subset of bands p ^k In addition to X _p B, the number of the wave bands of the wave band subset is represented, and the wave band with the minimum square mean value in the wave band subset is used as a clustering center wave band of the wave band subset;

s2.4: grouping arbitrary bands p ^k Non-clustering center band X _i ∈(X ₁ ,X ₂ ,..,X _i ,X _p-1 ,X _p+1 ,..,X _B ) With other groups p ⁱ Calculating the square mean value of the relative entropy sum according to S2.3 by using the clustering center wave bands of i belonging to (1,2, 1, i, K-1, K +1, K), and if the clustering center is not changed, taking the K clustering center wave bands as the feature image set M _i If i ∈ (1, 2.. K) is not satisfied, the cluster centers are re-determined, that is, each cluster center is equal to the average value of the corresponding cluster band, and the formula is

Wherein i ∈ (1, 2.. K), j ∈ (1, 2.. B).

3. The hyperspectral image classification algorithm based on band clustering and improved domain transform recursive filtering according to claim 2, wherein the step 6 specifically comprises:

s6.1: first, the filtered image set F obtained in S5 _i [N]i ∈ (1, 2.. K), and F ∈ (K) _i [N]Randomly dividing the training samples into training samples train _ samples and test _ samples, and performing cross validation on the training samples train _ samples, wherein the specific method of the cross validation comprises the steps of firstly randomly generating a penalty coefficient p' and two parameter sets, namely parameters, namely an rbf kernel function g, and then training a support vector machine by using the generated parameter sets and the training samples train _ samples;

s6.2: the function of the training support vector machine can be set as svmtrain (train _ samples, parameter), the output result is a structural body model containing parameters, the model and the test sample test _ samples are predicted by using the support vector machine, and the finally predicted ground object classification map is output;

s6.3: calculating a confusion matrix fusion _ matrix according to the predicted prediction label and the real label, and then calculating the classification precision N of the predicted ground object classification map according to the confusion matrix _i ，i∈(1,2,...,K)；

S6.4: a decision fusion strategy based on majority voting is applied to K classification precisions to obtain a final classification result; the concrete formula is

Where l represents the label of one of the G possible feature classes for the test pixel, j is the classifier index, N (e) represents the number of times class e is predicted in the support vector machine, I represents the exponential function, b,

Indicating the strength of the vote, this step uses equal voting weights.

Images