CN106529484A - Combined spectrum and laser radar data classification method based on class-fixed multinucleated learning - Google Patents
Combined spectrum and laser radar data classification method based on class-fixed multinucleated learning Download PDFInfo
- Publication number
- CN106529484A CN106529484A CN201611009935.1A CN201611009935A CN106529484A CN 106529484 A CN106529484 A CN 106529484A CN 201611009935 A CN201611009935 A CN 201611009935A CN 106529484 A CN106529484 A CN 106529484A
- Authority
- CN
- China
- Prior art keywords
- class
- sample
- classification
- ground object
- nuclear matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention relates to a combined spectrum and laser radar data classification method based on class-fixed multinucleated learning, relates to combined spectrum image and laser radar data classification and solves problems that self data characteristics of each category are not considered and optimization efficiency is low existing in a multinucleated learning in the prior art. The method specifically comprises steps that 1, the spectrum information and the space information of ground objects are extracted on multiple/high spectrum images, and the height information of the ground objects is extracted from the laser radar data; 2, a total characteristic vector corresponding to each ground object sample is formed in a characteristic stacking mode according to the extracted multiple/high spectrum images and the laser radar data; 3, an expression described as the specifications is sequentially inputted to a Gaussian kernel under different kernel parameters to generate a kernel matrix set; 4, a kernel alignment model is utilized to select a corresponding kernel matrix and a kernel parameter from the kernel set for two classes of ground object samples; and 5, lastly, sample classes are determined through utilizing a majority voting principle, and a classification theme graph is acquired. The method is applicable to the digital image processing field.
Description
Technical field
The invention belongs to digital image processing field, is related to spectrum picture and laser radar data joint classification.
Background technology
The abundant spectral information of atural object that spectrum picture is provided and the object height information that laser radar data is provided have letter
The complementary feature of breath.Joint can obtain target empty-spectrum in scene-elevation various dimensions Heterogeneous Information using two kinds of data sources, therefore join
Conjunction has wide application demand in fields such as urban mapping, agriculture test, Disaster Assessments using two kinds of data sources.
Multiple Kernel Learning model does not consider the physical meaning between each feature during model foundation, therefore has good
Heterogeneous characteristic fusion faculty.Existing Multiple Kernel Learning method can be summarized as three classes.The first kind is the multinuclear based on fixed criterion
Learning method, such method using the fixed criterion such as adding, taking advantage of be combined to base core, are not accounted for each during Multiple Kernel Learning
Individual base core itself similarity measurements flow characteristic and any Optimality Criteria is utilized, therefore classification performance is relatively low.Equations of The Second Kind is a step
Parameter in base core weight parameter and grader in many nuclear models is solved by optimization Multiple Kernel Learning method, such method simultaneously, because
This amount of calculation is very big, and the learning efficiency is very low, and its classification performance is relatively strong.3rd class is two steps optimization Multiple Kernel Learning side
Base core weight in many nuclear models and classifier parameters are separately independently solved by method, such method, therefore relative Equations of The Second Kind method
Tend to improve computational efficiency.Generally in the case of base core scale is less, Equations of The Second Kind Multiple Kernel Learning classification performance is better than the
Three class Multiple Kernel Learning methods;But in the case of base core is larger, the 3rd class Multiple Kernel Learning classification performance is better than second
Class Multiple Kernel Learning method.What existing Multiple Kernel Learning method was considered during model solution is directed under given Optimality Criteria
The optimal solution of whole data set all categories, without considering each classification own characteristic;While existing Multiple Kernel Learning method
Generally optimization problem is converted into Semidefinite Programming to solve, due to being related to the quick of many nuclear matrix during parametric solution
Solve, higher-dimension nuclear matrix is decomposed, cause Multiple Kernel Learning efficiency very low the problems such as iterationses are high.
Existing Multiple Kernel Learning method is present at present, does not account for each classification itself in multinuclear model solution process special
Point, and optimization problem complexity, the low problem of the learning efficiency.For this problem for existing, the present invention utilizes each classification number
Nuclear matrix selection is carried out according to own characteristic, improves classification performance, simultaneously because solution need not be iterated, therefore efficiency of algorithm
It is high.
The content of the invention
The present invention is to solve in existing Multiple Kernel Learning method, the low problem of image classification accuracy, and proposes that one kind is based on class
The spectrum and laser radar data joint classification method of specified Multiple Kernel Learning.
A kind of spectrum and laser radar data joint classification method that Multiple Kernel Learning is specified based on class, is entered according to the following steps
OK:
First, spectral information and spatial information are extracted on many/high spectrum image, height is extracted on laser radar data and is believed
Breath:
Wherein spectral information includes spectral band and by spectral band inverting normalized differential vegetation index out, normalizing
Change building index, normalization water body index, the feature for representing spectral information is designated asWherein
Representative sample sum, NcRepresent c class sample numbers, dSpeRepresent spectral signature dimension;Spatial information include by using average,
Variance, morphology and Gabor spatial filters are filtered the space characteristics of generation to many/high spectrum image, will represent space
The feature of information is designated asWherein dSpaRepresent space characteristics dimension;Atural object is being extracted to laser radar data
Before elevation information, by laser radar point cloud altitude information be tiled into many/high spectrum image have same spatial resolution and
The image of spacial alignment.Then normalization digital surface model is extracted again as elevation information, will represent the feature of elevation information
It is designated asWherein dEleRepresent altitude feature dimension.
2nd, spectrum, space and the altitude feature of many/high spectrum image and laser radar data will be extracted from feature storehouse
Mode forms the corresponding general characteristic vector of each ground object sample, is designated asdSta=dSpe+
dSpa+dEle。
3rd, the storehouse characteristic vector for obtaining every two classes sample will be extractedIt is sequentially inputted in different kernel functions
Generate nuclear matrix collection.
4th, using core Alignment model, to be every two classes atural object concentrate from nuclear matrix selects corresponding nuclear matrix and nuclear parameter,
NoteIt is pth and q classes in nuclear parameter isσmWhen nuclear matrix.
5th, select to solve the problems, such as the support vector machine of multicategory classification as grader with " one-to-one " strategy, " one-to-one "
Classified using its corresponding nuclear matrix in categorizing process, finally sample class is determined using most Voting principles, most throughout one's life
Constituent class thematic map.
The present invention includes following beneficial effect:
1st, due to considering each classification its data feature during Multiple Kernel Learning, and with " one-to-one " classification policy
Combine so that the nuclear matrix used in categorizing process is adapted with each classification, therefore can improve classification performance, compares monokaryon
Classification precision improves 2.5% or so, improves 1.5% or so than the Multiple Kernel Learning classification precision of main flow;
2nd, due to directly determining classification nuclear matrix using core technique of alignment, it is not necessary to be optimized solution to many nuclear models,
Therefore algorithm operation efficiency can be improved, being enhanced about more than once based on the Multiple Kernel Learning method operation efficiency of optimization than main flow.
Description of the drawings
Fig. 1 is the method for the invention schematic flow sheet;
Fig. 2 is BayviewPark data sets RGB pseudo color composing figures (left side) and the true figure (right side) of atural object distribution;
Fig. 3 is Recology data sets RGB pseudo color composing figures (left side) and the true figure (right side) of atural object distribution;
Fig. 4 a are Bayview Park data set nicety of grading figures;
Fig. 4 b are Bayview Park data set Kappa coefficient figures;
Fig. 5 a are Recology data set nicety of grading figures;
Fig. 5 b are Recology data set Kappa coefficient figures;
Fig. 6 a are Bayview Park data set SK classification results thematic map;
Fig. 6 b are Bayview Park data set SimpleMKL classification results thematic map;
Fig. 6 c are Bayview Park data set MeanMKL classification results thematic map;
Fig. 6 d are Bayview Park data set RMKL classification results thematic map;
Fig. 6 e are Bayview Park data set C2MKL classification results thematic map;
Fig. 6 f are the true atural object scattergram classification results thematic map of Bayview Park data sets;
Fig. 7 a are Recology data set classification results thematic map;
Fig. 7 b are Recology data set SK classification results thematic map;
Fig. 7 c are Recology data set SimpleMKL classification results thematic map;
Fig. 7 d are Recology data set MeanMKL classification results thematic map;
Fig. 7 e are Recology data set RMKL classification results thematic map;
Fig. 7 f are Recology data set C2MKL classification results thematic map.
Specific embodiment
Specific embodiment one, present embodiment is illustrated with reference to Fig. 1, present embodiment specifies Multiple Kernel Learning based on class
Spectrum and laser radar data joint classification method are carried out according to the following steps:
Step one, on many/high spectrum image extract atural object spectral information and spatial information, on laser radar data
Extract object height information:
Wherein, spectral information includes spectral band and by spectral band inverting normalized differential vegetation index out, normalizing
Change building index, normalization water body index, computing formula is as follows
In formula, reflectance of the NIR near infrared band, near infrared band scope is 780~2500nm, and R is red band
Reflectance, red band scope is 640~680nm, and MIR is the reflectance of middle-infrared band, and middle-infrared band scope is 2.5
~6um, G are the reflections of green band, and green band scope is 500~540nm;
The feature for representing spectral information is designated asWhereinRepresent ground object sample sum, Nc
Represent c class ground object sample numbers, dSpeSpectral signature dimension is represented,For react i-th ground object sample spectral information feature to
Amount,Dimension is represented as dSpeSpace, C be ground object sample classification number;
Spatial information includes carrying out many/high spectrum image by using average, variance, morphology and Gabor filter
The spatial information feature that filtering is generated, the feature for representing spatial information is designated asWherein dSpaRepresent space
Intrinsic dimensionality,Dimension is represented as dSpaSpace,To react the characteristic vector of i-th ground object sample spatial information;
Before object height information is extracted to laser radar data, by laser radar point cloud altitude information (laser radar number
According to what is carried, this is input data) it is tiled into the figure that there is same spatial resolution and spacial alignment with many/high spectrum image
Picture, then again to the image zooming-out normalization digital surface model that obtains after laser radar point cloud altitude information rasterizing as height
Degree information, the feature for representing elevation information is designated asWherein dEleAltitude feature dimension is represented,Represent
Dimension is dEleSpace,To react the characteristic vector of i-th ground object sample elevation information;
Step 2, the feature of the spectral information of many/high spectrum image and laser radar data, spatial information will be extracted from
The feature of feature and elevation information forms the corresponding general characteristic vector of each ground object sample with feature stacked manner, is designated asdSta=dSpe+dSpa+dEle;General characteristic for i-th ground object sample is vectorial,Generation
Table dimension is dStaSpace, dStaFor the total dimension of feature of ground object sample;
Step 3, extraction is obtained every two class (assume total C classes, then just there are (C × (C-1))/2 every two classes, be not
Each two atural object two, is the general characteristic collection of every two classes ground object sample composition.Here there are many samples per class atural object.)
The storehouse characteristic vector of thing training sampleThe Gauss karyogenesis nuclear matrix collection being sequentially inputted under different nuclear parameters;
NpRepresent pth class ground object sample number, NqRepresent q class ground object sample numbers;It is the sample being made up of pth and two class ground object samples of q
The storehouse characteristic vector of i-th sample of this concentration;
Step 4, using core Alignment model be every two classes ground object sample concentrate from nuclear matrix select corresponding nuclear matrix and
Nuclear parameter, noteIt is pth and q classes in nuclear parameter isσmWhen nuclear matrix;
Step 5, selection solve the problems, such as the support vector machine of multicategory classification as grader with one-to-one strategy, at " a pair
One " classified using its corresponding nuclear matrix in categorizing process, finally sample class is determined using most Voting principles, finally
Generate classification scheme figure.
Multispectral image is the image for same target being shot repeatedly and being attained the Way by multiple wave bands, due in target
Each object is different to the sensitivity of the same band, therefore also has certain difference between each image in multispectral image
Part.
Spectrum picture of the spectral resolution in the range of the 10l orders of magnitude is referred to as high spectrum image (Hyperspectral
Image)。
Present embodiment includes following beneficial effect:
1st, due to considering each classification its data feature during Multiple Kernel Learning, and with " one-to-one " classification policy
Combine so that the nuclear matrix used in categorizing process is adapted with each classification, therefore can improve classification performance, compares monokaryon
Classification precision improves 2.5% or so, improves 1.5% or so than the Multiple Kernel Learning classification precision of main flow;
2nd, due to directly determining classification nuclear matrix using core technique of alignment, it is not necessary to be optimized solution to many nuclear models,
Therefore algorithm operation efficiency can be improved, being enhanced about more than once based on the Multiple Kernel Learning method operation efficiency of optimization than main flow.
Specific embodiment two:Present embodiment from unlike specific embodiment one:To extract in the step 3
Obtain the storehouse feature of every two class (assuming total C classes, then just have (C × (C-1))/2 every two classes) atural object training sample to
AmountThe Gauss karyogenesis nuclear matrix collection being input under different nuclear parameters;Detailed process is:
Described nuclear matrix collection is the storehouse characteristic vector of every two classes atural object training sampleIn different nuclear parameters
Under gaussian kernel, computing formula is:
The data set of every two classes ground object sample compositionIt is the input data of the gaussian kernel of different nuclear parameters, it is defeated
Go out result for nuclear matrix collection;
WhereinI-th ground object sample in the data set being made up of pth and two class atural object training samples of q is represented,
J-th ground object sample in the data set being made up of pth and two class atural object training samples of q is represented, i, j span is 1 to Np+
Nq, Np+NqIt is two class atural object number of training sum of pth and q, σmFor m-th nuclear parameter,It is that pth and q classes exist
Nuclear parameter isσmWhen nuclear matrix;NpFor positive integer, NqFor positive integer;
In assuming image, (image here refers to the scene of many/EO-1 hyperion and laser radar data acquisition, C classes here
Atural object refers to that how many class atural object in scene, many/high spectrum image and laser radar are to carry out data acquisition to Same Scene)
Total C classes ground object sample, then have (C × (C-1))/2 pairs of atural objects, forms M × ((C × (C-1))/2) individual nuclear matrix altogether
Other steps and parameter are identical with specific embodiment one.
Specific embodiment three:Present embodiment from unlike specific embodiment one or two:Profit in the step 4
It is that every two classes atural object selects corresponding nuclear matrix and nuclear parameter from nuclear matrix concentration with core Alignment model, remembersFor pth and q
Class is σ in nuclear parametermWhen nuclear matrix;Detailed process is:
Described core technique of alignment is:
Wherein<·,·>FRepresent Frobenius (not Luo Beini crows this norm/Euclid norms/F- norms/E- norms)
Product,<Ki,KIdeal>F=Tr (Ki×KIdeal),<Ki,Ki>F=Tr (Ki×Ki),<KIdeal,KIdeal>F=Tr (KIdeal×KIdeal),
Tr () represents the mark for taking matrix, KiFor nuclear matrix collectionIn i-th nuclear matrix, KIdealIt is by training sample mark
The preferable nuclear matrix for constituting is signed, its building method is as follows:
Wherein, yiFor ground object sample xiClass label, yjFor ground object sample xjClass label;
Using above-mentioned core technique of alignment to the data set that is made up of pth and two class ground object samples of qIn each core
The nuclear matrix generated under parameter is calculated, and obtains the alignment value KA of each nuclear parameter correspondence nuclear matrix and preferable core, further according to
KA values maximum principle selects pth and the corresponding nuclear matrix of two classes of qWith core yardstickσm。
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:The nuclear parameter
σmFor [0.05:0.05:2], nuclear matrix is 40.
Other steps and parameter are identical with one of specific embodiment one to three.
Specific embodiment five, present embodiment is to unlike one of specific embodiment one to four:The step
One-to-one strategy in five referred to when multicategory classification is carried out, and multicategory classification problem is changed into multiple two classes classification problems, most
Multicategory classification is realized by multiple two classes classification results are carried out with most ballots afterwards.
Such as one C class classification problem, will be converted to (C × (C-1))/2 two classes classification under " one-to-one " strategy
Processed.
Selection solves the problems, such as the support vector machine of multicategory classification as grader with " one-to-one " strategy, in " one-to-one " classification
During classified using its corresponding nuclear matrix, finally determine sample class using most Voting principles, ultimately generate point
Class thematic map.
Specific embodiment six, present embodiment is to unlike one of specific embodiment one to five:The step
Classification is carried out using its corresponding nuclear matrix during one-against-one in five to refer to for a ground object sample to be sorted, when
Assume which belongs to two class of pth and q, and when classifying to which, using what is selected in step 4Nuclear matrix and core chi
DegreeσmClassified;Classification is carried out as follows:
A, hypothesis ground object sample x to be sorted are a class of two apoplexy due to endogenous wind of pth and q, solve following primal-dual optimization problem
WhereinFor supporting vector weight sets, n is supporting vector number,For training sample tally set;αiTo prop up
Hold vector xiWeighted value, αjFor supporting vector xjWeighted value, Q (αi,αj) for the majorized function of support vector machine classifier, Np
For pth class atural object number of training, NqFor q class atural object number of training, yiFor training sample xiClass label, yjFor instruction
Practice sample xjClass label,Be pth and two class atural object training sample set of q classes be σ in core scale parametermWhen nuclear matrix,
N is two class atural object number of training sum of pth and q classes,Represent i and j and take 1 Arbitrary Digit in N;N is for just
Integer;
Here it is the training process of support vector machine classifier, only uses training sample, is not related to sample to be sorted, here
Purpose is to solve for the supporting vector weight sets for obtainingFinally according to this supporting vector weight setsDetermine to be sorted
Ground object sample classification.
B, classified using following categorised decision function pair sample x to be sorted
WhereinIt is to solve the supporting vector collection for obtaining in step a;For supporting vector xiWeight, b*To bias,
b*It is calculated by given criterion;X is ground object sample to be sorted, xiAnd xjIt is training sample, using training sample to be sorted
Ground object sample is classified;N is positive integer;
C, p and q is recycled to into C, repeat step a and b from 1 respectively;
D, using step a, the two class classification results f of (C × (C-1))/2 obtained in b and cp,qMost ballots are carried out, most
The classification of sample x is determined eventually.
Beneficial effects of the present invention are verified using following examples:
The present embodiment based on class specify Multiple Kernel Learning spectrum and laser radar data joint classification method specifically according to
Prepared by following steps:
The effect of the present invention can be further illustrated by following experiment.
1st, experimental data
This experiment adopt two group data sets be BayviewPark and Recology data sets, this two group data set from
Two pieces of subgraphs in san francisco, usa urban district, wherein Bayview Park data sets derive from one, the urban district public affairs for being named as Bayview
Garden, Recology data sets are from a factory for being named as Recology.This two group data set is used for IEEE GRSS numbers in 2012
According to fusion contest.Spectrum picture is shot on October 9th, 2011 by WorldView2 satellites, and spectral region is from 400-
1040nm, obtains 8 wave bands.LiDAR data is taken in June, 2010.Two width remote sensing image spatial resolution are 1.8
Rice.Data area is observed by Google Earth, classifies to observing atural object in scene according to object height and material,
Bayview Park data sets are divided into 7 class atural objects, are respectively:It is building 1, building 2, building 3, road, trees, naked
Ground and sea water;Recology data sets are divided into 11 classes, are respectively:Building 1, building 2, building 3, building 4, build
Build thing 5, building 6, building 7, trees, parking lot, bare area and meadow.Fig. 2 and Fig. 3 respectively show Bayview Park and
The RGB pseudo color composings figure of Recology data sets and the true figure of atural object distribution.The feature for being used includes 8 original wave bands,
Normalization digital surface model (nDSM) and 8 morphological features (MPs), when morphological feature is extracted, specially using 3 ×
3rd, 5 × 5 rectangular configuration element carries out space characteristics to spectrum first principal component and normalization digital surface model (nDSM) and carries
Take, through open and close computing, 8 morphological features are obtained per group data set, as shown in table 1.
1 different pieces of information collection feature service condition of table
2nd, experiment content
In an experiment, classified using (being designated as C2MKL) of the invention and existing four kinds of methods two groups of data to more than.
There are four kinds of methods to include monokaryon method (SK), simple Multiple Kernel Learning (SimpleMKL), representative Multiple Kernel Learning (RMKL) and averagely
Multiple Kernel Learning (MeanMKL).The core range scale of all Multiple Kernel Learning algorithms is set to [0.05:0.05:2], penalty coefficient C and
Slack variable ξ is obtained by cross validation.Select to be classified by 10,15,20,30,40,50,100 per class number of training respectively
Experiment.Ten operation result averages of record are used as final result, and count ten operation result variances.The assessment classification knot for using
The index of fruit includes overall accuracy (OverallAccuracy, OA) and Kappa coefficients.Two groups of value data classification results are as schemed
Classification results thematic map such as Fig. 6 a, Fig. 6 b, Fig. 6 c, Fig. 6 d, figure shown in 4a, Fig. 4 b and 5a, Fig. 5 b, when training sample is 50
Shown in 6e, Fig. 6 f and 7a, Fig. 7 b, Fig. 7 c, Fig. 7 d, Fig. 7 e, Fig. 7 f, in figure at rectangle frame for algorithms of different classification results difference compared with
Big part.
From experimental result it can be seen that the present invention can significantly improve nicety of grading, improve classification performance.
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 (6)
1. the spectrum and laser radar data joint classification method of Multiple Kernel Learning are specified based on class, it is characterised in that:Referred to based on class
The spectrum and laser radar data joint classification method detailed process for determining Multiple Kernel Learning is:
Step one, on many/high spectrum image extract atural object spectral information and spatial information, on laser radar data extract
Object height information:
Wherein, spectral information includes spectral band and is built by spectral band inverting normalized differential vegetation index out, normalization
Index, normalization water body index are built, computing formula is as follows
In formula, reflectance of the NIR near infrared band, near infrared band scope are 780~2500nm, and R is the anti-of red band
Penetrate rate, red band scope is 640~680nm, MIR is the reflectance of middle-infrared band, middle-infrared band scope is 2.5~
6um, G are the reflections of green band, and green band scope is 500~540nm;
The feature for representing spectral information is designated asWhereinGround object sample sum is represented,
NcRepresent c class ground object sample numbers, dSpeSpectral signature dimension is represented,To react the feature of i-th ground object sample spectral information
Vector,Dimension is represented as dSpeSpace, C be ground object sample classification number;
Spatial information includes being filtered many/high spectrum image by using average, variance, morphology and Gabor filter
The spatial information feature of generation, the feature for representing spatial information is designated asWherein dSpaRepresent space
Intrinsic dimensionality,Dimension is represented as dSpaSpace,To react the characteristic vector of i-th ground object sample spatial information;
Before object height information is extracted to laser radar data, laser radar point cloud altitude information is tiled into and many/height
Spectrum picture has the image of same spatial resolution and spacial alignment, then again to laser radar point cloud altitude information rasterizing
The feature for representing elevation information is designated as by the image zooming-out normalization digital surface model for obtaining afterwards as elevation informationWherein dEleAltitude feature dimension is represented,Dimension is represented as dEleSpace,For reaction i-th
The characteristic vector of individual ground object sample elevation information;
Step 2, the feature of the feature of the spectral information of many/high spectrum image and laser radar data, spatial information will be extracted from
The corresponding general characteristic vector of each ground object sample is formed with feature stacked manner with the feature of elevation information, is designated as For the general characteristic of i-th ground object sample
Vector,Dimension is represented as dStaSpace, dStaFor the total dimension of feature of ground object sample;
Step 3, will extract and obtain the storehouse characteristic vector of every two classes atural object training sampleIt is sequentially inputted to difference
Gauss karyogenesis nuclear matrix collection under nuclear parameter;NpRepresent pth class ground object sample number, NqRepresent q class ground object sample numbers;For
The storehouse characteristic vector of i-th sample in the sample set be made up of pth and two class ground object samples of q;
Step 4, it is that every two classes ground object sample is concentrated from nuclear matrix and selects corresponding nuclear matrix and core ginseng using core Alignment model
Number, noteBe pth and q classes be σ in nuclear parametermWhen nuclear matrix;
Step 5, selection solve the problems, such as the support vector machine of multicategory classification as grader with one-to-one strategy, in one-against-one
During classified using its corresponding nuclear matrix, finally determine sample class using most Voting principles, ultimately generate point
Class thematic map.
2. the spectrum and laser radar data joint classification method of Multiple Kernel Learning are specified based on class according to claim 1, its
It is characterised by:The storehouse characteristic vector for obtaining every two classes atural object training sample will be extracted in the step 3Input
Gauss karyogenesis nuclear matrix collection under different nuclear parameters;Detailed process is:
Described nuclear matrix collection is the storehouse characteristic vector of every two classes atural object training sampleUnder different nuclear parameters
Gaussian kernel, computing formula is:
The data set of every two classes ground object sample compositionIt is the input data of the gaussian kernel of different nuclear parameters, output knot
Fruit is nuclear matrix collection;
WhereinI-th ground object sample in the data set being made up of pth and two class atural object training samples of q is represented,Represent
J-th ground object sample in the data set be made up of pth and two class atural object training samples of q, i, j span are 1 to Np+Nq, Np
+NqIt is two class atural object number of training sum of pth and q, σmFor m-th nuclear parameter,It is that pth and q classes are joined in core
Number is σmWhen nuclear matrix;
When total C class ground object samples in high spectrum image, then have (C × (C-1))/2 pairs of atural objects, form M × ((C × (C- altogether
1))/2) individual nuclear matrix
3. the spectrum and laser radar data joint classification method of Multiple Kernel Learning are specified based on class according to claim 2, its
It is characterised by:In the step 4, using core Alignment model, to be every two classes atural object concentrate from nuclear matrix selects corresponding nuclear matrix
And nuclear parameter, noteBe pth and q classes be σ in nuclear parametermWhen nuclear matrix;Detailed process is:
Described core Alignment model is:
Wherein<·,·>FFrobenius products are represented,<Ki,KIdeal>F=Tr (Ki×KIdeal),<Ki,Ki>F=Tr (Ki×Ki),<
KIdeal,KIdeal>F=Tr (KIdeal×KIdeal), Tr () represents the mark for taking matrix, KiFor nuclear matrix collectionIn
I-th nuclear matrix, KIdealIt is the preferable nuclear matrix being made up of training sample label, its building method is as follows:
Wherein, yiFor ground object sample xiClass label, yjFor ground object sample xjClass label;
Using above-mentioned core technique of alignment to the data set that is made up of pth and two class ground object samples of qJoin in each core
The nuclear matrix of several lower generations is calculated, and obtains the alignment value KA of each nuclear parameter correspondence nuclear matrix and preferable core, further according to KA
Value maximum principle selects pth and the corresponding nuclear matrix of two classes of qWith core yardstick σm。
4. the spectrum and laser radar data joint classification method of Multiple Kernel Learning are specified based on class according to claim 3, its
It is characterised by:Nuclear parameter σmFor [0.05:0.05:2], nuclear matrix is 40.
5. the spectrum and laser radar data joint classification method of Multiple Kernel Learning are specified based on class according to claim 4, its
It is characterised by:One-to-one strategy in the step 5 is referred to when multicategory classification is carried out, multicategory classification problem is changed into many
Individual two classes classification problem, realizes multicategory classification finally by multiple two classes classification results are carried out with most ballots.
6. the spectrum and laser radar data joint classification method of Multiple Kernel Learning are specified based on class according to claim 5, its
It is characterised by:Classification is carried out using its corresponding nuclear matrix during one-against-one in the step 5 to refer to for one
Ground object sample to be sorted, when hypothesis, which belongs to two class of pth and q, and when classifying to which, using what is selected in step 4Nuclear matrix and core yardstick σmClassified;Classification is carried out as follows:
A, hypothesis ground object sample x to be sorted are a class of two apoplexy due to endogenous wind of pth and q, solve following primal-dual optimization problem
WhereinFor supporting vector weight sets, n is supporting vector number,For training sample tally set;αiFor support to
Amount xiWeighted value, αjFor supporting vector xjWeighted value, Q (αi,αj) for the majorized function of support vector machine classifier, NpFor
P class atural object number of training, NqFor q class atural object number of training, yiFor training sample xiClass label, yjTo train sample
This xjClass label,Be pth and two class atural object training sample set of q classes be σ in core scale parametermWhen nuclear matrix, N is
Two class atural object number of training sum of pth and q classes,Represent i and j and take 1 Arbitrary Digit in N;
B, classified using following categorised decision function pair sample x to be sorted
WhereinIt is to solve the supporting vector collection for obtaining in step a;For supporting vector xiWeight, b*For biasing;X is to treat
Classification ground object sample, xiAnd xjBe training sample, classification ground object sample is treated using training sample and classified;N is positive integer;
C, p and q is recycled to into C, repeat step a and b from 1 respectively;
D, using step a, the two class classification results f of (C × (C-1))/2 obtained in b and cp,qMost ballots are carried out, it is final true
The classification of random sample this x.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611009935.1A CN106529484A (en) | 2016-11-16 | 2016-11-16 | Combined spectrum and laser radar data classification method based on class-fixed multinucleated learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611009935.1A CN106529484A (en) | 2016-11-16 | 2016-11-16 | Combined spectrum and laser radar data classification method based on class-fixed multinucleated learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106529484A true CN106529484A (en) | 2017-03-22 |
Family
ID=58353370
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611009935.1A Pending CN106529484A (en) | 2016-11-16 | 2016-11-16 | Combined spectrum and laser radar data classification method based on class-fixed multinucleated learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106529484A (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108846352A (en) * | 2018-06-08 | 2018-11-20 | 广东电网有限责任公司 | A kind of vegetation classification and recognition methods |
CN109101977A (en) * | 2018-06-27 | 2018-12-28 | 深圳大学 | A kind of method and device of the data processing based on unmanned plane |
CN110163101A (en) * | 2019-04-17 | 2019-08-23 | 湖南省中医药研究院 | The difference of Chinese medicine seed and grade quick discrimination method |
WO2020000271A1 (en) * | 2018-06-27 | 2020-01-02 | 深圳大学 | Unmanned aerial vehicle-based data processing method and device |
CN111487643A (en) * | 2020-04-13 | 2020-08-04 | 中国科学院空天信息创新研究院 | Building detection method based on laser radar point cloud and near-infrared image |
CN111738298A (en) * | 2020-05-27 | 2020-10-02 | 哈尔滨工业大学 | Data classification method based on depth-width-variable multi-core learning |
CN112130169A (en) * | 2020-09-23 | 2020-12-25 | 广东工业大学 | Point cloud level fusion method for laser radar data and hyperspectral image |
CN113589257A (en) * | 2021-07-28 | 2021-11-02 | 哈尔滨工业大学 | Global feature-based classification method for full-waveform laser echo signals |
CN116933141A (en) * | 2023-09-15 | 2023-10-24 | 昆明理工大学 | Multispectral laser radar point cloud classification method based on multicore graph learning |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102708373A (en) * | 2012-01-06 | 2012-10-03 | 香港理工大学 | Method and device for classifying remote images by integrating space information and spectral information |
CN103034863A (en) * | 2012-12-24 | 2013-04-10 | 重庆市勘测院 | Remote-sensing image road acquisition method combined with kernel Fisher and multi-scale extraction |
CN103514456A (en) * | 2013-06-30 | 2014-01-15 | 安科智慧城市技术(中国)有限公司 | Image classification method and device based on compressed sensing multi-core learning |
CN103839084A (en) * | 2014-03-12 | 2014-06-04 | 湖州师范学院 | Multi-kernel support vector machine multi-instance learning algorithm applied to pedestrian re-identification |
CN103903007A (en) * | 2014-03-10 | 2014-07-02 | 哈尔滨工程大学 | Hyperspectral semi-supervised classification method based on space-spectral information |
CN104751477A (en) * | 2015-04-17 | 2015-07-01 | 薛笑荣 | Space domain and frequency domain characteristic based parallel SAR (synthetic aperture radar) image classification method |
CN106056591A (en) * | 2016-05-25 | 2016-10-26 | 哈尔滨工业大学 | Method for estimating urban density through fusion of optical spectrum image and laser radar data |
-
2016
- 2016-11-16 CN CN201611009935.1A patent/CN106529484A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102708373A (en) * | 2012-01-06 | 2012-10-03 | 香港理工大学 | Method and device for classifying remote images by integrating space information and spectral information |
CN103034863A (en) * | 2012-12-24 | 2013-04-10 | 重庆市勘测院 | Remote-sensing image road acquisition method combined with kernel Fisher and multi-scale extraction |
CN103514456A (en) * | 2013-06-30 | 2014-01-15 | 安科智慧城市技术(中国)有限公司 | Image classification method and device based on compressed sensing multi-core learning |
CN103903007A (en) * | 2014-03-10 | 2014-07-02 | 哈尔滨工程大学 | Hyperspectral semi-supervised classification method based on space-spectral information |
CN103839084A (en) * | 2014-03-12 | 2014-06-04 | 湖州师范学院 | Multi-kernel support vector machine multi-instance learning algorithm applied to pedestrian re-identification |
CN104751477A (en) * | 2015-04-17 | 2015-07-01 | 薛笑荣 | Space domain and frequency domain characteristic based parallel SAR (synthetic aperture radar) image classification method |
CN106056591A (en) * | 2016-05-25 | 2016-10-26 | 哈尔滨工业大学 | Method for estimating urban density through fusion of optical spectrum image and laser radar data |
Non-Patent Citations (4)
Title |
---|
WANG QINGWANG等: ""Discriminating multiple kernel learning for joint classification of optical and LiDAR data in urban area"", 《2015 7TH WORKSHOP ON HYPERSPECTRAL IMAGE AND SIGNAL PROCESSING: EVOLUTION IN REMOTE SENSING (WHISPERS)》 * |
YANFENG GU等: ""A Novel MKL Model of Integrating LiDAR Data and MSI for Urban Area Classification"", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》 * |
冯凯: ""基于多核学习的多/高光谱图像与激光雷达数据联合分类研究"", 《中国优秀硕士学位论文全文数据库 电子期刊》 * |
时彧: "《机械故障诊断技术与应用》", 30 April 2014 * |
Cited By (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108846352B (en) * | 2018-06-08 | 2020-07-14 | 广东电网有限责任公司 | Vegetation classification and identification method |
CN108846352A (en) * | 2018-06-08 | 2018-11-20 | 广东电网有限责任公司 | A kind of vegetation classification and recognition methods |
CN109101977A (en) * | 2018-06-27 | 2018-12-28 | 深圳大学 | A kind of method and device of the data processing based on unmanned plane |
WO2020000271A1 (en) * | 2018-06-27 | 2020-01-02 | 深圳大学 | Unmanned aerial vehicle-based data processing method and device |
CN110163101A (en) * | 2019-04-17 | 2019-08-23 | 湖南省中医药研究院 | The difference of Chinese medicine seed and grade quick discrimination method |
CN110163101B (en) * | 2019-04-17 | 2022-09-23 | 湖南省中医药研究院 | Method for rapidly distinguishing seeds of traditional Chinese medicinal materials and rapidly judging grades of seeds |
CN111487643B (en) * | 2020-04-13 | 2021-06-08 | 中国科学院空天信息创新研究院 | Building detection method based on laser radar point cloud and near-infrared image |
CN111487643A (en) * | 2020-04-13 | 2020-08-04 | 中国科学院空天信息创新研究院 | Building detection method based on laser radar point cloud and near-infrared image |
CN111738298A (en) * | 2020-05-27 | 2020-10-02 | 哈尔滨工业大学 | Data classification method based on depth-width-variable multi-core learning |
CN111738298B (en) * | 2020-05-27 | 2023-09-12 | 哈尔滨工业大学 | MNIST handwriting digital data classification method based on deep-wide variable multi-core learning |
CN112130169B (en) * | 2020-09-23 | 2022-09-16 | 广东工业大学 | Point cloud level fusion method for laser radar data and hyperspectral image |
CN112130169A (en) * | 2020-09-23 | 2020-12-25 | 广东工业大学 | Point cloud level fusion method for laser radar data and hyperspectral image |
CN113589257A (en) * | 2021-07-28 | 2021-11-02 | 哈尔滨工业大学 | Global feature-based classification method for full-waveform laser echo signals |
CN116933141A (en) * | 2023-09-15 | 2023-10-24 | 昆明理工大学 | Multispectral laser radar point cloud classification method based on multicore graph learning |
CN116933141B (en) * | 2023-09-15 | 2023-11-17 | 昆明理工大学 | Multispectral laser radar point cloud classification method based on multicore graph learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106529484A (en) | Combined spectrum and laser radar data classification method based on class-fixed multinucleated learning | |
Li et al. | Adaptive multiscale deep fusion residual network for remote sensing image classification | |
CN105740799B (en) | Classification of hyperspectral remote sensing image method and system based on the selection of three-dimensional Gabor characteristic | |
CN103971123B (en) | Hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL) | |
CN107292343A (en) | A kind of Classification of hyperspectral remote sensing image method based on six layers of convolutional neural networks and spectral space information consolidation | |
CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
CN103440505B (en) | The Classification of hyperspectral remote sensing image method of space neighborhood information weighting | |
CN106023065A (en) | Tensor hyperspectral image spectrum-space dimensionality reduction method based on deep convolutional neural network | |
CN104732244B (en) | The Classifying Method in Remote Sensing Image integrated based on wavelet transformation, how tactful PSO and SVM | |
CN104268593A (en) | Multiple-sparse-representation face recognition method for solving small sample size problem | |
CN108090447A (en) | Hyperspectral image classification method and device under double branch's deep structures | |
CN104517284A (en) | Polarimetric SAR (synthetic aperture radar) image segmentation based on DBN (deep belief network) | |
CN108460391A (en) | Based on the unsupervised feature extracting method of high spectrum image for generating confrontation network | |
CN110020693B (en) | Polarimetric SAR image classification method based on feature attention and feature improvement network | |
CN103413151A (en) | Hyperspectral image classification method based on image regular low-rank expression dimensionality reduction | |
CN110533077A (en) | Form adaptive convolution deep neural network method for classification hyperspectral imagery | |
CN102902956A (en) | Ground-based visible cloud image recognition processing method | |
CN107563442A (en) | Hyperspectral image classification method based on sparse low-rank regular graph qualified insertion | |
CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
CN105760900A (en) | Hyperspectral image classification method based on affinity propagation clustering and sparse multiple kernel learning | |
CN104239902A (en) | Hyper-spectral image classification method based on non-local similarity and sparse coding | |
CN107194423A (en) | The hyperspectral image classification method of the integrated learning machine that transfinites of feature based random sampling | |
CN110060273A (en) | Remote sensing image landslide plotting method based on deep neural network | |
CN108197650A (en) | The high spectrum image extreme learning machine clustering method that local similarity is kept | |
CN103049767A (en) | Aurora image classification method based on biological stimulation characteristic and manifold learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170322 |