CN108280486B - Hyperspectral image unmixing method based on end member cluster - Google Patents
Hyperspectral image unmixing method based on end member cluster Download PDFInfo
- Publication number
- CN108280486B CN108280486B CN201810104197.1A CN201810104197A CN108280486B CN 108280486 B CN108280486 B CN 108280486B CN 201810104197 A CN201810104197 A CN 201810104197A CN 108280486 B CN108280486 B CN 108280486B
- Authority
- CN
- China
- Prior art keywords
- end member
- cluster
- image
- spectrum
- pixel
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
- G06F18/2193—Validation; Performance evaluation; Active pattern learning techniques based on specific statistical tests
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A novel hyperspectral image unmixing method based on end member clusters achieves self-adaptive accurate unmixing of hyperspectral images through two steps of end member cluster extraction and end member cluster-based abundance inversion. The method comprises the following steps: (1) sparse representation based on a global image; (2) alternative end member screening based on voting; (3) constructing an end member cluster by extracting spectral shape characteristics; (4) constructing a block overcomplete dictionary; (5) traversing the whole image, and selecting the optimal end member based on block sparsity aiming at each pixel; (6) and performing abundance estimation by using a fully constrained least square method, and outputting end member cluster spectrum and abundance results. The method mainly aims at the problem of spectral variation existing in the hyperspectral image, can effectively reduce errors caused by the spectral variation, improves the unmixing precision, and meanwhile, the self-adaptability of the algorithm is strong due to the end member cluster extraction based on the image.
Description
Technical Field
The invention relates to a novel hyperspectral image unmixing method based on an end member cluster, which can effectively reduce errors caused by spectral variation during unmixing through technical means and belongs to the field of remote sensing image processing.
Background
The hyperspectral remote sensing refers to a remote sensing technology with hyperspectral resolution, and the detected waveband covers spectral regions (0.4-2.5 μm) including ultraviolet, visible light, near infrared, middle infrared and thermal infrared with spectral resolution of nanometer level. The hyperspectral remote sensing can continuously and finely describe the spectrum of the ground object and has outstanding advantages in the aspects of exploration, detection and identification. The development potential of hyperspectral remote sensing is huge, and the hyperspectral remote sensing and imaging radar are considered as two most important technical breakthroughs in remote sensing technology together since the 20 th century and the 80 th generation. Since the 90 s of the 20 th century, hyperspectral remote sensing has gradually become the mainstream direction of international photoelectric remote sensing and the hot topic of remote sensing technology. China is one of a few countries in the world with independent autonomous hyperspectral remote sensing technology intellectual property rights. In recent years, the development of hyperspectral remote sensing technology is greatly supported by the nation, and a lot of scientific researchers have gained remarkable results in the research fields of resource investigation, environmental monitoring, engineering construction, agricultural identification, medical diagnosis and the like under the support of the tasks of 863 projects, national natural science fund and the like.
Under the limitation of the spatial resolution of a spectrometer and the influence of the diversity of ground feature distribution, one pixel in a hyperspectral image often contains various ground features, and the pixel is called as a mixed pixel. In contrast to mixed pixels, we refer to pixels that contain only one type of terrain as clear pixels, or end members, while in mixed pixels the percentage of each substance is referred to as abundance. The purpose of hyperspectral unmixing is to acquire the characteristic spectrum of the end member in the image and then calculate the proportion of each ground feature in the mixed image element. The hyperspectral unmixing can solve the problems of ground object distribution detection and material identification caused by mixing pixels, so that the hyperspectral unmixing plays an important role in theoretical research and application of hyperspectral remote sensing images.
The existing hyperspectral unmixing algorithm generally represents an end member by calculating or selecting a spectral curve. However, there are limitations to representing an end member with only one spectral curve. In one image, due to factors of uneven illumination, different mineral particle size distribution and different contained organic matters and impurities, the spectrum of the same substance is easy to have a spectrum variation phenomenon. Spectral variations cause variations in the spectral values, which in turn lead to unmixing errors. In order to solve the problem of spectrum variation in the hyperspectral unmixing process in a complex environment, the invention provides a brand-new efficient and accurate hyperspectral unmixing model by utilizing the concept of an end member cluster, namely, representing a substance by using a group of spectra and combining a sparse representation theoretical model.
Disclosure of Invention
Aiming at the spectrum variation phenomenon of an image, the invention designs a brand-new hyperspectral unmixing method, which has the core of the method that the end member cluster is constructed in a self-adaptive manner and unmixing is carried out based on the end member cluster, and unmixing errors caused by the spectrum variation phenomenon are weakened through the omnibearing expression of the end member cluster to the end member. The method can directly extract the end member cluster from the image and unmix the end member cluster without using a spectrum library as prior information, has strong adaptability and wide applicable environment, and can effectively improve the unmixing precision.
In order to achieve the purpose, the invention adopts the technical scheme that: a brand-new hyperspectral unmixing method mainly comprises the following steps: extracting alternative end members, constructing end member clusters and estimating abundance. The purpose of the alternative end member extraction is to find out points with the properties of end members from the image, wherein the variation spectrum of each end member is contained. The construction of end member clusters requires the aggregation of spectra of the same kind under the unsupervised condition, and the invention completes the accurate spectrum classification by extracting the shape characteristics of the spectra. And finally, on the basis of the extracted end member clusters, finding out the optimal matching spectral combination for each pixel to carry out abundance estimation.
The method flow related by the invention comprises the following steps: (1) sparse representation based on a global image; (2) alternative end member screening based on voting; (3) constructing an end member cluster by extracting spectral shape characteristics; (4) constructing a block overcomplete dictionary; (5) traversing the whole image, and selecting the optimal end member based on block sparsity aiming at each pixel; (6) and performing abundance estimation by using a fully constrained least square method, and outputting end member cluster spectrum and abundance results.
The process steps of the process are described in detail below.
(1) Sparse representation based on global images
A hyperspectral image X with m rows and n columns and l wave bands is given, wherein X is { X ═ X }1,x2,...,xi,...,xm×nIn which xi={xi1,xi2,...,xil}. Then, each pixel is sparsely represented by a matching pursuit method, and the pixel x is aimed atiIs constructed asThe number of the material types in the image, namely the number of the end members is k, and the sparsity in the sparse representation is also set to be k. Sparse expression is carried out by utilizing a matching pursuit method to obtain a sparse coefficient yiThe method comprises the following specific steps:
2) Selecting the nearest to the residual, i.e. inner productMaximum atom, whose index λ is recordedtThen update the set of indices Λt=Λt-1∪{λtAnd Atom set Atomt=[Atomt-1;dλt]。
λt=arg max|<rt-1,Di>| (1)
3) Obtaining estimated coefficients by a least squares problem and updating the residual rt
rt=xi-εtAtomt(3)
4) Update t +1 and return to step 2) until the residual converges gradually or the iteration stops when t K.
5) Finally obtaining the sparse coefficient yi=εt。
(2) Alternative end member screening based on voting
For pixel xiAnd its corresponding sparse coefficient yiWe find the sparse coefficient yiThe image comprises m × n pixels, namely m × n votes are cast, the number of votes obtained by each atom is accumulated, all atoms are sorted according to the votes, the number cn of alternative end members is calculated according to the number k of the end members, wherein cn is k × 5, the atom with cn before the number of votes is selected as the alternative end member, and an alternative end member set X is obtainedcand。
(3) End member cluster construction by extracting spectral shape features
In the step, the characteristic extraction is carried out on the alternative end members obtained in the step 2, so that the aim of accurately constructing an end member cluster is fulfilled. For a spectrum containing l bandsThe spectrum is first cut into N spectra of length l0In which N ═ l/l0]. If l cannot be replaced by l0Dividing by the remaining wave length in the last part of the spectrumThe number of stages.
Then, a straight line is used for fitting each spectrum section, the slope of the straight line is extracted to represent the shape of the spectrum section, and then the slope vector of the whole curve is obtained
Least square fitting method for straight line fitting
That is, alternative end members X can be obtainedcandCharacteristic F ofcand=[f1;f2;...;fc...;fcn]The features are then input into an unsupervised classifier to obtain k classes Fbundle=[Fb1;Fb2...;Fbi;...;Fbk]. Then forming an end member cluster X by the corresponding alternative end members according to the classification resultbundle=[Xb1;Xb2...;Xbi;...Xbk],Xbi=[xh|fh∈Fbi]And outputs the result.
(4) Constructing a segmented overcomplete dictionary
For the obtained end member cluster Xbundle=[Xb1;Xb2...;Xbi;...Xbk]Each block of the overcomplete dictionary consists of a cluster of end-members, i.e., D ═ Db1;Db2;...;Dbi;...;DbK],Dbi=Xbi。
(5) Traversing the whole image, and selecting the optimal end member based on block sparsity for each pixel
For arbitrary pixel xiThe specific calculation steps are as follows:
1) initialization residual r0=xiSet of indicesTerminal unit setDictionary D0And D, setting the iteration number t to be 1 and the iteration termination number K.
2) Finding the exponent λ in the whole dictionarytSo that it satisfies (6)
λt=arg max|<rt-1,Dt-1>| (6)
3) Solving the optimization problem in (7) by using least square method, and updating residual rt。
4) If the residual is less than the set threshold, the iteration is stopped, otherwise t +1 is increased and 2 is returned) until t K.
(6) And performing abundance estimation by using a fully constrained least square method, and outputting end member cluster spectrum and abundance results.
Drawings
FIG. 1 is a flow chart of a hyperspectral image unmixing method based on end member clusters.
FIG. 2 is a schematic diagram of curve shape feature extraction: a) is the original curve, b) is the extracted feature.
FIG. 3 is the lunar surface hyperspectral data and the results obtained after unmixing, a) is a schematic diagram of the processed image and its position on the lunar surface, b) is the obtained end-member cluster spectrum, c) is the abundance distribution diagram of each substance obtained by unmixing.
Detailed Description
The method of application of the present invention is further illustrated below with reference to examples.
The hyperspectral data processed at this time is 300 × 128 in size, has a wavelength range of 480-.
(1) Sparse representation based on global images
An overcomplete dictionary of size 38400 × 20 was built and sparse representation was performed for each pixel, the sparsity of the sparse representation in this example was set to 3 and the number of candidate end-members was set to 15.
(2) Alternative end member screening based on voting
And voting the atom corresponding to the item with the maximum sparse coefficient for each pixel and the sparse coefficient obtained by the pixel. The votes of all atoms are counted, and the first 15 pixels are selected as alternative end members in the example.
(3) End member cluster construction by extracting spectral shape features
Segmenting the spectrum of the alternative end member, wherein the length of each segment is 2 wave bands in the example, then extracting the shape characteristics of the curve to obtain a characteristic vectorThe curve is fitted based on least square method, wherein the solution method of the slope coefficient is
After the slope characteristics are obtained, the extracted characteristics are grouped into 3 types by using a k-means method, and the construction of the end member cluster is completed.
(4) Constructing a segmented overcomplete dictionary
Constructing a block dictionary D ═ Db with the size of 15 × 20 by using the extracted 3-type end member cluster spectrum1;Db2;Db3]。
(5) Traversing the whole image, and selecting the optimal end member based on block sparsity for each pixel
For arbitrary pixel xi,
1) Initialization residual r0=xiSet of indicesTerminal unit setDictionary D0And D, setting the iteration number t to be 1 and the iteration termination number K.
2) Finding the exponent λ in the whole dictionarytSo that it satisfies (8)
λt=arg max|<rt-1,Dt-1>| (8)
3) Solving the optimization problem in (9) by using least square method, and updating residual error rt。
4) If the residual is less than the set threshold, the iteration is stopped, otherwise t +1 is increased and 2 is returned) until t K.
(6) And performing abundance estimation by using a fully constrained least square method, and outputting end member cluster spectrum and abundance results.
Claims (2)
1. An end member cluster extraction method is characterized in that: analyzing the relationship among pixels based on sparse representation, voting and screening alternative end members for the pixels by using a sparse coefficient, extracting spectral shape characteristics and constructing an end member cluster, wherein the method can realize automatic extraction of the end member cluster based on an image and comprises the following steps:
step 1, sparse representation based on global image
A hyperspectral image X with m rows and n columns and l wave bands is given, wherein X is { X ═ X }1,x2,...,xi,...,xm×nIn which xi={xi1,xi2,...,xil}; firstly, each pixel is sparsely represented by a matching pursuit method aiming at a pixel xiIs constructed asThe number of the material types in the image, namely the number of the end members is k, the sparsity in the sparse representation is also set as k, then the sparse representation is carried out by utilizing a matching tracking method, and a sparse coefficient y is obtainediThe sparse coefficient solving method comprises the following specific steps:
2) selecting the atom closest to the residual, i.e. with the largest inner product, and recording the index lambda thereoftThen updating the exponent set Λt=Λt-1∪{λtAnd Atom set Atomt=[Atomt-1;dλt];
λt=arg max|<rt-1,Di>| (1)
3) Obtaining estimated sparseness by a least squares problem and updating residual rt;
rt=xi-εtAtomt(3)
4) Updating t +1 and returning to step 2), stopping iteration until the residual converges gradually or t K;
5) finally obtaining the sparse coefficient yi=εt;
Step 2, alternative end member screening based on voting
For pixel xiAnd its corresponding sparse coefficient yiWe find the sparse coefficient yiThe method comprises the steps of selecting atoms with maximum absolute values, voting, leading an image to comprise m × n pixels, namely casting m × n votes, accumulating the number of votes obtained by each atom, sequencing all atoms according to the votes, calculating the number cn of alternative end members by the number k of the end members, wherein the cn is k × 5, selecting the atom with cn before the number of votes as the alternative end member, and obtaining an alternative end member set Xcand;
Step 3, constructing end member clusters by extracting spectral shape characteristics
In the step, the characteristic extraction is carried out on the alternative end members obtained in the step 2 so as to achieve the aim of accurately constructing an end member cluster; for a spectrum containing l bandsThe spectrum is first cut into N spectra of length l0In which N ═ l/l0](ii) a If l cannot be replaced by l0Dividing, the length of the last section of the spectrum is the number of the rest wave bands;
then, a straight line is used for fitting each spectrum section, the slope of the straight line is extracted to represent the shape of the spectrum section, and then the slope vector of the whole curve is obtained
Least square fitting method for straight line fitting
That is, alternative end members X can be obtainedcandCharacteristic F ofcand=[f1;f2;...;fc...;fcn](ii) a The features are then input into an unsupervised classifier to obtain k classes Fbundle=[Fb1;Fb2...;Fbi;...;Fbk]Then, the corresponding alternative end members are formed into an end member cluster X according to the classification resultbundle=[Xb1;Xb2...;Xbi;...Xbk],Xbi=[xh|fh∈Fbi]And outputs the result.
2. An abundance estimation method based on end-member clusters is characterized in that: the method is characterized in that a known end member cluster spectrum library is utilized, a block sparse thought is combined, a most matched end member combination is selected for each pixel, and inversion is carried out by utilizing a fully constrained least square method to obtain abundance distribution, wherein the method is different from the traditional method that only a single end member is extracted for one kind of ground, a plurality of candidate end members are extracted for one kind of ground to form a cluster, namely an end member cluster, and the abundance estimation method based on the end member cluster comprises the following specific steps:
step 1, constructing a block overcomplete dictionary
For the obtained end member cluster Xbundle=[Xb1;Xb2...;Xbi;...Xbk]Each block of the overcomplete dictionary consists of a cluster of end-members, i.e., D ═ Db1;Db2;...;Dbi;...;DbK],Dbi=Xbi;
Step 2, traversing the whole image, and selecting the optimal end member based on block sparsity aiming at each pixel
For arbitrary pixel xiThe specific calculation steps are as follows:
1) initialization residual r0=xiSet of indicesTerminal unit setDictionary D0Setting the iteration time t as 1 and the iteration termination time K as D;
2) finding the exponent λ in the whole dictionarytSo that it satisfies (6)
λt=arg max|<rt-1,Dt-1>| (6)
3) Solving the optimization problem in (7) by using least square method, and updating residual rt;
4) If the residual error is smaller than a set threshold value, stopping iteration, otherwise, increasing t to t +1, and returning to 2) until t to K;
and 3, carrying out abundance estimation by using a fully constrained least square method, and outputting end member cluster spectra and abundance results.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810104197.1A CN108280486B (en) | 2018-02-02 | 2018-02-02 | Hyperspectral image unmixing method based on end member cluster |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810104197.1A CN108280486B (en) | 2018-02-02 | 2018-02-02 | Hyperspectral image unmixing method based on end member cluster |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108280486A CN108280486A (en) | 2018-07-13 |
CN108280486B true CN108280486B (en) | 2020-06-23 |
Family
ID=62807325
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810104197.1A Active CN108280486B (en) | 2018-02-02 | 2018-02-02 | Hyperspectral image unmixing method based on end member cluster |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108280486B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109696406B (en) * | 2018-11-29 | 2020-10-02 | 北京航空航天大学 | Moon table hyperspectral image shadow region unmixing method based on composite end member |
CN109840544B (en) * | 2018-12-19 | 2023-05-16 | 国网浙江省电力有限公司嘉兴供电公司 | Hyperspectral image multi-end member spectrum hybrid analysis method and device |
CN112733867B (en) * | 2021-02-04 | 2023-08-08 | 大连民族大学 | End member extraction method of hyperspectral image |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102314685A (en) * | 2011-07-23 | 2012-01-11 | 北京航空航天大学 | Hyperspectral image sparse unmixing method based on random projection |
CN103020912A (en) * | 2012-12-18 | 2013-04-03 | 武汉大学 | Remote sensing image restoration method combining wave-band clustering with sparse representation |
CN104036509A (en) * | 2014-06-19 | 2014-09-10 | 哈尔滨工业大学 | Method for unmixing hyperspectral mixed pixel based on compressed sensing |
CN105320959A (en) * | 2015-09-30 | 2016-02-10 | 西安电子科技大学 | End member learning based hyperspectral image sparse unmixing method |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7505608B2 (en) * | 2005-04-15 | 2009-03-17 | The Boeing Company | Methods and apparatus for adaptive foreground background analysis |
-
2018
- 2018-02-02 CN CN201810104197.1A patent/CN108280486B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102314685A (en) * | 2011-07-23 | 2012-01-11 | 北京航空航天大学 | Hyperspectral image sparse unmixing method based on random projection |
CN103020912A (en) * | 2012-12-18 | 2013-04-03 | 武汉大学 | Remote sensing image restoration method combining wave-band clustering with sparse representation |
CN104036509A (en) * | 2014-06-19 | 2014-09-10 | 哈尔滨工业大学 | Method for unmixing hyperspectral mixed pixel based on compressed sensing |
CN105320959A (en) * | 2015-09-30 | 2016-02-10 | 西安电子科技大学 | End member learning based hyperspectral image sparse unmixing method |
Also Published As
Publication number | Publication date |
---|---|
CN108280486A (en) | 2018-07-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111860612B (en) | Unsupervised hyperspectral image hidden low-rank projection learning feature extraction method | |
CN110321963B (en) | Hyperspectral image classification method based on fusion of multi-scale and multi-dimensional space spectrum features | |
US10691974B2 (en) | Superpixel classification method based on semi-supervised K-SVD and multiscale sparse representation | |
WO2017190542A1 (en) | Partition-based vca end-member extraction method | |
CN109978041B (en) | Hyperspectral image classification method based on alternative updating convolutional neural network | |
Sun et al. | UL-Isomap based nonlinear dimensionality reduction for hyperspectral imagery classification | |
Schulz et al. | Learning Object-Class Segmentation with Convolutional Neural Networks. | |
CN110533077B (en) | Shape adaptive convolution depth neural network method for hyperspectral image classification | |
CN110619263B (en) | Hyperspectral remote sensing image anomaly detection method based on low-rank joint collaborative representation | |
CN108280486B (en) | Hyperspectral image unmixing method based on end member cluster | |
CN108734199B (en) | Hyperspectral image robust classification method based on segmented depth features and low-rank representation | |
CN111914909B (en) | Hyperspectral change detection method based on space-spectrum combined three-direction convolution network | |
Aggarwal et al. | Segmentation of crop images for crop yield prediction | |
Zhang et al. | Object-oriented subspace analysis for airborne hyperspectral remote sensing imagery | |
CN112396029A (en) | Clustering segmentation and coupling end member extraction synergistic hyperspectral coastal wetland subpixel change detection method | |
Mukhtar et al. | Wheat plant counting using UAV images based on semi-supervised semantic segmentation | |
Haq et al. | 3D-CNNHSR: A 3-Dimensional Convolutional Neural Network for Hyperspectral Super-Resolution. | |
Guo et al. | CNN‐combined graph residual network with multilevel feature fusion for hyperspectral image classification | |
Bao et al. | Method for wheat ear counting based on frequency domain decomposition of MSVF-ISCT | |
CN109271874B (en) | Hyperspectral image feature extraction method fusing spatial and spectral information | |
Hsu et al. | Hyperspectral image classification via joint sparse representation | |
Zhang et al. | Spatially constrained Bag-of-Visual-Words for hyperspectral image classification | |
CN110852304B (en) | Hyperspectral data processing method based on deep learning method | |
Schulz et al. | Object-class segmentation using deep convolutional neural networks | |
Liu et al. | Remote sensing image classification based on convolutional neural networks with two-fold sparse regularization |
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 |