CN106651861A - Method for automatically and optimally selecting remote sensing image segmentation parameters based on regional inconsistency evaluation - Google Patents

Abstract

The invention discloses a method for automatically and optimally selecting remote sensing image segmentation parameters based on regional inconsistency evaluation. A segmentation data set is constructed in turn based on five groups of equidistant scale parameters and fixed shape factors and compactness factors, then an ED2 array is obtained from the segmentation data set and a corresponding reference data set based on PSE-NSR-ED2 inconsistency segmentation parameter evaluation system, the distribution mode of the five values of the ED2 array changing along with the scale parameters is analyzed, constant iteration is performed through transformation of the five scale parameters until the minimum ED2 value of the bottom part of an oblique U-shaped ED2-SP curve is found, and the scale parameters corresponding to the minimum ED2 value are the optimal scale parameters; and finally constant iteration of the process is performed on the basis of different combinations of the shape factors and the compactness factors. Blindness of parameter selection can be avoided so that the uncertainty caused by a trial-and-error method and time consumption caused by a method of exhaustion can be solved, and the accuracy and the degree of automation of object-oriented remote sensing image processing and analysis can also be enhanced.

Description

The preferably method of Remote Sensing Image Segmentation parameter is evaluated automatically based on region inconsistency

Technical field

The invention belongs to the side such as space statistical analysis and pattern-recognition is learned on geoscience research field, more particularly to remote sensing ground Face, and in particular to the preferably method of Remote Sensing Image Segmentation parameter is evaluated automatically based on region inconsistency.

Background technology

In remote sensing fields, for different application and purpose, different image segmentation algorithms are constantly suggested, and multiresolution The appearance of segmentation (Multiresolution Segmentation) algorithm a, it is considered to be mileage of Remote Sensing Image Segmentation Upright stone tablet.The algorithm combines the spectral information and spatial information of image in Image Segmentation, can produce internal homogeney highest shadow As object, the various combination that its major parameter has yardstick, form factor, the compactness factor, these parameters can produce different dividing Cut result.The process for selecting top-quality segmentation result is referred to as parameter preferably, and parameter preferably has to the problem for solving It is the concrete evaluation to splitting quality.Therefore, optimum segmentation parameter combination how is obtained to evaluate Image Segmentation quality be OBIA In a problem must solving.

Inconsistency segmentation result evaluation method is based on reference to polygon (Reference Polygon) and corresponding matching Split the inconsistency (Discrepancy) between polygon (Corresponding Polygon) to measure parameter current group The quality of the partition data produced by closing.It is a kind of objectively empirical evaluation method (Empirical Method), and it is With geometry inconsistency tolerance is to refer to polygon and the difference for matching area between polygon, and arithmetic inconsistency is measured Be both polygon quantity difference.

In inconsistency appraisement system (Potential Segmentation Error, PSE-Number-of- Segments Ratio, NSR-Euclidean Distance 2, ED2) in, PSE is potential segmentation area of error ratio, and NSR is Segmentation polygon quantity ratio, ED2 is the Euclidean distance of PSE and NSR.Oblique U-shaped (Euclidean Distance2, ED2- Scale Patterns, SP) pattern is based on to PSE-SP, what the analysis of NSR-SP, ED2-SP curve was proposed.In given shape In the case of shape, compactness parameter, Image Segmentation cell-average area is with the form that the increasing of scale parameter is in approximately power function Monotonic increase, the area of cutting unit is in approximately power function with the incremental of scale parameter.Correspondingly, the quantity of cutting unit is with chi Degree parameter is successively decreased approximately in power function.ED2 can be inclined as the combining form of PSE and NSR as the change of scale parameter is presented U-shaped curve form, as shown in Figure 1.

2nd, prior art

The method of segmentation quality evaluation has inconsistency method and goodness method.Based on inconsistency evaluation method by comparing ginseng Data set and partitioned data set are examined, comprehensive commenting is carried out to splitting quality in terms of geometry inconsistency and arithmetic inconsistency two Valency.Main inconsistency evaluation index has a series of segmentation quality evaluation index QR of the propositions such as Clinton etc., Weidner (Quality Rate), UR (Under-Segmentation Rate), OR (Over-Segmentation Rate) and ED1； Liu etc. proposes ED2 (Euclidean Distance 2) index based on geometry inconsistency and arithmetic inconsistency；And Yang Deng by analyzing ED1, ED2 serial evaluations index, ED3, ED3-Modified and SEI (Segmentation is further provided Evaluation Index).Meanwhile, Zhang etc. proposes F-measure indexs and MOA (Multiscale Object ) and BCA (Bidirectional Consistency Accuracy) Accuracy.Commented based on the partitioning parameters quality of goodness method What valency index was mainly built by local variance；The local variance for proposing Kim Deng (2010) is evaluated most Optimal sorting cuts the method for yardstick and is automated, and constructs the instrument ESP of optimal scale parameter selection；It is right Deng (2014) ESP instruments are improved.

By analyzing and summarizing optimal scale partitioning parameters system of selection, propose that yardstick evaluation should be with reference to shape, texture Etc. information, the automatic selection of multi-scale segmentation parameter is finally realized, but at present automatic mode mostly is non-supervisory segmentation, certain The partitioning parameters yardstick that there is selection in degree is bigger than normal, and over the ground the other specific aim of differentiating of species is not strong, and less divided phenomenon is obvious etc. Problem, this can bring detrimental effect to subsequent images classification.

The content of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art, there is provided one kind evaluates automatically excellent based on region inconsistency The method for selecting Remote Sensing Image Segmentation parameter combination.

Technical scheme is as follows：

The invention discloses a kind of evaluate automatically the preferably method of Remote Sensing Image Segmentation parameter, bag based on region inconsistency Include following steps：

Step one：Input remote sensing image to be split and reference data set, initiation parameter, described initiation parameter bag Include given initial gauges partitioning parameters interval value s₁And s₅, and s₅>s₁, give minimum step d between multi-scale segmentation parameter_min, give ED2_minMaximum L, form factor, the compactness factor and ED2_minMinimum ζ；

Step 2：If s₅-s₁>4d_min, then in initial gauges partitioning parameters value s₁And s₅On the basis of arrange 5 yardsticks Partitioning parameters and its step pitch, otherwise need to reset initial gauges partitioning parameters value s₁And s₅, d here>d_minFor constraints, s₁, s₂, s₃, s₄, s₅Five multi-scale segmentation parameters are specifically calculated as follows：

s₂=s₁+d

s₄=s₅-d

Step 3：According to s₁, s₂, s₃, s₄, s₅Respective partitioned data set and reference data set in multi-scale segmentation parameter, point Do not count it and refer to polygon quantity, segmentation polygon quantity, overlapping area, over-segmentation area, less divided area, calculate successively Go out the corresponding PSE of the corresponding inconsistency metric parameter of each multi-scale segmentation parameter_i, NSR_iAnd ED2_i, wherein i=2,3,4, 5}.Concrete calculation process is as follows：

In each multi-scale segmentation parameter, with R={ r_i:I=1,2 ..., m } represent that m refers to polygonal set, S ={ s_j:J=1,2 ..., n } the polygonal set of segmentation is represented, | r_i∩s_j| represent and refer to polygon r_iWith segmentation polygon s_j The area of intersection, | r_i| and | s_j| respectively with reference to polygon r_iWith segmentation polygon s_jArea, S'={ s_k:K=1, 2 ..., v } represent the set of the partitioned data set corresponding with reference data set.Two subsets that Sa and Sb is set S are defined, And meet matching criterior：

That is matching criterior is with reference to polygon and splits the area of polygon intersection at least with reference to polygon Shape matches polygonal half of the area, then with reference to the polygonal set S' of segmentation that polygon matches be just Sa and Sb's and Collection.∑ is defined simultaneously | R_i| the polygonal gross area is referred to for m, ∑ | S_k| it is many with reference to the segmentation that polygon matches with m The side shape gross area, ∑ | R_i∩S_k| for reference data set and matching partitioned data set overlapping area, | r_i-s_j|=| r_i|-|r_i∩ s_j| be matching polygon outside part with reference to polygonal area be over-segmentation area, | s_j-r_i|=| s_j|-|r_i∩s_j | it is that to match polygonal area be less divided area for part outside with reference to polygon.Therefore, PSE, NSR and ED2 can be with It is expressed as：

Finally obtain the ED2 minimum of a values and maximum calculated between 5 points：

ED2_min=min { ED2₁,ED2₂,ED2₃,ED2₄,ED2₅}

ED2_max=max { ED2₁,ED2₂,ED2₃,ED2₄,ED2₅}

Step 4:Pattern matching process

ED2 is with s for analysis₁～s₅The trend and dynamic five multi-scale segmentation parameters of adjustment of multi-scale segmentation Parameters variation, constantly Matching inconsistent evaluation model Case a～17 kinds of the Case q changing patteries of PSE-NSR-ED2, obtain in the solid shape factor and The optimal scale partitioning parameters of remote sensing image under the compactness factor.

Step 5：Successively according to form factor=0.1,0.2 ..., 0.9, the compactness factor=0.1,0.2 ..., 0.9 Combination carries out step one to the interative computation of step 4, obtains the optimal scale partitioning parameters in 81 parameter combinations；So This 81 groups of optimal scale partitioning parameters are sorted again afterwards, the corresponding multi-scale segmentation parameter of minimum ED2 groups therefrom selected, shape Three parameter combinations of the shape factor and the compactness factor are optimum segmentation parameter combination.

Preferably, described step four ED2 has Case a～Case q17 kinds with the trend of multi-scale segmentation Parameters variation Pattern：

Case a.ED2_iDo not change with multi-scale segmentation parameter, i.e. ED2₁=ED2₂=ED2₃=ED2₄=ED2₅, or Maximum therebetween, the difference of minimum of a value are less than default minimum ζ：

|ED2_max-ED2_min| ＜ ζ

If so and ED2_max>=L, then mean that default multi-scale segmentation parameter is substantially bigger than normal to ED2 value changes range of instability Domain, needs multi-scale segmentation parameter area to little value direction, by s₅Move left to s₁Position, s₁Move to left 4 step pitches, s₃For dynamic S after adjustment₁And s₅Value half, s₂And s₄Also 4 step pitches are accordingly moved to left：

s₅←s₁

s₁←s₁-4·d

s₂←s₁+d

s₄←s₅-d

Wherein " ← " represents the variable that the value of arrow right-hand member is assigned to arrow left end, and as multi-scale segmentation dynamic state of parameters is adjusted Whole process, the adjustment of five multi-scale segmentation parameters is synchronously to carry out, dividing without priority；S1～s5 on the right of arrow is tune Multi-scale segmentation parameter value before whole；With s₅←s₁As a example by；Represent former s₁Value be assigned to new s₅, that is, represent s₅Move left to Former s₁Position.

The adjustment of parameter is specifically according to the trend of multi-scale segmentation Parameters variation to s₁～s₅Dynamic adjustment, complete corresponding Multi-scale segmentation parameter " moving to left ", " moving to right ", " amplification " and " diminution " etc., all " ← " for mentioning represent this meaning in the present invention Think.

If s₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For S after dynamic adjustment₁And s₅Value 1/2nd, s₂And s₄Move to left 1 step pitch for newly obtaining：

s₁←5

s₅←s₄

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

Otherwise when meeting condition ED2_max<Expand multi-scale segmentation parameter search scope, corresponding s during L₃It is constant, s₁Reduce former The 2 times of step pitches come, s₅Expand 2 times of original step pitches, s₂And s₄Original s is replaced respectively₁And s₅)

s₂←s₁

s₄←s₅

s₁←s₂-2·d

s₅←s₄+2·d

If there is s in expansion₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value four/ One, s₃For s after dynamic adjustment₁And s₅Value 1/2nd, s₂And s₄Move to left 1 step pitch for newly obtaining：

s₁←5

s₅←s₄

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

If expanding the search of multi-scale segmentation parameter area, still there is ED2_iDo not change with scale parameter, or therebetween Situation of the difference of maximum, minimum of a value less than one of default minimum ζ：

|ED2_max-ED2_min| ＜ ζ

And ED2_max≤ L, then optimal scale partitioning parameters may occupy first, s_opt∈[s₁,s₅]；Need to further sentence It is disconnected：If in front and back there is above-mentioned condition in computing twice, computing ED twice before and after contrast_2min, take ED2_minLess once conduct The scope of optimal scale partitioning parameters, s_opt∈[s₁,s₅], then s_opt=s₅, end of run simultaneously reports result.

Case b.ED2_iSuccessively decrease with the increase of multi-scale segmentation parameter, i.e. ED2₁≥ED2₂≥ED2₃≥ED2₄≥ED2₅, then Optimal scale partitioning parameters are more than s₅, multi-scale segmentation parameter setting is adjusted, by s₁It is shifted to the right to s₃Position, s₂It is shifted to the right to s₄Position Put, s₃It is shifted to the right to s₅Position, s₂And s₄Also 2 step pitches are accordingly moved to right：

s₁←s₃

s₂←s₄

s₃←s₅

s₄←s₃+2·d

s₅←s₄+2·d

Return to step 3 continuation；

Case c.ED2_iIn s₄Place is presented minimum of a value flex point, i.e. ED2₁≥ED2₂≥ED2₃≥ED2₄And ED2₄≤ED2₅, Then optimal scale partitioning parameters are likely to be at s₃And s₅Between, by multi-scale segmentation parameter area to little value direction, then make s₁It is shifted to the right to s₂Position, s₂It is shifted to the right to s₃Position, s₃It is shifted to the right to s₄Position, s₄It is shifted to the right to s₅Position, s₅Move to right 1 step pitch：

s₁←s₂

s₂←s₃

s₃←s₄

s₄←s₅

s₅←s₄+d

Return to step 3 continuation；

Case d.ED2_iIn s₃Place is presented minimum of a value flex point, i.e. ED2₁≥ED2₂≥ED2₃, and ED2₃≤ED2₄≤ED2₅, Then optimal scale partitioning parameters are in s₂And s₄Between；And if then ED2₃>=L, then mean that default multi-scale segmentation parameter is obvious It is bigger than normal to ED2 value changes unstable regions, need multi-scale segmentation scope to little value direction, by s₅Move left to s₁Position, s₁It is left Move 4 step pitches, s₃For s after dynamic adjustment₅And s₁1/2nd, s₂And s₄Also 4 step pitches are accordingly moved to left：

s₅←s₁

s₁←s₁-4·d

s₂←s₁+d

s₄←s₂-d

If s1<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For S after dynamic adjustment₁And s₅Value 1/2nd, s₂And s₄Also 1 step pitch for newly obtaining accordingly is moved to left：

s₁←5

s₅←s₄

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

Otherwise ED2₃<L, and d ＞ d_min, then need in s₂And s₄Between encryption search, corresponding s₃It is constant, by s₁Move to left To s₂Position, s₅Move left to s₄Position, s₂And s₄Also 1/2nd step pitches are accordingly moved to left：

s₁←s₂

s₅←s₄

Until d≤d_min, and meet ED2₃<L, then s_opt=s₃。

Case e.ED2_iIn s₂Place is presented minimum of a value flex point, i.e. ED2₁≥ED2₂, and ED2₂≤ED2₃≤ED2₄≤ED2₄, Then optimal scale partitioning parameters are likely to be at s₁And s₃Between, if ED2₂<L, and d<d_min, then s₂For optimized parameter；Else if ED2₂>=L, then need multi-scale segmentation parameter area to little value direction, then make s₅Move left to s₄Position, s₄Move left to s₃Position Put, s₃Move left to s₂Position, s₂Move left to s₁Position, s₁Move to left 1 step pitch：

s₅←s₄

s₄←s₃

s₃←s₂

s₂←s₁

s₁←s₁-d

If s₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For S after dynamic adjustment₁And s₅Value 1/2nd, s₂And s₄Also 1 step pitch for newly obtaining accordingly is moved to left：

s₁←5

s₅←s₄

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

Otherwise work as ED2₂<L, and d>d_min, in s₁And s₃Place's encryption, then need in s₂And s₄Between encryption search, corresponding s₃ It is constant, by s₁Move left to s₂Position, s₅Move left to s₄Position, s₂And s₄Also 1/2nd step pitches are accordingly moved to left：

s₁←s₂

s₅←s₄

Return to step 3 continuation；

Case f.ED2_iIt is incremented by with multi-scale segmentation parameter, i.e. ED2₁≤ED2₂≤ED2₃≤ED2₄≤ED2₂, then optimal scale Partitioning parameters are less than s₁Or neighbouring s₁, then s is made₅Move left to s₃Position, s₃Move left to s₁Position, s₄Move left to s₂Position, s₂Move to left 2 step pitches, s₁Move to left 2 step pitches：

s₅←s₃

s₃←s₁

s₄←s₂

s₂←s₃-d

s₁←s₂-d

If s₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For S after dynamic adjustment₁And s₅Value 1/2nd, s₂And s₄Also 1 step pitch for newly obtaining accordingly is moved to left：

s₁←5

s₅←s₄

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

Case g-Case q. are incremented by with multi-scale segmentation parameter, ED2_iTend to unstable with the variation tendency of multi-scale segmentation parameter It is fixed, ED2i values unstable region that arrives with multi-scale segmentation Parameters variation, and ED2_min>=L, then mean default multi-scale segmentation parameter It is substantially bigger than normal to ED2_iValue changes unstable region, needs to move to left segmentation scale parameter scope to little value direction, by s₅Move left to s₁Position, s₁Move to left 4 step pitches, s₃For s after dynamic adjustment₁And s₅Value half, s₂And s₄Also 4 step pitches are accordingly moved to left：

s₅←s₁

s₁←s₁-4·d

s₂←s₁+d

s₄←s₂-d

If s₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For S after dynamic adjustment₁And s₅Value 1/2nd, s₂And s₄Also accordingly move to left 1 and newly obtain step pitch：

s₁←5

s₅←s₄

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation:Otherwise run into ED2_maxDuring≤L, this group of data of acquisition will be calculated after amplification with upper one group The ED2 of data_minIt is compared, if there is continuing to amplify less than the situation of upper one group of minimum of a value, until finding minimum ED2_min, amplification terminates, it is determined that minimum ED2_minInterval, be then encrypted in smallest interval, be encrypted into d≤d_min, encryption knot Beam, takes its ED2_minCorresponding multi-scale segmentation parameter is optimal scale partitioning parameters.

Preferably, minimum step d between multi-scale segmentation parameter is given in described step one_min=1, give ED2_minMaximum L=1.0, form factor=0.1, the compactness factor=0.1, ED2_minMinimum ζ=0.0001.

The present invention ground spatial analysis and pattern match are used in particular problem, it is determined that optimum segmentation parameter can land used Spatial analysis is learned and based on geometry inconsistency and arithmetic inconsistency definite principle ground be further explained, with higher theory Confidence level.

The inventive method, is not limited to Remote Sensing Image Segmentation parameter optimization algorithm, can be widely applied to the oblique U of any obedience The model of type curve distribution asks the problem of optimal value, the method to have wide applicability and using value.

The prediction of partitioning parameters before the inventive method, substantially remote sensing images analysis, it is to avoid selection parameter it is blind Mesh, solves uncertainty that trial-and-error method brings and that the method for exhaustion is brought is time-consuming, while having taken into account operating efficiency, also improves Object-oriented remote sensing image processes the precision and automaticity with analysis.

Description of the drawings

Fig. 1 is given form factor and the oblique U-shaped ED2-SP pattern diagrams of the compactness factor；

Fig. 2 is 17 kind basic model schematic diagrames of the ED2 with multi-scale segmentation Parameters variation；

Fig. 3 evaluates automatically the preferably method of Remote Sensing Image Segmentation parameter for the embodiment of the present invention based on region inconsistency Procedure Procedure figure；

Fig. 4 is the track pair that the embodiment of the present invention and the method for exhaustion select the automatic search of remote sensing image optimum segmentation parameter combination Than figure；

Fig. 5 applies the segmentation in the Quick bird high-resolution remote sensing images of Dongguan City, Guangdong Province for the embodiment of the present invention Design sketch；

Fig. 6 is applied in Hui Nationality in Ningxia Hui Nationality Autonomy autonomous prefecture Zhongwei City World view2 high-definition remote sensing shadows for the embodiment of the present invention Segmentation effect figure as in.

Specific embodiment

As shown in figure 3, evaluating automatically preferably Remote Sensing Image Segmentation ginseng based on region inconsistency for the embodiment of the present invention Several method flow flow charts；In all embodiments of the invention, step one gives minimum step d between multi-scale segmentation parameter_min =1, give ED2_minMaximum L=1.0, form factor=0.1, the compactness factor=0.1, ED2_minMinimum ζ=0.0001.

It is illustrated in figure 2 17 kind basic model schematic diagrames of the ED2 with multi-scale segmentation Parameters variation.

1. the method for the present invention and the method for exhaustion select the contrast of optimum segmentation parameter combination

The embodiment of the present invention is applied into Dongguan City, Guangdong Province and Hui Nationality in Ningxia Hui Nationality Autonomy autonomous prefecture Zhongwei City three kinds of sensors (Quick bird, Alos, World view2) it is multispectral with fusion 6 scape high-resolution remote sensing images in 3 kinds of different atural objects The selection of (arable land FL, residential building RB, swag WB) optimum segmentation parameter combination, and compared with the result of method of exhaustion acquisition Compared with, tested with the computer for similarly configuring in experiment, initial gauges parameter area is unified, it is ensured that two methods have There is comparativity, obtain experimental result, wherein as shown in table 1, method of the present invention result is as shown in table 2 for method of exhaustion experimental result.

Contrast Tables 1 and 2, it can be found that the embodiment of the present invention substantially can under identical initial gauges parameter area The partitioned data set compared with small data magnitude is produced, the spent time is also relatively also shorter, and step-length can be accurate to 1, obtained The value of more accurate ED2 is determining the optimum segmentation parameter combination of remote sensing image, but the currently a popular method of exhaustion needs The partitioned data set that step-length specifies mechanically is produced, step-length is once it is determined that dynamic can not adjust, the initial gauges parameter for finding Scope, as initial gauges parameter area considers not comprehensively, not being enumerated to the scope of optimum segmentation parameter combination presence, Specific restriction is brought to parameter is found.The contrast tested by this, embodiment of the present invention performance is good, not only can be with Optimal scale partitioning parameters are quickly obtained by arranging suitable initial gauges parameter area, and step can be changed with dynamic It is long, can precisely predict that PSE-NSR-ED2 inconsistencies evaluation model obtains the interval of optimal scale partitioning parameters, save the time While avoid select optimum segmentation parameter combination blindness.

The method of exhaustion result statistical form of table 1

The inventive method result statistical form of table 2

Obtain from Tables 1 and 2 when one group of form factor and the compactness factor are extracted in substantial amounts of data being all 0.1, select The track of optimal scale partitioning parameters is analyzed, and can be clearly seen that in the embodiment of the present invention can split in initial gauges Then fast prediction starts encryption and determines optimum to the interval of the presence of optimal scale partitioning parameters in the range of parameter, and poor Act method does not have any acquisition track, needs the artificial judgment tiltedly optimum position of U-shaped curve, as shown in Figure 4.

2. the method for the present invention and other automatic segmentation parameter preferred kit contrasts

By the embodiment of the present invention withThe instrument ESP of the automatic preferred remote sensing image scale parameter proposed Deng (2014) Contrasted, experimental data be Dongguan City, Guangdong Province and Hui Nationality in Ningxia Hui Nationality Autonomy autonomous prefecture Zhongwei City three kinds of sensors (Quick bird, Alos, World view2) 3 kinds of different atural objects (arable land, forest land, swag) in 3 scape high-resolution remote sensing images, because ESP can only Optimal scale partitioning parameters are selected, using the form factor found in the embodiment of the present invention and the compactness factor as standard, and is set Putting identical initial gauges parameter and step-length carries out contrast experiment.According to the regulation in segmentation precision evaluation index, producer's essence Degree and user's precision all close 1, the optimum principle of segmentation effect weighs the result in table 3, it can be found that user's essence of ESP instruments Degree is very unstable, minimum to reach 0.11, and the optimal scale partitioning parameters for finding are substantially bigger than normal, and partitioned data set magnitude is very Greatly, whole efficiency is poor.

The optimum segmentation parametric results of table 3 and precision evaluation contrast statistical form

3. Fig. 5 is the embodiment of the present invention and the optimum segmentation parameter combination of method of exhaustion acquisition to Dongguan City, Guangdong Province Quick Segmentation effect of the bird high-resolution remote sensing images to water body, it can be seen that can well divide the region of water body, too Cut and less divided phenomenon is all solved well；Fig. 6 is the embodiment of the present invention and the optimum segmentation parameter combination of method of exhaustion acquisition to peaceful Segmentation effect of summer autonomous prefecture of the Hui ethnic group Zhongwei City World view2 high-resolution remote sensing images to arable land, hence it is evident that it can be seen that this Inventive embodiments can predict the optimal scale partitioning parameters scope of such atural object than larger, but method of exhaustion ignorant of the economics Mesh thinks initial gauges parameter area it has been estimated that maximum magnitude, and so artificial interpretation has a serious subjectivity.

Claims (3)

1. a kind of method that automatically preferably Remote Sensing Image Segmentation parameter is evaluated based on region inconsistency, it is characterised in that include as Lower step：

Step one：Input remote sensing image to be split and reference data set, initiation parameter, described initiation parameter include to Determine initial gauges partitioning parameters interval value s₁And s₅, and s₅>s₁, give minimum step d between multi-scale segmentation parameter_min, give ED2_min Maximum L, form factor, the compactness factor and ED2_minMinimum ζ；Step 2：If s₅-s₁>4d_min, then in initial gauges point Cut parameter value s₁And s₅On the basis of 5 multi-scale segmentation parameters and its step pitch are set, otherwise need to reset initial gauges segmentation ginseng Numerical value s₁And s₅, d here>d_minFor constraints, s₁, s₂, s₃, s₄, s₅Five multi-scale segmentation parameters are specifically calculated as follows：

$d=\frac{1}{4}(s_5-s_1)$

$s_3=\frac{1}{2}(s_1+s_5)$

s₂=s₁+d

s₄=s₅-d

Step 3：According to s₁, s₂, s₃, s₄, s₅Respective partitioned data set and reference data set, unite respectively in multi-scale segmentation parameter Count it and refer to polygon quantity, segmentation polygon quantity, overlapping area, over-segmentation area, less divided area, calculate successively it is every The corresponding PSE of the corresponding inconsistency metric parameter of one multi-scale segmentation parameter_i, NSR_iAnd ED2_i, wherein i={ 2,3,4,5 }. Concrete calculation process is as follows：

In each multi-scale segmentation parameter, with R={ r_i:I=1,2 ..., m } represent that m refers to polygonal set, S={ s_j: J=1,2 ..., n } the polygonal set of segmentation is represented, | r_i∩s_j| represent and refer to polygon r_iWith segmentation polygon s_jCross-shaped portion The area for dividing, | r_i| and | s_j| respectively with reference to polygon r_iWith segmentation polygon s_jArea, S'={ s_k:K=1,2 ..., v } Represent the set of the partitioned data set corresponding with reference data set.Define two subsets that Sa and Sb is set S, and satisfaction With criterion：

$Sa=\{s_j:\frac{\left|r_i\&cap;s_j\right|}{\left|s_j\right|}>0.5\}$

$Sb=\{s_j:\frac{\left|r_i\&cap;s_j\right|}{\left|r_i\right|}>0.5\}$

That is matching criterior be with reference to polygon and segmentation polygon intersection area at least with reference to polygon or Polygonal half of the area is matched, then the polygonal set S' of segmentation for matching with reference to polygon is just the union of Sa and Sb. ∑ is defined simultaneously | R_i| the polygonal gross area is referred to for m, ∑ | S_k| it is polygon with reference to the segmentation that polygon matches with m The shape gross area, ∑ | R_i∩S_k| for reference data set and matching partitioned data set overlapping area, | r_i-s_j|=| r_i|-|r_i∩s_j | be matching polygon outside part with reference to polygonal area be over-segmentation area, | s_j-r_i|=| s_j|-|r_i∩s_j| It is that to match polygonal area be less divided area for part outside with reference to polygon.Therefore, PSE, NSR and ED2 can be with It is expressed as：

$PSE=\frac{\mathrm{\&Sigma;}\left|S_k\right|-\mathrm{\&Sigma;}\left|S_k\&cap;R_i\right|}{\mathrm{\&Sigma;}\left|R_i\right|}$

$NSR=\frac{abs(m-n)}{m}$

$ED2=\sqrt{{\mathrm{PSE}}^2+{\mathrm{NSR}}^2}$

Finally obtain the ED2 minimum of a values and maximum calculated between 5 points：

ED2_min=min { ED2₁,ED2₂,ED2₃,ED2₄,ED2₅}

ED2_max=max { ED2₁,ED2₂,ED2₃,ED2₄,ED2₅}

Step 4:Pattern matching process

ED2 is with s for analysis₁～s₅The trend and dynamic five multi-scale segmentation parameters of adjustment of multi-scale segmentation Parameters variation, constantly matches The inconsistent evaluation model Case a of PSE-NSR-ED2～17 kinds of Case q changing patteries, obtain in the solid shape factor and compact The optimal scale partitioning parameters of remote sensing image under the degree factor.

Step 5：Successively according to form factor=0.1,0.2 ..., 0.9, the compactness factor=0.1,0.2 ..., 0.9 combination Mode carries out step one to the interative computation of step 4, obtains the optimal scale partitioning parameters in 81 parameter combinations；Then it is right This 81 groups of optimal scale partitioning parameters are sorted again, the corresponding multi-scale segmentation parameter of minimum ED2 groups therefrom selected, shape because Son and three parameter combinations of the compactness factor are optimum segmentation parameter combination.

2. method according to claim 1, it is characterised in that described step four ED2 becoming with multi-scale segmentation Parameters variation Gesture has Case a～17 kinds of Case q patterns：

Case a.ED2_iDo not change with multi-scale segmentation parameter, i.e. ED2₁=ED2₂=ED2₃=ED2₄=ED2₅, or therebetween Maximum, the difference of minimum of a value be less than default minimum ζ：

|ED2_max-ED2_min| ＜ ζ

If so and ED2_max>=L, then mean that default multi-scale segmentation parameter is substantially bigger than normal to ED2 value changes unstable regions, needs By multi-scale segmentation parameter area to little value direction, by s₅Move left to s₁Position, s₁Move to left 4 step pitches, s₃After dynamic adjustment s₁And s₅Value half, s₂And s₄Also 4 step pitches are accordingly moved to left：

s₅←s₁

s₁←s₁-4·d

$s_3\&LeftArrow;\frac{1}{2}(s_1+s_5)$

s₂←s₁+d

s₄←s₅-d

If s₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For dynamic tune Whole rear s₁And s₅Value 1/2nd, s₂And s₄Move to left 1 step pitch for newly obtaining：

s₁←5

s₅←s₄

$d\&LeftArrow;\frac{1}{4}(s_5-s_1)$

$s_3\&LeftArrow;\frac{1}{2}(s_5+s_1)$

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

Otherwise when meeting condition ED2_max<Expand multi-scale segmentation parameter search scope, corresponding s during L₃It is constant, s₁Reduce original 2 Times step pitch, s₅Expand 2 times of original step pitches, s₂And s₄Original s is replaced respectively₁And s₅)

s₂←s₁

s₄←s₅

s₁←s₂-2·d

s₅←s₄+2·d

If there is s in expansion₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃ For s after dynamic adjustment₁And s₅Value 1/2nd, s₂And s₄Move to left 1 step pitch for newly obtaining：

s₁←5

s₅←s₄

$d\&LeftArrow;\frac{1}{4}(s_5-s_1)$

$s_3\&LeftArrow;\frac{1}{2}(s_5+s_1)$

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

If expanding the search of multi-scale segmentation parameter area, still there is ED2_iDo not change with scale parameter, or maximum therebetween, Situation of the difference of minimum of a value less than one of default minimum ζ：

|ED2_max-ED2_min| ＜ ζ

And ED2_max≤ L, then optimal scale partitioning parameters may occupy first, s_opt∈[s₁,s₅]；Further judgement need to be made：Such as There is above-mentioned condition in computing twice before and after fruit, computing ED twice before and after contrast_2min, take ED2_mIn is less once as optimum The scope of multi-scale segmentation parameter, s_opt∈[s₁,s₅], then s_opt=s₅, end of run simultaneously reports result.

Case b.ED2_iSuccessively decrease with the increase of multi-scale segmentation parameter, i.e. ED2₁≥

ED2₂≥ED2₃≥ED2₄≥ED2₅, then optimal scale partitioning parameters are more than s₅, multi-scale segmentation parameter setting is adjusted, by s₁It is right Move on to s₃Position, s₂It is shifted to the right to s₄Position, s₃It is shifted to the right to s₅Position, s₂And s₄Also 2 step pitches are accordingly moved to right：

s₁←s₃

s₂←s₄

s₃←s₅

s₄←s₃+2·d

s₅←s₄+2·d

Return to step 3 continuation；

Case c.ED2_iIn s₄Place is presented minimum of a value flex point, i.e. ED2₁≥ED2₂≥ED2₃≥ED2₄And ED2₄≤ED2₅, then most Excellent multi-scale segmentation parameter is likely to be at s₃And s₅Between, by multi-scale segmentation parameter area to little value direction, then make s₁It is shifted to the right to s₂'s Position, s₂It is shifted to the right to s₃Position, s₃It is shifted to the right to s₄Position, s₄It is shifted to the right to s₅Position, s₅Move to right 1 step pitch：

s₁←s₂

s₂←s₃

s₃←s₄

s₄←s₅

s₅←s₄+d

Return to step 3 continuation；

Case d.ED2_iIn s₃Place is presented minimum of a value flex point, i.e. ED2₁≥ED2₂≥ED2₃, and ED2₃≤ED2₄≤ED2₅, then most Excellent multi-scale segmentation parameter is in s₂And s₄Between；And if then ED2₃>=L, then mean that default multi-scale segmentation parameter is substantially bigger than normal To ED2 value changes unstable regions, need multi-scale segmentation scope to little value direction, by s₅Move left to s₁Position, s₁Move to left 4 Individual step pitch, s₃For s after dynamic adjustment₅And s₁1/2nd, s₂And s₄Also 4 step pitches are accordingly moved to left：

s₅←s₁

s₁←s₁-4·d

$s_3\&LeftArrow;\frac{1}{2}(s_1+s_2)$

s₂←s₁+d

s₄←s₂-d

If s1<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For dynamic S after adjustment₁And s₅Value 1/2nd, s₂And s₄Also 1 step pitch for newly obtaining accordingly is moved to left：

s₁←5

s₅←s₄

$d\&LeftArrow;\frac{1}{4}(s_5-s_1)$

$s_3\&LeftArrow;\frac{1}{2}(s_5+s_1)$

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

Otherwise ED2₃<L, and d ＞ d_min, then need in s₂And s₄Between encryption search, corresponding s₃It is constant, by s₁Move left to s₂ Position, s₅Move left to s₄Position, s₂And s₄Also 1/2nd step pitches are accordingly moved to left：

s₁←s₂

s₅←s₄

$s_2\&LeftArrow;s_1+\frac{1}{2}\&CenterDot;d$

$s_4\&LeftArrow;s_5-\frac{1}{2}\&CenterDot;d$

Until d≤d_min, and meet ED2₃<L, then s_opt=s₃。

Case e.ED2_iIn s₂Place is presented minimum of a value flex point, i.e. ED2₁≥ED2₂, and ED2₂≤ED2₃≤ED2₄≤ED2₄, then most Excellent multi-scale segmentation parameter is likely to be at s₁And s₃Between, if ED2₂<L, and d<d_min, then s₂For optimized parameter；Else if ED2₂≥ L, then need multi-scale segmentation parameter area to little value direction, then make s₅Move left to s₄Position, s₄Move left to s₃Position, s₃It is left Move on to s₂Position, s₂Move left to s₁Position, s₁Move to left 1 step pitch：

s₅←s₄

s₄←s₃

s₃←s₂

s₂←s₁

s₁←s₁-d

If s₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For dynamic tune Whole rear s₁And s₅Value 1/2nd, s₂And s₄Also 1 step pitch for newly obtaining accordingly is moved to left：

s₁←5

s₅←s₄

$d\&LeftArrow;\frac{1}{4}(s_5-s_1)$

$s_3\&LeftArrow;\frac{1}{2}(s_5+s_1)$

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

Otherwise work as ED2₂<L, and d>d_min, in s₁And s₃Place's encryption, then need in s₂And s₄Between encryption search, corresponding s₃No Become, by s₁Move left to s₂Position, s₅Move left to s₄Position, s₂And s₄Also 1/2nd step pitches are accordingly moved to left：

s₁←s₂

s₅←s₄

$s_2\&LeftArrow;s_1+\frac{1}{2}\&CenterDot;d$

$s_5\&LeftArrow;s_2-\frac{1}{2}\&CenterDot;d$

Return to step 3 continuation；

Case f.ED2_iIt is incremented by with multi-scale segmentation parameter, i.e. ED2₁≤ED2₂≤ED2₃≤ED2₄≤ED2₂, then optimal scale segmentation Parameter is less than s₁Or neighbouring s₁, then s is made₅Move left to s₃Position, s₃Move left to s₁Position, s₄Move left to s₂Position, s₂It is left Move 2 step pitches, s₁Move to left 2 step pitches：

s₅←s₃

s₃←s₁

s₄←s₂

s₂←s₃-d

s₁←s₂-d

If s₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For dynamic tune Whole rear s₁And s₅Value 1/2nd, s₂And s₄Also 1 step pitch for newly obtaining accordingly is moved to left：

s₁←5

s₅←s₄

$d\&LeftArrow;\frac{1}{4}(s_5-s_1)$

$s_3\&LeftArrow;\frac{1}{2}(s_5+s_1)$

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation；

Case g-Case q. are incremented by with multi-scale segmentation parameter, ED2_iTend to unstable with the variation tendency of multi-scale segmentation parameter, ED2i values unstable region that arrives with multi-scale segmentation Parameters variation, and ED2_min>=L, then mean that default multi-scale segmentation parameter is bright It is aobvious bigger than normal to ED2_iValue changes unstable region, needs to move to left segmentation scale parameter scope to little value direction, by s₅Move left to s₁ Position, s₁Move to left 4 step pitches, s₃For s after dynamic adjustment₁And s₅Value half, s₂And s₄Also 4 step pitches are accordingly moved to left：

s₅←s₁

s₁←s₁-4·d

$s_3\&LeftArrow;\frac{1}{2}(s_1+s_2)$

s₂←s₁+d

s₄←s₂-d

If s₁<0, then make s₁=5, s₅Move to left s₄Position, d for dynamic adjustment after s₁And s₅Value a quarter, s₃For dynamic tune Whole rear s₁And s₅Value 1/2nd, s₂And s₄Also accordingly move to left 1 and newly obtain step pitch：

s₁←5

s₅←s₄

$d\&LeftArrow;\frac{1}{4}(s_5-s_1)$

$s_3\&LeftArrow;\frac{1}{2}(s_5+s_1)$

s₂←s₁+d

s₄←s₅-d

Return to step 3 continuation:Otherwise run into ED2_maxDuring≤L, this group of data and upper one group of data of acquisition will be calculated after amplification ED2_minIt is compared, if there is continuing to amplify less than the situation of upper one group of minimum of a value, until finding minimum ED2_min, Amplification terminates, it is determined that minimum ED2_minInterval, be then encrypted in smallest interval, be encrypted into d≤d_min, encryption terminates, takes Its ED2_minCorresponding multi-scale segmentation parameter is optimal scale partitioning parameters.

3. method according to claim 1, it is characterised in that give in described step one minimum between multi-scale segmentation parameter Step pitch d_min=1, give ED2_minMaximum L=1.0, form factor=0.1, the compactness factor=0.1, ED2_minMinimum ζ= 0.0001。