CN105956612B - Hyperspectral image classification method based on Active Learning and semi-supervised learning - Google Patents
Hyperspectral image classification method based on Active Learning and semi-supervised learning Download PDFInfo
- Publication number
- CN105956612B CN105956612B CN201610261010.XA CN201610261010A CN105956612B CN 105956612 B CN105956612 B CN 105956612B CN 201610261010 A CN201610261010 A CN 201610261010A CN 105956612 B CN105956612 B CN 105956612B
- Authority
- CN
- China
- Prior art keywords
- sample
- region
- dendrogram
- mclu
- class
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
The invention discloses a kind of hyperspectral image classification method based on Active Learning and semi-supervised learning, the low problem of the amounts of specimen information for overcoming and marking sample number deficiency under the prior art, choose.The step of present invention implements are as follows: (1) input data;(2) dendrogram is obtained;(3) setting maximum number of iterations is 22;(4) classify;(5) region division;(6) area sample marks;(7) sample set is updated;(8) judge whether maximum number of iterations is 0, if so, thening follow the steps (9), otherwise, execute step (4) after subtracting 1 for maximum number of iterations;(9) classified using updated training sample set training SVM SVM to test sample collection, export classification hyperspectral imagery figure.Present invention utilizes region division strategies to choose the high sample of information content, meanwhile, the quantity of marker samples is decreased, nicety of grading is promoted and rapidly improves.
Description
Technical field
The invention belongs to image classification process fields, further relate to one of field of target recognition while utilizing master
The hyperspectral image classification method of dynamic study and semi-supervised learning.This method can be used for map making, ocean remote sensing, vegetation tune
It looks into, atmospheric research, agricultural remote sensing, the fields such as environmental monitoring carry out ground object target identification.
Background technique
In the numerous areas of high spectrum image research, the classification of high spectrum image is played an extremely important role.It is high
Spectrum picture processing and classification have following two: 1) spatial diversity of the every class atural object of high spectrum image is big;2)Hughes
(under the premise of keeping precision constant, with the increase of wave band number, number of training needed for classifier also exists for the presence of phenomenon
Increase).In many classifiers, SVM (Support Vector Machine, SVM) classifier is shown one's talent,
And it is widely used in data mining, pattern-recognition and computer vision field, it is in processing small sample, higher-dimension
There is very big advantage when data, seldom sample can be used and obtain higher nicety of grading.
As a kind of Typical Representative of supervised learning, its main thought is Active Learning: utilizing the training sample of initialization
Collection goes to train classifier, then iteratively chooses the sample to contain much information and is added to training sample concentration, utilizes updated instruction
Practice sample training classifier, until meeting some termination condition, wherein the sample that per generation is chosen all assigns its true atural object
Label.
Li Jun et al. is in paper " the Semisupervised Self-Learning for Hyperspectral delivered
Image Classification”(IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING,
VOL.51, NO.7, JULY 2013) in propose a kind of hyperspectral image classification method based on semi-supervised autonomous learning.The party
Method initializes some marker samples first, its four neighborhoods sample is then looked for centered on exemplar, if point of neighborhood sample
The label of class is identical as central sample label, then adds it to training sample, utilizes updated training sample re -training
Classifier, until reaching termination condition.Deficiency existing for this method is that the information content of the sample of selection is low.
Patent " hyperspectral image classification method that based on semi-supervised dimension is about subtracted " of the Xian Electronics Science and Technology University in application
It proposes and a kind of is about subtracted in (number of patent application: 201110199094.6, publication number: CN102208034A) based on semi-supervised dimension
Hyperspectral image classification method.This method constructs the dissimilar matrix in part first, then construction feature equation, obtains projection square
Battle array, then marked sample and unmarked sample are projected into lower dimensional space, new training set and test set is obtained, support is reused
Vector machine SVM classifier is classified.But deficiency existing for this method is, the sample diversity of selection is not high.
Summary of the invention
It is an object of the invention to be directed to the deficiency of above-mentioned prior art, propose a kind of based on Active Learning and semi-supervised
The hyperspectral image classification method of study allows the two to have complementary advantages in conjunction with Active Learning and semi-supervised learning method, in label sample
Very high nicety of grading can be also obtained in the case that this is seldom.
The present invention realizes that the concrete thought of above-mentioned purpose is: inputting high spectrum image to be sorted and EO-1 hyperion to be sorted
Image data set;Cluster operation is carried out to high spectrum image to be sorted, obtains dendrogram;Using training sample training support to
Amount machine SVM, obtains classification chart;All areas are divided into trusted area and insincere region by using area division methods;Every
A sample is chosen in a region, marks selected sample using area sample labeling method;Update sample set;Judge termination condition
Whether reach;Using updated sample set training SVM SVM, final classification figure is obtained.
To achieve the above object, the present invention includes the following steps:
(1) input data:
(1a) inputs a high spectrum image to be sorted;
(1b) inputs hyperspectral image data collection corresponding with high spectrum image to be sorted;
(1c) chooses and hyperspectral image data collection atural object classification point from the sample of the hyperspectral image data collection of input
Not corresponding training sample set;
Hyperspectral image data is concentrated remaining sample as test sample collection by (1d);
(2) dendrogram is obtained:
Average drifting method is analyzed using iterative data, cluster operation is carried out to high spectrum image to be sorted, is obtained wait divide
The dendrogram of the high spectrum image of class;
(3) setting maximum number of iterations is 22;
(4) classify:
Classified using training sample set training SVM SVM to test sample collection, obtains EO-1 hyperion to be sorted
The classification chart of image;
(5) region division:
All areas in dendrogram are divided into region to be divided and insincere region using global partitioning by (5a);
Region division to be divided is trusted area and insincere region using local partitioning by (5b);
(5c) closes in the insincere region in the insincere region of dendrogram and region to be divided in the form of gathering and seek union
And insincere regional ensemble is obtained, using the trusted area in region to be divided as trusted area set;
(6) area sample marks:
(6a) chooses the smallest sample of multi-class uncertainty MCLU value from each region of trusted area set, will divide
Class label of the tag along sort corresponding with the sample as the sample in class figure.
(6b) never chooses the smallest sample of multi-class uncertainty MCLU value in each region of trusted area set, will
The corresponding hyperspectral image data of high spectrum image to be sorted concentrates true class label corresponding with the sample as the sample
Class label.
(7) sample set is updated;
(7a) is by training sample set and the selected the smallest sample of multi-class uncertainty MCLU value from each region
Merge in the form of set seeks union, obtains updated training sample set;
(7b) subtracts the smallest with multi-class uncertainty MCLU value selected from each region from test sample concentration
The identical sample of sample, obtains updated test sample collection;
(8) judge whether maximum number of iterations is 0, if so, thening follow the steps (9), otherwise, maximum number of iterations is subtracted 1
Execute step (4) afterwards;
(9) classified using updated training sample set training SVM SVM to test sample collection, output is high
Spectrum picture classification chart.
The present invention has the advantage that compared with prior art
1st, since the present invention utilizes region partitioning method, so that selected sample has diversity, overcome existing
The low problem of the diversity of selected training sample, allows the present invention to make full use of the diversity of sample under technology, into
And obtain preferable classifying quality.
2nd, due to overcoming present invention utilizes area sample labeling method, marker samples under the prior art are few, mark
Sample problem at high cost, so that the present invention can also obtain higher classification results in the case where training sample is less.
Detailed description of the invention
Fig. 1 is flow chart of the invention;
Fig. 2 is the flow chart of the region partitioning method in the present invention;
Fig. 3 is the classifying quality figure of the present invention and the prior art;
Fig. 4 is the nicety of grading curve graph of the present invention with the prior art.
Specific embodiment
Present invention is described with reference to the accompanying drawing.
Referring to Fig.1, the specific steps of the present invention are as follows:
Step 1, input data.
Input a high spectrum image to be sorted.
Input hyperspectral image data collection corresponding with high spectrum image to be sorted.
From the sample of the hyperspectral image data collection of input, it is right respectively with hyperspectral image data collection atural object classification to choose
The training sample set answered.
Concentrate remaining sample as test sample collection hyperspectral image data.
In embodiments of the present invention, input be 145 × 145 pixels high spectrum image to be sorted, input with to point
The corresponding hyperspectral image data collection to be sorted of the high spectrum image of class includes 20125 samples, and each sample is 200 dimensions, will
160 samples that hyperspectral image data to be sorted is concentrated are concentrated surplus as training sample set, hyperspectral image data to be sorted
Remaining conduct test sample collection.
Step 2, dendrogram is obtained.
Average drifting method is analyzed using iterative data, cluster operation is carried out to high spectrum image to be sorted, is obtained wait divide
The dendrogram of the high spectrum image of class.
In embodiments of the present invention, cluster operation is carried out to image using the data analysis average drifting method of iteration.The party
Method sets region of search radius first, by calculating the mean shift amount of each region, cluster centre vector is updated, to obtain
Reasonable cluster result.The sample that cluster operation uses all is the vector of 202 dimensions, wherein 200 dimensions are spectral Dimensions information, 2 dimensions
It is Spatial Dimension information, the radius of region of search space segment is set as 6, and the radius of spectra part is set as 2000.
According to the following formula, the mean-shift vector of each region is calculated:
M=M-N
Wherein, m indicates the mean-shift vector of each region, and M indicates the sample mean vector that each region is included, N
Indicate current cluster centre vector;
The rule for updating cluster centre is as follows:
If m > ε, sample that M is assigned to N as the cluster centre vector of next iteration, in more new search area radius
Mean vector M all samples in the region of search radius centered on each cluster centre vector are otherwise divided into phase
In the cluster areas answered, final cluster result is obtained.Wherein: m indicates the mean-shift vector of each region, and ε indicates default
Threshold value, M indicate the sample mean vector that each region is included.
Step 3, setting maximum number of iterations is 22.
Step 4, classify.
Classified using training sample set training SVM SVM to test sample collection, obtains classification chart.
In an embodiment of the present invention, classification task is executed using the tool box LIBSVM, using the support vector for having supervision
Machine SVM, classifies to test sample collection, obtains classification results.
Step 5, region division.
Referring to Fig. 2, the concrete operations of this step are as follows:
Firstly, all areas in dendrogram are divided into region to be divided and insincere region using global partitioning.
The global partitioning, the specific steps are as follows:
Step 1, count in classification hyperspectral imagery figure to be sorted of all categories is formed by each region in dendrogram
Total sample number;
Step 2, according to the following formula, the sample number that the maximum classification of each region is included in calculating dendrogram is in the region sample
Shared ratio in this sum:
Wherein, the sample number that the maximum classification of each region is included in p expression dendrogram is in area sample sum
Shared ratio, nmaxIndicate that the other sample number of maximum kind in each region, C indicate the classification sum of high spectrum image to be sorted,
∑ indicates overlap-add operation, niIndicate the sum of the i-th class sample in each region in dendrogram;
Step 3 calculates multi-class uncertainty corresponding to all samples of each region in dendrogram according to the following formula
The mean value of MCLU value:
Wherein, μ indicates the mean value of multi-class uncertainty MCLU corresponding to all samples of each region in dendrogram,
N indicates that the sum of sample in each region in dendrogram, ∑ indicate overlap-add operation, diIt indicates in dendrogram i-th in each region
The multi-class uncertainty MCLU value of a sample;
Step 4 calculates multi-class uncertainty corresponding to all samples of each region in dendrogram according to the following formula
The standard deviation of MCLU value:
Wherein, σ indicates the standard of multi-class uncertainty MCLU corresponding to all samples of each region in dendrogram
Difference,Indicate extraction of square root operation, n indicates that the sum of sample in each region in dendrogram, ∑ indicate overlap-add operation, ()2
Indicate squared operation, diIndicate the MCLU value of i-th of sample in each region in dendrogram, μ indicates each area in dendrogram
The mean value of multi-class uncertainty MCLU corresponding to all samples in domain, n indicate in dendrogram the total of sample in each region
Number;
Step 5 calculates the global of each region in dendrogram and divides evaluation of estimate according to the following formula:
Wherein, t indicates that the global of each region divides evaluation of estimate in dendrogram, and p indicates that each region is most in dendrogram
The sample number that big classification is included ratio shared in area sample sum, * indicate multiplication operations, and μ is indicated in dendrogram
The mean value of multi-class uncertainty MCLU corresponding to all samples of each region, σ indicate the institute of each region in dendrogram
There is the standard deviation of multi-class uncertainty MCLU corresponding to sample;
Step 6, by the global region for dividing evaluation of estimate and being more than or equal to 1 of each region in dendrogram, as area to be divided
The global of each region in dendrogram is divided region of the evaluation of estimate less than 1, as insincere region by domain.
Secondly, being trusted area and insincere region by region division to be divided using local partitioning.
The local partitioning, the specific steps are as follows:
Step 1 successively selects a multi-class uncertainty MCLU value most in each of dendrogram region to be divided
Small sample;
Step 2 is chosen and class corresponding to the selected the smallest sample of multi-class uncertainty MCLU value from classification chart
Distinguishing label, using such distinguishing label as the tag along sort of the sample;
Step 3 calculates classification chart and corresponds to the sample that each classification in each of dendrogram region to be divided is included
Sum obtains the most class label of total sample number;
Step 4, using include in region to be divided each in dendrogram the most classification of total sample number class label as
The cluster labels of selected sample from the region to be divided;
Step 5 compares the smallest sample of multi-class uncertainty MCLU value selected by each region to be divided in dendrogram
This cluster labels and tag along sort when label is identical, using corresponding region as trusted area, are otherwise used as insincere region.
Finally, the insincere region in the insincere region of dendrogram and region to be divided is closed in the form of gathering and seeking union
And insincere regional ensemble is obtained, using the trusted area in region to be divided as trusted area set.
The multi-class uncertainty MCLU value mentioned in this step is acquired by multi-class uncertainty MCLU method.
According to the following formula, the multi-class uncertainty MCLU value of sample is calculated:
Wherein, r1Indicate serial number apart from maximum value of the sample relative to classifying face, x is sample to be calculated, fi(x) table
Show that sample x is at a distance from i-th of classifying face in SVM SVM, k indicates the classification number that sample set includes, r2Indicate sample
Originally the serial number relative to the second largest value of the distance of classifying face, d (x) indicate the multi-class uncertainty MCLU value of sample.
Step 6, area sample marks.
The smallest sample of multi-class uncertainty MCLU value is chosen from each region of trusted area set, by classification chart
In class label of the tag along sort corresponding with the sample as the sample.
The smallest sample of multi-class uncertainty MCLU value is never chosen in each region of trusted area set, it will be wait divide
The corresponding hyperspectral image data of class high spectrum image concentrates class of the true class label corresponding with the sample as the sample
Distinguishing label.
Step 7, sample set is updated.
By training sample set and the selected the smallest sample of multi-class uncertainty MCLU value from each region to collect
Conjunction asks the form of union to merge, and obtains updated training sample set.
It is subtracted and the smallest sample of selected multi-class uncertainty MCLU value from each region from test sample concentration
Identical sample obtains updated test sample collection.
Step 8, judge whether maximum number of iterations is 0, if so, (9) are thened follow the steps, otherwise, by maximum number of iterations
Step (4) are executed after subtracting 1.
Step 9, classified using updated training sample set training SVM SVM to test sample collection, it is defeated
Classification hyperspectral imagery figure out.
Effect of the invention is described further below by emulation experiment.
1, emulation experiment condition:
The condition of emulation experiment of the present invention is are as follows: CPU:Intel Core I5 2430M, basic frequency 2.4GHz, memory
2G, MATLAB MatlabR2011b.
The high spectrum image Indiana Pines that the present invention uses, it is by the airborne visible/red of National Space Agency
Imaging of the outer imaging spectrometer (AVIRIS) to the Indian remote sensing test block in the U.S., the northwestward, the state Indiana.Indiana Pine
Comprising atural object in 16 including trees, meadow and crops etc., totally 10366 pixels, 375-2200 μm of spectral region, space
Resolution ratio is 20m.The image is 145 × 145 × 220 data cube, 220 wave bands is shared, due to water mist and atmosphere etc.
The influence of noise eliminates 104-108,150-163,220 wave bands, and emulation experiment uses remaining 200 wave bands.
2, emulation experiment content and interpretation of result:
Emulation experiment of the present invention is to select Gauss radial based on one-to-many SVM SVM (OAA) classifier
Base kernel function (RBF) is the kernel function of emulation experiment.Cross validation grid search parameter γ and C are rolled over by 5- in emulation experiment
Value, C is the normalized parameter of model, and γ is the nuclear parameter of gaussian kernel function, C=31.62 in the present invention, γ=0.32.Intersect
The method of verifying is to find the optimal value for training SVM SVM.The threshold value that the global criteria for classifying is arranged is
1.For the present invention, it is 160 that setting, which initializes marked number of training, the sample that per generation is chosen all be it is adaptive,
Maximum number of iterations 22 is set, and the algorithm finally proposed will converge to higher precision quickly.
In order to verify the validity of the proposed method of the present invention, by two kinds of Active Learnings of method and the prior art of the invention
Method compared.Both methods is the active based on MCLU-ECBD sampling that Beg ü m Demir et al. is proposed respectively
The LORSAL-AL Active Learning Method that learning method and Li Jun et al. propose, both the above method have used Active Learning
Thought, difference is to have used the different method of samplings to choose sample.In addition, also with the Active Learning Method of some classics into
It has gone comparison, has such as been based on the Active Learning of stochastical sampling (RS), the Active Learning based on classification uncertain (MCLU) sampling.
The prior art used in emulation experiment of the invention is all made of 160 initial marked training samples, according to each
The method of sampling of a prior art itself chooses 30 new marker samples in per generation and is added to marked training sample concentration, repeatedly
Generation number is 30.That is, the sample number that human expert marks during Active Learning is 1060, remaining unmarked sample
As test sample.
The nicety of grading comparison of the present invention and the prior art when marked sample number is 610 is as shown in table 1.In table 1
RS indicates that the Active Learning Method sampled based on RS, MCLU indicate the Active Learning Method sampled based on MCLU, LORSAL-AL table
Show that LORSAL-AL Active Learning Method, MCLU-ECBD indicate the Active Learning Method sampled based on MCLU-ECBD.
Table 1
Algorithm | Nicety of grading (%) |
RS | 80.65 |
MCLU | 84.68 |
LORSAL-AL | 85.06 |
MCLU-ECBD | 86.03 |
The present invention | 89.92 |
As it can be seen from table 1 nicety of grading of the present invention is apparently higher than other prior arts.
The classifying quality figure of the present invention and the prior art is as shown in Figure 3, wherein Fig. 3 (a) be high spectrum image practically
Species are not schemed, and Fig. 3 (b) is the classifying quality figure of the Active Learning Method based on RS sampling, and Fig. 3 (c) is sampled based on MCLU
The classifying quality figure of Active Learning Method, Fig. 3 (d) are the classifying quality figures of LORSAL-AL Active Learning Method, and Fig. 3 (e) is base
In the classifying quality figure of the Active Learning Method of MCLU-ECBD sampling, Fig. 3 (f) is classifying quality figure of the invention.
Table 2
Algorithm | Marked sample number |
RS | —— |
MCLU | 1060 |
LORSAL-AL | 1030 |
MCLU-ECBD | 1030 |
The present invention | 620 |
The present invention and the prior art marked sample number required when nicety of grading reaches 90% or more are compared such as 2 institute of table
Show.RS in table 2 indicates that the Active Learning Method sampled based on RS, MCLU indicate the Active Learning Method sampled based on MCLU,
LORSAL-AL indicates that LORSAL-AL Active Learning Method, MCLU-ECBD indicate the Active Learning side sampled based on MCLU-ECBD
Method.
From table 2 it can be seen that the present invention only needs 620 marked samples just nicety of grading to may make to reach 90% or more,
And it Active Learning Method that the prior art is sampled based on MCLU, LORSAL-AL Active Learning Method and is sampled based on MCLU-ECBD
Active Learning Method at least need 1000 or more marked sample, the marker samples that the present invention has clearly a need for are less, and show
Not up to 90% classification essence when the marked sample of the Active Learning Method maximum as defined in us for thering is technology to sample based on RS
Therefore degree uses --- it indicates.
Fig. 4 describes the nicety of grading curve graph of the present invention with the prior art.Wherein, abscissa indicates marked sample
Quantity, ordinate indicate overall classification accuracy.Indicate bent using nicety of grading of the invention in figure with the curve that diamond shape indicates
Line chart is indicated with the curve that triangle indicates using the nicety of grading curve based on the RS Active Learning Method sampled, with star
The curve of mark indicates to use the nicety of grading curve based on the MCLU Active Learning Method sampled, with the curve table of cross mark
Show the nicety of grading curve using LORSAL-AL Active Learning Method, is indicated with the curve that box indicates using based on MCLU-
The nicety of grading curve of the Active Learning Method of ECBD sampling.
Figure 4, it is seen that the present invention it is each instead of between abscissa be spaced in and be gradually reduced, illustrate that every generation is chosen
Marker samples number gradually decreasing, but precision is continuing to increase, and illustrates effectiveness of the invention, promotes the receipts of algorithm
It holds back.
Claims (3)
1. a kind of hyperspectral image classification method based on Active Learning and semi-supervised learning, comprising the following steps:
(1) input data:
(1a) inputs a high spectrum image to be sorted;
(1b) inputs hyperspectral image data collection corresponding with high spectrum image to be sorted;
(1c) from the sample of the hyperspectral image data collection of input, it is right respectively with hyperspectral image data collection atural object classification to choose
The training sample set answered;
Hyperspectral image data is concentrated remaining sample as test sample collection by (1d);
(2) dendrogram is obtained:
Average drifting method is analyzed using iterative data, cluster operation is carried out to high spectrum image to be sorted, is obtained to be sorted
The dendrogram of high spectrum image;
(3) setting maximum number of iterations is 22;
(4) classify:
Classified using training sample set training SVM SVM to test sample collection, obtains high spectrum image to be sorted
Classification chart;
(5) region division:
All areas in dendrogram are divided into region to be divided and insincere region using global partitioning by (5a);
Region division to be divided is trusted area and insincere region using local partitioning by (5b);
(5c) merges in the insincere region in the insincere region of dendrogram and region to be divided in the form of gathering and seek union, obtains
To insincere regional ensemble, using the trusted area in region to be divided as trusted area set;
(6) area sample marks:
(6a) chooses the smallest sample of multi-class uncertainty MCLU value from each region of trusted area set, by classification chart
In class label of the tag along sort corresponding with the sample as the sample;
(6b) never chooses the smallest sample of multi-class uncertainty MCLU value in each region of trusted area set, will to point
The corresponding hyperspectral image data of the high spectrum image of class concentrates true class label corresponding with the sample as the sample
Class label;
(7) sample set is updated;
(7a) is by training sample set and the selected the smallest sample of multi-class uncertainty MCLU value from each region to collect
Conjunction asks the form of union to merge, and obtains updated training sample set;
(7b) is subtracted and the smallest sample of selected multi-class uncertainty MCLU value from each region from test sample concentration
Identical sample obtains updated test sample collection;
(8) judge whether maximum number of iterations is 0, if so, thening follow the steps (9), otherwise, held after subtracting 1 for maximum number of iterations
Row step (4);
(9) classified using updated training sample set training SVM SVM to test sample collection, export EO-1 hyperion
Image classification figure.
2. the hyperspectral image classification method according to claim 1 based on Active Learning and semi-supervised learning, special
Sign is: specific step is as follows for global partitioning described in step (5a):
Step 1 counts the sample of all categories being formed by each region in dendrogram in classification hyperspectral imagery figure to be sorted
This sum;
Step 2, according to the following formula, the sample number that the maximum classification of each region is included in calculating dendrogram are total in the area sample
Shared ratio in number:
Wherein, p indicates that the sample number that the maximum classification of each region in dendrogram is included is shared in area sample sum
Ratio, nmaxIndicate that the other sample number of maximum kind in each region, C indicate the classification sum of high spectrum image to be sorted, ∑ table
Show overlap-add operation, niIndicate the sum of the i-th class sample in each region in dendrogram;
Step 3 calculates multi-class uncertainty MCLU value corresponding to all samples of each region in dendrogram according to the following formula
Mean value:
Wherein, μ indicates the mean value of multi-class uncertainty MCLU corresponding to all samples of each region in dendrogram, n table
Show that the sum of sample in each region in dendrogram, ∑ indicate overlap-add operation, diI-th of sample in each region in expression dendrogram
This multi-class uncertainty MCLU value;
Step 4 calculates multi-class uncertainty MCLU value corresponding to all samples of each region in dendrogram according to the following formula
Standard deviation:
Wherein, σ indicates the standard deviation of multi-class uncertainty MCLU corresponding to all samples of each region in dendrogram,Indicate extraction of square root operation, n indicates that the sum of sample in each region in dendrogram, ∑ indicate overlap-add operation, ()2Table
Show squared operation, diIndicate the MCLU value of i-th of sample in each region in dendrogram, μ indicates each region in dendrogram
All samples corresponding to multi-class uncertainty MCLU mean value, n indicates the sum of sample in each region in dendrogram;
Step 5 calculates the global of each region in dendrogram and divides evaluation of estimate according to the following formula:
Wherein, t indicates that the global of each region divides evaluation of estimate in dendrogram, and p indicates the maximum kind of each region in dendrogram
Other included sample number ratio shared in area sample sum, * indicate multiplication operations, and μ indicates each in dendrogram
The mean value of multi-class uncertainty MCLU corresponding to all samples in region, σ indicate all samples of each region in dendrogram
The standard deviation of multi-class uncertainty MCLU corresponding to this;
Step 6, will as region to be divided by the global region for dividing evaluation of estimate and being more than or equal to 1 of each region in dendrogram
The global of each region divides region of the evaluation of estimate less than 1 in dendrogram, as insincere region.
3. the hyperspectral image classification method according to claim 1 based on Active Learning and semi-supervised learning, special
Sign is: specific step is as follows for local partitioning described in step (5b):
It is the smallest successively to select a multi-class uncertainty MCLU value in each of dendrogram region to be divided for step 1
Sample;
Step 2 is chosen and classification mark corresponding to the selected the smallest sample of multi-class uncertainty MCLU value from classification chart
Label, using such distinguishing label as the tag along sort of the sample;
Step 3 calculates classification chart and corresponds to the total sample number that each classification in each of dendrogram region to be divided is included,
Obtain the most class label of total sample number;
Step 4, using include in region to be divided each in dendrogram the most classification of total sample number class label as from this
The cluster labels of selected sample in region to be divided;
Step 5 compares the smallest sample of multi-class uncertainty MCLU value selected by each region to be divided in dendrogram
Cluster labels and tag along sort, when label is identical, using corresponding region as trusted area, otherwise, as insincere region.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610261010.XA CN105956612B (en) | 2016-04-25 | 2016-04-25 | Hyperspectral image classification method based on Active Learning and semi-supervised learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610261010.XA CN105956612B (en) | 2016-04-25 | 2016-04-25 | Hyperspectral image classification method based on Active Learning and semi-supervised learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105956612A CN105956612A (en) | 2016-09-21 |
CN105956612B true CN105956612B (en) | 2019-03-26 |
Family
ID=56915746
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610261010.XA Active CN105956612B (en) | 2016-04-25 | 2016-04-25 | Hyperspectral image classification method based on Active Learning and semi-supervised learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105956612B (en) |
Families Citing this family (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107977668A (en) * | 2017-07-28 | 2018-05-01 | 北京物灵智能科技有限公司 | A kind of robot graphics' recognition methods and system |
CN107451614B (en) * | 2017-08-01 | 2019-12-24 | 西安电子科技大学 | Hyperspectral classification method based on fusion of space coordinates and space spectrum features |
CN108108769B (en) * | 2017-12-29 | 2020-08-25 | 咪咕文化科技有限公司 | Data classification method and device and storage medium |
CN108197666A (en) * | 2018-01-30 | 2018-06-22 | 咪咕文化科技有限公司 | A kind of processing method, device and the storage medium of image classification model |
CN108985360B (en) * | 2018-06-29 | 2022-04-08 | 西安电子科技大学 | Hyperspectral classification method based on extended morphology and active learning |
CN109827908B (en) * | 2018-12-24 | 2021-04-13 | 核工业北京地质研究院 | Method for judging rot degree of Fuji apples by using spectral data |
CN110222793B (en) * | 2019-06-20 | 2021-06-22 | 中国科学院自动化研究所 | Online semi-supervised classification method and system based on multi-view active learning |
CN110309868A (en) * | 2019-06-24 | 2019-10-08 | 西北工业大学 | In conjunction with the hyperspectral image classification method of unsupervised learning |
CN110490061B (en) * | 2019-07-11 | 2021-10-22 | 武汉大学 | Uncertainty modeling and measuring method for remote sensing image characteristics |
CN110501290B (en) * | 2019-08-16 | 2021-09-24 | 安徽优思天成智能科技有限公司 | Ship exhaust gas spectral image segmentation and pollution prediction method |
CN111860508A (en) | 2020-07-28 | 2020-10-30 | 平安科技(深圳)有限公司 | Image sample selection method and related equipment |
CN114841214B (en) * | 2022-05-18 | 2023-06-02 | 杭州电子科技大学 | Pulse data classification method and device based on semi-supervised discrimination projection |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103914705A (en) * | 2014-03-20 | 2014-07-09 | 西安电子科技大学 | Hyperspectral image classification and wave band selection method based on multi-target immune cloning |
CN103996047A (en) * | 2014-03-04 | 2014-08-20 | 西安电子科技大学 | Hyperspectral image classification method based on compression spectrum clustering integration |
CN104484681A (en) * | 2014-10-24 | 2015-04-01 | 西安电子科技大学 | Hyperspectral remote sensing image classification method based on space information and ensemble learning |
CN104680185A (en) * | 2015-03-15 | 2015-06-03 | 西安电子科技大学 | Hyperspectral image classification method based on boundary point reclassification |
-
2016
- 2016-04-25 CN CN201610261010.XA patent/CN105956612B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103996047A (en) * | 2014-03-04 | 2014-08-20 | 西安电子科技大学 | Hyperspectral image classification method based on compression spectrum clustering integration |
CN103914705A (en) * | 2014-03-20 | 2014-07-09 | 西安电子科技大学 | Hyperspectral image classification and wave band selection method based on multi-target immune cloning |
CN104484681A (en) * | 2014-10-24 | 2015-04-01 | 西安电子科技大学 | Hyperspectral remote sensing image classification method based on space information and ensemble learning |
CN104680185A (en) * | 2015-03-15 | 2015-06-03 | 西安电子科技大学 | Hyperspectral image classification method based on boundary point reclassification |
Non-Patent Citations (1)
Title |
---|
Collaborative active and semisupervised learning for hyperspectral remote sensing image classification;Wan L,et al.;《Geoscience and Remote Sensing,IEEE Transactions on,2015》;20150531;第53卷(第5期);Pages 2384-2396 |
Also Published As
Publication number | Publication date |
---|---|
CN105956612A (en) | 2016-09-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105956612B (en) | Hyperspectral image classification method based on Active Learning and semi-supervised learning | |
Polewski et al. | Detection of fallen trees in ALS point clouds using a Normalized Cut approach trained by simulation | |
CN101551809B (en) | Search method of SAR images classified based on Gauss hybrid model | |
CN106339674B (en) | The Hyperspectral Image Classification method that model is cut with figure is kept based on edge | |
CN106503739A (en) | The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics | |
CN108830312B (en) | Integrated learning method based on sample adaptive expansion | |
CN110443139B (en) | Hyperspectral remote sensing image noise band detection method for classification | |
CN103208011A (en) | Hyperspectral image space-spectral domain classification method based on mean value drifting and group sparse coding | |
Chen et al. | Locating crop plant centers from UAV-based RGB imagery | |
Polewski et al. | Active learning approach to detecting standing dead trees from ALS point clouds combined with aerial infrared imagery | |
CN104182767A (en) | Active learning and neighborhood information combined hyperspectral image classification method | |
Hao et al. | An object-based change detection approach using uncertainty analysis for VHR images | |
Hakula et al. | Individual tree segmentation and species classification using high-density close-range multispectral laser scanning data | |
Azadbakht et al. | Improved urban scene classification using full-waveform LiDAR | |
Li et al. | A new combination classification of pixel-and object-based methods | |
CN105069471A (en) | Hyperspectral data subspace projection and classification method based on fuzzy label | |
Alburshaid et al. | Palm trees detection using the integration between gis and deep learning | |
CN110175638B (en) | Raise dust source monitoring method | |
Khanday et al. | Change detection in hyper spectral images | |
Zhang et al. | Classification of imbalanced hyperspectral imagery data using support vector sampling | |
RoyChowdhury et al. | Distinguishing weather phenomena from bird migration patterns in radar imagery | |
Yang et al. | Multiscale integration approach for land cover classification based on minimal entropy of posterior probability | |
Chen et al. | Context-aware lane marking detection on urban roads | |
Zhang et al. | A rough set decision tree based MLP-CNN for very high resolution remotely sensed image classification | |
Kutluk et al. | Classification of hyperspectral images using mixture of probabilistic PCA models |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |