CN106033545B - Wave band selection method of determinant point process - Google Patents

Abstract

The invention provides a wave band selection method of a determinant point process, which comprises the following steps: 1, dividing wave bands of original hyperspectral data to generate different candidate wave band subsets; 2, evaluating the candidate wave band subset, and selecting the candidate wave band subset which accords with the determinant point process; and 3, calculating the probability of the occurrence of the candidate waveband subset which accords with the determinant point process according to the property of the elementary determinant point process, wherein the probability is higher, so that the probability is higher, namely the probability of the occurrence is the best candidate waveband subset. The method solves the problems of high redundancy in the original hyperspectral data, improvement of data expressiveness and poor discriminability of the existing method.

Description

Wave band selection method of determinant point process

Technical Field

The invention belongs to the technical field of image processing, particularly relates to hyperspectral waveband selection, and particularly relates to a waveband selection method for a determinant point process.

Background

With the development of remote sensing technology and imaging spectrometers, the application of hyperspectral remote sensing images is more and more extensive, but the characteristics of more wave bands, huge data volume and the like bring great difficulty to the classification, identification and the like of hyperspectral images. For example, the information redundancy is high, the space required by data storage is large, the processing time is long, and a dimension disaster phenomenon easily occurs due to the large number of wave bands of the hyperspectral image, namely, the classification precision is reduced. Therefore, it is very necessary to reduce the data amount and save resources in order to ensure the feature classification recognition rate. The wave band selection is to select a wave band subset which plays a main role from all wave bands of the hyperspectral image, so that the data dimension of the hyperspectral image can be greatly reduced, useful information can be completely reserved, and the method has a special meaning. The wave band selection of the hyperspectral image is a very complicated wave band combination optimization problem, and the selected wave band combination is required to have better performance, namely, the wave band combination with larger information content, smaller relevance and better category separability is selected.

The development of the band selection of the hyperspectral image can be regarded as the development of the problem of searching for optimal band subsets, namely, the band subsets enabling the criterion function to be optimal are searched for in all bands of the hyperspectral image according to a certain searching method. The band selection method can be classified into an optimal search band selection method and a suboptimal search band selection method according to the search method.

The optimal search waveband selection method is to search a group of waveband subsets optimal in a certain sense in all waveband features of an image, so that the waveband subsets can maximally retain ground object target information under the condition of reducing the dimensionality of a hyperspectral image and the same size. Hongjun Su et al propose particle group based Band Selection methods in the literature "Hongjun Su, Qian Du, Genshe Chen, Peijun Du, and M.Cristani. optimized Hyperspectral Band Selection Using particle Swarm optimization. IEEE Journal of Selected Topics in Applied Earth updates and remotes Sensing, pages 1659-. This method has the disadvantage that it is computationally expensive. Due to the fact that the number of the wave bands of the hyperspectral image is large, the number of wave band combinations is large, and therefore finding of the optimal wave band subset is quite difficult.

Due to the complexity of the optimal search algorithm, the hyperspectral image band selection usually adopts a suboptimal search algorithm. According to a certain evaluation criterion function, a group of band subsets with better performance is selected by adopting a specific search strategy, but the band subsets are not necessarily the optimal band subsets. Due to the complexity of the optimal search algorithm, the hyperspectral image band selection usually adopts a suboptimal search algorithm. According to the evaluation criterion function, a group of wave band subsets with better performance is selected by adopting a specific search strategy, but the wave band subsets are not necessarily the optimal wave band subsets. The method has the defects that due to the fact that the number of wave bands of the hyperspectral image is large, the data volume is large, the performance of the obtained wave band subset is still low in the conventional suboptimal search algorithm, or the calculation amount is large, the processing time is long, and the satisfactory effect is not achieved.

Disclosure of Invention

The invention provides a wave band selection method for a determinant point process to solve the technical problems, and the method solves the problems of high redundancy in original hyperspectral data, improvement of data expressiveness and poor discriminability of the existing method.

The technical scheme of the invention is as follows:

a wave band selection method of determinant point process is characterized by comprising the following steps:

1, dividing wave bands of original hyperspectral data to generate different candidate wave band subsets;

2, evaluating the candidate wave band subset, and selecting the candidate wave band subset which accords with the determinant point process;

and 3, calculating the probability of the occurrence of the candidate waveband subset which accords with the determinant point process according to the property of the elementary determinant point process, wherein the probability is higher, so that the probability is higher, namely the probability of the occurrence is the best candidate waveband subset.

The step 1 is specifically: suppose that the band of the original hyperspectral data is divided into p bands B ═ B₁,B₂,...B_p]Selecting K wave bands from the p wave bands to form different candidate wave band subsets; the candidate band subset is represented by a binary with a length of N, 1 represents selection, 0 represents non-selection, and the number of 1 in the p-band subset is K.

The step 3 is specifically: let X be the original hyperspectral data, and the size is p × q, where p is the number of bands, q is the total number of pixels in each band, a is the flag vector of the candidate band subset, Y is an index value other than 0 in a, that is, the index of the selected band, and the length is the number of the selected band subset, and if a ═ 0101], Y ═ 2,4} indicates that two bands are selected from four bands: the 2 nd and 4 th bands; probability of occurrence of candidate band subset a:

P(Y)＝det(L_Y)/det(L+I)

where p (Y) is the probability that the candidate band subset a corresponds to the index value Y. Matrix L_Y≡[L_ij]_i,j∈Y，det(L_Y) Representation solving matrix L_YThe matrix I is a unit matrix of size p × p, and L is a metric matrix between bands, which can be simply expressed as: l is XX', so Σ det (L)_Y)＝det(L+I)。

Repeating the step 3, performing terrain classification on the optimal candidate wave band subset after the wave band selection, and counting the accuracy of the classification to obtain the accuracy of different wave bands:

wherein tp (true positive) represents the number of correctly classified positive samples, tn (true negative) is the number of correctly classified negative samples, P is the number of all positive samples, N is the number of all negative samples, and acc represents the probability, i.e. accuracy, that correctly matched positive and negative samples occupy all samples.

The invention has the advantages that:

the method utilizes a determinant point process to select the wave band subsets with differences, and effectively solves the problems of large calculation amount and long processing time of original data; effective information of the original data is used for guiding the ground feature classification process, so that a better result is obtained; the data dimension of the hyperspectral image is reduced, and useful information can be completely reserved, so that the hyperspectral image can be used for the aspects of geographic national situation investigation, military reconnaissance, environmental monitoring and the like.

Drawings

FIG. 1 is a flowchart of the band selection process based on determinant point process of the present invention.

Detailed Description

Referring to fig. 1, the steps implemented by the present invention are as follows:

and step 1, generating different candidate waveband subsets.

(1a) The set of images in the raw hyperspectral data is divided into different band subsets. A flag vector is established for the selected band, which is marked with 1, otherwise 0.

(1b) Randomly selecting K wave bands as an initial test wave band subset, wherein the number of 1 in the sign vector is K;

and 2, establishing an evaluation criterion for evaluating the quality of the wave band subset. Assuming that the candidate subset conforms to the determinant point process, the bands in the subset have strong diversity.

A determinant point process probability of the subset is established. Let X be the original hyperspectral data and be p × q in size, where p is the number of bands and q is the total number of pixels in each band. A is a flag vector for selecting a subset of bands, and Y is an index value in A that is not 0, i.e., an index for selecting a band. If a ═ 0101], Y ═ 2,4 }. The probability that subset a is selected is:

P(Y)＝det(L_Y)/det(L+I)

wherein L is_Y≡[L_ij]_i,j∈YAnd I is a unit matrix of size p × p. L is a metric matrix between bands, which can be simply expressed as: l ═ XX'. Obviously, Σ det (L)_Y)＝det(L+I)。

And 3, searching the optimal subset in different subset spaces according to the determinant point process.

According to the property of the determinant point process, the wave bands conforming to the determinant point process have strong difference, and simultaneously, the original wave bands can be well represented. The method comprises the following steps:

(3a) inputting the metric matrix L to obtain the feature vector v_nAnd a characteristic value lambda_n. Let J be phi and Y be phi.

(3b) According to the probability Pr_n＝λ_n/(λ_n+1), n ═ 1., p, the bands being chosen such that J ═ J utou { n }, the corresponding eigenvectors V ═ { V { n } being chosen accordingly_n}_n∈J。

(3c) According to the probability Pr_i＝Σ_v∈V(v^Te_i) Selecting band B from band set B_iUpdating Y and V:

Y＝Y∪{i}，

wherein

Is orthogonal to e_iA subset of feature vectors.

(3d) Judging that the | V | is larger than 0, outputting the obtained subset Y, and continuing (3b) (3c) if not.

And 4, calculating the accuracy (acc).

And (4) repeating the step (3), carrying out ground object classification on the data after the wave band selection, and counting the accuracy of the classification to obtain the accuracy of the wave band after the wave band selection.

The effects of the present invention can be further explained by the following experiments.

1. Simulation conditions

The invention uses MATLAB software to simulate the central processing unit of Intel (R) Core i3-21303.4GHZ and the memory 16G, WINDOWS 8 operating system.

2. Emulated content

Firstly, the hyperspectral data is used as a real hyperspectral image. Acquired in 1992 by the National Aeronautics and Space Administration (NASA) airborne visible/infrared imaging spectrometer (AVIRIS) in the remote indian area, northwest, indiana, usa. The spatial resolution of the image is 20m multiplied by 20m, the size of the image is 145 multiplied by 145, and the wavelength range is 400-2500 nm. The original AVIRIS collector contains 224 bands, but 4 of them contain only 0 values and are therefore filtered out. In most cases, the known spectral image is affected by the absorption band in the atmosphere in some spectra, and in IndianPine, such bands include 20 bands, 104 and 108 bands, and 200 bands remain after 150 and 163 and 220 bands are removed. The entire image data contains 16 types of vegetation, but not all pixels belong to these 16 types, with many irrelevant areas divided into backgrounds.

Experiments of the algorithm of the present invention (band selection based on determinant point process) were done on IndianPine datasets. In order to prove the effectiveness of the algorithm and comprehensively consider the popularity and the novelty of the algorithm, 6 comparison methods of CEM-BCM/BDM, CEM-BCC/BDC, LCMV-BCM/BDM, LCMV-BCC/BDC, CBBS-MI and CBBS-KLD are selected for comparison. Wherein CEM-BCM/BDM, CEM-BCC/BDC, LCMV-BCM/BDM and LCMV-BCC/BDC are proposed in the publications "C. -I.Chang and S.Wang.Constrained band selection for hyperspectral image, IEEE Transactions on society and Remote Sensing,44(6): 1575-. CBBS-MI and CBBS-KLD are described in detail in "A.M.us 'o, F.Pla, J.M.Sotoca, and P.Garc' ia-seville. Cluster-based hyper-spectral band selection using information measures. IEEE Transactions on Geoscience and Remote Sensing,45(12):4158, 4171, 2007".

And measuring the effect of the wave band selection by using the accuracy of the ground feature classification. The first 50 samples were selected for training in the IndianPine dataset and the remaining samples were tested.

Secondly, according to the distance measurement formula in the step 3, the distance between the test set and the prototype image set is calculated, and the accumulated accuracy is calculated.

The above steps were repeated 10 times, and the average accuracy precision was calculated, and the results are shown in table 1.

The 4 classifiers commonly used were selected: CART, KNN, Naive and SVM. As can be seen from table 1, the classification accuracy of the present invention is higher than that of the existing band selection method. Therefore, the method is more effective and robust than other methods.

TABLE 1 Effect of band selection on ground object Classification

The invention can provide accurate false color image, has obvious ground object distinguishing degree and visualization effect, and is convenient for target detection and ground object observation in later period.

Claims (2)

1. A wave band selection method of determinant point process is characterized by comprising the following steps:

1, dividing wave bands of original hyperspectral data to generate different candidate wave band subsets;

suppose that the band of the original hyperspectral data is divided into p bands B ═ B₁,B₂,...B_p]Selecting K wave bands from the p wave bands to form different candidate wave band subsets; the candidate waveband subsets are represented by a binary system with the length of N, 1 represents selection, 0 represents non-selection, and the number of 1 in the p waveband subsets is K;

2, evaluating the candidate wave band subset, and selecting the candidate wave band subset which accords with the determinant point process;

let X be the original hyperspectral data, and the size is p × q, where p is the number of bands, q is the total number of pixels in each band, a is the flag vector of the candidate band subset, Y is an index value other than 0 in a, that is, the index of the selected band, and the length is the number of the selected band subset, and if a ═ 0101], Y ═ 2,4} indicates that two bands are selected from four bands: the 2 nd and 4 th bands; probability of occurrence of candidate band subset a:

P(Y)＝det(L_Y)/det(L+I)

wherein P (Y) is the probability of index value Y corresponding to candidate waveband subset A, and matrix L_Y≡[L_ij]_i,j∈Y，det(L_Y) Representation solving matrix L_YThe matrix I is a unit matrix with the size of p multiplied by p; l is a metric matrix between bands, which can be simply expressed as: l ═ XX', therefore, Σ det (L)_Y)＝det(L+I)；

Calculating the probability of the occurrence of the candidate waveband subsets conforming to the determinant point process according to the property of the elementary determinant point process, wherein the probability is higher, and the candidate waveband subset with the maximum probability is the optimal candidate waveband subset;

3.1, inputting a measurement matrix L to obtain a feature vector v_nAnd a characteristic value lambda_nLet J be phi and Y be phi;

3.2 according to the probability Pr_n＝λ_n/(λ_n+1), n ═ 1., p, the bands being chosen such that J ═ J utou { n }, the corresponding eigenvectors V ═ { V { n } being chosen accordingly_n}_n∈J；

3.3 according to probability

Selecting band B from band set B_iUpdating Y and V:

Y＝Y∪{i}，

wherein

Is orthogonal to e_iA subset of feature vectors of (a);

and 3.4, judging | V | 0, meeting the output of the obtained subset Y, and otherwise, continuing to 3.2 and 3.3.

2. The method of band selection for a determinant point process as in claim 1, wherein: and (3) carrying out ground object classification on the optimal candidate waveband subset after the waveband selection, and counting the classification accuracy to obtain the accuracy of different wavebands:

wherein tp (true positive) represents the number of correctly classified positive samples, tn (true negative) is the number of correctly classified negative samples, P is the number of all positive samples, N is the number of all negative samples, and acc represents the probability, i.e. accuracy, that correctly matched positive and negative samples occupy all samples.

Images