CN110135448B - Hyperspectral image virtual dimension estimation method based on ridge ratio shrinkage - Google Patents
Hyperspectral image virtual dimension estimation method based on ridge ratio shrinkage Download PDFInfo
- Publication number
- CN110135448B CN110135448B CN201811595570.4A CN201811595570A CN110135448B CN 110135448 B CN110135448 B CN 110135448B CN 201811595570 A CN201811595570 A CN 201811595570A CN 110135448 B CN110135448 B CN 110135448B
- Authority
- CN
- China
- Prior art keywords
- matrix
- hyperspectral image
- sample
- virtual dimension
- ratio
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/751—Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a method for estimating a virtual dimension of a hyperspectral image based on ridge ratio shrinkage, which comprises the following steps of: extracting a sample matrix of a hyperspectral image to be processed; calculating a covariance matrix and a correlation matrix of the obtained sample matrix, calculating characteristic values of the covariance matrix and the correlation matrix, and sequencing the characteristic values; constructing a ratio sequence according to the characteristic value sequence; and comparing the sequence value in the ratio sequence with a preset constraint value to obtain the virtual dimension of the hyperspectral image to be processed. The estimation method of the invention abandons the idea of hypothesis test, and directly adopts the algorithm for constructing the ratio series for estimation, thereby improving the accuracy of the virtual dimension estimation of the hyperspectral image.
Description
Technical Field
The invention belongs to the technical field of virtual dimension estimation of hyperspectral images, and particularly relates to a virtual dimension estimation method of a hyperspectral image based on ridge ratio shrinkage.
Background
Spectral images with spectral resolution in the order of 10nm are called hyperspectral images. By means of high spectrum sensors carried on different space platforms, namely an imaging spectrometer, a target area is imaged simultaneously in tens of to hundreds of continuous and subdivided spectral bands in ultraviolet, visible light, near infrared and mid-infrared areas of an electromagnetic spectrum; the earth surface image information is obtained, and simultaneously the spectrum information is also obtained, so that the combination of the spectrum and the image is really realized. Compared with multispectral remote sensing images, the hyperspectral images not only greatly improve the information abundance, but also provide possibility for more reasonable and effective analysis and processing of the type of spectral data in the aspect of processing technology. The influence and the development potential of the hyperspectral image technology are incomparable with the development stages of the prior art. It not only draws attention from the remote sensing world, but also draws great interest in other fields (such as medicine, agriculture, etc.). In practical application, direct analysis of data faces dimensionality disasters, and the problem of dimensionality reduction of a hyperspectral image is an important link.
The hyperspectral image is composed of a series of pixel points, and each pixel point can be expressed as an L-dimensional vector, wherein L represents the number of channels. In the actual processing of the image, the information of all the pixel points of the image is highly correlated and can be expanded by a group of low-dimensional signal end members. Defining an image Virtual Dimension (VD) as a dimension of a minimum signal end metabase expanded into hyperspectral image data, wherein the value is usually far smaller than L, and thus great possibility is provided for dimension reduction processing of the user; by processing the hyperspectral image in a lossless dimension reduction way, the calculation time and the storage space can be greatly reduced.
In order to define the features of hyperspectral images, the concept of virtual dimensions is now commonly used. The conventional method is to adopt an HFC method, the HFC method is simple and effective based on characteristic value analysis and a Neyman-Pearson detection theory, but the assumed inspection thought cannot ensure the estimated consistency, and the first type of error needs to be selected in an attempt, so that the accuracy of the hyperspectral image virtual dimension estimation is poor.
In summary, a new hyperspectral image virtual dimension estimation algorithm is needed.
Disclosure of Invention
The invention aims to provide a method for estimating a virtual dimension of a hyperspectral image based on ridge ratio shrinkage, so as to solve the existing technical problems. The method abandons the idea of hypothesis test, and directly adopts an algorithm for constructing a ratio array to carry out estimation, so that the accuracy of the virtual dimension estimation of the hyperspectral image can be improved.
In order to achieve the purpose, the invention adopts the following technical scheme:
a virtual dimension estimation method of a hyperspectral image based on ridge ratio shrinkage comprises the following steps:
Step 2, obtaining a sample matrix according to the step 1Calculating to obtain a sample covariance matrixAnd a sample correlation matrix
Step 3, respectively calculating and obtaining the covariance matrix of the sample obtained in the step 2And a sample correlation matrixThe eigenvalues of (a) are ranked;
step 4, sorting according to the characteristic values obtained in the step 3, and constructing a ratio series;
and 5, comparing the sequence value in the ratio sequence with a preset constraint value to obtain the virtual dimension of the hyperspectral image to be processed.
wherein, the matrixIs an L x n matrix, L is the number of bands, n is the number of pixel points,each column vector of (a) represents an end-member,each column vector of the matrix represents the abundance of the corresponding pixel point under the end member,is white noise.
Further, in step 2, the covariance matrix of the sampleAnd a sample correlation matrixThe specific expressions of (a) are respectively:
wherein the content of the first and second substances,representing the corresponding pixel points;representing the sample mean vector.
Further, step 3 specifically includes calculating and obtaining a sample covariance matrixIs positive ordering of the eigenvalues ofSample correlation matrixIs positive ordering of the eigenvalues of
furthermore, k is more than or equal to 0.2 and less than or equal to 0.5.
Further, in step 5, the expression of the estimated amount of the virtual dimension is:
constructing a ratio sequenceIs provided withThe constraint of τ is: 0<τ<1; selecting the index j from small to large, and selecting the corresponding sequence valueComparing with tau, and determining the last index value smaller than tau as the virtual dimension; or the index j is selected from large to small, and the index corresponding to the first ratio sequence item smaller than tau is the solved virtual dimension.
Furthermore, the value range of tau is more than or equal to 0.4 and less than or equal to 0.6.
Further, τ is 0.5.
Further, a noise reduction processing step is arranged between the step 1 and the step 2, and the noise reduction processing method is to carry out noise reduction processing on the obtained signalCarrying out pretreatment; the pretreatment comprises the following steps: de-averaging and whitening.
Compared with the prior art, the invention has the following beneficial effects:
the method of the invention inherits the simple effectiveness of the HFC algorithm; in the face of huge data of a hyperspectral image, the method can greatly reduce the processing time and the storage space under the condition of not losing information, has greater significance to practical application, and simultaneously lays a foundation for wide popularization; the method has no bias of an HFC algorithm, abandons the idea of hypothesis test, directly adopts the method of constructing a ratio series, is quick, simple and convenient, and can greatly improve the accuracy.
Furthermore, the parameters are determined according to theoretical proof and actual inspection, a new parameter selection problem does not need to be considered when the method is used, and the method has strong operability and reproducibility; the selection of the ridge function is data-driven, and has great universality and reasonableness.
Drawings
FIG. 1 is a schematic block diagram of a flow chart of a virtual dimension estimation method of a hyperspectral image based on ridge ratio shrinkage according to the invention.
Detailed Description
The invention is described in further detail below with reference to the figures and specific examples.
Comparative example
Existing HFC processes include:
step 1, extracting a sample matrix of a hyperspectral image to be processedSpecifically, the hyperspectral image to be processed has n pixel points and L wave bands, and each pixel point is composed of a series of end members and white noise. Each pixel point is an L-dimensional column vector and sample matrixThe expression of (a) is:
step 2, obtaining a sample covariance matrix of the sample matrixAnd a sample correlation matrixThe expressions are respectively:
step 3, calculating and obtaining the covariance matrix of the sample obtained in the step 2And a sample correlation matrixA characteristic value of (d);are respectivelyAndpositive ordering of eigenvalues of.
The default source signal is a positive constant, the noise is white,andthe characteristic values of (a) have the following properties:
where VD represents the number of features.
The above comparative example is the basis of the HFC method, and according to the calculation method, roughly speaking, the signal component affects the eigenvalue of the correlation matrix but does not affect the eigenvalue of the covariance matrix, and noise affects both eigenvalues identically. Therefore, if a certain component does not contain a feature, the eigenvalues of the covariance matrix and the correlation matrix are the same. By using this feature, a hypothesis testing problem is derived. The HFC algorithm does not work well in practice. This is mainly because the hypothesis testing concept does not guarantee the consistency of the estimation, and the size of the first error needs to try to select itself, which also increases the difficulty of estimation.
Referring to fig. 1, the method for estimating a virtual dimension of a hyperspectral image based on ridge ratio shrinkage of the present invention discards a hypothesis testing concept for HFC defects, constructs a ratio sequence method to estimate a virtual dimension, and continues to use HFC symbols, and specifically includes the following steps:
step 1, extracting a sample matrix of a hyperspectral image to be processedSample matrixExpressed as:
likewise, the matrixEach column vector of (a) represents an end-member (which can be considered a basis),each column vector of the matrix represents the abundance of the corresponding pixel point under the end member,is white noise.
Step 2, obtaining a sample matrix according to the step 1Calculating to obtain a corresponding covariance matrixAnd a correlation matrixThe specific expressions of the two are respectively as follows:
Step 3, calculating and obtaining the covariance matrix obtained in the step 2And a correlation matrixPositive ordering of eigenvalues of;is positive ordering of the eigenvalues of Is positive ordering of the eigenvalues of
Step 4, obtaining the covariance matrix according to the step 3And a correlation matrixPositive ordering of characteristic values of (1), constructing a ratio arrayTo solve the situation where 0/0 may occur, a ridge function is added, specifically:
wherein k is an unknown parameter to be determined, and the preferable value range is that k is more than or equal to 0.2 and less than or equal to 0.5. When k is greater than 0.5, the composition,
this results in a small estimate of the virtual dimension. When k is less than 0.2, the composition,
it is not general enough.
And 5, constructing the obtained ratio number series according to the step 4, and comparing and calculating to obtain the number of the virtual dimensions.
The expression for the estimator of the virtual dimension is:
the constraint of τ is: 0< τ < 1.
When n goes to infinity, it has been demonstratedVD represents the true value of the virtual dimension in the overall sense.Is an estimate of VD.
Since the general virtual dimension is much smaller than L/4, for the convenience of calculation, calculation can be performed from [ L/4] onwards. The value of j corresponding to the first term smaller than τ is the required number p of virtual dimensions.
According to the idea of plug-in, tau is preferably selected to be within a value range of 0.4-0.6; further preferably τ is 0.5.
It should be noted that, regarding the selection of the parameter k, after a series of simulations and practical experiments, it is found that the method of the present invention is not sensitive to the selection of k, and finally, we select k as 1/4 to obtain the best effect.
Principle analysis of the invention
In view of the fact that when n is sufficiently large,andapproaching (lambda) at a rate of one-half the root number niAnd gammaiRespectively representing the true values of the corresponding characteristic values in the overall sense). Structural ridge item cnAnd requires cnGoing to infinity at n is going to 0 at a speed that is less than one-n times the root number. According to theorem guarantee thatWe can determine VD by looking for the last minimum point.
Example 1
The first example uses analog data. Comparing the TRR method and the NWTRR method proposed by the present invention with other methods, wherein the NWTRR method is a method in which the TRR method is used after whitening noise. The noise whitening method is as follows: to pairAnd (4) carrying out pretreatment. The pre-processing includes de-averaging and whitening. The correlation between each observation is removed by linearly transforming the observation data vector to make the mean value zero and the variance 1.
In generating the simulation data, we set the virtual dimension of the image to be 5, the white noise to be Legendre white noise, and the SNR to take four different values of 20, 40, 60, and 80 to increase the accuracy of the comparison, the results are shown in Table 1:
TABLE 1 comparative data for example 1
Table 1 analysis shows that the above method performs substantially well at a signal to noise ratio of 20. With the improvement of the signal to noise ratio, the TRR and the NWTRR of the denoising improved version have very remarkable accuracy in simulation data compared with other classical methods HySime, HySURE, ELM, HFC and NWHFC. Even with the denoised modified version of nwwfc, its accuracy and NWTRR, even TRR, are not comparable. This represents a disadvantage of the HFC process, in contrast to the superiority of the TRR process of the present invention. As the virtual dimension increases, the gap will only grow larger.
Example 2:
in a second example we apply the method TRR of the present invention together with the other methods used in example 1 to four real images for virtual dimension estimation, the results are shown in table 2:
TABLE 2 comparative data for example 2
The above four actual images are commonly used hyperspectral datasets. As can be seen from Table 2, the TRR method of the present invention performs well in four groups of images, has a small floating range near the virtual dimension, and estimates the real virtual dimension substantially correctly. While other methods, such as ELM, which perform well in the analog data set, present significant errors. The performance of HFC and denoised modified version of NWHFC is also unsatisfactory in each set of images. Thus, in general, other methods have large errors.
In summary, in the current general theory of HFC hypothesis testing, we find it to be an inappropriate, biased estimate with poor estimation accuracy. In practice, it is more often found that the results have a non-negligible error. The method of the invention abandons the thought of hypothesis test and directly adopts the method of constructing ratio series. The present invention uses the series construction ratio of the differences. Considering that 0/0 is encountered in practical operation, the parameter sequence c is addednThe obtained new ratio sequence can be compared with τ, and excellent results can be obtained when τ is 0.5. The algorithm of the invention has higher accuracy than the existing conventional HFC method; the invention removes hypothesis test and bias from the source, and in actual operation, a large number of experiments prove that the algorithm of the invention has more accuracy than the HFC algorithm.
Claims (6)
1. A virtual dimension estimation method of a hyperspectral image based on ridge ratio shrinkage is characterized by comprising the following steps:
Step 2, obtaining a sample matrix according to the step 1Calculating to obtain a sample covariance matrixAnd a sample correlation matrix
Step 3, respectively calculating and obtaining the covariance matrix of the sample obtained in the step 2And a sample correlation matrixThe eigenvalues of (a) are ranked;
step 4, sorting according to the characteristic values obtained in the step 3, and constructing a ratio series;
step 5, comparing the sequence value in the ratio sequence with a preset constraint value to obtain a virtual dimension of the hyperspectral image to be processed;
wherein, the matrixIs an L x n matrix, L is the number of bands, n is the number of pixel points,each column vector of (a) represents an end-member,each column vector of the matrix represents the abundance of the corresponding pixel point under the end member,is white noise;
in step 2, the sample covariance matrixAnd a sample correlation matrixThe specific expressions of (a) are respectively:
wherein the content of the first and second substances,representing the corresponding pixel points;representing a sample mean vector;
step 3 specifically includes calculating and obtaining a sample covariance matrixIs positive ordering of the eigenvalues ofSample correlation matrixIs positive ordering of the eigenvalues of
2. the method for estimating the virtual dimension of the hyperspectral image based on ridge ratio shrinkage as claimed in claim 1, wherein k is in a range of 0.2-0.5.
3. The method for estimating the virtual dimension of the hyperspectral image based on ridge ratio shrinkage as claimed in claim 1, wherein in step 5, the expression of the estimated amount of the virtual dimension is as follows:
constructing a ratio sequenceIs provided withThe constraint of τ is: tau is more than 0 and less than 1; selecting the index j from small to large, and selecting the corresponding sequence valueComparing with tau, and determining the last index value smaller than tau as the virtual dimension; or the index j is selected from large to small, and the index corresponding to the first ratio sequence item smaller than tau is the solved virtual dimension.
4. The method for estimating the virtual dimension of the hyperspectral image based on ridge ratio shrinkage as claimed in claim 3, wherein τ is in a range of 0.4-0.6.
5. The method for estimating the virtual dimension of the hyperspectral image based on ridge ratio shrinkage as claimed in claim 4, wherein τ is 0.5.
