CN1252588C - High spectrum remote sensing image combined weighting random sorting method - Google Patents

Abstract

The present invention relates to a random classifying method for combined weighting of hyperspectral remote sensing images. The random classifying method gives a certain weighting coefficient through development language provided by a remote sensing image processing developing platform ENVI/IDL, and aiming at the sizes of hyperspectral image files and the processing speed of a computer, roughly classifies the hyperspectral image files with the improved self-adapting minimum distance method, then takes an error criterion sum as a criterion function, completes fine classification according to another weight coefficient and with cluster repeated iteration, and finally splices the classifying results of the divided images to obtain the classified images of the whole hyperspectral remote sensing images. The method of the present invention can raise the total classification accuracy of the obtained result by about 10% no matter according to visual explanation or quantitative evaluation, remarkably raise the computer processing speed, widely serve for classifying hyperspectral remote sensing image data, and obtain more reliable remote sensing classification images under the condition of raising the processing speed.

Description

High-spectrum remote sensing associating Weighted random sorting technique

Technical field:

The present invention relates to a kind of high-spectrum remote sensing associating Weighted random sorting technique, pretreated high-spectrum remote sensing is carried out the high precision Fast Classification, be a core technology of classification of remote-sensing images decision-making, in fields such as all kinds of military or civilian Remote Sensing Information Processing System, digital city space information system, all can be widely used.

Background technology:

A plurality of target areas contained in the areal one or more remote sensing images separately just are called the classification of remote sensing images, or categorised decision merges, its target can be summarized as substantially under the condition few or no specimen of trying one's best, and can try one's best and discern type of ground objects exactly.This technology has important effect for aspects such as area planning, face of land ecosystem researchs.For example, certain is specified ground object target, the information characteristics of its remote sensing images is not unalterable, and the variation with season and meteorological condition is not only arranged, and the variation of human factor is also arranged.Like this, select the processing of classifying of the high spectrum image in area interested, again to not simultaneously the classification results of phase compare, whether variation has taken place with regard to the decidable characters of ground object, realize environmental monitoring.In order to reach the classification results of expection, at first determine class categories according to application purpose and picture characteristics, find out the characteristic quantity of these classifications then and extract the training data of each classification, determine the classification benchmark, use this benchmark at last each pixel is sorted out.Simultaneously known training data, class categories and classification results are done contrast, determine result's reliability.

Classification mechanism to remote sensing images is concluded, and can think that it comprises two basic requirements: the speed of classification and nicety of grading.The former is under the condition that improves the computer hardware physical property, and algorithm design reduces manual intervention as far as possible.Because reduce any automatic processing procedure of manual intervention, all comprising the intension of rapidity.The latter then need consider every possible angle when algorithm design, and combining geographic information system (GIS) waits other hi-tech simultaneously.Such as, utilize the spatial data of GIS just can improve the nicety of grading of remote sensing images.Realize its Target Recognition if reflect the difference of various atural objects indirectly by the image spectrum characteristic, " the different spectrum of jljl " and " foreign matter is with spectrum " phenomenon can usually take place, the precision of reduction classification.And introduce the GIS spatial data, and, just can improve the confidence level of classification with its supplementary as the image classification processing.

At present, classification methods of remote sensing images is a lot, and sorter commonly used has maximum likelihood method, hierarchial-cluster analysis, multistage cutting sorter, decision tree classification device, minimum distance method, the charting of spectrum angle, fuzzy classification device and the neural network classifier etc. based on bayesian criterion.Generally speaking, the approach of these classifier design is on the basis of random theory, carries out in conjunction with other application technology, and can be divided into non-supervision and two kinds of fundamental types of supervised classification.Wherein, unsupervised classification does not need to select known zone to train, and is simple, but the result is not necessarily best; Supervised classification will have known zone to train in advance, is used for whole study area then, and the result who obtains like this explains well, and the precision of classification is also high.

At present, along with improving constantly of remote sensor performance, high spectrum, high-resolution remote sensing image become research and divide one of the most frequently used object of time-like, and it has very high image definition and the high characteristics of spectral resolution.Therefore, during processing, be not enough only with a kind of sorting algorithm, it can not satisfy the nicety of grading requirement to image.In addition, because the size of high spectrum image is generally the order of magnitude of hundreds of million, when existing these methods of application were carried out the image classification decision-making, processing speed also was the problem that a palpus solves.Tracing it to its cause, is because the limitation of existing method itself causes, so need the new algorithm of research.

Summary of the invention:

The objective of the invention is to above-mentioned deficiency at prior art, a kind of high-spectrum remote sensing associating Weighted random sorting technique is provided, not only can improves nicety of grading most effectively, and can also handle by the piecemeal of video data block, the operational speed of a computer is accelerated, satisfied application demand.

For realizing such purpose, in the technical scheme of the present invention, the development language that utilizes remote sensing image processing and development platform ENVI/IDL to provide, file size and computer processing speed at high spectrum image, provide certain weight coefficient, it is carried out rough sort with improved self-adaptation minimum distance method.Then, according to another weighting coefficient, with error criterion be criterion function, utilize iterating of cluster to finish smart classification.Last basis is spliced the classification results of block image, obtains the classified image of view picture high-spectrum remote sensing.

Method of the present invention comprises following concrete steps:

1. the extraction of high spectrum image data characteristic: according to demands of applications, high spectrum image file to be classified behind the acquisition registration, the interface function that provides by ENVI/IDL remote sensing image processing and development platform, obtain their performance data, comprise file size, the wave band number of spatial resolution and image high-spectrum remote sensing etc.

2. the deblocking of high spectrum image: the spatial resolution with high spectrum image is a standard, api function by WINDOWS2000 operating system, automatically after obtaining indexs such as the internal memory of computing machine and CPU processing speed, it is divided into the data block (number is high spectrum image data volume/calculator memory amount) of several successive according to the big young pathbreaker of the data volume of high spectrum image file correspondence.And each data block is numbered.

3. corresponding high spectrum image data block is called in circulation, calculate the center and the radius of classified sample set according to formula, and remember classification number, then with all sample points as treating branch, carry out the rough sort first time by improved self-adaptation minimum distance method criterion function.

4. according to the rough sort result, in conjunction with original image, given smart nicety of grading is carried out the smart classification of the average data block sample of K-, during, must iterate, reach classificating requirement up to the nicety of grading of each sample set.

5. the splicing again of classified image data block is spliced the classification block of high spectrum image data block again by the aforesaid order of cutting apart, and forms new classified image.

Characteristic of the present invention is, proposes the secondary treating algorithm of high-spectrum remote-sensing rough sort and smart classification, to obtained result, no matter is visual interpretation or quantitative evaluation, and it can improve about 10% overall nicety of grading.Simultaneously,, each data block is called in calculator memory successively, can improve Computer Processing speed significantly owing to carried out deblocking.The method that the present invention proposes can be widely used in the high-spectrum remote sensing classification of Data and handles, and obtains remote sensing classified image more reliably under the condition that improves processing speed.

Description of drawings:

Fig. 1 is the The general frame of disposal route proposed by the invention.

As shown in Figure 1, the interface function that at first utilizes the ENVI/IDL development platform based on Windows to provide obtains the performance data of high spectrum and full-colour image.The high spectrum image performance data is carried out deblocking according to a certain percentage.The high spectrum image data block is called in circulation, each corresponding data block is carried out the rough sort of self-adaptation minimum distance method by the certain criterion function.According to this classification results, in conjunction with original image, given smart nicety of grading is implemented the smart classification of secondary of video data block.At last,, carry out connecing again of classification results, obtain final total classification image by original image segmentation order.

Fig. 2 is the syncretizing effect contrast of method and other method of the present invention's employing.

Wherein, the OMIS high-spectrum remote sensing of Fig. 2 (a), this example are got its wave band 117, file size 662 pixels * 686 pixels, spatial resolution 20m.Comprise multiple atural objects such as soil (farmland), buildings, greenery patches and water body in the image.The image classification result that Fig. 2 (b) utilizes unsupervised classification to carry out.Fig. 2 (c) is the classification results that utilizes maximum likelihood method, and Fig. 2 (d) is the classification results that the inventive method is carried out, and therefrom as can be seen, the visual effect of image that the present invention obtains obviously is better than other two kinds of methods.

Embodiment:

In order to understand technical scheme of the present invention better, embodiments of the present invention are further described below in conjunction with accompanying drawing.

A kind of The general frame that is used for high-spectrum remote sensing classification processing that Fig. 1 proposes for the present invention, the input of data file is the OMIS high-spectrum remote sensing of Fig. 2 (a), gets its wave band 117, file size 662 pixels * 686 pixels, spatial resolution 20m.Comprise multiple atural objects such as soil (farmland), buildings, greenery patches and water body in the image.The concrete implementation detail of each several part is as follows:

1. the extraction of data characteristic, the performance data that obtains high spectrum and full-colour image is utilized the ENVI_DISPLAY_BANDS function of ENVI/IDL development platform, can demonstrate size, wave band number and the spatial resolution etc. of file.

2. the deblocking of high spectrum image, utilize the C language interface of ENVI/IDLIDL to call the api function that WINDOWS2000 operating system provides, obtain the ardware feature of computing machine, behind CPU processing speed, memory size, carry out the piecemeal of high-spectral data file in the ratio of high spectrum file data amount/internal memory.

3. improve the rough sort of self-adaptation minimum distance method, the high spectrum image data block is called in circulation, calculates the center and the radius of classified sample set according to following calculating formula, and finishes classification number, utilizes the criterion function of minimum distance method to carry out the rough sort first time.

In each data block, establish training sample and be divided into N classification W ₁, W ₂..., W _n, the master sample of each classification is the center of this classification training sample in the training sample, promptly

$C_1=\frac{1}{m_i}\underset{j=1}{\overset{m_i}{\mathrm{\Σ}}}{X}_i---\left(1\right)$

M in the formula _i---W _iThe number of samples of class;

X _i---total sample number;

Point P to be classified may be defined as to certain any distance

D _i(P)＝‖P-C _i‖i＝1，2，...，n (2)

(x y), introduces the Euclidean distance of the input space and improves, then to formula (2) because piece image can be seen a two-dimensional function f as

$d(X,Y)=\left|\right|X-Y\left|\right|=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(x_i-y_i)}^2---\left(3\right)$

Therefore, the criterion function of classification is

d _i(P)＜d _j(P)，j≠iP∈W _i (4)

4. the average essence classification of data block K-

According to the rough sort result, choose the core of some representative points by the principle of minor increment as cluster, then remaining is treated that branch is assigned to according to judgment criterion and go to finish preliminary classification in all kinds of.Recomputate each cluster centre afterwards, carry out next iteration, until satisfying accuracy requirement.Wherein, the comparative result of smart sorted plot quality and unsupervised classification and maximum likelihood method classification is as shown in table 1.Calculating formula is:

N sample is divided into K cluster, calculates the average M of each cluster ₁, M ₂..., M _kAnd J _c,

$M_i=\frac{1}{N_i}\underset{y{\∈w}_i}{\mathrm{\Σ}}y---\left(5\right)$

$J_e=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\underset{{y\∈w}_i}{\mathrm{\Σ}}\left|\right|y-M_i|{|}^2---\left(6\right)$

N in the formula _i---i cluster W _iIn number of samples;

M _i---the average of sample;

J _c---the error sum of squares criterion function of sample set y and classification collection Ω.

The precision of the various sorting techniques of table 1 and speed contrast

Method	(s) consuming time	Classification accuracy rate (%)
							Soil (farmland)	Buildings	Road	The greenery patches	Water body
		Unsupervised classification	22	75	92.5	86.5						82.2	67
Maximum likelihood method	125						82	94	78.6	80.6	75.4
		The classification of associating Weighted random	95	89	95	92.3						87.5	82.2

Formula (5) has in fact illustrated the choosing of initial mean vectors, preliminary classification has been recomputated cluster centre M _l, and then revise cluster centre, so that carry out next iteration, its process is.

(1) chooses a training sample y, establish y and be in W _lIn, if, N _l=1, class W is described _lDo not exist, recomputate by formula (5), (6), on the contrary y from W _iPut into W _jIn, be calculated as follows the change amount of error sum of squares:

${\mathrm{\ΔJ}}_J=J_i-J_i^{\′}=\frac{N_i}{N_i-1}{\left|\right|y-M_i\left|\right|}^2---\left(7\right)$

${\mathrm{\ΔJ}}_k=J_k^{\′}-J_k=\frac{N_k}{N_k+1}{\left|\right|y-M_k\left|\right|}^2---\left(8\right)$

If Δ J _l＞Δ J _kSet up, then allow to change over to next sample, (1) selectes new samples again set by step, otherwise, carry out next step.

(2) for all j, if ${\mathrm{\ΔJ}}_k=\min \left[\underset{J^{\′}}{\mathrm{\Δ}{J}_J}\right],$ Then y from W _iMove on to W _jIn go.

(3) recomputate M _iAnd M _kValue, and revise J _c

(4) if after subsequent iteration arrives N time, J _eNo longer change, then iteration stopping is determined classification, otherwise changes step (1).

5. the splicing again of classified image data block

The classification block of high spectrum image data block is spliced again by the aforesaid order of cutting apart, form new classified image, the result is shown in Fig. 2 (d).

Claims (1)

1, a kind of high-spectrum remote sensing associating Weighted random sorting technique is characterized in that comprising following concrete steps:

1) extraction of high spectrum image data characteristic: according to demands of applications, high spectrum image file to be classified behind the acquisition registration, the interface function that provides by remote sensing image processing and development platform ENVI/IDL, obtain their performance data, comprise file size, the wave band number of spatial resolution and image high-spectrum remote sensing;

2) deblocking of high spectrum image: the spatial resolution with high spectrum image is a standard, api function by WINDOWS2000 operating system, automatically after obtaining the internal memory and CPU processing speed index of computing machine, it is divided into the data block of several successive according to the big young pathbreaker of the data volume of high spectrum image file correspondence, and each data block is numbered, the number of data block is high spectrum image data volume/calculator memory amount;

3) corresponding high spectrum image data block is called in circulation, in each data block, establishes training sample and is divided into N classification W ₁, W ₂..., W _n, according to formula $C_i=\frac{1}{m_i}\underset{j=1}{\overset{m_i}{\mathrm{\Σ}}}{X}_i$ Calculate the center C of classified sample set _iAnd radius,

And remember classification number, and wherein, m _iBe W _iThe number of samples of class, X _iBe total sample number, then with all sample points as treating branch, by improved self-adaptation minimum distance method criterion function d _i(P)＜d _j(P), j ≠ i  P ∈ W _iCarry out the rough sort first time, P is a picture element to be classified here;

4) according to the rough sort result, in conjunction with original image, given smart nicety of grading is divided into K cluster with N sample, according to formula $M_i=\frac{1}{N_i}\underset{y\∈w_i}{\mathrm{\Σ}}y,J_e=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\underset{y\∈w_i}{\mathrm{\Σ}}{\left|\right|y-M_i\left|\right|}^2$ And $\mathrm{\Δ}{J}_k=\underset{J}{\min \left[\mathrm{\Δ}{J}_j\right]}$ Carry out the average smart classified calculating of K-of data block sample, here M _iBe the average of each cluster, N _iBe i cluster W _iIn number of samples, J _eError sum of squares criterion function for sample set y and classification collection Ω; During this time, must iterate, reach classificating requirement up to the nicety of grading of each sample set;

5) splicing again of classified image data block: the classification block of high spectrum image data block is spliced again by the aforesaid order of cutting apart, form new classified image.

Images