CN109726641A - A kind of remote sensing image cyclic sort method based on training sample Automatic Optimal - Google Patents
A kind of remote sensing image cyclic sort method based on training sample Automatic Optimal Download PDFInfo
- Publication number
- CN109726641A CN109726641A CN201811503807.1A CN201811503807A CN109726641A CN 109726641 A CN109726641 A CN 109726641A CN 201811503807 A CN201811503807 A CN 201811503807A CN 109726641 A CN109726641 A CN 109726641A
- Authority
- CN
- China
- Prior art keywords
- training sample
- classification
- remote sensing
- sensing image
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of cyclic sort methods based on training sample Automatic Optimal.This method makes full use of precision information provided by previous subseries result constantly to adjust and optimizes the training sample for next subseries, is then classified automatically to remote sensing images again using the classification based training sample after optimization;It is thusly-formed a cyclic sort process, nicety of grading twice tends towards stability until front and back, stops classifying.This method has the characteristics that Automatic Optimal training sample, isodose training sample maximally utilize, compared to same sorting algorithm precision height, strong applicability.Pass through verifying, most supervised classification methods of this method suitable for Classification in Remote Sensing Image application;Especially when containing many noises in training sample, the case where automatically optimizing training sample to improve nicety of grading is needed.In practical applications, this method, which is mainly used in, is related to the Classification in Remote Sensing Image of large-scale land cover pattern/utilization or special topic (for example, crops).
Description
Technical field
The invention belongs to remote sensing machine learning classification application fields, in particular to using remote sensing image carry out land cover pattern/
Utilize the engineer application of, Crop classification.
Background technique
Supervised classification method is one of common method in the application of the Classification in Remote Sensing Image such as land cover pattern/utilization, crops.Supervision
The conventional method of classification method is that suitable training sample is first selected on image, then selects suitable Machine learning classifiers
Classify to image, finally carries out precision evaluation.Research shows that: the precision and training sample quality of Classification in Remote Sensing Image achievement are divided
The factors such as class method, post-classification comparison Method means are highly relevant.However, after being continuously improved and classify with the performance of classifier
Processing means diversification, at present improve nicety of grading Main way be quality [Zhu Xiufang, Pan for how improving training sample
Credit is loyal, Zhang Jinshui, Wang Shuan, and Gu Xiaohe, Xu Chao (2007) training sample influences TM scale wheat planting area measurement precision
Study (I) --- nicety of grading response relation studies remote sensing journal, 826-837 between training sample and classification method].However,
Even the classification samples chosen meticulously during actual classification, there is also a large amount of noise and errors, severely impact
The precision of classification.Similar research for example, Zhu Xiufang respectively from the angle analysis of quality and quantity training sample to TM scale
Winter wheat measurement accuracy influence, and point out different responses of the different classifications device to the training sample under different quality and quantity
[Zhu Xiufang, Pan Yaozhong, Wang Shuan, Han Lijian, Xu Chao (2009) training sample is to TM scale wheat planting area measurement precision shadow
Ring research (II) --- impact analysis Surveying and mapping of the sample quality to wheat measurement accuracy, 132-135].Currently, from training
The angle of sample improves there are mainly two types of the methods of Classification in Remote Sensing Image precision:
(1) select the pixel inside plot as training sample
The hypotheses of this method are that the pixel spectrum inside plot is opposite " pure ", and noise is less.Basic way
It is to select the pixel of plot immediate vicinity as training sample when selection training sample for the first time.However, this method is deposited
In many drawbacks, first is that human factor is higher, many ground class center still has a large amount of mixed pixel (noise);Second is that right
The classifier of mixed pixel is adapted in part, cannot accomplish Automatic Optimal.