6. The method for estimating the virtual dimension of the hyperspectral image based on ridge ratio shrinkage as claimed in claim 1, wherein a noise reduction processing step is further provided between step 1 and step 2, and the noise reduction processing method is to pairCarrying out pretreatment; the pre-processing includes de-averaging and whitening.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811595570.4A CN110135448B (en) | 2018-12-25 | 2018-12-25 | Hyperspectral image virtual dimension estimation method based on ridge ratio shrinkage |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811595570.4A CN110135448B (en) | 2018-12-25 | 2018-12-25 | Hyperspectral image virtual dimension estimation method based on ridge ratio shrinkage |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110135448A CN110135448A (en) | 2019-08-16 |
CN110135448B true CN110135448B (en) | 2021-03-23 |
Family
ID=67568390
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811595570.4A Active CN110135448B (en) | 2018-12-25 | 2018-12-25 | Hyperspectral image virtual dimension estimation method based on ridge ratio shrinkage |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110135448B (en) |
Family Cites Families (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101807301B (en) * | 2010-03-17 | 2012-11-07 | 北京航空航天大学 | High spectral image target detection method based on high order statistic |
CN102136067B (en) * | 2011-03-23 | 2013-02-27 | 复旦大学 | Cayley-Menger determinant-based hyperspectral remote sensing image end member extracting method |
CN102542288B (en) * | 2011-11-28 | 2013-09-25 | 北京航空航天大学 | Construction and merging classification method for high spectrum data multi-characteristic space |
CN104331880B (en) * | 2014-10-20 | 2017-02-15 | 西安电子科技大学 | Hyper-spectral mixed pixel decomposition method based on geometric spatial spectral structure information |
-
2018
- 2018-12-25 CN CN201811595570.4A patent/CN110135448B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN110135448A (en) | 2019-08-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Hong et al. | SULoRA: Subspace unmixing with low-rank attribute embedding for hyperspectral data analysis | |
Hong et al. | An augmented linear mixing model to address spectral variability for hyperspectral unmixing | |
Karoui et al. | Blind spatial unmixing of multispectral images: New methods combining sparse component analysis, clustering and non-negativity constraints | |
Zhang et al. | Spatial-temporal fraction map fusion with multi-scale remotely sensed images | |
CN107316309B (en) | Hyperspectral image saliency target detection method based on matrix decomposition | |
CN107194936B (en) | Hyperspectral image target detection method based on superpixel combined sparse representation | |
CN111680579B (en) | Remote sensing image classification method for self-adaptive weight multi-view measurement learning | |
US9036889B2 (en) | Method and apparatus for stain separation using vector analysis | |
CN111292266B (en) | GF-5 remote sensing image mixed noise removing method based on double low rank matrix decomposition | |
CN110097499B (en) | Single-frame image super-resolution reconstruction method based on spectrum mixing kernel Gaussian process regression | |
CN109271874B (en) | Hyperspectral image feature extraction method fusing spatial and spectral information | |
CN113446998B (en) | Hyperspectral target detection data-based dynamic unmixing method | |
CN110276782B (en) | Hyperspectral target tracking method combining spatial spectral features and related filtering | |
CN113421198A (en) | Hyperspectral image denoising method based on subspace non-local low-rank tensor decomposition | |
CN110135448B (en) | Hyperspectral image virtual dimension estimation method based on ridge ratio shrinkage | |
Antoniadis et al. | Peaks detection and alignment for mass spectrometry data | |
CN108875798B (en) | Super-pixel-level feature extraction method based on spatial pyramid pooling | |
CN116137067A (en) | Hyperspectral image anomaly detection method based on coefficient analysis collaborative representation | |
Falini et al. | Approximated iterative QLP for change detection in hyperspectral images | |
Dong et al. | A novel VHR image change detection algorithm based on image fusion and fuzzy C-means clustering | |
CN112329654B (en) | Hyperspectral image data classification method and system based on multi-manifold learning algorithm | |
Geng et al. | A novel color image segmentation algorithm based on JSEG and Normalized Cuts | |
Sigurdsson et al. | Total variation and ℓ q based hyperspectral unmixing for feature extraction and classification | |
Vafadar et al. | Hyperspectral anomaly detection using Modified Principal component analysis reconstruction error | |
Vargas et al. | Spectral unmixing approach in hyperspectral remote sensing: a tool for oil palm mapping |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |