CN101710387A - Intelligent method for classifying high-resolution remote sensing images - Google Patents

Abstract

In view of the characteristics of the high-resolution remote sensing images, the invention provides a practical intelligent method for classifying images, which comprises the following six steps: generating the image segmentation result of a full-color image; acquiring the segmentation result of a multispectral image by utilizing space mapping; determining whether the segmented regions of the full-colour image are insufficiently segmented; resegmenting the detected insufficiently-segmented region; generating the regional feature space; and classifying the image by using a classifier. The invention solves the problem that the insufficiently-segmented regions frequently influence the image classification precision in the image classification process. The method is suitable for high-resolution images from remote sensing satellites, such as IKONOS, QUICKBID and the like, and plays an important role in extracting application messages, such as target identification, resource environment survey, land utilization trends, disaster monitoring, disaster situation evaluation and the like.

Description

A kind of intelligent method for classifying high-resolution remote sensing images

Technical field

The present invention is a kind of intelligent method for classifying high-resolution remote sensing images of practicality, be applicable to high spatial resolution satellite remote sensing images such as IKONOS and QUICKBIRD, and be widely used in the investigation of Target Recognition, resource environment, soil utilization dynamically, research and applications such as disaster assessment.

Background technology

The classification of remote sensing images is one of basic problems of remote sensing images information processing and application, and the decipher of remotely-sensed data, analysis and ground are learned to use often all to need to handle by classification of remote-sensing images and realized.

In the evolution of remote sensing image classification technology, development along with the high resolution sensor technology, the spatial resolution of satellite remote sensing images is more and more higher, very big change has also taken place in the content and form of image thereupon, image space information is more and more abundanter, and this has brought new challenge for traditional image classification technology.Traditional image classification technical research be the sorting technique of a large amount of low resolution remote sensing images that exist of mixed pixel, the pixel characteristic that this technology is extracted is more single, only comprises the information of image spectrum.Along with the raising of remote sensing pattern space resolution, remote sensing images have had gem-pure structure, and image pixel no longer has been the elementary cell of image, can cause a large amount of mistake branches if still use with pixel as the sorting algorithm of research unit.In addition, the raising of image spatial resolution is accompanied by weakening of image randomness, simple low order markov random file model can not effectively have been simulated a lot of behaviors of high-definition picture, and traditional image classification algorithms is applied in the high-resolution remote sensing image has too many difficulties to cope with.

The present invention has utilized specific image format of high-resolution remote sensing image and picture material, thereby adopts no processing policy to improve the automatization level and the nicety of grading of the classification of remote sensing images at different data sources.

Summary of the invention

The present invention is a kind of intelligent method for classifying of high-resolution remote sensing image, by setting several simple sorting parameters, makes computing machine to be divided into different zones to high-resolution remote sensing image automatically according to different characters of ground object.Same zone has identical characters of ground object, and gives identical color.

Concrete method step is:

The first step: cut apart full-colour image, generate multiple dimensioned expression, and according to the full-colour image segmentation result under suitable yardstick of goal in research selection.

1. the present invention selects to estimate based on mathematical morphology operators the watershed divide image segmentation algorithm of ground object location.This algorithm has improved traditional watershed algorithm, principle is simple, fast operation, solved the over-segmentation problem preferably, can obtain the segmentation result of image in the short period of time, this point is extremely important in the research of the remote sensing images segmentation problem of big data quantity.Algorithm can access the cut zone border of single pixel, and image segmentation result can be used for the identical marking image of image size and represent, these advantages are obtained the multispectral image segmentation result for the follow-up spatial mappings of further processing condition is provided.

2. utilize the Gaussian filter of different scale original image filtering result to be formed the multiple dimensioned expression of image, the image employing watershed algorithm of different scale can obtain the segmentation result under the different scale, can be according to the image segmentation result under concrete suitable yardstick of goal in research selection.

Second step: the ecbatic of cutting apart according to full-colour image utilizes the spatial mappings technology to obtain the segmentation result of low spatial resolution multispectral image.

1. the mapping relations in space are expressed as:

Multispectral image is upper is changed to (i, pixel j) has been represented image block following on the full-colour image:

2. according to this mapping relations, the pixel of the optional position on the multispectral image all can find the image block above the corresponding with it full-colour image.

3. according to the segmentation result of full-colour image in the first step, determine on the spectrum picture segmentation result of pixel arbitrarily: find out the image block in the full-colour image of this pixel correspondence, and select the dividing mark value of that maximum dividing mark value of number as this pixel on the spectrum picture.

The 3rd step: according to the multispectral data of each cut zone, whether the segmentation result of full-colour image is differentiated on automated intelligent ground one by one correct: promptly cut zone is single atural object, still mixes atural object (needing in this case to divide again).

Whether adopt following step to differentiate cut zone for cut zone arbitrarily is single atural object:

1. the hypothesis zone is single atural object, utilizes the regional multispectral information that obtains in second step, finds the averaged spectrum center and the distribution parameter thereof of this cut zone.

Wherein, the calculation procedure of averaged spectrum center and distribution parameter is:

[1] calculates initial cluster center

[2] calculate each some P _iDistance center point apart from r _i, estimate statistical parameter then

Thereby obtained probability density function f (r)

$r_i=\sqrt{{(x-\hat{x})}^2+{(y-\hat{y})}^2}$

${\hat{R}}^2=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{r_i}^2$

[3] calculate each some P according to probability density function f (r) _iThe contribution to central point (perhaps being probability of happening) p _i

$p\left(x\right)=\frac{1}{\mathrm{\σ}\sqrt{2\mathrm{\π}}}{e}^{\frac{-{(x-\mathrm{\μ})}^2}{2{\mathrm{\σ}}^2}}$

[4] each some contribution p different to the center _iCalculate new classification center

$\left\{\begin{array}{c}\hat{x}=\frac{1}{N}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_i=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\frac{1}{N}\×x_i=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{p}_i{x}_i\\ \hat{y}=\frac{1}{N}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{y}_i=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\frac{1}{N}\×y_i=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{p}_i{y}_i\end{array}\right.$

[5] distance of calculating new and old classification center judges whether iteration stops, otherwise continues 2 work

2. the hypothesis zone is to mix subdivisible ground object area, utilizes regional multispectral information equally, does the divisional processing again of two classes in multispectral feature space, and calculates the averaged spectrum center and the distribution parameter thereof of two classes respectively.

Wherein, two classes of cut zone again the divisional processing step be:

[1] determines initial sets S ₁And S ₂: at first distribute an initial point S to each set ₁={ P} and S ₂={ P ' } calculates the central point distance of each point and these two set among the remaining point set P then respectively, distributes according to bee-line to belong to which set.

[2] according to the algorithm that proposes above, difference set of computations S ₁The classification center

And S set ₁Some probability of happening function f ₂(r).Set of computations S ₂The classification center

And S set ₂Some probability of happening function f ₂(r).

[3] repartition two S set ₁And S ₂, for each some P _i∈ P calculates respectively and S set ₁The center

Apart from r ₁And calculating and S set ₂The center

Apart from r ₂Compare f ₁(r ₁) and f ₂(r ₂) size, if f ₁(r ₁) greater than f ₂(r ₂), think a P so _iFor S set ₁The probability that takes place is bigger, so P _i∈ S ₁'.Otherwise P _i∈ S ₂'.Will form new S set like this ₁' and S ₂'.

[4] newer S set ₁' and S ₂' whether with original S set ₁And S ₂Identical, if difference then continue step 2.

Wherein, 1 content during the calculating of averaged spectrum center and distribution thereof goes on foot with the 3rd.

3. calculate the quantitative differences of ideal distribution and actual distribution according to following formula:

$d=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}|F\left(a_i\right)-\hat{F}\left(a_i\right)|=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}|F\left(a_i\right)-\frac{M\left(a_i\right)}{N}|$

Wherein F (x) is the gaussian probability distribution function,

Be the probability distribution function that actual count obtains, a _iBe sampled point, M is the number of sampled point, generally gets 4 and is advisable.

4. calculate respectively, the quantitative differences in the quantitative differences and 2 in 1 relatively, the less situation of selection differences makes decisions, if the quantitative differences that is: in 2 is bigger, illustrates that so it is irrational that cut zone is divided again.So this cut zone is differentiated is single ground object area.

The 4th step: for cut zone is the situation of mixing atural object, cuts apart under the driving of full-colour image data again, and the segmentation result replay is mapped to multispectral image gets on.

Those that the 4th step was obtained mix the cut zone of atural objects, carry out divisional processing again, and its step is as follows:

1. the mapping of multispectral image data qualification result on full-colour image.From full-colour image, extract region R p to be split, with classified regions Rm (Rm be decision-making then be divided into the multispectral zone of the correspondence of two classes) in the multispectral image, Rm is not that those pixels of 0 are represented corresponding multispectral data, wherein those pixel representatives of Rm=1 belong to the first kind, and those pixel representatives of Rm=2 belong to second class.Those values of non-zero among the Rm are mapped directly to the Rp zone of full-colour image according to the proportionate relationship of resolution.Those pixels of Rp non-zero are made up of three parts like this, are respectively Rp=1, the first kind of representative mapping; Rp=2, second class after the representative mapping; Rp=-1, when those pixels that representative is not mapped to, these pixels are mapped in the multispectral image owing to initial full-colour image segmentation result often, the position of the multispectral pixel correspondence that those of ninsolid color are rejected.

2. first kind sets of pixel values C ₁=I (x, y) | Rp (x, y)=1}, add up its grey scale pixel value center I ₁And the parameter of calculating distribution F

The second class sets of pixel values C ₂=I (x, y) | Rp (x, y)=2}, statistical pixel gray-scale value center I ₂, and calculate distribution parameter

For unfiled sets of pixel values C ₀=I (x, y) | Rp (x, y)=-1}, wherein any one some P (x, y) ∈ C ₀, calculate f ₁(P-I ₁) and f ₂(P-I ₂) size judges that the probability that belongs to which set is bigger, formed new set C like this ₁And C ₂

3. define a neighborhood system

At new set C ₂And C ₂Down, if for a P (x, y) ∈ C ₁, and have some Q (x, y) ∈ C ₁, (x is y) in that (x is in the neighborhood system at center y), thinks that so (x y) belongs to C to a P with a P as fruit dot Q ₁And C ₂Adjacent point.(x y), constitutes new set C to find all such some P ₀

4. just formed new two classes set C like this ₁And C ₂, and C is gathered in the new classification for the treatment of ₀Recomputate the classification center and reclassify according to step 2 then.Till twice iteration front and back set content no longer changes.

The 5th step: extract the provincial characteristics of cut zone, form feature space.

Process through the 4th step and the 5th step, the zone of mixing atural object in the image segmentation process has obtained correct branch again, the 5th step mainly was to extract provincial characteristics, this method adopts the mode of the average multispectral centre data feature of the regional multispectral image of extraction as provincial characteristics, can simplify calculating like this, improve counting yield.

The 6th step: the sorter according to design is realized classification, and obtains sorting result.

The design part of sorter among the present invention, employing be average drifting Cluster Classification device, this sorter utilization be that the Density Distribution of point in feature space finished Cluster Classification automatically, method simple computation speed is fast.

Description of drawings

Fig. 1 is the designed classification process figure of the present invention.

Fig. 2 is specific embodiments of the invention.

Embodiment

The high spatial resolution remote sense image classification result of Fig. 2 for adopting the inventive method to obtain.

It is 1024 * 1024 view data that test has intercepted the full-colour image size.

Figure a is the full-colour image of intercepting, and figure b is the result after the image segmentation.Come as can be seen from the image that merges, there is the phenomenon of regional less divided in image segmentation result.

Figure c is the result images that the zone cut apart is endowed different colours.Here use " circle " to mark the zone of visual less divided, the situation of less divided can obviously be found out from the image that merges.

Figure d is the figure as a result behind the less divided zone subdivision.Can see that the zone of four apparent in view less divideds has all obtained dividing accurately again among the figure c in figure d.

Figure e is the image after final divided area is endowed multispectral feature.The result who divides again in the less divided zone and divide the multispectral feature difference of latter two subregion can be more here from seeing.

Figure f is final classification results figure.

Claims (1)

1. intelligent method for classifying high-resolution remote sensing images, the specific implementation step is:

The first step: cut apart full-colour image, generate multiple dimensioned expression, and according to the full-colour image segmentation result under suitable yardstick of goal in research selection.

Second step: the segmentation result according to full-colour image utilizes the spatial mappings technology to obtain the segmentation result of the multispectral image of low spatial resolution

The 3rd step: according to the multispectral data of each cut zone, whether the segmentation result of full-colour image is differentiated on automated intelligent ground one by one correct: promptly cut zone is single atural object, still mixes atural object (needing in this case to divide again).

The 4th step: for cut zone is the situation of mixing atural object, cuts apart under the driving of full-colour image data again, and the segmentation result replay is mapped to multispectral image gets on.

The 5th step: extract the provincial characteristics of cut zone, form feature space.

The 6th step: the sorter according to design is realized classification, and obtains sorting result.

Images