(2) pixel analysis of spectrum threshold
The hypotheses of this method are that training sample is presented on spectral space and is just distributed very much, mainly every by calculating
Then the mean value and variance of the training sample of a type reject mean value and add the training sample outside three times variance.This method plays
Certain effect, the example of research is for example, Arai (1992) is removed in training sample using spectral information and spatial information
The biggish pixel of spectrum variation, makes the overall accuracy of Maximum likelihood classification improve 11.9% [Arai, K. (1992) .A
supervised thematic mapper classification with a purification of training
samples.International Journal of Remote Sensing,13,2039-2049].Wu Jian equality (1996)
Training sample is optimized using the method for spectrum threshold and space variance, nicety of grading improves 6% [Wu Jianping, Yang Xingwei
(1996) in remotely-sensed data supervised classification training sample purifying land resources remote sensing, 36-41].Wang Yi etc. (2007) is used
Spectrum searching algorithm, given threshold eliminate the impure part in training sample, and nicety of grading significantly improves [Wang Yi, Zhang Liang
Training, Purified Algorithm for Training Samples Wuhan University Journal (letter of Li Pingxiang (2007) based on automatic search and Spectral Matching Technique
Cease scientific version), 216-219].Kavzoglu (2009) successively using visualization and histogram method eliminate edge mixed pixel it
Afterwards, the nicety of grading of green tea is improved significantly.Nicety of grading can generally be improved 5%-10% or so by this method, be played
Certain effect, but there are certain drawbacks: the hypothesis being just distributed very much is not suitable for all regions, and some areas are same
A type of ground objects spectrum variation is larger, and analysis of spectrum threshold can remove the pixel that should belong to training sample in the form of noise;
The spectrum more homogeneous of the same type of ground objects in some areas, will lead in training sample using analysis of spectrum threshold still have it is very much
Noise [Kavzoglu, T. (2009) .Increasing the accuracy of neural network
classification using refined training data.Environmental Modelling&Software,
24,850-858]。
In conclusion the quality for improving classification based training sample is a kind of highly effective method of raising nicety of grading.At present
The method for improving training sample quality relies primarily on artificial means or single spectrum threshold index to be screened, not tangible
At the optimization method of automation.
Summary of the invention
The technical problem to be solved by the present invention is the optimization of training sample relies primarily on craft during conventional sorting methods
Adjustment, adjustment there are blindness, take time and effort the problems such as.This patent devises one kind for these problems and is based on training sample certainly
The cyclic sort method of dynamic optimization, this method make full use of precision information provided by previous subseries result constantly adjust with
Optimize the training sample for being used for next subseries, then remote sensing images are carried out certainly again using the classification based training sample after optimization
Dynamic classification;It is thusly-formed a cyclic sort process, nicety of grading twice tends towards stability until front and back, stops classifying.Invention
The purpose of this method is to improve the precision of the efficiency of training sample optimization, the degree of automation of Classification in Remote Sensing Image and classification results.
The technical solution adopted by the present invention to solve the technical problems is: a kind of remote sensing based on training sample Automatic Optimal
Image cyclic sort method comprising the steps of:
Step 1): selection classification of remote-sensing images training sample;
Step 2): the selection of classification of remote-sensing images device and classification;
Step 3): classification overall accuracy evaluation;
Step 4): the classification correct probability of each pixel in classification based training sample is calculated;
Step 5): setting probability threshold value optimizes training sample;
Step 6): automatic cycle 2-5 step, until the difference of the resulting classification overall accuracy of front and back step 3) calculating twice
Just stop recycling less than precision threshold.
A kind of each step of remote sensing image cyclic sort method based on training sample Automatic Optimal specifically:
Step 1): classification of remote-sensing images training sample is selected: to be sorted for each on remote sensing image to be sorted
Type selects training sample, and the quantity of the training sample of k-th of type is denoted asIt is training sample that wherein symbol T, which represents this,;
Wherein type to be sorted can be forest, meadow, crops, building, water body etc.;
Step 2): the selection of classification of remote-sensing images device and classification: on the basis of selection sort training sample, to remote sensing image
Selection machine learning algorithm is classified, and Classification in Remote Sensing Image result figure is obtained, it is assumed that pixel sum is N;
Step 3): the precision evaluation sample independently of classification based training sample, sample classification overall accuracy evaluation: are randomly choosed
Quantity be denoted as nA, it is nicety of grading evaluation sample that wherein symbol A, which represents this,;Then, building confusion matrix is in step 2)
The result of 1st subseries carries out precision evaluation, marks whether each sample classifies correctly, and correctly label is 1 ", the mark of mistake
It is denoted as " 0 ";Calculate the overall accuracy OA that type in the 1st subseries result is c1(c);Pay attention to precision evaluation sample be it is fixed,
Once random select will not change later in entire assorting process;Wherein, c is type forest to be sorted, meadow, agriculture
One of crop, building, water body;
Step 4): the classification correct probability of each pixel in classification based training sample is calculated: the sample evaluated using nicety of grading
This, the classification accuracy rate of each pixel on image is calculated using logistic regression model.
Wherein, the regression model of classification accuracy rate is as follows:
In formula, β0、βkFor regression coefficient, the acquisition modes of the coefficient are known in the sample evaluated by nicety of grading
2 independent variable (maximum a posteriori probability Pmax(j) and the posterior probability information entropy H of pixelc(k).P (c) is dependent variable) and strain
Amount (whether classification is correct (0 or 1)) carries out logistics recurrence and obtains.xk(c) it is independent variable, is maximum a posteriori probability respectively
Pmax(j) and the posterior probability information entropy H of pixelc(k).P (c) is dependent variable, represents the classification accuracy rate of pixel;
Step 5): setting probability threshold value optimizes training sample: setting probability threshold value, in the mistake optimized to training sample
Cheng Zhong, it is assumed that the c class threshold value of the training sample of i-th classification is Pi(c), the threshold value of the training sample of i+1 subseries is
Pi+1(c), the rule of the optimization of training sample is as follows: there was only type in i+1 subseries is c and when the accuracy of classification is big
In Pi+1(c) pixel could enter in i+1 subseries algorithm as training sample, the threshold value of i+1 subseries training sample
It is as follows with the relation formula of the threshold value of i-th classification based training sample:
Pi+1(c)=Pi(c)+ΔP(c)
Wherein, Δ P (c) is changes of threshold value obtained by being determined by i-th classification results, notices that the value of Δ P (c) can be
Positive value or negative value.
Step 6): automatic cycle 2-5 step, until the difference of the resulting classification overall accuracy of front and back step 3) calculating twice
Just stop recycling less than precision threshold: assuming that the overall accuracy of c class is in i-th cyclic sort during cyclic sort
OAi(c), the overall accuracy of i+1 subseries is OAi+1(c).The stop condition design of cyclic sort is as follows: preceding for type c
The precision threshold e (c) that the absolute value of the difference of double classification overall accuracy is less than setting afterwards just stops classification.Calculation formula is as follows
It is shown.
|OAi+1(c)-OAi(c)|<e(c)
Specifically, in step 1) when selection training sample, the quantity of each type kNot less than 1000 pixels, in sky
Between on be uniformly distributed.
Specifically, there is the optional supervised classification method of machine learning classification method in step 2): maximum likelihood method, support to
Amount machine, spectral modeling scheduling algorithm;The number of types of classification is at least more than 2 classes (containing).
Specifically, precision evaluation sample must use arbitrary sampling method in step 3), and optional random device has: simple
Random sampling, stratified randon sampling and systematic sampling etc..The ratio of sampling samples is generally 0.5%-2%, each class
Random sample is at least at 120 or more.
Specifically, the dependent variable in step 4) is " 0-1 " value in step 3), and independent variable is the assorting process from step 2)
With the maximum a posteriori probability (P being calculated in classification resultsmaxAnd posterior probability information entropy H (k))c(k)。
Maximum a posteriori probability (the P of pixelmax(j)) method calculated is as follows:
Pmax(j)=Max (P (ωi|xj))
In formula, xjIt is the spectral vector of j-th of pixel on remote sensing images;P(ωi|xj) it is by machine learning remote sensing point
X after classjBelong to ωiThe posterior probability of class;Max(P(ωi|xj)) it is xjBelong to the maximum value in inhomogeneous posterior probability.Most
Big posterior probability Pmax(j) value is bigger, belongs to ωiA possibility that a possibility that class is higher, and mistake occurs in assorting process is got over
It is low.
The posterior probability information entropy H of pixelc(k) calculation method is as follows:
In formula, m represents number of types of the remote sensing images after machine learning classification, ωiIt is i-th in m type
A type;xjIt is the spectral vector of j-th of pixel on remote sensing images;P(ωi|xj) it is the x after machine learning classificationjBelong to
ωiThe posterior probability of class;logP(ωi|xj) it is P (ωi|xj) logarithm, HcIt (j) is posterior probability information entropy.Hc(j) value
It is bigger, then illustrate that j-th of pixel belongs to inhomogeneous posterior probability and be closer to, classification ownership process occur false judgment can
Energy property is higher.
(1) specifically, the P in step 5)i(c) initial value P1(c) it is traditionally arranged to be 0.85.The rule of the setting of Δ P (c)
Then as follows: if the overall accuracy of i+1 subseries result is higher than i-th, the value of Δ P (c) is set as 0.05;Otherwise it sets
It is set to -0.05.
(2) specifically, the setting of the threshold value e (c) in step 6), is traditionally arranged to be 0.03.
Beneficial effects of the present invention mainly have following 3 aspects:
(1) when to plot be crushed region carry out land cover pattern/utilize Classification in Remote Sensing Image when, compared using the method for this patent
Conventional method, overall accuracy improve 7.1%-8.7%.This explanation, this patent are suitable for broken with respect to plot in spatial landscape
The land cover pattern in region/utilize Classification in Remote Sensing Image.
(2) when it is single to type, block type be distributed regular area and carry out Land Cover Classification when, specially using this
The method of benefit compares conventional method, and overall accuracy improves 9.3%-10.2%.This explanation, this patent are suitable in spatial landscape
Relatively regular, type single area Land Cover Classification.
(3) core of the invention innovative point is to propose a kind of remote sensing image circulation based on training sample Automatic Optimal
Classification method.The basic ideas of innovation are when the precision of classification results cannot reach required precision, before making full use of once
Precision information provided by classification results constantly adjusts and optimizes the training sample of next subseries, after then using optimization
Classification based training sample again classifies automatically to remote sensing images, is thusly-formed a cyclic sort process, until front and back twice
Nicety of grading tend towards stability just stop classification.There is this method Automatic Optimal training sample, isodose training sample to maximize
The features such as using, strong applicability high compared to same sorting algorithm precision.By verifying, this method is suitable for Classification in Remote Sensing Image application
In most supervised classification methods;Especially when containing many noises in training sample, need automatically to carry out training sample excellent
Change the case where adjustment is to improve nicety of grading.In practical applications, this method be mainly used in be related to large-scale land cover pattern/
Using or thematic (for example, crops) Classification in Remote Sensing Image.
Detailed description of the invention
Present invention will be further explained below with reference to the attached drawings and examples.
Fig. 1 (a) be Xinghua County's on March 27th, 2018 8 remote sensing image of a scape Landsat (spatial resolution:
30 meters).Fig. 1 (b) is after being standardized pretreatment (projection transform, radiation calibration, atmospheric correction etc.), using this patent skill
The sorted land cover classification figure of art.Winter wheat, rapeseed, other crops, forest, grass are broadly divided on remote sensing images
8 ground, bare area, water body, building types.
Fig. 2 (a) is on 2 7th, a 2018 scape Landsat 8 remote sensing image (spatial resolutions: 30 in Dongtai City, Jiangsu Province
Rice).Fig. 2 (b) is after being standardized pretreatment (projection transform, radiation calibration, atmospheric correction etc.), using the art of this patent
Sorted land cover classification figure.Be broadly divided on remote sensing images winter wheat, rapeseed, other crops, forest, meadow,
8 bare area, water body, building types.
Table 1 is that this patent method is obtained with traditional maximum likelihood method, support vector machines, Spectral angle mapper classification method respectively
The comparison of the overall accuracy (OA) of classification results.
Specific embodiment
In conjunction with the accompanying drawings, the present invention is further explained in detail.These attached drawings are simplified schematic diagram, only with
Illustration illustrates basic structure of the invention, therefore it only shows the composition relevant to the invention.
Case study on implementation 1: applied to the Land Cover Classification scene in spatial landscape opposed breaker region, test area with
For Xinghua County.It is winter wheat after the artificial initial selected of training sample, rapeseed, other crops, forest, meadow, naked
Ground, water body, 8 types of building quantity be respectively 1421,1384,1731,1302,1972,1544,1407,1845.Supervision
The algorithms selection of classification selects three kinds of maximum likelihood method, support vector machines, spectral modeling algorithms respectively.Initial value P1(c) in this case
0.85 is set as in example.Δ P's (c) is provided that if the overall accuracy of i+1 subseries result is higher than i-th, Δ
The value of P (c) is set as 0.05;It is otherwise provided as -0.05.Precision evaluation uses each class of method of stratified random smapling by type
Type has extracted 150 samples.Classification outage threshold e (c) is set as 0.03.Finally with the overall essence in Classification in Remote Sensing Image precision evaluation
Spend (OA) as assess this patent sorting algorithm index, and with traditional maximum likelihood method, support vector machines, spectral modeling three
Kind algorithm obtains precision index and is compared.
Test result shows: carrying out Land Cover Classification for the Xinghua City in spatial landscape opposed breaker region, adopts
With the method for traditional analysis of spectrum threshold optimization training sample, maximum likelihood method, support vector machines, the overall essence of classification of spectral modeling
Degree is 71.2%, 75.7%, 72.0% respectively;Sample Automatic Optimal is trained using the method for this patent, three kinds of methods obtain
Overall accuracy to classification results is respectively 79.9%, 83.4%, 79.1%, and overall accuracy has been respectively increased 8.7%, 7.7%,
7.1%.This explanation is obviously improved than conventional method in the classification method that spatial landscape opposed breaker region this patent proposes
The precision of 7.1%-8.7%.
Case study on implementation 2:, the Land Cover Classification field in type more single area relatively regular applied to spatial landscape
Scape, test area is by taking Dongtai City, Jiangsu Province as an example.Winter wheat, other crops, forest, grass after the artificial initial selected of training sample
Ground, bare area, water body, 7 types of building quantity be respectively 1181,1078,1213,1102,1411,1123,1022,
1430.The algorithms selection of supervised classification selects three kinds of maximum likelihood method, support vector machines, spectral modeling algorithms respectively.Initial value P1
(c) 0.85 is set as in present case.Δ P's (c) is provided that if the overall accuracy of i+1 subseries result is higher than i-th
It is secondary, then the value of Δ P (c) is set as 0.05;It is otherwise provided as -0.05.Precision evaluation is using stratified random smapling by type
The each type of method has extracted 150 samples.Classification outage threshold e (c) is set as 0.03.Finally with Classification in Remote Sensing Image precision evaluation
In overall accuracy (OA) as assessment this patent sorting algorithm index, and with traditional maximum likelihood method, supporting vector
Three kinds of machine, spectral modeling algorithms obtain precision index and are compared.
As shown in Table 1, test result shows: it is distant that the Dongtai City in region relatively regular for spatial landscape carries out land cover pattern
Sense classification, using the method for traditional analysis of spectrum threshold optimization training sample, maximum likelihood method, support vector machines, u spectral modeling
Overall accuracy of classifying is 78.1%, 82.7%, 76.5% respectively;Sample Automatic Optimal is trained using the method for this patent,
The overall accuracy that three kinds of methods obtain classification results is respectively 87.4%, 92.9%, 86.3%, and overall accuracy is respectively increased
9.3%, 10.2%, 9.8%.This explanation obviously compares in the classification method that the relatively regular region this patent of spatial landscape proposes
Conventional method improves the precision of 9.3%-10.2%.
Table 1: this patent method is obtained with traditional maximum likelihood method, support vector machines, Spectral angle mapper classification method respectively
The comparison of the overall accuracy (OA) of classification results
Taking the above-mentioned ideal embodiment according to the present invention as inspiration, through the above description, relevant staff is complete
Various changes and amendments can be carried out without departing from the scope of the technological thought of the present invention' entirely.The technology of this invention
Property range is not limited to the contents of the specification, it is necessary to which the technical scope thereof is determined according to the scope of the claim.
Claims (6)
1. a kind of remote sensing image cyclic sort method based on training sample Automatic Optimal, mainly comprises the steps of:
Step 1): selection classification of remote-sensing images training sample;
Step 2): the selection of classification of remote-sensing images device and classification;
Step 3): classification overall accuracy evaluation;
Step 4): the classification correct probability of each pixel in classification based training sample is calculated;
Step 5): setting probability threshold value optimizes training sample;
Step 6): automatic cycle 2-5 step, until front and back, the difference that step 3) calculates resulting classification overall accuracy twice is less than
Precision threshold just stops recycling.
2. a kind of remote sensing image cyclic sort method based on training sample Automatic Optimal as described in claim 1, feature
It is:
The step 1) is to step 6) specifically:
Step 1): it selects classification of remote-sensing images training sample: being each type to be sorted on remote sensing image to be sorted
Training sample is selected, the quantity of the training sample of k-th of type is denoted asIt is training sample that wherein symbol T, which represents this,;
Step 2): the selection of classification of remote-sensing images device and classification: on the basis of selection sort training sample, remote sensing image is selected
Machine learning algorithm is classified, and Classification in Remote Sensing Image result figure is obtained, it is assumed that pixel sum is N;
Step 3): the precision evaluation sample independently of classification based training sample, the number of sample classification overall accuracy evaluation: are randomly choosed
Amount is denoted as nA, it is nicety of grading evaluation sample that wherein symbol A, which represents this,;Then, building confusion matrix is to the 1st time in step 2)
The result of classification carries out precision evaluation, marks whether each sample classifies correctly, and correctly label is 1 ", the label of mistake for
" 0 " calculates the overall accuracy OA that the 1st subseries type is c1(c);
Step 4): calculate the classification correct probability of each pixel in classification based training sample: the sample evaluated using nicety of grading is adopted
The classification accuracy rate of each pixel on image is calculated with logistic regression model,
Wherein, the regression model of accuracy is as follows:
In formula, β0、βkFor regression coefficient, xkIt (c) is independent variable, P (c) is the accuracy of pixel;
Step 5): setting probability threshold value optimizes training sample: setting probability threshold value, in the process optimized to training sample
In, it is assumed that the c class threshold value of the training sample of i-th classification is Pi(c), the threshold value of the training sample of i+1 subseries is Pi+1
(c), the rule of the optimization of training sample is as follows: there was only type in i+1 subseries is c and when the accuracy of classification is greater than
Pi+1(c) pixel could enter in i+1 subseries algorithm as training sample, the threshold value of i+1 subseries training sample with
The relation formula of the threshold value of i-th classification based training sample is as follows:
Pi+1(c)=Pi(c)+ΔP(c)
Wherein, Δ P (c) is changes of threshold value obtained by being determined by i-th classification results;
Step 6): automatic cycle 2-5 step, until front and back, the difference that step 3) calculates resulting classification overall accuracy twice is less than
Precision threshold just stops recycling: assuming that the overall accuracy of c class is OA in i-th classification during cyclic sorti(c), i-th
The overall accuracy of+1 subseries is OAi+1(c);The stop condition design of cyclic sort is as follows: for type c, front and back double classification
The precision threshold e (c) that the absolute value of the difference of overall accuracy is less than setting just stops classification, and calculation formula is as follows:
|OAi+1(c)-OAi(c)|<e(c)。
3. a kind of remote sensing image cyclic sort method based on training sample Automatic Optimal as claimed in claim 2, feature
It is: when selection training sample, the quantity of each type kNot less than 1000 pixels, are spatially uniformly distributed.
4. a kind of remote sensing image cyclic sort method based on training sample Automatic Optimal as claimed in claim 2, feature
It is: the Pi(c) initial value P1(c) it is set as 0.85;The rule of the setting of the Δ P (c) is as follows: if i+1 time point
The overall accuracy of class result is higher than i-th, then the value of Δ P (c) is set as 0.05;It is otherwise provided as -0.05.
5. a kind of remote sensing image cyclic sort method based on training sample Automatic Optimal as claimed in claim 2, feature
Be: the threshold value e (c) is set as 0.03.
6. a kind of remote sensing image cyclic sort method based on training sample Automatic Optimal as claimed in claim 2, feature
Be: the machine learning algorithm is maximum likelihood method, support vector machines or spectrum horn cupping.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811503807.1A CN109726641B (en) | 2019-01-24 | 2019-01-24 | Remote sensing image cyclic classification method based on automatic optimization of training samples |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811503807.1A CN109726641B (en) | 2019-01-24 | 2019-01-24 | Remote sensing image cyclic classification method based on automatic optimization of training samples |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109726641A true CN109726641A (en) | 2019-05-07 |
CN109726641B CN109726641B (en) | 2023-02-28 |
Family
ID=66295916
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811503807.1A Active CN109726641B (en) | 2019-01-24 | 2019-01-24 | Remote sensing image cyclic classification method based on automatic optimization of training samples |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109726641B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113222005A (en) * | 2021-05-08 | 2021-08-06 | 兰州交通大学 | Automatic updating method for land coverage |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103489005A (en) * | 2013-09-30 | 2014-01-01 | 河海大学 | High-resolution remote sensing image classifying method based on fusion of multiple classifiers |
CN104063718A (en) * | 2014-04-01 | 2014-09-24 | 中国科学院遥感与数字地球研究所 | Method for selecting remote sensing data and classification algorithms in crop identification and area estimation |
CN104331716A (en) * | 2014-11-20 | 2015-02-04 | 武汉图歌信息技术有限责任公司 | SVM active learning classification algorithm for large-scale training data |
CN105303198A (en) * | 2015-11-17 | 2016-02-03 | 福州大学 | Remote-sensing image semi-supervision classification method based on customized step-size learning |
CN108038448A (en) * | 2017-12-13 | 2018-05-15 | 河南理工大学 | Semi-supervised random forest Hyperspectral Remote Sensing Imagery Classification method based on weighted entropy |
CN108490344A (en) * | 2018-03-28 | 2018-09-04 | 电子科技大学 | Analog-circuit fault diagnosis method based on training sample optimization |
-
2019
- 2019-01-24 CN CN201811503807.1A patent/CN109726641B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103489005A (en) * | 2013-09-30 | 2014-01-01 | 河海大学 | High-resolution remote sensing image classifying method based on fusion of multiple classifiers |
CN104063718A (en) * | 2014-04-01 | 2014-09-24 | 中国科学院遥感与数字地球研究所 | Method for selecting remote sensing data and classification algorithms in crop identification and area estimation |
CN104331716A (en) * | 2014-11-20 | 2015-02-04 | 武汉图歌信息技术有限责任公司 | SVM active learning classification algorithm for large-scale training data |
CN105303198A (en) * | 2015-11-17 | 2016-02-03 | 福州大学 | Remote-sensing image semi-supervision classification method based on customized step-size learning |
CN108038448A (en) * | 2017-12-13 | 2018-05-15 | 河南理工大学 | Semi-supervised random forest Hyperspectral Remote Sensing Imagery Classification method based on weighted entropy |
CN108490344A (en) * | 2018-03-28 | 2018-09-04 | 电子科技大学 | Analog-circuit fault diagnosis method based on training sample optimization |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113222005A (en) * | 2021-05-08 | 2021-08-06 | 兰州交通大学 | Automatic updating method for land coverage |
CN113222005B (en) * | 2021-05-08 | 2022-05-27 | 兰州交通大学 | Automatic updating method for land coverage |
Also Published As
Publication number | Publication date |
---|---|
CN109726641B (en) | 2023-02-28 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022160771A1 (en) | Method for classifying hyperspectral images on basis of adaptive multi-scale feature extraction model | |
CN107229917B (en) | A kind of several remote sensing image general character well-marked target detection methods based on iteration cluster | |
EP2681715B1 (en) | Method and software for analysing microbial growth | |
CN104599275B (en) | The RGB-D scene understanding methods of imparametrization based on probability graph model | |
CN104484681B (en) | Hyperspectral Remote Sensing Imagery Classification method based on spatial information and integrated study | |
CN103473786B (en) | Gray level image segmentation method based on multi-objective fuzzy clustering | |
CN104732244B (en) | The Classifying Method in Remote Sensing Image integrated based on wavelet transformation, how tactful PSO and SVM | |
CN104123561B (en) | Fuzzy C-mean algorithm remote sensing image automatic classification method based on spatial attraction model | |
CN109063754B (en) | Remote sensing image multi-feature joint classification method based on OpenStreetMap | |
CN109684906B (en) | Method for detecting red fat bark beetles based on deep learning | |
CN110766058B (en) | Battlefield target detection method based on optimized RPN (resilient packet network) | |
CN105320970B (en) | A kind of potato disease diagnostic device, diagnostic system and diagnostic method | |
CN109684922A (en) | A kind of recognition methods based on the multi-model of convolutional neural networks to finished product dish | |
CN111310756A (en) | Damaged corn particle detection and classification method based on deep learning | |
CN108428220A (en) | Satellite sequence remote sensing image sea island reef region automatic geometric correction method | |
CN108960404A (en) | A kind of people counting method and equipment based on image | |
CN108921210B (en) | Cloud classification method based on convolutional neural network | |
CN109544538A (en) | Wheat scab disease grade grading method and device | |
CN109034233A (en) | A kind of high-resolution remote sensing image multi classifier combination classification method of combination OpenStreetMap | |
CN101980251A (en) | Remote sensing classification method for binary tree multi-category support vector machines | |
CN115994907B (en) | Intelligent processing system and method for comprehensive information of food detection mechanism | |
CN109145848A (en) | A kind of wheat head method of counting | |
CN108154158A (en) | A kind of building image partition method applied towards augmented reality | |
CN110503140A (en) | Classification method based on depth migration study and neighborhood noise reduction | |
CN109671019A (en) | A kind of remote sensing image sub-pixed mapping drafting method based on multi-objective optimization algorithm and sparse expression |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |