CN106780503A - Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method - Google Patents

Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method Download PDF

Info

Publication number
CN106780503A
CN106780503A CN201611253668.2A CN201611253668A CN106780503A CN 106780503 A CN106780503 A CN 106780503A CN 201611253668 A CN201611253668 A CN 201611253668A CN 106780503 A CN106780503 A CN 106780503A
Authority
CN
China
Prior art keywords
segmentation
yardstick
posterior probability
dividing body
pixel
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
Application number
CN201611253668.2A
Other languages
Chinese (zh)
Inventor
曹鑫
许飞
陈学泓
崔喜红
陈晋
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Normal University
Original Assignee
Beijing Normal University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing Normal University filed Critical Beijing Normal University
Priority to CN201611253668.2A priority Critical patent/CN106780503A/en
Publication of CN106780503A publication Critical patent/CN106780503A/en
Pending legal-status Critical Current

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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30181—Earth observation

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

Determine method the invention provides a kind of remote sensing images optimum segmentation yardstick based on posterior probability information entropy.Methods described is included the multi-scale division of remote sensing image, the optimum segmentation yardstick of single body is automatically selected based on comentropy change indicator.The method obtains the posterior probability vector of each pixel based on SVM classifier, then calculate different scale under each dividing body average posterior probability vector, calculate again it is different segmentation yardsticks under each dividing body comentropy, obtain the entropy curve of each pixel affiliated dividing body under different segmentation yardsticks, the first-order difference of entropy and segmentation yardstick is calculated according to entropy curve and the maximum yardstick of changes of entropy is found, finally using the yardstick as the optimum segmentation yardstick of current pixel.The method effectively overcomes " over-segmentation " and " low segmentation " problem in object-oriented segmentation, is a kind of effective ground target automatic division method while realizing automatically selecting for object optimal scale.

Description

Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method
Technical field
The present invention relates to the processing method of remote sensing image, the remote sensing images most optimal sorting of posterior probability information entropy is based particularly on Cut yardstick and determine method, belong to image processing field.
Background technology
High spatial resolution remote sense image is the significant data that the mankind precisely monitor ground mulching, the image based on pixel point The problem of analysis method generally existing " spiced salt phenomenon ", so-called " spiced salt phenomenon " refers to object spectrum image difference fragmentation of the same race Phenomenon.The phenomenon seriously constrains the ground target extraction accuracy of high spatial resolution remote sense image.In order to solve this problem, Image analysis methods based on object-oriented are gradually paid attention to and are used by researcher in the industry.
With the development of remote sensing science and technology, the spaceborne high spatial resolution image of business, airborne high spatial resolution image It has been widely used in urban planning, ground mulching drawing.Image Segmentation is that a series of dividing bodies are divided into spectrum image, Each dividing body is made up of one group of homogeneous pixel.Wherein, segmentation yardstick is the weight for determining dividing body size and homogeneous degree Index is wanted, segmentation yardstick is smaller, and the area of dividing body is also smaller, and homogeneous degree is also higher；Segmentation yardstick is bigger, the face of dividing body Product is also bigger, and homogeneous degree is also lower.In current dividing method, people rely primarily on the method choice of experience and trial and error most Excellent segmentation yardstick, this method hinders the universality of dividing method, reduces the efficiency of dividing method.
In high spatial resolution image, there is their actual physical yardstick on every house, each piece of lawn, i.e., completely Cover the size of the dividing body of these atural objects.But during Image Segmentation, any single segmentation yardstick cannot all be accorded with simultaneously The actual physical yardstick of group photo all ground objects as in.Therefore multi-scale segmentation method is more suitable for solution automatic target extraction and asks Topic, is that each ground object in image finds optimal segmentation yardstick, will greatly improve the accuracy of identification of ground target.
In the last few years, numerous researchers were devoted to solve the select permeability of Image Segmentation optimal scale.Lucian Dragut etc. proposes to automatically determine the optimum segmentation yardstick of image using local variance, referring to《For estimating many of remotely-sensed data The instrument of the scale parameter of image in different resolution segmentation》(a tool to estimate scale parameter for Multiresolution of remotely sensed data) (Dragut L, international Geographical Information Sciences magazine, 2010, 24(6)：859-871).The method uses same optimal scale for all of object on image.But in majority of case Under, the object size in a width image is different, and same segmentation scale dimension applications are split clearly in different size of object It is irrational.Additionally, T-Esch etc. proposes a kind of method for automatically determining the respective optimum segmentation yardstick of different objects, referring to 《The improvement of the image segmentation precision based on multi_dimension optimization》(Improvement of image segmentation Accuracy based on multiscale optimization procedure) (Esch T, geoscience and remote sensing Report, IEEE, 2008,5 (3)：463-467).But substantial amounts of parameter is used for the determination of optimal scale, this causes whole algorithm Become suitable complexity.For simultaneously for different application purposes, even if there is also different optimum segmentations with piece image Yardstick.Such as, when using house as need extract dividing body when need a relatively large yardstick, but when using automobile as Need but to need a relatively small yardstick when extracting dividing body.Therefore, we can automatically determine shadow in the urgent need to one kind The method of different object optimum segmentation yardsticks as in.
The content of the invention
For this determines method the invention provides a kind of remote sensing images optimum segmentation yardstick based on posterior probability information entropy, The method can be reduced or avoided problem noted earlier.
To solve the above problems, the remote sensing images optimum segmentation yardstick based on comentropy change indicator that the present invention is provided is true Determine method, it comprises the following steps：
Step A, carries out multi-scale division and calculates pixel level posterior probability vector to remote sensing image：The step is utilized The softwares of eCognition 8.9 obtain multiple dimensioned segmentation result；The posterior probability for calculating each pixel using SVM classifier is sweared Amount, calculates the average posterior probability of each dividing body under different segmentation yardsticks.
Specifically, first, carrying out the segmentation of multiple yardsticks to remote sensing image using the softwares of eCognition 8.9, obtain many The segmentation result of yardstick.Secondly, selection pixel level training sample, and using SVM classifier calculate original multiband image each The posterior probability vector of pixel.3rd, according to multi-scale division result, calculate the average of each dividing body under each segmentation yardstick Posterior probability.
Step B, the optimum segmentation yardstick of single dividing body is selected based on comentropy change indicator：It is general that the step is based on posteriority Rate calculate comentropy as dividing body classification Evaluation of Uncertainty index, according to single object it is different segmentation yardsticks under entropy Change determines its optimum segmentation yardstick.
According to the multiple dimensioned posterior probability vector that step A is obtained, posterior probability information entropy is calculated to each dividing body, The calculating of entropy is with reference to as follows：
Wherein Pj,iRepresent that j-th dividing body under s segmentation yardsticks belongs to the probability of the i-th class, n represents classification number, Es,jGeneration J-th information entropy of dividing body under table s segmentation yardsticks.
Based on the comentropy under each yardstick, comentropy is calculated with the first-order difference Δ E of adjacent segmentation dimensional variation, will The segmentation yardstick of the maximum value position of the comentropy change of acquisition is used as optimum segmentation yardstick.
The invention has the advantages that：
The inventive method is automatically determined often based on ground object comentropy change indicator of dividing body under different scale Individual dividing body optimum segmentation yardstick.
Other features and advantages of the present invention will be illustrated in the following description, also, the partial change from specification Obtain it is clear that or being understood by implementing the present invention.The purpose of the present invention and other advantages can be by the explanations write Specifically noted structure is realized and obtained in book, claims and accompanying drawing.
Brief description of the drawings
The following drawings is only intended to, in schematic illustration and explanation is done to the present invention, not delimit the scope of the invention.Wherein,
Fig. 1 is the Remote Sensing Image Segmentation yardstick based on posterior probability information entropy according to a specific embodiment of the invention Determine the schematic flow sheet of method；
Fig. 2 is the inclusion relation schematic diagram between dividing body under eCognition adjacent segmentation yardsticks；
The principle schematic of the method that Fig. 3 is provided for the present invention；
Fig. 4 (a) and Fig. 4 (b) are respectively optimal according to the remote sensing images based on posterior probability information entropy of present invention offer Segmentation yardstick determines two design sketch of example of method.
Specific embodiment
In order to be more clearly understood to technical characteristic of the invention, purpose and effect, now illustrate of the invention specific Implementation method.But it will be appreciated by those skilled in the art that, following examples are not to the unique of technical solution of the present invention work Restriction, every any equivalents done under technical solution of the present invention Spirit Essence or change, are regarded as belonging to this hair Bright protection domain.
Fig. 1 is the Remote Sensing Image Segmentation yardstick based on posterior probability information entropy according to a specific embodiment of the invention Determine the schematic flow sheet of method；Shown in reference picture 1, referred to based on comentropy change the following detailed description of being provided according to the present invention Target remote sensing images optimum segmentation yardstick determines the principle of method, and methods described includes following two big steps：
Step A, carries out multi-scale division and calculates pixel level posterior probability vector to remote sensing image.
Step B, optimum segmentation yardstick is selected based on comentropy change indicator.
For remote sensing image, before processing procedure of the present invention is carried out, it may be preferred to carry out atmospheric correction to image, Removal air image.In addition, for target in hyperspectral remotely sensed image, in order to reduce amount of calculation, treatment of the present invention can be being carried out Before process, Hyperspectral imaging is carried out preferably by principal component analysis (PCA, Principle Component Analysis) Dimension-reduction treatment.During the feature extraction of image is carried out, when the intrinsic dimensionality of extraction often causes characteristic matching too much It is excessively complicated, consume system resource, therefore method frequently with Feature Dimension Reduction is processed Hyperspectral imaging.So-called feature drop Dimension, i.e., represent high-dimensional using a feature for low dimensional.Feature Dimension Reduction generally carries out feature extraction with PCA to carry out dimensionality reduction Treatment.Above preferred steps were used before process step of the invention, for example, first can carry out Atmospheric Correction then Dimension-reduction treatment is carried out, or can first carry out dimension-reduction treatment to carry out atmospheric correction again, the two step is successively for after of the invention The effect of continuous treatment is similar to.
Step A and step B are discussed in detail separately below：
Step A, carries out multi-scale division and calculates pixel level posterior probability vector to remote sensing image
This step first has to carry out multi-scale division to remote sensing image, then calculates each segmentation in multi-scale division result The posterior probability vector of body.
The present invention carries out the segmentation of multiple yardsticks using the softwares of eCognition 8.9 to remote sensing image first.During segmentation, Form factor is set to 0.2, and degree of the compacting factor is set to 0.5.This example is step-length with 10, and 1000 are incremented to from 50, selection Multi-stage division yardstick is that 50,60,70,80 ..., 970,980,990,1000 (the segmentation principle of eCognition softwares is referring to figure 2).Wherein, segmentation yardstick is proportionate with the area of dividing body, and the homogenieity with dividing body interior pel species is negatively correlated.
Although eCognition softwares can produce it is different segmentation yardsticks segmentation result (such as 50,60 ..., 980, 990,1000) all atural objects, but in the same segmentation multi-scale segmentation image of these segmentation results use.In actual conditions, Atural object not of the same race has the geometric parameters such as different physical sizes, i.e. area, girth and diameter.Such as house and meadow, the two Plaque area in image is of different sizes, therefore should be split with different segmentation yardsticks in Image Segmentation.To sum up Described, the segmentation result that eCognition softwares are directly produced cannot be directly used to the ground target identification of remote sensing image.
To solve this technical problem, the present invention by calculating comentropy change indicator, comprehensive utilization eCognition from " over-segmentation " arrives the segmentation result (correspondence segmentation yardstick 50~1000) of " low segmentation ", automatically selects out every kind of atural object in image Optimum segmentation yardstick, final production goes out a multi-scale division result, wherein, as shown in Fig. 2 yardstick from big to small adjacent point Cutting the segmentation result of yardstick has the relation of complete or collected works and subset.
That is, segmentation when, it is (regular that the segmentation range scale according to selection creates rule set in the softwares of eCognition 8.9 The creation method of collection is discussed in detail in eCognition software documents), and split step by step from small to large according to yardstick.Point After cutting, the segmentation result of the different scale of eCognition Software Creates is exported as into several Raster Images (Tiff forms Or img forms), the corresponding Raster Images of segmentation result of wherein each segmentation yardstick.In each Raster Images, each Pixel in dividing body is all endowed same label value DN, that is, the DN values are the label of the dividing body.It is excellent at one Select in embodiment, the selection of multi-stage division yardstick can also determine according to the scope of the DN values of image or reflectivity.
Followed by calculate the posterior probability vector of each pixel.The present invention uses SVMs (SVM, support Vector machine) the single pixel of classifier calculated posterior probability vector.SVM classifier is a kind of complicated classification times It is engaged in being mapped by kernel function and is allowed to be converted into a problem for constructing linear classification hyperplane in high-dimensional feature space, most optimal sorting Class hyperplane can be obtained by solving a quadratic programming problem.Different classes of sample point and hyperplane it is vertical European away from From the posterior probability that can be converted to such sample, the vector of the posterior probability composition of all categories is referred to as posterior probability arrow Amount.SVM classifier is a kind of sorting technique of mature, and the code, principle of the method can be by public network platforms It is free to obtain, referring to《The application study that SVMs and its remote sensing image space characteristics are extracted and classified》(Luo Jiancheng etc., it is distant Sense journal, 2010,24 (6)：859-871).
Namely using SVM classifier calculate multi-scale division result in each dividing body posterior probability vector it Before, the posterior probability vector of each pixel is obtained using sample training, concretely comprise the following steps, selection pixel level training sample is used SVM classifier calculates the posterior probability vector of each pixel, such as every kind of ground mulching type from from original multiband image 3000-5000 pixel of middle selection is sweared as training sample, the posterior probability for being calculated each pixel using SVM classifier Amount.
Subsequently, the multi-scale division result of the grid for being obtained based on above-mentioned steps, and each pixel Posterior probability vector, calculates the average posterior probability vector of each dividing body under each segmentation yardstick respectively.Assuming that s is split J-th dividing body includes n pixel under yardstick, then the average posterior probability vector of the dividing bodyComputational methods it is as follows：
WhereinThe average posterior probability of k-th classification of current dividing body is represented, C represents total classification number, and n is represented and worked as The number of all pixels, p in preceding dividing bodyi,kRepresent i-th posterior probability of k-th classification of pixel.According to the above method, Calculate the posterior probability vector of each object in all segmentation yardsticks.
Step B, optimum segmentation yardstick is selected based on comentropy change indicator
The multiple dimensioned posterior probability vector that the step according to step A is obtained can be obtained：When dividing body is in over-segmentation (when i.e. segmentation yardstick is smaller), the pixel essentially same category included in dividing body, the classification homogenieity of interior pel is more It is high.It is more comprising different types of pixel in dividing body when dividing body is in low segmentation (when segmentation yardstick is larger), it is internal The classification homogenieity of pixel is lower.Based on above thought, the present invention propose by the use of comentropy as weigh dividing body inner classes The index of other homogenieity.Multilayer posterior probability (the segmentation knot of yardstick adjacent segmentation yardstick from big to small tried to achieve according to step A The segmentation result that fruit has the relation of complete or collected works and subset, difference segmentation yardstick is represented with stratiform, referring to Fig. 2), to each segmentation Body is calculated the entropy of posterior probability by the incremental order of segmentation yardstick respectively, and the computing formula of entropy is as follows：
Wherein Pj,iRepresent that j-th dividing body under s segmentation yardsticks belongs to the probability of the i-th class, n represents classification number, Es,jGeneration J-th information entropy of dividing body under table s segmentation yardsticks.
When object is in over-segmentation state, dividing body internal sort is single, and information entropy is relatively low；Equally, at object When low segmentation, dividing body is larger, and the mixing of plurality of classes causes entropy larger.Now, based on it is different segmentation yardsticks under point The comentropy of body is cut, can be calculated and be obtained comentropy as the first-order difference of adjacent segmentation dimensional variation is (for example, when in image (posterior probability is respectively p two kinds of classifications1And p2), wherein p1It is the posterior probability of the true classification of Target scalar, p2It is target ground The posterior probability of other classifications around thing).The first-order difference of comentropy is expressed as follows：
The wherein first-order difference of Δ E representative informations entropy,It is that comentropy seeks partial derivative to every kind of classification.Work as p2When minimum, Δ E is maximum.This means during comentropy change maximum, dividing body contains neighbouring impurity pixel.Therefore, with segmentation yardstick In the comentropy curve of change, according to above-mentioned derivation, positioned at the maximum Δ E of comentropy changemaxPosition segmentation yardstick make Be optimum segmentation yardstick, and it is final segmentation result to make the dividing body of the yardstick, i.e. abovementioned steps (50,60, 70 ..., 990,1000) in multi-scale division result, the first-order difference of adjacent segmentation dimensional variation is calculated, and obtains the letter of maximum Breath entropy changing value Δ Emax:
Finally with Δ EmaxSegmentation yardstick as optimum segmentation yardstick.Fig. 3 is that the present invention selects single object most Optimal sorting cuts the schematic diagram of yardstick.
In order to better illustrate technique effect of the invention, for a panel height spectrum image, proposition of the present invention is utilized respectively The remote sensing images optimum segmentation yardstick based on posterior probability information entropy determine method, using multiple dimensioned point of variance stable region Cut algorithm and single multi-scale segmentation algorithm is tested, then the result after segmentation is compared.Wherein, single yardstick point It is artificial selection to cut the yardstick that algorithm uses：50,60,70 are produced respectively ..., the segmentation result under 990,1000 segmentation yardsticks, Selection segmentation precision highest yardstick.Additionally, the stability in order to verify the method, the present invention respectively in different sample sizes and Test of many times has been carried out under different segmentation step elongate members.Different categories of samples size is with 2000 for interval increases to from 1000 pixels 9000 pixels.To reduce amount of calculation, before the treatment of method of offer of the present invention is provided, calculated first with principal component analysis (principal component analytical method is a kind of mathematical method being commonly used to method, and the software of the method, code and principle can be public Network platform Free Acquisition) by image dimension-reduction treatment, and preceding 8 wave bands are extracted for the treated of the method for present invention offer Journey.
Compared to using variance stable region multi-scale division algorithm, by using according to the present invention provide based on information The remote sensing images optimum segmentation yardstick of entropy change indicator determines that method carries out the multi-scale division result of remote sensing image, effectively goes Except the segmentation error that ' spiced salt phenomenon ' noise is brought, overwhelming majority building and lawn are all presented with actual physical yardstick.
Fig. 4 (a) and Fig. 4 (b) are respectively the remote sensing images optimum segmentation based on posterior probability information entropy of present invention offer Yardstick determines two result figures of sample that method is obtained.From fig. 4, it can be seen that the method that the present invention is provided can find each The optimum segmentation yardstick of atural object, the dividing body under all optimum segmentation yardsticks constitutes multi-scale division result.
Table 1 is the remote sensing images optimum segmentation yardstick determination side based on posterior probability information entropy provided according to the present invention Method, single multi-scale segmentation algorithm, and using the multi-scale division algorithm of variance stable region, carry out multi-scale division result Overall accuracy statistical form.Wherein, be may refer to the algorithm of method of the present invention relevant comparative and segmentation principle, term etc. 《The multi-scale segmentation method classified for object-oriented ground mulching using high spatial resolution image》(Multi-scale segmentation approach for object-based land-cover classification using high- Resolution imagery) (etc., remote sensing bulletin, 2014,5 (1):73-82).
From table 1 it follows that the remote sensing images optimum segmentation based on posterior probability information entropy provided using the present invention Yardstick determines the segmentation precision highest that method is obtained.Wherein, multi-scale division algorithm proposed by the present invention, in F-measure and All more outstanding than other two methods under two kinds of test ratings of BCI, both test ratings can be in " over-segmentation ", " most optimal sorting Cut ", effective evaluation is carried out to dividing body in " low segmentation " three kinds of cutting states, referring to《Precision based on region and recall measure Segmentation quality evaluation is carried out to remote sensing image》(Segmentation quality evaluation using region- Based precision and recall measures for remote sensing images) (etc., ISPRS photographies Measurement and remote sensing journal, 2015,102 (5), 73-84).
Table 1
Table 2 is the remote sensing images optimum segmentation yardstick determination side based on posterior probability information entropy provided according to the present invention Method, the multi-scale division result accuracy assessment under different segmentation yardstick step-lengths and different training sample amount sizes；Can be with from table 2 Find out, the method that this research is provided can keep the segmentation precision relatively stablized under different segmentation yardstick intervals；Reached in sample size During to 5000 pixels, segmentation precision highest.Additionally, under the conditions of all segmentation step-lengths and training sample, what the present invention was provided The segmentation precision that method is obtained all is higher than other two methods, and the method that this explanation is originally researched and proposed is to carry out having for Image Segmentation Efficacious prescriptions method.
Table 2
Be can be seen that by above-mentioned graphical example, it is optimal according to the remote sensing images based on posterior probability information entropy that the present invention is provided Segmentation yardstick determines that method can effectively split the ground object in high score image.
Although it will be appreciated by those skilled in the art that the present invention is described according to the mode of multiple embodiments, together The Shi Jinhang across comparisons of several method, but not each embodiment only includes independent technical scheme.Specification In so narration just for the sake of for the sake of clear, those skilled in the art should be understood specification as an entirety, And technical scheme involved in each embodiment being regarded as, can be mutually combined into the mode of different embodiments understands this hair Bright protection domain.
Schematical specific embodiment of the invention is the foregoing is only, the scope of the present invention is not limited to.It is any Those skilled in the art, the equivalent variations made on the premise of design of the invention and principle is not departed from, modification and combination, The scope of protection of the invention all should be belonged to.

Claims (7)

1. a kind of remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method, it is characterised in that method bag Include following steps：
Step A, carries out multi-scale division and calculates pixel level posterior probability vector to remote sensing image：
The step is segmentation step-length with 10 using the softwares of eCognition 8.9, successively with (50,60,70 ..., 990,1000) point Cut yardstick to split remote sensing images, obtain multi-scale division result；
Selection pixel level training sample, the posterior probability vector of each pixel is calculated using SVM classifier, based on described multiple dimensioned Segmentation, and each pixel posterior probability vector, respectively result calculate each segmentation yardstick under each dividing body it is average Posterior probability vector.
Step B, optimum segmentation yardstick is selected based on comentropy change indicator：
The multiple dimensioned posterior probability vector that the step is obtained according to step A, is incremented by each dividing body by segmentation yardstick Order calculates the entropy of posterior probability respectively, and the computing formula of entropy is as follows：
E s , j = - Σ i = 1 n p s , j , i × log 2 ( P s , j , i )
Wherein Pj,iRepresent that j-th dividing body under s segmentation yardsticks belongs to the probability of the i-th class, n represents classification number, Es,jRepresent s points Cut j-th information entropy of dividing body under yardstick.
Based on the comentropy under each yardstick, comentropy is calculated with the first-order difference Δ E of adjacent segmentation dimensional variation, will obtain Comentropy change maximum value position segmentation yardstick as optimum segmentation yardstick.
2. method according to claim 1, it is characterised in that methods described also includes advancing with principal component analytical method The process of dimension-reduction treatment is carried out to target in hyperspectral remotely sensed image.
3. method according to claim 1 and 2, it is characterised in that the step A is further included：Utilizing svm classifier Before device calculates the posterior probability vector of each dividing body in multi-scale division result, training is chosen from original multiband image Sample, the posterior probability vector of each pixel is calculated using SVM classifier.
4. method according to claim 1 and 2, it is characterised in that the step A is further included：Utilizing When the softwares of eCognition 8.9 carry out multi-scale division to remote sensing image, the selection of multi-split yardstick according to the DN values of image or The scope of reflectivity determines that during segmentation, form factor is set to 0.2, and degree of the compacting factor is set to 0.5.
5. method according to claim 1 and 2, it is characterised in that the step A is further included：According to dividing for selection Cut range scale and rule set is created in the softwares of eCognition 8.9, and split step by step from small to large according to yardstick.
6. method according to claim 1 and 2, it is characterised in that the step A is further included：After segmentation, by life Into the segmentation result of different scale export as Raster Images, the pixel in each Raster Images in each dividing body is endowed Same label value as the dividing body label.
7. method according to claim 1 and 2, it is characterised in that the step A is further included：Based on above-mentioned steps The multi-scale division result of the grid of acquisition, and each pixel posterior probability vector, each is calculated respectively The average posterior probability vector of each dividing body under segmentation yardstick.Assuming that j-th dividing body includes n picture under s segmentation yardsticks Unit, the then average posterior probability vector of the dividing bodyComputational methods it is as follows：
P ‾ = [ p ‾ 1 , ... , p ‾ k , ... , p ‾ C ]
p ‾ k = Σ i = 1 n p i , k n
WhereinThe average posterior probability of k-th classification of current dividing body is represented, C represents total classification number, and n represents current point Cut the number of internal all pixels, pi,kRepresent i-th posterior probability of k-th classification of pixel.According to the above method, calculate Go out the posterior probability vector of each object in all segmentation yardsticks.
CN201611253668.2A 2016-12-30 2016-12-30 Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method Pending CN106780503A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201611253668.2A CN106780503A (en) 2016-12-30 2016-12-30 Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201611253668.2A CN106780503A (en) 2016-12-30 2016-12-30 Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method

Publications (1)

Publication Number Publication Date
CN106780503A true CN106780503A (en) 2017-05-31

Family

ID=58954031

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201611253668.2A Pending CN106780503A (en) 2016-12-30 2016-12-30 Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method

Country Status (1)

Country Link
CN (1) CN106780503A (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108053406A (en) * 2017-09-11 2018-05-18 同济大学 Ground mulching drafting method and device based on multiple-spatial resolution remote sensing image
CN108230329A (en) * 2017-12-18 2018-06-29 孙颖 Semantic segmentation method based on multiple dimensioned convolutional neural networks
CN108319717A (en) * 2018-02-11 2018-07-24 福州大学 The computational methods of optimum lattice size range when a kind of vector point is converted to grid point
CN108561127A (en) * 2018-03-26 2018-09-21 上海电力学院 A kind of Formation pressure prediction method based on stochastic simulation
CN109063577A (en) * 2018-07-05 2018-12-21 浙江大学 Method is determined based on the satellite image segmentation best segmental scale of information gain-ratio
CN109583516A (en) * 2018-12-24 2019-04-05 天津珞雍空间信息研究院有限公司 A kind of space and time continuous PM2.5 inversion method based on ground and moonscope
CN110097026A (en) * 2019-05-13 2019-08-06 北京邮电大学 A kind of paragraph correlation rule evaluation method based on multidimensional element Video segmentation
CN110197135A (en) * 2019-05-13 2019-09-03 北京邮电大学 A kind of video structural method based on multidimensional segmentation
CN117496309A (en) * 2024-01-03 2024-02-02 华中科技大学 Building scene point cloud segmentation uncertainty evaluation method and system and electronic equipment

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060088207A1 (en) * 2004-10-22 2006-04-27 Henry Schneiderman Object recognizer and detector for two-dimensional images using bayesian network based classifier
CN104881677A (en) * 2015-05-08 2015-09-02 北京师范大学 Optimum segmentation dimension determining method for remotely-sensed image land cover classification
CN104915676A (en) * 2015-05-19 2015-09-16 西安电子科技大学 Deep-level feature learning and watershed-based synthetic aperture radar (SAR) image classification method

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060088207A1 (en) * 2004-10-22 2006-04-27 Henry Schneiderman Object recognizer and detector for two-dimensional images using bayesian network based classifier
CN104881677A (en) * 2015-05-08 2015-09-02 北京师范大学 Optimum segmentation dimension determining method for remotely-sensed image land cover classification
CN104915676A (en) * 2015-05-19 2015-09-16 西安电子科技大学 Deep-level feature learning and watershed-based synthetic aperture radar (SAR) image classification method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
陈杰等: "基于最优尺度选择的高分辨率遥感影像丘陵农田提取", 《农业工程学报》 *

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108053406A (en) * 2017-09-11 2018-05-18 同济大学 Ground mulching drafting method and device based on multiple-spatial resolution remote sensing image
CN108053406B (en) * 2017-09-11 2021-05-11 同济大学 Surface coverage mapping method and device based on multi-resolution remote sensing image
CN108230329A (en) * 2017-12-18 2018-06-29 孙颖 Semantic segmentation method based on multiple dimensioned convolutional neural networks
CN108230329B (en) * 2017-12-18 2021-09-21 孙颖 Semantic segmentation method based on multi-scale convolution neural network
CN108319717B (en) * 2018-02-11 2019-10-25 福州大学 The calculation method of optimum lattice size range when a kind of vector point is converted to grid point
CN108319717A (en) * 2018-02-11 2018-07-24 福州大学 The computational methods of optimum lattice size range when a kind of vector point is converted to grid point
CN108561127A (en) * 2018-03-26 2018-09-21 上海电力学院 A kind of Formation pressure prediction method based on stochastic simulation
CN108561127B (en) * 2018-03-26 2022-04-01 上海电力学院 Stratum pressure prediction method based on random simulation
CN109063577A (en) * 2018-07-05 2018-12-21 浙江大学 Method is determined based on the satellite image segmentation best segmental scale of information gain-ratio
CN109063577B (en) * 2018-07-05 2021-08-31 浙江大学 Satellite image segmentation optimal segmentation scale determination method based on information gain rate
CN109583516A (en) * 2018-12-24 2019-04-05 天津珞雍空间信息研究院有限公司 A kind of space and time continuous PM2.5 inversion method based on ground and moonscope
CN110097026A (en) * 2019-05-13 2019-08-06 北京邮电大学 A kind of paragraph correlation rule evaluation method based on multidimensional element Video segmentation
CN110197135A (en) * 2019-05-13 2019-09-03 北京邮电大学 A kind of video structural method based on multidimensional segmentation
CN110197135B (en) * 2019-05-13 2021-01-08 北京邮电大学 Video structuring method based on multi-dimensional segmentation
CN117496309A (en) * 2024-01-03 2024-02-02 华中科技大学 Building scene point cloud segmentation uncertainty evaluation method and system and electronic equipment
CN117496309B (en) * 2024-01-03 2024-03-26 华中科技大学 Building scene point cloud segmentation uncertainty evaluation method and system and electronic equipment

Similar Documents

Publication Publication Date Title
CN106780503A (en) Remote sensing images optimum segmentation yardstick based on posterior probability information entropy determines method
CN109829399B (en) Vehicle-mounted road scene point cloud automatic classification method based on deep learning
CN101877007B (en) Remote sensing image retrieval method with integration of spatial direction relation semanteme
CN102005034B (en) Remote sensing image segmentation method based on region clustering
CN108830870B (en) Satellite image high-precision farmland boundary extraction method based on multi-scale structure learning
CN107067405B (en) Remote sensing image segmentation method based on scale optimization
CN103413151B (en) Hyperspectral image classification method based on figure canonical low-rank representation Dimensionality Reduction
CN106503739A (en) The target in hyperspectral remotely sensed image svm classifier method and system of combined spectral and textural characteristics
US9064151B2 (en) Device and method for detecting plantation rows
CN102521624B (en) Classification method for land use types and system
CN104599275A (en) Understanding method of non-parametric RGB-D scene based on probabilistic graphical model
CN104182767B (en) The hyperspectral image classification method that Active Learning and neighborhood information are combined
CN101599120A (en) A kind of identification method of remote sensing image building
CN105069796B (en) SAR image segmentation method based on small echo both scatternets
CN113223042B (en) Intelligent acquisition method and equipment for remote sensing image deep learning sample
CN102073867B (en) Sorting method and device for remote sensing images
CN103208011A (en) Hyperspectral image space-spectral domain classification method based on mean value drifting and group sparse coding
CN115965812B (en) Evaluation method for classification of unmanned aerial vehicle images on wetland vegetation species and land features
CN106228136A (en) Panorama streetscape method for secret protection based on converging channels feature
CN114821102A (en) Intensive citrus quantity detection method, equipment, storage medium and device
CN112906537A (en) Crop identification method and system based on convolutional neural network
CN109740504B (en) Method for extracting sea area resources based on remote sensing image
CN111985575B (en) Hyperspectral image classification method based on convolutional neural network
CN115719453B (en) Deep learning-based remote sensing extraction method for rice planting structure
CN110175638B (en) Raise dust source monitoring method

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20170531