LU502854B1 - A hyperspectral image band selection method and system based on nearest neighbor subspace division - Google Patents
A hyperspectral image band selection method and system based on nearest neighbor subspace division Download PDFInfo
- Publication number
- LU502854B1 LU502854B1 LU502854A LU502854A LU502854B1 LU 502854 B1 LU502854 B1 LU 502854B1 LU 502854 A LU502854 A LU 502854A LU 502854 A LU502854 A LU 502854A LU 502854 B1 LU502854 B1 LU 502854B1
- Authority
- LU
- Luxembourg
- Prior art keywords
- denotes
- bands
- band
- subspace
- hyperspectral image
- Prior art date
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/194—Terrestrial scenes using hyperspectral data, i.e. more or other wavelengths than RGB
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/01—Arrangements or apparatus for facilitating the optical investigation
-
- 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
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- 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
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N2021/1793—Remote sensing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A40/00—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production
- Y02A40/10—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production in agriculture
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Analytical Chemistry (AREA)
- Immunology (AREA)
- Pathology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Evolutionary Computation (AREA)
- Chemical & Material Sciences (AREA)
- Biochemistry (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Remote Sensing (AREA)
- Image Analysis (AREA)
Abstract
The present application discloses a method and system for selecting hyperspectral image bands based on nearest neighbor subspace division, wherein a method for selecting hyperspectral image bands based on nearest neighbor subspace division is involved, comprising: S11. Input the hyperspectral image cube and calculate the correlation coefficients between any two adjacent bands contained in the input hyperspectral image cube to obtain the vector of correlation coefficients; S12. Find all the correlation coefficient extrema based on the vector of correlation coefficients, and filter the correlation coefficient minima among all the correlation coefficient extrema found, and determine the optimal subspace of hyperspectral bands by the filtered correlation coefficient minima; S13. The subspaces are sorted according to the size of the number of bands in the subspace, and the information entropy of the bands in each subspace is calculated, and the required number of characteristic bands are selected from the subspace according to the calculated information entropy. The present application ensures that the final selected bands contain relatively complete information.
Description
BL-5568
LU502854
A HYPERSPECTRAL IMAGE BAND SELECTION METHOD AND SYSTEM
BASED ON NEAREST NEIGHBOR SUBSPACE DIVISION
The present application relates to the technical field of hyperspectral remote sensing image band selection, and in particular to a method and system for selecting hyperspectral image bands based on nearest neighbor subspace division.
Compared with traditional RGB images, hyperspectral images are characterized by rich information, large number of bands and high resolution, so it is widely used in target detection, environmental monitoring, mineral exploration, agricultural resource investigation and marine research. Since it contains a large amount of band information and the feature similarity between neighboring bands is high, this increases the computational complexity of hyperspectral image classification to a certain extent and also affects the final classification accuracy of the classifier.
Therefore, hyperspectral images need to be downscaled to solve the dimensional disaster problem.
Feature extraction and feature selection have been the hot research topics in the field of data dimensionality reduction, and these two methods are also commonly applied to hyperspectral image dimensionality reduction. Among other things, in hyperspectral imagery, feature selection can also be referred to as band selection.
As it stands, most of the clustering-based band selection algorithms suffer from a strong sensitivity to the number of cluster centers. Specifically, the accuracy is low when the number of selected cluster centers is small; when the number of selected cluster centers exceeds a certain number, the accuracy instead shows a decreasing trend as the number of cluster centers increases.
These two cases clearly do not meet the practical needs.
It is an object of the present application to provide a method and a system for selecting a hyperspectral image band based on a nearest neighbor subspace division, in response to a defect in 1
BL-5568
LU502854 the prior art.
To achieve the above, this application uses the following technical solution:
A hyperspectral image band selection method based on nearest neighbor subspace division, comprising:
S1. Input the hyperspectral image cube and calculate the correlation coefficients between any two adjacent bands contained in the input hyperspectral image cube to obtain the vector of correlation coefficients;
S2. Find all the correlation coefficient extrema based on the vector of correlation coefficients, and filter the correlation coefficient minima among all the correlation coefficient extrema found, and determine the optimal subspace of hyperspectral bands by the filtered correlation coefficient minima;
S3. The subspaces are sorted according to the size of the number of bands in the subspace, and the information entropy of the bands in each subspace is calculated, and the required number of characteristic bands are selected from the subspace according to the calculated information entropy.
Further, said step S1 in calculating the correlation coefficient between any two adjacent bands contained in the input hyperspectral image cube to obtain a vector of correlation coefficients, denoted as:
W H
ZZ (Km) X )(X (mm) =X)
RX, X,)= CE
W OH _ 2 = 2X, (mn) X)
Where, X € R"" denotes the hyperspectral image cube, /, denotes the number of bands contained in the hyperspectral image cube, W and H denotes the width and height of each band respectively; X, denotes the 7 -th band in the hyperspectral image data set; X ; denotes the j-th band in the hyperspectral image data set; X, (m,n) denotes the value of the m -th row and 72-th column of X, ; R denotes the vector of correlation coefficients; V; denotes the pixel difference of the 7 -th band; V; denotes the pixel difference of the j-th band; VW, denotes the pixel difference of the o-th band; X, denotes the o-th band in the hyperspectral image data set; x, 2
BL-5568 denotes the pixel mean of the o-th band; O = {i, J} 10906856
Further, said step S2 of filtering out the correlation coefficient minima 1s done by using the toolbox of MATLAB.
Further, said filtering by MATLAB's toolbox 1s denoted as: y = findpeaks(—R)
Where, findpeaks denotes the MATLAB built-in function; R denotes the vector of correlation coefficients between adjacent bands of the whole data set; and y denotes the vector containing all minimal values.
Further, said step S3 in calculating the information entropy of the bands in each subspace, denoted as:
Q
H =-) p,logp, i=1
Where, © denotes the entire sample space, p(i) denotes the probability of the sample
I appearing in the image, and H denotes the information entropy of the sample 7.
Further, said step S3 in which a desired number of feature bands are selected from the subspace based on the calculated information entropy, denoted as:
S = Z xK
L
Where, Z denotes the number of bands contained in each band subspace; [ denotes the number of bands contained in the whole data set; K denotes the number of feature bands to be selected; S denotes the selected bands.
Accordingly, there is also provided a hyperspectral image band selection system based on nearest neighbor subspace division, comprising:
A calculation module for inputting a hyperspectral image cube and calculating a correlation coefficient between any two adjacent bands contained in the input hyperspectral image cube to obtain a vector of correlation coefficients;
A screening module for finding all correlation coefficient extreme value points based on the obtained vector of correlation coefficients, and screening out correlation coefficient extreme value points among all correlation coefficient extreme value points found, and determining an optimally 3
BL-5568
LU502854 divided hyperspectral band subspace; by the screened correlation coefficient extreme value points
A selection module for sorting the subspaces according to the size of the number of bands in the subspaces and calculating the information entropy of the bands in each subspace, and selecting a desired number of feature bands from the subspaces based on the calculated information entropy.
Further, said computing module in calculating the correlation coefficient between any two adjacent bands contained in the input hyperspectral image cube to obtain a vector of correlation coefficients, denoted as:
W OH
ZZ (Km) X )(X (mm) =X)
RX, X,)= FT
W H _ 2 = 2X, (mn) X)
Where, X € R"" denotes the hyperspectral image cube, /, denotes the number of bands contained in the hyperspectral image cube, W and H denotes the width and height of each band respectively; X, denotes the 7 -th band in the hyperspectral image data set; X ; denotes the j-th band in the hyperspectral image data set; À, (m,n) denotes the value of the m -th row and 72-th column of X, ; R denotes the vector of correlation coefficients; V; denotes the pixel difference of the 7 -th band; V; denotes the pixel difference of the j-th band; VW, denotes the pixel difference of the o-th band; X, denotes the o-th band in the hyperspectral image data set; x, denotes the pixel mean of the o-th band; O = {i , J}
Further, said information entropy of the bands in each subspace is calculated in said selection module, denoted as:
Q
H= 72 p,logp,
Where, © denotes the entire sample space, p(i) denotes the probability of the sample appearing in the image, and H denotes the information entropy of the sample /.
Further, said selecting module in which a desired number of feature bands are selected from the subspace based on the calculated information entropy, denoted as: 4
BL-5568
LU502854
S = Z xK
L
Where, Z denotes the number of bands contained in each band subspace; [ denotes the number of bands contained in the whole data set; K denotes the number of feature bands to be selected; S denotes the selected bands.
Compared with the prior art, this application finds the grouping critical points by calculating the correlation coefficients between adjacent bands, and then uses the number of critical points as the final determined number of clusters. In addition, information entropy is used as the evaluation index for waveband selection. When band selection is performed for each clustering cluster, the information entropy is used as a weight to rank each band, and then the band with the highest information entropy is selected as the feature band, which ensures that the final selected band contains relatively complete information.
Illustration of the attached figure
Figure 1 is a flow chart of a hyperspectral image band selection method based on nearest neighbor subspace division provided in Embodiment 1;
Fig. 2 is a block diagram of a hyperspectral image band selection based on nearest neighbor subspace division provided in Embodiment 1;
Figure 3 is a schematic diagram of an example of nearest neighbor subspace partitioning provided in Embodiment 1;
Figure 4 is a schematic diagram of the example information entropy of all bands on the Indian
Pines dataset provided in Embodiment I;
Figure 5 is a schematic diagram of the Overall-Accuracy curves of the different hyperspectral image band selection methods provided in Embodiment II on the Botswana dataset;
Figure 6 is a schematic diagram of the Overall-Accuracy curves of the different hyperspectral image band selection methods provided in Embodiment II on the Salinas dataset;
Figure 7 is a schematic diagram of the Overall-Accuracy curves of the different hyperspectral image band selection methods provided by Embodiment II on the Indian Pines dataset.
BL-5568
LU502854
The following illustrates the steps of this application by specific concrete examples, and other advantages and efficacy of this application can be readily understood by a person of ordinary skill in the art by what 1s disclosed in this specification. The present application may also be implemented or applied by additionally different specific embodiments, and the details in this specification may also be modified or changed in various ways without departing from the spirit of the present application based on different views and applications. It is to be noted that the following embodiments and the features in the embodiments can be combined with each other without conflict.
The present application provides a method and system for hyperspectral image band selection based on nearest neighbor subspace division in response to existing defects.
Embodiment I
The present embodiment provides a hyperspectral image band selection method based on nearest neighbor subspace division, as shown in FIGS. 1 and 2, comprising:
S11. Input the hyperspectral image cube and calculate the correlation coefficient between any two adjacent bands contained in the input hyperspectral image cube to obtain the vector of correlation coefficients:
S12. Find all the correlation coefficient extreme value points based on the obtained vector of correlation coefficients, and screen out the correlation coefficient extreme value points among all the correlation coefficient extreme value points found, and determine the hyperspectral band subspace; for optimal division by the screened correlation coefficient extreme value points
S13. Sort the subspaces according to the size of the number of bands in the subspace and calculate the information entropy of the bands in each subspace, and select the desired number of feature bands from the subspace according to the calculated information entropy.
In the present embodiment, based on the idea that the bands are ordered among themselves and that the correlation between bands that are farther away is low, it is considered that dividing the bands according to the ordered bands not only reduces the correlation between the subspaces of the bands, but also avoids the problem of high correlation between the last selected feature bands.
Based on this, this embodiment proposes a hyperspectral image band selection method based on 6
BL-5568
LU502854 nearest neighbor subspace division, referred to as SEASP. Among them, SEASP uses a combination of clustering and ranking to solve the problem of low accuracy when the number of clustering centers is small.
In step S11, the hyperspectral image cube is input, and the correlation coefficient between any two adjacent bands contained in the input hyperspectral image cube is calculated to obtain a vector of correlation coefficients.
For a given input hyperspectral image cube, first calculate the correlation coefficients between any two neighboring bands contained therein to obtain the corresponding correlation coefficient vectors.
Among them, the specific steps for calculating the correlation coefficient between any two bands contained in the hyperspectral image cube are:
Given a hyperspectral image cube À € RSI where represents the number of bands contained in the hyperspectral image cube, W and H represent the width and height of each band respectively.
In general, a hyperspectral band image can be regarded as a two-dimensional matrix consisting of pixel points with different spectral values, so let X, represent the / th band in the hyperspectral image dataset. Then the correlation coefficient between bands X, and X, is calculated as follows.
W H _ -
ZZ (Km) X )(X (mm) =X)
RX, X,)= CE
W H _ 2 = 2X, (mn) X)
Where, X € R"" denotes the hyperspectral image cube, /, denotes the number of bands contained in the hyperspectral image cube, W and H denote the width and height of each band, respectively; X, denotes the 7 -th band in the hyperspectral image dataset, X ; denotes the j-th band in the hyperspectral image dataset, À, (m,n) denotes the value of the #77 -th row and -th column of X, ; R denotes the vector of correlation coefficients; Vi denotes the pixel 7
BL-5568
LU502854 difference of the j-th band; Jj denotes the pixel difference of the j-th band; Vo denotes the pixel difference of the o-th band; Xo denotes the o-th band in the hyperspectral image dataset; x, denotes the pixel mean of the o-th band; 0 = {i , J} x, are calculated as follows:
W H
__ LAN (mm) _ m=l n=l +o = Wx H
Considering the orderliness between hyperspectral bands and the weak correlation in non- neighborhood space, only the correlation coefficients between adjacent bands are calculated, and finally the correlation coefficient vector is obtained.
In this embodiment, for the input hyperspectral image cube y = p”<<L , the band image it contains is regarded as a two-dimensional matrix consisting of pixel points with different spectral values, and then the correlation coefficient between any two adjacent bands is found according to the above calculation method to obtain the correlation coefficient vector.
In step S12, all correlation coefficient extreme value points are searched based on the obtained vector of correlation coefficients, and the correlation coefficient extreme value points are screened out among all correlation coefficient extreme value points searched out, and the optimally divided hyperspectral band subspace is determined by the screened correlation coefficient extreme value points.
All correlation coefficient extreme value points are searched from the correlation coefficient vector, and the optimal segmented hyperspectral band subspace is determined by the selection of the extreme minor value points.
In the idea of clustering algorithm, the criterion for measuring the division is usually to make the highest intra-class correlation and the lowest inter-class correlation. Based on this, in SEASP, the magnitude of correlation between groupings is determined mainly based on the rate of change of correlation coefficients. Specifically, if the correlation between two bands is the smallest in some interval, i.e., the rate of change of the correlation coefficient is the largest in that interval, it means that these two bands do not belong to the same group with a high probability, and are segmentation points between two adjacent groupings. Considering the ordered nature between the bands and the weak correlation in the non-neighboring space, the implementation only calculates the correlation 8
BL-5568
LU502854 coefficients of the adjacent bands, at which point the problem of cluster delineation is transformed into the problem of finding the intermittent points between groupings in the ordered data set. So all the correlation coefficient minima it contains are calculated based on the obtained correlation coefficient vector to obtain the intermittent points of all subspaces.
The specific steps to find all the correlation coefficient minima from the correlation coefficient vector are:
In the ordered hyperspectral bands, cluster segmentation then translates into finding interruptions between groupings in the ordered dataset. For the problem of finding the minimal value point of correlation coefficient, the existing toolbox of MATLAB can be used to solve the problem, and the specific calculation method is shown below: y = findpeaks(—R)
Where, findpeaks denotes the MATLAB built-in function; R denotes the vector of correlation coefficients between adjacent bands of the whole data set; and y denotes the vector containing all minimal values.
To be specific, the function findpeaks in determining whether a point is a peak point, the main function value at the point and its left and right neighboring function values for comparison, if it is the largest, it will be judged as the waveform of a peak point. Therefore, when the inverse operation is performed on the input parameters, the peak point is the trough point of the original data. Now assume that the function value at point X is / (x ), then when [+ (x ) satisfies both
F(x) < F(x- 1) and F(x) < F(x+1) , it will be used as a trough point. Through the selection of the minimal value point to determine the interruption point between the ordered bands, so as to obtain the subspace of the divided bands, the specific example is shown in Figure 3.
In step S13, the subspaces are sorted according to the size of the number of bands in the subspaces, and the information entropy of the bands in each subspace 1s calculated, and a desired number of feature bands are selected from the subspaces according to the calculated information entropy.
The subspaces are sorted according to the size of the number of bands contained in each subspace, and the information entropy of each band is calculated, and finally the desired number of feature bands are selected from these several subspaces according to the information entropy 9
BL-5568
LU502854 size.
The specific steps to calculate the information entropy size of each band and select the desired number of feature bands from it are:
Information entropy is used to measure the size of the average amount of information contained in an image; the higher the information entropy of an image, the richer the information contained in that image. Therefore, information entropy is more appropriate as a metric to measure the degree of importance of the bands. The details of the calculation are shown below:
Q
H=-) p,logp, j=1
Where, © denotes the entire sample space, p(i) denotes the probability of the sample
I appearing in the image, and H denotes the information entropy of the sample 7.
An example of information entropy of all bands in the Indian Pines dataset is given in Fig. 4 as shown in Fig. 4.
In the process of ranking all bands according to the information entropy, the incoming parameters, i.e., the variation in the number of bands required, are taken into account, and each band subspace divided is ranked according to the size of the number of bands it contains to obtain the ranked band subspace.
When the number of required bands is less than or equal to the number of subspaces, the corresponding number of subspaces is selected from the sorted subspaces, and then the band with the largest information entropy is selected as the feature band from the selected subspaces; when the number of required bands is greater than the number of subspaces, S bands are selected from each subspace as the feature band.
To be Specific: when K is less than or equal to the number of subspaces, the band with the largest information entropy is selected from each subspace as the feature band; when K is greater than the number of subspaces, S bands are selected from each subspace according to the magnitude of information entropy. The specific calculation method of S is as follows:
S = Z xK
L
Where, Z denotes the number of bands contained in each band subspace; [ denotes the
BL-5568
LU502854 number of bands contained in the whole data set; K denotes the number of feature bands to be selected; S denotes the selected bands.
Through the calculation of the above equation, the desired subset of feature bands Y can be finally obtained.
Compared with the prior art, the present application finds the grouping critical points by calculating the correlation coefficients between adjacent bands, and finds the grouping critical points according to the curvature change of the correlation coefficients, and then uses the number of critical points as the final determined number of clusters. In addition, information entropy 1s used as the evaluation index for waveband selection. When band selection is performed for each clustering cluster, the information entropy 1s used as a weight to rank each band, and then the band with the highest information entropy 1s selected as the feature band, which ensures that the final selected band contains relatively complete information.
Embodiment II
The present embodiment provides a hyperspectral image band selection method based on nearest-neighbor subspace division that differs from Embodiment I in that:
In the present embodiment, a large number of relevant experiments are conducted to verify the effectiveness of the proposed SEASP algorithm for hyperspectral band selection.
Three public datasets were selected for the experiments, namely, the Botswana dataset, the
Salinas dataset and the Indian Pines dataset. The selected classifiers are KNN, SVM and LDA. The parameter kK of KNN classifier is set to 5, the distance function uniformly used in SVM is
Gaussian kernel function, the penalty factor of both Indian Pines and Botswana datasets is 10000, and the penalty factor of Salinas dataset is 100. Since all the above three classifiers are supervised classification methods. Therefore, when conducting experiments, 10% of the entire dataset 1s randomly selected as training samples and the rest of the data is used as test samples. In addition, this embodiment is compared with several contemporary more advanced algorithms, which are
ASPS MN, ASPS IE, TOF, UBS, and FNGBS. For the three public hyperspectral image datasets, the number of bands is selected in the range of 5 to 50 in this experiment because their optimal number of bands is currently unknown. in order to fully demonstrate the randomness of the band number setting Three metrics were used to analyze the classification results: overall accuracy (OA), 11
BL-5568
LU502854 average overall accuracy (AOA), and kappa coefficient (Kappa). All experiments were run on
MATLAB 2016a with a CPU of 17-5500U, 2.40 GHz and 8 Gb of memory.
Table 1 shows the AOA and Kappa exhibited by the competing algorithms on the three public data sets. Among them, the number of bands selected for the three datasets is 36, 5 and 5, respectively, and AOA is the range of the results of 10 runs when the test comparison is performed.
To clearly show the performance of the compared algorithms on the three datasets, the two better classification results are marked in bold form in Table 1.
Dataset Classifier ASPS_MN ASPS IE TOF FNGBS UBS SEASP
KNN(AOA) 85.75+0.29 85.12+031 8515+082 8562+053 85.10+080 86.52+0.31
KNN(Kappa) 84.89+0.30 8421+032 8426x085 8474+054 8420+084 85.68+ 0.32
SVM(AOA) 90.75£0.75 9020+0.69 8918+044 9028+0.61 9036+042 91.37+0.79
Botswana SVM(Kappa) 90.12+£0.79 89.57+0.70 8847+047 89.62+065 8971+045 90.78 0.82
LDA(AOA) 9035+038 90.39+046 89.16+049 9025+090 90.26+0.73 90.91+0.81
LDA(Kappa) 8969+040 90.39* 0.46 8845+0.52 89.60+095 89.60+0.78 90.291 0.84
KNN(AOA) 86.55+034 8437x047 386.97* 0.08 86.10+0.60 8536+031 87.31+£0.30
KNN(Kappa) 85.48+0.35 83.18+050 8592%£0.08 8501+063 8421+034 86.31+0.31
SVM(AOA) 8752+045 86.64+0.07 8948+0.23 8701+038 85.79+035 389.19+ 0.35
Salinas SVM(Kappa) 86.46+048 8534+0.08 88.57+0.25 85.944042 8458+038 88.26+ 0.38
LDA(AOA) 8093+0.87 8127+023 8341+0.30 82.00+045 7877045 84.31+0.16
LDA(Kappa) 79381094 79791025 82.08+0.32 80.60+0.57 77.05+055 83.00+0.18
KNN(AOA) 59.92+0.76 5543+024 3996x078 6191+0.64 5256+032 60.16 0.08
KNN(Kappa) 57.09+0.54 5256+025 5655+022 58.69+055 5031+031 57.61+0.10
SVM(AOA) 56.95+087 5298+0.17 62.71+0.25 6095+0.66 5737+032 62.504 0.42
Indian SVM(Kappa) 5409+089 4913+015 59.69+0.66 57.77+0.65 5443+037 59.57+047
Pines LDA(AOA) 5451+025 5231x017 56.62+0.15 5522+0.75 5251+045 55.83+0.10
LDA(Kappa) 51.63+040 _4922+023 53.53+0.20 _52.12+080 4826+0.21 53.22+0.18
Table 1
As can be seen from Figures 5, 6, and 7, the algorithm proposed in the present embodiment outperforms other algorithms for some specific number of bands. In Fig. 5, SEASP always maintains a stability when the number of selected bands is greater than 25, while the other algorithms fluctuate up and down. In addition, it can be seen from the results of all classifiers that some algorithms do not give as good results as others when the number of selected bands is small, which reflects their strong sensitivity to the number of selected bands. It can show a better performance only when a certain number of bands are selected. As a whole, with the increasing number of selected bands, SEASP does not show as much variation in the results as the other algorithms, which is a side verification that it is less sensitive to the number of selected bands and shows that it can perform well with a small number of selected bands. In Figs. 6 and 7, SEASP is basically in a stable upward state with the increase of the number of selected bands, and there is no 12
BL-5568
LU502854 obvious rebound. The other algorithms are less stable, and the opposite result tends to decrease when the number of selected bands becomes larger. This reflects that as the number of selected bands increases, all these algorithms inevitably select some redundant bands, which leads to a decrease in classification accuracy. The algorithm proposed in the present embodiment, on the other hand, does not show this obvious phenomenon, which also shows that SEASP fully takes into account the strong correlation between the final selected target bands when selecting the bands, thus correcting the shortcomings of most clustering-based band selection algorithms.
To further validate the feasibility of the proposed method, the times of all competing algorithms were also calculated on the wave selection of the three data sets. Where, above the number of bands selected for the three data sets, they are set to 10, 15, and 20, showing an increasing trend. The time taken by different algorithms to select the same number of bands on different datasets is given in Table 2. As can be seen from the results in the table, the time consumed by the algorithm proposed in the present embodiment is smaller and comparable to other algorithms.
Dataset ASPS-MN ASPS-IE TOF FNGBS UBS SEASP
Botswana 1.3690 2.2856 1.2859 1.0356 0.0001 1.9196
Salinas 0.6145 0.9672 0.8009 0.4422 0.0001 0.6754
Indian 0.1917 0.2348 0.4477 0.1021 0.0001 0.1363
Table 2
From the above series of experiments, it is clear that the algorithm proposed in the present embodiment is not only simple in principle, but also comparable or even better than other algorithms in terms of classification performance for the three common data sets, and in addition its execution speed is faster, thus verifying the effectiveness and feasibility of the algorithm.
Embodiment III
The present embodiment provides a hyperspectral image band selection system based on nearest neighbor subspace division, comprising:
A calculation module for inputting a hyperspectral image cube and calculating a correlation coefficient between any two adjacent bands contained in the input hyperspectral image cube to obtain a vector of correlation coefficients;
A screening module for finding all correlation coefficient extreme value points based on the obtained vector of correlation coefficients, and screening out correlation coefficient extreme value 13
BL-5568
LU502854 points among all correlation coefficient extreme value points found, and determining an optimally divided hyperspectral band subspace; by the screened correlation coefficient extreme value points
A selection module for sorting the subspaces according to the size of the number of bands in the subspaces and calculating the information entropy of the bands in each subspace, and selecting a desired number of feature bands from the subspaces based on the calculated information entropy.
Further, said calculation module in which the correlation coefficient between any two adjacent bands contained in the input hyperspectral image cube 1s calculated, denoted as:
W OH
ZZ (Km) X )(X (mm) =X)
RX, X,)= FT
W H _ 2 = 2X, (mn) X)
Where, X € R"" denotes the hyperspectral image cube, /, denotes the number of bands contained in the hyperspectral image cube, W and H denotes the width and height of each band respectively; X, denotes the 7 -th band in the hyperspectral image data set; X ; denotes the j-th band in the hyperspectral image dataset; X, (m,n) denotes the value of the M -th row and N-th column of X,; R(X,,X,) denotes the correlation coefficient between two adjacent bands; J; denotes the pixel difference of the 7 -th band; Jj denotes the pixel difference of the j-th band; Vo denotes the pixel difference of the o-th band; Xo denotes the o-th band in the hyperspectral image dataset; x, denotes the pixel mean value of the o-th band; O = {i, J}
Further, said information entropy of the bands in each subspace is calculated in said selection module, expressed as:
Q
H= 72 p,logp,
Where, © denotes the entire sample space, p(i) denotes the probability of the sample appearing in the image, and H denotes the information entropy of the sample /.
Further, said selecting module in which a desired number of feature bands are selected from the subspace based on the calculated information entropy, denoted as: 14
BL-5568
LU502854
S = Z xK
L
Where, Z denotes the number of bands contained in each band subspace; [ denotes the number of bands contained in the whole data set; K denotes the number of feature bands to be selected; S denotes the selected bands.
It should be noted that this embodiment provides a hyperspectral image band selection system based on nearest-neighbor subspace division similar to embodiment 1, and will not be repeated here.
Compared with the prior art, the present application finds the grouping critical points by calculating the correlation coefficients between adjacent bands, and finds the grouping critical points according to the curvature change of the correlation coefficients, and then uses the number of critical points as the final determined number of clusters. In addition, information entropy is used as the evaluation index for waveband selection. When band selection is performed for each clustering cluster, the information entropy is used as a weight to rank each band, and then the band with the highest information entropy is selected as the feature band, which ensures that the final selected band contains relatively complete information.
Note that the above is only a preferred embodiment of the present application and the technical principles applied. It will be understood by those skilled in the art that this application is not limited to the particular embodiments described herein, and that various variations, readjustments, and substitutions are apparent to those skilled in the art without departing from the scope of protection of this application. Therefore, although the present application is described in some detail by the above embodiments, the present application is not limited to the above embodiments, but can include more other equivalent embodiments without departing from the conception of the present application, and the scope of the present application is determined by the scope of the appended claims.
Claims (10)
1. Ahyperspectral image band selection method based on nearest neighbor subspace division, characterized in that it comprises:
S1. Input the hyperspectral image cube and calculate the correlation coefficients between any two adjacent bands contained in the input hyperspectral image cube to obtain the vector of correlation coefficients:
S2. Find all the correlation coefficient extrema based on the vector of correlation coefficients, and filter the correlation coefficient minima among all the correlation coefficient extrema found, and determine the optimal subspace of hyperspectral bands by the filtered correlation coefficient minima:
S3. The subspaces are sorted according to the size ofthe number of bands in the subspace, and the information entropy of the bands in each subspace is calculated, and the required number of characteristic bands are selected from the subspace according to the calculated information entropy.
2. A method for selecting a hyperspectral image band based on the division of a nearest neighbor subspace according to claim 1, characterized in that said step S1 calculates a correlation coefficient between any two adjacent bands contained in the input hyperspectral image cube to obtain a vector of correlation coefficients, denoted as: W H ZZ (Km) X )(X (mm) =X) RX X ) = A JT W OH _ 2 = 2X, (mn) X) Where, X € R"" denotes the hyperspectral image cube, /, denotes the number of bands contained in the hyperspectral image cube, W and H denotes the width and height of each band respectively; X, denotes the 7 -th band in the hyperspectral image data set; X ; denotes the j-th band in the hyperspectral image data set; X, (m,n) denotes the value of the m -th row and 72-th column of X, ; R denotes the vector of correlation coefficients; V; denotes the pixel difference of the 7 -th band; V; denotes the pixel difference of the j-th band; VW, denotes the pixel difference of the o-th band; X, denotes the o-th band in the hyperspectral image data set; x, 16
BL-5568 denotes the pixel mean of the o-th band; O = {i, J} 10906856
3. A method for selecting hyperspectral image bands based on nearest neighbor subspace division according to claim 2, characterized in that said step S2 of screening out the points with very small values of correlation coefficients is performed by means of a toolbox of MATLAB.
4. A method for selecting hyperspectral image bands based on nearest neighbor subspace division according to claim 3, characterized in that said screening is performed over a toolbox of MATLAB, denoted as: y = findpeaks(—R) Where, findpeaks denotes the MATLAB built-in function; R denotes the vector of correlation coefficients between adjacent bands of the whole data set; and y denotes the vector containing all minimal values.
5. A method for selecting hyperspectral image bands based on nearest neighbor subspace division according to claim 4, characterized in that said step S3 calculates the information entropy of the bands in each subspace, denoted as: Q H =-) p,logp, i=1 Where, © denotes the entire sample space, p(i) denotes the probability of the sample I appearing in the image, and H denotes the information entropy of the sample 7.
6. A method for selecting hyperspectral image bands based on the division of the nearest neighbor subspace according to claim 5, characterized in that said step S3 selects a desired number of feature bands from the subspace based on the calculated information entropy, denoted as: S = Z xK L Where, Z denotes the number of bands contained in each band subspace; [ denotes the number of bands contained in the whole data set; K denotes the number of feature bands to be selected; S denotes the selected bands.
7. A system for segmenting hyperspectral image bands based on a nearest neighbor subspace, characterized in that it comprises: A calculation module for inputting a hyperspectral image cube and calculating a correlation 17
BL-5568 LU502854 coefficient between any two adjacent bands contained in the input hyperspectral image cube to obtain a vector of correlation coefficients: À screening module for finding all correlation coefficient extreme value points based on the obtained vector of correlation coefficients, and screening out correlation coefficient extreme value points among all correlation coefficient extreme value points found, and determining an optimally divided hyperspectral band subspace; by the screened correlation coefficient extreme value points A selection module for sorting the subspaces according to the size of the number of bands in the subspaces and calculating the information entropy of the bands in each subspace, and selecting a desired number of feature bands from the subspaces based on the calculated information entropy.
8. A hyperspectral image band selection system based on nearest neighbor subspace division according to claim 7, characterized in that said calculation module in which the correlation coefficients between any two adjacent bands contained in the input hyperspectral image cube are calculated to obtain a vector of correlation coefficients, denoted as: W H _ - ZZ (Km) X )(X (mm) =X) RX X) A JT W H 2 = 2X, (mn) X) Where, X € R"" denotes the hyperspectral image cube, /, denotes the number of bands contained in the hyperspectral image cube, W and H denotes the width and height of each band respectively; X, denotes the first band in the hyperspectral image data set; denotes the j-th band in the hyperspectral image data set; X, (m,n) denotes the value of the 7 -th row and n-th column of X, ; R denotes the vector of correlation coefficients; J; denotes the pixel difference of the 7 -th band; V; denotes the pixel difference of the j-th band; Vo denotes the pixel difference of the o-th band; X, denotes the o-th band in the hyperspectral image data set; x, denotes the pixel mean of the o-th band; O = {i , J}
9. Ahyperspectral image band selection system based on nearest neighbor subspace division according to claim 8, characterized in that said selecting module in which the information entropy of the bands in each subspace is calculated, denoted as: 18
BL-5568 LU502854 Q H=-) p,logp, j=1 Where, © denotes the entire sample space, p(i) denotes the probability of the sample I appearing in the image, and H denotes the information entropy of the sample 7.
10. A hyperspectral image band selection system based on the division of the nearest neighbor subspace according to claim 9, characterized in that said selection module selects a desired number of feature bands from the subspace based on the calculated information entropy, denoted as: Z S=—xK L Where, 7 denotes the number of bands contained in each band subspace; 7, denotes the number of bands contained in the whole data set; K denotes the number of feature bands to be selected; S denotes the selected bands. 19
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110174636.8A CN113075129B (en) | 2021-02-07 | 2021-02-07 | Hyperspectral image band selection method and system based on neighbor subspace division |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| LU502854B1 true LU502854B1 (en) | 2023-01-30 |
Family
ID=76609942
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| LU502854A LU502854B1 (en) | 2021-02-07 | 2021-12-07 | A hyperspectral image band selection method and system based on nearest neighbor subspace division |
Country Status (4)
| Country | Link |
|---|---|
| CN (1) | CN113075129B (en) |
| LU (1) | LU502854B1 (en) |
| WO (1) | WO2022166363A1 (en) |
| ZA (1) | ZA202207737B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113075129B (en) * | 2021-02-07 | 2023-03-31 | 浙江师范大学 | Hyperspectral image band selection method and system based on neighbor subspace division |
| CN113486876A (en) * | 2021-09-08 | 2021-10-08 | 中国地质大学（武汉） | Hyperspectral image band selection method, device and system |
| CN117435940B (en) * | 2023-12-20 | 2024-03-05 | 龙建路桥股份有限公司 | Spectrum detection method for winter concrete curing process |
Family Cites Families (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101131734A (en) * | 2007-06-25 | 2008-02-27 | 北京航空航天大学 | Automatic waveband selection method for high optical spectrum remote sensing image |
| KR100963797B1 (en) * | 2008-02-27 | 2010-06-17 | 아주대학교산학협력단 | Method for realtime target detection based on reduced complexity hyperspectral processing |
| CN103065293A (en) * | 2012-12-31 | 2013-04-24 | 中国科学院东北地理与农业生态研究所 | Correlation weighted remote-sensing image fusion method and fusion effect evaluation method thereof |
| CN103886334A (en) * | 2014-04-08 | 2014-06-25 | 河海大学 | Multi-index fused hyperspectral remote sensing image dimensionality reduction method |
| CN104122210B (en) * | 2014-07-02 | 2017-01-25 | 中国林业科学研究院林业研究所 | Hyperspectral waveband extraction method based on optimum index factor-correlation coefficient method |
| CN104751179B (en) * | 2015-04-01 | 2018-02-06 | 河海大学 | A kind of multiple target target in hyperspectral remotely sensed image band selection method based on game theory |
| CN107124612B (en) * | 2017-04-26 | 2019-06-14 | 东北大学 | Method for compressing high spectrum image based on distributed compression perception |
| CN108154094B (en) * | 2017-12-14 | 2020-04-24 | 浙江工业大学 | Hyperspectral image unsupervised waveband selection method based on subinterval division |
| WO2019183136A1 (en) * | 2018-03-20 | 2019-09-26 | SafetySpect, Inc. | Apparatus and method for multimode analytical sensing of items such as food |
| CN110751142B (en) * | 2019-09-25 | 2022-07-26 | 河海大学 | Improved hyperspectral remote sensing image band selection method |
| CN113075129B (en) * | 2021-02-07 | 2023-03-31 | 浙江师范大学 | Hyperspectral image band selection method and system based on neighbor subspace division |
-
2021
- 2021-02-07 CN CN202110174636.8A patent/CN113075129B/en active Active
- 2021-12-07 WO PCT/CN2021/135928 patent/WO2022166363A1/en active Application Filing
- 2021-12-07 LU LU502854A patent/LU502854B1/en active IP Right Grant
-
2022
- 2022-07-12 ZA ZA2022/07737A patent/ZA202207737B/en unknown
Also Published As
| Publication number | Publication date |
|---|---|
| WO2022166363A1 (en) | 2022-08-11 |
| ZA202207737B (en) | 2022-07-27 |
| CN113075129A (en) | 2021-07-06 |
| CN113075129B (en) | 2023-03-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| LU502854B1 (en) | A hyperspectral image band selection method and system based on nearest neighbor subspace division | |
| Wang et al. | Hyperspectral band selection via adaptive subspace partition strategy | |
| Dang et al. | Local outlier detection with interpretation | |
| Chen et al. | A clustering algorithm for multiple data streams based on spectral component similarity | |
| Alshammari et al. | Refining a k-nearest neighbor graph for a computationally efficient spectral clustering | |
| CN106778680B (en) | A kind of hyperspectral image band selection method and device based on critical bands extraction | |
| CN103403704A (en) | Method and device for finding nearest neighbor | |
| Mondal et al. | Video shot boundary detection using multiscale geometric analysis of nsct and least squares support vector machine | |
| CN108205657A (en) | Method, storage medium and the mobile terminal of video lens segmentation | |
| Duin et al. | Mode seeking clustering by KNN and mean shift evaluated | |
| Dang et al. | Outlier detection with space transformation and spectral analysis | |
| CN109871768B (en) | Hyperspectral optimal waveband selection method based on shared nearest neighbor | |
| Ding et al. | Classification of hyperspectral remote sensing images with support vector machines and particle swarm optimization | |
| Datta et al. | Clustering based band selection for hyperspectral images | |
| CN112884035A (en) | Noise robust hyperspectral image band selection method | |
| CN104123538B (en) | A kind of bad image detecting method of network of view-based access control model bag of words | |
| Le Bris et al. | A random forest class memberships based wrapper band selection criterion: Application to hyperspectral | |
| CN114663770A (en) | Hyperspectral image classification method and system based on integrated clustering waveband selection | |
| Jain et al. | An unsupervised band selection method for hyperspectral images using mutual information based dependence index | |
| Moon et al. | Image patch analysis and clustering of sunspots: A dimensionality reduction approach | |
| Demidova et al. | Improving the accuracy of the SVM classification using the Parzen classifier | |
| Borhani et al. | Hyperspectral image classification based on non-uniform spatial-spectral kernels | |
| Gu et al. | Spatial-spectral multiple kernel learning for hyperspectral image classification | |
| CN111539470A (en) | Image processing method, image processing device, computer equipment and storage medium | |
| Zhang et al. | SVM based feature screening applied to hierarchical cervical cancer detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| FG | Patent granted |
Effective date: 20230130 |