CN113486876A - Hyperspectral image band selection method, device and system - Google Patents
Hyperspectral image band selection method, device and system Download PDFInfo
- Publication number
- CN113486876A CN113486876A CN202111050606.2A CN202111050606A CN113486876A CN 113486876 A CN113486876 A CN 113486876A CN 202111050606 A CN202111050606 A CN 202111050606A CN 113486876 A CN113486876 A CN 113486876A
- Authority
- CN
- China
- Prior art keywords
- hyperspectral image
- similarity
- hyperspectral
- super
- 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.)
- Pending
Links
- 238000010187 selection method Methods 0.000 title claims abstract description 46
- 230000003595 spectral effect Effects 0.000 claims abstract description 30
- 238000009792 diffusion process Methods 0.000 claims abstract description 24
- 230000004927 fusion Effects 0.000 claims abstract description 16
- 238000000034 method Methods 0.000 claims abstract description 16
- 238000004422 calculation algorithm Methods 0.000 claims description 26
- 230000011218 segmentation Effects 0.000 claims description 16
- 238000004590 computer program Methods 0.000 claims description 11
- 239000013598 vector Substances 0.000 claims description 7
- 238000001228 spectrum Methods 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 3
- 238000003709 image segmentation Methods 0.000 claims description 3
- 238000012847 principal component analysis method Methods 0.000 claims 1
- 238000007499 fusion processing Methods 0.000 abstract description 4
- 238000010586 diagram Methods 0.000 description 10
- 230000006870 function Effects 0.000 description 8
- 238000000513 principal component analysis Methods 0.000 description 8
- 235000008331 Pinus X rigitaeda Nutrition 0.000 description 5
- 235000011613 Pinus brutia Nutrition 0.000 description 5
- 241000018646 Pinus brutia Species 0.000 description 5
- 239000011159 matrix material Substances 0.000 description 5
- 241001466077 Salina Species 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 238000012549 training Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 208000037540 Alveolar soft tissue sarcoma Diseases 0.000 description 2
- 208000008524 alveolar soft part sarcoma Diseases 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000012880 independent component analysis Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000002310 reflectometry Methods 0.000 description 2
- 230000035945 sensitivity Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000010191 image analysis Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
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
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a hyperspectral image band selection method, device and system. The method comprises the following steps: acquiring a hyperspectral image cube, and generating a plurality of superpixels according to the hyperspectral image cube; for each super pixel, constructing a similarity map reflecting the similarity between the wave bands contained in the super pixel; generating a unified similar graph according to all similar graphs by adopting a multi-graph diffusion fusion strategy; performing spectral clustering on original hyperspectral data according to the unified similarity map to obtain a plurality of hyperspectral subcubes; the band with the smallest noise value is selected from each hyperspectral sub-cube as the characteristic band to determine the optimal band subset. The spatial and spectral information of each super-pixel is embedded in the fusion process, so that the super-pixels with higher similarity have higher contribution rate, and the ground feature information of different areas in the hyperspectral image can be fused, so that the interrelation among different wave bands can be described more accurately, and further the selection of the hyperspectral image wave bands can be realized more accurately.
Description
Technical Field
The invention relates to the technical field of hyperspectral remote sensing image band selection, in particular to a hyperspectral image band selection method, device and system.
Background
Compared with the traditional RGB image, the hyperspectral remote sensing image records the reflectivity of a target scene by using different electromagnetic waves, and the land coverage information is more abundant because the hyperspectral remote sensing image contains hundreds of wave bands. The hyperspectral dimension reduction methods mainly comprise two modes, namely feature extraction and wave band selection.
In the existing band selection method, clustering is widely applied to hyperspectral band selection as a common unsupervised band selection strategy in the past years, but most of the current clustering algorithms still have two problems. Firstly, they only consider the correlation between adjacent bands and ignore the global information between all bands; secondly, most of the conventional clustering algorithms take a certain waveband as a whole, stretch the waveband and reshape the waveband into a feature vector, and the way ignores that different ground objects have different spectral characteristics. The above problem results in the current algorithm being inaccurate for describing the interrelationship between different bands.
Disclosure of Invention
The invention solves the problem of how to more accurately describe the mutual relation among different wave bands so as to more accurately realize the selection of the wave bands of the hyperspectral image.
In order to solve the above problems, the present invention provides a method for selecting a hyperspectral image band, comprising: acquiring a hyperspectral image cube, and generating a plurality of superpixels according to the hyperspectral image cube; for each super pixel, constructing a similarity map reflecting the similarity between the wave bands contained in the super pixel; generating a unified similar graph according to all the similar graphs by adopting a multi-graph diffusion fusion strategy; performing spectral clustering on original hyperspectral data according to the unified similarity map to obtain a plurality of hyperspectral subcubes; and selecting the wave band with the minimum noise value from each high spectrum sub-cube as a characteristic wave band to determine an optimal wave band subset.
According to the hyperspectral image band selection method, the spatial and spectral information of each superpixel is embedded in the fusion process, so that the superpixel with higher similarity has higher contribution rate, and the ground feature information of different areas in the hyperspectral image can be fused, so that the interrelation among different bands can be more accurately described, and further, the hyperspectral image band selection can be more accurately realized.
Optionally, the generating a plurality of superpixels from the hyperspectral imagery cube comprises: and extracting a first principal component of the hyperspectral image cube by PCA (principal component analysis), and dividing the first principal component into a plurality of superpixels by superpixel division.
According to the hyperspectral image band selection method, the first main component of the original hyperspectral image cube is divided into a plurality of uniform areas through the superpixel division algorithm, and the distinguishing information of different ground objects can be considered, so that the similarity among the bands can be described more accurately, and the hyperspectral image band selection can be realized more accurately.
Optionally, the super-pixel segmentation adopts an ERS entropy rate super-pixel segmentation algorithm, and the ERS entropy rate super-pixel segmentation algorithm segments the first principal component by using a first formula, where the first formula includes:
whereinA first principal component representing the hyperspectral imagery cube; n represents the number of the super pixels;andrespectively representing the p-th and q-th superpixels.
According to the hyperspectral image band selection method, the first main component is segmented through the ERS entropy rate superpixel segmentation algorithm, and the distinguishing information of different ground objects can be considered, so that the similarity among bands can be described more accurately, and the hyperspectral image band selection can be realized more accurately.
Optionally, the constructing a similarity map reflecting similarities between bands included in each of the superpixels includes: determining the similarity between the bands contained in the superpixels according to a second formula, wherein the second formula comprises:
whereinRepresenting a similarity between an ith band and a jth band;andrespectively representing the ith and jth bands in the pth super-pixel;andrespectively representAnda set of k nearest neighbor bands;is a width parameter in the gaussian kernel function.
According to the hyperspectral image band selection method, the similarity graphs reflecting the similarity between the bands contained in the segmented super pixels are constructed according to the second formula, and then the uniform similarity graphs reflecting the similarity of different hyperspectral bands can be obtained.
Optionally, the obtaining a unified similarity graph according to the similarity graph by using a multi-graph diffusion fusion strategy includes:
updating and diffusing the similarity graph according to a third formula, wherein the third formula comprises:
wherein,is a balance parameter;representing the contribution rate of the similar graph corresponding to the qth super pixel to the similar graph corresponding to the pth super pixel;andrespectively representing the corresponding similar graphs of the qth super-pixel and the pth super-pixel in the t iteration;representing a similar graph corresponding to the p-th super pixel in the t +1 th iteration; a. thet pShowing the normalized similarity graph corresponding to the pth superpixel region at time t.
According to the hyperspectral image band selection method, the similar graphs are updated and diffused through the third formula to obtain the unified similar graph, and then spectral clustering can be performed according to the unified similar graph to construct the optimal band subset, so that the interrelation among different bands can be described more accurately, and the hyperspectral image band selection can be realized more accurately.
wherein,andrespectively representing the p-th and q-th feature vectors of the first principal component of the hyperspectral image cube;andrespectively representing the central pixel positions of the p-th and q-th superpixels in the first main component of the hyperspectral image cube;anda gaussian kernel function representing a distance metric; and N is the number of the super pixels.
According to the hyperspectral image band selection method, the spectrum similarity and the space similarity of the non-superpixel are used for weighting the diffusion process, the complementarity among a plurality of similar graphs can be better utilized, so that the mutual relation among different bands can be more accurately described, and further, the hyperspectral image band selection can be more accurately realized.
Optionally, the unified similarity map is determined by a fifth formula, the fifth formula comprising:
whereinRepresenting a unified similarity graph;a corresponding similar graph of the pth super pixel at the Tth iteration; and N is the number of the super pixels.
According to the hyperspectral image band selection method, the unified similar graph reflecting the similarity of different hyperspectral bands is obtained according to the fifth formula, and then spectral clustering can be carried out according to the unified similar graph so as to construct the optimal band subset, so that the interrelation among different bands can be described more accurately, and the hyperspectral image band selection can be realized more accurately.
The invention also provides a hyperspectral image band selection device, which comprises: the hyperspectral image segmentation module is used for acquiring a hyperspectral image cube and generating a plurality of superpixels according to the hyperspectral image cube; a super-pixel similarity map construction module, configured to construct, for each super-pixel, a similarity map reflecting similarity between bands included in the super-pixel; the similar graph fusion module is used for generating a unified similar graph according to all the similar graphs by adopting a multi-graph diffusion fusion strategy; the spectral clustering module is used for performing spectral clustering on the original hyperspectral data according to the unified similar graph to obtain a plurality of hyperspectral subcubes; and the hyperspectral waveband selecting module is used for selecting a waveband with the minimum noise value from each hyperspectral sub-cube as a characteristic waveband so as to determine an optimal waveband subset. Compared with the prior art, the hyperspectral image band selection device and the hyperspectral image band selection method have the same advantages, and are not repeated herein.
The invention also provides a hyperspectral image band selection system which comprises a computer readable storage medium and a processor, wherein the computer readable storage medium is stored with a computer program, and the computer program is read by the processor and runs to realize the hyperspectral image band selection method. Compared with the prior art, the hyperspectral image band selection system and the hyperspectral image band selection method have the same advantages, and are not repeated herein.
The invention also provides a computer readable storage medium, which stores a computer program, and when the computer program is read and executed by a processor, the hyperspectral image band selection method is realized. The advantages of the computer-readable storage medium and the hyperspectral image band selection method over the prior art are the same, and are not described herein again.
Drawings
FIG. 1 is a flowchart of a hyperspectral image band selection method according to an embodiment of the invention;
FIG. 2 is a block diagram of a hyperspectral image band selection method according to an embodiment of the invention;
FIG. 3 is a schematic diagram of an Overall-Accuracy curve of different hyperspectral image band selection methods on an Indian Pines dataset according to an embodiment of the present invention;
FIG. 4 is a second schematic diagram of an Overall-Accuracy curve of a different hyperspectral image band selection method on an Indian Pines dataset according to an embodiment of the present invention;
FIG. 5 is a third schematic diagram of an Overall-Accuracy curve of a different hyperspectral image band selection method on an Indian Pines dataset according to an embodiment of the present invention;
FIG. 6 is a schematic diagram of an Overall-Accuracy curve of a different hyperspectral image band selection method on a Pavia University dataset according to an embodiment of the invention;
FIG. 7 is a second schematic diagram of an Overall-Accuracy curve on a Pavia University dataset according to a different hyperspectral image band selection method of the embodiment of the invention;
FIG. 8 is a third schematic diagram of an Overall-Accuracy curve on a Pavia University dataset according to a different hyperspectral image band selection method of the embodiment of the invention;
FIG. 9 is a schematic diagram of an Overall-Accuracy curve on a Salinas dataset according to a different hyperspectral image band selection method in an embodiment of the invention;
FIG. 10 is a second schematic diagram of an Overall-Accuracy curve on a Salinas dataset according to a different hyperspectral image band selection method of the embodiment of the invention;
FIG. 11 is a third schematic diagram of an Overall-Accuracy curve on a Salinas dataset according to a different hyperspectral image band selection method of the embodiment of the invention.
Detailed Description
With the rapid development of the hyperspectral remote sensing imaging technology, a large amount of hyperspectral remote sensing image data can be easily obtained. Compared with the traditional RGB image, the hyperspectral remote sensing image records the reflectivity of a target scene by using different electromagnetic waves, and the land coverage information is more abundant because the hyperspectral remote sensing image contains hundreds of wave bands. Based on the above, the hyperspectral remote sensing image is widely applied to the fields of environment detection, vegetation coverage estimation, landslide monitoring and the like. However, since the original hyperspectral image cube is mixed with non-negligible redundant information and noise, this not only degrades the information expression capability thereof, but also causes the computational complexity of the subsequent hyperspectral image analysis to become high. Therefore, it is necessary to denoise it and reduce redundant information.
In the past decades, there are two main ways of hyperspectral dimension reduction, which are feature extraction and band selection. For the former, the original high-dimensional spatial data is usually projected into the new low-dimensional feature space according to a certain mathematical transformation criterion. The transformation criteria are mainly used to learn more distinctive features in a new feature space, and typical methods are Linear Discriminant Analysis (LDA), Principal Component Analysis (PCA), Independent Component Analysis (ICA), and maximum noise ratio (MNF). However, the physical meaning of the new feature space obtained by feature projection is often not well defined and is not explanatory. In the latter case, some representative bands are selected from the raw data to represent the whole hyperspectral data set. In this way the intrinsic properties of the band and its physical significance can be well preserved. We are mainly concerned here with band selection.
According to the separability of the sample class labels, the existing band selection methods can be further divided into supervised band selection and unsupervised band selection. For the former, they require sample label training classifiers to select the optimal band. In addition, it is often difficult to design a suitable training regimen and the labels for the samples are not readily available. Compared to the supervised band selection method, the unsupervised band selection method is more flexible because it does not require a sample tag. In addition, for the unsupervised waveband selection method, a learning model with a certain standard can be constructed only by some specific priori knowledge of original hyperspectral data, such as information divergence, sample similarity, maximum ellipsoid volume and the like.
In unsupervised hyperspectral band selection, clustering-based and ranking-based are two common selection strategies. The hyperspectral waveband selection method based on ranking mainly comprises two steps of evaluating the importance of each waveband and then selecting the most important waveband from the importance of each waveband as a characteristic waveband. Typical evaluation criteria include information entropy, information divergence and maximum variance. Clustering has been widely used in hyperspectral band selection as another common unsupervised band selection strategy in the past few years, and has achieved satisfactory results. The clustering-based band selection method firstly divides all original bands into a plurality of groups, and then selects characteristic bands from each group, thereby constructing an optimal band subset. Therefore, the quality of the clustering performance is crucial to the final result.
However, most current clustering algorithms still have two problems. Firstly, they only consider the correlation between adjacent bands and ignore the global information between all bands; secondly, most of the conventional clustering algorithms take a certain waveband as a whole, stretch the waveband and reshape the waveband into a feature vector, and the way ignores that different ground objects have different spectral characteristics. Therefore, it is not appropriate to treat each band directly as a single feature vector. In order to solve the above problems, the present invention proposes the following solutions.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in detail below.
As shown in fig. 1, an embodiment of the present invention provides a method for selecting a hyperspectral image band, including: acquiring a hyperspectral image cube, and generating a plurality of superpixels according to the hyperspectral image cube; for each super pixel, constructing a similarity map reflecting the similarity between the wave bands contained in the super pixel; generating a unified similar graph according to all the similar graphs by adopting a multi-graph diffusion fusion strategy; performing spectral clustering on original hyperspectral data according to the unified similarity map to obtain a plurality of hyperspectral subcubes; and selecting the wave band with the minimum noise value from each high spectrum sub-cube as a characteristic wave band to determine an optimal wave band subset.
Specifically, the spectral resolution is 10-2The spectral images within the λ order range are called hyperspectral images, and in this embodiment, the hyperspectral image band selection method is entirely called a hyperspectral image band selection method based on spectral clustering and spatial-spectral weighting multi-region fusion, which is abbreviated as RMGF in english. Referring to fig. 2, first, a first principal component of an original hyperspectral image cube is divided into a plurality of uniform regions, and for all the divided superpixels, similarity maps are constructed respectively to reflect the similarity between bands. Then, a unified similar graph of the whole hyperspectral is learned through a theoretically converged multi-graph diffusion fusion strategy. In the fusion process, the super-pixels with higher similarity have larger contribution rate by embedding the spatial and spectral information of each super-pixel. And finally, performing spectral clustering on the original hyperspectral data by using the learned uniform similarity map to obtain a plurality of divided hyperspectral subcubes, and selecting a wave band with the minimum noise value from each subcube as a characteristic wave band to construct an optimal wave band subset. Through the mode, the ground feature information of different areas in the hyperspectral image can be fused, so that the mutual relation among different wave bands can be more accurately described.
In order to verify the validity of the proposed RMGF algorithm in hyperspectral band selection, a number of correlation experiments were performed, as illustrated in fig. 3 to 11.
The experiment is mainly carried out on three public data setsAre the Indian Pines dataset, the Pavia University dataset, and the Salinas dataset, respectively. To evaluate the performance of RMGF and other competitors, we selected three classifiers, including KNN, SVM, and LDA. In the experiment, the value of the parameter k of the KNN classifier is set to be 3; the SVM uniformly adopts a radial basis kernel function, and the corresponding penalty coefficient and gamma value are respectively set asAnd 0.5. For each classifier, 10% of the entire data set was randomly selected as training samples, the rest as test samples. Furthermore, since the number of best selected bands per data set is unknown, bands are selected at 5 intervals in the range of 5-50. In the experiment, each algorithm was performed 10 times and the average results were presented, reducing the impact of randomly selecting training and testing samples. To verify the validity of the proposed algorithm, three different classifiers were used to compare it with the current state-of-the-art algorithm. By selecting the optimum number of selected bands, the overall accuracy of the different methods over three different data sets is shown in table 1 below. With the best results highlighted in bold. It can be seen that the RMGF algorithm consistently outperforms other competitors on different classifiers for all datasets.
The other algorithms are shown below.
(1) UBS: band selection based on equal band number splitting.
(2) TOF: band selection based on optimized clustering.
(3) ASPS _ MN: based on band selection to minimize noise.
(4) ASPS _ IE: and selecting a waveband based on the information entropy.
(5) FNGBS: band selection based on fast neighbor clustering.
(6) ONR: band selection based on optimal neighborhood reconstruction (not shown in table 1).
TABLE 1
For the Indian Pines dataset, the RMGF algorithm achieved an improvement of over 7% accuracy over the other algorithms. RMGF may achieve an improvement of accuracy of more than 10%, especially when LDA is used as classifier. For the Pavia University dataset, RMGF also yielded over 5% improvement over other methods when LDA was used as the classifier.
As previously mentioned, in practical applications, it is difficult to determine the number of bands that different data sets are desired to select. The OA curves for different numbers of selected bands are plotted for each data set in fig. 3-11 (where the abscissa represents the number of different bands ultimately selected and the ordinate represents the overall accuracy of the hyperspectral image classification from the corresponding selected band subset). It can be seen that the RMGF method outperforms other algorithms in most of the number of bands selected. As shown in fig. 3-8, RMGF always achieves higher OA for KNN and LDA classifiers over the selected band. In fig. 9 to 11, RMGF is always able to select more distinct bands for different classifiers.
In this embodiment, the spatial and spectral information of each super-pixel is embedded in the fusion process to enable the super-pixel with higher similarity to have higher contribution rate, and the ground feature information of different areas in the hyperspectral image can be fused, so that the interrelation among different wave bands can be described more accurately, and further the selection of the wave bands of the hyperspectral image can be realized more accurately.
Optionally, the generating a plurality of superpixels from the hyperspectral imagery cube comprises: and extracting a first principal component of the hyperspectral image cube by PCA (principal component analysis), and dividing the first principal component into a plurality of superpixels by superpixel division.
Specifically, in this embodiment, a PCA (principal component analysis) is used to extract a first principal component of the hyperspectral image cube, and a superpixel segmentation algorithm is used to segment the first principal component of the original hyperspectral image cube into a plurality of uniform regions in consideration that different targets have different reflection characteristics. In this embodiment, a hyper-pixel segmentation method is used to segment the first main component of the hyper-spectrum into a plurality of small regions, and different regions represent different surface feature information, so that a similarity map between bands can be respectively constructed for each region, and then a plurality of similarity maps are fused to obtain an overall band similarity map of the spectral image. Compared with the traditional method of considering each wave band as a whole, the method can give consideration to the distinguishing information of different ground objects, so that the similarity among the wave bands can be more accurately described.
In the embodiment, the first principal component of the original hyperspectral image cube is divided into a plurality of uniform areas through a superpixel division algorithm, and the distinguishing information of different ground objects can be considered, so that the similarity among the wave bands can be described more accurately, and the selection of the hyperspectral image wave bands can be realized more accurately.
Optionally, the super-pixel segmentation adopts an ERS entropy rate super-pixel segmentation algorithm, and the ERS entropy rate super-pixel segmentation algorithm segments the first principal component by using a first formula, where the first formula includes:
whereinA first principal component representing the hyperspectral imagery cube; n represents the number of the super pixels;andrespectively representing the p-th and q-th superpixels.
Specifically, in this embodiment, the super-pixel segmentation adopted in this embodiment is an ERS entropy rate super-pixel segmentation algorithm, and the segmentation specifically includes the following steps:
whereinRepresenting a first principal component of a hyperspectral image cube; n represents the number of superpixels;andrespectively representing a pth super pixel and a qth super pixel; s.t. representation is limited to, i.e. a constraint.
In the embodiment, the first principal component is segmented through the ERS entropy rate superpixel segmentation algorithm, and the information for distinguishing different ground objects can be considered, so that the similarity among wave bands can be more accurately described, and the selection of the wave bands of the hyperspectral image can be more accurately realized.
Optionally, the constructing a similarity map reflecting similarities between bands included in each of the superpixels includes: determining the similarity between the bands contained in the superpixels according to a second formula, wherein the second formula comprises:
whereinRepresenting a similarity between an ith band and a jth band;andrespectively representing the ith and jth bands in the pth super-pixel;andrespectively representAnda set of k nearest neighbor bands;is a width parameter in the gaussian kernel function.
Specifically, in this embodiment, when constructing the similarity map, the similarity map reflecting the similarity between the bands included in the segmented superpixels is constructed for each of the segmented superpixels, and the specific calculation method is as follows:
whereinRepresenting a similarity between an ith band and a jth band;andrespectively representing the ith and jth bands in the pth super-pixel;andrespectively representAnda set of k nearest neighbor bands;is a width parameter in the gaussian kernel function.
In this embodiment, a similarity graph reflecting similarity between bands included in the segmented plurality of superpixels is constructed according to a second formula, and then a unified similarity graph reflecting similarity between different hyperspectral bands can be obtained.
Optionally, the obtaining a unified similarity graph according to the similarity graph by using a multi-graph diffusion fusion strategy includes:
updating and diffusing the similarity graph according to a third formula, wherein the third formula comprises:
wherein,is a balance parameter;representing the contribution rate of the similar graph corresponding to the qth super pixel to the similar graph corresponding to the pth super pixel;andrespectively representing the corresponding similar graphs of the qth super-pixel and the pth super-pixel in the t iteration;representing a similar graph corresponding to the p-th super pixel in the t +1 th iteration; a. thet pShowing the normalized similarity graph corresponding to the pth superpixel region at time t.
Specifically, in this embodiment, there may be a certain deviation in the initial similarity graph corresponding to each super pixel obtained according to the second formula, and in general, a plurality of super pixels may represent the same or similar target objects. In this regard, the present embodiment designs a diffusion fusion strategy, and updates the initial similarity graph by using the supplementary information implied in the different similarity graphs.
For convenience of explanation, a diffusion updating method for the similarity of nodes of a single similar graph is introduced, and then the method is extended to the case of multi-similar graph diffusion updating. According to the popularity ranking theory, an initial similarity graph representing the similarity between M sample points is obtainedThe update of the similarity between the nodes can be expressed as the following optimization problem:
wherein M represents the number of nodes of the similar graph;representing the similarity between the ith and jth nodes, SpqRepresenting the similarity between the p-th node and the q-th node; d is a degree matrix corresponding to the similarity graph S,represents its ith row and ith column element, its value is,DjjDenotes the jth row and jth column element, DqqRepresents the q-th row and q-th column elements;representing the updated similarity graph;the similarity between the ith node and the jth node in the updated similarity graph is obtained;is a regularization parameter. As can be seen from the above formula, it is mainly composed of two terms. The first term is the analogy constraint of local similarity and global consistency, i.e. in the original similarity graph, if nodes i, j are similar and nodes p, q are similar, then the new learned similarity point isAndshould also be similar; the second term of the above equation can be regarded as a fitting term that constrains the learned new similarity graph from changing too much from the initial state.
For the objective function proposed above, it has a solution of the form:
wherein S is a similar graph before updating;is the updated similarity graph;;means an operator for vectorizing each column superposition of the input matrix, and the corresponding inverse operator is ;Is an identity matrix;is the kronecker product of A, i.e.,Is a Laplace matrix having values of. In the process of iteratively updating the similarity graph, the similarity propagation can be expressed as:
where t represents the tth iteration. With the above equation, it is a disadvantage that only a single similar graph can be updated at a time.
Based on the above problems, the present embodiment designs a multi-image diffusion fusion method to update multiple similar images of a hyperspectral image, thereby fully utilizing complementarity between different similar images. Considering that in a certain wave band of a hyperspectral image, superpixels with similar spectral values or close spatial positions have great probability of representing the same target. Therefore, in the diffusion process, spectral values are close or spatially located and should have a higher contribution ratio than similar maps corresponding to neighboring superpixels. In this regard, we have devised an update diffusion process that resembles the following graph:
wherein,is a balance parameter;representing the p-th superpixel corresponding to the q-th superpixelThe contribution rate of the similar graph corresponding to the pixel;andrespectively representing the corresponding similar graphs of the qth super-pixel and the pth super-pixel in the t iteration;representing a similar graph corresponding to the p-th super pixel in the t +1 th iteration; a. thet pThe normalized similarity graph corresponding to the p-th super-pixel area at the time t is defined as,Is composed ofThe corresponding diagonal matrix, the ith diagonal element of which is defined as:. Wherein B is the wave band number of the original hyperspectral data.
By using the above-mentioned update diffusion of the similarity maps of all superpixels, we can iteratively exchange the connection information in different maps, thereby obtaining the final unified similarity map. On one hand, similarity values of different similarity graphs are mutually propagated through an iterative diffusion process; on the other hand, parametersPart of the information of the original similar graph is retained. Moreover, by weighting the diffusion process without using spectral and spatial similarities of superpixels, the complementarity between multiple similar maps can be better exploited. After diffusion t times, the corresponding similarity graph of each super pixel can be updated to a stable state.
In this embodiment, the similarity map is updated and diffused through the third formula to obtain a unified similarity map, and then spectral clustering can be performed according to the unified similarity map to construct an optimal waveband subset, so that the interrelation among different wavebands can be more accurately described, and further hyperspectral image waveband selection can be more accurately realized.
wherein,andrespectively representing the p-th and q-th feature vectors of the first principal component of the hyperspectral image cube;andrespectively representing the central pixel positions of the p-th and q-th superpixels in the first main component of the hyperspectral image cube;anda gaussian kernel function representing a distance metric; and N is the number of the super pixels.
Specifically, in the present embodiment,representing the contribution rate of the qth similarity map to the p similarity maps,the specific solution of (a) is as follows:
wherein,andrespectively representing the p-th and q-th feature vectors of the first principal component of the hyperspectral image cube;andrespectively representing the central pixel positions of the p-th and q-th superpixels in the first main component of the hyperspectral image cube;anda Gaussian kernel function representing a distance metric controls the spectral value sensitivity and spatial distance sensitivity, respectively, of the superpixel region. In the invention, the values of both parameters are 1; n is the number of superpixels. For the above equation, the first term is used mainly to measure the spectral similarity of two superpixels, while the second term is used mainly to measure the spatial distance of two superpixels.
In the embodiment, the spectrum similarity and the space similarity without using the superpixel are used for weighting the diffusion process, and the complementarity among a plurality of similar graphs can be better utilized, so that the mutual relation among different wave bands can be more accurately described, and further the selection of the wave bands of the hyperspectral image can be more accurately realized.
Optionally, the unified similarity map is determined by a fifth formula, the fifth formula comprising:
whereinRepresenting a unified similarity graph;a corresponding similar graph of the pth super pixel at the Tth iteration; and N is the number of the super pixels.
Specifically, in this embodiment, after T diffusion times through the third formula, the similarity map corresponding to each super pixel may be updated to a stable state, and finally, a unified similarity map reflecting similarities of different hyperspectral bands may be obtained.
In the embodiment, the unified similar graph reflecting the similarity of different hyperspectral wave bands is obtained according to the fifth formula, and then spectral clustering can be performed according to the unified similar graph so as to construct the optimal wave band subset, so that the interrelation among different wave bands can be more accurately described, and further the selection of the hyperspectral image wave bands can be more accurately realized.
Another embodiment of the present invention provides a hyperspectral image band selection apparatus, including: the hyperspectral image segmentation module is used for acquiring a hyperspectral image cube and generating a plurality of superpixels according to the hyperspectral image cube; a super-pixel similarity map construction module, configured to construct, for each super-pixel, a similarity map reflecting similarity between bands included in the super-pixel; the similar graph fusion module is used for generating a unified similar graph according to all the similar graphs by adopting a multi-graph diffusion fusion strategy; the spectral clustering module is used for performing spectral clustering on the original hyperspectral data according to the unified similar graph to obtain a plurality of hyperspectral subcubes; and the hyperspectral waveband selecting module is used for selecting a waveband with the minimum noise value from each hyperspectral sub-cube as a characteristic waveband so as to determine an optimal waveband subset.
Another embodiment of the present invention provides a hyperspectral image band selection system, including a computer-readable storage medium storing a computer program and a processor, where the computer program is read by the processor and executed to implement the hyperspectral image band selection method.
Another embodiment of the present invention provides a computer-readable storage medium, which stores a computer program, and when the computer program is read and executed by a processor, the method for selecting a hyperspectral image band is implemented as described above.
Although the present disclosure has been described above, the scope of the present disclosure is not limited thereto. Various changes and modifications may be effected therein by one of ordinary skill in the pertinent art without departing from the spirit and scope of the present disclosure, and these changes and modifications are intended to be within the scope of the present disclosure.
Claims (10)
1. A hyperspectral image band selection method is characterized by comprising the following steps:
acquiring a hyperspectral image cube, and generating a plurality of superpixels according to the hyperspectral image cube;
for each super pixel, constructing a similarity map reflecting the similarity between the wave bands contained in the super pixel;
generating a unified similar graph according to all the similar graphs by adopting a multi-graph diffusion fusion strategy;
performing spectral clustering on original hyperspectral data according to the unified similarity map to obtain a plurality of hyperspectral subcubes;
and selecting the wave band with the minimum noise value from each high spectrum sub-cube as a characteristic wave band to determine an optimal wave band subset.
2. The hyperspectral image band selection method of claim 1, wherein the generating a plurality of superpixels from the hyperspectral image cube comprises:
and extracting a first principal component of the hyperspectral image cube by adopting a principal component analysis method, and dividing the first principal component into a plurality of superpixels by superpixel division.
3. The hyperspectral image band selection method according to claim 2, wherein the superpixel segmentation adopts an ERS entropy rate superpixel segmentation algorithm, and the ERS entropy rate superpixel segmentation algorithm adopts a first formula to segment the first principal component, wherein the first formula comprises:
4. The method for selecting the wavelength band of the hyperspectral image according to claim 1, wherein constructing the similarity map reflecting the similarity between the wavelength bands included in the superpixel comprises: determining the similarity between the bands contained in the superpixels according to a second formula, wherein the second formula comprises:
5. The hyperspectral image band selection method according to claim 1, wherein the obtaining a unified similarity map from the similarity map by using a multi-map diffusion fusion strategy comprises:
updating and diffusing the similarity graph according to a third formula, wherein the third formula comprises:
wherein,is flatWeighing parameters;representing the contribution rate of the similar graph corresponding to the qth super pixel to the similar graph corresponding to the pth super pixel;andrespectively representing the corresponding similar graphs of the qth super-pixel and the pth super-pixel in the t iteration;representing a similar graph corresponding to the p-th super pixel in the t +1 th iteration; a. thet pShowing the normalized similarity graph corresponding to the pth superpixel region at time t.
6. The hyperspectral image band selection method according to claim 5, wherein the hyperspectral image band selection methodDetermined by a fourth formula comprising:
wherein,andrespectively representing the p-th and q-th feature vectors of the first principal component of the hyperspectral image cube;andrespectively representing the central pixel positions of the p-th and q-th superpixels in the first main component of the hyperspectral image cube;anda gaussian kernel function representing a distance metric; and N is the number of the super pixels.
7. The hyperspectral image band selection method of claim 5, wherein the unified similarity map is determined by a fifth formula comprising:
8. A hyperspectral image band selection device, comprising:
the hyperspectral image segmentation module is used for acquiring a hyperspectral image cube and generating a plurality of superpixels according to the hyperspectral image cube;
a super-pixel similarity map construction module, configured to construct, for each super-pixel, a similarity map reflecting similarity between bands included in the super-pixel;
the similar graph fusion module is used for generating a unified similar graph according to all the similar graphs by adopting a multi-graph diffusion fusion strategy;
the spectral clustering module is used for performing spectral clustering on the original hyperspectral data according to the unified similar graph to obtain a plurality of hyperspectral subcubes;
and the hyperspectral waveband selecting module is used for selecting a waveband with the minimum noise value from each hyperspectral sub-cube as a characteristic waveband so as to determine an optimal waveband subset.
9. A hyperspectral image band selection system comprising a computer readable storage medium and a processor, wherein the computer readable storage medium stores a computer program, and the computer program is read by the processor and when executed implements the hyperspectral image band selection method according to any of claims 1 to 7.
10. A computer-readable storage medium, wherein a computer program is stored, which when read and executed by a processor, implements the hyperspectral image band selection method according to any of claims 1 to 7.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111050606.2A CN113486876A (en) | 2021-09-08 | 2021-09-08 | Hyperspectral image band selection method, device and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202111050606.2A CN113486876A (en) | 2021-09-08 | 2021-09-08 | Hyperspectral image band selection method, device and system |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113486876A true CN113486876A (en) | 2021-10-08 |
Family
ID=77946740
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202111050606.2A Pending CN113486876A (en) | 2021-09-08 | 2021-09-08 | Hyperspectral image band selection method, device and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113486876A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116754511A (en) * | 2023-08-18 | 2023-09-15 | 天津博霆光电技术有限公司 | Indocyanine green detection method and system based on spectrum technology |
CN117423004A (en) * | 2023-12-19 | 2024-01-19 | 深圳大学 | Band selection method, device, terminal and storage medium for hyperspectral image |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113075129A (en) * | 2021-02-07 | 2021-07-06 | 浙江师范大学 | Hyperspectral image band selection method and system based on neighbor subspace division |
-
2021
- 2021-09-08 CN CN202111050606.2A patent/CN113486876A/en active Pending
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113075129A (en) * | 2021-02-07 | 2021-07-06 | 浙江师范大学 | Hyperspectral image band selection method and system based on neighbor subspace division |
Non-Patent Citations (1)
Title |
---|
CHANG TANG ET AL.: "Hyperspectral Band Selection via Spatial-Spectral Weighted Region-wise Multiple Graph Fusion-Based Spectral Clustering", 《 PROCEEDINGS OF THE THIRTIETH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116754511A (en) * | 2023-08-18 | 2023-09-15 | 天津博霆光电技术有限公司 | Indocyanine green detection method and system based on spectrum technology |
CN116754511B (en) * | 2023-08-18 | 2023-10-27 | 天津博霆光电技术有限公司 | Indocyanine green detection method and system based on spectrum technology |
CN117423004A (en) * | 2023-12-19 | 2024-01-19 | 深圳大学 | Band selection method, device, terminal and storage medium for hyperspectral image |
CN117423004B (en) * | 2023-12-19 | 2024-04-02 | 深圳大学 | Band selection method, device, terminal and storage medium for hyperspectral image |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhai et al. | Laplacian-regularized low-rank subspace clustering for hyperspectral image band selection | |
Zhou et al. | Polarimetric SAR image classification using deep convolutional neural networks | |
Zhao et al. | Band-subset-based clustering and fusion for hyperspectral imagery classification | |
Kong et al. | Spectral–spatial feature extraction for HSI classification based on supervised hypergraph and sample expanded CNN | |
Sun et al. | An MRF model-based active learning framework for the spectral-spatial classification of hyperspectral imagery | |
Zheng et al. | Hyperspectral image classification with small training sample size using superpixel-guided training sample enlargement | |
Appice et al. | A novel spectral-spatial co-training algorithm for the transductive classification of hyperspectral imagery data | |
CN113486876A (en) | Hyperspectral image band selection method, device and system | |
CN114419381B (en) | Semantic segmentation method and road ponding detection method and device applying same | |
Jia et al. | A multiscale superpixel-level group clustering framework for hyperspectral band selection | |
Chen et al. | Hyperspectral remote sensing image classification based on dense residual three-dimensional convolutional neural network | |
Feng et al. | Dual-graph convolutional network based on band attention and sparse constraint for hyperspectral band selection | |
Tang et al. | Hyperspectral Band Selection via Spatial-Spectral Weighted Region-wise Multiple Graph Fusion-Based Spectral Clustering. | |
Lu et al. | Comparison between several feature extraction/classification methods for mapping complicated agricultural land use patches using airborne hyperspectral data | |
CN111680579A (en) | Remote sensing image classification method for adaptive weight multi-view metric learning | |
CN104392454B (en) | The merging method based on the scoring of atural object classification degree of membership under the empty spectrum combining classification framework of high-spectrum remote sensing | |
Zhao et al. | Superpixel-level global and local similarity graph-based clustering for large hyperspectral images | |
Hanachi et al. | Semi-supervised classification of hyperspectral image through deep encoder-decoder and graph neural networks | |
Azar et al. | Linear mixing model with scaled bundle dictionary for hyperspectral unmixing with spectral variability | |
Sellami et al. | SHCNet: A semi-supervised hypergraph convolutional networks based on relevant feature selection for hyperspectral image classification | |
Li et al. | A label propagation method using spatial-spectral consistency for hyperspectral image classification | |
Cao et al. | SDFL-FC: Semisupervised deep feature learning with feature consistency for hyperspectral image classification | |
Islam et al. | Improving hyperspectral image classification through spectral-spatial feature reduction with a hybrid approach and deep learning | |
Ge et al. | A semi-supervised learning method for hyperspectral imagery based on self-training and local-based affinity propagation | |
Sharma et al. | Classification of hyperspectral remote sensing image via rotation-invariant local binary pattern-based weighted generalized closest neighbor |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20211008 |