CN106650765A - Hyperspectral data classification method through converting hyperspectral data to gray image based on convolutional neural network - Google Patents
Hyperspectral data classification method through converting hyperspectral data to gray image based on convolutional neural network Download PDFInfo
- Publication number
- CN106650765A CN106650765A CN201610805454.5A CN201610805454A CN106650765A CN 106650765 A CN106650765 A CN 106650765A CN 201610805454 A CN201610805454 A CN 201610805454A CN 106650765 A CN106650765 A CN 106650765A
- Authority
- CN
- China
- Prior art keywords
- data
- hyperion
- convolutional neural
- sample
- gray
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
Abstract
The invention discloses a hyperspectral data classification method through converting hyperspectral data to a gray image based on a convolutional neural network, which relates to a hyperspectral data classification method and aims at solving problems that the existing high-spectral dimension image analysis and recognition precision is not matched with actual application demands, a mathematical model does not conform to the actual ground object distribution regulation and logic lacks. The hyperspectral data classification method through converting hyperspectral data to the gray image based on the convolutional neural network comprises steps: 1, original hyperspectral data are pre-processed to obtain hyperspectral vectors, and the hyperspectral vectors are converted to gray image data; and 2, texture features of samples in a sample set are learnt automatically through a convolution classification model, and gray image data samples are classified. The method is applied to the hyperspectral data classification field.
Description
Technical field
The present invention relates to Hyperspectral data classification method, more particularly to turns gray-scale maps based on the modal data of convolutional neural networks
Hyperspectral data classification method.
Background technology
Hyperspectral data classification is an application of high-spectrum remote-sensing, is all have spectral characteristic, light using all objects
Spectrum picture of the spectral resolution in the range of the 10l orders of magnitude is referred to as high spectrum image (Hyperspectral Image).And same
The situation of one spectral regions objects reaction is different, and reaction of the same object to different spectrum also has significant difference this feature to come to height
Spectral remote sensing data are classified.Past Hyperspectral data classification method is mainly by following two thinkings：Based on spectrum
The image classification matched somebody with somebody, the classification based on data statistics.Although having many sorting techniques answer by the two thinkings
On high-spectrum remote sensing, but, the aspect of many deficiencies in actual application and operation, is still shown, is such as instructed
Practice the demand of high cost, high resolution information waste, the graphical analyses of existing high spectrum dimension and identification accuracy and practical application
Not the problems such as mismatch, mathematical model do not meet actual atural object regularity of distribution shortcoming logic.Especially spectral Dimensions are continuously increased,
So that existing data analysis capabilities gradually do not catch up with the step of high spectrum dimension quantity of information.
The content of the invention
The invention aims to solve graphical analyses and identification accuracy and the practical application of existing high spectrum dimension
Not the problems such as demand mismatch, mathematical model do not meet the actual atural object regularity of distribution and are short of logic, and proposition is refreshing based on convolution
The high-spectral data of Jing networks turns the Hyperspectral data classification method of gray-scale maps.
Above-mentioned goal of the invention is achieved through the following technical solutions：
Step one, the spectrum vector that pretreatment obtains EO-1 hyperion is carried out to EO-1 hyperion initial data, by the spectrum of EO-1 hyperion vector
Data conversion becomes greyscale image data；
Step carries out successively normalization one by one, to EO-1 hyperion initial data, obtains the high-spectral data after normalization；
The all-wave segment value of each pixel in high-spectral data after step one two, extraction normalization, by each pixel
All-wave segment value composition one-dimensional vector data
In step one three, high-spectral data after normalization, other pixel repeat steps one or two are traveled through, extract altogether W ×
The one-dimensional pixel wave spectrum vector of L 1 × H obtains the spectrum vector of EO-1 hyperion；
Step one four, will EO-1 hyperion spectrum vector in each pixel wave spectrum vector be 1 × H the equal transposition of one-dimensional vector
Become two-dimensional matrix and obtain W × L two-dimensional matrix, and W × L two-dimensional matrix preservation is become containing W × L gray scale picture
Sample set is greyscale image data；
Step 2, by the textural characteristics of sample in convolution disaggregated model autonomic learning sample set, to greyscale image data
Sample is classified.
Invention effect
The present invention is the Hyperspectral data classification method for turning gray-scale maps based on the modal data of convolutional neural networks, in EO-1 hyperion
The theoretical and model of convolutional neural networks is introduced in image classification task, spectrum dimensional vector data conversion is become into gray scale diagram form
2-D data, and attempt to go to understand the abundant information amount entrained by high-spectral data with the language of image conversion, to EO-1 hyperion number
According to being classified.The method classification accuracy rate is high, goes forward side by side for hyperspectral information and autonomic learning its abstract characteristics are made full use of
Row classification has great significance.
Spectrum information is converted into gray scale pictorial information can become the data conversion of traditional vector form with stricture of vagina
The two dimensional image of reason feature, its abundant textural characteristics can be very good to react the data variation between high-spectral data spectral coverage, together
When, vector data is converted into into two-dimensional image data, the dimensionality reduction carried out by reducing data volume can be avoided to operate.The opposing party
Face, is processed to the texture information after conversion using convolutional network, can pass through multilamellar convolutional network independently learning data
Entrained abstract characteristics.The method contributes to while the correct classification of high-spectral data is realized, improves the correct of classification
The utilization rate of the abundant information entrained by rate, EO-1 hyperion.
Table 4 respectively in KSC, Pavia U data using CNN- gray-scale maps, CNN- oscillograms, linear kernel support vector machine,
RBF kernel support vectors machines, PCA conversion support vector machine, automatic encoding and logistic regression SAE-LR methods are classified, and contrast
Classification results.The average accuracy and overall accuracy of each classification are given in two above form, and is given respectively
Go out the classification accuracy rate of all subdivision species in two datasets, in order that result is apparent, in the classification side that invention is proposed
Method is obtained and is shown with runic at optimum.
Based on the sorting technique of the spectrum information gray level image of convolutional neural networks obtain on Pavia U data sets it is correct
Rate has surmounted the sorting technique based on support vector machine；However, the performance on KSC data sets is barely satisfactory, trace it to its cause,
It is that the selected network number of plies and parameter are all based on most stable equilibrium to consider, rather than obtains under each data set
Optimal result the network number of plies and parameter.In addition, convolutional neural networks are a probabilistic models, its classification accuracy rate substantially with
The increase of classification species and reduce.
The advantage of design
(1) one-dimensional data is converted into into 2-D data, bigger data volume and spectral Dimensions can be accommodated.
(2) by the method, improve the classification accuracy rate of high spectrum image.
(3) time used by categorizing process is shortened using GPU.
(4) using the abstract characteristics of depth convolutional network autonomic learning data, it is to avoid using the data model of shortcoming logic.
The present invention carries out pretreatment to data using Matlab, and one-dimensional data is converted into two dimensional gray image data,
Realize that convolutional neural networks model is classified to the high-spectral data after conversion by Caffe platforms and (SuSE) Linux OS,
And make use of GPU to accelerate experiment, to reduce the operation time consumption that huge amount of calculation is brought, shorten classification institute's used time
Between.
Description of the drawings
Fig. 1 is that the process of data preprocessing based on spectrum information gray-scale maps sorting technique that specific embodiment one is proposed is illustrated
Figure；Wherein, width of the W for EO-1 hyperion initial data；Length of the L for EO-1 hyperion initial data；Depths of the H for EO-1 hyperion initial data
Degree；Width of a for two-dimensional matrix；
Fig. 2 is that overall accuracy of the various methods of embodiment proposition on KSC data sets is rolled over number of training change
Line chart；
Fig. 3 is that overall accuracy of the various methods of embodiment proposition on Pavia U data sets becomes with number of training
Change broken line graph；
Fig. 4 (a) for embodiment propose by taking Indianpines data sets as an example, No. 1 Alfalfa (alfalfa) gray scale
The visualization result figure of figure；
Fig. 4 (b) for embodiment propose by taking Indianpines data sets as an example, No. 4 Corn (Semen Maydiss) gray-scale maps it is visual
Change result figure；
Fig. 4 (c) for embodiment propose by taking Indianpines data sets as an example, No. 5 Grass-pasture (basic unit's agricultures
) the visualization result figure of gray-scale maps；
Fig. 4 (d) for embodiment propose by taking Indianpines data sets as an example, No. 6 Grass-trees (grass-trees) ash
The visualization result figure of degree figure；
Fig. 4 (e) for embodiment propose by taking Indianpines data sets as an example, No. 7 Grass-pasture-mowed (grass
Wooden grass cuts) the visualization result figure of gray-scale maps；
Specific embodiment
Specific embodiment one：The bloom for turning gray-scale maps based on the high-spectral data of convolutional neural networks of present embodiment
Modal data sorting technique, specifically prepares according to following steps：
Step one, the spectrum vector that pretreatment obtains EO-1 hyperion is carried out to EO-1 hyperion initial data, by the spectrum of EO-1 hyperion vector
Data conversion becomes greyscale image data；(input of convolutional neural networks needs two-dimensional image data), is by convolutional Neural net
Network is applied in Hyperspectral data classification task；
The spectrum information introduced based on convolutional neural networks below by the deciphering to process of data preprocessing is converted into ash
The sorting technique of degree image.Process of data preprocessing is as shown in Figure 1：
Step carries out successively normalization one by one, to EO-1 hyperion initial data, obtains the high-spectral data after normalization；
The all-wave segment value of each pixel in high-spectral data after step one two, extraction normalization, by each pixel
All-wave segment value composition one-dimensional vector data
In step one three, high-spectral data after normalization, other pixel repeat steps one or two are traveled through, extract altogether W ×
The one-dimensional pixel wave spectrum vector of L 1 × H obtains the spectrum vector of EO-1 hyperion；
Hyperspectral image data classification is carried out by target of pixel, and the one-dimensional vector for extracting represents certain pixel
In the data message of all-wave spectral coverage.
Step one four, will EO-1 hyperion spectrum vector in each pixel wave spectrum vector be 1 × H the equal transposition of one-dimensional vector
Become two-dimensional matrix and obtain W × L two-dimensional matrix, and by W × L two-dimensional matrix imwrite (during imwrite is matlab
One function) preserve and become the sample set i.e. greyscale image data containing W × L gray scale pictures, each pictures in sample set
The all-wave spectral coverage information of certain pixel is represented, certain pixel of single picture represents data of the pixel under specific band
Value is characterized.The textural characteristics of whole pictures are extremely obvious, it is sufficient to which the data message of the goal pels is expressed, and its texture
Feature has also reflected the change of data inter-layer information；
Step 2, by the textural characteristics of sample in convolution disaggregated model autonomic learning sample set, to greyscale image data
Sample is classified.
Present embodiment effect：
Present embodiment is the Hyperspectral data classification method for turning gray-scale maps based on the modal data of convolutional neural networks, in height
The theoretical and model of convolutional neural networks is introduced in spectrum picture classification task, spectrum dimensional vector data conversion is become into gray-scale maps
The 2-D data of form, and attempt to go to understand the abundant information amount entrained by high-spectral data with the language of image conversion, to bloom
Modal data is classified.The method classification accuracy rate is high, for making full use of hyperspectral information and autonomic learning its abstract characteristics
And carry out classification and have great significance.
Spectrum information is converted into gray scale pictorial information can become the data conversion of traditional vector form with stricture of vagina
The two dimensional image of reason feature, its abundant textural characteristics can be very good to react the data variation between high-spectral data spectral coverage, together
When, vector data is converted into into two-dimensional image data, the dimensionality reduction carried out by reducing data volume can be avoided to operate.The opposing party
Face, is processed to the texture information after conversion using convolutional network, can pass through multilamellar convolutional network independently learning data
Entrained abstract characteristics.The method contributes to while the correct classification of high-spectral data is realized, improves the correct of classification
The utilization rate of the abundant information entrained by rate, EO-1 hyperion.
Table 4 respectively in KSC, Pavia U data using CNN- gray-scale maps, CNN- oscillograms, linear kernel support vector machine,
RBF kernel support vectors machines, PCA conversion support vector machine, automatic encoding and logistic regression SAE-LR methods are classified, and contrast
Classification results.The average accuracy and overall accuracy of each classification are given in two above form, and is given respectively
Go out the classification accuracy rate of all subdivision species in two datasets, in order that result is apparent, in the classification side that invention is proposed
Method is obtained and is shown with runic at optimum.
Based on the sorting technique of the spectrum information gray level image of convolutional neural networks obtain on Pavia U data sets it is correct
Rate has surmounted the sorting technique based on support vector machine；However, the performance on KSC data sets is barely satisfactory, trace it to its cause,
It is that the selected network number of plies and parameter are all based on most stable equilibrium to consider, rather than obtains under each data set
Optimal result the network number of plies and parameter.In addition, convolutional neural networks are a probabilistic models, its classification accuracy rate substantially with
The increase of classification species and reduce.
The advantage of design
(1) one-dimensional data is converted into into 2-D data, bigger data volume and spectral Dimensions can be accommodated.
(2) by the method, improve the classification accuracy rate of high spectrum image.
(3) time used by categorizing process is shortened using GPU.
(4) using the abstract characteristics of depth convolutional network autonomic learning data, it is to avoid using the data model of shortcoming logic.
Present embodiment carries out pretreatment to data using Matlab, and one-dimensional data is converted into two dimensional gray picture number
According to realizing that convolutional neural networks model is carried out to the high-spectral data after conversion point by Caffe platforms and (SuSE) Linux OS
Class, and make use of GPU to accelerate experiment, to reduce the operation time consumption that huge amount of calculation is brought, shorten classification used
Time.
Specific embodiment two：Present embodiment from unlike specific embodiment one：Step one by one described in height
Spectral data carries out pretreatment and is specially：Successively normalization is carried out to EO-1 hyperion initial data：
In formula,For the high-spectral data after normalization；For the EO-1 hyperion original number of kth layer (i, j) position
According to；Width of the W for EO-1 hyperion initial data；Length of the L for EO-1 hyperion initial data；Depth of the H for EO-1 hyperion initial data；W、
L, H value is positive integer；
The normalization mode of the present invention selects successively inner linear normalization, reason to have two：First, by data normalization 0
In the range of～1, it is easy to data imwrite subsequently into picture；Second, due to the image-forming principle of EO-1 hyperion, the number of different spectral coverage
It is excessive according to dimensional discrepancy, pixel wave spectrum transposition are being extracted into after image, partial spectrum segment information can be caused to be ignored.Other steps
And parameter is identical with specific embodiment one.
Specific embodiment three：Present embodiment from unlike specific embodiment one or two：Constitute in step one two
One-dimensional vector dataSpecially：
Wherein,For depth H (i, j) position EO-1 hyperion initial data；For the high-spectral data after normalization
In be located at (i, j) position pixel wave spectrum vector.Other steps and parameter are identical with specific embodiment one or two.
Specific embodiment four：Unlike one of present embodiment and specific embodiment one to three：It is right in step 2
Greyscale image data sample is classified specially：
The Small Sample Database of gray level image is classified using convolutional neural networks (CNN)；Classifying quality is carried out
Analysis, finally completes the comparison of the spectrum picture classification based on CNN and additive method；Sample data includes Small Sample Database and big
Sample data.Other steps and parameter are identical with one of specific embodiment one to three.
Beneficial effects of the present invention are verified using following examples：
Embodiment one：
The present embodiment turns the Hyperspectral data classification method of gray-scale maps, the party based on the high-spectral data of convolutional neural networks
What method was specifically prepared according to following steps：
Step one, the spectrum vector that pretreatment obtains EO-1 hyperion is carried out to EO-1 hyperion initial data, by the spectrum of EO-1 hyperion vector
Data conversion becomes greyscale image data；(input of convolutional neural networks needs two-dimensional image data), is by convolutional Neural net
Network is applied in Hyperspectral data classification task；
The spectrum information introduced based on convolutional neural networks below by the deciphering to process of data preprocessing is converted into ash
The sorting technique of degree image.Process of data preprocessing is as shown in Figure 1：
Step carries out successively normalization one by one, to EO-1 hyperion initial data, obtains the high-spectral data after normalization；
It is described pretreatment is carried out to EO-1 hyperion initial data to be specially：Successively normalization is carried out to EO-1 hyperion initial data：
In formula,For the high-spectral data after normalization；For the EO-1 hyperion original number of kth layer (i, j) position
According to；Width of the W for EO-1 hyperion initial data；Length of the L for EO-1 hyperion initial data；Depth of the H for EO-1 hyperion initial data；W、
L, H value is positive integer；
The normalization mode of the present invention selects successively inner linear normalization, reason to have two：First, by data normalization 0
In the range of～1, it is easy to data imwrite subsequently into picture；Second, due to the image-forming principle of EO-1 hyperion, the number of different spectral coverage
It is excessive according to dimensional discrepancy, pixel wave spectrum transposition are being extracted into after image, partial spectrum segment information can be caused to be ignored.
The all-wave segment value of each pixel in high-spectral data after step one two, extraction normalization, by each pixel
All-wave segment value composition one-dimensional vector dataSpecially：
Wherein,For depth H (i, j) position EO-1 hyperion initial data；For the high-spectral data after normalization
In be located at (i, j) position pixel wave spectrum vector.
In step one three, high-spectral data after normalization, other pixel repeat steps one or two are traveled through, extract altogether W ×
The one-dimensional pixel wave spectrum vector of L 1 × H obtains the spectrum vector of EO-1 hyperion；
Hyperspectral image data classification is carried out by target of pixel, and the one-dimensional vector for extracting represents certain pixel
In the data message of all-wave spectral coverage.
Step one four, will EO-1 hyperion spectrum vector in each pixel wave spectrum vector be 1 × H the equal transposition of one-dimensional vector
Become two-dimensional matrix and obtain W × L two-dimensional matrix, and by W × L two-dimensional matrix imwrite (during imwrite is matlab
One function) preserve and become the sample set i.e. greyscale image data containing W × L gray scale pictures, each pictures in sample set
The all-wave spectral coverage information of certain pixel is represented, certain pixel of single picture represents data of the pixel under specific band
Value is characterized.The textural characteristics of whole pictures are extremely obvious, it is sufficient to which the data message of the goal pels is expressed, and its texture
Feature has also reflected the change of data inter-layer information；
Step 2, by the textural characteristics of sample in convolution disaggregated model autonomic learning sample set, to greyscale image data
Sample is classified；
The Small Sample Database of 50～200 gray level images is classified using convolutional neural networks (CNN)；To classification
Effect is analyzed, and finally completes the comparison of the spectrum picture classification based on CNN and additive method.
Experimental program and data prediction result
The high-spectral data storehouse that experimental section is intended using is respectively：KSC、Pavia U、Indianpines.
The Hyperspectral data classification method based on CNN introduced using the present invention, the present embodiment is by little training sample
When classification accuracy rate based on CNN methods be compared with KNMF, RBF SVM methods, analyze the method based on CNN in little training
Performance situation during sample.6 are taken afterwards:2:2 data sample division proportions are tested, and are based on by contrast experiment
The research of classification capacity of the CNN methods in terms of Hyperspectral data classification, the training set of all generations, checking collection, survey in experiment
The picture that examination is concentrated is and is randomly assigned.
The gray level image obtained through series of preprocessing has obvious textural characteristics and fluctuation characteristic, with
As a example by the pre-processed results of Indianpines data sets, visualization result such as Fig. 4 (a)～(e) of its gray-scale map：
Small sample is analyzed
In order to analyze sensitivity of the disaggregated model based on convolutional neural networks to training samples number, following reality is carried out
Test.Compare the performance situations of the method in little training sample such as two methods based on CNN and KNMF, RBF SVM, experiment first
Carry out on above mentioned Pavia U data sets with KSC data sets, the training sample number of each classification elects 50 as,
100,150,200 are tested.KSC experimental results as shown in table 1, are depicted as broken line graph for Fig. 2.Pavia U are tested
As a result as shown in table 2, broken line graph is depicted as Fig. 3.
In 1 KSC data of table, CCN is contrasted with the overall accuracy of each method
In 2 Pavia U data of table, CNN is contrasted with the overall accuracy of each method
Can see from table 1 and Fig. 2 on KSC data sets two methods of CNN, KNMF, RBF-SVM it is overall correct
Rate all increases with the increase of training sample.Among this, when training sample is 50, the effect of RBF-SVM is still best
, the gray-scale maps method of CNN, the training result of KNMF and RBF-SVM is with the increase of training sample, its basic phase of trend for rising
Together, its each training sample number is very nearly the same, is because while from the point of view of at this moment each class training sample is individually seldom,
But KSC data sets effective sample itself is just seldom, when it is training sample that each class takes 200, which has accounted for population proportion very greatly,
So at this moment very nearly the same with the method such as RBF-SVM.
Can see on Pavia U data sets from table 2 and Fig. 3, the overall accuracy of CNN, KNMF and RBF-SVM method
All increase with the increase of training sample.Among this when training sample is 50 the effect of CNN gray-scale maps method be it is best,
But most fast, the RBF SVM when training sample increases to 100 that the training result of RBF-SVM rises with the increase of training sample
Overall accuracy with CNN gray-scale maps maintain an equal level.6 are adopted in next section Pavia U data sets:2:During 2 model split by
It is very big in training sample, it can be seen that the method that the method for CNN will exceed RBF-SVM.
Test more than comprehensive, as CNN is, based on probabilistic model, compared with Small Sample Database is taken, to take big
Sample data volume, increase amount of input information just can more play the advantage of the hyperspectral classification method based on convolutional neural networks.
Contrast with conventional sorting methods
The parameters of network in order to carry out contrast experiment, are adjusted on the basis of five layer networks of final choice, is passed through
Training obtains a CNN disaggregated model for having stability, and which is preferably linear with classification hyperspectral imagery performance respectively
The classification such as core SVM, the supporting vector machine model RBF-SVM based on automatic coding machine model SAE-LR, RBF core for proposing in the recent period
Method is contrasted.CNN disaggregated models will be classified to test data after repeatedly being trained by 20000 times.
Various sorting algorithm accuracy statistics on 3 KSC data sets of table
Table 3, table 4 respectively in KSC, Pavia U data using CNN- gray-scale maps, CNN- oscillograms, linear kernel support to
Amount machine, RBF kernel support vectors machines, PCA conversion support vector machine, automatic encoding and logistic regression SAE-LR methods are classified,
And comparison-of-pair sorting's result.The average accuracy and overall accuracy of each classification are given in two above form, and
The classification accuracy rate of all subdivision species in two datasets is sets forth, in order that result is apparent, is proposed in invention
Sorting technique is obtained and is shown with runic at optimum.
Various sorting algorithm accuracy statistics on 4 Pavia U data sets of table
Based on the sorting technique of the spectrum information gray level image of convolutional neural networks obtain on Pavia U data sets it is correct
Rate has surmounted the sorting technique based on support vector machine；However, the performance on KSC data sets is barely satisfactory, trace it to its cause,
It is that the selected network number of plies and parameter are all based on most stable equilibrium to consider, rather than obtains under each data set
Optimal result the network number of plies and parameter.In addition, convolutional neural networks are a probabilistic models, its classification accuracy rate substantially with
The increase of classification species and reduce.
The advantage of design
(1) one-dimensional data is converted into into 2-D data, bigger data volume and spectral Dimensions can be accommodated.
(2) by the method, improve the classification accuracy rate of high spectrum image.
(3) time used by categorizing process is shortened using GPU.
(4) using the abstract characteristics of depth convolutional network autonomic learning data, it is to avoid using the data model of shortcoming logic.
The present invention can also have other various embodiments, in the case of without departing substantially from spirit of the invention and its essence, this area
Technical staff works as and can make various corresponding changes and deformation according to the present invention, but these corresponding changes and deformation should all belong to
The protection domain of appended claims of the invention.
Claims (5)
1. the Hyperspectral data classification method of gray-scale maps is turned based on the high-spectral data of convolutional neural networks, it is characterised in that should
What method was specifically followed the steps below：
Step one, EO-1 hyperion initial data is carried out pretreatment obtain EO-1 hyperion spectrum vector, by the spectrum vector data of EO-1 hyperion
It is converted into greyscale image data；
Step carries out successively normalization one by one, to EO-1 hyperion initial data, obtains the high-spectral data after normalization；
The all-wave segment value of each pixel in high-spectral data after step one two, extraction normalization, by the complete of each pixel
Band value constitutes one-dimensional vector data
In step one three, high-spectral data after normalization, other pixel repeat steps one or two are traveled through, and extract W × L 1 altogether
The one-dimensional pixel wave spectrum vector of × H obtains the spectrum vector of EO-1 hyperion；
Step one four, will EO-1 hyperion spectrum vector in each pixel wave spectrum vector be 1 × H the equal transposition of one-dimensional vector become
Two-dimensional matrix obtains W × L two-dimensional matrix, and W × L two-dimensional matrix preservation is become the sample containing W × L gray scale picture
Collection is greyscale image data；
Step 2, by the textural characteristics of sample in convolution disaggregated model autonomic learning sample set, to greyscale image data sample
Classified.
2. the Hyperspectral data classification side of gray-scale maps is turned based on the high-spectral data of convolutional neural networks according to claim 1
Method, it is characterised in that：Step one by one described in pretreatment carried out to EO-1 hyperion initial data be specially：
Successively normalization is carried out to EO-1 hyperion initial data：
In formula,For the high-spectral data after normalization；For the EO-1 hyperion initial data of kth layer (i, j) position；W
For the width of EO-1 hyperion initial data；Length of the L for EO-1 hyperion initial data；Depth of the H for EO-1 hyperion initial data；W、L、H
Value is positive integer.
3. the Hyperspectral data classification side of gray-scale maps is turned based on the high-spectral data of convolutional neural networks according to claim 2
Method, it is characterised in that：One-dimensional vector data are constituted in step one twoSpecially：
Wherein,For depth H (i, j) position EO-1 hyperion initial data；To be located in the high-spectral data after normalization
The wave spectrum vector of (i, j) position pixel.
4. the Hyperspectral data classification side of gray-scale maps is turned based on the high-spectral data of convolutional neural networks according to claim 3
Method, it is characterised in that：Greyscale image data sample is classified specially in step 2：
The sample data of gray level image is classified using convolutional neural networks；Sample data includes Small Sample Database and big
Sample data.
5. the high-spectral data based on convolutional neural networks according to claim 1 or claim 2 turns the high-spectral data point of gray-scale maps
Class method, it is characterised in that：Greyscale image data sample is classified specially in step 2：
The sample data of gray level image is classified using convolutional neural networks.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610805454.5A CN106650765B (en) | 2016-09-06 | 2016-09-06 | High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610805454.5A CN106650765B (en) | 2016-09-06 | 2016-09-06 | High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106650765A true CN106650765A (en) | 2017-05-10 |
| CN106650765B CN106650765B (en) | 2019-06-11 |
Family
ID=58851895
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610805454.5A Active CN106650765B (en) | 2016-09-06 | 2016-09-06 | High-spectral data based on convolutional neural networks turns the Hyperspectral data classification method of grayscale image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106650765B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109447184A (en) * | 2018-11-28 | 2019-03-08 | 南京理工大学 | Android application network behavior classification method and system based on deep learning |
| CN109961073A (en) * | 2019-03-29 | 2019-07-02 | 云南电网有限责任公司电力科学研究院 | The acquisition methods and device of a kind of transmission line of electricity and shaft tower information |
| CN110998286A (en) * | 2017-07-31 | 2020-04-10 | 史密斯探测公司 | System for determining the presence of a substance of interest in a sample |
| CN111178425A (en) * | 2019-12-26 | 2020-05-19 | 哈尔滨工业大学 | Surface type classification method for large-scale high-speed rotation equipment based on convolutional neural network |
| CN114076637A (en) * | 2020-08-12 | 2022-02-22 | 舜宇光学（浙江）研究院有限公司 | Hyperspectral acquisition method and system, electronic equipment and coding wide-spectrum imaging device |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104102929A (en) * | 2014-07-25 | 2014-10-15 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
| WO2015017796A3 (en) * | 2013-08-02 | 2015-04-02 | Digimarc Corporation | Learning systems and methods |
| WO2015077493A1 (en) * | 2013-11-20 | 2015-05-28 | Digimarc Corporation | Sensor-synchronized spectrally-structured-light imaging |
| CN105320965A (en) * | 2015-10-23 | 2016-02-10 | 西北工业大学 | Hyperspectral image classification method based on spectral-spatial cooperation of deep convolutional neural network |
| CN105865624A (en) * | 2016-03-22 | 2016-08-17 | 清华大学 | Spectrum extraction method and device for hyperspectral collection system |
-
2016
- 2016-09-06 CN CN201610805454.5A patent/CN106650765B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2015017796A3 (en) * | 2013-08-02 | 2015-04-02 | Digimarc Corporation | Learning systems and methods |
| WO2015077493A1 (en) * | 2013-11-20 | 2015-05-28 | Digimarc Corporation | Sensor-synchronized spectrally-structured-light imaging |
| CN104102929A (en) * | 2014-07-25 | 2014-10-15 | 哈尔滨工业大学 | Hyperspectral remote sensing data classification method based on deep learning |
| CN105320965A (en) * | 2015-10-23 | 2016-02-10 | 西北工业大学 | Hyperspectral image classification method based on spectral-spatial cooperation of deep convolutional neural network |
| CN105865624A (en) * | 2016-03-22 | 2016-08-17 | 清华大学 | Spectrum extraction method and device for hyperspectral collection system |
Non-Patent Citations (1)
| Title |
|---|
| 谢传奇: ""基于高光谱成像技术的茄子叶片色差值检测和早疫病识别方法研究"", 《中国博士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110998286A (en) * | 2017-07-31 | 2020-04-10 | 史密斯探测公司 | System for determining the presence of a substance of interest in a sample |
| CN110998286B (en) * | 2017-07-31 | 2023-06-13 | 史密斯探测公司 | System for determining the presence of a substance of interest in a sample |
| CN109447184A (en) * | 2018-11-28 | 2019-03-08 | 南京理工大学 | Android application network behavior classification method and system based on deep learning |
| CN109961073A (en) * | 2019-03-29 | 2019-07-02 | 云南电网有限责任公司电力科学研究院 | The acquisition methods and device of a kind of transmission line of electricity and shaft tower information |
| CN111178425A (en) * | 2019-12-26 | 2020-05-19 | 哈尔滨工业大学 | Surface type classification method for large-scale high-speed rotation equipment based on convolutional neural network |
| CN114076637A (en) * | 2020-08-12 | 2022-02-22 | 舜宇光学（浙江）研究院有限公司 | Hyperspectral acquisition method and system, electronic equipment and coding wide-spectrum imaging device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106650765B (en) | 2019-06-11 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106650765A (en) | Hyperspectral data classification method through converting hyperspectral data to gray image based on convolutional neural network | |
| Cecotti et al. | Grape detection with convolutional neural networks | |
| CN112836610B (en) | Land use change and carbon reserve quantitative estimation method based on remote sensing data | |
| CN105608433B (en) | A kind of hyperspectral image classification method based on nuclear coordination expression | |
| CN105046276B (en) | Hyperspectral image band selection method based on low-rank representation | |
| CN110119753B (en) | Lithology recognition method by reconstructed texture | |
| CN111639587B (en) | Hyperspectral image classification method based on multi-scale spectrum space convolution neural network | |
| CN103336942A (en) | Traditional Chinese painting identification method based on Radon BEMD (bidimensional empirical mode decomposition) transformation | |
| CN107392140A (en) | Polarization SAR terrain classification method with incremental learning is decomposed based on low-rank sparse | |
| CN110689065A (en) | Hyperspectral image classification method based on flat mixed convolution neural network | |
| CN113435254A (en) | Sentinel second image-based farmland deep learning extraction method | |
| Liu et al. | Deep learning based research on quality classification of shiitake mushrooms | |
| Shantkumari et al. | Grape leaf image classification based on machine learning technique for accurate leaf disease detection | |
| CN113486869B (en) | Method, device and medium for lithology identification based on unsupervised feature extraction | |
| CN104268535A (en) | Method for extracting features of two-dimensional image | |
| Gouiffès et al. | Body color sets: A compact and reliable representation of images | |
| CN109299295B (en) | Blue printing layout database searching method | |
| CN109460788B (en) | Hyperspectral image classification method based on low-rank-sparse information combination network | |
| Anavyanto et al. | EfficientNetV2M for Image Classification of Tomato Leaf Deseases | |
| Sarimole et al. | Classification of Durian Types Using Features Extraction Gray Level Co-Occurrence Matrix (Glcm) and K-Nearest Neighbors (Knn) | |
| CN113887656B (en) | Hyperspectral image classification method combining deep learning and sparse representation | |
| Zheng et al. | A large-scale hyperspectral dataset for flower classification | |
| CN112966781A (en) | Hyperspectral image classification method based on triple loss and convolutional neural network | |
| Xie et al. | Image retrieval based on the combination of region and orientation correlation descriptors | |
| Jin et al. | Sunflower seeds classification based on self-attention Focusing algorithm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |